Method for enhancing quality of bare optical fiber sensing signal in vehicle-mounted vibration environment

Abstract

This application discloses a method for enhancing the quality of bare fiber sensing signals in vehicle-mounted vibration environments. It relates to the field of fiber optic sensing signal quality enhancement, and utilizes a multi-axis inertial measurement unit and a vibration sensing network to construct a spatiotemporal feature tensor, enabling three-dimensional dynamic perception of the vibration environment. Filtering parameters and gain are dynamically configured based on vibration characteristics to ensure stable signal acquisition under strong vibration conditions. A deep feature decoupling network, combined with vibration environment information, effectively separates noise from the target signal, improving the signal-to-noise ratio. A dynamic compensation module, based on a physical heuristic model, corrects vibration-induced transmission distortion from three dimensions: polarization state, birefringence, and microbending loss, ensuring the physical integrity of the signal. A closed-loop collaborative control mechanism dynamically optimizes the parameters of each module through real-time signal quality feedback, forming an adaptive optimization closed loop. Through spatiotemporal multi-dimensional collaborative processing, high-quality and robust enhancement of bare fiber sensing signals in vehicle-mounted vibration environments is achieved.

Description

Technical Field

[0001] This application relates to the field of fiber optic sensing signal quality enhancement, and in particular to a method for enhancing the quality of bare fiber optic sensing signals for vehicle vibration environments. Background Technology

[0002] With the development of intelligent cockpits and health monitoring systems, the need for non-intrusive and continuous monitoring of vital signs of drivers or passengers during vehicle operation is becoming increasingly urgent. Fiber optic sensing technology, especially bare optical fibers sensitive to micro-strain, is considered for monitoring vital signs (such as heartbeat and respiration) in vehicle seats due to its advantages such as resistance to electromagnetic interference, small size, and ability to be embedded in fabric. However, the complex and intense vibration noise generated during vehicle operation can be coupled to the fiber optic sensors through the seat structure. Its amplitude is often much greater than the micro-strain signal caused by vital signs, resulting in an extremely low signal-to-noise ratio and making it extremely difficult to directly extract vital signs.

[0003] In existing technologies, for single-channel fiber optic sensors, fixed-bandband bandpass filters or blind source separation at the signal processing backend are typically used. However, the former is difficult to adapt to the non-stationary nature of vehicle vibration and noise and the possibility that the spectrum may overlap with the frequency band of vital signs. The latter has limited separation effect under low signal-to-noise ratio conditions and is prone to losing weak vital sign information. There are also schemes that use reference sensors (such as accelerometers) for active noise cancellation, but this increases the complexity of the system, and the signal of the reference sensor may not be able to perfectly match the vibration and noise sensed by the fiber optic cable. Summary of the Invention

[0004] The purpose of this invention is to provide a method for enhancing the quality of bare fiber sensing signals in vehicle vibration environments to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for enhancing the quality of bare optical fiber sensing signals in vehicle-mounted vibration environments, comprising:

[0006] Vibration environment perception step: It is used to collect multi-dimensional vibration acceleration, angular velocity and spectral characteristic data of key measurement points of the vehicle in real time through multi-axis inertial measurement unit and vibration sensing network, and construct the spatiotemporal characteristic tensor of vehicle vibration environment field;

[0007] Signal Synchronization Acquisition and Conditioning Steps: The raw signal output by the bare fiber sensing system is synchronously acquired, and the preprocessing parameters are adaptively configured according to the spatiotemporal characteristic tensor of the vehicle vibration environment field to perform preliminary noise reduction and normalization on the raw signal, so as to obtain the synchronized sensing signal stream and vibration environment stream.

[0008] Feature enhancement and extraction steps: used to perform spatiotemporal alignment and feature-level fusion of synchronized sensing signal streams and vibration environment streams. Through a pre-trained vibration noise and signal decoupling model, noise components strongly correlated with the vibration environment and potential target signal components are separated, and the target signal components are enhanced and reconstructed.

[0009] Dynamic compensation step: Based on the vibration environment feature tensor, it is used to solve the model of the influence of vibration on the polarization state, birefringence and microbending additional loss of bare optical fiber in real time, generate dynamic compensation vector, and compensate for the transmission characteristic distortion of the enhanced signal.

[0010] Signal quality assessment steps: used to calculate the time-frequency domain quality indicators of the compensated enhanced signal, and compare them with the preset signal-to-noise ratio and distortion thresholds to generate real-time signal quality assessment factors and confidence labels;

[0011] Cooperative control steps: These steps are used to receive signal quality assessment factors, dynamically adjust the parameter weights of the vibration noise-signal decoupling model, the coefficients of the dynamic compensation model, and the preprocessing strategies of the signal synchronous acquisition and conditioning steps, forming a closed-loop control flow with the goal of optimizing the output signal quality.

[0012] In a preferred embodiment of this scheme, the vibration environment perception step, when performing the real-time acquisition of multi-dimensional vibration data from key vehicle measurement points via a multi-axis inertial measurement unit and vibration sensor network to construct the spatiotemporal characteristic tensor of the vehicle vibration environment field, is specifically executed as follows:

[0013] Three-axis accelerometers and gyroscopes are deployed at the vehicle chassis suspension connection points, powertrain mounting points, and key sections of the body longitudinal beams to form a vibration sensing network.

[0014] Using the vehicle CAN bus time as a reference, the data of each node in the vibration sensing network is synchronized in time.

[0015] Short-time Fourier transform is performed on the triaxial acceleration and angular velocity signals of each node to extract the energy, dominant frequency, and spectral kurtosis features of each node within a preset frequency band, forming a node vibration feature vector;

[0016] Based on the spatial coordinates of the sensor nodes, the vibration feature vectors of all nodes at the same time are spatially interpolated and meshed to generate a spatial feature map describing the vibration energy and mode distribution of the whole vehicle.

[0017] By stacking spatial feature maps according to time series, a spatiotemporal feature tensor of the vehicle vibration environment field with time, space, two dimensions, and feature channel dimensions is formed.

[0018] In the preferred embodiment of this scheme, the signal synchronous acquisition and conditioning step, when performing preliminary noise reduction and normalization of the original signal by adaptively configuring preprocessing parameters based on the spatiotemporal characteristic tensor of the vehicle vibration environment field, is specifically executed as follows:

[0019] Extract the current dominant vibration frequency band range and vibration energy level from the spatiotemporal characteristic tensor of the vehicle vibration environment field;

[0020] Based on the dominant vibration frequency band, the basis functions and threshold parameters of the band-stop filter bank or wavelet threshold denoising are adaptively configured to perform targeted frequency domain notch filtering or time-frequency domain filtering on the original fiber optic sensing signal.

[0021] Based on the vibration energy level, the gain of the signal amplification circuit or the scaling factor of the digital signal is dynamically adjusted to normalize the signal amplitude to the preset dynamic range center interval.

