Fiber-optic accelerometer signal noise suppression method based on deep learning

Abstract

The application relates to the field of optical fiber signal processing, and discloses a deep learning-based optical fiber accelerometer signal noise suppression method, which comprises the following steps: collecting corresponding original noise-containing signals and pure reference signals during the working of an optical fiber accelerometer, modeling noise characteristics, and designing a neural network for suppressing signal noise according to the noise characteristics. The network is run, and the network is optimized according to problems existing in the running process. The application can improve the signal noise suppression effect and efficiency of the optical fiber accelerometer, and improve the accuracy of the monitoring signal output by the optical fiber accelerometer.

Description

Technical Field

[0001] This invention relates to the field of fiber optic signal processing, and in particular to a method for suppressing noise in fiber optic accelerometer signals based on deep learning. Background Technology

[0002] A fiber optic accelerometer is a sensor based on the principle of optical interferometry, measuring acceleration by detecting changes in the phase / intensity of light signals within an optical fiber. Its core structure typically includes an optical fiber coil, a mass block, a laser source, and a photodetector. During operation, fiber optic accelerometers generate optical noise, mechanical noise, and circuit noise. Optical noise includes laser phase noise, shot noise, and polarization fading noise; mechanical noise includes thermomechanical noise caused by microscopic vibrations of the fiber optic cable due to Brownian motion, and acoustic coupling noise generated by ambient sound wave vibrations; circuit noise includes dark current noise from the photodetector and amplifier noise. Noise can be detrimental, affecting measurement accuracy, characteristic frequencies, and system stability. Suppressing noise can transform fiber optic accelerometers from "precision laboratory instruments" into "reliable industrial sensors," a key technological breakthrough for their large-scale application. Therefore, a deep learning-based method for suppressing signal noise in fiber optic accelerometers is proposed. Summary of the Invention

[0003] This invention overcomes the shortcomings of the prior art and provides a method for suppressing signal noise of fiber optic accelerometers based on deep learning.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] The first aspect of this invention provides a method for suppressing noise in fiber optic accelerometer signals based on deep learning, comprising the following steps:

[0006] The original noisy signal and the clean reference signal from the fiber optic accelerometer are collected and combined to model the noise characteristics and construct a target noisy signal library.

[0007] A deep neural signal noise suppression network is designed by combining a target noise information database with the corresponding stored target composite noise signals.

[0008] By combining a deep neural signal noise suppression network, the target composite noisy signal in the target noisy signal library is segmented, and the signal of the target composite noisy signal sample is suppressed at the same time.

[0009] Sample analysis was performed on the target composite noise reduction signal samples, and the deep neural signal noise suppression network was modified and optimized based on the sample analysis results.

[0010] Furthermore, in a preferred embodiment of the present invention, the acquisition of the original noisy signal and the clean reference signal from the fiber optic accelerometer, combined with noise characteristic modeling, to construct a target noisy signal library, specifically involves:

[0011] Obtain the fiber optic accelerometer that needs to be suppressed for signal noise, calibrate it as the target fiber optic accelerometer, run the target fiber optic accelerometer, and collect the original noisy signal of the target fiber optic accelerometer in real time during the operation, and calibrate it as the target original noisy signal;

[0012] The target's original noisy signal includes environmental noise and its own output electrical signal. At the same time, the target fiber optic accelerometer's clean reference signal is collected and calibrated as the target's clean reference signal.

[0013] The target pure reference signal needs to be collected in an anechoic chamber and serves as the reference signal for the target fiber optic accelerometer.

[0014] Spectral analysis is performed on the original noisy target signal to measure its noise characteristics, wherein the noise characteristics of the original noisy target signal include Gaussianity, stationarity, and temperature drift characteristics.

[0015] Based on the noise characteristics of the original noisy signal of the target, noise modeling is performed on the original noisy signal of the target to obtain the noise library of the target fiber optic accelerometer, which is calibrated as the target noise library. The target noise library stores the original noisy signal of the target acquired in real time and the corresponding noise characteristics.

