FMCW laser light source nonlinear pre-correction method and system based on generative adversarial network

Abstract

The invention discloses an FMCW laser light source non-linear pre-correction method and system based on a generative adversarial network, belongs to the technical field of laser radars and the like, and solves the problem that the prior art depends on complex physical modeling and parameter adjustment and calibration and is often low in efficiency when facing a non-linear problem. Uncorrected beat frequency time domain data are acquired, nonlinear correction debugging of an FMCW light source is performed by using iteration and photoelectric phase-locked loop technologies based on an FMCW laser radar measurement system and a data processing module, and expert-level modulation current modulation data are acquired to construct a data set to train a conditional adversarial generative neural network with constraints; and obtaining beat frequency data generated by the expert modulation network model guidance system, carrying out nonlinear evaluation to obtain a final expert modulation network model, and applying the final expert modulation network model to the FMCW laser radar measurement system and the data processing module to complete nonlinear pre-correction of the FMCW laser light source. The FMCW laser radar nonlinear pre-correction method is used for FMCW laser radar nonlinear pre-correction based on the generative adversarial network.

Description

Technical Field

[0001] A method and system for nonlinear precorrection of an FMCW laser light source based on a generative adversarial network are provided, which are used for nonlinear precorrection of an FMCW laser radar based on a generative adversarial network and belong to the technical fields of laser radar, laser ranging and artificial intelligence. Background Art

[0002] FMCW laser measurement technology uses linear frequency modulation (LMCW) to measure distance and speed by analyzing the frequency difference and frequency shift characteristics between the echo and transmitted signals. FMCW distance measurement systems calculate target distance based on the frequency difference, while FMCW speed measurement systems leverage the Doppler effect to infer target speed by analyzing the difference between the reflected and transmitted light frequencies. Combined with beat frequency signal processing, FMCW laser measurement technology can achieve millimeter-per-second or higher speed accuracy for high-speed moving targets. This technology is particularly effective in areas such as transportation, autonomous driving, and industrial production line speed monitoring. For example, in autonomous vehicles, FMCW lidar not only enables high-precision three-dimensional scanning of the environment but also detects the target's speed in real time, providing precise data support for vehicle path planning and obstacle avoidance. FMCW laser measurement technology offers several significant advantages over traditional measurement methods. First, its high-frequency modulation provides superior resolution, enabling precise distinction between the motion states of multiple targets even in complex scenes. Second, FMCW laser systems are more robust to environmental noise and can operate stably in the presence of optical interference and harsh environments. In addition, since the system does not require mechanical scanning components, it has a fast measurement speed and a short response time, and can adapt to real-time speed measurement needs in dynamic scenes.

[0003] Despite the numerous advantages of FMCW laser measurement technology, achieving high precision still faces challenges. Regarding frequency modulation linearization, the nonlinear characteristics and thermal effects of existing tunable lasers can lead to a decrease in frequency modulation linearity, which in turn affects the accuracy of distance and speed measurements. While the use of phase-locked loops (PLLs), optical / resampling, optical frequency combs, iterative algorithms, and sideband modulation can address these technical issues, the following still remain:

[0004] 1. Traditional methods rely on complex physical modeling and parameter tuning, and are often inefficient when faced with nonlinear problems.

[0005] 2. For data-driven nonlinear correction methods, such as reinforcement learning methods based on Mapping-Rel ati on, due to the incompleteness of the data set, the system description is incomplete and the random adaptability is not strong, which limits the linearization performance. Summary of the Invention

[0006] In response to the above research problems, the purpose of the present invention is to provide a FMCW lidar nonlinear pre-correction method and system based on a generative adversarial network to solve the problem that the existing technology relies on complex physical modeling and parameter adjustment, which is often inefficient when facing nonlinear problems.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A nonlinear pre-correction method for FMCW lidar based on a generative adversarial network includes the following steps:

[0009] Step 1: Build an FMCW lidar measurement system and a data processing module for processing input data and output data of the FMCW lidar measurement system;

[0010] Step 2: Obtain uncorrected beat frequency time-domain data. Simultaneously, based on the FMCW lidar measurement system and data processing module, use iteration and optoelectronic phase-locked loop technology to perform nonlinear correction and debugging of the FMCW light source. This completes expert-level debugging of the FMCW light source linearization and collects expert-level modulation current data with application value.

[0011] Step 3: Based on the beat frequency time domain data and expert-level modulation current modulation data obtained in step 2, a data set is constructed to train a constrained conditional adversarial generative neural network to obtain an expert modulation network model;

[0012] Step 4: Apply the trained expert modulation network model to the FMCW lidar measurement system and data processing module to complete the nonlinear pre-correction of the FMCW light source, and obtain the beat frequency data generated by the system under the guidance of the expert modulation network model;

[0013] Step 5: Perform nonlinear evaluation on the beat frequency data generated by the FMCW lidar measurement system and data processing module under the guidance of the expert modulation network model. If the requirements are met, the final expert modulation network model is obtained. Otherwise, step 4 is repeated.

[0014] Step 6: Apply the final expert modulation network model to the FMCW lidar measurement system and data processing module to complete the nonlinear pre-correction of the FMCW laser light source.

[0015] Furthermore, the FMCW lidar measurement system in step 1 includes an FPGA, a laser driver module driven by the FPGA, a laser driven by the laser driver module to generate an FMCW laser signal, an isolator, an attenuator, an interferometer MZI for generating a beat frequency signal, and a photoelectric detector PD for detecting the optical signal, which are sequentially connected to the laser. The PD transmits the detected optical signal back to the FPGA for processing into beat frequency time domain data. The laser driver module includes a current controller and a temperature controller.

[0016] The data processing module includes a data processing module A that converts the modulation current slope into modulation current data and writes the modulation current data into the FPGA, and a data processing module B that processes the beat frequency time domain data output by the FPGA into beat frequency data through Hilbert transform and phase difference;

[0017] The FPGA includes an FPGA mainboard, a laser current driver, a light source control module, a signal transmitting and receiving module, a digital-to-analog converter DAC, and an analog-to-digital converter ADC arranged on the FPGA mainboard, the light source control module is connected to the digital-to-analog converter DAC, and the signal transmitting and receiving module is connected to the analog-to-digital converter ADC;

[0018] The laser is a 1550 nm frequency-swept laser, which is either a VSCEL or a DFB laser.