[0022] The filtered and normalized sensor signal stream and the corresponding vibration environment stream are given a unified timestamp.

[0023] In a preferred embodiment of this scheme, the feature enhancement and extraction steps, when executing the pre-trained vibration noise and signal decoupling model to separate the noise component and the target signal component, are specifically executed as follows:

[0024] A deep feature decoupling network is constructed, which takes the spatiotemporal feature tensor of the vehicle vibration environment field as the conditional input and the original sensing signal stream as the main input. The deep feature decoupling network includes a shared encoder, a noise-dedicated decoding branch and a signal-dedicated decoding branch.

[0025] The network is pre-trained using vehicle-mounted bench test data, which includes bare fiber sensing signals under the combined action of known vibration input and known target signal input, as well as pure noise signals under individual vibration excitation.

[0026] The real-time acquired synchronous sensing signal stream and the spatiotemporal characteristic tensor of the vehicle vibration environment field are input into the trained decoupled model. The noise-dedicated decoding branch outputs the estimated vibration-induced noise component, and the signal-dedicated decoding branch outputs the initially enhanced target signal component.

[0027] The estimated noise component is subtracted from the original sensing signal stream and then weighted and fused with the output of the dedicated signal decoding branch to obtain the enhanced and reconstructed target signal component.

[0028] In the preferred embodiment of this scheme, the dynamic compensation step, when executing the model for real-time calculation of the impact of vibration on the polarization state, birefringence, and microbending-related losses of bare optical fiber, and generating the dynamic compensation vector, is specifically executed as follows:

[0029] A simplified physical heuristic model is established, with vibration acceleration and spectrum as inputs and polarization state rotation matrix, equivalent birefringence change, and microbending loss coefficient as outputs.

[0030] The spatiotemporal characteristic tensor of the vehicle vibration environment field is spatially mapped along the bare optical fiber deployment path to obtain the local vibration excitation vector distributed along the fiber length direction.

[0031] By inputting the local vibration excitation vector into the influence model, the polarization state disturbance vector, birefringence change vector, and additional loss vector at each point along the fiber length direction are calculated.

[0032] The polarization fading compensation and phase deblurring correction of the enhanced signal are performed by using the polarization state perturbation vector and the birefringence change vector. The signal strength is then compensated by using the additional loss vector. Finally, the dynamic compensation vector is output and applied to the feature-enhanced signal.

[0033] In a preferred embodiment of this scheme, the signal quality assessment step, when calculating the time-frequency domain quality index of the compensated enhanced signal, is executed as follows:

[0034] Under quiet operating conditions where the target signal is known, a reference signal is acquired, and its power spectrum is calculated as the reference spectrum.

[0035] For the signal enhanced by the dynamic compensation vector, calculate its spectral distortion compared to the reference spectrum, the signal-to-noise ratio in the target characteristic frequency band, and the short-time zero-crossing rate or envelope smoothness of the signal.

[0036] The calculated spectral distortion, signal-to-noise ratio, and smoothness indices are compared with the high-quality signal thresholds preset according to the application scenario.

[0037] If all indicators are better than the threshold, high-quality evaluation factors and high-confidence labels are generated; if any indicator is lower than the threshold, medium- and low-quality evaluation factors and corresponding low-confidence labels are generated according to the degree of deviation, and the type of indicator that does not meet the standard is fed back to the collaborative control steps.

[0038] In the preferred embodiment of this scheme, the collaborative control step, when dynamically adjusting the parameters of each step to form a closed-loop control flow, is executed in the following manner:

[0039] The quality assessment factors and types of non-compliant indicators fed back from the received signal quality assessment steps;

[0040] If the signal-to-noise ratio does not meet the standard, increase the weight of the loss function of the noise-dedicated decoding branch in the feature enhancement and extraction steps, or guide the band-stop filter to tighten the bandwidth in the signal synchronization acquisition and conditioning steps.

[0041] If the spectral distortion does not meet the standard, adjust the parameters that affect the model in the dynamic compensation step, or reduce the intensity of noise separation to prevent signal distortion.

[0042] If the evaluation factors remain of high quality, the processing parameters will be gradually relaxed to reduce the computational complexity of the system and achieve energy-saving operation.

[0043] The adjusted parameter set is sent to the corresponding steps in real time, and the changes in the signal quality assessment factor in the next cycle are monitored to form an adaptive optimization closed loop with quality assessment as feedback.

[0044] Compared with the prior art, the beneficial effects of the present invention are:

[0045] The bare fiber sensing signal quality enhancement method proposed in this invention for vehicle vibration environments has significant technical advantages. It constructs a spatiotemporal feature tensor using a multi-axis inertial measurement unit and a vibration sensing network to achieve three-dimensional dynamic perception of the vibration environment, providing accurate environmental context for subsequent processing. An adaptive signal conditioning module dynamically configures filtering parameters and gain based on vibration characteristics, ensuring stable signal acquisition under strong vibration conditions. A deep feature decoupling network, combined with vibration environment information, effectively separates noise from the target signal, improving the signal-to-noise ratio. A dynamic compensation module, based on a physical heuristic model, corrects vibration-induced transmission distortion from three dimensions: polarization state, birefringence, and microbending loss, ensuring the physical integrity of the signal. A closed-loop collaborative control mechanism dynamically optimizes the parameters of each module through real-time signal quality feedback, forming an adaptive optimization closed loop that reduces system energy consumption while ensuring signal quality. This method achieves high-quality and robust enhancement of bare fiber sensing signals in vehicle vibration environments through spatiotemporal multi-dimensional collaborative processing. Attached Figure Description

[0046] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0047] Figure 1 This is a schematic diagram illustrating the connection steps in an embodiment of the present invention. Detailed Implementation

[0048] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0049] Please see Figure 1This invention provides a method for enhancing the quality of bare fiber sensing signals for vehicle vibration environments. The method includes a vibration environment sensing step, a signal synchronous acquisition and conditioning step, a feature enhancement and extraction step, a dynamic compensation step, a signal quality assessment step, and a collaborative control step.

[0050] Vibration environment perception step: It is used to collect multi-dimensional vibration acceleration, angular velocity and spectral characteristic data of key measurement points of the vehicle in real time through multi-axis inertial measurement unit and vibration sensing network, and construct the spatiotemporal characteristic tensor of vehicle vibration environment field;

[0051] Furthermore, the vibration environment perception step, when executing the real-time acquisition of multi-dimensional vibration data from key vehicle measurement points via a multi-axis inertial measurement unit and vibration sensor network to construct the spatiotemporal characteristic tensor of the vehicle vibration environment field, is specifically executed as follows:

[0052] Three-axis accelerometers and gyroscopes are deployed at the vehicle chassis suspension connection points, powertrain mounting points, and key sections of the body longitudinal beams to form a vibration sensing network.

[0053] Using the vehicle CAN bus time as a reference, the data of each node in the vibration sensing network is synchronized in time.

