FMCW laser ranging signal denoising method based on neural network

Abstract

The invention discloses an FMCW laser ranging signal denoising method based on a neural network, and belongs to the technical field of laser ranging. The method comprises the following steps: firstly, performing high-pass filtering, demodulation and normalization processing on an original signal to eliminate system difference; extracting local frequency features by using MLP and expansion convolution; signal features are optimized through a residual network and jump connection; and finally, respectively outputting the purified signal and the target distance value by adopting a double-branch decoding structure. The simulation training data is generated based on Monte Carlo simulation and Mie scattering theories, and the aerosol scattering environment is truly restored. According to the method, the neural network is creatively used for processing the frequency domain characteristics of the FMCW signals, real target signals and environmental noise can still be accurately distinguished under the low visibility condition, and the accuracy and environmental adaptability of a ranging system are remarkably improved; the method has the advantages of end-to-end processing, strong anti-interference capability, good practicability and the like, and is particularly suitable for laser ranging application under severe weather conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of laser ranging, and in particular relates to a FMCW laser ranging signal denoising method based on a neural network. Background Art

[0002] Laser ranging (LDR) technology uses laser beams to detect echoes for high-precision distance perception. Laser ranging boasts high measurement accuracy and long detection range, making it widely used in industrial manufacturing, remote sensing, and other fields. However, in inclement weather such as fog, particles in the medium can scatter the laser beam emitted by the LiDAR, resulting in multiple echoes that affect accurate perception of the target object, limiting the application scope and accuracy of laser ranging and radar technologies.

[0003] Currently, existing lidar noise suppression research primarily focuses on pulsed lidar. Traditionally, adaptive filtering methods can reduce noise by approximately 6dB by constructing a reference signal model. In recent years, deep learning-based noise suppression methods, such as convolutional neural networks (CNNs), have emerged, enabling end-to-end learning of noise characteristics and achieving over 90% target echo recognition accuracy in smoky environments.

[0004] However, relatively little research has been conducted on noise suppression in FMCW lidar. FMCW laser ranging technology, based on coherent light interferometry, offers high ranging accuracy and robust interference immunity. However, because distance and velocity information are encoded in frequency components, signal processing complexity is increased. Scattered noise in inclement weather can introduce additional frequency components, leading to spectral aliasing and further complicating noise reduction.

[0005] Existing noise suppression methods for FMCW lidars fall into two main categories: threshold-based methods exploit the differences in the scattering properties of dual-wavelength lasers in a medium to distinguish targets from noise using a preset threshold; and cyclic detection algorithms screen for stable target signals through multiple observations and signal correlation analysis. For example, the correlation entropy spectral density (CSD) method achieves a 91% target recognition rate in low visibility. However, the performance of these existing methods degrades significantly in highly concentrated media, and more robust noise suppression methods are still needed. Summary of the Invention

[0006] The purpose of the present invention is to overcome the above-mentioned defects of the prior art and provide a method for denoising FMCW laser ranging signals based on a neural network, thereby suppressing the signal noise generated by FMCW laser ranging in an aerosol medium and simultaneously sensing the true distance of an object through a multi-task network structure.

[0007] The technical problem proposed by the present invention is solved as follows:

[0008] A method for denoising an FMCW laser ranging signal based on a neural network comprises the following steps:

[0009] Step 1: Regularization of laser ranging signals

[0010] The laser ranging electrical signal collected by the photodetector in the FMCW laser ranging system is regularized. First, high-pass filtering is performed to eliminate interference from the low-frequency components and collimating lens in the optical fiber. Then, signal demodulation is performed to eliminate the effect of the optical path difference between the reference and measurement optical paths of different FMCW laser ranging systems on the signal frequency. Low-pass filtering is performed to eliminate the sum frequency component of the demodulated signal while retaining the difference frequency component. Finally, the signal amplitude and variance are normalized.

[0011] Step 2: Signal local frequency perception

[0012] The regularized signal is input to the input layer of the neural network. The input layer includes a cascade of multi-layer perceptrons and dilated convolutional layers. The regularized signal is encoded and dilated convolutionalized to sense the local frequency of the signal.