[0016] The target clean reference signal and the target original noisy signal stored in the target noise library are combined and superimposed to form a composite noise. The number of data samples in the target noise library is expanded and determined. The mixing ratio of the superimposed composite noise is dynamically adjusted in the target noise library to ensure that the signal-to-noise ratio of the superimposed composite noise is maintained within a preset value, thus obtaining the target noisy signal library and the target composite noisy signal.

[0017] Furthermore, in a preferred embodiment of the present invention, the design of a deep neural signal noise suppression network by combining the target noise information database and the corresponding stored target composite noisy signal specifically includes:

[0018] A big data network is introduced, and all network architectures of general deep neural networks for noise suppression of target composite noisy signals are retrieved from the big data network. At the same time, the network architecture with the highest historical usage rate is selected and labeled as the target deep neural network architecture.

[0019] The target deep neural network architecture includes an input layer, a feature extraction layer, a regularization layer, and an output layer.

[0020] The architecture design of the target deep neural network includes determining the maximum processable sample dimension of the input layer based on the number of data samples of the target composite noisy signal in the target noisy signal library, while controlling the maximum processable sample dimension of the input layer and the output layer to be equal.

[0021] Within the regularization layer, the regularization method is determined to be total variational regularization, and within the feature extraction layer, a multi-objective loss function is determined. The multi-objective loss function is determined based on the noise characteristics of the original noisy target signal, including a time-domain loss function and a frequency-domain loss function.

[0022] The designed target deep neural network architecture is named Deep Neural Signal Noise Suppression Network.

[0023] Furthermore, in a preferred embodiment of the present invention, the step of combining a deep neural signal noise suppression network to segment the target composite noisy signal within the target noisy signal library, and simultaneously suppressing the signal of the target composite noisy signal samples, specifically involves:

[0024] Within the target noisy signal library, target composite noisy signal samples are extracted, and the target composite noisy signal samples are segmented to obtain target composite noisy signal segmented samples.

[0025] A deep neural signal noise suppression network was run to perform normalized time-frequency domain joint enhancement processing on segmented samples of different target composite noisy signals.

[0026] A bandwidth constraint term is introduced, which is the start and stop of the enhancement process based on the bandwidth energy in the deep neural signal noise suppression network during the normalized time-frequency domain joint enhancement processing of different target composite noisy signal segment samples.

[0027] Real-time bandwidth energy is calculated within the deep neural signal-noise suppression network, and the maximum bandwidth energy and target augmentation time are preset.

[0028] If, within the target enhancement time, the real-time bandwidth energy within the deep neural signal noise suppression network exceeds the maximum bandwidth energy, then the normalized time-frequency domain joint enhancement processing for different target composite noisy signal segment samples is stopped.

[0029] By combining a deep neural signal-noise suppression network, signal-noise suppression processing is performed on segmented samples of different target composite noisy signals after normalized time-frequency domain enhancement processing.

[0030] Furthermore, in a preferred embodiment of the present invention, the step of combining a deep neural signal-noise suppression network to perform signal-noise suppression processing on different target composite noisy signal segment samples after normalized time-frequency domain enhancement processing specifically involves:

[0031] During the target enhancement time, after normalized time-frequency domain joint enhancement processing is performed on the segmented samples of different target composite noisy signals, the segmented samples of target composite noisy signals are trained in segments.

[0032] Among them, segmented training involves using a multi-objective loss function to segment the noisy composite signal of the target, and monitoring the noise frequency of different segments of the noisy composite signal of the target in real time during the segmented denoising process.

[0033] A preset training stop mechanism is provided. The training stop mechanism is as follows: when the noise frequency of the segmented samples of different target composite noisy signals is less than a preset value after five consecutive monitoring, the segmented noise reduction of the target composite noisy signal segments is stopped, and the segmented noise reduction of the target composite noisy signal segments is sampled and synthesized to obtain the initial target composite noise reduction signal sample.

