Polarimetric ranging method and system based on convolutional auto-encoding model for noise elimination

Abstract

The present application relates to the technical field of polarization signal noise reduction, and particularly relates to a polarization ranging method and system for eliminating noise based on a convolutional auto-encoding model, comprising obtaining a polarization detection signal, generating a to-be-detected optical signal by using a phase modulator, and introducing a sinusoidal phase delay in the phase modulator; converting the to-be-detected optical signal into an analog electrical signal by using a photodetector; performing noise reduction on the to-be-detected optical signal by using a band-pass filter; the convolutional auto-encoding model is composed of an encoder and a decoder, the encoder is stacked by a convolutional layer and a ReLU activation function; the decoder is stacked by a transposed convolution and a ReLU function, and the encoder and the decoder are connected through a bottleneck layer; the in-phase frequency at two continuous light intensity minimum values is obtained from the signal denoised by the convolutional auto-encoding model, and the to-be-detected distance is calculated. The present application solves the problem that the effect is not ideal when the existing method denoises the noise at the wave peaks and wave troughs in the polarization signal.

Description

Technical Field

[0001] The present invention relates to the technical field of polarization signal noise reduction, and in particular to a polarization ranging method and system for eliminating noise based on a convolutional autoencoder model. Background Art

[0002] There are many different laser ranging technologies in use in the world today, but each of these ranging technologies has its own drawbacks. Although the timed pulse method is widely used, its resolution is limited and therefore not suitable for industrial applications.

[0003] The phase comparison method has fast response and strong anti-interference ability, but is affected by challenges such as 2-claw entanglement and intensity interference and is not suitable for high-precision measurement.

[0004] FMCW ranging has the advantages of high accuracy and no ranging blind spots, but is limited by the modulation bandwidth and light source linearity, resulting in a limited measurement range; in contrast, the femtosecond optical comb ranging method shows exceptional ranging accuracy and ranging capabilities; however, the complexity of the system and the associated maintenance costs make it less suitable for engineering applications.

[0005] Polarization modulation ranging technology has important engineering applicability because it eliminates the need for phase recognition, reduces distance ambiguity, and exhibits strong anti-interference capabilities. Polarization modulation ranging technology uses an electro-optical technology modulator to perform two consecutive frequency-sweep-polarization modulations on the measuring light, followed by interference demodulation. However, when the polarization modulation ranging device draws the curve that forms the detection image, noise is generated at the troughs and peaks, which will affect the numerical value of the trough position and thus the ranging accuracy.

[0006] Common noise reduction methods include: using Gaussian filtering, median filtering, and mean filtering to remove noise; it is also beneficial to filter the spatial noise by performing singular value decomposition (SVD) or principal component analysis (PCA) on the low-frequency sub-matrix.

[0007] However, Gaussian filtering is less effective in processing random noise or non-Gaussian noise, as it only smoothes the distribution of noise but does not remove the noise itself very well. Median filtering is less effective in processing continuous noise. Mean filtering is better at processing signals with Gaussian noise, but does not focus on signal correlation when the noise intensity is large. SVD is less effective when the noise is large. PCA is more sensitive to linearly correlated signals and noise, but has poor results in processing nonlinearly correlated or high-order correlated signals and noise. Summary of the Invention

[0008] In view of the shortcomings of the existing methods, the present invention constructs a convolutional autoencoder model with a simple structure, which can effectively remove the noise at the trough during polarization modulation ranging, and the denoising effect is better than the existing methods.

[0009] The technical solution adopted by the present invention is: a polarization ranging method for eliminating noise based on a convolutional autoencoder model comprises the following steps:

[0010] Step 1: Obtain a polarization detection signal, use a phase modulator to generate a light signal to be measured, and introduce a sinusoidal phase delay into the phase modulator; use a photodetector to convert the light signal to be measured into an analog electrical signal;

[0011] Step 2: Using a bandpass filter to perform denoising and normalization on the light signal to be measured;

[0012] As a preferred embodiment of the present invention, the bandpass filter includes: a second-order bandpass filter, a resonant bandpass filter, a SAW bandpass filter, and a BAW bandpass filter.

[0013] Step 3: Input the measured optical signal data with noise at the trough and / or peak positions into the convolutional autoencoder model. The convolutional autoencoder model consists of an encoder and a decoder. The encoder is a stack of convolutional layers and ReLU activation functions; the decoder is a stack of transposed convolution and ReLU functions. The encoder and decoder are connected by a bottleneck layer.

[0014] As a preferred embodiment of the present invention, the convolutional autoencoder model includes:

[0015] The encoder converts the light signal X to be measured, which contains noise i Compressed into the feature F1 of the latent space, the bottleneck layer extracts the i The most relevant feature F2; the decoder receives the screening results of the bottleneck layer, maps the encoded data back to the original input layer and decodes the features; calculates the reconstruction loss, and optimizes the network parameters through back propagation to minimize the reconstruction error.

[0016] As a preferred embodiment of the present invention, the encoder formula is:

[0017] F1 k =σ(X i *w k +b k ) (1)

[0018] Among them, k means there are k convolution kernels, each of which is determined by the parameter w k and b k Composition, F1 k Represents the output of the encoder, and σ() is the Relu activation function.