[0013] Step 3: Network feature refinement

[0014] The output signal of the input layer is input to the middle layer of the neural network. The middle layer includes several residual blocks and skip connection operations cascaded in sequence. The residual blocks refine the signal features and output them through weighted skip connections.

[0015] Step 4: Denoising result output and distance perception

[0016] The output signal of the middle layer of the neural network is input to the output layer of the neural network; the output layer includes a denoising module and a distance extraction module. The denoising module is implemented using a multi-layer perceptron, and the distance extraction module is implemented using a cascaded multi-layer perceptron and pooling operation. In the denoising module, a decoding operation is performed on the multi-layer perceptron to obtain the denoised laser ranging signal; in the distance extraction module, a decoding operation and a pooling operation are performed on the multi-layer perceptron to output the predicted true distance of the object after removing the noise.

[0017] Furthermore, the specific process of step 1 is:

[0018] Step 1-1, high-pass filtering the laser ranging signal collected by the photoelectric detector;

[0019] A high-pass filter is used to filter the laser ranging signal collected by the photoelectric detector. The cutoff frequency of the filter is f re , f re=k0L0, k0 is the modulation speed of the tuning laser, and L0 is the optical path difference of the fiber part of the measurement light path and the reference light path;

[0020] Step 1-2, demodulating the high-pass filtered signal;

[0021] demodulating the high-pass filtered laser ranging signal, and the demodulation process can be represented as:

[0022] s b (n)=s0(n)·cos(2πnf re / f s )

[0023] wherein s b (n) represents the demodulated signal, s0(n) represents the high-pass filtered signal, f s is the sampling frequency of the signal s0(n), n is a sampling point, 1≤n≤N, and N is the total number of sampling points;

[0024] A low-pass frequency domain filter with a cutoff frequency of f re is set to filter the demodulated signal to generate a signal s′ b (n);

[0025] Step 1-3, normalizing the amplitude of the demodulated signal;

[0026] The z-score standardization method is selected to standardize the signal s′ b (n), which is represented as:

[0027]

[0028] wherein u and σ are the mean and variance of the signal s′ b (n), respectively, and s i (n) represents the normalized signal.

[0029] Further, the specific process of step 2 is as follows:

[0030] The input layer includes a plurality of layers of perception machines and dilated convolution layers connected in sequence, and is used to extract the local frequency information of the signal s i (n); the signal s i (n) is encoded by the plurality of layers of perception machines to expand the feature dimension to generate a vector x(n); the local frequency perception is realized by the convolution layer with a dilated rate of d; and the output signal x c (n) of the input layer is represented as:

[0031]

[0032] wherein w c(k) represents the kth convolution kernel of the dilated convolution layer, 0≤k≤K, K is the convolution kernel length of the dilated convolution layer, and d is the convolution dilation rate.

[0033] Furthermore, the specific process of step 3 is:

[0034] Each residual block refines the signal features through convolution and gated filtering, where the refinement process of the residual block in the i-th layer can be expressed as:

[0035] x r，i (n) = MLP l，i (tanh(x r，i-1 (n)*w i，h )·σ(x r，i-1 (n)*w i，g ))+x r，i-1 (n)

[0036] Among them, x r，i (n) is the output signal of the i-th residual block, x r，i-1 (n) represents the output signal of the i-1th residual block; when i=1, let x r，i-1 (n) = x c (n); w i，h and w i，g They represent the independent convolution kernels in the i-th residual block, σ(·) and tanh(·) are activation functions, and * represents the convolution operation;

[0037] Each residual block jumps to connect as the output signal y(n) of the intermediate layer, expressed as:

[0038]

[0039] Among them, α i is the weight of the i-th residual block, 1≤i≤M, and M is the number of residual blocks.

[0040] Furthermore, the specific process of step 4 is:

[0041] In step 4, in the denoising module, the output signal y(n) of the middle layer is decoded by the multi-layer perceptron to generate the denoised signal s(n); in the distance extraction module, the output signal y(n) of the middle layer is pooled after decoding to achieve distance perception and output the object distance information L. The process is expressed as follows:

[0042] s(n)=MLP d (y(n))

[0043]

[0044] Among them, MLP d and MLP frespectively represent the multi-layer perceptron in the de-noising module and the distance extraction module.

