An online self-calibration and drift compensation method for a vehicle-mounted small radar

Abstract

The present application relates to the technical field of automatic control, and further relates to a kind of online self-calibration and drift compensation method of vehicle-mounted small radar, method includes the following steps: step 1: in the front-end package module of vehicle-mounted small radar, output loopback beat signal and obtain the digital sample sequence of loopback beat signal by analog-digital conversion;Step 2, the digital sample sequence of loopback beat signal is executed Hilbert transform to obtain analytic signal sequence, and adaptive segmented interval division is executed to instantaneous frequency deviation sequence and the drift feature vector of each segmented interval is extracted;Step 3, the signal link between the input from the transmitting digital baseband signal to the instantaneous frequency deviation output is defined as closed-loop equivalent nonlinear system, and after point-by-point addition with transmitting digital baseband signal, it is output to transmitting channel by digital-analog converter.The present application considers the compensation accuracy and the real-time constraint of vehicle-mounted embedded platform.

Description

Technical Field

[0001] This invention belongs to the field of automatic control technology, specifically relating to an online self-calibration and drift compensation method for a vehicle-mounted small radar. Background Technology

[0002] Vehicle-mounted small radars are widely used in advanced driver assistance systems such as adaptive cruise control, forward collision warning, and automatic emergency braking. Among them, X-band radars using the FMCW (Frequency-Modulated Wave) system have become the mainstream solution due to their combined ranging and speed measurement capabilities. The ranging accuracy and range resolution of FMCW radar largely depend on the frequency-time linearity of the transmitted chirped signal. Any nonlinear distortion that deviates from the ideal linear relationship will broaden the spectrum of the beat signal and increase the range sidelobes, thereby reducing the ability to detect weak targets at close range.

[0003] In actual vehicle-mounted operating environments, radar front-end devices are subjected to severe temperature fluctuations, mechanical vibrations, and power supply voltage fluctuations over long periods. The tuning characteristics of the voltage-controlled oscillator, the gain compression point of the power amplifier, and the conversion losses of the mixer all continuously drift with operating conditions, causing the nonlinear characteristics of the transmit chirp to change constantly in time and frequency bands. Currently, the commonly used compensation methods in engineering mainly fall into two categories. The first category is an open-loop compensation method based on temperature sensors and lookup tables. This method calibrates the frequency offset at different temperatures during the factory stage and stores it as a lookup table. During operation, the correction is made directly by looking up the table based on the temperature reading. This method is simple to implement, but it can only compensate for slow drifts related to a single variable of temperature. It cannot cope with complex nonlinear changes caused by device aging, power supply fluctuations, and multi-factor coupling. Furthermore, the accuracy of the lookup table is limited by the calibration temperature interval, and the correction lag is significant when the temperature changes rapidly. The second category is a closed-loop method based on external calibration targets. This method uses corner reflectors installed on the vehicle body or fixed ground objects at known distances as reference targets. The system drift is inferred by comparing the deviation between the measured echo and the expected echo. While these methods can capture comprehensive system biases, they are strictly dependent on the existence and stability of external reference targets. Calibration reliability drops significantly when targets are blocked, multipath interference occurs, or the scene changes. Furthermore, they cannot distinguish between errors introduced by nonlinearity of the transmission link and those introduced by the propagation environment.

[0004] Neither of the above two methods establishes the frequency band resolution capability for nonlinear drift of the transmit link. They treat the entire chirped bandwidth as a uniform compensation object, ignoring the differences in drift characteristics of different frequency bands, making it difficult to achieve high-precision full-band nonlinear compensation with limited model complexity. Summary of the Invention

[0005] The main objective of this invention is to provide an online self-calibration and drift compensation method for a vehicle-mounted small radar. The calibration process is completed entirely within a closed loop inside the module, without relying on external targets, and has continuous availability under all operating conditions. The segmented compensation strategy independently matches the inverse model for the differentiated drift characteristics of different frequency bands, taking into account both the compensation accuracy and the real-time constraints of the vehicle-mounted embedded platform.

[0006] To solve the above problems, the technical solution of the present invention is implemented as follows:

[0007] A method for online self-calibration and drift compensation of a vehicle-mounted small radar, the method comprising the following steps:

[0008] Step 1: In the front-end packaging module of the vehicle-mounted small radar, the output of the power amplifier of the transmitting channel is connected to the calibration input port of the mixer of the receiving channel via a directional coupler and a known dielectric delay line to form an FMCW chirped loopback calibration channel; during the calibration time slot, the FMCW chirped transmitting signal is coupled by the directional coupler and loops back to the mixer of the receiving channel via the known dielectric delay line for down-conversion, outputting the loopback beat signal and obtaining the digital sampling sequence of the loopback beat signal through analog-to-digital conversion;

[0009] Step 2: Perform Hilbert transform on the digital sampling sequence of the loop hysteresis signal to obtain the analytical signal sequence. Extract the instantaneous frequency sequence from the analytical signal sequence. Subtract the instantaneous frequency sequence from the ideal instantaneous frequency reference sequence recorded during factory calibration point by point to obtain the instantaneous frequency deviation sequence. Perform adaptive segmentation on the instantaneous frequency deviation sequence and extract the drift feature vector of each segment.

[0010] Step 3: Define the signal link from the input of the transmitted digital baseband signal to the output of the instantaneous frequency deviation as a closed-loop equivalent nonlinear system. Select the corresponding Volterra inverse model kernel parameter set from the drift feature classification table according to the drift feature vector of each segment interval. Perform Volterra inverse filtering operation on the transmitted digital baseband signal segment corresponding to each segment interval to obtain the compensation segment. Concatenate all the compensation segments into a compensation sequence, add it point by point with the transmitted digital baseband signal, and output it to the transmission channel through the digital-to-analog converter.

[0011] Furthermore, in step 1, the directional coupler is a Lange microstrip directional coupler; it is known that the dielectric delay line adopts a microstrip transmission line structure embedded in a low-temperature co-fired ceramic substrate, and has a fixed propagation delay value pre-calibrated at the factory.

[0012] Furthermore, in step 1, a platinum thin-film resistance temperature sensor is set at both the input and output ends of the known dielectric delay line to collect the temperature values ​​at both ends of the known dielectric delay line in real time and take the arithmetic mean as the delay line temperature measurement value; the temperature-corrected loop propagation delay value is obtained by looking up the temperature-delay lookup table pre-calibrated at the factory based on the delay line temperature measurement value.

