Phase generated carrier demodulation method based on end-to-end neural network

Abstract

The invention discloses a phase generated carrier demodulation method based on an end-to-end neural network, belongs to the technical field of optical fiber interference sensing signal processing, and is used for a phase demodulation process of an optical fiber interference signal. According to the method, a data set with a physical label is generated through simulation, a deep neural network is trained to realize direct regression prediction of the phase, and dependence on multi-stage parameters and a complex demodulation process in a traditional phase generation carrier demodulation method is avoided. By constructing a regression network model comprising an input layer, a hidden layer and an output layer, high-precision effective demodulation can be performed on interference signals under different parameter conditions. The method has strong robustness, is suitable for various optical fiber sensing application scenes, and especially has good practical value in a dynamic environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical fiber interference sensing signal processing, and specifically relates to a phase generation carrier demodulation method based on an end-to-end neural network. Background Art

[0002] Fiber-optic interferometric sensing technology, due to its high sensitivity, strong resistance to electromagnetic interference, and long transmission distances, is widely used in bridge structure monitoring, oil and gas pipeline safety testing, and geological disaster early warning. Phase Generated Carrier (PGC) technology, a mainstream phase demodulation method, offers excellent linearity and dynamic range, making it the current mainstream demodulation approach. This technology introduces a high-frequency carrier during the modulation process, shifting low-frequency phase information to a high-frequency band. The original phase is then restored through a series of processes, including frequency mixing and low-pass filtering.

[0003] However, traditional PGC methods are highly dependent on system parameters, requiring strict calibration of carrier frequency and modulation depth. The demodulation process also involves multiple nonlinear processing steps, resulting in system complexity, poor adaptability, and sensitivity to noise and system non-idealities. Furthermore, the demodulation algorithm is not portable across diverse system architectures and sensing environments.

[0004] With the development of deep learning technology, neural networks have demonstrated superior performance in time series signal modeling, nonlinear mapping, and pattern recognition. Introducing deep learning into fiber interferometry signal processing promises to overcome the limitations of traditional methods, establishing a direct mapping relationship between interfering light intensity and the measured signal, avoiding intermediate transformations and improving system robustness and demodulation efficiency. Therefore, developing a novel phase demodulation method incorporating deep neural networks has clear practical needs and engineering significance. Summary of the Invention

[0005] In order to address the deficiencies of the above-mentioned prior art, the present invention provides a phase generation carrier demodulation technology based on an end-to-end neural network, so as to directly and quickly extract phase information from the original interference signal.

[0006] The purpose of the present invention can be achieved through the following technical solutions: The phase generation carrier demodulation method based on end-to-end neural network includes the following steps: Step 1: Randomly generate parameters using the interferometer output formula in the PGC demodulation sensor system to simulate the one-dimensional continuous output light intensity; Step 2: Use the generated parameters as labels and the corresponding simulation signals as data to form a data set; Step 3: Set the structure, parameters, activation function, and loss function of the neural network model, and send the data set to the model for training; Step 4: Use the evaluation index to determine whether the error meets the requirements. If so, save the model. Otherwise, readjust the model and then train the model again. Step 5: The interference signal detected by the actual fiber optic sensing system is sent to the saved model for prediction. The detected data is corrected according to the prediction result to restore the signal to be measured.

[0007] As a further solution of the present invention, the specific operations of step 1 are as follows: According to the interferometer output formula, the output signal parameters are randomly generated, including DC component, AC component, phase difference, carrier frequency, and phase modulation depth. The parameter generation range must meet the requirements of randomness and actual distribution. The interferometer output formula is: ; Where A is the DC component in the interferometer output signal; B is the amplitude of the AC component in the interferometer output signal; C is the phase modulation depth of the interferometer, ω is the carrier signal frequency; φ(t) is the signal to be measured; and φ0 is the initial phase difference of the system.

[0008] As a further solution of the present invention, the above-mentioned signal to be measured is expressed as: Where D is the amplitude of the signal to be measured, f s is the frequency (range) of the signal to be measured.

[0009] As a further solution of the present invention, the specific operations of step 2 are as follows: Randomly generated parameters are used as labels, including DC component, AC term amplitude, phase difference, carrier frequency, and phase modulation depth; The simulated signal is used as data to form a data set and used as a training data set. A test data set is generated in the same way. 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.

[0010] As a further solution of the present invention, the specific operations of step 3 are as follows: The neural network model structure 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 the prediction result of the regression task, that is, the phase to be measured, in the output layer.

[0011] As a further solution of the present invention, the neural network model structure in step 3 is as follows: Input layer: The model receives an input whose number of features is determined by the number of samples; First hidden layer: The input data undergoes feature extraction through the first linear transformation and then outputs through the activation function. The first hidden layer has 128 neurons; Second hidden layer: The data undergoes feature extraction through a second linear transformation and is then output through an activation function. The second hidden layer has 64 neurons. Output layer: The data undergoes a third linear transformation and the output is the predicted result, i.e. the phase to be measured.