[0019] Furthermore, the specific steps of step 2 are:

[0020] Step 2.1. Obtain uncorrected beat frequency time-domain data and input the modulated current slope into the constructed FMCW lidar measurement system and data processing module. System debugging experts use iterative or optoelectronic phase-locked loop technology to complete the nonlinear correction of the FMCW laser light source, obtaining expert-level current modulation data corresponding to the beat frequency signal nonlinearity below 5kHz.

[0021] Step 2.2: Observe the results obtained in step 2.1 using an oscilloscope, collect expert-level current modulation data corresponding to the beat frequency signal with a nonlinearity below 5 kHz, and obtain uncorrected beat frequency time domain data, wherein the number of collected data is greater than 50.

[0022] Furthermore, the specific steps of step 3 are:

[0023] Step 3.1. Construct a constrained conditional adversarial generative network, including a generator network and a discriminator network. The generator network consists of an amplification layer, a self-attention mechanism layer, a mask layer, a SoftMax activation function layer, and a convolution layer that discards the second dimension data to obtain the final generated data. The convolution layer includes four groups of convolution operations connected in sequence. Each group of convolution operations includes a 1×1 Conv layer and a ReLu activation function layer connected in sequence. The amplification layer amplifies the input data into a three-dimensional tensor, and then the data enters the self-attention mechanism layer. The self-attention mechanism layer linearly maps the input at each time step, calculates the attention weight along the time step, multiplies the calculated attention weight by the original input element by element, and outputs the intermediate data after the attention mechanism. The unnecessary positions are then shielded by the pre-constructed mask layer and processed by the SoftMax activation function layer before entering the convolution layer. The discriminator network uses a multi-layer perceptron MLP to complete the discrimination between generated data and real data.

[0024] Step 3.2: Construct the constraint loss function L of the conditional adversarial generation network with constraints c (D, G), the formula is:

[0025] L c (D, G) = L(D, G) + λL1 + αL2

[0026] L1=d(p data , p G(z) )

[0027]

[0028] Among them, L(D, G) represents the adversarial loss of the standard GAN, and L1 loss is the true data distribution p data and generate data distribution p G(z) Similarity distance, λ is the coefficient of L1, L2 is the coefficient of generating data x i ~P G(z) , xi represents the i-th generated data, which comes from the generated data distribution, max and min represent the upper and lower limits of the generated data, α is the coefficient of L2, d(p data , p G(z) ) represents the similarity distance between the real data distribution and the generated data distribution, θ G represents the generation network parameters of the conditional adversarial generation network, θ D represents the discriminator parameters of the conditional adversarial generation network, Represents the data from the dataset p x The training sample x in the expected value, Represents the distribution of latent variable samples p from the training process zThe sample z in the expectation, D represents the discriminator network, G table generator network, is calculated by Jensen-Shannon (JS) divergence distance, d(p data , p G(z) ), the formula is:

[0029]

[0030] Where KL() represents the Kullback-Leibler divergence;

[0031] Step 3.3: Read the current modulation data value from the expert-level modulated current modulation data obtained in step 2 according to the output dimension of the constrained conditional adversarial generative network, convert the current modulation data value from hexadecimal data to decimal data, and calculate the current modulation slope. Finally, obtain the current modulation slope of the R0I region, where the R0I region represents the region of interest, which refers to the percentage of the intercepted signal cycle containing the nonlinear portion in the total cycle.

[0032] Step 3.4: Read the sampling rate and beat signal time domain data points from the uncorrected beat signal time domain data obtained in step 2 according to the input dimension of the constrained conditional adversarial generative network, perform HiIbert transform and phase difference on the beat signal time domain data to obtain beat frequency data, perform average superposition by period to obtain beat frequency data points for a single period, finally obtain beat frequency data points with a data magnitude of 100,000 in the R0I region, and reduce the data magnitude of the beat frequency data points to the hundredth digit;

[0033] Step 3.5: Based on the results obtained in steps 3.3 and 3.4, construct a set Set = {frequency, slope_current} as a data set, where frequency represents the beat frequency data, slope_current represents the current modulation slope, and frequency is used as the conditional input of the constrained conditional adversarial network, that is, as the input of the generator network, and slope_current is used as the real data in the training of the constrained conditional adversarial network;

[0034] Step 3.6: Use the Adam optimization algorithm and the dataset to train the constrained conditional adversarial generative network, repeat the iteration and adjust the hyperparameters until convergence to obtain the expert modulation network model.

[0035] Furthermore, the specific steps of step 4 are:

[0036] Step 4.1, inputting a new modulation current slope into the FMCW lidar measurement system and data processing module to obtain beat frequency data and inputting the data into the expert modulation network model to obtain output data, i.e., current modulation slope data generated by the expert modulation network model;

[0037] Step 4.2: Convert the generated current modulation slope data into current modulation data points, and then convert them into hexadecimal current modulation data as the current driving data of the FMCW lidar system;

[0038] Step 4.3: Collect the current drive data to drive the FMCW lidar measurement system and the beat frequency data generated again by the data processing module.

[0039] Furthermore, the specific steps of step 5 are:

[0040] Step 5.1: Perform nonlinear evaluation on the beat frequency data generated by the system under the guidance of the expert modulation network model. The nonlinearity is defined as:

[0041] H=K*|f b (t)-f b |

[0042] Among them, K is a normalized constant, f b (t) is the beat frequency data generated by the system under the guidance of the expert modulation network model, f b is the target beat frequency data, H represents the nonlinear evaluation value;

[0043] Step 5.2: When H is less than or equal to the given threshold, the final expert modulation network model is obtained. When H is greater than the given threshold, the parameters of the expert modulation network model need to be adjusted, and step 4 is re-executed for training again.

[0044] A nonlinear pre-correction system for FMCW lidar based on a generative adversarial network, comprising:

[0045] An FMCW laser radar measurement system and a data processing module for processing input data and output data of the FMCW laser radar measurement system;

[0046] Expert-level debugging module: This module obtains uncorrected beat frequency time-domain data and, based on the FMCW lidar measurement system and data processing module, uses iteration and optoelectronic phase-locked loop technology to perform nonlinear correction and debugging of the FMCW light source. This module completes expert-level debugging of the FMCW light source linearization and collects expert-level modulation current data with application value.