[0045] Further, the neural network comprises an input layer, an intermediate layer and an output layer which are sequentially cascaded; and the training data for training the neural network model is generated by using a simulation method, and a physical transmission process of laser in an aerosol medium is simulated based on a Monte Carlo simulation method to generate a simulated FMCW laser ranging signal.

[0046] Further, the FMCW laser ranging signal is generated based on the Monte Carlo simulation method to generate a training data set for training the neural network; multiple echoes are generated due to scattering of the aerosol medium, and according to the FMCW laser ranging principle, the laser ranging signal s under the multiple echoes is written as:

[0047]

[0048] wherein 1<=q<=Q, Q is the number of return light beams, E r and E m，q are the optical fields of the reference light and the qth measurement light respectively, I q is the intensity of the qth measurement light, L q is the reflection position of the qth measurement light, and s0 is a local oscillator signal characteristic parameter of the laser.

[0049] The training data in the training sample is a laser ranging signal under the condition that there is an aerosol medium, and the sample label is a laser ranging signal and object distance information under the condition that there is no aerosol medium.

[0050] The beneficial effects of the present application are:

[0051] (1) The present application uses a neural network-based method to suppress noise and estimate the position of the FMCW laser ranging signal affected by the aerosol medium in the ranging channel, which cannot accurately identify the object position. Compared with the prior art, the starting point of the method is to suppress noise based on the aerosol characteristics, and accurate perception can be achieved through single observation, which improves the effectiveness and accuracy of the algorithm. In addition, the designed deep network structure adapts to the complex characteristics of the FMCW ranging signal, and is more robust.

[0052] (2) The method described in the present application has been verified by experiments, and can accurately identify the distance of the target object under lower visibility, which shows that the method described in the present application can adapt to a wider application environment and ensure the stability of the laser ranging technology in a more severe environment.

[0053] (3) The method described in the present application is simple to apply, and is an end-to-end perception technology, which not only has a simple implementation and is easy to operate, but also has strong practicality and is easy to popularize.

[0054] (4) The method of the present application uses FMCW laser signal simulation generation to discuss the characteristics of FMCW laser ranging signals and the influence of aerosol medium environment on laser signals, providing a reference for such research. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 The flowchart of the method of the present application is shown in the figure.

[0056] Figure 2 The actual environment laser ranging signal processing result display graph of the method of the present application is shown in the figure. DETAILED DESCRIPTION

[0057] The present application will be further described below in conjunction with the drawings and examples.

[0058] The present embodiment provides a neural network-based FMCW laser ranging signal denoising method, which uses neural network to perceive the influence of aerosol medium on laser ranging signal and realizes noise removal. This neural network-based FMCW laser ranging signal denoising algorithm has simple implementation, lightweight network structure, easy operation, strong practicality and generalizability. And it is verified by experiment that it still has good perception ability in the same low visibility environment.

[0059] The flowchart of the method of the present embodiment is shown in the figure Figure 1 as shown, comprising the following steps:

[0060] Step 1, laser ranging signal regularization

[0061] In order to standardize the signals generated by different FMCW laser ranging systems and realize accurate distance perception, the laser ranging electrical signals collected by the photodetector in the FMCW laser ranging system are subjected to regularization processing.

[0062] First, high-pass filtering is performed to eliminate the interference of the low-frequency part in the optical fiber and the collimating lens part on the FMCW laser ranging system; then, demodulation signal is performed to eliminate the influence of the optical path difference between the reference light path and the measurement light path of different FMCW laser ranging systems on the signal frequency, low-pass filtering is performed to eliminate the sum frequency component of the demodulated signal and retain the difference frequency component; finally, the amplitude and variance of the signal are normalized.

[0063] Further, the specific process of step 1 is as follows:

[0064] The FMCW laser ranging signal obtained by the photodetector is a sinusoid-like signal containing multiple frequency components, and the object distance information to be perceived is contained in the frequency. The laser ranging signals generated by different FMCW laser ranging systems are different, and regularization is used to realize the standardization and unification of the laser ranging signals.