[0019] As a preferred embodiment of the present invention, the formula of the bottleneck layer is:

[0020] F2 k =g(F1k *W+b) (2)

[0021] Among them, W is the weight of the bottleneck layer, b is the bias term of the bottleneck layer, and g() is the nonlinear activation function.

[0022] As a preferred embodiment of the present invention, the formula of the decoder is:

[0023]

[0024] in, Represents a flip operation on the two dimensions of the weight, and C is the same transpose for each input channel.

[0025] Step 4: Obtain the in-phase frequencies at two consecutive light intensity minima from the denoised signal of the convolutional autoencoder model and calculate the distance to be measured;

[0026] As a preferred embodiment of the present invention, the formula for calculating the distance to be measured is:

[0027]

[0028] Where f1 and f2 are the in-phase frequencies at two consecutive light intensity minima, c is the speed of light, and [] is a rounding operation.

[0029] As a preferred embodiment of the present invention, a polarization ranging system for eliminating noise based on a convolutional autoencoder model includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement a polarization ranging method for eliminating noise based on a convolutional autoencoder model.

[0030] As a preferred embodiment of the present invention, a computer-readable medium stores computer program code, and when the computer program code is executed by a processor, a polarization ranging method for eliminating noise based on a convolutional autoencoder model is implemented.

[0031] Beneficial effects of the present invention:

[0032] A convolutional autoencoder model is constructed to denoise the noisy polarization detection signal, and the denoising effect is obvious compared with the existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is the block diagram of the polarization detection system;

[0034] Figure 2 This is a flow chart of the polarization ranging method for eliminating noise using a convolutional autoencoder model of the present invention;

[0035] Figure 3 This is a schematic diagram of the convolutional autoencoder model;

[0036] Figure 4 This is a detailed structure diagram of the convolutional autoencoder model;

[0037] Figure 5 3 is a comparison chart between the method of the present invention and the existing noise reduction method. DETAILED DESCRIPTION

[0038] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore only shows the components related to the present invention.

[0039] like Figure 1 、 2 As shown, the polarization ranging method for eliminating noise based on the convolutional autoencoder model includes the following steps:

[0040] Step 1: Obtain a polarization detection signal and generate a signal to be measured by modulation of a phase modulator.

[0041] The continuous linearly polarized light emitted by the laser passes through an isolator and a polarization beam splitter at a 45-degree angle to the x-axis, generating equal-amplitude polarization components on the x- and y-axes. The phase modulator introduces a sinusoidal phase delay between the isolator and the polarization beam splitter, causing the frequency of the sinusoidal phase-delayed signal to change linearly. The fast axis of the quarter-wave plate is at a 45-degree angle to the x-axis, allowing the polarized light to propagate to the target and return along the same path before being demodulated by the phase modulator. After the demodulated light reaches the polarization beam splitter, polarization interference occurs. The resulting optical signal is then converted into an analog electrical signal by a photodetector. The analog electrical signal is converted by an A / D sampling module and then calculated by a computing unit to measure the distance.

[0042] The optical signal is modulated twice on a phase modulator, and the phase difference between the two modulated signals on the phase modulator is demodulated by a polarization beam splitter and received by a photodetector. Different modulation frequencies correspond to different wavelengths of the modulated signal. Signal errors detected by the photodetector include complete demodulation of the measured light at the phase zero point during polarization modulation ranging, causing the measured light to be immersed in background noise. This includes the 1 / f noise and dark current of the photodetector, thermal noise and shot noise within the electronic circuit, etc. Among them, 1 / f refers to a specific type of noise in order to express the inverse relationship between power spectral density and frequency. In the waveform output by the photodetector, the amplitude, signal-to-noise ratio and measurement resolution reach minimum values.

[0043] Step 2: Use a bandpass filter to process the signal output by the photodetector to remove the noise component within the frequency range of the signal to be measured, and perform normalization preprocessing on the filtered signal;

[0044] Bandpass filtering can use second-order bandpass filters, resonant bandpass filters, SAW bandpass filters, and BAW bandpass filters.

[0045] During the normalization preprocessing of the filtered signal, the signal can be subjected to zero-mean normalization preprocessing to ensure that the signal has a uniform zero mean before being input into the convolutional autoencoder model, which helps to improve the stability and training effect of the model while ensuring the consistency of the input data.

[0046] Step 3: Build a convolutional autoencoder model and put the preprocessed sample data into the convolutional autoencoder model;

[0047] like Figure 3 、 4 As shown in the figure, the convolutional autoencoder model is composed of an encoder and a decoder in the computing unit. The encoder is composed of a stack of convolutional layers and ReLU activation functions; the decoder is composed of a stack of transposed convolution and ReLU functions. The encoder and decoder are connected by a bottleneck layer.

[0048] The training data uses the preprocessed polarization detection signal X containing noise i As input, the encoder takes X i Compressed into the feature F1 of the latent space, the bottleneck layer extracts the feature F1 from the input data X i The most relevant feature F2, the decoder receives the screening results of the bottleneck layer, maps the encoded data back to the original input layer and decodes the features into a form closest to the real signal; calculates the reconstruction loss, and optimizes the network parameters through back propagation to minimize the reconstruction error.