[0013] Furthermore, in step 1, a calibration time slot is inserted between the normal operating time slots of the transmit channel and the receive channel. During the calibration time slot, the mixer input of the receive channel is switched from the antenna port to the calibration input port using an RF single-pole double-throw switch.

[0014] Furthermore, in step 2, the method for performing Hilbert transform on the digital sampling sequence of the loop hysteresis signal to obtain the analytic signal sequence is as follows: Perform an N-point discrete Fourier transform on the digital sampling sequence of the loop hysteresis signal to obtain a frequency domain complex sequence of length N; perform a one-sided spectrum construction process on the frequency domain complex sequence, multiplying the complex values ​​of the 2nd to N / 2th frequency points by 2 respectively, setting the complex values ​​of the N / 2+1th to Nth frequency points to 0, and keeping the complex values ​​of the 1st and N / 2+1th frequency points unchanged; perform an N-point inverse discrete Fourier transform on the frequency domain complex sequence after the one-sided spectrum construction process to obtain the analytic signal sequence.

[0015] Furthermore, in step 2, the instantaneous frequency sequence is extracted from the analytical signal sequence as follows: For each complex value of the analytical signal sequence, a four-quadrant arctangent operation is performed with its real and imaginary parts as inputs to obtain the wrapped phase value; along the direction of the sampling point number, when the absolute value of the difference between the wrapped phase values ​​of two adjacent sampling points exceeds pi, the phase values ​​of subsequent sampling points are accumulated by adding or subtracting one complete circumference to obtain a continuously unfolded instantaneous phase sequence; the difference between the instantaneous phase values ​​of two adjacent sampling points in the instantaneous phase sequence is calculated and divided by the sampling time interval to obtain the instantaneous frequency value, thus forming the instantaneous frequency sequence.

[0016] Furthermore, in step 2, after extracting the instantaneous frequency sequence and before subtracting it point by point from the ideal instantaneous frequency reference sequence, the instantaneous frequency sequence is subjected to descrambling and smoothing: sampling points in the analytical signal sequence with amplitudes lower than a preset amplitude threshold are marked, and the instantaneous frequency values ​​of the marked sampling points are replaced with the arithmetic mean of the instantaneous frequency values ​​of the unmarked sampling points within each preset number of neighboring sampling points; the replaced instantaneous frequency sequence is subjected to a sliding median filter with a preset window length to obtain a smoothed instantaneous frequency sequence.

[0017] Furthermore, in step 2, before subtracting the instantaneous frequency sequence from the ideal instantaneous frequency reference sequence point by point, the time axis coordinates of the instantaneous frequency sequence are converted to the transmit frequency coordinates point by point using the temperature-corrected loopback propagation delay and the FMCW chirped frequency modulation slope. The adaptive segmentation interval is divided as follows: an entry threshold and an exit threshold are set, with the exit threshold being less than the entry threshold; the absolute value of the difference between the instantaneous frequency deviation values ​​of adjacent sampling points in the instantaneous frequency deviation sequence is calculated point by point as the point-by-point deviation change; when the point-by-point deviation change rises from below the entry threshold to above the entry threshold for the first time, the candidate segmentation boundary state is entered; in the candidate segmentation boundary state, the number of sampling points whose point-by-point deviation change remains above the exit threshold is continuously counted. When the number of consecutive judgment points reaches a preset number, the position of the first sampling point that continuously exceeds the limit is marked as the segment boundary point. When the deviation change at each point decreases to below the exit threshold, the candidate segment boundary state is exited. All segment boundary points, together with the start and end points of the transmission frequency coordinates, define the entire segment interval. Segment intervals with fewer than the preset minimum segment length are merged with adjacent segment intervals with fewer than the preset minimum segment length into one segment interval. The drift feature vector consists of the deviation mean feature and the deviation slope feature. The deviation mean feature is the arithmetic mean of the absolute values ​​of the instantaneous frequency deviations of all sampling points in the segment interval. The deviation slope feature is the value obtained by dividing the difference of the instantaneous frequency deviation values ​​of the first and last sampling points in the segment interval by the number of sampling points in the segment interval.

[0018] Furthermore, in step 3, the Volterra inverse model kernel parameter set is a 3rd-order Volterra inverse model kernel parameter set with finite memory depth, containing a 1st-order linear kernel coefficient vector, a 2nd-order pruned nonlinear kernel coefficient matrix, and a 3rd-order pruned nonlinear kernel coefficient tensor. The 2nd-order pruned nonlinear kernel coefficient matrix retains only elements within the order range of the main diagonal and its two preset neighborhoods, and the 3rd-order pruned nonlinear kernel coefficient tensor retains only elements within the order range of the main diagonal and its two preset neighborhoods. The drift feature classification table divides the joint value range of the deviation mean feature and the deviation slope feature into multiple non-overlapping and continuously covered drift feature regions. Each drift feature region corresponds to storing one set of 3rd-order Volterra inverse model kernel parameter sets with finite memory depth. During the factory calibration stage, the temperature of the front-end packaging module is adjusted by the temperature control platform, a broadband excitation signal is injected into the transmit digital baseband port, and the output of the closed-loop equivalent nonlinear system is acquired through the FMCW chirped loopback calibration channel. The process involves performing a 3rd-order Volterra forward kernel identification with finite memory depth, followed by a 3rd-order inverse kernel solution and pruning of the 2nd and 3rd-order inverse kernels. The Volterra inverse filtering operation is as follows: a 3rd-order Volterra inverse filter is constructed using the 3rd-order Volterra inverse model kernel parameter set with finite memory depth. The 3rd-order Volterra inverse filter has three parallel processing branches. The first branch uses a 1st-order linear kernel coefficient vector to perform a linear discrete convolution operation on the transmitted digital baseband signal segment to obtain a 1st-order output sequence. The second branch uses a 2nd-order pruned nonlinear kernel coefficient matrix to perform a 2nd-order discrete Volterra convolution operation on the transmitted digital baseband signal segment to obtain a 2nd-order output sequence. The third branch uses a 3rd-order pruned nonlinear kernel coefficient tensor to perform a 3rd-order discrete Volterra convolution operation on the transmitted digital baseband signal segment to obtain a 3rd-order output sequence. The 1st, 2nd, and 3rd-order output sequences are summed point-by-point, and then subtracted point-by-point from the transmitted digital baseband signal segment to obtain the compensation segment.