[0034] The surrounding environmental parameters of the target fiber optic accelerometer are monitored in real time, and the rate of change of the surrounding environmental parameters within a predetermined time is calculated. If the rate of change of the surrounding environmental parameters within the predetermined time is greater than a preset value, the initial target composite noise reduction signal sample is fine-tuned in the deep neural signal noise suppression network to obtain the target composite noise reduction signal sample.

[0035] Furthermore, in a preferred embodiment of the present invention, the step of performing sample analysis on the target composite noise-reduced signal samples and optimizing the deep neural signal-noise suppression network based on the sample analysis results specifically includes:

[0036] Sample analysis is performed on the target composite noise-reduced signal sample, wherein the sample analysis is to analyze the distortion rate and waveform fidelity of the target composite noise-reduced signal sample;

[0037] If the distortion rate and waveform fidelity of the target composite noise reduction signal sample are not lower than the preset values, the deep neural signal noise suppression network is calibrated as a qualified deep neural signal noise suppression network.

[0038] If the distortion rate and waveform fidelity of the target composite noise-reduced signal sample are lower than the preset value, a feature frequency protection term is added to the multi-target loss function in the deep neural signal noise suppression network. The feature frequency protection term prioritizes protecting the distortion rate and waveform fidelity of the target composite noise-reduced signal sample, so that the distortion rate and waveform fidelity of the target composite noise-reduced signal sample are not lower than the preset value.

[0039] A qualified deep neural signal noise suppression network is used to perform signal noise suppression processing on all target composite noisy signals in the target noisy signal library.

[0040] A second aspect of the present invention also provides a deep learning-based fiber optic accelerometer signal noise suppression system, the system comprising a memory and a processor, wherein the memory stores a fiber optic accelerometer signal noise suppression method, and when the processor executes the fiber optic accelerometer signal noise suppression method, the following steps are implemented:

[0041] The original noisy signal and the clean reference signal from the fiber optic accelerometer are collected and combined to model the noise characteristics and construct a target noisy signal library.

[0042] A deep neural signal noise suppression network is designed by combining a target noise information database with the corresponding stored target composite noise signals.

[0043] By combining a deep neural signal noise suppression network, the target composite noisy signal in the target noisy signal library is segmented, and the signal of the target composite noisy signal sample is suppressed at the same time.

[0044] Sample analysis was performed on the target composite noise reduction signal samples, and the deep neural signal noise suppression network was modified and optimized based on the sample analysis results.

[0045] This invention addresses the technical deficiencies in the prior art and offers the following advantages: During the operation of the fiber optic accelerometer, it acquires the original noisy signal and a clean reference signal, models the noise characteristics, and designs a neural network to suppress signal noise based on these characteristics. The network is then run, and improvements are made to address any issues encountered during operation. This invention improves the signal noise suppression effect and efficiency of the fiber optic accelerometer, thereby enhancing the accuracy of the monitoring signal output by the fiber optic accelerometer. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0047] Figure 1 A flowchart of a deep learning-based method for suppressing noise in fiber optic accelerometer signals is shown.

[0048] Figure 2 A flowchart of a method for suppressing signals from samples of a target composite noisy signal is shown.

[0049] Figure 3 A program view of a deep learning-based fiber optic accelerometer signal noise suppression system is shown. Detailed Implementation

[0050] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0051] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0052] Figure 1 A flowchart illustrating a deep learning-based method for suppressing noise in fiber optic accelerometer signals is shown, including the following steps:

[0053] S102: Collect the raw noisy signal and clean reference signal from the fiber optic accelerometer, and combine them to model the noise characteristics and construct a target noisy signal library;

[0054] S104: Design a deep neural signal noise suppression network by combining the target noise information database and the corresponding stored target composite noise signal;

[0055] S106: Combining a deep neural signal noise suppression network, the target composite noisy signal in the target noisy signal library is segmented into samples, and the signal of the target composite noisy signal samples is suppressed at the same time.

[0056] S108: Perform sample analysis on the target composite noise reduction signal samples, and optimize the deep neural signal noise suppression network based on the sample analysis results.

