A fiber optic signal demodulation algorithm based on fully connected neural network

Abstract

The present invention belongs to the field of fiber Bragg grating sensing phase demodulation technology, and specifically relates to a fiber signal demodulation algorithm based on a fully connected neural network, and the steps are as follows: for a sensing system composed of a phase modulation type Michelson interferometer using a 3×3 coupler, randomly generate parameters according to the interferometer output formula, and simulate three-way one-dimensional continuous output light intensity; use the generated parameters as labels, and use the obtained simulation signals as data to form a data set; set the structure, parameters, activation function, and loss function of the fully connected neural network model, and send the data set to the model for training; judge whether the error meets the requirements through evaluation indicators, and if it meets the requirements, save the model, otherwise readjust the model and retrain the model; send the actual detected three-way interference output light intensity to the saved model for prediction, and correct the detected data according to the prediction results to restore the signal to be measured. The present invention reduces the time of data processing and improves the demodulation speed.

Description

Technical Field

[0001] The invention belongs to the technical field of fiber grating sensing phase demodulation, and in particular relates to a fiber optic signal demodulation algorithm based on a fully connected neural network. Background Art

[0002] In the field of fiber Bragg grating sensing, interferometric phase demodulation technology is favored for its high sensitivity and good stability. It aims to detect tiny physical changes in the external environment, such as temperature, stress, and sound waves. At present, many high-precision interferometric phase demodulation schemes have been proposed. In order to ensure high-precision demodulation effects, the amount of data processing and the data processing flow are often increased, resulting in a decrease in demodulation speed. With the expansion of application scenarios, there are new requirements for demodulation speed efficiency. In order to solve this problem, the demodulation speed is improved while ensuring accuracy as much as possible. A fiber optic signal demodulation algorithm based on a fully connected neural network was invented.

[0003] In the prior art, the 3×3 coupler phase demodulation technology based on ellipse fitting is one of the most common phase demodulation methods. This method has high demodulation accuracy and can effectively eliminate the influence of 3×3 coupler asymmetry. However, the traditional ellipse fitting method uses the least squares method, which requires huge data processing and reduces the demodulation speed. It is not suitable for demodulation scenarios that require high speed. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a fiber optic signal demodulation algorithm based on a fully connected neural network, which not only ensures the demodulation accuracy but also improves the demodulation speed.

[0005] The present invention is achieved through the following technical solutions:

[0006] A fiber optic signal demodulation algorithm based on a fully connected neural network.

[0007] The steps include:

[0008] Step S1, for a sensing system formed by a phase modulation Michelson interferometer composed of 3×3 couplers, randomly generating parameters according to an interferometer output formula, and simulating three-way one-dimensional continuous output light intensity;

[0009] Step S2, using the generated parameters as labels and the corresponding simulation signals as data to form a data set;

[0010] Step S3, setting the structure, parameters, activation function, and loss function of the fully connected neural network model, and sending the data set to the model for training;

[0011] Step S4, judging whether the error meets the requirement through the evaluation index, if so, saving the model, otherwise readjusting the model and then training the model;

[0012] Step S5, the interference three-way output light intensity detected by the actual optical fiber sensing system is sent to the saved model for prediction, and the detected data is corrected according to the prediction result to restore the signal to be measured.

[0013] Furthermore, the specific operations of step S1 are as follows:

[0014] According to the interferometer output formula, three-way output light intensity parameters are randomly generated, including DC component, AC term amplitude, and phase difference. The parameter generation range must meet the requirements of randomness and actual distribution.

[0015] The interferometer output formula in step 1 is as follows:

[0016]

[0017] in, A 1. A 2. A 3 are the DC components of the three randomly generated signals, B 1. B 2. B 3 are the AC amplitudes of the three randomly generated signals, f 1. f 2. f 3 are the phase differences of the three signals caused by the environment and generated randomly, f （t） is the signal to be tested.

[0018] Furthermore, the signal to be tested can be expressed as:

[0019]

[0020] in f （t） is the signal to be measured, A is the amplitude of the signal to be measured, F s is the frequency of the signal to be measured.

[0021] Furthermore, the specific operations of step S2 are as follows:

[0022] The randomly generated parameters obtained in step S1 are used as labels, including the DC component, AC term amplitude, phase difference, and the simulated signal as data to form a data set as a training data set. A test data set is generated in the same way, where the training data set is used for model training and the test data set is used to test the model performance and evaluate whether the model meets the requirements.