[0047] Model building module: Based on the beat frequency time domain data and expert-level modulation current modulation data obtained by the expert-level debugging module, a data set is constructed to train the constrained conditional adversarial generative neural network to obtain the expert modulation network model;

[0048] Guidance module: This module applies the trained expert modulation network model to the FMCW lidar measurement system and data processing module to complete the nonlinear pre-correction of the FMCW light source and obtain the beat frequency data generated by the system under the guidance of the expert modulation network model.

[0049] Evaluation module: Performs nonlinear evaluation on the beat frequency data generated by the FMCW lidar measurement system and data processing module under the guidance of the expert modulation network model. If the requirements are met, the final expert modulation network model is obtained. Otherwise, the guidance module is re-executed.

[0050] Nonlinear pre-correction module: Apply the final expert modulation network model to the FMCW lidar measurement system and data processing module to complete the nonlinear pre-correction of the FMCW laser light source.

[0051] Furthermore, the FMCW laser radar measurement system includes an FPGA, a laser driver module driven by the FPGA, a laser driven by the laser driver module to generate an FMCW laser signal, an isolator, an attenuator, an interferometer MZI for generating a beat frequency signal, and a photoelectric detector PD for detecting the optical signal, which are sequentially connected to the laser. The PD transmits the detected optical signal back to the FPGA for processing into beat frequency time domain data. The laser driver module includes a current controller and a temperature controller.

[0052] The data processing module includes a data processing module A that converts the modulation current slope into modulation current data and writes the modulation current data into the FPGA, and a data processing module B that processes the beat frequency time domain data output by the FPGA into beat frequency data through Hilbert transform and phase difference;

[0053] The FPGA includes an FPGA mainboard, a laser current driver, a light source control module, a signal transmitting and receiving module, a digital-to-analog converter DAC, and an analog-to-digital converter ADC arranged on the FPGA mainboard, the light source control module is connected to the digital-to-analog converter DAC, and the signal transmitting and receiving module is connected to the analog-to-digital converter ADC;

[0054] The laser is a 1550 nm frequency-swept laser, which is either a VSCEL or a DFB laser.

[0055] Furthermore, the specific implementation steps of the expert-level debugging module are:

[0056] Step 2.1. Obtain uncorrected beat frequency time-domain data and input the modulated current slope into the constructed FMCW lidar measurement system and data processing module. System debugging experts use iterative or optoelectronic phase-locked loop technology to complete the nonlinear correction of the FMCW laser light source, obtaining expert-level current modulation data corresponding to the beat frequency signal nonlinearity below 5kHz.

[0057] Step 2.2: Observe the results obtained in step 2.1 using an oscilloscope, collect expert-level current modulation data corresponding to the beat frequency signal with a nonlinearity below 5 kHz, and obtain uncorrected beat frequency time domain data, wherein the number of collected data is greater than 50.

[0058] Furthermore, the specific implementation steps of the model building module are:

[0059] Step 3.1. Construct a constrained conditional adversarial generative network, including a generator network and a discriminator network. The generator network consists of an amplification layer, a self-attention mechanism layer, a mask layer, a SoftMax activation function layer, and a convolution layer that discards the second dimension data to obtain the final generated data. The convolution layer includes four groups of convolution operations connected in sequence. Each group of convolution operations includes a 1×1 Conv layer and a ReLu activation function layer connected in sequence. The amplification layer amplifies the input data into a three-dimensional tensor, and then the data enters the self-attention mechanism layer. The self-attention mechanism layer linearly maps the input at each time step, calculates the attention weight along the time step, multiplies the calculated attention weight by the original input element by element, and outputs the intermediate data after the attention mechanism. The unnecessary positions are then shielded by the pre-constructed mask layer and processed by the SoftMax activation function layer before entering the convolution layer. The discriminator network uses a multi-layer perceptron MLP to complete the discrimination between generated data and real data.

[0060] Step 3.2: Construct the constraint loss function L of the conditional adversarial generation network with constraints c (D, G), the formula is:

[0061] L c (D, G) = L(D, G) + λL1 + αL2

[0062] L1=d(p data , p G(z) )

[0063]

[0064] Among them, L(D,G) represents the adversarial loss of the standard GAN, and L1 loss is the real data distribution p data and generate data distribution p G(z) Similarity distance, λ is the coefficient of L1, L2 is the coefficient of generating data xi ~p G(z) Soft constraint, x i represents the i-th generated data, which comes from the generated data distribution, max and min represent the upper and lower limits of the generated data, α is the coefficient of L2, d(p data , p G(z) ) represents the similarity distance between the real data distribution and the generated data distribution, θ G represents the generation network parameters of the conditional adversarial generation network, θ D represents the discriminator parameters of the conditional adversarial generation network, Represents the data from the dataset p x The training sample x in the expected value, Represents the distribution of latent variable samples p from the training process z The sample z in the expectation, D represents the discriminator network, G table generator network, is calculated by Jensen-Shannon (JS) divergence distance, d(p data , p G(z) ), the formula is:

[0065]

[0066] Where KL() represents the Kullback-Leibler divergence;

[0067] Step 3.3: Read the current modulation data value from the expert-level modulated current modulation data obtained in step 2 according to the output dimension of the constrained conditional adversarial generative network, convert the current modulation data value from hexadecimal data to decimal data, and calculate the current modulation slope. Finally, obtain the current modulation slope of the R0I region, where the R0I region represents the region of interest, which refers to the percentage of the intercepted signal cycle containing the nonlinear portion in the total cycle.

[0068] Step 3.4: Read the sampling rate and beat signal time domain data points from the uncorrected beat signal time domain data obtained in step 2 according to the input dimension of the constrained conditional adversarial generative network, perform Hilbert transform and phase difference on the beat signal time domain data to obtain beat frequency data, perform average superposition by period to obtain beat frequency data points for a single period, and finally obtain beat frequency data points with a data magnitude of 100,000 in the R0I region, and reduce the data magnitude of the beat frequency data points to the hundredth digit;

[0069] Step 3.5: Based on the results obtained in steps 3.3 and 3.4, construct a set Set = {frequency, slope_current} as a data set, where frequency represents the beat frequency data and slope_current represents the current modulation slope. Frequency is used as the conditional input of the constrained conditional adversarial network, that is, as the input of the generator network, and slope_current is used as the real data in the training of the constrained conditional adversarial network.