[0065] Step 1-1, high-pass filtering the laser ranging signal collected by the photoelectric detector;

[0066] According to the principle of FMCW laser ranging, the spectrum of the ranging signal represents the scattered reflection characteristics of the fiber part and the free space part from low to high. In order to eliminate the influence of the partial reflection in the fiber on the signal spectrum, the cutoff frequency is set to f re A high-pass frequency domain filter. re It is set according to the physical system characteristics of laser ranging and can be expressed as f re =k0L0. Where k0 is the modulation speed of the tuned laser, L0 is the optical path difference between the measurement optical path and the reference optical path, that is, when the ranging system determines f re is a constant.

[0067] Step 1-2, demodulate the high-pass filtered signal;

[0068] Since the laser ranging distance information is contained in the frequency, the optical path difference in different system optical fibers will lead to different fundamental frequencies, which will affect the perception results. Therefore, the laser ranging signal after high-pass filtering is demodulated. The demodulation process can be expressed as:

[0069] s b (n)=s0(n)·cos(2πnf re / f s )

[0070] Among them, s b (n) represents the demodulated signal, s0(n) represents the high-pass filtered signal, and f s is the sampling frequency of signal s0(n), n is the sampling point, 1≤n≤N, and N is the total number of sampling points. Since the signal demodulation process will produce sum frequency components and difference frequency components, it is necessary to remove the high-frequency sum frequency component. When the optical path difference of the optical fiber part is much larger than the optical path of the ranging signal, set the cutoff frequency to f re A low-pass frequency domain filter is used to suppress the image.

[0071] Step 1-3, normalize the amplitude of the demodulated signal;

[0072] The z-score normalization method is used to process the demodulated signal so that the mean and variance of the demodulated signal are consistent. The normalization process can be expressed as:

[0073]

[0074] Where u and σ are the demodulated signal s′ b The mean and variance of (n), s i(n) represents the signal after regularization processing. So far, the standardization process is completed and the output signal s i (n).

[0075] In the FMCW laser ranging system being built, a coupler couples the reference light and the measurement light used for free-space ranging. This light is then converted into a beat frequency signal by a photodetector. This beat frequency signal is converted into a digital signal by an analog-to-digital converter, and the noise suppression algorithm is processed on the host computer.

[0076] The FMCW laser ranging beat signal is based on the different modulation methods of the system, and the signal frequency will have unstable parts in the time domain. Before regularization, it is necessary to intercept the stable part of the signal according to the time domain characteristics. In this embodiment, the laser modulation frequency f m =57kHz, modulation bandwidth B = 1.508GHz, Where c and n' are system constants of the speed of light and the refractive index in the optical fiber, respectively. The modulated carrier wave is a triangle wave. The reference optical path length in the ranging system is L r =2m, the optical path in the measuring optical system is L mi =40m. Therefore, the high-pass filter cutoff frequency is f re =k0(L mi -L r )=34.05MHz. The center frequency of the signal demodulation part is also set to f re After demodulation, the signal needs to be filtered to suppress the image frequency. In this embodiment, since the visibility of the test is low and the ranging range is much smaller than the optical path difference, the cutoff frequency of the low-pass filter can be set to f re In the signal normalization part, z-score is used to process the signal.

[0077] Step 2: Signal local frequency perception

[0078] The regularized signal is input into the input layer of the neural network, and the local frequency of the signal is perceived by encoding and dilating the signal.

[0079] Furthermore, the specific process of step 2 is:

[0080] The input layer consists of a cascade of multilayer perceptrons and dilated convolutional layers (Multilayer Perceptron, MLP) to extract the signal s i (n) local frequency information. Since the signal s i (n) is a one-dimensional signal. First, the signal s is processed by a multi-layer perceptron. i(n) is encoded to expand the feature dimension to generate a vector x(n); local frequency perception is then achieved through a convolutional layer with an expansion rate of d, where the output signal x of the input layer is c (n) can be expressed as:

[0081]

[0082] Among them, w c (k) represents the kth convolution kernel of the dilated convolution layer, 0≤k≤K, K is the convolution kernel length of the dilated convolution layer, and d is the convolution dilation rate.