[0054] Short-time Fourier transform is performed on the triaxial acceleration and angular velocity signals of each node to extract the energy, dominant frequency, and spectral kurtosis features of each node within a preset frequency band, forming a node vibration feature vector;

[0055] Based on the spatial coordinates of the sensor nodes, the vibration feature vectors of all nodes at the same time are spatially interpolated and meshed to generate a spatial feature map describing the vibration energy and mode distribution of the whole vehicle.

[0056] By stacking spatial feature maps according to time series, a spatiotemporal feature tensor of the vehicle vibration environment field with time, space, two dimensions, and feature channel dimensions is formed.

[0057] Vibration data acquisition and synchronization: Inertial measurement units (IMUs) integrating three-axis accelerometers and three-axis gyroscopes are deployed at representative locations of vehicle vibration response, such as chassis suspension connection points (e.g., the connection between the front left and right lower control arms), powertrain mounting points (e.g., the engine and transmission mounts), and key sections of the body longitudinal beams (e.g., the longitudinal beam below the B-pillar). This forms a sensor network covering the key vibration transmission paths of the entire vehicle. All IMU nodes receive and follow the same high-precision time message (e.g., a synchronization clock from the vehicle gateway) via the vehicle's CAN bus, and timestamp the three-axis acceleration (X, Y, Z direction linear vibration) and three-axis angular velocity (roll, pitch, yaw angle vibration) signals they collect, achieving microsecond-level synchronization of all network data under the vehicle's global time reference.

[0058] Secondly, node feature vector extraction is performed: the raw time-domain vibration signal synchronously acquired by each IMU node is slid across a fixed time window (e.g., 100 milliseconds), and the data within the window is subjected to a short-time Fourier transform (STFT) to convert the time-domain signal into a time spectrum. For preset frequency bands where vehicle vibration energy is concentrated (e.g., the 0-200Hz body structure vibration band, the 200-1000Hz powertrain high-frequency vibration band), the following are calculated in each frequency band: 1. Energy feature (the sum of squares of the spectral amplitudes in this frequency band); 2. Dominant frequency feature (the frequency component with the highest energy in this frequency band); 3. Spectral kurtosis feature (characterizing the intensity of the signal impact component). For a node, the multiple feature values ​​calculated on all preset frequency bands (e.g., 3 frequency bands × 3 features = 9 dimensions) are arranged in a fixed order to form the node vibration feature vector of that node. This vector comprehensively characterizes the vibration intensity, dominant frequency, and impact characteristics of the measuring point in the current time window.

[0059] Next, spatial feature map generation is performed: the three-dimensional spatial coordinates of all IMU nodes are obtained (in the vehicle coordinate system, such as with the vehicle's center of mass as the origin). For any given moment, the vibration feature vectors of all nodes are used as the attribute values ​​at that spatial coordinate point. Spatial interpolation algorithms (such as Kriging interpolation or inverse distance weighted interpolation) are used to estimate the feature vector values ​​at discrete nodes onto each grid point in the entire vehicle space (such as a regular three-dimensional grid covering the vehicle body). In this way, for a specific vibration feature (such as the first frequency band energy), a spatial feature map describing the distribution of that feature in the vehicle space can be generated. This process is repeated for all vibration feature dimensions.

[0060] Finally, a spatiotemporal feature tensor is constructed: the spatial feature maps of all feature dimensions generated in each time window are stacked according to the dimension of the feature channels (for example, 9 feature channels correspond to 9 spatial feature maps) to form a three-dimensional data block (two spatial dimensions and one feature channel dimension). Then, according to the time series, these three-dimensional data blocks are stacked sequentially to add the time dimension. Finally, a four-dimensional spatiotemporal feature tensor of the vehicle vibration environment field is formed, with dimensions of [time series length, spatial grid height, spatial grid width, number of feature channels]. This tensor systematically describes the dynamic evolution process of the vehicle vibration energy and vibration mode from both time and space dimensions, providing comprehensive environmental context information for subsequent steps.

[0061] Specific details of nodal vibration eigenvector calculation:

[0062] Energy characteristics: For the energy Ei of node i within the preset frequency band [flow, fhigh], it is obtained by summing the squares of the short-time Fourier transform amplitudes |xi(k)| corresponding to all discrete frequency points k within the frequency band:

[0063] Dominant frequency characteristics: Within the same frequency band [flow, fhigh], find the frequency point kmax that makes |x1(k)|^2 reach its maximum value. The frequency f{kmax} corresponding to this point is the dominant frequency fmaini.

[0064] Spectral kurtosis feature: Spectral kurtosis Ki is defined as a normalized fourth-order spectral cumulant, used to detect transient impulses in a signal;

[0065] The calculation formula is: ki=(∑{k}|X1(k)|^4) / (∑{k}|xi(k)|^2)^2, where the summation range is also within the target frequency band. The larger the value, the more significant the impulse component in the signal.

[0066] 2. Specification of spatiotemporal feature tensor construction:

[0067] Spatial interpolation: Taking a two-dimensional planar grid (corresponding to a vehicle top view) as an example, the value Fc(x,y) of the i-th feature channel at each grid point (x,y) is obtained through inverse distance weighted (IDW) interpolation:

[0068] Fc(x,y)=[∑{j=1}^{N}(wj*f{c,j})] / Z{j=1}^{N}wj, where N is the number of surrounding IMU nodes, f{c,j} is the feature value at the j-th node, wj=1 / dj^p is the weight, dj is the Euclidean distance from the grid point (x,y) to node j, and P is the power parameter (usually taken as 2). Each feature channel is interpolated independently to generate a single spatial feature map.

[0069] Tensor stacking: Assuming the time window sliding step is 50ms, 20 time slices can be generated every 1 second. The three-dimensional data blocks corresponding to these 20 time slices (each block has a size of [H, W, c], where H and W are the height and width of the spatial grid, and C is the number of feature channels) are stacked in time order to finally form a spatiotemporal feature tensor with a size of [20, H, W, c].

[0070] Signal Synchronization Acquisition and Conditioning Steps: The raw signal output by the bare fiber sensing system is synchronously acquired, and the preprocessing parameters are adaptively configured according to the spatiotemporal characteristic tensor of the vehicle vibration environment field to perform preliminary noise reduction and normalization on the raw signal, so as to obtain the synchronized sensing signal stream and vibration environment stream.

[0071] Furthermore, the specific execution method of the signal synchronous acquisition and conditioning step, when performing preliminary noise reduction and normalization of the original signal by adaptively configuring preprocessing parameters based on the spatiotemporal characteristic tensor of the vehicle vibration environment field, is as follows:

[0072] Extract the current dominant vibration frequency band range and vibration energy level from the spatiotemporal characteristic tensor of the vehicle vibration environment field;

[0073] Based on the dominant vibration frequency band, the basis functions and threshold parameters of the band-stop filter bank or wavelet threshold denoising are adaptively configured to perform targeted frequency domain notch filtering or time-frequency domain filtering on the original fiber optic sensing signal.