[0057] Furthermore, in a preferred embodiment of the present invention, the acquisition of the original noisy signal and the clean reference signal from the fiber optic accelerometer, combined with noise characteristic modeling, to construct a target noisy signal library, specifically involves:

[0058] Obtain the fiber optic accelerometer that needs to be suppressed for signal noise, calibrate it as the target fiber optic accelerometer, run the target fiber optic accelerometer, and collect the original noisy signal of the target fiber optic accelerometer in real time during the operation, and calibrate it as the target original noisy signal;

[0059] The target's original noisy signal includes environmental noise and its own output electrical signal. At the same time, the target fiber optic accelerometer's clean reference signal is collected and calibrated as the target's clean reference signal.

[0060] The target pure reference signal needs to be collected in an anechoic chamber and serves as the reference signal for the target fiber optic accelerometer.

[0061] Spectral analysis is performed on the original noisy target signal to measure its noise characteristics, wherein the noise characteristics of the original noisy target signal include Gaussianity, stationarity, and temperature drift characteristics.

[0062] Based on the noise characteristics of the original noisy signal of the target, noise modeling is performed on the original noisy signal of the target to obtain the noise library of the target fiber optic accelerometer, which is calibrated as the target noise library. The target noise library stores the original noisy signal of the target acquired in real time and the corresponding noise characteristics.

[0063] The target clean reference signal and the target original noisy signal stored in the target noise library are combined and superimposed to form a composite noise. The number of data samples in the target noise library is expanded and determined. The mixing ratio of the superimposed composite noise is dynamically adjusted in the target noise library to ensure that the signal-to-noise ratio of the superimposed composite noise is maintained within a preset value, thus obtaining the target noisy signal library and the target composite noisy signal.

[0064] It is important to note that two sets of data need to be acquired simultaneously: the original noisy signal and the clean reference signal. The original noisy signal is the output signal of the actual fiber optic accelerometer, including environmental and system noise. It reflects the acceleration changes of the measured object and is also an electrical signal obtained after photoelectric conversion of the optical phase / intensity modulation. Its core function is to monitor weak mechanical vibrations, operate in scenarios where traditional electrical sensors fail, such as strong electromagnetic interference, high temperatures, and flammable and explosive environments, and for early fault warning in bridges, aircraft engines, and oil and gas pipelines. The clean reference signal is a reference signal acquired in an anechoic chamber or vacuum environment. Spectral analysis is performed on the original noisy target signal to measure its noise characteristics. Gaussian characteristics, stationarity, and temperature drift characteristics all reflect the noise characteristics and provide conditions for subsequent noise processing. The purpose of superimposing the clean reference signal and the original noisy signal is to adjust the signal-to-noise ratio and the noise mixing ratio. This allows for the expansion of the dataset during the fiber optic accelerometer signal noise suppression process, used for data augmentation, and to prevent data loss during filtering and noise reduction.

[0065] Furthermore, in a preferred embodiment of the present invention, the design of a deep neural signal noise suppression network by combining the target noise information database and the corresponding stored target composite noisy signal specifically includes:

[0066] A big data network is introduced, and all network architectures of general deep neural networks for noise suppression of target composite noisy signals are retrieved from the big data network. At the same time, the network architecture with the highest historical usage rate is selected and labeled as the target deep neural network architecture.

[0067] The target deep neural network architecture includes an input layer, a feature extraction layer, a regularization layer, and an output layer.

[0068] The architecture design of the target deep neural network includes determining the maximum processable sample dimension of the input layer based on the number of data samples of the target composite noisy signal in the target noisy signal library, while controlling the maximum processable sample dimension of the input layer and the output layer to be equal.

[0069] Within the regularization layer, the regularization method is determined to be total variational regularization, and within the feature extraction layer, a multi-objective loss function is determined. The multi-objective loss function is determined based on the noise characteristics of the original noisy target signal, including a time-domain loss function and a frequency-domain loss function.

[0070] The designed target deep neural network architecture is named Deep Neural Signal Noise Suppression Network.

