Novel FMCW laser radar signal processing method based on lightweight neural network

Abstract

The invention discloses a novel FMCW laser radar signal processing method based on a lightweight neural network. The novel FMCW laser radar signal processing method comprises the steps of performing data preprocessing on an acquired non-stationary FMCW laser radar intermediate frequency signal by using variational mode decomposition; calculating the characteristic quantity of each group of components as an index for evaluating component characteristics, establishing a composite criterion factor for component selection according to different weights of the index, and screening out the component with the maximum criterion factor; designing a lightweight long-short-term memory neural network to carry out multi-step prediction on the screened component signals; fFT and CFAR algorithms are used to carry out frequency spectrum estimation and target detection on LSTM multi-step prediction results, and target distance information is calculated through an FMCW laser radar detection principle. Aiming at the problem of low detection precision of the FMCW laser radar under limited laser frequency modulation bandwidth and limited data length, the target distance information is extracted with lower cost and higher efficiency, so that the FMCW laser radar system has larger equivalent bandwidth, better distance measurement resolution and higher signal-to-noise ratio.

Description

Technical Field

[0001] The present invention belongs to the field of frequency-modulated continuous-wave lidar detection technology, and specifically relates to a novel FMCW lidar signal processing method based on a lightweight neural network. Background Art

[0002] In recent years, with the development of fields such as manufacturing inspection, autonomous driving, and aerospace, the demand for high-precision detection has been increasing. Among the existing detection methods, lidar stands out for its non-contact measurement ability and high resolution. Currently, lidar has been widely used in fields such as long-distance ranging, atmospheric detection, three-dimensional modeling, and autonomous driving.

[0003] However, TOF (Time-of-Flight technology) lidar is easily affected by ambient light, reducing the signal-to-noise ratio and even potentially causing mismeasurements. In contrast, frequency-modulated continuous-wave (FMCW) lidar is based on the principle of coherent detection and has natural anti-interference to ambient light. In addition, FMCW lidar has the advantages of high resolution and high precision, and can simultaneously calculate target distance and velocity information in a single measurement. Therefore, FMCW lidar has become a research hotspot in the field of high-precision detection.

[0004] For most FMCW lidar systems, ranging resolution is one of the most important performance indicators, which is usually determined by the effective frequency modulation bandwidth of the laser and the observation time of the signal for spectrum estimation (i.e., the effective number of sampling points). In most cases, when the observation time is less than or equal to one or half of the scan period, a larger scan bandwidth means higher resolution. The most direct way to improve ranging resolution is to use a better tunable laser with a larger frequency modulation bandwidth, which usually results in higher hardware costs. According to Fourier transform theory, when the frequency of light is linearly modulated at a fixed scan speed, due to the limited frequency modulation bandwidth, the observation time length of the detection signal within a single scan period is limited, resulting in target spectrum leakage and reduced resolution.

[0005] Currently, in addition to optimizing the lidar hardware system, there already exist some back-end processing algorithms to improve the quality and robustness of detection under the condition of limited observation time. In the past few years, many signal splicing methods have been proposed. By extending the data length to increase the observation time, the time window of FFT and the ranging resolution of the FMCW lidar system have been improved. Shi et al. [1.G.Shi,F.M.Zhang,X.H.Qu,Improvement spatial resolution offrequency modulated continuous wave laser ranging system by splicing equaloptical frequency interval sampled signal,In Ninth International Symposium onPrecision Engineering Measurement and Instrumentation(2015)] spliced multiple groups of FMCW ranging signals collected in different time intervals to increase the signal length, and the measurement resolution was correspondingly improved. DiLazaro et al. [2.D.Thomas,N.George,Large-volume,low-cost,high-precision FMCW tomographyusing stitched DFBs,Opt.Express 26(2018)2891–2904.] adopted an integrated light source containing 12 distributed feedback laser (DFB) elements to obtain multiple ranging signals in one measurement. By splicing multiple ranging signals, the scanning bandwidth was broadened and the ranging resolution was improved. However, all existing methods require manual signal splicing, and phase matching errors still exist, resulting in a reduced signal-to-noise ratio and long time consumption.

[0006] Although the method based on signal splicing has been widely used, there are still many disadvantages that are difficult to overcome, such as high cost, manual operation, and phase mismatch. To sum up, the existing methods for improving the detection resolution of the FMCW lidar system either require additional equipment or a cumbersome manual signal processing process, with high cost and low efficiency. Summary of the Invention

[0007] The purpose of the present invention is to provide a novel FMCW lidar signal processing method based on a lightweight neural network to solve the problems of low detection accuracy, large error, and low signal-to-noise ratio of traditional methods under the condition of limited data length, so as to obtain an effective and stable detection result with higher detection accuracy, stronger robustness, and higher signal-to-noise ratio with a smaller system complexity.