[0070] Step 3.6: Use the Adam optimization algorithm and the dataset to train the constrained conditional adversarial generative network, iterate repeatedly and adjust the hyperparameters until convergence to obtain the expert modulation network model;

[0071] The specific implementation steps of the guidance module are:

[0072] Step 4.1: Input the new modulation current slope to the FMCW lidar measurement system and data processing module to obtain beat frequency data and input it into the expert modulation network model to obtain output data, that is, the current modulation slope data generated by the expert modulation network model:

[0073] Step 4.2: Convert the generated current modulation slope data into current modulation data points, and then convert them into hexadecimal current modulation data as the current driving data of the FMCW lidar system;

[0074] Step 4.3: Collect the current drive data to drive the FMCW lidar measurement system and the beat frequency data generated by the data processing module:

[0075] The specific implementation steps of the evaluation module are:

[0076] Step 5.1: Perform nonlinear evaluation on the beat frequency data generated by the system under the guidance of the expert modulation network model. The nonlinearity is defined as:

[0077] H=K*|f b (t)-f b |

[0078] Among them, K is a normalized constant, f b (t) is the beat frequency data generated by the system under the guidance of the expert modulation network model, f b is the target beat frequency data, H represents the nonlinear evaluation value;

[0079] Step 5.2: When H is less than or equal to the given threshold, the final expert modulation network model is obtained. When H is greater than the given threshold, the parameters of the expert modulation network model need to be adjusted, and step 4 is re-executed for training again.

[0080] Compared with the prior art, the present invention has the following beneficial effects:

[0081] The present invention does not rely on complex physical modeling and parameter adjustment, and is highly efficient when dealing with nonlinear problems. In addition, the present invention provides a complete system description, strong random adaptability, and improved linearization performance, which is specifically reflected in the following aspects:

[0082] First, the generator module of the constrained conditional adversarial generative network in the present invention consists of a self-attention mechanism layer, a mask layer, and a convolution layer. The self-attention mechanism layer calculates weights, and the mask layer shields irrelevant or interfering information in the input data, successfully fitting the temporal characteristics of expert-modulated current data and uncorrected beat frequency time domain data. The multi-layer CNN network (i.e., the convolution layer) can capture local and global features of the input data through convolution and pooling operations at different levels, realizing multi-scale feature extraction, which is conducive to capturing data characteristics at different levels, thereby improving the generalization ability and classification accuracy of the model.

[0083] 2. The constraint loss function introduced in this invention, L1 penalty term forces the generation of distribution P G(z) In the Statistical Moment Space, data Manifold projection, and can provide supplementary gradient direction when the gradient signal of the discriminator D fails; G(z) With P data When the difference is significant, the penalty term L1 dominates the optimization direction, which can promote the generated distribution to quickly cover the multimodal structure of the real distribution. As the distributions gradually align, the impact of the penalty term will weaken, and the adversarial gradient provided by the discriminator D will dominate the fine-tuning. The L2 penalty term constrains the range of the data generated by the generator G, where max and min represent the upper and lower limits of the current that the laser can withstand. For out-of-bounds data, the square term imposes a quadratic penalty on the degree of out-of-bounds, ensuring that the data generated by the generator is within a reasonable range of the laser modulation current.

[0084] 3. The present invention proposes an expert model based on a constrained generative adversarial network (GAN), which realizes efficient correction of nonlinear errors in the FMCW measurement process. Compared with traditional nonlinear compensation methods that rely on high-precision hardware or static modeling, the present invention adopts an intelligent control strategy, which significantly reduces the dependence on high-precision devices, thereby effectively reducing the overall cost of the system. In addition, the method has self-learning and generalization capabilities, better adapting to complex and changeable operating environments and working conditions, and reflecting stronger environmental adaptability and system robustness. By precisely controlling the frequency tuning process, the inventive solution can effectively suppress nonlinear distortion in the frequency modulation process, significantly improving the measurement accuracy of the FMCW radar system, and outperforming traditional solutions based on linear fitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 It is a framework structure diagram of the present invention;

[0086] Figure 2 Flowchart of the present invention;

[0087] Figure 3 The structure diagram of the FMCW lidar measurement system of the present invention;

[0088] Figure 4 The beat frequency time domain data effect diagram obtained by the FMCW lidar measurement system of the present invention;

[0089] Figure 5 : This is a structural diagram of the constrained conditional adversarial generative network in the present invention, where Generator represents the generator network, Discriminator represents the discriminator network, Beat Frequency represents the beat frequency data, RealExpert Sample represents the real expert data, GeneratedFakeSample represents the generated fake data, Distance represents the distance, Statistics represents the data distribution, Real / Fake represents true or false, FineTune Training represents the training, InputExpend represents the amplification layer, Self-Attention ion represents the self-attention mechanism layer, Mark represents the mask layer, SoftMax represents the SoftMax activation function, Conv represents the convolution kernel, and Relu represents the Relu activation function;

[0090] Figure 6 This is a diagram showing the effect of the beat frequency signal obtained by the FMCW lidar measurement system on an oscilloscope when driven by the modulation current data output by the expert modulation network model obtained in the present invention. DETAILED DESCRIPTION

[0091] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0092] Although FMCW laser measurement technology has many advantages, its high-precision implementation still faces challenges, especially in frequency modulation linearization. The nonlinear characteristics and thermal effects of tunable lasers will cause the frequency modulation linearity to decrease, which in turn affects the accuracy of ranging and speed measurement. In order to deal with the impact of nonlinear factors and other factors on measurement accuracy in FMCW laser measurement, researchers at home and abroad are actively exploring various methods and technologies. Traditional nonlinear correction methods include frequency modulation nonlinear correction methods based on phase-locked principle, sampling principle, frequency comb, iterative algorithm, sideband modulation technology and other mechanisms. However, traditional methods rely on complex physical modeling and parameter adjustment, generally rely on manual adjustment and fine operation by experts, and have defects such as low automation and intelligence. They are often inefficient when facing nonlinear problems. In this regard, the present invention proposes a nonlinear pre-correction method for FMCW laser radar based on a generative adversarial network.