[0023] Furthermore, the fully connected neural network model structure in step S3 is as follows:

[0024] The fully connected neural network model adopts a supervised learning method and consists of four parts: an input layer, a first hidden layer, a second hidden layer, and an output layer. The input layer receives a piece of data, extracts features through two hidden layers, and finally outputs three results in the output layer, corresponding to the prediction results of the regression task.

[0025] Furthermore, the fully connected neural network model structure in step S3 is as follows:

[0026] Input layer: The model accepts an input whose number of features is determined by the number of samples;

[0027] First hidden layer: The input data is subjected to a linear transformation for feature extraction, and then activated by the ReLU function. The first hidden layer has 128 neurons.

[0028] Second hidden layer: The data is subjected to the second linear transformation for feature extraction, and then passes through the ReLU activation function. The second hidden layer has 64 neurons;

[0029] Output layer: The data passes through the third linear transformation and outputs three prediction results.

[0030] Furthermore, step S4 determines whether the error meets the requirement through the evaluation index. If it meets the requirement, the model is saved. Otherwise, the model is readjusted and then trained. The specific operations are as follows:

[0031] The evaluation index uses variance to quantify the error. When the demodulation result can be accurate to two decimal places, it meets the error requirements. Otherwise, the model is readjusted, including adjusting the structure and parameters of the fully connected neural network model and setting the data set. The adjustment of the model includes the number of layers of the model and the number of neurons in each layer. The adjustment of the model parameters includes the learning rate, batch size, and the number of training rounds. The setting of the data set includes expanding the data set and adjusting the distribution range of the data set.

[0032] Compared with the prior art, the present invention has the following technical effects:

[0033] (1) The three-way one-dimensional continuous output light intensity in the fiber optic sensing system is simulated and the simulation data is transformed into a tensor suitable for model input. The simulation data input for model training has a reasonable structure and takes into account the characteristic that the 3×3 coupler is difficult to be completely symmetrical.

[0034] (2) The interferometric three-way output light intensity detected by the actual fiber optic sensing system is sent to the saved model for prediction. The prediction speed is more than twice that of the least squares ellipse fitting speed, and the prediction time consumed by the model is more stable. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flow chart of the optical fiber signal demodulation algorithm based on the fully connected neural network of the present invention.

[0036] Figure 2 It is a structural diagram of the fully connected neural network model.

[0037] Figure 3 This is a comparison chart of the time consumed in calculating parameters for 100 groups of data with a length of 10,000.

[0038] Figure 4 It is a comparison chart between the demodulation result and the signal to be measured. DETAILED DESCRIPTION

[0039] The optical fiber signal demodulation algorithm of the present invention is based on a fully connected neural network model, and includes a simulation data generation algorithm and a fully connected neural network model. The three-way output light intensity of the optical fiber sensing system is simulated by randomly generating parameters according to the interferometer output formula, and sent to the fully connected neural network model for training; the interferometer output light intensity detected by the actual optical fiber sensing system is sent to the trained model for prediction, and the returned prediction results are used to correct the detected data and restore the signal to be measured.

[0040] An optical fiber signal demodulation algorithm based on a fully connected neural network comprises the following steps:

[0041] Step S1, for a sensing system formed by a phase modulation Michelson interferometer composed of 3×3 couplers, randomly generating parameters according to an interferometer output formula, and simulating three-way one-dimensional continuous output light intensity;

[0042] Step S2, using the generated parameters as labels and the corresponding simulation signals as data to form a data set;

[0043] Step S3, setting the structure, parameters, activation function, loss function, etc. of the fully connected neural network model, and sending the data set to the model training;

[0044] Step S4, judging whether the error meets the requirement through the evaluation index, if so, saving the model, otherwise readjusting the model and then training the model;

[0045] Step S5, the interference three-way output light intensity detected by the actual optical fiber sensing system is sent to the saved model for prediction, and the detected data is corrected according to the prediction result to restore the signal to be measured.

[0046] Furthermore, in step 1, the three-way one-dimensional continuous output light intensity of the sensing system of the phase-modulated Michelson interferometer composed of a 3×3 coupler is simulated, and the parameters are randomly generated according to the interferometer output formula, as follows:

[0047] According to the interferometer output formula, three output light intensity parameters are randomly generated, including DC component, AC term amplitude, and phase difference. The parameter generation range must meet the requirements of randomness and actual distribution.