[0049] The encoder formula can be written as follows:

[0050] F1 k =σ(X i *w k +b k ) (1)

[0051] Among them, k means there are k convolution kernels, each of which is determined by the parameter w k and b k Composition, F1 k Represents the output of the encoder, and σ() is the Relu activation function.

[0052] The bottleneck layer formula can be written as follows:

[0053] F2 k =g(F1 k *w+b) (2)

[0054] Among them, W is the weight of the bottleneck layer, b is the bias term of the bottleneck layer, and g() is the nonlinear activation function.

[0055] The g() activation function can introduce nonlinear factors. Without the activation function, each layer of the convolutional autoencoder model only makes linear changes, but the expressiveness of the linear model is not strong enough. By introducing the activation function using nonlinear factors, the output range can be controlled.

[0056] The decoder will get the feature F2 k Perform feature reconstruction, the formula is as follows:

[0057]

[0058] in, Represents a flip operation on the two dimensions of the weight, and C is the same transpose for each input channel.

[0059] Finally, the reconstruction loss is calculated, and the network parameters are optimized through back propagation to minimize the reconstruction error.

[0060] Repeat the above steps until the convolutional autoencoder model can accurately reconstruct the input data to obtain a noise-free signal; the convolutional autoencoder can capture the local correlation features of the signal and effectively learn the local structure in the signal.

[0061] Step 4: Based on the denoised signal, obtain the in-phase frequencies at two consecutive light intensity minima to calculate the distance to be measured;

[0062] The formula for calculating the distance to be measured is:

[0063]

[0064]

[0065]

[0066] Where L is the distance to be measured; f1 and f2 are the in-phase frequencies at two consecutive light intensity minima; is the rounding operation; and c is the speed of light.

[0067] like Figure 5 As shown in the figure, Input is the noise data after preprocessing; Output is the output of the SVD model, wavelet change model, mean filter model and convolutional autoencoder model; Y_true: noise-free signal data; SVD, wavelet change and mean filter models are compared with the model of the present invention; it can be seen that the denoising effect of the present invention at peaks and troughs is significantly better than that of the existing method.

[0068] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the contents of the specification and must be determined according to the scope of the claims.

Claims

1. A polarization ranging method for eliminating noise based on a convolutional autoencoder model, characterized in that: The following steps are involved: Step 1: Obtain a polarization detection signal, use a phase modulator to generate a light signal to be measured, and introduce a sinusoidal phase delay into the phase modulator; use a photodetector to convert the light signal to be measured into an analog electrical signal; Step 2: Using a bandpass filter to perform denoising and normalization on the light signal to be measured; Step 3: Input the measured optical signal data with noise at the trough and / or peak positions into a convolutional autoencoder model. The convolutional autoencoder model consists of an encoder and a decoder. The encoder is a stack of 3x3 convolutional layers and ReLU activation functions. The decoder is composed of a stack of 3x3 transposed convolution and ReLU functions, and the encoder and decoder are connected by a bottleneck layer; The convolutional autoencoder model includes: The encoder will contain noise to measure the optical signal X i Compressed into the feature F1 of the latent space, the bottleneck layer extracts the X i The most relevant feature F2; the decoder receives the screening results of the bottleneck layer, maps the encoded data back to the original input layer and decodes the features; calculates the reconstruction loss, and optimizes the network parameters through backpropagation to minimize the reconstruction error; The formula for the encoder is: （1） in, k Indicates that there is k convolution kernels, each of which is parameterized by and composition, represents the output of the encoder, is the Relu activation function; The formula for the bottleneck layer is: （2） in, W is the weight of the bottleneck layer, b is the bias term of the bottleneck layer, is a nonlinear activation function; The decoder formula is: （3） in, Indicates flipping the two dimensions of the weight. c The same transpose for each input channel; Step 4: Obtain the in-phase frequencies at two consecutive light intensity minima from the denoised signal of the convolutional autoencoder model and calculate the distance to be measured.

2. The polarization ranging method for eliminating noise based on a convolutional autoencoder model according to claim 1, wherein: The formula for calculating the distance to be measured is: in, and are the in-phase frequencies of two consecutive light intensity minima, c is the speed of light, This is the rounding operation.

3. The polarization ranging method for eliminating noise based on a convolutional autoencoder model according to claim 1, wherein: Bandpass filters include: second-order bandpass filters, resonant bandpass filters, SAW bandpass filters, and BAW bandpass filters.

4. A polarization ranging system based on a convolutional autoencoder model to eliminate noise, characterized in that: include: a memory for storing instructions executable by the processor; A processor, configured to execute instructions to implement the polarization ranging method for eliminating noise based on a convolutional autoencoder model as described in any one of claims 1 to 3.

5. A computer-readable medium storing computer program code, characterized in that When the computer program code is executed by a processor, the polarization ranging method for eliminating noise based on a convolutional autoencoder model is implemented as described in any one of claims 1 to 3.

Images