[0093] A nonlinear pre-correction method for FMCW lidar based on a generative adversarial network includes the following steps:

[0094] Step 1: Build an FMCW lidar measurement system and a data processing module for processing input data and output data of the FMCW lidar measurement system;

[0095] The FMCW lidar measurement system includes an FPGA, a laser driver module driven by the FPGA, a laser driven by the laser driver module to generate an FMCW laser signal, an isolator, an attenuator, an interferometer MZI that generates a beat frequency signal, and a photoelectric detector PD that detects the optical signal, which are sequentially connected to the laser. The PD transmits the detected optical signal back to the FPGA for processing into beat frequency time domain data. The laser driver module includes a current controller and a temperature controller.

[0096] The data processing module includes a data processing module A that converts the modulation current slope into modulation current data and writes the modulation current data into the FPGA, and a data processing module B that processes the beat frequency time domain data output by the FPGA into beat frequency data through Hilbert transform and phase difference;

[0097] The FPGA includes an FPGA mainboard, a laser current driver, a light source control module, a signal transmitting and receiving module, a digital-to-analog converter DAC, and an analog-to-digital converter ADC arranged on the FPGA mainboard, the light source control module is connected to the digital-to-analog converter DAC, and the signal transmitting and receiving module is connected to the analog-to-digital converter ADC;

[0098] The laser is a 1550nm band swept laser, which is a VSCEL or DFB laser, and has a swept bandwidth exceeding 140 GHz.

[0099] Step 2: Obtain uncorrected beat frequency time-domain data. Simultaneously, based on the FMCW lidar measurement system and data processing module, use iteration and optoelectronic phase-locked loop technology to perform nonlinear correction and debugging of the FMCW light source. This completes expert-level debugging of the FMCW light source linearization and collects expert-level modulation current data with application value.

[0100] The specific steps are:

[0101] Step 2.1. Obtain uncorrected beat frequency time-domain data and input the modulated current slope into the constructed FMCW lidar measurement system and data processing module. System debugging experts use iterative or optoelectronic phase-locked loop technology to complete the nonlinear correction of the FMCW laser light source, obtaining expert-level current modulation data corresponding to the beat frequency signal nonlinearity below 5kHz.

[0102] Step 2.2: Observe the results obtained in step 2.1 using an oscilloscope, collect expert-level current modulation data corresponding to the beat frequency signal with a nonlinearity below 5 kHz, and obtain uncorrected beat frequency time domain data, wherein the number of collected data is greater than 50.

[0103] Step 3: Based on the beat frequency time domain data and expert-level modulation current modulation data obtained in step 2, a data set is constructed to train a constrained conditional adversarial generative neural network to obtain an expert modulation network model;

[0104] The specific steps are:

[0105] Step 3.1. Construct a constrained conditional adversarial generative network, including a generator network and a discriminator network. The generator network consists of an amplification layer, a self-attention mechanism layer, a mask layer, a SoftMax activation function layer, and a convolution layer that discards the second dimension data to obtain the final generated data. The convolution layer includes four groups of convolution operations connected in sequence. Each group of convolution operations includes a 1×1 Conv layer and a ReLu activation function layer connected in sequence. The amplification layer amplifies the input data into a three-dimensional tensor, and then the data enters the self-attention mechanism layer. The self-attention mechanism layer linearly maps the input (3400 dimensions) for each time step, calculates the attention weight along the time step, multiplies the calculated attention weight by the original input element by element, and outputs the intermediate data after the attention mechanism. The mask layer constructed in advance masks out the unnecessary positions and then enters the convolution layer after processing through the SoftMax activation function layer. The discriminator network uses a multi-layer perceptron MLP to complete the discrimination between generated data and real data.

[0106] Step 3.2: Construct the constraint loss function L of the conditional adversarial generation network with constraints c (D, G), the formula is:

[0107] Lc (D, G) = L(D, G) + 2L1 + αL2

[0108] L1=d(p data , p G(z) )

[0109]

[0110] Among them, L(D,G) represents the adversarial loss of the standard GAN, and L1 loss is the real data distribution p data and generate data distribution p G(z) Similarity distance, λ is the coefficient of L1, L2 is the coefficient of generating data x i ~p G(z) Soft constraint, x i represents the i-th generated data, which comes from the generated data distribution, max and min represent the upper and lower limits of the generated data, α is the coefficient of L2, d(p data , p G(z) ) represents the similarity distance between the real data distribution and the generated data distribution, θ G represents the generation network parameters of the conditional adversarial generation network, θ D represents the discriminator parameters of the conditional adversarial generation network, Indicates the expectation of the training sample x from the dataset px, Represents the distribution of latent variable samples p from the training process z Find the expected value of sample z in D

[0111] Represents the discriminator network and the G-table generator network, which are calculated by Jensen-Shannon (JS) divergence distance, d(p data , p G(z) ), the formula is:

[0112]

[0113] Where KL() represents the Kullback-Leibler divergence;

[0114] Step 3.3: Read the current modulation data values ​​(4096 data points in total) from the expert-level modulated current modulation data obtained in Step 2 based on the output dimension of the constrained conditional adversarial generative network. Convert the current modulation data values ​​from hexadecimal to decimal and calculate the current modulation slope. Finally, obtain the current modulation slope of the R0I region (3400 data points in total). The R0I region represents the region of interest, which refers to the percentage of the intercepted signal cycle containing the nonlinear portion in the total cycle.

[0115] Step 3.4: Read the sampling rate and beat signal time domain data points (a total of 10020 data points) from the beat signal time domain data obtained in step 2 according to the input dimension of the constrained conditional adversarial generative network. Perform Hilbert transform and phase difference on the beat signal time domain data to obtain the beat frequency data. Perform average superposition by period to obtain the beat frequency data points of a single period (a total of 4096 data points), as shown in the following example: Figure 4 As shown, finally, the beat frequency data points with a data level of one hundred thousand in the ROI area (a total of 3400 data points) are obtained, and the data level of the beat frequency data points is reduced to the hundredth place;

[0116] Step 3.5: Based on the results obtained in steps 3.3 and 3.4, construct a set Set = {frequency, slope_current} as a data set, where frequency represents the beat frequency data and slope_current represents the current modulation slope. Use frequency as the conditional input of the constrained conditional adversarial network, that is, as the input of the generator network, and use slope_current as the real data in the training of the constrained conditional adversarial network.