[0048] Furthermore, the interferometer output formula in step S1 is as follows:

[0049]

[0050] in, A 1. A 2. A 3 are the DC components of the three randomly generated signals, B 1. B 2. B 3 are the AC amplitudes of the three randomly generated signals, f 1. f 2. f 3 are the phase differences of the three signals caused by the environment and generated randomly, f （t） is the signal to be tested.

[0051] Furthermore, the signal to be tested can be expressed as:

[0052]

[0053] in f （t） is the signal to be measured, A is the amplitude of the signal to be measured, F s is the frequency of the signal to be measured.

[0054] Furthermore, in step S2, the generated parameters are used as labels, and the corresponding simulation signals are used as data to form a data set, as follows:

[0055] The randomly generated parameters obtained in step 1 are used as labels, including the DC component, AC term amplitude, and phase difference, and the simulated signal is used as data to form a data set as a training data set.

[0056] A test dataset is generated in the same way, where the training dataset is used for model training and the test dataset is used to test the model performance and evaluate whether the model meets the requirements.

[0057] Furthermore, the fully connected neural network model structure in step S3 is as follows:

[0058] The model adopts a supervised learning method and consists of four parts: input layer, first hidden layer, second hidden layer, and output layer. The input layer receives a data, extracts features through two hidden layers, and finally outputs three results in the output layer, corresponding to the prediction results of the regression task.

[0059] Furthermore, the fully connected neural network model structure in step S3 is as follows:

[0060] Input layer: The model accepts an input whose number of features is determined by the number of samples.

[0061] First hidden layer: The input data passes through a linear transformation and then a ReLU activation function. This layer has 128 neurons.

[0062] Second hidden layer: The data passes through a second linear transformation and then a ReLU activation function. This layer has 64 neurons.

[0063] Output layer: The data passes through the third linear transformation and outputs three prediction results.

[0064] Example 1

[0065] Combination Figure 1 , a fiber demodulation algorithm based on a fully connected neural network model, including

[0066] Step S1, randomly generating parameters according to the interferometer output formula, and simulating the three-way one-dimensional continuous output light intensity;

[0067] Furthermore, the method steps are as follows:

[0068] A 1. A 2. A 3 are the DC components of the three randomly generated signals, B 1. B 2. B 3 are the AC amplitudes of the three randomly generated signals, f 1. f 2. f 3

[0069] Randomly generate 10,000 sets of data, each set of data includes A 1. A 2. A 3. B 1. B 2. B 3. f 1. f 2. f 3.

[0070] in, A 1.A 2. A 3 (range 1 ± 0.5), B 1. B 2. B 3 (range 1 ± 0.5), f 1 (Range )、 f 2 (Range )、 f 3 (Range ).

[0071] Substitute the measured signal and randomly generated data into the interferometer output formula to obtain the simulated signal.

[0072] Step S2, using the generated parameters as labels and the corresponding simulation signals as data to form a data set;

[0073] Furthermore, the method steps are as follows:

[0074] The randomly generated data in step 1 is used as labels and the corresponding simulation signals as data to form a training data set. The test data set is generated in the same way.

[0075] Step S3, setting the structure, parameters, activation function, loss function, etc. of the fully connected neural network model.

[0076] Convert the data set in step 2 into a tensor suitable for model input and feed it into the model for training;

[0077] The fully connected neural network model has a structure as follows Figure 2 shown.

[0078] The data enters the input layer to extract features, then passes through two hidden layers to further extract features, and finally outputs three prediction results after linear transformation. The two hidden layers of the fully connected neural network model contain a linear transformation Linear and an activation function Relu, and the number of output features is 128 and 64 respectively.

[0079] Step S4: After the model is trained, it is set to evaluation mode and sent to the test set for prediction. The prediction results are compared with the labels for error evaluation. After the evaluation requirements are met, the model is saved. Otherwise, it returns to step S3 to re-adjust the model structure, parameters and data set, and train the adjusted model again.

[0080] In order to better demonstrate the improvement in demodulation speed of the present invention, in this example, 100 sets of simulation data are generated by using step S1, and parameters are obtained using the fully connected neural network model and the least squares ellipse fitting algorithm respectively, and the time consumed is compared. The comparison results are shown in FIG. Figure 3As shown in the figure, the method of predicting parameters using a fully connected neural network model not only has an advantage in average speed, but also consumes more stable time, making the entire demodulation system smoother.