[0074] Based on the vibration energy level, the gain of the signal amplification circuit or the scaling factor of the digital signal is dynamically adjusted to normalize the signal amplitude to the preset dynamic range center interval.

[0075] The filtered and normalized sensor signal stream and the corresponding vibration environment stream are given a unified timestamp.

[0076] First, extract the dominant vibration features: from the newly generated spatiotemporal feature tensor, analyze the vibration frequency band with the strongest energy in the entire vehicle space at the current moment, and record it as the dominant vibration frequency band (for example, the current energy is concentrated in 50-80Hz). At the same time, calculate the average energy value of the spatial feature map of all current feature channels and map it to the vibration energy level (for example, low, medium and high levels).

[0077] Secondly, adaptive frequency domain filtering: Based on the extracted dominant vibration frequency band, preprocessing parameters are dynamically configured. If the dominant vibration frequency band is narrow, a set of band-stop filters with adjustable center frequencies are activated, with the center frequency set to the center frequency of the dominant frequency band and the bandwidth set according to the frequency band range. Targeted notch filtering is performed on the original bare fiber sensing signal to filter out common-mode interference that may be introduced by this strong vibration frequency band. If the dominant vibration frequency band is wide or exhibits broadband characteristics, wavelet threshold denoising is activated, and a matching wavelet basis function is adaptively selected according to the frequency band characteristics (e.g., 'db4' for impact components and 'sym8' for stationary vibrations). The threshold parameter is set according to the vibration energy level (the higher the energy level, the larger the threshold). The signal is filtered in the time-frequency domain.

[0078] Next, dynamic amplitude normalization: Based on the extracted vibration energy level, the gain of the signal conditioning stage is adjusted. On the analog side, the amplification factor is dynamically adjusted through a programmable gain amplifier (PGA) to ensure that the signal amplitude falls within the optimal quantization range of the ADC. On the digital side, the digital gain of the digital signal sampled by the ADC is adjusted according to the scaling factor corresponding to the energy level. The goal is to normalize the amplitude of the conditioned sensor signal to a preset, stable dynamic range center range (e.g., -1V to +1V or digital value -2048 to 2048) regardless of the strength of the external vibration excitation, thus avoiding signal saturation or excessively low signal-to-noise ratio.

[0079] Finally, data stream synchronization: Add high-precision timestamps from the vehicle CAN bus, which are the same as those for the vibration environment data acquisition, to the data stream of the sensing signal after the above filtering and normalization processes. Thus, two data streams with strictly synchronized time are obtained: the raw optical fiber sensing signal stream after preliminary conditioning and the spatio-temporal feature tensor stream of the vehicle vibration environment field, laying a foundation for subsequent joint processing.

[0080] Calculation of the dominant vibration frequency band and energy level:

[0081] Dominant vibration frequency band: Perform spatial averaging on the energy characteristics of all nodes to obtain the average energy Eavgband of each frequency band within the entire vehicle. The dominant vibration frequency band is the one with the largest Eavgband value.

[0082] Vibration energy level: Calculate the total energy mean Etotalavg of all nodes and all frequency bands. Preset two thresholds Thlow and Thhigh (calibrated through experiments). If Etotalavg < Thlow, it is a low level; if Thlow ≤ Etotalavg < Thhigh, it is a medium level; if Etotalavg ≥ Thhigh, it is a high level.

[0083] Adaptive configuration of wavelet threshold denoising parameters:

[0084] Threshold calculation: Use a general threshold (VisuShrink) or a threshold based on the estimation of the noise variance of the current signal segment;

[0085] For example, for a high vibration energy level, the threshold can be set as: = *sqrt(2*1og(M))*β, where is the standard deviation of the noise estimated from the detail coefficients of the highest layer of wavelet decomposition, M is the signal length, and β is a scaling factor set according to the energy level (e.g., low: 0.8, medium: 1.0, high: 1.2), which is used for more aggressive filtering during strong vibrations.

[0086] Feature enhancement and extraction steps: Used to perform spatio-temporal alignment and feature-level fusion on the synchronized sensing signal stream and vibration environment stream. Through a pre-trained vibration noise and signal decoupling model, separate the noise components strongly related to the vibration environment and potential target signal components, and enhance and reconstruct the target signal components;

[0087] Furthermore, when the feature enhancement and extraction steps execute the pre-trained vibration noise and signal decoupling model to separate the noise components and target signal components, the specific execution method is as follows:

[0088] A deep feature decoupling network is constructed, which takes the spatiotemporal feature tensor of the vehicle vibration environment field as the conditional input and the original sensing signal stream as the main input. The deep feature decoupling network includes a shared encoder, a noise-dedicated decoding branch and a signal-dedicated decoding branch.

[0089] The network is pre-trained using vehicle-mounted bench test data, which includes bare fiber sensing signals under the combined action of known vibration input and known target signal input, as well as pure noise signals under individual vibration excitation.

[0090] The real-time acquired synchronous sensing signal stream and the spatiotemporal characteristic tensor of the vehicle vibration environment field are input into the trained decoupled model. The noise-dedicated decoding branch outputs the estimated vibration-induced noise component, and the signal-dedicated decoding branch outputs the initially enhanced target signal component.

[0091] The estimated noise component is subtracted from the original sensing signal stream and then weighted and fused with the output of the dedicated signal decoding branch to obtain the enhanced and reconstructed target signal component.

[0092] First, a deep feature decoupling network is constructed: a neural network model is built with the spatiotemporal feature tensor of the vibration environment as the conditional input and the pre-conditioned sensor signal fragment as the main input. The core structure of this model includes: a shared encoder for extracting deep mixed features from the original signal; a noise-dedicated decoding branch, whose input is the features of the shared encoder and the vibration environment tensor (as conditional information), and outputs the estimated noise signal component induced by the current vibration environment; and a signal-dedicated decoding branch, whose input is also the features of the shared encoder and the vibration environment tensor, and outputs the pre-reconstructed target signal component.

[0093] Secondly, pre-training is performed using bench test data: During the network training phase, data obtained from vehicle-mounted bench tests is used. Specifically, on the bench, known and reproducible vibration excitations (simulating different road conditions) are applied to the vehicle, and the output signal of the bare optical fiber at this time is collected as a pure noise signal. On the other hand, when the vehicle is stationary (quiet environment), a known target test signal (such as sinusoidal strain at a specific frequency) is input to the system under test (such as a bare optical fiber used to monitor structural strain), and the output is collected as an ideal signal. Then, while applying vibration excitation, the input target test signal is superimposed, and the output at this time is collected as a mixed signal. The training data pair is: input = mixed signal + corresponding vibration environment spatiotemporal feature tensor, noise branch target output = pure noise signal, signal branch target output = ideal signal. By training the network with a large amount of such data, it learns to decouple noise and signal from the mixed signal under given vibration environment characteristics.

