Method for correcting frequency modulation linearity of FMCW laser radar laser source

Abstract

The invention discloses a method for correcting the frequency modulation linearity of a laser light source of an FMCW laser radar, comprising: describing the hysteresis effect of a piezoelectric ceramic based on a Prandtl-Ishinskii model to obtain a hysteresis model, inverting the hysteresis model to obtain an inverse hysteresis model; establishing a CNN convolutional neural network model, obtaining training data through a laser beat frequency experiment, training the CNN convolutional neural network model based on the training data, using the output parameters of the inverse hysteresis model as input parameters of the trained CNN convolutional neural network model to obtain a combined model of the inverse hysteresis model and the CNN convolutional neural network model, and using the combined model to control the light source parameters to achieve compensation for the PZT hysteresis effect and frequency path response characteristics, thereby increasing the generalization degree of the combined model.

Description

Technical Field

[0001] The invention belongs to the field of signal regulation, and in particular relates to a method for correcting the frequency modulation linearity of a laser light source of an FMCW laser radar. Background Art

[0002] There is a strong demand for high-precision active imaging at ultra-long distances (over km) in modern remote sensing imaging, geographic information acquisition, marine detection and target identification, high-speed unmanned driving and other application fields. At present, ultra-long-distance imaging mainly uses ultra-high energy pulsed lasers as light sources, uses Geiger mode APDs as receiving elements, and uses mechanical scanning devices or imaging lenses in conjunction with Geiger mode APD arrays for imaging. However, it has poor suppression of ambient light and is difficult to work normally under strong light. It needs to be imaged at night or in low light environments, which greatly limits the application scenarios. In order to overcome this problem, there are patents and related reports that use gate control circuits to only open the detection APD for a certain period of time, which can effectively suppress the influence of background light and improve the signal-to-noise ratio. However, the switch gate technology limits the dynamic range of imaging, making it difficult to achieve effective target detection and identification within the entire detection range in front of the radar (e.g., 10m-1km).

[0003] Frequency Modulation Continuous Wave Lidar (FMCW Lidar) is a ranging technology that greatly improves the radar's ability to resist ambient light and the dynamic range of measurement. FMCW Lidar is essentially a heterodyne interferometry measurement. The principle is to linearly modulate the laser frequency and split the laser into two, one for the local oscillator light and the other for the signal light. The signal light is collimated and reflected by the target into the receiving system, interfering with the local oscillator light to generate a beat frequency signal. The signal carries the frequency tuning amount generated by the flight time and the Doppler frequency shift generated by the relative motion. The beat frequency is measured to calculate the distance and speed information. This method can effectively resist the interference of ambient light, its detection capability is close to the quantum limit, and it is basically immune to crosstalk between different radars.

[0004] At the same time, FMCW LiDAR also has strict requirements on light source and optical path in actual use:

[0005] 1. High requirements on light source, narrow linewidth laser with large coherence length is required;

[0006] 2. The modulation is difficult and requires high modulation frequency, and it is necessary to achieve modulation of GHz bandwidth of the order of magnitude of us at optical frequency;

[0007] 3. The signal-to-noise ratio of long-distance signals is difficult to control, and the modulation frequency linearity needs to be extremely high;

[0008] External cavity lasers can meet the requirements of FMCW laser radar for narrow line width and high modulation bandwidth. However, Littrow external cavity lasers generally achieve laser frequency modulation by controlling the piezoelectric ceramic bias device to change the length of the external cavity. Piezoelectric ceramics have hysteresis characteristics and creep characteristics of multi-value mapping, and there are nonlinear characteristics between their input and output, among which the hysteresis characteristic has the most serious impact. Domestic and foreign scholars have conducted a lot of research on the hysteresis nonlinearity problem of piezoelectric ceramic actuators. Among them, the control method based on the open-loop feedforward model inverse compensation is the most effective and convenient. This method is based on the mathematical modeling of the hysteresis characteristics of the piezoelectric ceramic actuator to obtain the corresponding relationship between the piezoelectric ceramic bias device control voltage and the output displacement.