[0071] It should be noted that various network architectures exist for signal and noise suppression of fiber optic accelerometers. The network architecture with the highest historical usage is selected and labeled as the target deep neural network architecture. This target deep neural network architecture typically uses a CRNN network architecture, whose core structure employs a CNN neural network for feature extraction and LSTM for temporal modeling to suppress composite noise. This network architecture includes an input layer, a feature extraction layer, a regularization layer, and an output layer. The input layer receives the input signal; the feature extraction layer performs signal feature scale matching using convolutional kernels and a loss function; the regularization layer is used for subsequent normalized time-frequency domain joint enhancement processing; and the output layer outputs the noise-suppressed signal. The resulting deep neural signal and noise suppression network can achieve signal and noise suppression for fiber optic accelerometers.

[0072] Furthermore, in a preferred embodiment of the present invention, the step of performing sample analysis on the target composite noise-reduced signal samples and optimizing the deep neural signal-noise suppression network based on the sample analysis results specifically includes:

[0073] Sample analysis is performed on the target composite noise-reduced signal sample, wherein the sample analysis is to analyze the distortion rate and waveform fidelity of the target composite noise-reduced signal sample;

[0074] If the distortion rate and waveform fidelity of the target composite noise reduction signal sample are not lower than the preset values, the deep neural signal noise suppression network is calibrated as a qualified deep neural signal noise suppression network.

[0075] If the distortion rate and waveform fidelity of the target composite noise-reduced signal sample are lower than the preset value, a feature frequency protection term is added to the multi-target loss function in the deep neural signal noise suppression network. The feature frequency protection term prioritizes protecting the distortion rate and waveform fidelity of the target composite noise-reduced signal sample, so that the distortion rate and waveform fidelity of the target composite noise-reduced signal sample are not lower than the preset value.

[0076] A qualified deep neural signal noise suppression network is used to perform signal noise suppression processing on all target composite noisy signals in the target noisy signal library.

[0077] It should be noted that the target composite denoised signal sample has already been denoised after it is generated. If the distortion rate and waveform fidelity of the target composite denoised signal sample are lower than the preset values, it indicates that the deep neural signal noise suppression network needs to be optimized. A feature frequency protection term is added to protect the feature change rate during the denoising process, ensuring that the distortion rate and waveform fidelity of the target composite denoised signal sample do not fall below the preset values.

[0078] Figure 2 A flowchart of a method for signal suppression of samples of a target composite noisy signal is shown, including the following steps:

[0079] S202: Combining a deep neural signal noise suppression network, the target composite noisy signal in the target noisy signal library is segmented into samples, and the signal of the target composite noisy signal samples is suppressed at the same time.

[0080] S204: Combining a deep neural signal-noise suppression network, signal-noise suppression processing is performed on segmented samples of different target composite noisy signals after normalized time-frequency domain enhancement processing.

[0081] Furthermore, in a preferred embodiment of the present invention, the step of combining a deep neural signal noise suppression network to segment the target composite noisy signal within the target noisy signal library, and simultaneously suppressing the signal of the target composite noisy signal samples, specifically involves:

[0082] Within the target noisy signal library, target composite noisy signal samples are extracted, and the target composite noisy signal samples are segmented to obtain target composite noisy signal segmented samples.

[0083] A deep neural signal noise suppression network was run to perform normalized time-frequency domain joint enhancement processing on segmented samples of different target composite noisy signals.

[0084] A bandwidth constraint term is introduced, which is the start and stop of the enhancement process based on the bandwidth energy in the deep neural signal noise suppression network during the normalized time-frequency domain joint enhancement processing of different target composite noisy signal segment samples.

[0085] Real-time bandwidth energy is calculated within the deep neural signal-noise suppression network, and the maximum bandwidth energy and target augmentation time are preset.

[0086] If, within the target enhancement time, the real-time bandwidth energy within the deep neural signal noise suppression network exceeds the maximum bandwidth energy, then the normalized time-frequency domain joint enhancement processing for different target composite noisy signal segment samples is stopped.