[0012] As a further solution of the present invention, the specific operation of step 4 is: The evaluation index uses variance. When the demodulation result can be accurate to two decimal places, it meets the error requirement. Otherwise, the model is adjusted. Adjustments include adjusting the model's structure and parameters and resetting the data set. Adjustments to the model's structure include the number of model layers and the number of neurons in each layer; adjustments to the model's parameters include the learning rate, batch size, and number of training rounds; and data set settings include expanding the data set and adjusting the data set distribution range.

[0013] Beneficial effects of the present invention: The present invention covers a variety of parameter combinations through large-scale data simulation, making the model robust and eliminating the need for precise modeling of system parameters; The present invention directly establishes the mapping relationship between the interference signal and the phase quantity through a neural network model, avoiding the traditional PGC demodulation method's reliance on complex algorithm chain processes (such as harmonic extraction and filtering), and improving the overall stability and simplicity of the demodulation process; The present invention can be applied to various phase-generated carrier technology solutions and has universal applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The present invention will be further described below with reference to the accompanying drawings.

[0015] Figure 1 1 is a flow chart of a phase generation carrier demodulation method based on an end-to-end neural network according to the present invention; Figure 2 It is a structural diagram of the neural network model. DETAILED DESCRIPTION

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0017] A phase-generated carrier demodulation method based on an end-to-end neural network. This method directly obtains the signal to be measured by performing end-to-end phase extraction without complex signal processing. Specifically, such as Figure 1 As shown, the following steps are included: Step 1: Randomly generate parameters using the interferometer output formula in the PGC demodulation sensor system to simulate the one-dimensional continuous output light intensity; The specific operations of step 1 are as follows: According to the interferometer output formula, the output signal parameters are randomly generated, including DC component, AC component, phase difference, carrier frequency and phase modulation depth. The parameter generation range must meet the requirements of randomness and actual distribution. The interferometer output formula in step 1 is as follows: Where A is the DC component in the interferometer output signal; B is the AC component amplitude in the interferometer output signal; C is the phase modulation depth of the interferometer, ω is the carrier signal frequency; φ(t) is the signal to be measured; φ0 is the initial phase difference of the system; The signal to be measured is expressed as: Where D is the amplitude of the signal to be measured, f s is the frequency (range) of the signal to be measured; Step 2: Use the generated parameters as labels and the corresponding simulation signals as data to form a data set; Specifically, randomly generated parameters are used as labels, including DC component, AC term amplitude, phase difference, carrier frequency, and phase modulation depth; Use the simulated signal as data to form a data set, which is used as the training data set. A test data set is generated in the same way. 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. Step 3: Set the structure, parameters, activation function, and loss function of the neural network model, and send the data set to the model for training; Specifically, the neural network model structure in step 3 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 data, extracts features through two hidden layers, and finally outputs the prediction result of the regression task, i.e., the phase to be measured, in the output layer. like Figure 2 As shown, the neural network model structure is as follows: Input layer: The model receives an input whose number of features is determined by the number of samples; First hidden layer: The input data undergoes feature extraction through the first linear transformation and then outputs through the activation function. The first hidden layer has 128 neurons; Second hidden layer: The data undergoes feature extraction through a second linear transformation and is then output through an activation function. The second hidden layer has 64 neurons. Output layer: The data undergoes a third linear transformation and the output is the predicted result, i.e. the phase to be measured.

[0018] Step 4: Use the evaluation index to determine whether the error meets the requirements. If so, save the model. Otherwise, readjust the model and then train the model again. This step uses the evaluation indicators to determine whether the error meets the requirements. If it does, the model is saved. Otherwise, the model needs to be readjusted and trained again. 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 adjusted, including adjusting the model structure and parameters and resetting the data set. The structural adjustment of the model includes the number of layers and the number of neurons in each layer; the adjustment of model parameters includes learning rate, batch size, and number of training rounds; the setting of the data set includes expanding the data set and adjusting the distribution range of the data set.

[0019] Step 5: The interference signal detected by the actual fiber optic sensing system is sent to the saved model for prediction. The detected data is corrected according to the prediction result to restore the signal to be measured.

[0020] Here are some examples: The interferometer output formula in step 1 is as follows: Where A is the DC component in the interferometer output signal; B is the AC component amplitude in the interferometer output signal; C is the phase modulation depth of the interferometer, ω is the carrier signal frequency; φ(t) is the signal to be measured; φ0 is the initial phase difference of the system; The signal to be measured is expressed as: Where D is the amplitude of the signal to be measured, f s is the frequency (range) of the signal to be measured; Set the sampling rate to 100kHz; A (range 0.5-2.0); B (range 0.5-2.0); C (range 1.0–3.5); D (range 0.01-0.1).