[0094] Next, real-time inference performs feature decoupling: During actual vehicle operation, the synchronously acquired sensor signal stream (current segment) and the corresponding vibration environment spatiotemporal feature tensor are input into the pre-trained decoupling model. The model workflow is as follows: The shared encoder first encodes the sensor signal; then, the vibration environment tensor, as conditional information, guides the noise decoding branch and the signal decoding branch to decode, respectively. The noise-dedicated decoding branch outputs the estimated vibration-induced noise component Nest in the current signal, and the signal-dedicated decoding branch outputs a preliminary enhanced target signal component Spre. Finally, weighted fusion is performed to obtain the enhanced signal: To obtain a more robust enhancement result, a fusion strategy is adopted: On the one hand, the estimated noise component Nest is directly subtracted from the original sensor signal sraw to obtain the residual signal sresidue = Sraw - Nest; on the other hand, the output spre of the signal decoding branch is used, and the final enhanced reconstructed target signal component senhanced is obtained through weighted fusion.

[0095] senhanced=a*Sresidue+(1-a)*Spre, where the weight α can be dynamically adjusted according to the feedback of the signal quality assessment step (initially set to 0.5). The core of this step is to intelligently strip away the sound that is strongly correlated with vibration by using the pre-trained model and real-time vibration environment information, while retaining and enhancing the potential target sensing signal.

[0096] Deep Feature Decoupling Network Structure and Training:

[0097] Network structure: The shared encoder consists of three one-dimensional convolutional layers (kernel size = 5, number of channels = 16, 32, 64) and a max pooling layer. The two decoding branches have a symmetrical structure, including transposed convolutional layers and skip connections (connected to the corresponding layers of the encoder). The vibration environment tensor is compressed into a conditional vector c through a fully connected network and concatenated onto the feature map at each layer of the decoder.

[0098] Training data acquisition: In vehicle-mounted bench tests:

[0099] Pure noise signal (Npure): Apply a known vibration excitation (such as white noise input from a hydraulic exciter or actual road spectrum) to the vehicle and record the output from the bare optical fiber.

[0100] Ideal signal s1dea1: In a quiet environment, a bare optical fiber is excited by a known target signal generator (such as strain introduced by a piezoelectric ceramic at a known frequency), and the output is recorded.

[0101] Mixed signal (Smixed): Simultaneously apply Vexcitation and input target signal, and record the output.

[0102] The training sample pairs are: Input = (Smixed segment, corresponding Tensor), Noise branch label = Npure segment, Signal branch label = Sideal segment.

[0103] Loss Function and Training: Total Loss

[0104] Ltotal = Lmse(Spre,Sideal) + Q * Lmse(Nest,Npure), where Q is the weight of the noise branch (initially 1.0), and the Adam optimizer is used for training.

[0105] 2. Real-time inference and weighted fusion:

[0106] Input: The synchronized sensing signal segment ssync(t) and the corresponding vibration environment tensor Tensor(t).

[0107] Output: After forward propagation of the model, Nest(t) and spre(t) are obtained.

[0108] Weighted fusion: Initial weight a=0.5 (prior assumption that noise and signal decoding branches are equally important), enhanced signal sentanced(t)=a*(Ssync(t)-Nest(t))+(1-a)*Spre(t).

[0109] Dynamic compensation step: Based on the vibration environment feature tensor, it is used to solve the model of the influence of vibration on the polarization state, birefringence and microbending additional loss of bare optical fiber in real time, generate dynamic compensation vector, and compensate for the transmission characteristic distortion of the enhanced signal.

[0110] Furthermore, the dynamic compensation step, when executing the model for real-time calculation of the impact of vibration on the polarization state, birefringence, and microbending-related losses of bare optical fibers, and generating dynamic compensation vectors, is specifically executed as follows:

[0111] A simplified physical heuristic model is established, with vibration acceleration and spectrum as inputs and polarization state rotation matrix, equivalent birefringence change, and microbending loss coefficient as outputs.

[0112] The spatiotemporal characteristic tensor of the vehicle vibration environment field is spatially mapped along the bare optical fiber deployment path to obtain the local vibration excitation vector distributed along the fiber length direction.

[0113] By inputting the local vibration excitation vector into the influence model, the polarization state disturbance vector, birefringence change vector, and additional loss vector at each point along the fiber length direction are calculated.

[0114] The polarization fading compensation and phase deblurring correction of the enhanced signal are performed by using the polarization state perturbation vector and the birefringence change vector. The signal strength is then compensated by using the additional loss vector. Finally, the dynamic compensation vector is output and applied to the feature-enhanced signal.

[0115] First, a simplified physical heuristic model is established: Based on fiber optics theory, a set of parameterized simplified mathematical models is established. This model takes local vibration information (acceleration amplitude, spectral centroid) as input and outputs three physical quantities: 1. Polarization state rotation matrix, which describes the random rotation of the polarization state caused by the change in biaxial stress of the fiber due to vibration; 2. Equivalent birefringence change, which describes the instantaneous change in the refractive index distribution inside the fiber caused by vibration; 3. Microbending loss coefficient, which describes the additional light intensity attenuation caused by the microbending of the fiber due to vibration. The model parameters are determined through laboratory calibration (measuring the polarization, phase, and light intensity response of bare fiber under different vibration conditions).

[0116] Secondly, spatial mapping of the vibration environment: Based on the actual deployment path of the bare optical fiber on the vehicle (e.g., pasted along the door frame), the three-dimensional coordinates of a series of discrete points on the optical fiber path are obtained. The spatiotemporal characteristic tensor of the vehicle vibration environment field (which contains spatial distribution information) is converted into a local vibration excitation vector distributed along the length of the optical fiber deployment path through coordinate mapping and interpolation. This vector describes the vibration acceleration and spectral characteristics felt at each point on the optical fiber.

[0117] Next, the physical perturbation vector along the optical fiber is calculated: the local vibration excitation vectors obtained above for each discrete point on the optical fiber path are sequentially input into the physical heuristic simplified model. The model calculates and outputs the following for each point along the length of the optical fiber: 1. Polarization state perturbation vector (a Jones matrix or Stokes parameter change that characterizes polarization rotation); 2. Birefringence change vector (a scalar that represents the change in phase delay); 3. Additional loss vector (a scalar that represents the attenuation coefficient of light intensity).

[0118] Finally, a dynamic compensation vector is generated and applied: the calculated perturbation vector is used to compensate the Sensitized signal obtained in the feature enhancement step. Specifically: using the polarization state perturbation vector, polarization fading compensation is performed on the signal through a digital polarization control algorithm (such as digital signal processing based on polarization diversity reception) to stabilize the signal amplitude; using the birefringence change vector, phase deambiguation correction is performed on the sensing signal involving phase demodulation (such as based on the interference principle) to correct the additional phase noise introduced by vibration; and using the additional loss vector, inverse gain compensation is performed on the signal intensity (e.g., proportional amplification compensation of the signal amplitude based on the estimated loss coefficient). All the above compensation operations are integrated into a comprehensive dynamic compensation vector (including amplitude adjustment coefficient, phase correction amount, and polarization compensation parameters), which is directly applied to the Sensitized signal to obtain... The enhanced signal after dynamic compensation is scompensated. This step aims to actively counteract or mitigate the negative impact of vibration on the fiber optic sensing mechanism itself (rather than just the signal content) at the physical level.