[0087] By combining a deep neural signal-noise suppression network, signal-noise suppression processing is performed on segmented samples of different target composite noisy signals after normalized time-frequency domain enhancement processing.

[0088] It should be noted that extracting and segmenting the target composite noisy signal samples serves to determine whether the deep neural signal noise suppression network (DNN) needs optimization. Segmentation allows for more detailed noise suppression each time. Normalized time-frequency domain joint enhancement processing is performed on different segments of the target composite noisy signal. Segmented normalization allows each signal segment to be calculated independently, avoiding the destruction of subtle acceleration features by global normalization, and preventing the distortion of the signal's dynamic range by strong noise segments. Time-frequency domain joint enhancement processing enhances the data of the samples, avoiding data loss during suppression. Bandwidth constraints are used to start and stop enhancement processing based on the bandwidth energy within the DNN during the normalized time-frequency domain joint enhancement processing of different segments of the target composite noisy signal. Excessive energy requires stopping the enhancement process because, according to the law of conservation of energy, high bandwidth energy indicates high energy loss in the time-frequency domain, which can lead to suppression bias.

[0089] Furthermore, in a preferred embodiment of the present invention, the step of combining a deep neural signal-noise suppression network to perform signal-noise suppression processing on different target composite noisy signal segment samples after normalized time-frequency domain enhancement processing specifically involves:

[0090] During the target enhancement time, after normalized time-frequency domain joint enhancement processing is performed on the segmented samples of different target composite noisy signals, the segmented samples of target composite noisy signals are trained in segments.

[0091] Among them, segmented training involves using a multi-objective loss function to segment the noisy composite signal of the target, and monitoring the noise frequency of different segments of the noisy composite signal of the target in real time during the segmented denoising process.

[0092] A preset training stop mechanism is provided. The training stop mechanism is as follows: when the noise frequency of the segmented samples of different target composite noisy signals is less than a preset value after five consecutive monitoring, the segmented noise reduction of the target composite noisy signal segments is stopped, and the segmented noise reduction of the target composite noisy signal segments is sampled and synthesized to obtain the initial target composite noise reduction signal sample.

[0093] The surrounding environmental parameters of the target fiber optic accelerometer are monitored in real time, and the rate of change of the surrounding environmental parameters within a predetermined time is calculated. If the rate of change of the surrounding environmental parameters within the predetermined time is greater than a preset value, the initial target composite noise reduction signal sample is fine-tuned in the deep neural signal noise suppression network to obtain the target composite noise reduction signal sample.

[0094] It should be noted that segmented training involves using a multi-objective loss function to denoise segmented samples of the target composite noisy signal. During the segmented denoising process, the noise frequency of different target composite noisy signal segments is monitored in real time. A high noise frequency indicates the continued presence of noise and ineffective signal noise suppression. When the noise frequency of different target composite noisy signal segments falls below a preset value after five consecutive monitoring iterations, segmented denoising is stopped. This indicates that the segmented denoising effect is significant, and the denoised target composite noisy signal segments can be synthesized to obtain an initial target composite denoised signal sample. At this point, environmental parameters can affect the synthesized denoised signal; even a slight change can regenerate noise or alter its frequency. Therefore, fine-tuning of the initial target composite denoised signal sample is necessary to obtain the final target composite denoised signal sample. This fine-tuning involves processing samples in small batches, freezing the underlying signal parameters, and fine-tuning the top-level signal parameters to prevent overfitting.

[0095] like Figure 3 As shown, a second aspect of the present invention also provides a deep learning-based fiber optic accelerometer signal noise suppression system. The fiber optic accelerometer signal noise suppression system includes a memory 31 and a processor 32. The memory 31 stores a fiber optic accelerometer signal noise suppression method. When the processor 32 executes the fiber optic accelerometer signal noise suppression method, it performs the following steps:

[0096] The original noisy signal and the clean reference signal from the fiber optic accelerometer are collected and combined to model the noise characteristics and construct a target noisy signal library.