[0009] For example, Jiaqiang E et al. designed a hysteresis modeling method based on the Prandtl_Ishinski model and fuzzy least squares support vector machine, which effectively eliminated the hysteresis phenomenon. In addition, Antonino Laudani et al. used a new hybrid heuristic algorithm called metric topology evolutionary optimization. Experimental results verified that the algorithm can accurately identify the parameters of the Bouc-Wen hysteresis model. Zhi Li et al. designed an inverse compensation method for the Prandtl_Ishinski model using the inverse multiplication structure. By reconstructing the Prandtl_Ishinski model from the implicit form of the input signal to an explicit expression, a better compensation effect was achieved.

[0010] Although these methods can accurately establish hysteresis models, many models do not consider the memory characteristics and frequency correlation of hysteresis behavior, which leads to weak generalization ability of the model, that is, it can only have a good effect on trained data, but is not applicable to untrained data. Summary of the invention

[0011] The purpose of the present invention is to provide a method for correcting the frequency modulation linearity of a FMCW laser radar laser light source to solve the problems existing in the above-mentioned prior art.

[0012] To achieve the above object, the present invention provides a method for correcting the frequency modulation linearity of a FMCW laser radar laser light source, comprising:

[0013] Establish a frequency modulated continuous wave FMCW laser radar ranging system, and analyze the ranging system to obtain light source parameters that affect the frequency sweep linearity of the ranging system;

[0014] The hysteresis effect of the piezoelectric ceramic PZT is described based on the Prandtl_Ishinskii model to obtain the Prandtl_Ishinskii hysteresis model, and the Prandtl_Ishinskii hysteresis model is inverted to obtain the Prandtl_Ishinskii hysteresis inverse model;

[0015] A CNN convolutional neural network model is established, training data is obtained through a laser beat frequency experiment, the CNN convolutional neural network model is trained based on the training data to obtain a trained CNN convolutional neural network model, the output parameters of the Prandtl_Ishinskii hysteresis inverse model are used as input parameters of the trained CNN convolutional neural network model to obtain a combined model of the Prandtl_Ishinskii hysteresis inverse model and the CNN convolutional neural network model, and the light source parameters are corrected using the combined model.

[0016] Optionally, the light source of the FMCW lidar ranging system is an external cavity laser.

[0017] Optionally, the light source parameters include hysteresis effect and frequency response characteristics of PZT.

[0018] Optionally, the lens barrel of the external cavity laser has an external cavity mirror connected by a PZT module, and the frequency modulation function is achieved by changing the pressure point voltage of the PZT and the length of the external cavity.

[0019] Optionally, the Prandtl-Ishinskii model summarizes the hysteresis loop as a parallel connection of independent relays, and uses a weighted hysteresis unit to simulate the function of the relay during the implementation process. Let H r [v](t) is the time variable, and the output result w(t) of the play operator of the Prandtl_Ishinskii model related to v(t) is expressed as:

[0020] w(t)=H r [v](t)=f r (v(t),w(tk)); t k ≤t<t k+1 ,0≤k≤N,

[0021] Where k is the number of operators, v(t) is the voltage value at time t, and w(t k ) is time t k The output voltage value of the play operator;

[0022] f r (v,w)=max{vr,min(v+r,w)},

[0023] w(t) is the output voltage value of the play operator, and the initial value w(0) is defined as follows:

[0024] w(0)=H r [v](0)=f r (v(0),0),

[0025] The width of the hysteresis loop of the play operator mainly depends on the threshold r, and the size of the threshold r determines the bandwidth of the play operator.

[0026] Finally, the final output of the Prandtl_Ishinski model is the weighted integration of the play operator, as shown below:

[0027]

[0028] Where y(t) is the output of the play operator and p(r) is the integrated density function.