[0019] Furthermore, in step 3, when splicing all compensation amount segments into a compensation amount sequence, a linear gradient transition zone with a preset number of transition sampling points is set at the splicing boundary of the compensation amount segments between two adjacent segment intervals. Within the linear gradient transition zone, for each transition sampling point, the value of the compensation amount segment of the previous segment interval at the corresponding sampling point is linearly interpolated and mixed with the value of the compensation amount segment of the next segment interval at the corresponding sampling point. The interpolation ratio linearly transitions along the linear gradient transition zone from completely taking the output of the previous segment interval to completely taking the output of the next segment interval.

[0020] This invention provides an online self-calibration and drift compensation method for a vehicle-mounted small radar, which has the following advantages: By constructing an FMCW chirped loopback calibration channel within the radar front-end encapsulation module, a portion of the transmitted chirped signal is directly looped back to the receiving channel via a known dielectric delay line for mixing and sampling. This allows the calibration process to be completed entirely within a closed loop within the module, independent of any external target or specific scattering environment. This eliminates the risk of failure associated with traditional external calibration methods when targets are obstructed or the scene changes, ensuring the continuous availability of calibration data under any operating conditions. By applying real-time temperature monitoring and temperature-delay lookup table correction using a platinum thin-film resistor temperature sensor to the known dielectric delay line, the interference of the delay line's own thermal drift on the calibration reference is effectively isolated, ensuring that the propagation characteristics of the loopback reference path remain known and reliable. At the signal processing level, Hilbert transform is used to extract point-by-point instantaneous frequency information from the loopback beat signal and subtract it from the factory reference. This allows for a fine-grained characterization of the nonlinear deviation distribution of the transmitted chirp across the entire frequency sweep range at the sampling point level, demonstrating significantly higher time-frequency resolution compared to traditional global spectrum analysis methods. A hysteresis dual-threshold adaptive segmentation mechanism is introduced. Based on the drastic point-by-point changes in instantaneous frequency deviation, the frequency band boundaries where drift characteristics change are automatically identified. The entire frequency band is divided into several segmented intervals with similar drift characteristics. This allows the subsequent compensation model to independently match the specific drift pattern of each interval, avoiding the trade-off problem of fitting accuracy when a single global model has significant drift differences across different frequency bands. By selecting the corresponding pruned Volterra inverse model kernel parameter set for each segmented interval based on the drift feature vector for pre-distortion compensation, the computational complexity is kept within the tolerance range of the vehicle-mounted embedded platform while effectively capturing the higher-order nonlinear effects of the transmit link, achieving a balance between compensation accuracy and real-time performance. The linear gradual transition processing at the segment boundaries ensures the continuous smoothness of the compensation sequence across the entire frequency band, avoiding additional spectral spurious signals introduced by segment splicing abrupt changes. Attached Figure Description

[0021] Figure 1 A schematic diagram of the hardware connection relationship of the FMCW chirped loopback calibration channel in an embodiment of the present invention is provided for this embodiment.

[0022] Figure 2 A schematic diagram illustrating the distribution of the instantaneous frequency deviation sequence along the transmission frequency coordinate direction and the results of adaptive segmentation interval division provided in this embodiment of the invention;

[0023] Figure 3 This is a frequency domain schematic diagram illustrating the single-side spectrum construction process performed on the digital sampling sequence of the loop hysteresis signal provided in an embodiment of the present invention. Detailed Implementation

[0024] A method for online self-calibration and drift compensation of a vehicle-mounted small radar, the method comprising the following steps:

[0025] Step 1: In the front-end packaging module of the vehicle-mounted small radar, the output of the power amplifier of the transmitting channel is connected to the calibration input port of the mixer of the receiving channel via a directional coupler and a known dielectric delay line to form an FMCW chirped loopback calibration channel; during the calibration time slot, the FMCW chirped transmitting signal is coupled by the directional coupler and loops back to the mixer of the receiving channel via the known dielectric delay line for down-conversion, outputting the loopback beat signal and obtaining the digital sampling sequence of the loopback beat signal through analog-to-digital conversion;

[0026] Step 2: Perform Hilbert transform on the digital sampling sequence of the loop hysteresis signal to obtain the analytical signal sequence. Extract the instantaneous frequency sequence from the analytical signal sequence. Subtract the instantaneous frequency sequence from the ideal instantaneous frequency reference sequence recorded during factory calibration point by point to obtain the instantaneous frequency deviation sequence. Perform adaptive segmentation on the instantaneous frequency deviation sequence and extract the drift feature vector of each segment.

[0027] Step 3: Define the signal link from the input of the transmitted digital baseband signal to the output of the instantaneous frequency deviation as a closed-loop equivalent nonlinear system. Select the corresponding Volterra inverse model kernel parameter set from the drift feature classification table according to the drift feature vector of each segment interval. Perform Volterra inverse filtering operation on the transmitted digital baseband signal segment corresponding to each segment interval to obtain the compensation segment. Concatenate all the compensation segments into a compensation sequence, add it point by point with the transmitted digital baseband signal, and output it to the transmission channel through the digital-to-analog converter.

[0028] In practical engineering applications of vehicle-mounted small radars, the frequency-time characteristics of the FMCW chirped transmission signal deviate from the ideal linear relationship calibrated at the factory due to factors such as changes in ambient temperature, component aging, and fluctuations in power supply voltage. This results in spectral broadening of the beat signal, increase in range sidelobes, and a decrease in ranging accuracy. To detect and compensate for the aforementioned nonlinear drift in real time during continuous radar operation, a closed-loop reference path independent of external targets needs to be established within the radar front-end package module. The core idea of ​​this reference path is to directly send a small portion of the chirped signal output from the transmission channel to the receiving channel for mixing after passing through a delay medium with known and stable propagation characteristics. The resulting beat signal is entirely determined by the nonlinearity and drift of the transmission link itself, unaffected by interference from the external scattering environment, thus providing a clean internal reference standard for subsequent calibration algorithms.