[0008] The technical solution for achieving the object of the present invention is as follows: A novel FMCW lidar signal processing method based on a lightweight neural network, and the specific steps are as follows:

[0009] Step 1: Use variational mode decomposition to perform data preprocessing on the collected non-stationary FMCW lidar intermediate frequency signal to obtain intermediate frequency signal components containing different features;

[0010] Step 2: Calculate the sample entropy, kurtosis, cosine similarity entropy, and Spearman correlation coefficient of each group of components as indicators for evaluating component features, and establish a composite criterion factor for component selection according to the different weights of the indicators, and select the component with the largest criterion factor;

[0011] Step 3: Design a lightweight long short-term memory neural network (LSTM) to perform multi-step prediction on the selected component signal to increase the effective data length of the component signal and achieve higher-precision detection;

[0012] Step 4: Use the FFT and CFAR algorithms to perform spectrum estimation and target detection on the results of the LSTM multi-step prediction, and calculate the target distance information through the FMCW lidar detection principle.

[0013] Preferably, the specific method for using variational mode decomposition to perform data preprocessing on the collected non-stationary FMCW lidar intermediate frequency signal to obtain intermediate frequency signal components containing different features is as follows:

[0014] Step 1.1: Split the FMCW lidar target intermediate frequency signal into the superposition of component signals:

[0015]

[0016] where u k (t) is the k-th intrinsic mode function IMF, A k (t) is the amplitude of the k-th intrinsic mode function, φ k (t) is the phase of the k-th intrinsic mode function, represents the sum of K intrinsic mode functions;

[0017] Perform Hilbert transform on the k-th IMF and construct an analytic signal:

[0018]

[0019] where δ(t) represents the Dirac function, and * represents the convolution operation;

[0020] Step 1.2: The estimated central frequency of the k-th IMF The modulation signal of the analytic signal is obtained by multiplying with the k-th analytic signal of the construction:

[0021]

[0022] Step 1.3: Estimate the bandwidth of each IMF by calculating the gradient L 2 norm of the modulation signal in Step 1.2:

[0023]

[0024] Step 1.4: By introducing the Lagrange multiplier λ and the penalty factor α, transform the problem of solving the variational constraint in Step 1.3 into the problem of solving the Lagrangian maximum problem:

[0025]

[0026] Step 1.5: Solve the Lagrangian maximum problem by sampling the alternating direction method of multipliers to obtain the K decomposed component signals, which are specifically implemented through the following iterative steps 1.5.1 to 1.5.5;

[0027] Step 1.5.1: Initialize and Preset the decomposition layer number K, the quadratic penalty term factor α, the maximum number of iterations N, the convergence tolerance ε, and execute Step 1.5.2;

[0028] Step 1.5.2: n←n + 1, k = 1, and execute Step 1.5.3;

[0029] Step 1.5.3: Solve and

[0030]

[0031] and execute Step 1.5.4;

[0032] Step 1.5.4: If k ≤ K, then k←k + 1 and return to execute Step 1.5.3; otherwise execute Step 1.5.5;

[0033] Step 1.5.5: Solve

[0034]

[0035] and calculate whether the convergence condition is satisfied:

[0036]

[0037] If the above convergence condition is satisfied or n = N, then complete the solution process, This is the required intrinsic mode function IMF, i.e., the component signal; otherwise, execute step 1.5.2.

[0038] Preferably, variational mode decomposition is used to preprocess the non-stationary FMCW lidar intermediate frequency signal collected. Calculate the sample entropy, kurtosis, cosine similarity entropy, and Spearman correlation coefficient of each group of components as indicators to evaluate the component characteristics, and establish a composite criterion factor for component selection according to the different weights of the indicators. The specific method for screening out the component with the largest criterion factor is as follows:

[0039] Step 2.1: Calculate the characteristic parameters T = {T1, T2, T3, T4} of the component signal decomposed in step 1, where represents the reciprocal of the sample entropy, where m represents the embedding dimension, r represents the similarity tolerance, N is the data length of the component signal, and B m (r) represents the probability that two sequences match m points under the similarity tolerance r;

[0040] represents the reciprocal of the kurtosis, where N is the data length of the component signal, σ represents the standard deviation of the component signal, represents the mean of the component signal, x i represents the i-th number of the component signal;