[0021] In step 2: A simulation platform built using the Python programming language generates a large number of data samples with different parameter combinations based on the aforementioned interference signal model. Each set of samples contains the input light intensity and the corresponding φ(t) label value. Gaussian white noise can be introduced into the training data to enhance robustness.

[0022] In step 3: A feedforward neural network was constructed, where the input passed through a first hidden layer of 128 nodes and a second hidden layer of 64 nodes and then output a single phase value. The activation function uses ReLU, and the output layer is a linear regression node; The model was implemented using the PyTorch framework. The mean square error (MSE) loss function was used, the optimizer was Adam, the initial learning rate was 0.001, the number of training rounds was 200, and the batch size was 64.

[0023] In step 4: The test set is used to monitor errors in real time and evaluate model fitting performance. Model training is considered satisfactory when the MSE error is less than 1e-3 and the mean absolute error (MAE) is accurate to within 0.01 rad.

[0024] In step 5: The interference signal detected by the actual fiber optic sensing system is sent to the saved model for prediction. The detected data is corrected according to the prediction result to restore the signal to be measured.

[0025] The above contents are merely examples and explanations of the present invention. Those skilled in the art may make various modifications or additions to the described specific embodiments or replace them in similar ways. As long as they do not deviate from the invention or exceed the scope defined by the claims, they should all fall within the scope of protection of the present invention.

Claims

1. A phase generation carrier demodulation method based on an end-to-end neural network, characterized in that: The steps include: Step 1: Randomly generate parameters using the interferometer output formula in the PGC demodulation sensor system to simulate the one-dimensional continuous output light intensity; Step 2: Use the generated parameters as labels and the corresponding simulation signals as data to form a data set; Step 3: Set the structure, parameters, activation function, and loss function of the neural network model, and send the data set to the model for training; Step 4: Use the evaluation index to determine whether the error meets the requirements. If so, save the model. Otherwise, readjust the model and then train the model again. Step 5: The interference signal detected by the actual fiber optic sensing system is sent to the saved model for prediction. The detected data is corrected according to the prediction result to restore the signal to be measured.

2. The phase generation carrier demodulation method based on end-to-end neural network according to claim 1 is characterized in that The specific operations of step 1 are as follows: According to the interferometer output formula, the output signal parameters are randomly generated, including DC component, AC component, phase difference, carrier frequency, and phase modulation depth. The parameter generation range must meet the requirements of randomness and actual distribution. The interferometer output formula is: ; Where A is the DC component in the interferometer output signal; B is the amplitude of the AC component in the interferometer output signal; C is the phase modulation depth of the interferometer, ω is the carrier signal frequency; φ(t) is the signal to be measured; and φ0 is the initial phase difference of the system.

3. The phase generation carrier demodulation method based on end-to-end neural network according to claim 2 is characterized in that The signal to be measured is expressed as: Where D is the amplitude of the signal to be measured, f s is the frequency range of the signal to be measured.

4. The phase generation carrier demodulation method based on end-to-end neural network according to claim 1 is characterized in that The specific operations of step 2 are as follows: Randomly generated parameters are used as labels, including DC component, AC term amplitude, phase difference, carrier frequency, and phase modulation depth; The simulated signal is used as data to form a data set and used as a training data set. A test data set is generated in the same way. 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.

5. The phase generation carrier demodulation method based on end-to-end neural network according to claim 1 is characterized in that The specific operations of step 3 are as follows: The neural network model structure 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 the prediction result of the regression task, that is, the phase to be measured, in the output layer.

6. The phase generation carrier demodulation method based on end-to-end neural network according to claim 5 is characterized in that The neural network model structure in step 3 is as follows: Input layer: The model receives an input whose number of features is determined by the number of samples; First hidden layer: The input data undergoes feature extraction through the first linear transformation and then outputs through the activation function. The first hidden layer has 128 neurons; Second hidden layer: The data undergoes feature extraction through a second linear transformation and is then output through an activation function. The second hidden layer has 64 neurons. Output layer: The data undergoes a third linear transformation and the output is the predicted result, i.e. the phase to be measured.

7. The phase generation carrier demodulation method based on end-to-end neural network according to claim 1 is characterized in that The specific operations of step 4 are: The evaluation index uses variance. When the demodulation result can be accurate to two decimal places, it meets the error requirement. Otherwise, the model is adjusted. Adjustments include adjusting the model's structure and parameters and resetting the data set. Adjustments to the model's structure include the number of model layers and the number of neurons in each layer; adjustments to the model's parameters include the learning rate, batch size, and number of training rounds; and data set settings include expanding the data set and adjusting the data set distribution range.

Images