[0029] Specifically, the entire FMCW chirped loop calibration channel is constructed within the front-end package module of the vehicle-mounted miniature radar. The front-end package module employs system-in-package (SiP) technology, integrating the FMCW chirped signal generator, power amplifier, low-noise amplifier for the receiving channel, mixer, analog-to-digital converter, and all passive components of the calibration channel into a single package. In the transmitting channel, the FMCW chirped signal generator produces a chirped baseband signal with a frequency that increases linearly with time. After up-conversion and power amplification, this signal is output to the antenna port for target detection. A directional coupler is installed at the output of the power amplifier. The through port of the directional coupler connects to the transmitting antenna feed network, while the coupling port outputs a low-amplitude copy of the chirped signal as the excitation source for the calibration channel. The isolation port of the directional coupler is connected to a matched load to absorb reverse leakage energy.

[0030] The coupling port of the directional coupler is connected to the input of the known dielectric delay line via a transmission line trace inside the package module. The output of the known dielectric delay line is connected to the calibration input port of the mixer in the receiving channel via a transmission line trace inside the package module. The mixer in the receiving channel has two RF input ports: one is an antenna port, used to receive the target echo signal in normal operating mode; the other is a calibration input port, specifically used to receive the loopback chirp signal from the known dielectric delay line in calibration mode. The complete signal path from the coupling port of the directional coupler through the known dielectric delay line to the calibration input port of the mixer in the receiving channel constitutes the FMCW chirped loopback calibration channel.

[0031] In a preferred embodiment, the directional coupler uses a Lange microstrip directional coupler structure for selection and implementation. The Lange microstrip directional coupler consists of multiple parallel microstrip conductors arranged interdigitally and bridged by bonding wires, utilizing electromagnetic coupling between adjacent conductors to achieve power distribution. Compared to ordinary parallel-line couplers, the interdigital structure of the Lange microstrip directional coupler increases the difference between odd-mode and even-mode impedances, enabling the maintenance of a stable coupling amount close to the preset coupling ratio over a wide frequency range, while maintaining high directivity. In the typical operating frequency band of 9.3GHz to 9.5GHz for automotive small radars, the coupling fluctuation of the Lange microstrip directional coupler can be controlled within ±0.5dB, with directivity better than 20dB. Setting the coupling ratio requires balancing the trade-off between calibration signal strength and transmit power loss: excessive coupling significantly reduces the effective power transmitted to the antenna, thus shortening the radar detection range; insufficient coupling results in an inadequate signal-to-noise ratio for the loopback beat signal to support subsequent phase extraction accuracy. In a typical implementation, the coupling degree of the directional coupler is set to -20dB, meaning that only about one percent of the transmit power output from the power amplifier is taken into the calibration channel, resulting in an impact of less than 0.1dB on the effective radiated power of the transmit channel. In another optional implementation, the directional coupler can also employ other microstrip coupling structures such as branch-line couplers or capacitor-compensated couplers, as long as the requirements for coupling flatness and directivity are met within the radar operating frequency band.

[0032] The purpose of a known dielectric delay line is to introduce a known and stable propagation delay into the loopback chirp signal, creating a clear time difference between the loopback chirp signal and the local oscillator reference signal, thus forming a beat signal with a defined frequency at the mixer output. Without a known dielectric delay line, the time delay between the loopback chirp signal and the local oscillator reference signal would be generated solely by the extremely short transmission line trace within the package module, resulting in a very low beat frequency, even close to DC, making it difficult to distinguish from the mixer's DC offset and low-frequency flicker noise. By introducing a known dielectric delay line, the loopback beat frequency is raised to a frequency band far from DC, effectively avoiding areas of high low-frequency noise in subsequent digital processing.

[0033] In a preferred embodiment, the known dielectric delay line employs a microstrip transmission line structure embedded in a low-temperature co-fired ceramic substrate. The low-temperature co-fired ceramic substrate is formed by stacking and co-firing multiple layers of ceramic green sheets, with metal conductor patterns embedded between each layer. The embedded microstrip transmission line is a meandering, folded microstrip conductor trace formed within the inner layers of the low-temperature co-fired ceramic substrate, utilizing the high dielectric constant of the ceramic medium to achieve the required electrical length within a limited physical size. The dielectric constant of low-temperature co-fired ceramic materials is typically in the range of 7.0 to 7.8, and the propagation speed of electromagnetic waves within it is approximately 37% to 38% of the speed of light in free space. Therefore, a propagation delay much greater than that of an air transmission line can be obtained for the same physical length. In a typical embodiment, the physical fold length of the known dielectric delay line is approximately 15 mm, corresponding to a one-way propagation delay of approximately 0.5 ns. The resulting loopback frequency in the 9.5 GHz band is on the order of several hundred Hz to several kHz, falling within the effective sampling bandwidth of the analog-to-digital converter and far from DC. The fixed propagation delay value of the known dielectric delay line is precisely measured and recorded at the factory using a vector network analyzer, with a measurement accuracy better than ±0.01 ns. In another optional embodiment, the known dielectric delay line can also adopt a stripline structure embedded in the package substrate, a thin-film multilayer dielectric transmission line structure, or a ceramic waveguide delay line structure, as long as the long-term stability of the propagation delay within the radar operating temperature range meets the calibration accuracy requirements.

[0034] Because the dielectric constant of low-temperature co-fired ceramic materials drifts slowly with temperature, the propagation delay value of a known dielectric delay line also shifts with temperature. Without correction, this shift will be mistaken for a nonlinear drift in the transmission link, leading to inaccurate compensation. Therefore, a platinum thin-film resistance temperature sensor is installed at both the input and output ends of the known dielectric delay line. The platinum thin-film resistance temperature sensor utilizes the physical property that the resistance of a platinum metal thin film changes approximately linearly with temperature to achieve temperature measurement, offering advantages such as good long-term stability and high measurement accuracy. In a typical implementation, the measurement accuracy of the platinum thin-film resistance temperature sensor is better than ±0.5 degrees Celsius, and the response time is less than 50 ms. The arithmetic mean of the temperature readings from the platinum thin-film resistance temperature sensor at the input and output ends of the known dielectric delay line is taken as the delay line temperature measurement value, used to characterize the overall temperature level of the known dielectric delay line. The reason for taking the arithmetic mean of the temperature readings from the input and output temperature sensors instead of just a single point temperature is that the winding trace of the dielectric delay line may cross different heat distribution areas inside the package module. A single point temperature cannot represent the equivalent temperature state of the entire delay line, while the arithmetic mean of the two points at the input and output ends can more reasonably characterize the average temperature of the delay line within an engineering-acceptable accuracy range.