[0029] Optionally, the combined model includes: a Fourier transform layer, a nonlinear layer and a hysteresis model layer.

[0030] Optionally, the combined model analyzes the input signal through a Fourier transform layer to obtain the frequency information of the input voltage, uses the frequency information as the input of the nonlinear layer, generates frequency-related nonlinear terms through nonlinear changes and inputs them into the hysteresis model layer to obtain a frequency-related Prandtl_Ishinski hysteresis model.

[0031] Optionally, the nonlinear change is as shown in the following formula:

[0032]

[0033] Among them, f U is the frequency information of the input voltage, ω l , l is the coefficient of the reference transformation, and l is the number of γ operators selected in the Preisach algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0035] Figure 1 This is a principle diagram of frequency modulated continuous wave laser ranging and speed measurement in an embodiment of the present invention;

[0036] Figure 2 Schematic diagram of the structure of an external cavity laser in an embodiment of the present invention;

[0037] Figure 3 is a block diagram of system composition in an embodiment of the present invention;

[0038] Figure 4 It is a flow chart of the Prandtl-Ishinski model system based on convolutional neural network compensation in an embodiment of the present invention;

[0039] Figure 5 Schematic diagram of basic operators of the Prandtl-Ishinski model in an embodiment of the present invention;

[0040] Figure 6 The output graphs of the Prandtl_Ishinskii model and the inverse Prandtl_Ishinskii model in the embodiments of the present invention are shown. DETAILED DESCRIPTION

[0041] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0042] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0043] like Figure 1-6 As shown, this embodiment provides a method for correcting the frequency modulation linearity of a FMCW laser radar laser light source, including:

[0044] Frequency modulated continuous wave (FMCW) laser radar, which measures the flight time of pulses, is essentially a heterodyne interferometry device. The light source uses a narrow linewidth laser (larger coherence length) and linearly changes the wavelength of the laser through internal or external modulation, thereby achieving linear modulation of the laser frequency.

[0045] The frequency components include time-of-flight heterodyne frequency shift and Doppler frequency shift, and the target distance and axial velocity measurements can be achieved through simple calculations.

[0046] according to Figure 1It can be found that the laser is subjected to triangular wave linear frequency modulation, the red color is the local oscillator light, and the cyan color is the echo. The echo and the local oscillator light interfere to generate a beat frequency signal. To ensure the frequency estimation accuracy of the beat frequency signal, at least one frequency modulation cycle of measurement time is required. In one triangular wave frequency modulation cycle, there are two frequency components of the rising and falling triangular wave. If the object to be measured and the laser transceiver are stationary, the two frequencies are equal, that is, the frequency shift generated during the echo flight time, and the frequency shift is the frequency modulation slope multiplied by the echo flight time. If the object to be measured and the transceiver move axially, a Doppler shift will be generated. Compared with microwave radar, the Doppler shift of laser radar is obvious. If the two move in opposite directions, the beat frequency will decrease (red shift), and if they move in opposite directions, the beat frequency will increase (blue shift).

[0047] Therefore, the frequency of distance and speed can be obtained through the beat frequency information generated by interference. The common mode frequency of the beat frequency of the rising and falling edges is the heterodyne frequency generated by the distance, and the differential mode frequency is the Doppler frequency of the axial movement. The heterodyne frequency is positively correlated with the modulation frequency slope and the flight time, so the flight time, that is, the distance information, can be obtained; and the Doppler frequency shift is proportional to the axial speed and inversely proportional to the wavelength, so the speed information can be obtained.