[0097] A deep neural signal noise suppression network is designed by combining a target noise information database with the corresponding stored target composite noise signals.

[0098] By combining a deep neural signal noise suppression network, the target composite noisy signal in the target noisy signal library is segmented, and the signal of the target composite noisy signal sample is suppressed at the same time.

[0099] Sample analysis was performed on the target composite noise reduction signal samples, and the deep neural signal noise suppression network was modified and optimized based on the sample analysis results.

[0100] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for suppressing signal noise from fiber optic accelerometers based on deep learning, characterized in that, Includes the following steps: The original noisy signal and the clean reference signal from the fiber optic accelerometer are collected and combined to model the noise characteristics and construct a target noisy signal library. A deep neural signal noise suppression network is designed by combining a target noise information database with the corresponding stored target composite noise signals. By combining a deep neural signal noise suppression network, the target composite noisy signal in the target noisy signal library is segmented, and the signal of the target composite noisy signal sample is suppressed at the same time. Sample analysis was performed on the target composite noise reduction signal samples, and the deep neural signal noise suppression network was modified and optimized based on the sample analysis results. Specifically, the design of a deep neural signal noise suppression network, which combines a target noise information database with corresponding stored target composite noise signals, involves: A big data network is introduced, and all network architectures of general deep neural networks for noise suppression of target composite noisy signals are retrieved from the big data network. At the same time, the network architecture with the highest historical usage rate is selected and labeled as the target deep neural network architecture. The target deep neural network architecture includes an input layer, a feature extraction layer, a regularization layer, and an output layer. The architecture design of the target deep neural network includes determining the maximum processable sample dimension of the input layer based on the number of data samples of the target composite noisy signal in the target noisy signal library, while controlling the maximum processable sample dimension of the input layer and the output layer to be equal. Within the regularization layer, the regularization method is determined to be total variational regularization, and within the feature extraction layer, a multi-objective loss function is determined. The multi-objective loss function is determined based on the noise characteristics of the original noisy target signal, including a time-domain loss function and a frequency-domain loss function. The designed target deep neural network architecture is named Deep Neural Signal Noise Suppression Network. Specifically, the deep neural signal noise suppression network is used to segment the target composite noisy signals in the target noisy signal library and simultaneously suppress the signals of the target composite noisy signals. Within the target noisy signal library, target composite noisy signal samples are extracted, and the target composite noisy signal samples are segmented to obtain target composite noisy signal segmented samples. A deep neural signal noise suppression network was run to perform normalized time-frequency domain joint enhancement processing on segmented samples of different target composite noisy signals. A bandwidth constraint term is introduced, which is the start and stop of the enhancement process based on the bandwidth energy in the deep neural signal noise suppression network during the normalized time-frequency domain joint enhancement processing of different target composite noisy signal segment samples. Real-time bandwidth energy is calculated within the deep neural signal-noise suppression network, and the maximum bandwidth energy and target augmentation time are preset. If, within the target enhancement time, the real-time bandwidth energy within the deep neural signal noise suppression network exceeds the maximum bandwidth energy, then the normalized time-frequency domain joint enhancement processing for different target composite noisy signal segment samples is stopped. By combining a deep neural signal-noise suppression network, signal-noise suppression processing is performed on segmented samples of different target composite noisy signals after normalized time-frequency domain enhancement processing.