[0035] During the factory calibration phase, the radar front-end packaged module is placed in a programmable temperature chamber. The chamber temperature is adjusted point-by-point within a preset temperature range at preset temperature step intervals. After each temperature point stabilizes, the propagation delay value of a known medium delay line at that temperature is accurately measured using an external vector network analyzer. All temperature points and their corresponding propagation delay values ​​are stored as a temperature-delay lookup table and written to the radar's on-chip non-volatile memory. In a typical implementation, the temperature range covers -40 degrees Celsius to 85 degrees Celsius, with a temperature step interval of 5 degrees Celsius, totaling 26 temperature calibration points. During radar online operation, at each calibration time slot, a linear interpolation lookup is performed in the temperature-delay lookup table based on the current delay line temperature measurement to obtain the temperature-corrected loopback propagation delay value. This temperature-corrected loopback propagation delay value will be used in subsequent steps to accurately convert the time axis coordinates of the beat signal into the instantaneous transmission frequency coordinates of the FMCW chirped transmission signal.

[0036] refer to Figure 1 In the front-end packaging module of the vehicle-mounted small radar, the transmission channel sequentially includes an FMCW chirped signal generator, an up-converter, and a power amplifier. These three components are connected in series to form a complete transmission link from the digital baseband signal to the radio frequency transmission signal. The FMCW chirped signal generator produces a chirped baseband signal with a frequency that increases linearly with time. After being shifted to the X-band operating frequency by the up-converter, it is sent to the power amplifier for power amplification. The amplified signal is then sent to the transmitting antenna for radiation into the external space via the output of the power amplifier. A Lange microstrip directional coupler is set at the output of the power amplifier. The Lange microstrip directional coupler has four ports: an input port, a through port, a coupling port, and an isolation port. The input port is connected to the output of the power amplifier, the through port is connected to the transmitting antenna, and most of the transmission power is sent to the antenna for normal radiation via the through port. The coupling port extracts a low-amplitude copy of the transmitted signal according to a preset coupling ratio as the excitation source for the calibration channel. The isolation port is connected to a matching load to absorb reverse leakage energy. Figure 1 The grounding symbol indicates the matched load. The coupling port of the Lange microstrip directional coupler is connected to the input of a known dielectric delay line via a transmission line trace inside the package module. The known dielectric delay line employs a microstrip transmission line structure embedded in a low-temperature co-fired ceramic substrate. Figure 1 LTCC stands for Low Temperature Co-fired Ceramics. The output of the known dielectric delay line is connected to the calibration input port of the mixer in the receiving channel. The complete signal path from the coupling end of the Lange microstrip directional coupler through the known dielectric delay line to the calibration input port of the mixer constitutes the FMCW chirped loopback calibration channel. Figure 1The range of devices covered by this channel is marked with a dashed box. Two platinum thin-film resistance temperature sensors, RTD-1 and RTD-2, are installed at the input and output ends of the known dielectric delay line, respectively. The temperature values ​​at both ends of the delay line are acquired in real time, and after arithmetic averaging, the temperature-corrected loopback propagation delay value is obtained by consulting a temperature-delay lookup table. The receiving channel is located at... Figure 1 The lower section comprises an RF single-pole double-throw switch, a mixer, a low-pass anti-aliasing filter, and an ADC. The RF single-pole double-throw switch has two switching positions: an antenna port and a calibration input port. During normal operation, it connects to the antenna port to receive external target echoes; during calibration, it switches to the calibration input port to receive loopback chirped signals from a known dielectric delay line. The mixer down-converts the loopback chirped signal received at the calibration input port with the local oscillator reference signal, outputting a loopback beat signal. This loopback beat signal is then processed sequentially by the low-pass anti-aliasing filter and the ADC to obtain a digital sampling sequence of the loopback beat signal, which serves as input data for subsequent steps. The entire channel forms a closed-loop reference path from sampling at the transmitter, looping back through a known delay, to digital acquisition at the receiver. All calibration signal generation and acquisition are completed within the front-end package module, independent of external targets.

[0037] Regarding the scheduling method of calibration time slots, calibration time slots are inserted between the normal operating time slots of the transmit and receive channels. Normal operating time slots refer to the regular working cycle in which the radar transmits chirped signals into external space and receives target echoes for ranging and velocity measurement. Calibration time slots refer to dedicated cycles in which the radar suspends external detection and performs internal loopback self-checks. The two are alternated in time, and the duration of a calibration time slot is typically the length of one complete chirped cycle, approximately tens of microseconds. In a typical implementation, one calibration time slot is inserted every 100 normal operating time slots, with a duty cycle of approximately one percent, having a negligible impact on the radar's normal detection performance. During the calibration time slot, the mixer input of the receive channel is switched from the antenna port to the calibration input port using an RF single-pole double-throw switch. The radio frequency (RF) single-pole double-throw (SPDT) switch is an active microwave switching device with two mutually exclusive on / off states: during the normal operation time slot, the RF SPDT switch connects the mixer input to the antenna port to receive external echo signals; during the calibration time slot, the RF SPDT switch switches the mixer input to the calibration input port to receive loopback chirped signals from a known dielectric delay line. The switching time of the RF SPDT switch is typically on the order of 10 ns, much shorter than the chirp period, so the switching process does not affect the acquisition of the calibration signal. In another alternative implementation, the RF SPDT switch can be omitted, and a separate calibration receiving mixer can be set in the receiving channel, configured in parallel with the signal receiving mixer used for normal operation. This eliminates the need to switch the antenna path during calibration, but this approach increases the hardware complexity and chip area of ​​the receiving channel.

[0038] In the specific signal flow of the calibration time slot, the FMCW chirp signal generator generates a chirped baseband signal using the exact same frequency modulation parameters as the normal operating time slot. This is crucial for the effectiveness of the calibration, as the purpose of calibration is to observe the nonlinear response of the transmit link under excitation conditions completely consistent with actual operation. If different frequency modulation parameters are used in the calibration time slot, the measured nonlinear characteristics will not represent the normal operating state. After up-conversion and amplification by a power amplifier, a portion of the power of the chirped baseband signal is led out through the coupling port of the Lange microstrip directional coupler into a known dielectric delay line. The loopback chirped signal, after undergoing a fixed propagation delay in the known dielectric delay line, reaches the calibration input port of the mixer in the receiving channel. The mixer in the receiving channel performs a down-conversion mixing operation between the loopback chirped signal received at the calibration input port and the local oscillator reference signal, outputting a loopback beat signal. In the FMCW system, the instantaneous frequency of the loopback beat signal is proportional to the time delay difference between the loopback chirp signal and the local oscillator reference signal, and this time delay difference is mainly determined by the propagation time delay of the known dielectric delay line. If the transmit chirp is perfectly linear, the loopback beat signal should be a single-tone signal with a constant frequency; in reality, due to the nonlinear drift of the transmit link, the instantaneous frequency of the loopback beat signal will fluctuate around the nominal single-tone frequency, and these fluctuations are precisely the target quantities that need to be extracted and compensated for in subsequent steps.