[0081] Step S5, the interference three-way output light intensity detected by the actual optical fiber sensing system is sent to the saved model for prediction, and the detected data is corrected according to the prediction result to restore the signal to be measured.

[0082] Compare the demodulation result with the signal to be tested. The comparison result is as follows: Figure 4 As shown in the figure, it can be seen that the invention basically restores the signal to be measured, wherein the frequency of the signal to be measured is accurately restored, and the amplitude error is small, thus meeting the demodulation requirements.

[0083] In summary, the present invention can effectively simulate the three-way output light intensity of the interferometer, construct a data set in a reasonable way, and send the data set to a fully connected deep network model for training. The present invention is suitable for demodulation scenarios with large data throughput.

Claims

1. A fiber optic signal demodulation algorithm based on a fully connected neural network, characterized in that: The steps include: Step S1, for a sensing system formed by a phase modulation Michelson interferometer composed of 3×3 couplers, randomly generating parameters according to an interferometer output formula, and simulating three-way one-dimensional continuous output light intensity; Step S2, using the generated parameters as labels and the corresponding simulation signals as data to form a data set; Step S3, setting the structure, parameters, activation function, and loss function of the fully connected neural network model, and sending the data set to the model for training; Step S4, judging whether the error meets the requirement through the evaluation index, if so, saving the model, otherwise readjusting the model and then training the model; Step S5, the interference three-way output light intensity detected by the actual optical fiber sensing system is sent to the saved model for prediction, and the detected data is corrected according to the prediction result to restore the signal to be measured; The specific operations of step S1 are as follows: According to the interferometer output formula, three-way output light intensity parameters are randomly generated, including DC component, AC term amplitude, and phase difference. The parameter generation range must meet the requirements of randomness and actual distribution. The interferometer output formula in step S1 is as follows: ， in, A 1. A 2. A 3 are the DC components of the three randomly generated signals, B 1. B 2. B 3 are the AC amplitudes of the three randomly generated signals, φ 1. φ 2. φ 3 are the phase differences of the three signals caused by the environment and are randomly generated, φ （t） is the signal to be tested; The signal to be measured is expressed as: ， where φ （t） is the signal to be measured, A is the amplitude of the signal to be measured, F s is the frequency of the signal to be measured; The randomly generated parameters obtained in step S1 are used as labels, including the DC component, AC term amplitude, phase difference, and the simulated signal as data to form a data set as a training data set. A test data set is generated in the same way, where the training data set is used for model training and the test data set is used to test the model performance and evaluate whether the model meets the requirements.

2. The optical fiber signal demodulation algorithm based on a fully connected neural network according to claim 1, characterized in that: The fully connected neural network model structure in step S3 is as follows: The fully connected neural network model adopts a supervised learning method and consists of four parts: an input layer, a first hidden layer, a second hidden layer, and an output layer. The input layer receives a piece of data, extracts features through two hidden layers, and finally outputs three results in the output layer, corresponding to the prediction results of the regression task.

3. The optical fiber signal demodulation algorithm based on a fully connected neural network according to claim 2, characterized in that: The fully connected neural network model structure in step S3 is as follows: Input layer: The model accepts an input whose number of features is determined by the number of samples; First hidden layer: The input data is subjected to a linear transformation for feature extraction, and then activated by the ReLU function. The first hidden layer has 128 neurons. Second hidden layer: The data is subjected to the second linear transformation for feature extraction, and then passes through the ReLU activation function. The second hidden layer has 64 neurons; Output layer: The data passes through the third linear transformation and outputs three prediction results.

4. The optical fiber signal demodulation algorithm based on a fully connected neural network according to claim 1, characterized in that: Step S4 determines whether the error meets the requirement through the evaluation index. If it does, the model is saved. Otherwise, the model is readjusted and then trained. The specific operations are as follows: The evaluation index uses variance to quantify the error. When the demodulation result can be accurate to two decimal places, it meets the error requirements. Otherwise, the model is readjusted, including adjusting the structure and parameters of the fully connected neural network model, and setting the data set. The adjustment of the model includes the number of layers of the model and the number of neurons in each layer. The adjustment of the model parameters includes the learning rate, batch size, and the number of training rounds. The setting of the data set includes expanding the data set and adjusting the distribution range of the data set.

Images