[0083] In this embodiment, since the FMCW signal's distance perception information is contained in the frequency component, this component primarily implements feature dimensionality increase, simulating local frequency perception through convolution. In this feature dimensionality increase, an MLP network structure is set to (1, 32). The convolution component uses a dilated convolution with a kernel size of 3 and a dilation rate of 2. This process transforms the signal from a (6000, 1) array to a (6000, 32) array.

[0084] Step 3: Network feature refinement

[0085] The output signal of the input layer is input to the middle layer of the neural network. The middle layer includes several residual blocks and jump connection operations cascaded in sequence. The residual block achieves the purpose of refining the signal features and outputs them through weighted jump connections.

[0086] Furthermore, the specific process of step 3 is:

[0087] Each residual block refines the signal features through convolution and gated filtering, where the refinement process of the residual block in the i-th layer can be expressed as:

[0088] x r，i (n) = MLP l，i (tanh(x r，i-1 (n)*w i，h )·σ(x r，i-1 (n)*w i，g ))+x r，i-1 (n)

[0089] Among them, x r，i (n) is the output signal of the i-th residual block, x r，i-1 (n) represents the output signal of the i-1th residual block; when i=1, let x r，i-1 (n) = x c (n); w i，h and w i，gThey represent the independent convolution kernels in the i-th residual block, σ(·) and tanh(·) are activation functions, and * represents the convolution operation.

[0090] The output signal of each residual block will be output to the next layer to achieve cascade. After each residual block jumps to connect, the output signal y(n) of the middle layer can be expressed as:

[0091]

[0092] Among them, α i is the weight of the i-th residual block, 1≤i≤M, and M is the number of residual blocks.

[0093] In this embodiment, each residual block is implemented by convolution and gated filtering, where the convolution kernel size is 3 and the number of residual blocks is 4. In the jump connection part, MLP l The size is (32, 32, 64). After the deep network signal is converted to an array of (6000, 64).

[0094] Step 4: Denoising result output and distance perception

[0095] The output signal after the skip connection is input into the output layer of the neural network. The output layer includes a denoising module and a distance extraction module. The denoising module is implemented using a multi-layer perceptron, while the distance extraction module is implemented using a cascade of multi-layer perceptrons and pooling operations. The multi-layer perceptron performs decoding operations to obtain the denoised laser ranging signal. Furthermore, the multi-layer perceptron performs decoding operations and pooling operations to output the predicted true distance to the object after noise removal.

[0096] Furthermore, the specific process of step 4 is:

[0097] In step 4, noise removal and object position prediction are achieved. In the noise removal part, the output signal y(n) of the middle layer is decoded by a multi-layer perceptron to generate a denoised signal. In the object distance prediction part, the output signal y(n) of the middle layer is decoded and pooled to achieve distance perception. The process can be expressed as:

[0098] s(n)=MLP d (y(n))

[0099]

[0100] Among them, s(n) is the denoised laser ranging signal output by the denoising module, and L is the object distance information output by the distance extraction module; MLP d and MLP f They represent the multilayer perceptrons in the denoising module and the distance extraction module respectively. The denoised signal s(n) is a one-dimensional vector, and L is a scalar representing the object distance.

[0101] In this embodiment, in the noise suppression part, the MLP size is set to (64, 1). That is, the output signal is a denoised signal of (6000, 1). After the above process, noise suppression is achieved. In the distance perception part, the MLP size is also (64, 1). The output signal is then pooled and a scalar is output. The result represents the location of the real object in this distance measurement process estimated by the network. Figure 2 The figure shows the time-frequency results of the FMCW laser ranging signal and the signal after network suppression under the aerosol channel collected based on real scenes.

[0102] In the method described in this embodiment, the training data for training the neural network model is generated by a simulation method, and the physical transmission process of the laser in the aerosol medium is simulated based on the Monte Carlo simulation method to generate a simulated FMCW laser ranging signal.