[0039] After the loopback beat signal is output from the mixer, it passes through a low-pass anti-aliasing filter in the receiving channel to filter out high-frequency mixing products and out-of-band noise, retaining the useful signal components near the beat frequency. The cutoff frequency of the low-pass anti-aliasing filter is set to less than half of the analog-to-digital converter (ADC) sampling frequency to satisfy the Nyquist sampling theorem and prevent high-frequency components from folding into the beat frequency band and generating spurious frequency components. The loopback beat signal after passing through the low-pass anti-aliasing filter enters the ADC for uniform time interval sampling. The ADC continuously samples at a fixed sampling clock for the duration of one complete chirped cycle, obtaining a digital sampling sequence of the loopback beat signal containing all sampling points. In a typical implementation, the ADC sampling frequency is 20MHz, the effective resolution is 12 bits, and a total of 800 sampling points are collected within a single chirped cycle of 40 microseconds. This digital sampling sequence of the loopback beat signal completely records the time-domain waveform of the loopback beat signal within one chirped cycle and serves as the input data for subsequent steps of performing Hilbert transform and instantaneous frequency extraction.

[0040] After obtaining the digital sampling sequence of the loop hysteresis signal in step 1, it is necessary to accurately extract the instantaneous frequency deviation information reflecting the nonlinear drift of the transmit link from this set of time-domain sampled data. While direct spectral analysis of the digital sampling sequence of the loop hysteresis signal can reveal spectral broadening, spectral analysis only provides global frequency distribution characteristics and cannot reveal the trajectory of instantaneous frequency changes over time at each sampling point. The essence of transmit chirp nonlinear drift lies precisely in the point-by-point variation of instantaneous frequency deviating from the ideal value; therefore, an analytical method capable of providing point-by-point time-frequency information must be employed. The Hilbert transform provides a classic path for constructing an analytic signal from a real-valued time-domain signal and then extracting the instantaneous phase and frequency. Its computational structure is regular and suitable for efficient implementation on digital signal processing hardware, making it particularly suitable for embedded real-time calibration scenarios.

[0041] The specific process of performing a Hilbert transform on the digitally sampled sequence of the loop hysteresis signal to obtain an analytic signal sequence is as follows. Suppose the digitally sampled sequence of the loop hysteresis signal contains a total of... There are 1 sampling points, among which It is a positive integer, in a typical implementation Take 800, which corresponds to the total number of sampling points of the analog-to-digital converter within a single chirp cycle in step 1. First, for this... Execute at each sampling point The point-based discrete Fourier transform converts a real-valued sampled sequence in the time domain to the frequency domain, resulting in a sequence of length . A frequency domain complex sequence. Each frequency point in the frequency domain complex sequence corresponds to a complex value, which contains the amplitude and phase information of that frequency component.

[0042] Subsequently, a one-sided spectrum construction process is performed on the frequency domain complex sequence. The purpose of the one-sided spectrum construction process is to eliminate negative frequency components, so that the signal obtained after the inverse transform back to the time domain becomes a complex analytic signal. The real part of the analytic signal is the original signal itself, and the imaginary part is the Hilbert transform result of the original signal. The two form a pair of orthogonal components, which together describe the instantaneous amplitude and instantaneous phase of the signal at each moment. The operation rule of the one-sided spectrum construction process is: the first frequency point (corresponding to the DC component) and the... Adding one frequency point (corresponding to the Nyquist frequency component) keeps the complex value unchanged; the second to the third... The complex values ​​of each frequency point are multiplied by 2 to compensate for the energy loss of the positive frequency components by doubling their original value; the energy loss of the negative frequency components is compensated by multiplying the complex values ​​of each frequency point by 2. Add 1 to the number The complex values ​​of each frequency point are set to 0 to completely eliminate negative frequency components. The reason for multiplying the positive frequency components by 2 is that the spectrum of a real-valued signal has conjugate symmetry, with the positive and negative frequencies each carrying half of the signal energy. If the negative frequencies are not compensated after removal, the amplitude of the analytical signal will be reduced to half of the original signal, affecting the subsequent judgment of instantaneous amplitude.

[0043] Perform on the frequency domain complex sequence after single-sided spectrum construction. The point-wise discrete Fourier inverse transform, returning to the time domain, yields the analytic signal sequence of the loop hysteresis signal. Each sampling point in the analytic signal sequence is a complex value, denoted as the i-th... The complex value of each sampling point is ,in The sampling point number ranges from 0 to... . The real part of the original loop hysteresis signal is the first hysteresis signal. The sampled values ​​at the nth sampling point, the imaginary part being the original signal after Hilbert transform at the nth sampling point. The value of each sampling point.

[0044] Extracting the instantaneous frequency sequence from an analytic signal sequence requires three steps: instantaneous phase extraction, phase unwrapping, and point-by-point differencing. This involves processing the complex values ​​of each sampling point in the analytic signal sequence. In fact and the virtual part Perform a four-quadrant arctangent operation on the input to obtain the wrap-around phase value for each sampling point. .here Representing complex values The real part, Representing complex values The imaginary part, Indicates the first The phase value of each sampling point. The four-quadrant arctangent operation differs from the ordinary arctangent operation, which only takes the ratio of the imaginary part to the real part as input, and its output range is limited to... arrive It cannot distinguish between quadrant differences where both the real and imaginary parts are simultaneously positive or simultaneously negative; the four-quadrant arctangent operation considers the signs of both the real and imaginary parts simultaneously, and the output range covers the entire quadrant. arrive It can correctly determine the phase angle in all four quadrants of the complex plane. Using the four-quadrant arctangent operation can avoid phase jumps caused by quadrant ambiguity, reducing unnecessary processing burden for subsequent phase untangling.