[0041] represents the reciprocal of the cosine similarity entropy, where m represents the embedding dimension, τ represents the delay time, r represents the angular threshold, N is the data length of the component signal, and B m (r) represents the global probability of the occurrence of similar patterns;

[0042] represents the Spearman correlation coefficient, where represents the mean of the component signal, x i represents the i-th data of the component signal, represents the mean of the original signal before VMD decomposition, y i represents the i-th data of the original signal before VMD decomposition;

[0043] Step 2.2: According to step 2.1, calculate the characteristic parameters T i ={T i1 ,T i2 ,T i3 ,T i4} for each intrinsic mode function, where i represents the i-th component, i = 1, 2, 3.....K;

[0044] Step 2.3: Calculate the composite criterion factor w jRepresents the j-th component of the weight w;

[0045] Step 2.4: Sort each component signal in descending order according to the composite criterion factor value, and select the component with the largest composite criterion factor value as the original signal.

[0046] Preferably, variational mode decomposition is used to preprocess the collected non-stationary FMCW lidar intermediate frequency signal, and multiple different LSTM units are interconnected to form a lightweight long short-term memory neural network. The unit structure of the lightweight long short-term memory neural network includes a forget gate, an input gate, and an output gate. The input of the basic LSTM unit structure includes the cell state C at the previous moment t-1 , the hidden state h t-1 and the input data x at the current moment t . The output of the basic LSTM unit structure includes the cell state C at the current moment t and the hidden state h t .

[0047] Preferably, variational mode decomposition is used to preprocess the collected non-stationary FMCW lidar intermediate frequency signal. The specific process of the lightweight long short-term memory neural network for multi-step prediction of the component signals selected in step 2 is as follows:

[0048] Feed the input data x at the current time t and the hidden state h at the previous time t-1 into the LSTM unit as data together;

[0049] Calculate the output i of the input gate t : i t =σ(W ix x t +W ih h t-1 +b i ), where σ is the Sigmid activation function, W ix and W ih are the weights corresponding to the input gate, and b i is the threshold of the input gate;

[0050] Calculate the output f of the forget gate t : f t =σ(W fx x t +W fh h t-1 +b f ), where W fx and W fh are the weights corresponding to the forget gate, and b f is the threshold of the forget gate;

[0051] Calculate the output o of the output gate t : o t = σ(W ox x t + W oh h t-1 + b o ), where W ox and W oh are the weights corresponding to the output gate, and b o is the threshold of the output gate;

[0052] Calculate the cell candidate state where tanh is the tanh activation function, and are the weights corresponding to the cell state, is the threshold of the cell state;

[0053] Calculate the cell state C t : ⊙ is the element-wise multiplication operation, C t-1 represents the cell state at time t - 1, i t is the output result of the input gate, f t is the output result of the forget gate, is the candidate cell state;

[0054] Calculate the hidden state h t : h t = o t ⊙ tanh(C t ), o t is the output result of the output gate in the middle;

[0055] Calculate the actual estimated output at the current time: where h t is the hidden state, is the weight of the hidden state to the output link, is the corresponding threshold.

[0056] Preferably, variational mode decomposition is used to preprocess the collected non-stationary FMCW lidar intermediate frequency signal, and the FFT and CFAR algorithms are used to perform spectral estimation and target detection on the results of LSTM multi-step prediction. The specific method for calculating the target distance information through the FMCW lidar detection principle is as follows:

[0057] Step 4.1: Perform a fast Fourier transform on the signal output by LSTM multi-step prediction to obtain the corresponding spectrum and power spectrum;

[0058] Step 4.2: Perform a constant false alarm rate detection (CFAR) on the power spectrum obtained in Step 4.1 to obtain the frequency value f corresponding to the target;

[0059] Step 4.3: Using the target frequency obtained by CFAR and the ranging principle formula of the FMCW lidar Calculate the distance value R of the target, where T is the frequency modulation period, c is the speed of light, and B is the frequency modulation bandwidth.