[0103] Specifically, the FMCW laser ranging signal is generated based on the Monte Carlo simulation method to generate the training data set used for neural network training. The scattering of the aerosol medium will produce multiple echoes. According to the principle of FMCW laser ranging, the laser ranging signal s under multiple echoes can be written as:

[0104]

[0105] Where, 1≤q≤Q, Q is the number of return beams, E r and E m，q are the light fields of the reference light and the qth measurement light, I q is the intensity of the qth beam of measured light, L q is the reflection position of the qth beam of measurement light; s0 is the characteristic parameter of the laser's local oscillator signal, which is related to factors such as the laser line width.

[0106] The training data in the training samples are laser ranging signals in the presence of an aerosol medium, and the sample labels are laser ranging signals and object distance information in the absence of an aerosol medium.

[0107] During the training process for this example, 3,000 FMCW laser ranging signals were generated under simulated rain and fog conditions. Each signal was simulated using a Monte Carlo simulation with 10,000 photons. In the aerosol interference portion, the object positions in the simulated signals in this example were uniformly distributed between 0.5 and 15 meters. Aerosols are water-soluble media with a uniform size distribution between 800 and 1600 nm. The extinction coefficient was uniformly distributed between 0.1 and 2.4.

[0108] To sum up, the application uses a neural network-based method to suppress noise and estimate the position of FMCW laser ranging signals affected by aerosol medium in the measured ranging channel, which cannot accurately identify the position of the object. Compared with the existing method, the starting point of the present method is to suppress noise based on the characteristics of aerosol, and accurate perception can be achieved through a single observation, which improves the effectiveness and accuracy of the algorithm. In addition, the designed deep network structure adapts to the complex characteristics of the FMCW ranging signal, and is more perceptive and robust. In addition, it is verified by experiment that the distance of the target object can still be accurately identified in the case of lower visibility. This shows that the present method can adapt to a wider application environment and ensure the stability of the laser ranging technology in a more severe environment. The present method is simple to apply and is an end-to-end perception technology. It not only has a simple implementation and is easy to operate, but also has strong practicality and is suitable for popularization and use.

[0109] The structures listed above are only basic structures of the application as examples and should not be used to limit the protection scope of the application. Any modification or polishing without substantial meaning made within the main design idea and spirit of the application, which still solves the same technical problems as the application, should be included in the protection scope of the application.

Claims

1. A FMCW laser ranging signal denoising method based on neural network, characterized in that: The following steps are involved: Step 1: Regularization of laser ranging signals The laser ranging electrical signal collected by the photodetector in the FMCW laser ranging system is regularized. First, high-pass filtering is performed to eliminate interference from the low-frequency components and collimating lens in the optical fiber. Then, signal demodulation is performed to eliminate the effect of the optical path difference between the reference and measurement optical paths of different FMCW laser ranging systems on the signal frequency. Low-pass filtering is performed to eliminate the sum frequency component of the demodulated signal while retaining the difference frequency component. Finally, the signal amplitude and variance are normalized. Step 2: Signal local frequency perception The regularized signal is input to the input layer of the neural network. The input layer includes a cascade of multi-layer perceptrons and dilated convolutional layers. The regularized signal is encoded and dilated convolutionalized to sense the local frequency of the signal. Step 3: Network feature refinement The output signal of the input layer is input to the middle layer of the neural network. The middle layer includes several residual blocks and skip connection operations cascaded in sequence. The residual blocks refine the signal features and output them through weighted skip connections. Step 4: Denoising result output and distance perception The output signal of the middle layer of the neural network is input to the output layer of the neural network; the output layer includes a denoising module and a distance extraction module. The denoising module is implemented using a multi-layer perceptron, and the distance extraction module is implemented using a cascaded multi-layer perceptron and pooling operation. In the denoising module, a multi-layer perceptron is used to perform decoding operations to obtain the denoised laser ranging signal; in the distance extraction module, a multi-layer perceptron is used to perform decoding operations and pooling operations to output the predicted true distance of the object after removing the noise.