[0045] Package phase value The range of values ​​is limited to arrive Between these, the true instantaneous phase of the loopback beat signal continuously accumulates and increases over time, and its absolute value far exceeds... Whenever the true phase crosses or When the boundary is reached, the wrapping phase value will change. The abrupt jump in phase is called phase entanglement. To recover the true continuous instantaneous phase, phase unwinding is required along the sampling point index direction. Specifically, starting from index 0, the system scans point by point backwards, calculating the difference in the wrapped phase values ​​between adjacent sampling points. ,in Indicates the first The sampling point and the first The difference in phase value is encapsulated between each sampling point. When The absolute value exceeds At that time, it is determined that a phase entanglement transition has occurred here: if Negative values ​​and absolute values ​​exceeding This indicates that the true phase continuously increases in the positive direction, but the wrapped phase has changed from... Jump back The rollback, at this time for the first The phase values ​​of the first and all subsequent sampling points are accumulated. ;like It is a positive value and its absolute value exceeds Then for the first Phase value cumulative reduction of each and all subsequent sampling points After the above processing, the phase values ​​of all sampling points constitute a continuously expanded instantaneous phase sequence. ,in Indicates the first The instantaneous phase value after unwinding at each sampling point.

[0046] For instantaneous phase sequence Perform point-by-point difference operations to extract instantaneous frequencies. For the ... Each sampling point has an instantaneous frequency value. Divide the difference in instantaneous phase values ​​between two adjacent sampling points by the sampling time interval. Divide by To obtain, that is ,in Indicates the first The instantaneous frequency value of each sampling point. This represents the sampling time interval of the analog-to-digital converter, which is equal to the reciprocal of the sampling frequency. In a typical implementation with a sampling frequency of 20MHz, This is equal to 50 ns. The above difference operation is performed on all valid sampling points one by one to obtain the instantaneous frequency sequence. It should be noted that the 0th sampling point has no preceding sampling point available for difference, therefore the effective length of the instantaneous frequency sequence is... One point.

[0047] Instantaneous frequency sequences are obtained directly from phase difference. In regions where the analytic signal amplitude is high and phase changes are stable, the extraction accuracy of instantaneous frequencies is good. However, in actual acquisition, the loopback beat signal may experience instantaneous amplitude dips at certain times due to receiver channel gain fluctuations or non-ideal characteristics of the mixer, causing a significant decrease in the amplitude of the analytic signal at these sampling points. When the analytic signal amplitude approaches the noise floor, phase extraction at this point will be severely interfered with by noise. The output of the four-quadrant arctangent operation will be almost randomly distributed under the dominance of noise, resulting in sharp spikes in the differential frequency values. These spikes do not reflect the true frequency drift but are purely noise products. If left untreated, these spurious spikes will severely interfere with subsequent segmentation and drift feature extraction.

[0048] Therefore, after obtaining the instantaneous frequency sequence, it is first subjected to desqueezing and smoothing processing before subsequent calculations. The desqueezing process involves backtracking and analyzing the amplitude (i.e., complex value) of each sampling point in the signal sequence. The sampling points with amplitudes below a preset amplitude threshold are marked as unreliable sampling points with low signal-to-noise ratios. The preset amplitude threshold should be selected so that, under typical operating conditions, only a very small number of sampling points with severely distorted amplitudes of the analytic signal are marked, while sampling points with normal amplitudes remain unaffected. In a typical implementation, the preset amplitude threshold is set to 10% of the median amplitude of all sampling points in the analytic signal sequence. For each marked unreliable sampling point, its instantaneous frequency value is replaced with the arithmetic mean of the instantaneous frequency values ​​of all unmarked sampling points within a preset number of neighboring sampling points before and after that sampling point. In a typical implementation, the preset number of neighboring sampling points is 5, that is, the instantaneous frequency values ​​of 5 unmarked sampling points forward and 5 unmarked sampling points backward are averaged and then filled into the marked position. The advantage of this local replacement strategy is that it only repairs isolated points with severe noise pollution, without changing the frequency values ​​of the remaining normal sampling points, thus preserving the true drift information to the maximum extent.

[0049] After descrambling eliminates the spike transitions caused by amplitude traps, the instantaneous frequency sequence may still retain small high-frequency fluctuations introduced by phase noise and quantization noise. These fluctuations are usually much smaller than the spikes before descrambling, but if directly used for differential calculations of point-by-point deviation changes, they may still trigger false boundaries in the segmented interval division. Therefore, a sliding median filter with a preset window length is further applied to the descrambled instantaneous frequency sequence. Median filtering differs from mean filtering; it uses the median of all sampled values ​​within the sliding window as the output, naturally suppressing isolated outliers without blurring the abrupt edges of signal transitions like mean filtering. In a typical implementation, the preset window length is 7 sample points, a length sufficient to smooth the small fluctuations caused by quantization and phase noise without over-smoothing and masking the true drift characteristics. The smoothed instantaneous frequency sequence is obtained after sliding median filtering. In another alternative implementation, a Savitzky-Golay filter can be used instead of a sliding median filter. The Savitzky-Golay filter achieves smoothing through local polynomial fitting, which can better preserve the local trend shape of the signal, but the computational cost is slightly higher than that of median filtering.

[0050] Next, the time axis coordinates of the smoothed instantaneous frequency sequence need to be converted to the instantaneous transmit frequency coordinates of the FMCW chirped transmit signal. The purpose of this coordinate conversion is to establish a one-to-one correspondence between the sampling time of the loopback beat signal and the frequency sweep position of the transmit chirp, thereby enabling precise location of which frequency interval within the chirp sweep range each segment corresponds to in subsequent segmented compensation. The conversion is based on the mapping relationship between the beat signal time axis and the transmit frequency under the FMCW system: Under ideal linear chirp conditions, the beat signal's... The instantaneous transmission frequency corresponding to each sampling point Equal to the chirping initiation frequency Add chirp FM slope The product of the launch times corresponding to the sampling point, where the launch time is Adding the temperature-corrected loop propagation delay value Half of it. Here. The starting frequency for FMCW chirped transmission signals is typically 9.475 GHz or 9.325 GHz in X-band vehicle radar. The FMCW chirp modulation slope represents the rate of change of the chirp frequency over time. In a typical implementation... Approximately 10 MHz / microsecond; This is the temperature-corrected loop propagation delay value obtained in step 1. The reason for adding this is... Half of the time offset is corrected because the loop back beat signal is... Each sampling point actually corresponds to the transmitted signal being earlier than the current reception time. The mixing result between the frequency component emitted by the time and the current local oscillator frequency requires precise timing alignment to account for the propagation delay. After coordinate transformation, each sampling point in the smoothed instantaneous frequency sequence obtains its precise position label on the transmit frequency axis.