[0060] Compared with the prior art, the present invention has the following remarkable advantages: (1) VMD (Variational Mode Decomposition) is used in the intermediate frequency signal preprocessing of the FMCW lidar. VMD can effectively suppress the mode mixing phenomenon of the traditional EMD (Empirical Mode Decomposition), has strong robustness in denoising, and has a good processing effect on the non-stationary and non-linear intermediate frequency signals of the FMCW lidar in practice. (2) The screening and reconstruction of VMD components are the key to denoising and feature extraction. The present invention comprehensively evaluates from four dimensions: information complexity, signal sharpness or abnormality degree, structural complexity of autocorrelation, and correlation with the original signal, makes a more comprehensive judgment on the characteristics of the components, and enhances the robustness of component selection and signal reconstruction for different intermediate frequency signals and application scenarios. Sample entropy can characterize the chaotic properties of time series and has a strong ability to characterize the complexity of noise. The lower the sample entropy value, the more useful information the signal contains and the less noise. In the FMCW lidar system, kurtosis has high sensitivity to non-stationary components. Kurtosis can be used to detect sharp pulses or outliers in low signal-to-noise ratio signals, helping to identify noise and interference. When the signal-to-noise ratio is low, the noise component is larger and the kurtosis value is also larger. Cosine similarity entropy can quantify the structural complexity of autocorrelation. The more noise components the signal contains, the larger its cosine similarity entropy. The Spearman correlation coefficient is a method to describe the correlation between two sequences. Compared with the strict assumptions of other conditions such as normal distribution, linear constraint, and Pearson correlation coefficient, the Spearman correlation coefficient does not require strict parametric assumptions, has a wider application range, and is less affected by outliers. (3) In the data prediction process, a lightweight LSTM neural network structure is designed to solve the long-term dependence problem and the problem of gradient explosion or disappearance of the RNN. The forget gate in the LSTM cell can conditionally decide which information to forget, the input gate decides which information is worth being used to update the internal state, and the output gate decides what information should be output. The core of the LSTM is the cell state, which is used to ensure that the information does not change during the transmission process.

[0061] The present invention will be further described in detail below with reference to the accompanying drawings. Description of the Drawings

[0062] Figure 1 are the time domain diagram and frequency spectrum diagram of the FMCW lidar intermediate frequency signal with a sampling rate of 250 MHz and a length of 2048 points.

[0063] Figure 2 It is the basic unit structure of the lightweight LSTM neural network designed by the present invention.

[0064] Figure 3 It is the comparison result diagram between the novel FMCW lidar signal processing method based on the lightweight neural network of the present invention and the Fourier transform algorithm.

[0065] Figure 4 It is the flowchart of the novel FMCW lidar signal processing method based on the lightweight neural network of the present invention. Detailed implementation manners

[0066] A novel FMCW lidar signal processing method based on a lightweight neural network, the specific steps are as follows:

[0067] Step 1: Use variational mode decomposition (VMD) to perform data preprocessing on the non-stationary FMCW lidar intermediate frequency signal collected by ADC, reduce the complexity of the intermediate frequency signal, and obtain intermediate frequency signal components containing different features; the specific method is:

[0068] Step 1.1: Split the FMCW lidar target intermediate frequency signal into a superposition of component signals (using the intrinsic mode function IMF to refer to the component signals):

[0069]

[0070] where u k (t) is the k-th IMF, A k (t) is the amplitude of the k-th IMF, φ k (t) is the phase of the k-th IMF, represents the sum of all K IMFs;

[0071] Perform Hilbert transform on the k-th IMF and construct an analytic signal:

[0072]

[0073] where δ(t) represents the Dirac function, and * represents the convolution operation;

[0074] Step 1.2: Multiply the estimated center frequency of the k-th IMF with the constructed k-th analytic signal to obtain the modulation signal of the analytic signal:

[0075]

[0076] Step 1.3: Estimate the bandwidth of each IMF by calculating the gradient L 2 norm of the modulation signal in Step 1.2:

[0077]

[0078] Step 1.4: By introducing the Lagrange multiplier λ and the penalty factor α, transform the variational constraint problem in Step 1.3 into a Lagrangian maximum problem to be solved:

[0079]

[0080] Step 1.5: Solve the Lagrangian maximum problem by sampling the alternating direction method of multipliers to obtain the K decomposed component signals, which are specifically implemented through the following iterative steps 1.5.1 to 1.5.5;

[0081] Step 1.5.1: Initialization and Preset the decomposition layer number K, the quadratic penalty term factor α, the maximum number of iterations N, the convergence tolerance ε, and execute Step 1.5.2;

[0082] Step 1.5.2: n←n + 1, k = 1, and execute Step 1.5.3;

[0083] Step 1.5.3: Solve through the following expressions and

[0084]

[0085] Execute Step 1.5.4;

[0086] Step 1.5.4: If k ≤ K, then k←k + 1 and return to execute Step 1.5.3; otherwise execute Step 1.5.5;

[0087] Step 1.5.5: Solve through the following expressions

[0088]

[0089] and calculate whether the convergence condition is satisfied:

[0090]

[0091] If the above convergence condition is satisfied or n = N, then complete the solution process, which is the required IMF; otherwise execute Step 1.5.2.