[0119] Local vibration excitation vector mapping:

[0120] Suppose the bare optical fiber deployment path is discretized into M points, each with coordinates (xm, ym, zm). For each point, in the latest time slice of the spatiotemporal feature tensor, the dominant vibration frequency band bandm, dominant frequency fmainm, and energy Em at the spatial grid where the point is located are obtained through nearest neighbor or bilinear interpolation. The root mean square value of acceleration at the point, armsm = sqrt(Em), is estimated (assuming that the energy is proportional to the square of the acceleration). This forms the local vibration excitation vector Lm = [bandm, fmainm, armsm] at the point. The Lm values ​​of all M points constitute an Mx3 matrix.

[0121] Example of physics-based heuristic model computation:

[0122] Microbending loss coefficient Ym:

[0123] Ym = kloss * armsm^2.

[0124] Among them, k1oss is obtained through calibration experiments: In the laboratory, a single-frequency vibration with a known acceleration acal is applied to a section of bare optical fiber of the same type, and its transmitted optical power attenuation APcal is measured. Then kloss=APcal / (Pe*acal^2*Lfiber), where Pe is the input optical power and Lfiber is the length of the vibrated section of the optical fiber.

[0125] Polarization state rotation matrix Jm:

[0126] It simplifies to a random rotation related to the vibration frequency and amplitude. It can be modeled as a 2x2 Jones matrix generated by a pseudo-random number generator with the current time, fmainm, and armsm as seeds, whose parameters satisfy the unit modulus constraint.

[0127] Equivalent birefringence change ABm:

[0128] ABm=kb*armsm*sin(2π*fmainm*t+Φm).

[0129] kb is calibrated by applying vibration in the fiber interferometer and measuring the resulting phase change Ccal. Then kb = Ccal / (acal*Lfiber*sin(2πfcalt)), where Φm is the position-dependent random initial phase.

[0130] Dynamic compensation vector generation:

[0131] Amplitude compensation coefficient Gm: Gm = exp(+ym*Lsegment), where Lsegment represents the length of the fiber segment at each discrete point. It is used to compensate for light intensity attenuation.

[0132] Phase compensation amount compm: compm = -ABm * Lsegment. Used for subsequent digital signal processing.

[0133] Phase subtraction is performed in the process.

[0134] Polarization compensation parameter: the calculated inverse Jones matrix Jm{-1}, used to digitally rotate the received optical signal to cancel out disturbances.

[0135] Signal quality assessment steps: used to calculate the time-frequency domain quality indicators of the compensated enhanced signal, and compare them with the preset signal-to-noise ratio and distortion thresholds to generate real-time signal quality assessment factors and confidence labels;

[0136] Furthermore, the signal quality assessment step, when calculating the time-frequency domain quality index of the compensated enhanced signal, is executed as follows:

[0137] Under quiet operating conditions where the target signal is known, a reference signal is acquired, and its power spectrum is calculated as the reference spectrum.

[0138] For the signal enhanced by the dynamic compensation vector, calculate its spectral distortion compared to the reference spectrum, the signal-to-noise ratio in the target characteristic frequency band, and the short-time zero-crossing rate or envelope smoothness of the signal.

[0139] The calculated spectral distortion, signal-to-noise ratio, and smoothness indices are compared with the high-quality signal thresholds preset according to the application scenario.

[0140] If all indicators are better than the threshold, high-quality evaluation factors and high-confidence labels are generated; if any indicator is lower than the threshold, medium- and low-quality evaluation factors and corresponding low-confidence labels are generated according to the degree of deviation, and the type of indicator that does not meet the standard is fed back to the collaborative control steps.

[0141] First, establish a reference spectrum: Under the condition that the vehicle is stationary in a quiet environment (such as an underground garage) and the target signal source is known or controllable, collect a bare optical fiber sensing signal for a sufficiently long time as a reference signal, calculate the average power spectral density (PSD) of the reference signal, and use it as the reference spectrum, which represents the spectral shape of a high-quality signal under vibration-free interference.

[0142] Secondly, calculate real-time quality indicators: For the dynamically compensated and enhanced signal Scompensated obtained through real-time processing, calculate three core quality indicators for segments of fixed duration (e.g., 1 second): 1. Spectral distortion: Calculate the difference between the power spectrum of the current signal segment and the reference spectrum, commonly using spectral distortion distance (e.g., Itakura-Saito distance). The smaller the value, the closer the spectrum shape is to the ideal state. 2. Target band signal-to-noise ratio: Within a frequency band where the characteristics of the target signal are known (e.g., the characteristic frequency band used to monitor bearing faults), calculate the ratio of signal power to out-of-band noise power (expressed in decibels dB). 3. Signal smoothness: Calculate the short-time zero-crossing rate of the signal (sensitive to high-frequency noise) or the smoothness of the envelope (calculate its variance after obtaining the envelope through Hilbert transform). The lower the value, the smoother the signal, and the fewer the impulse noise and spikes.

[0143] Next, the index comparison and quality classification are performed: the calculated values ​​of spectral distortion, signal-to-noise ratio and smoothness index are compared with the high-quality signal thresholds preset according to the specific application scenario (such as structural health monitoring and acoustic detection). For example, the thresholds are set as follows: spectral distortion < 2.0, signal-to-noise ratio > 20dB, and smoothness index < 0.1.

[0144] Finally, evaluation factors and feedback are generated: if all indicators of the current signal segment are better than the preset threshold, a high-quality evaluation factor (e.g., value 1.0) is generated and marked as a high-confidence label. If any one or more indicators are lower than the threshold, a medium-quality (e.g., 0.5) or low-quality (e.g., 0.1) evaluation factor is generated based on the number of indicators lower than the threshold and the degree of deviation, and marked as a low-confidence label. At the same time, it is clearly recorded which specific indicators (e.g., signal-to-noise ratio, spectral distortion) did not meet the standard, and these types of indicators that did not meet the standard, together with the quality evaluation factor, are output and fed back to the collaborative control step.

[0145] Spectral Distortion DIS:

[0146] Calculate Pcurr(k) for a 1-second segment of Scompensated:

[0147] DIS = (1 / K) * ∑{k = 1}^{K}[Pcurr(k) / Pref(k) - log(Pcurr(k) / Pref(k)) - 1].

[0148] High-quality threshold: DISthresh = 2.0;

[0149] Target frequency band signal-to-noise ratio SNR:

[0150] Let the target signal frequency band be Bs = [9Hz, 11Hz] (corresponding to the 10Hz strain signal).