[0048] For long-distance FMCW laser radar, the frequency modulation linearity directly affects the ranging accuracy. Long-distance high-precision ranging imaging laser radar has higher requirements for frequency modulation linearity. The flight time of long-distance ranging brings a larger ranging scale. Good frequency modulation linearity is guaranteed in the entire large dynamic range. The external cavity laser used in this system is used as a tunable light source. The light source structure is as follows: Figure 2 As shown;

[0049] The external cavity laser is composed of a laser housing, a VECSEL chip, an aperture, a stop block, a collimating lens, a tuning block, a filter, a movable block, a reflector, a piezoelectric ceramic sheet, and a light-emitting lens barrel. The lens barrel has an external cavity mirror connected by a pzt module, and the external cavity length is changed by changing the pzt pressure point voltage. Due to the hysteresis effect of pzt, when a linearly changing voltage is applied, the displacement of pzt is nonlinear, which in turn causes the laser frequency sweep to be nonlinear, reducing the signal-to-noise ratio of the radar system.

[0050] The inverse model driving voltage of the Prandtl_Ishinski model modified based on deep learning is used to drive the PZT of the external cavity laser to achieve the frequency sweeping nonlinear correction of the external cavity laser. The system composition block diagram is shown in the figure. Figure 3 As shown;

[0051] The linear triangle wave is corrected by the PZT driving model, and the inverse signal of the model generated by the direct digital synthesizer (DDS) is input into the PZT module of the external cavity laser. The external cavity laser outputs a frequency-sweeping laser, which is split by the reference optical path and the signal optical path, and mixed after a certain degree of delay in the signal optical path to obtain a beat frequency signal. The beat frequency signal is compared with the theoretical signal to correct the parameter setting of the model.

[0052] Among them, the definition of the Prandtl_Ishinski model is as follows:

[0053] The Prandtl_Ishinski model summarizes the hysteresis loop as a parallel connection of independent relays, and in the implementation process, a weighted hysteresis unit is used to simulate the function of the relay. r [v](t) is the time variable, and the output result w(t) of the play operator of the Prandtl_Ishinskii model related to v(t) is expressed as:

[0054] w(t)=H r [v](t)=f r (v(t),w(tk)); t k ≤t<t k+1 ,0≤k≤N,

[0055] Where k is the number of operators. The value of k determines the accuracy of the fitting. It should be as large as possible under the condition that the computing power allows. A large k value consumes more computing power. v(t) is the voltage value at time t, and w(t k ) is time t k The output voltage value of the play operator;

[0056] f r (v,w)=max{vr,min(v+r,w)},

[0057] w(t) is the output voltage value of the play operator, and is defined as follows for the initial value w(0):

[0058] w(0)=H r [v](0)=f r (v(0),0),

[0059] The width of the hysteresis loop of the play operator mainly depends on the threshold r, and the size of the threshold r determines the bandwidth of the Play operator.

[0060] Finally, the final output of the Prandtl_Ishinski model is the weighted integration of the play operator, as follows:

[0061]

[0062] Where y(t) is the output of the play operator, and p(r) is the integral density function, which needs to be identified through experimental data. Considering the actual usage scenario, the frequency modulation period of the frequency modulated continuous wave lidar is generally in the Khz order of magnitude, but the Preisach model is rate-independent and is a model formed by the accumulation of symmetrical operators; as a result, the correction effect of the Preisach model in high-frequency application scenarios is not good, and the generalization ability of the model is also weak. This system intends to use a convolutional neural network to correct and increase the rate-related factors to improve the accuracy of the model in high-frequency conditions. The system flow is as follows: Figure 4 As shown:

[0063] The Fourier transform layer automatically obtains the frequency of the input voltage by analyzing the input signal, and uses the frequency as the input of the nonlinear layer. Through nonlinear transformation, frequency-related nonlinear terms are generated and input into the hysteresis model layer, thereby obtaining the frequency-related Prandtl_Ishinski hysteresis model. In addition, the deep learning layer obtains the deep features of the input signal by convolution and pooling the input voltage, and multiplies the feature vector by the output vector of the hysteresis model layer to generate the final output displacement. In order to perform convolution operation on the input signal and extract the depth information of the input signal, the input data is first arrayed, the input voltage data is arranged according to the matrix, and the size of the matrix is ​​determined according to the number of input signals; then, the input signal matrix is ​​convolved row by row and column by column through the pre-designed convolution kernel to obtain a matrix with one less row and column than the input matrix, which is used as the first feature matrix output by the first convolution layer; pooling processing is performed on the basis of the first convolution feature matrix, and a pooling layer is used to perform mean subsampling on the first feature matrix, thereby further reducing the complexity of the network and reducing the dimension of the feature matrix. The basic method is to select a small rectangular sliding window, and take the average of the eigenvalues ​​in the sliding window area row by column to obtain the pooled value corresponding to the area; select another convolution kernel, and perform pooling on the first feature matrix after convolution. The simplified convolution feature matrix obtained by the pooling layer is convolved for the second time to obtain the second-layer convolution feature extraction matrix; after obtaining the deep feature matrix of the input signal, the matrix needs to be converted into a vector form through the expansion layer, so as to further average the depth information, thereby obtaining the final depth information and acting on the traditional Preisach model. The expansion layer expands and reorganizes each output feature matrix to obtain the characteristic vector of the input voltage information; the frequency-related processing layer sorts the input voltage by time, and performs a discrete Fourier transform to obtain the frequency information of the input voltage, which is input to the nonlinear layer for nonlinear transformation. The nonlinear transformation formula is based on the hyperbolic function memory model and uses the hyperbolic function to obtain the nonlinear processing obtained by the discrete Fourier transform as shown in the following formula:

[0064]

[0065] where f U is the frequency information of the input voltage, ω l , l is the coefficient of the reference transformation, and l is the number of γ operators selected in the Preisach algorithm.

[0066] The nonlinear term is input into the Preisach model as a weight function, and the weight function is superimposed on the product of the hysteresis unit output result to obtain the output displacement vector. The calculation formula is as follows:

[0067]

[0068] Where γ is the hysteresis factor in the Prandtl_Ishinskii model, and the number of hysteresis factors determines the accuracy of the model;

[0069] Finally, the model output layer is the dot multiplication of the feature vector extracted by the convolution feature and the frequency-weighted output displacement vector, which is the output of the model. The calculation formula is as follows:

[0070]

[0071] The network is trained by using a capacitive displacement sensor to measure the actual displacement D of the piezoelectric scanner driven by an input voltage U, and the error between the actual displacement and the model output displacement is selected as the loss function for training.

[0072]

[0073] Calculate the loss function with respect to the parameter ω respectively l and λ l The gradient of , and the gradient descent method is used to continuously update the parameters:

[0074]

[0075]

[0076] Among them, z represents the number of training times, δ1 and δ2 are the training steps, and the training is stopped until the loss function value converges to the given threshold range. Then the model result is inverted to obtain the inverse solution of the model, such as Figure 5 As shown, by using the inverse model as the driving voltage, the PZT of the external cavity laser is driven through the DAC chip, thereby correcting the nonlinearity of the light source scanning caused by the PZT hysteresis.

[0077] The output of the Prandtl_Ishinskii hysteresis inverse model modified by the CNN neural network is shown in the following figure: Figure 6 As shown, an analog-to-digital converter is used to generate a control voltage, which is input into the PZT module of the laser to output a corrected swept-frequency laser with a high linearity of frequency response characteristics that corrects the hysteresis effect. By using the inverse model as the driving voltage, the PZT of the external cavity laser is driven through the DAC chip to correct the nonlinearity of the light source swept-frequency caused by the PZT hysteresis, thereby significantly improving the signal-to-noise ratio of the FMCW lidar and the detection distance.

[0078] The present invention adopts a method combined with a convolutional neural network to achieve the effect of correcting the frequency sweep linearity of an ECL light source based on a PZT. The correction model is a probability correlation model with good generalization performance, thereby achieving the effect of improving the effective detection distance and detection resolution of the FMCW laser radar.