[0092] Step 2: Calculate the sample entropy, kurtosis, cosine similarity entropy, and Spearman correlation coefficient of each group of components as indicators for evaluating component characteristics, and establish a composite criterion factor for component selection according to the different weights of the indicators, and select the component with the largest criterion factor. The specific method is as follows:

[0093] Step 2.1: Calculate the characteristic parameters \(T = \{T_1, T_2, T_3, T_4\}\) of the component signals obtained by decomposition in Step 1, where represents the reciprocal of sample entropy (where \(m\) represents the embedding dimension, \(r\) represents the similarity tolerance, \(N\) is the data length of the input component signal, and \(B\) m (r) represents the probability that two sequences match \(m\) points under the similarity tolerance \(r\));

[0094] represents the reciprocal of kurtosis (where \(N\) is the data length of the input component signal, \(\sigma\) represents the standard deviation of the component signal, represents the mean of the component signal, \(x\) i represents the \(i\)-th data of the component signal);

[0095] represents the reciprocal of cosine similarity entropy (where \(m\) represents the embedding dimension, \(\tau\) represents the delay time, \(r\) represents the angular threshold, \(N\) is the data length of the input component signal, and \(B\) m (r) represents the global probability of the occurrence of similar patterns);

[0096] represents the Spearman correlation coefficient (where represents the mean of the component signal, \(x\) i represents the \(i\)-th data of the component signal, represents the mean of the original signal before VMD decomposition, \(y\) i represents the \(i\)-th data of the original signal before VMD decomposition);

[0097] Step 2.2: According to Step 2.1, calculate the characteristic parameters \(T\) i of each IMF i1 =\{T i2 ,T i3 ,T i4 \}, where \(i\) represents the \(i\)-th component, \(i = 1, 2, 3.....K\);

[0098] Step 2.3: Calculate the composite criterion factor to quantify the "importance degree" of the characteristics of each component to the original signal. The higher the value of the composite criterion factor, the higher the importance degree, that is, the IMF contains more target feature information. The weights of different characteristic parameters are set as \(w=\{0.2, 0.2, 0.3, 0.3\}\), and \(w\) j represents the \(j\)-th component of \(w\);

[0099] Step 2.4: Sort each component signal in descending order according to the value of the composite criterion factor, and select the component with the largest value of the composite criterion factor as the original signal input into the LSTM neural network.

[0100] Step 3: Use the trained lightweight long short-term memory neural network (LSTM) to perform multi-step prediction on the original signal to increase the effective data length of the component signal and achieve higher-precision detection. The specific method is as follows:

[0101] To propagate effective features during training, multiple different LSTM units are linked together to form a deep LSTM network. The basic structure of the LSTM unit includes a forget gate, an input gate, and an output gate. The inputs to the basic structure of the LSTM unit include the cell state C at the previous moment t-1 , the hidden state h t-1 and the input data x at the current moment t . The outputs of the basic structure of the LSTM unit include the cell state C at the current moment t and the hidden state h t . The specific structure is as shown in Figure 2 . Use the theoretical model of the FMCW lidar signal to construct a dataset for training. When the number of iterations reaches the upper limit, stop training. Use the trained LSTM neural network to perform multi-step prediction on the component signals screened in Step 2 to strengthen the signal features and increase the effective data length of the component signals;

[0102] Furthermore, the specific process of the LSTM neural network performing multi-step prediction on the component signals screened in Step 2 is as follows:

[0103] Initialize the cell state C t and the hidden state h t . Take the component signals screened in Step 2 as the input data x at the current time t and input it into the LSTM neural network unit for prediction;

[0104] Take the input data x at the current time t and the hidden state h at the previous time t-1 as data and send them into the LSTM unit together. x t and h t-1 are processed by three fully connected layers with sigmoid activation functions, and at the same time, calculate the outputs of the input gate, the forget gate, and the output gate, as well as calculate the cell candidate state through the tanh function. The specific calculation is achieved through the following steps:

[0105] Calculate the output i of the input gate t : i t =σ(W ix x t +W ih h t-1 +b i ), where σ is the Sigmid activation function, W ix and W ih are the corresponding weights of the input gate, and bi is the threshold of the input gate;

[0106] Calculate the output f of the forget gate t : f t = σ(W fx x t + W fh h t-1 + b f ), where W fx and W fh are the weights corresponding to the forget gate, and b f is the threshold of the forget gate;