[0151] The noise bands are taken as Bnleft = [5Hz, 7Hz] and Bnright = [13Hz, 15Hz].

[0152] Psignal = mean(Pcurr(k) for k in Bs).

[0153] Pnoise = (mean(Pcurr(k) for k in Bnleft) + mean(Pcurr(k) for k in Bnright)) / 2.

[0154] SNR = 10 * log10(Psignal / Pnoise).

[0155] High-quality threshold: SNRthresh = 20dB.

[0156] Signal smoothness Smoothness:

[0157] Calculate the envelope e(n) of Scompensated.

[0158] Calculate the coefficient of variation of the envelope:

[0159] Smoothness = std(e(n)) / mean(e(n)).

[0160] High-quality threshold: Smoothnessthresh = 0.1.

[0161] Calculation of the quality assessment factor factor:

[0162] If DIS < DISthresh and SNR > SNRthresh AND;

[0163] Smoothness < Smoothnessthresh, then Qfactor = 1.0 (high quality).

[0164] Otherwise, factor = 0.5 * (number of qualified indicators / 3). For example, if only one indicator is qualified, then...

[0165] Qfactor = 0.5 * (1 / 3) is approximately equal to 0.17 (low quality).

[0166] Cooperative control steps: These steps are used to receive signal quality assessment factors, dynamically adjust the parameter weights of the vibration noise-signal decoupling model, the coefficients of the dynamic compensation model, and the preprocessing strategies of the signal synchronous acquisition and conditioning steps, forming a closed-loop control flow with the goal of optimizing the output signal quality.

[0167] Furthermore, the coordinated control step, when dynamically adjusting the parameters of each step to form a closed-loop control flow, is executed in the following specific manner:

[0168] The quality assessment factors and types of non-compliant indicators fed back from the received signal quality assessment steps;

[0169] If the signal-to-noise ratio does not meet the standard, increase the weight of the loss function of the noise-dedicated decoding branch in the feature enhancement and extraction steps, or guide the band-stop filter to tighten the bandwidth in the signal synchronization acquisition and conditioning steps.

[0170] If the spectral distortion does not meet the standard, adjust the parameters that affect the model in the dynamic compensation step, or reduce the intensity of noise separation to prevent signal distortion.

[0171] If the evaluation factors remain of high quality, the processing parameters will be gradually relaxed to reduce the computational complexity of the system and achieve energy-saving operation.

[0172] The adjusted parameter set is sent to the corresponding steps in real time, and the changes in the signal quality assessment factor in the next cycle are monitored to form an adaptive optimization closed loop with quality assessment as feedback.

[0173] First, receive quality feedback: continuously receive quality assessment factors and types of non-compliant indicators from the signal quality assessment steps.

[0174] Secondly, analyze the causes and decide on parameter adjustments: Based on the feedback information, analyze the possible causes and decide which steps to adjust the parameters in the preceding steps: 1. If the feedback indicates that the signal-to-noise ratio is not up to standard, it indicates that the vibration noise separation is incomplete or that additional noise has been introduced. In this case, guide the feature enhancement and extraction steps to increase the weight of the noise-dedicated decoding branch in the loss function (reflected in model fine-tuning or fusion weight α), forcing the model to pay more attention to noise suppression; or guide the signal synchronous acquisition and conditioning steps to tighten the bandwidth of the band-stop filter and perform more aggressive filtering. 2. If the feedback indicates that the spectral distortion is not up to standard, it indicates that the signal may have been overprocessed, resulting in distortion, or that the dynamic compensation model is inaccurate. In this case, guide the dynamic compensation steps to fine-tune the empirical parameters in the physical heuristic simplified model (such as adjusting the mapping relationship between the loss coefficient and vibration acceleration); or guide the feature enhancement and extraction steps to appropriately reduce the intensity of noise separation (such as reducing α) to prevent useful target signal features from being filtered out as noise. 3. If the feedback indicates that the smoothness is not up to standard, it may be due to impact noise or improper compensation. Adjustments can be made by referring to the strategies for signal-to-noise ratio or spectral distortion.

[0175] Next, optimization and energy-saving operation: If the quality assessment factor remains high for multiple cycles, it indicates that the current processing parameters are in a relatively good state. The system can gradually and slightly relax certain processing parameters (such as slightly increasing the filtering bandwidth and reducing the model calculation complexity) to reduce the overall computational load of the system while maintaining high quality, thereby achieving energy-saving operation.

[0176] Finally, a closed-loop control is formed: the adjusted parameter set determined after the above analysis and decision (such as new filter bandwidth, new model fusion weight α, and new physical model parameters) is sent to the corresponding steps (signal synchronous acquisition and conditioning, feature enhancement and extraction, and dynamic compensation) in real time. The system runs with the new parameters and performs signal quality evaluation again in the next processing cycle. Through this continuous evaluation-feedback-adjustment-re-evaluation cycle, an adaptive optimization closed loop with the final output signal quality as the optimization goal is formed, ensuring that the method can automatically maintain the best signal enhancement effect under different vibration environments.

[0177] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for enhancing the quality of bare optical fiber sensing signals in vehicle-mounted vibration environments, characterized by: It includes the following steps: vibration environment perception, signal synchronous acquisition and conditioning, feature enhancement and extraction, dynamic compensation, signal quality assessment, and collaborative control. Vibration environment perception step: It is used to collect multi-dimensional vibration acceleration, angular velocity and spectral characteristic data of key measurement points of the vehicle in real time through multi-axis inertial measurement unit and vibration sensing network, and construct the spatiotemporal characteristic tensor of vehicle vibration environment field; Signal Synchronization Acquisition and Conditioning Steps: The raw signal output by the bare fiber sensing system is synchronously acquired, and the preprocessing parameters are adaptively configured according to the spatiotemporal characteristic tensor of the vehicle vibration environment field to perform preliminary noise reduction and normalization on the raw signal, so as to obtain the synchronized sensing signal stream and vibration environment stream. Feature enhancement and extraction steps: used to perform spatiotemporal alignment and feature-level fusion of synchronized sensing signal streams and vibration environment streams. Through a pre-trained vibration noise and signal decoupling model, noise components strongly correlated with the vibration environment and potential target signal components are separated, and the target signal components are enhanced and reconstructed. Dynamic compensation step: Based on the vibration environment feature tensor, it is used to solve the model of the influence of vibration on the polarization state, birefringence and microbending additional loss of bare optical fiber in real time, generate dynamic compensation vector, and compensate for the transmission characteristic distortion of the enhanced signal. Signal quality assessment steps: used to calculate the time-frequency domain quality indicators of the compensated enhanced signal, and compare them with the preset signal-to-noise ratio and distortion thresholds to generate real-time signal quality assessment factors and confidence labels; Cooperative control steps: These steps are used to receive signal quality assessment factors, dynamically adjust the parameter weights of the vibration noise-signal decoupling model, the coefficients of the dynamic compensation model, and the preprocessing strategies of the signal synchronous acquisition and conditioning steps, forming a closed-loop control flow with the goal of optimizing the output signal quality.