[0079] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for correcting the frequency modulation linearity of a FMCW laser radar laser light source, characterized in that: The following steps are involved: Establish a frequency modulated continuous wave FMCW laser radar ranging system, and analyze the ranging system to obtain light source parameters that affect the frequency sweep linearity of the ranging system; The hysteresis effect of the piezoelectric ceramic PZT is described based on the PrandtI_Ishinskii model to obtain the PrandtI_Ishinskii hysteresis model, and the PrandtI_Ishinskii hysteresis model is inverted to obtain the PrandtI_Ishinskii hysteresis inverse model; A CNN convolutional neural network model is established, training data is obtained through a laser beat frequency experiment, the CNN convolutional neural network model is trained based on the training data to obtain a trained CNN convolutional neural network model, the output parameters of the PrandtI_Ishinskii hysteresis inverse model are used as input parameters of the trained CNN convolutional neural network model to obtain a combined model of the PrandtI_Ishinskii hysteresis inverse model and the CNN convolutional neural network model, and the light source parameters are corrected using the combined model.

2. The method for correcting the frequency modulation linearity of a FMCW laser radar laser light source according to claim 1, characterized in that: The light source of the FMCW laser radar ranging system is an external cavity laser.

3. The method for correcting the frequency modulation linearity of a FMCW laser radar laser light source according to claim 1, characterized in that: The light source parameters include the hysteresis effect and frequency response characteristics of the PZT.

4. The method for correcting the frequency modulation linearity of a FMCW laser radar laser light source according to claim 2, characterized in that: The lens barrel of the external cavity laser has an external cavity mirror connected by a PZT module. The frequency modulation function is achieved by changing the pressure point voltage of the PZT and the length of the external cavity.

5. The method for correcting the frequency modulation linearity of a FMCW laser radar laser light source according to claim 1, characterized in that: The PrandtI_Ishinskii model summarizes the hysteresis loop as a parallel connection of independent relays. In the implementation process, a weighted hysteresis unit is used to simulate the function of the relay. Let H r [v](t) is the time variable, and the output result w(t) of the play operator of the PrandtI_Ishinskii model related to v(t) is expressed as: w(t)=H r [v](t)=f r (v(t),w(t k ));t k ≤t<t k+1 ,0≤k≤N, Where: k is the number of operators, v(t) is the voltage value at time t, w(t k ) is time t k The output voltage value of the play operator; f r (v,w)=max{v-r,min(v+r,w)}, w(t) is the output voltage value of the play operator, and the initial value w(0) is defined as follows: w(0)=H r [v](0)=f r (v(0),0), The width of the hysteresis loop of the play operator mainly depends on the threshold r, and the size of the threshold r determines the bandwidth of the play operator. Finally, the final output of the Prandtl_Ishinski model is the weighted integration of the play operator, as shown below: Where y(t) is the output of the play operator and p(r) is the integrated density function.

6. The method for correcting the frequency modulation linearity of a FMCW laser radar laser light source according to claim 1, characterized in that: The combined model includes: a Fourier transform layer, a nonlinear layer and a hysteresis model layer.

7. The method for correcting the frequency modulation linearity of a FMCW laser radar laser light source according to claim 1, characterized in that: The combined model analyzes the input signal through the Fourier transform layer to obtain the frequency information of the input voltage, uses the frequency information as the input of the nonlinear layer, generates frequency-related nonlinear terms through nonlinear changes and inputs them into the hysteresis model layer to obtain the frequency-related Prandtl_Ishinski hysteresis model.

8. The method for correcting the frequency modulation linearity of a FMCW laser radar laser light source according to claim 7, characterized in that: The nonlinear change is shown in the following formula: Among them, f U is the frequency information of the input voltage, ωl and λl are the coefficients of the reference transformation, and l is the number of γ operators selected in the Preisach algorithm.

Images