[0107] Calculate the output o of the output gate t : o t = σ(W ox x t + W oh h t-1 + b o ),where W ox and W oh are the weights corresponding to the output gate, and b o is the threshold of the output gate;

[0108] Calculate the cell candidate state where tanh is the tanh activation function, and are the weights corresponding to the cell state, is the threshold of the cell state;

[0109] Calculate the cell state C t : ⊙ is the element-wise multiplication operation, C t-1 represents the cell state at time t-1, i t is the output result of the input gate, f t is the output result of the forget gate, is the candidate cell state;

[0110] Calculate the hidden state h t : h t = o t ⊙ tanh(C t ), o t is the output result of the output gate;

[0111] Calculate the actual estimated output at the current time: where h t is the hidden state, is the weight of the hidden state for the output link, is the corresponding threshold.

[0112] Step 4: Use the FFT and CFAR algorithms to perform spectral estimation and target detection on the results of LSTM multi-step prediction, and calculate the target distance information through the FMCW lidar detection principle. The specific method is as follows:

[0113] Step 4.1: Perform a fast Fourier transform (FFT) on the signal output by LSTM multi-step prediction to obtain the corresponding spectrum and power spectrum;

[0114] Step 4.2: Perform constant false alarm rate detection (CFAR) on the power spectrum obtained in Step 4.1 to obtain the frequency value f corresponding to the target;

[0115] Step 4.3: Use the target frequency obtained by CFAR and the FMCW lidar ranging principle formula to calculate the distance value R of the target, where T is the frequency modulation period, c is the speed of light, and B is the frequency modulation bandwidth.

[0116] The present invention proposes a prediction method based on machine learning to improve the resolution, establishes a lightweight long short-term memory (LSTM) neural network with a special structure, which is driven by the theoretical model of the frequency-modulated continuous wave detection signal for signal prediction. Then, a preprocessing algorithm based on variational mode decomposition (VMD) is designed, and a composite criterion factor that conforms to the characteristics of the FMCW lidar intermediate frequency signal is designed to remove the noise of the original data and enhance its features. Finally, the well-trained lightweight LSTM neural network is used to perform multi-step prediction on the input time series signal to increase the effective data length, thereby improving the resolution. The results show that the proposed method not only has the advantages of low cost and high efficiency, but also enables the FMCW lidar system to have a larger equivalent bandwidth, better ranging resolution, and higher signal-to-noise ratio (SNR), providing inspiration for the application of machine learning in the FMCW lidar system.

[0117] As Figure 1 shown, it is a 2048-point FMCW lidar intermediate frequency signal diagram collected by an ADC with a sampling rate of 250 MHz. Figure 1 The upper one is the time domain diagram, Figure 1 and the lower one is the frequency spectrum diagram. The target target is 10 m away from the lidar system, and the target frequency is about 35 MHz.

[0118] As Figure 2 shown, it is the basic unit structure of the designed lightweight LSTM neural network.

[0119] As Figure 3 shown, it is a comparison diagram of the prediction results of the novel FMCW lidar signal processing method based on the lightweight neural network of the present invention and the original data with the same number of collected points. Figure 3 (a) is the spectrum comparison diagram, Figure 3(b) is the CFAR target detection map of the prediction result.

[0120] Combined with Figure 4 , a novel FMCW lidar signal processing method based on a lightweight neural network. First, preprocess the intermediate frequency signal data of the FMCW lidar. On the one hand, use variational mode decomposition (VMD) to extract and separate the features of the non-linear and non-stationary intermediate frequency signal. The overall framework of VMD signal decomposition is to solve the variational problem to minimize the sum of the estimated bandwidths of each mode. Assume that each mode has a limited bandwidth and different center frequencies. To solve this variational problem, the alternating direction method of multipliers (ADMM) is used to continuously update the mode and its center frequency, so that each mode is gradually demodulated to the corresponding baseband. Finally, extract each mode and its corresponding center frequency. By introducing a quadratic penalty factor and a Lagrange multiplier operator, this algorithm transforms the variational problem into an unconstrained variational problem. The quadratic penalty factor can ensure the reconstruction accuracy of the signal under Gaussian noise. By presetting the hyperparameters of the decomposition layer number K and the quadratic penalty term factor α, the VMD signal decomposition and feature extraction capabilities can be adjusted.