[0117] Step 3.6: Use the Adam optimization algorithm and the dataset to train the constrained conditional adversarial generative network, repeat the iteration and adjust the hyperparameters until convergence to obtain the expert modulation network model.

[0118] Step 4: Apply the trained expert modulation network model to the FMCW lidar measurement system and data processing module to complete the nonlinear pre-correction of the FMCW light source, and obtain the beat frequency data generated by the system under the guidance of the expert modulation network model:

[0119] The specific steps are:

[0120] Step 4.1, inputting a new modulation current slope into the FMCW lidar measurement system and data processing module to obtain beat frequency data and inputting the data into the expert modulation network model to obtain output data, i.e., current modulation slope data generated by the expert modulation network model;

[0121] Step 4.2: Convert the generated current modulation slope data into current modulation data points, and then convert them into hexadecimal current modulation data as the current driving data of the FMCW lidar system;

[0122] Step 4.3: Collect current drive data to drive the FMCW lidar measurement system and data processing module to generate beat frequency data again.

[0123] Step 5: Perform nonlinear evaluation on the beat frequency data generated by the FMCW lidar measurement system and data processing module under the guidance of the expert modulation network model. If the requirements are met, the final expert modulation network model is obtained. Otherwise, step 4 is repeated.

[0124] The specific steps are:

[0125] Step 5.1: Perform nonlinear evaluation on the beat frequency data generated by the system under the guidance of the expert modulation network model. The nonlinearity is defined as:

[0126] H=K*|f b (t)-f b |

[0127] Among them, K is a normalized constant, f b (t) is the beat frequency data generated by the system (FMCW lidar measurement system and data processing module) under the guidance of the expert modulation network model, f b is the target beat frequency data, H represents the nonlinear evaluation value;

[0128] Step 5.2: When H is less than 5k (i.e., 5000), the final expert modulation network model is obtained. When H is greater than 5k, the parameters of the expert modulation network model need to be adjusted, and step 4 is re-executed and training is performed again.

[0129] Step 6: Apply the final expert modulation network model to the FMCW lidar measurement system and data processing module to complete the nonlinear pre-correction of the FMCW laser light source.

Claims

1. A nonlinear pre-correction method for FMCW lidar based on a generative adversarial network, characterized in that: The steps include: Step 1: Build an FMCW lidar measurement system and a data processing module for processing input data and output data of the FMCW lidar measurement system; Step 2: Obtain uncorrected beat frequency time-domain data. Simultaneously, based on the FMCW lidar measurement system and data processing module, use iteration and optoelectronic phase-locked loop technology to perform nonlinear correction and debugging of the FMCW light source. This completes expert-level debugging of the FMCW light source linearization and collects expert-level modulation current data with application value. Step 3: Based on the beat frequency time domain data and expert-level modulation current modulation data obtained in step 2, a data set is constructed to train a constrained conditional adversarial generative neural network to obtain an expert modulation network model; Step 4: Apply the trained expert modulation network model to the FMCW lidar measurement system and data processing module to complete the nonlinear pre-correction of the FMCW light source, and obtain the beat frequency data generated by the system under the guidance of the expert modulation network model; Step 5: Perform nonlinear evaluation on the beat frequency data generated by the FMCW lidar measurement system and data processing module under the guidance of the expert modulation network model. If the requirements are met, the final expert modulation network model is obtained. Otherwise, step 4 is repeated. Step 6: Apply the final expert modulation network model to the FMCW lidar measurement system and data processing module to complete the nonlinear pre-correction of the FMCW laser light source.

2. The nonlinear pre-correction method for FMCW lidar based on a generative adversarial network according to claim 1, characterized in that: The FMCW laser radar measurement system in step 1 includes an FPGA, a laser driver module driven by the FPGA, a laser driven by the laser driver module to generate an FMCW laser signal, an isolator, an attenuator, an interferometer MZI for generating a beat frequency signal, and a photoelectric detector PD for detecting the optical signal, which are sequentially connected to the laser. The PD transmits the detected optical signal back to the FPGA for processing into beat frequency time domain data. The laser driver module includes a current controller and a temperature controller. The data processing module includes a data processing module A that converts the modulation current slope into modulation current data and writes the modulation current data into the FPGA, and a data processing module B that processes the beat frequency time domain data output by the FPGA into beat frequency data through Hilbert transform and phase difference; The FPGA includes an FPGA mainboard, a laser current driver, a light source control module, a signal transmitting and receiving module, a digital-to-analog converter DAC, and an analog-to-digital converter ADC arranged on the FPGA mainboard, the light source control module is connected to the digital-to-analog converter DAC, and the signal transmitting and receiving module is connected to the analog-to-digital converter ADC; The laser is a 1550 nm frequency-swept laser, which is either a VSCEL or a DFB laser.

3. The nonlinear pre-correction method for FMCW lidar based on a generative adversarial network according to claim 2, characterized in that: The specific steps of step 2 are: Step 2.

1. Obtain uncorrected beat frequency time-domain data and input the modulated current slope into the constructed FMCW lidar measurement system and data processing module. System debugging experts use iterative or optoelectronic phase-locked loop technology to complete the nonlinear correction of the FMCW laser light source, obtaining expert-level current modulation data corresponding to the beat frequency signal nonlinearity below 5kHz. Step 2.2: Observe the results obtained in step 2.1 using an oscilloscope, collect expert-level current modulation data corresponding to the beat frequency signal with a nonlinearity below 5 kHz, and obtain uncorrected beat frequency time domain data, wherein the number of collected data is greater than 50.

4. The nonlinear pre-correction method for FMCW lidar based on a generative adversarial network according to claim 3, characterized in that: The specific steps of step 3 are: Step 3.