2. The FMCW laser ranging signal denoising method based on neural network according to claim 1, characterized in that: The specific process of step 1 is: Step 1-1, high-pass filtering the laser ranging signal collected by the photoelectric detector; A high-pass filter is used to filter the laser ranging signal collected by the photoelectric detector. The cutoff frequency of the filter is f re , f re =k0L0, k0 is the modulation speed of the tunable laser, L0 is the optical path difference between the measurement optical path and the reference optical path optical fiber part; Step 1-2, demodulate the high-pass filtered signal; The laser ranging signal after high-pass filtering is demodulated, and the demodulation process can be expressed as: s b (n)=s0(n)·cos(2πnf re / f s ) Among them, s b (n) represents the demodulated signal, s0(n) represents the high-pass filtered signal, f s is the sampling frequency of the signal s0(n), n is the sampling point, 1≤n≤N, N is the total number of sampling points; Set the cutoff frequency to f re The demodulated signal is filtered by a low-pass frequency domain filter to generate a signal s′ b (n); Step 1-3, normalize the amplitude of the demodulated signal; The z-score normalization method is used to normalize the signal s′b(n), which is expressed as: Where u and σ are the signal s′ b The mean and variance of (n), s i (n) represents the signal after regularization processing.

3. The FMCW laser ranging signal denoising method based on neural network according to claim 2, characterized in that: The specific process of step 2 is: The input layer consists of a cascade of multi-layer perceptrons and dilated convolutional layers to extract the signal s i (n) local frequency information; through the multi-layer perceptron to the signal s i (n) is encoded to expand the feature dimension to generate a vector x(n); local frequency perception is achieved through a convolutional layer with an expansion rate of d; the output signal x of the input layer is c (n) is expressed as: Among them, w c (k) represents the kth convolution kernel of the dilated convolution layer, 0≤k≤K, K is the convolution kernel length of the dilated convolution layer, and d is the convolution dilation rate.

4. The FMCW laser ranging signal denoising method based on neural network according to claim 3, characterized in that: The specific process of step 3 is: Each residual block refines the signal features through convolution and gated filtering, where the refinement process of the residual block in the i-th layer can be expressed as: x r，i (n)=MLP l，i (tanh(x r，i-1 (n)*w i，h )·σ(x r，i-1 (n)*w i,g ))+x r，i-1 (n) Among them, x r，i (n) is the output signal of the i-th residual block, x r,i-1 (n) represents the output signal of the i-1th residual block; when i=1, let x r，i-1 (n) = x c (n); w i，h and w i，g They represent the independent convolution kernels in the i-th residual block, σ(·) and tanh(·) are activation functions, and * represents the convolution operation; Each residual block jumps to connect as the output signal y(n) of the intermediate layer, expressed as: Among them, α i is the weight of the i-th residual block, 1≤i≤M, and M is the number of residual blocks.

5. The FMCW laser ranging signal denoising method based on neural network according to claim 4, characterized in that: The specific process of step 4 is: In step 4, in the denoising module, the output signal y(n) of the middle layer is decoded by the multi-layer perceptron to generate the denoised signal s(n); in the distance extraction module, the output signal y(n) of the middle layer is pooled after decoding to achieve distance perception and output the object distance information L. The process is expressed as follows: s(n)=MLP d (y(n)) Among them, MLP d and MLP f They represent the multi-layer perceptrons in the denoising module and the distance extraction module respectively.

6. The FMCW laser ranging signal denoising method based on neural network according to claim 1, characterized in that: The neural network includes an input layer, an intermediate layer, and an output layer that are cascaded in sequence. The training data for training the neural network model is generated by a simulation method. The physical transmission process of the laser in the aerosol medium is simulated based on the Monte Carlo simulation method to generate a simulated FMCW laser ranging signal.

7. The FMCW laser ranging signal denoising method based on neural network according to claim 6, characterized in that: The FMCW laser ranging signal is generated based on the Monte Carlo simulation method to generate the training data set used for neural network training. The scattering of the aerosol medium produces multiple echoes. According to the FMCW laser ranging principle, the laser ranging signal s under multiple echoes is written as: Where, 1≤q≤Q, Q is the number of return beams, E r and E m，q are the light fields of the reference light and the qth measurement light, I q is the intensity of the qth beam of measured light, L q is the reflection position of the qth beam of measurement light, s0 is the characteristic parameter of the local oscillator signal of the laser; The training data in the training samples are laser ranging signals in the presence of an aerosol medium, and the sample labels are laser ranging signals and object distance information in the absence of an aerosol medium.

Images