[0121] Secondly, the present invention designs a novel component selection strategy, using sample entropy, kurtosis, cosine similarity entropy, and Spearman correlation coefficient as indicators to evaluate the component features, and comprehensively evaluating from four dimensions: information complexity, signal sharpness or abnormality degree, structural complexity of autocorrelation, and correlation with the original signal, to make a more comprehensive judgment on the features of the components, and enhance the robustness of component selection and signal reconstruction after intermediate frequency signal preprocessing in different complex application scenarios.

[0122] Furthermore, a lightweight LSTM neural network structure is designed to predict the selected component data, solving the problems of long-term dependence and gradient explosion or disappearance of the RNN. Among them, the optimizer is a key parameter for LSTM network training, which determines the speed of gradient descent. It is found in the training that the adaptive moment estimation (ADAM) optimizer algorithm has a better convergence speed and convergence effect, so the ADAM optimizer algorithm is selected for network training. Among them, when the number of LSTM units is set to 200, it has the smallest RMSE. The learning rate is initialized to 0.01 and gradually decreased during the training process. The maximum number of training times is set to 1000 times. By adjusting these parameters and optimizing the network structure, a better feature prediction effect can be achieved.

[0123] Finally, perform FFT and CFAR on the prediction output of LSTM. When the false alarm rate is less than 10 -4 , it can still accurately identify the target target. And as the length of the effective data increases, the resolution of the spectrum estimation doubles, with smaller side lobes and a 4dB increase in the signal-to-noise ratio.

[0124] By using the above four steps, it is possible to solve the problems of low detection accuracy, large error, and low signal-to-noise ratio in the traditional method under the condition of limited data length, so as to obtain an effective and stable detection result with higher detection accuracy, stronger robustness, and higher signal-to-noise ratio with a smaller system complexity.

Claims

1. A novel FMCW lidar signal processing method based on lightweight neural network, characterized in that: The specific steps are: Step 1: Use variational mode decomposition to preprocess the collected non-stationary FMCW lidar intermediate frequency signal to obtain intermediate frequency signal components containing different characteristics; Step 2: Calculate the sample entropy, kurtosis, cosine similarity entropy and Spearman correlation coefficient of each component as indicators for evaluating component characteristics, and establish a composite criterion factor for component selection based on the different weights of the indicators to screen out the component with the largest criterion factor; Step 3: Design a lightweight long short-term memory neural network to perform multi-step prediction on the selected component signals; Step 4: Use FFT and CFAR algorithms to perform spectrum estimation and target detection on the results of multi-step prediction of the long short-term memory neural network, and calculate the target distance information through the FMCW lidar detection principle.

2. According to the novel FMCW laser radar signal processing method based on lightweight neural network according to claim 1, it is characterized in that: The specific method of using variational mode decomposition to preprocess the collected non-stationary FMCW lidar intermediate frequency signal to obtain intermediate frequency signal components with different characteristics is as follows: Step 1.1: Split the FMCW lidar target intermediate frequency signal into the superposition of component signals: In the formula, u k (t) is the kth intrinsic mode function IMF, A k (t) is the amplitude of the kth eigenmode function, φ k (t) is the phase of the kth eigenmode function, represents the sum of K intrinsic mode functions; Perform Hilbert transform on the kth IMF and construct the analytical signal: In the formula, δ(t) represents the Dirac function, and * represents the convolution operation; Step 1.2: Set the estimated center frequency of the kth IMF Multiplying the constructed k-th analytical signal obtains the modulation signal of the analytical signal: Step 1.3: By calculating the gradient L of the modulated signal in step 1.2 2 norm to estimate the bandwidth of each IMF: Step 1.4: By introducing the Lagrange multiplier λ and the penalty factor α, the variational constraint problem in step 1.3 is transformed into a Lagrange maximum problem: Step 1.5: Solve the Lagrangian maximum problem by using the sampling alternating direction multiplier method to obtain the decomposed K component signals, which is specifically achieved by the following iterative steps 1.5.1 to 1.5.5; Step 1.5.1: Initialization and Preset the number of decomposition levels K, the quadratic penalty factor α, the maximum number of iterations N, the convergence tolerance ε, and execute step 1.5.2; Step 1.5.2: n←n+1, k=1, execute step 1.5.3; Step 1.5.3: Solve the following expression and Follow step 1.5.

4. Step 1.5.4: If k≤K, then k←k+1 and return to step 1.5.3; otherwise, execute step 1.5.5; Step 1.5.5: Solve the following expression And calculate whether the convergence condition is met: If the above convergence conditions are met or n = N, the solution process is completed. This is the required intrinsic mode function IMF, that is, the component signal; otherwise, execute step 1.5.