1. Construct a constrained conditional adversarial generative network, including a generator network and a discriminator network. The generator network consists of an amplification layer, a self-attention mechanism layer, a mask layer, a SoftMax activation function layer, and a convolution layer that discards the second dimension data to obtain the final generated data. The convolution layer includes four groups of convolution operations connected in sequence. Each group of convolution operations includes a 1×1 Conv layer and a ReLu activation function layer connected in sequence. The amplification layer amplifies the input data into a three-dimensional tensor, and then the data enters the self-attention mechanism layer. The self-attention mechanism layer linearly maps the input at each time step, calculates the attention weight along the time step, multiplies the calculated attention weight by the original input element by element, and outputs the intermediate data after the attention mechanism. The unnecessary positions are then shielded by the pre-constructed mask layer and processed by the SoftMax activation function layer before entering the convolution layer. The discriminator network uses a multi-layer perceptron MLP to complete the discrimination between generated data and real data. Step 3.2: Construct the constraint loss function L of the conditional adversarial generation network with constraints c (D, G), the formula is: L c (D,G)=L(D,G)+λL1+αL2 L1=d(p data ,p G(z) ) Among them, L(D,G) represents the adversarial loss of the standard GAN, and L1 loss is the real data distribution p data and generate data distribution p G(z) Similarity distance, λ is the coefficient of L1, L2 is the coefficient of generating data x i ~p G(z) Soft constraint, x i represents the i-th generated data, which comes from the generated data distribution, max and min represent the upper and lower limits of the generated data, α is the coefficient of L2, d(p data , p G(z) ) represents the similarity distance between the real data distribution and the generated data distribution, θ G represents the generation network parameters of the conditional adversarial generation network, θ D represents the discriminator parameters of the conditional adversarial generation network, Represents the data from the dataset p x The training sample x in the expected value, Represents the distribution of latent variable samples p from the training process z The sample z in the expectation, D represents the discriminator network, G represents the generator network, and is calculated by Jensen-Shannon (JS) divergence distance, d(p data , p G(z) ), the formula is: Where KL() represents the Kullback-Leibler divergence; Step 3.3: Read the current modulation data value from the expert-level modulated current modulation data obtained in step 2 according to the output dimension of the constrained conditional adversarial generative network, convert the current modulation data value from hexadecimal data to decimal data, and calculate the current modulation slope. Finally, obtain the current modulation slope of the R0I region, where the R0I region represents the region of interest, which refers to the percentage of the intercepted signal cycle containing the nonlinear portion in the total cycle. Step 3.4: Read the sampling rate and beat signal time domain data points from the uncorrected beat signal time domain data obtained in step 2 according to the input dimension of the constrained conditional adversarial generative network, perform Hilbert transform and phase difference on the beat signal time domain data to obtain beat frequency data, perform average superposition by period to obtain beat frequency data points for a single period, and finally obtain beat frequency data points with a data magnitude of 100,000 in the R0I region, and reduce the data magnitude of the beat frequency data points to the hundredth digit; Step 3.5: Based on the results obtained in steps 3.3 and 3.4, construct a set Set = {frequency, slope_current} as a data set, where frequency represents the beat frequency data and slope_current represents the current modulation slope. Use frequency as the conditional input of the constrained conditional adversarial network, that is, as the input of the generator network, and use slope_current as the real data in the training of the constrained conditional adversarial network. Step 3.6: Use the Adam optimization algorithm and the dataset to train the constrained conditional adversarial generative network, repeat the iteration and adjust the hyperparameters until convergence to obtain the expert modulation network model.

5. The nonlinear pre-correction method for FMCW lidar based on a generative adversarial network according to claim 4, characterized in that: The specific steps of step 4 are: Step 4.1, input the new modulation current slope to the FMCW lidar measurement system and data processing module to obtain beat frequency data and input it into the expert modulation network model to obtain output data, that is, obtain the current modulation slope data generated by the expert modulation network model; Step 4.2: Convert the generated current modulation slope data into current modulation data points, and then convert them into hexadecimal current modulation data as the current driving data of the FMCW lidar system; Step 4.3: Collect the current drive data to drive the FMCW lidar measurement system and the beat frequency data generated again by the data processing module.

6. The nonlinear pre-correction method for FMCW lidar based on a generative adversarial network according to claim 5, characterized in that: The specific steps of step 5 are: Step 5.1: Perform nonlinear evaluation on the beat frequency data generated by the system under the guidance of the expert modulation network model. The nonlinearity is defined as: H=K*|f b (t)-f b | Among them, K is a normalized constant, f b (t) is the beat frequency data generated by the system under the guidance of the expert modulation network model, f b is the target beat frequency data, H represents the nonlinear evaluation value; Step 5.2: When H is less than or equal to the given threshold, the final expert modulation network model is obtained. When H is greater than the given threshold, the parameters of the expert modulation network model need to be adjusted, and step 4 is re-executed for training again.

7. A nonlinear pre-correction system for FMCW lidar based on a generative adversarial network, characterized in that: include: An FMCW laser radar measurement system and a data processing module for processing input data and output data of the FMCW laser radar measurement system; Expert-level debugging module: This module obtains uncorrected beat frequency time-domain data and, based on the FMCW lidar measurement system and data processing module, uses iteration and optoelectronic phase-locked loop technology to perform nonlinear correction and debugging of the FMCW light source. This module completes expert-level debugging of the FMCW light source linearization and collects expert-level modulation current data with application value. Model building module: Based on the beat frequency time domain data and expert-level modulation current modulation data obtained by the expert-level debugging module, a data set is constructed to train the constrained conditional adversarial generative neural network to obtain the expert modulation network model; Guidance module: This module applies the trained expert modulation network model to the FMCW lidar measurement system and data processing module to perform nonlinear pre-correction of the FMCW light source and obtain the beat frequency data generated by the system under the guidance of the expert modulation network model. Evaluation module: Performs nonlinear evaluation on the beat frequency data generated by the FMCW lidar measurement system and data processing module under the guidance of the expert modulation network model. If the requirements are met, the final expert modulation network model is obtained. Otherwise, the guidance module is re-executed. Nonlinear pre-correction module: Apply the final expert modulation network model to the FMCW lidar measurement system and data processing module to complete the nonlinear pre-correction of the FMCW laser light source.