2. The method for enhancing the quality of bare optical fiber sensing signals for vehicle-mounted vibration environments according to claim 1 is characterized in that, The vibration environment sensing step, when executing the real-time acquisition of multi-dimensional vibration data from key vehicle measurement points via a multi-axis inertial measurement unit and vibration sensing network, and constructing the spatiotemporal characteristic tensor of the vehicle vibration environment field, is specifically executed as follows: Three-axis accelerometers and gyroscopes are deployed at the vehicle chassis suspension connection points, powertrain mounting points, and key sections of the body longitudinal beams to form a vibration sensing network. Using the vehicle CAN bus time as a reference, the data of each node in the vibration sensing network is synchronized in time. Short-time Fourier transform is performed on the triaxial acceleration and angular velocity signals of each node to extract the energy, dominant frequency, and spectral kurtosis features of each node within a preset frequency band, forming a node vibration feature vector; Based on the spatial coordinates of the sensor nodes, the vibration feature vectors of all nodes at the same time are spatially interpolated and meshed to generate a spatial feature map describing the vibration energy and mode distribution of the whole vehicle. By stacking spatial feature maps according to time series, a spatiotemporal feature tensor of the vehicle vibration environment field with time, space, two dimensions, and feature channel dimensions is formed.

3. The method for enhancing the quality of bare optical fiber sensing signals for vehicle-mounted vibration environments according to claim 1 is characterized in that, The specific execution method of the signal synchronous acquisition and conditioning step, when performing preliminary noise reduction and normalization on the original signal by adaptively configuring preprocessing parameters based on the spatiotemporal characteristic tensor of the vehicle vibration environment field, is as follows: Extract the current dominant vibration frequency band range and vibration energy level from the spatiotemporal characteristic tensor of the vehicle vibration environment field; Based on the dominant vibration frequency band, the basis functions and threshold parameters of the band-stop filter bank or wavelet threshold denoising are adaptively configured to perform targeted frequency domain notch filtering or time-frequency domain filtering on the original fiber optic sensing signal. Based on the vibration energy level, the gain of the signal amplification circuit or the scaling factor of the digital signal is dynamically adjusted to normalize the signal amplitude to the preset dynamic range center interval. The filtered and normalized sensor signal stream and the corresponding vibration environment stream are given a unified timestamp.

4. The method for enhancing the quality of bare optical fiber sensing signals for vehicle-mounted vibration environments according to claim 1 is characterized in that, The feature enhancement and extraction step, when executing the pre-trained vibration noise and signal decoupling model to separate the noise component and the target signal component, is specifically executed as follows: A deep feature decoupling network is constructed, which takes the spatiotemporal feature tensor of the vehicle vibration environment field as the conditional input and the original sensing signal stream as the main input. The deep feature decoupling network includes a shared encoder, a noise-dedicated decoding branch and a signal-dedicated decoding branch. The network is pre-trained using vehicle-mounted bench test data, which includes bare fiber sensing signals under the combined action of known vibration input and known target signal input, as well as pure noise signals under individual vibration excitation. The real-time acquired synchronous sensing signal stream and the spatiotemporal characteristic tensor of the vehicle vibration environment field are input into the trained decoupled model. The noise-dedicated decoding branch outputs the estimated vibration-induced noise component, and the signal-dedicated decoding branch outputs the initially enhanced target signal component. The estimated noise component is subtracted from the original sensing signal stream and then weighted and fused with the output of the dedicated signal decoding branch to obtain the enhanced and reconstructed target signal component.

5. The method for enhancing the quality of bare optical fiber sensing signals for vehicle-mounted vibration environments according to claim 1 is characterized in that, The dynamic compensation step is executed as follows when performing real-time calculation of the impact model of vibration on the polarization state, birefringence, and microbending-related loss of bare optical fiber, and generating a dynamic compensation vector: A simplified physical heuristic model is established, with vibration acceleration and spectrum as inputs and polarization state rotation matrix, equivalent birefringence change, and microbending loss coefficient as outputs. The spatiotemporal characteristic tensor of the vehicle vibration environment field is spatially mapped along the bare optical fiber deployment path to obtain the local vibration excitation vector distributed along the fiber length direction. By inputting the local vibration excitation vector into the influence model, the polarization state disturbance vector, birefringence change vector, and additional loss vector at each point along the fiber length direction are calculated. The polarization fading compensation and phase deblurring correction of the enhanced signal are performed by using the polarization state perturbation vector and the birefringence change vector. The signal strength is then compensated by using the additional loss vector. Finally, the dynamic compensation vector is output and applied to the feature-enhanced signal.

6. The method for enhancing the quality of bare optical fiber sensing signals for vehicle-mounted vibration environments according to claim 1 is characterized in that, The signal quality assessment step, when calculating the time-frequency domain quality index of the compensated enhanced signal, is specifically executed as follows: Under quiet operating conditions where the target signal is known, a reference signal is acquired, and its power spectrum is calculated as the reference spectrum. For the signal enhanced by the dynamic compensation vector, calculate its spectral distortion compared to the reference spectrum, the signal-to-noise ratio in the target characteristic frequency band, and the short-time zero-crossing rate or envelope smoothness of the signal. The calculated spectral distortion, signal-to-noise ratio, and smoothness indices are compared with the high-quality signal thresholds preset according to the application scenario. If all indicators are better than the threshold, high-quality evaluation factors and high-confidence labels are generated; if any indicator is lower than the threshold, medium- and low-quality evaluation factors and corresponding low-confidence labels are generated according to the degree of deviation, and the type of indicator that does not meet the standard is fed back to the collaborative control steps.

7. The method for enhancing the quality of bare optical fiber sensing signals for vehicle-mounted vibration environments according to claim 6 is characterized in that, The coordinated control step, when dynamically adjusting the parameters of each step to form a closed-loop control flow, is executed in the following manner: The quality assessment factors and types of non-compliant indicators fed back from the received signal quality assessment steps; If the signal-to-noise ratio does not meet the standard, increase the weight of the loss function of the noise-dedicated decoding branch in the feature enhancement and extraction steps, or guide the band-stop filter to tighten the bandwidth in the signal synchronization acquisition and conditioning steps. If the spectral distortion does not meet the standard, adjust the parameters that affect the model in the dynamic compensation step, or reduce the intensity of noise separation to prevent signal distortion. If the evaluation factors remain of high quality, the processing parameters will be gradually relaxed to reduce the computational complexity of the system and achieve energy-saving operation. The adjusted parameter set is sent to the corresponding steps in real time, and the changes in the signal quality assessment factor in the next cycle are monitored to form an adaptive optimization closed loop with quality assessment as feedback.

Images