2.

3. According to the novel FMCW laser radar signal processing method based on lightweight neural network according to claim 1, it is characterized in that: The sample entropy, kurtosis, cosine similarity entropy and Spearman correlation coefficient of each component are calculated as indicators for evaluating component characteristics, and a composite criterion factor for component selection is established according to the different weights of the indicators. The specific method for screening out the component with the largest criterion factor is as follows: Step 2.1: Calculate the characteristic parameters T = {T1, T2, T3, T4} of the component signals decomposed in step 1, where: represents the inverse of sample entropy, where m represents the embedding dimension, r represents the similarity tolerance, N is the data length of the component signal, and B m (r) represents the probability that two sequences match m points under the similarity tolerance r; represents the inverse of the kurtosis, where N is the data length of the component signal, σ represents the standard deviation of the component signal, represents the mean value of the component signal, x i represents the i-th number of component signal; represents the inverse of the cosine similarity entropy, where m represents the embedding dimension, τ represents the delay time, r represents the angle threshold, N is the data length of the component signal, and B m (r) represents the global probability of similar patterns appearing; represents the Spearman correlation coefficient, where represents the mean value of the component signal, x i represents the i-th data of the component signal, represents the original signal mean before VMD decomposition, y i Represents the i-th data of the original signal before VMD decomposition; Step 2.2: According to step 2.1, calculate the characteristic parameter T of each eigenmode function i ={T i1 ,T i2 ,T i3 ,T i4 }, where i represents the i-th component, i = 1, 2, 3.....K; Step 2.3: Calculate the composite criterion factor w j represents the jth component of weight w; Step 2.4: Sort each component signal from large to small according to the composite criterion factor value, and select the component with the largest composite criterion factor value as the original signal.

4. The novel FMCW laser radar signal processing method based on lightweight neural network according to claim 1 is characterized in that: Multiple different LSTM units are linked to each other to form a lightweight long short-term memory neural network. The unit structure of the lightweight long short-term memory neural network includes a forget gate, an input gate, and an output gate. The input of the LSTM basic unit structure includes the cell state C at the previous moment. t-1 , hidden state h t-1 And the current input data x t The output of the LSTM basic unit structure includes the current cell state C t and hidden state h t .

5. According to claim 4, the novel FMCW laser radar signal processing method based on lightweight neural network is characterized in that: The specific process of the lightweight long short-term memory neural network for multi-step prediction of the component signals screened out in step 2 is as follows: The input data x at the current time t and the hidden state h at the previous time t-1 As data, they are fed into the LSTM unit; Calculate the output i of the input gate t :i t =σ(W ix x t +W ih h t-1 +b i ), where σ is the Sigmid activation function, W ix and W ih is the corresponding weight of the input gate, b i is the threshold of the input gate; Calculate the output f of the forget gate t :f t =σ(W fx x t +W fh h t-1 +b f ), where W fx and W fh is the weight corresponding to the forget gate, b f is the threshold of the forget gate; Calculate the output o of the output gate t :o t =σ(W ox x t +W oh h t-1 +b o ), where W ox and W oh is the corresponding weight of the output gate, b o is the threshold of the output gate; Calculate cell candidate states Where tanh is the tanh activation function, and is the weight corresponding to the cell state, is the threshold of the cell state; Calculate cell state C t : ⊙ is the element multiplication operation, C t-1 represents the cell state at time t-1, i t is the output result of the input gate, f t is the output of the forget gate, is the candidate cell state; Calculate the hidden state h t :h t =o t ⊙tanh(C t ), o t is the output result of the middle output gate; Calculate the actual estimated output for the current time: where h t is the hidden state, is the weight of the hidden state on the output link, is the corresponding threshold.

6. The novel FMCW laser radar signal processing method based on lightweight neural network according to claim 1 is characterized in that: The FFT and CFAR algorithms are used to perform spectrum estimation and target detection on the results of LSTM multi-step prediction. The specific method of calculating the target distance information based on the FMCW lidar detection principle is as follows: Step 4.1: Perform fast Fourier transform on the signal output by LSTM multi-step prediction to obtain the corresponding spectrum and power spectrum; Step 4.2: Perform constant false alarm rate detection on the power spectrum obtained in step 4.1 to obtain the frequency value f corresponding to the target; Step 4.3: Use the target frequency obtained by CFAR and the FMCW LiDAR ranging principle formula Calculate the distance value R of the target, where T is the frequency modulation period, c is the speed of light, and B is the frequency modulation bandwidth.

Images