2. The method for suppressing fiber optic accelerometer signal noise based on deep learning as described in claim 1, characterized in that, The process involves acquiring the raw noisy signal and the clean reference signal from the fiber optic accelerometer, and combining them to model the noise characteristics to construct a target noisy signal library. Specifically: Obtain the fiber optic accelerometer that needs to be suppressed for signal noise, calibrate it as the target fiber optic accelerometer, run the target fiber optic accelerometer, and collect the original noisy signal of the target fiber optic accelerometer in real time during the operation, and calibrate it as the target original noisy signal; The target's original noisy signal includes environmental noise and its own output electrical signal. At the same time, the target fiber optic accelerometer's clean reference signal is collected and calibrated as the target's clean reference signal. The target pure reference signal needs to be collected in an anechoic chamber and serves as the reference signal for the target fiber optic accelerometer. Spectral analysis is performed on the original noisy target signal to measure its noise characteristics, wherein the noise characteristics of the original noisy target signal include Gaussianity, stationarity, and temperature drift characteristics. Based on the noise characteristics of the original noisy signal of the target, noise modeling is performed on the original noisy signal of the target to obtain the noise library of the target fiber optic accelerometer, which is calibrated as the target noise library. The target noise library stores the original noisy signal of the target acquired in real time and the corresponding noise characteristics. The target clean reference signal and the target original noisy signal stored in the target noise library are combined and superimposed to form a composite noise. The number of data samples in the target noise library is expanded and determined. The mixing ratio of the superimposed composite noise is dynamically adjusted in the target noise library to ensure that the signal-to-noise ratio of the superimposed composite noise is maintained within a preset value, thus obtaining the target noisy signal library and the target composite noisy signal.

3. The method for suppressing fiber optic accelerometer signal noise based on deep learning as described in claim 1, characterized in that, The deep neural signal-noise suppression network is used to perform signal-noise suppression processing on segmented samples of different target composite noisy signals after normalized time-frequency domain enhancement processing, specifically as follows: During the target enhancement time, after normalized time-frequency domain joint enhancement processing is performed on the segmented samples of different target composite noisy signals, the segmented samples of target composite noisy signals are trained in segments. Among them, segmented training involves using a multi-objective loss function to segment the noisy composite signal of the target, and monitoring the noise frequency of different segments of the noisy composite signal of the target in real time during the segmented denoising process. A preset training stop mechanism is provided. The training stop mechanism is as follows: when the noise frequency of the segmented samples of different target composite noisy signals is less than a preset value after five consecutive monitoring, the segmented noise reduction of the target composite noisy signal segments is stopped, and the segmented noise reduction of the target composite noisy signal segments is sampled and synthesized to obtain the initial target composite noise reduction signal sample. The surrounding environmental parameters of the target fiber optic accelerometer are monitored in real time, and the rate of change of the surrounding environmental parameters within a predetermined time is calculated. If the rate of change of the surrounding environmental parameters within the predetermined time is greater than a preset value, the initial target composite noise reduction signal sample is fine-tuned in the deep neural signal noise suppression network to obtain the target composite noise reduction signal sample.

4. The method for suppressing fiber optic accelerometer signal noise based on deep learning as described in claim 1, characterized in that, The process of analyzing the target composite noise-reduced signal samples and then modifying and optimizing the deep neural signal noise suppression network based on the analysis results includes: Sample analysis is performed on the target composite noise-reduced signal sample, wherein the sample analysis is to analyze the distortion rate and waveform fidelity of the target composite noise-reduced signal sample; If the distortion rate and waveform fidelity of the target composite noise reduction signal sample are not lower than the preset values, the deep neural signal noise suppression network is calibrated as a qualified deep neural signal noise suppression network. If the distortion rate and waveform fidelity of the target composite noise-reduced signal sample are lower than the preset value, a feature frequency protection term is added to the multi-target loss function in the deep neural signal noise suppression network. The feature frequency protection term prioritizes protecting the distortion rate and waveform fidelity of the target composite noise-reduced signal sample, so that the distortion rate and waveform fidelity of the target composite noise-reduced signal sample are not lower than the preset value. A qualified deep neural signal noise suppression network is used to perform signal noise suppression processing on all target composite noisy signals in the target noisy signal library.

5. A deep learning-based fiber optic accelerometer signal noise suppression system, characterized in that, The fiber optic accelerometer signal noise suppression system includes a memory and a processor. The memory stores a program for a fiber optic accelerometer signal noise suppression method. When the program for a fiber optic accelerometer signal noise suppression method is executed by the processor, the steps of the fiber optic accelerometer signal noise suppression method as described in any one of claims 1-4 are implemented.

Images