8. The FMCW lidar nonlinear pre-correction system based on a generative adversarial network according to claim 7, characterized in that: The FMCW laser radar measurement system includes an FPGA, a laser driver module driven by the FPGA, a laser driven by the laser driver module to generate an FMCW laser signal, an isolator, an attenuator, an interferometer MZI for generating a beat frequency signal, and a photoelectric detector PD for detecting the optical signal, which are sequentially connected to the laser. The PD transmits the detected optical signal back to the FPGA for processing into beat frequency time domain data. The laser driver module includes a current controller and a temperature controller. The data processing module includes a data processing module A that converts the modulation current slope into modulation current data and writes the modulation current data into the FPGA, and a data processing module B that processes the beat frequency time domain data output by the FPGA into beat frequency data through Hilbert transform and phase difference; The FPGA includes an FPGA mainboard, a laser current driver, a light source control module, a signal transmitting and receiving module, a digital-to-analog converter DAC, and an analog-to-digital converter ADC arranged on the FPGA mainboard, the light source control module is connected to the digital-to-analog converter DAC, and the signal transmitting and receiving module is connected to the analog-to-digital converter ADC; The laser is a 1550 nm frequency-swept laser, which is either a VSCEL or a DFB laser.

9. The FMCW lidar nonlinear pre-correction system based on a generative adversarial network according to claim 8, characterized in that: The specific implementation steps of the expert-level debugging module are: Step 2.

1. Obtain uncorrected beat frequency time-domain data and input the modulated current slope into the constructed FMCW lidar measurement system and data processing module. System debugging experts use iterative or optoelectronic phase-locked loop technology to complete the nonlinear correction of the FMCW laser light source, obtaining expert-level current modulation data corresponding to the beat frequency signal nonlinearity below 5kHz. Step 2.2: Observe the results obtained in step 2.1 using an oscilloscope, collect expert-level current modulation data corresponding to the beat frequency signal with a nonlinearity below 5 kHz, and obtain uncorrected beat frequency time domain data, wherein the number of collected data is greater than 50.

10. The FMCW lidar nonlinear pre-correction system based on a generative adversarial network according to claim 9, characterized in that: The specific implementation steps of the model construction module are: Step 3.1, construct a constrained conditional adversarial generation network, including a generator network and a discriminator network. The generator network consists of an amplification layer, a self-attention mechanism layer, a mask layer, a SoftMax activation function layer, and a convolution layer that discards the second dimension data to obtain the final generated data. The convolution layer includes four groups of convolution operations connected in sequence, and each group of convolution operations includes 1×1Con v The amplification layer amplifies the input data into a three-dimensional tensor, and then the data enters the self-attention mechanism layer. The self-attention mechanism layer performs linear mapping on the input of each time step, calculates the attention weight along the time step, multiplies the calculated attention weight by the original input element by element, and outputs the intermediate data after the attention mechanism. The mask layer constructed in advance masks out the unnecessary positions, and then passes through the SoftMax activation function layer to enter the convolution layer. The discriminator network uses the multi-layer perceptron MLP to complete the discrimination between generated data and real data. Step 3.2: Construct the constraint loss function L of the conditional adversarial generation network with constraints c (D, G), the formula is: L c (D,G)=L(D,G)+2L1+αL2 L1=d(p data ,p G(z) ) Among them, L(D,G) represents the adversarial loss of the standard GAN, and L1 loss is the real data distribution p data and generate data distribution p G(z) Similarity distance, λ is the coefficient of L1, L2 is the coefficient of generating data x i ~p G(z) Soft constraint, x i represents the i-th generated data, which comes from the generated data distribution, max and min represent the upper and lower limits of the generated data, α is the coefficient of L2, d(p data , p G(z) ) represents the similarity distance between the real data distribution and the generated data distribution, θ G represents the generation network parameters of the conditional adversarial generation network, θ D represents the discriminator parameters of the conditional adversarial generation network, Represents the data from the dataset p x The training sample x in the expected value, Represents the distribution of latent variable samples p from the training process z The sample z in the expectation, D represents the discriminator network, G table generator network, is calculated by Jensen-Shannon (JS) divergence distance, d(p data , p G(z) ), the formula is: Where KL() represents the Kullback-Leibler divergence; Step 3.3: Read the current modulation data value from the expert-level modulated current modulation data obtained in step 2 according to the output dimension of the constrained conditional adversarial generative network, convert the current modulation data value from hexadecimal data to decimal data, and calculate the current modulation slope. Finally, obtain the current modulation slope of the ROI region, where the ROI region represents a region of interest, which refers to the percentage of the intercepted signal cycle containing the nonlinear portion in the total cycle. Step 3.4: Read the sampling rate and beat signal time domain data points from the uncorrected beat signal time domain data obtained in step 2 according to the input dimension of the constrained conditional adversarial generative network, perform HiIbert transform and phase difference on the beat signal time domain data to obtain beat frequency data, perform average superposition by period to obtain beat frequency data points for a single period, finally obtain beat frequency data points with a data magnitude of 100,000 in the ROI area, and reduce the data magnitude of the beat frequency data points to the hundredth digit; Step 3.5: Based on the results obtained in steps 3.3 and 3.4, construct a set Set = {frequency, slope_current} as a data set, where frequency represents the beat frequency data and slope_current represents the current modulation slope. Use frequency as the conditional input of the constrained conditional adversarial network, that is, as the input of the generator network, and use slope_current as the real data in the training of the constrained conditional adversarial network. Step 3.6: Use the Adam optimization algorithm and the dataset to train the constrained conditional adversarial generative network, iterate repeatedly and adjust the hyperparameters until convergence to obtain the expert modulation network model; The specific implementation steps of the guidance module are: Step 4.1, inputting a new modulation current slope into the FMCW lidar measurement system and data processing module to obtain beat frequency data and inputting the data into the expert modulation network model to obtain output data, i.e., current modulation slope data generated by the expert modulation network model; Step 4.2: Convert the generated current modulation slope data into current modulation data points, and then convert them into hexadecimal current modulation data as the current driving data of the FMCW lidar system; Step 4.3: Collect the current drive data to drive the FMCW lidar measurement system and the beat frequency data generated again by the data processing module; The specific implementation steps of the evaluation module are: Step 5.1: Perform nonlinear evaluation on the beat frequency data generated by the system under the guidance of the expert modulation network model. The nonlinearity is defined as: H=K*|f b (t)-f b | Among them, K is a normalized constant, f b (t) is the beat frequency data generated by the system under the guidance of the expert modulation network model, f b is the target beat frequency data, H represents the nonlinear evaluation value; Step 5.2: When H is less than or equal to the given threshold, the final expert modulation network model is obtained. When H is greater than the given threshold, the parameters of the expert modulation network model need to be adjusted, and step 4 is re-executed for training again.

Images