[0051] The instantaneous frequency deviation sequence is obtained by subtracting the smoothed instantaneous frequency sequence from the ideal instantaneous frequency reference sequence, which was recorded and stored under the same FMCW chirped loopback calibration channel during factory calibration. The ideal instantaneous frequency reference sequence is reference data acquired during factory calibration, when radar parameters are at their nominal operating points and the ambient temperature is the reference temperature (typically 25 degrees Celsius), using the same loopback channel and the same Hilbert transform process, and permanently stored in on-chip non-volatile memory. In the instantaneous frequency deviation sequence obtained after point-by-point subtraction, the value of each sampling point represents the degree to which the instantaneous frequency at that frequency position deviates from the factory reference under the current operating conditions. A positive deviation indicates that the current instantaneous frequency is higher than the ideal reference, and a negative deviation indicates that it is lower than the ideal reference. When the states of the devices in the transmit link do not drift, the instantaneous frequency deviation sequence should be close to zero everywhere; with temperature changes, device aging, or power supply fluctuations, the instantaneous frequency deviation sequence will exhibit a specific spatial distribution pattern.

[0052] Within the FMCW chirped frequency sweep range, the nonlinear drift characteristics of different frequency bands are often inconsistent. For example, in the low-frequency band near the chirped start frequency, the tuning sensitivity of the voltage-controlled oscillator (VCO) is high, and the frequency shift caused by temperature drift may manifest as a slow, monotonically bending motion. In the middle of the chirped range, the gain compression effect of the power amplifier may introduce local frequency fluctuations. In the high-frequency band near the chirped end frequency, the VCO may enter the nonlinear saturation region of the tuning curve, and the rate of change of frequency deviation will increase significantly. Attempting to cover the differentiated drift characteristics of all frequency bands with a single global compensation model inevitably faces a sharp contradiction between fitting accuracy and model complexity. Dividing the chirped frequency band into several segmented intervals and independently selecting the best-matching compensation model within each segmented interval can achieve high compensation accuracy with a lower single-interval model order. This is precisely the engineering starting point for implementing adaptive segmented interval division.

[0053] The adaptive segmentation interval is used to scan the instantaneous frequency deviation sequence point by point along the transmission frequency coordinate direction. The segmentation is based on the degree of change in instantaneous frequency deviation between adjacent sampling points: in regions where the drift characteristics are stable, the deviation changes of adjacent sampling points are small and can be classified into the same segment interval; when the deviation changes suddenly increase and continue to exceed a certain threshold, it is considered that the drift characteristics have changed significantly, and a new segment boundary should be drawn at this point.

[0054] To prevent accidental fluctuations caused by noise from triggering false segmentation boundaries, the segmentation algorithm introduces a hysteresis double-threshold mechanism. An entry threshold value is set. and exit threshold ,in The threshold for the point-by-point deviation change required to enter the candidate segment boundary state. The threshold for exiting the candidate segment boundary state. Less than This hysteresis design, where the entry threshold is higher than the exit threshold, gives the segmented decision-making a natural ability to resist noise jitter: even if a noise spike occasionally touches the entry threshold, it will quickly fall back below the exit threshold due to the change in deviation of subsequent sampling points, thus failing to accumulate enough consecutive decision points and forming an effective segment boundary. In a typical implementation, the entry threshold is set to 3 times the global standard deviation of the instantaneous frequency deviation sequence, and the exit threshold is set to 1.5 times the global standard deviation.

[0055] The specific process of point-by-point scanning is as follows: Calculate the absolute value of the difference between the instantaneous frequency deviation value of the current sampling point and the instantaneous frequency deviation value of the previous sampling point in the instantaneous frequency deviation sequence, and use this as the point-by-point deviation change. The initial state is normal scanning. When the point-by-point deviation change rises from below the entry threshold to above the entry threshold for the first time, the scanning state switches to candidate segment boundary state. In the candidate segment boundary state, the number of consecutive sampling points whose point-by-point deviation change remains above the exit threshold is continuously counted. When this number of consecutive sampling points reaches the preset number of consecutive judgment points—in a typical implementation, the preset number of consecutive judgment points is 5—the position of the first sampling point in this continuous excess is marked as a segment boundary point. Then, scanning continues. When the point-by-point deviation change drops below the exit threshold, the candidate segment boundary state is exited, the consecutive point counter is reset to zero, normal scanning state is restored, and subsequent scanning continues. If, in the candidate segment boundary state, the number of consecutive sampling points has not yet reached the preset number of consecutive judgment points, and the point-by-point deviation change drops below the exit threshold, then the candidate boundary is abandoned, no mark is made, the candidate segment boundary state is directly exited, and the scanning continues.

[0056] After scanning, all marked segment boundary points, along with the start and end points of the transmission frequency coordinates, are used as the boundaries of the segment intervals, with each pair of adjacent boundaries forming one segment interval. Since the adaptive segmentation process is entirely data-driven, extremely short segment intervals with too few sampling points may occur in some local areas where drift changes are extremely frequent. The sampling data within these extremely short segment intervals is insufficient to support reliable drift feature statistics and cannot provide sufficient context length for subsequent Volterra inverse filtering operations. Therefore, a preset minimum segment length is set as a constraint; in a typical implementation, the preset minimum segment length is 20 sampling points. Segments containing fewer sampling points than the preset minimum segment length are merged with their adjacent segment intervals containing fewer sampling points into a single segment interval. The merging operation always selects the shorter of the two adjacent intervals to maintain a balance in interval length as much as possible.

[0057] For each merged segment interval, a drift feature vector is extracted for subsequent selection of the Volterra inverse model kernel parameter set. The drift feature vector consists of two components: the mean deviation feature and the deviation slope feature. The mean deviation feature is equal to the arithmetic mean of the absolute values ​​of the instantaneous frequency deviations of all sampling points within the segment interval, reflecting the average magnitude of the drift within that interval. The deviation slope feature is equal to the difference between the instantaneous frequency deviation values ​​of the first and last sampling points within the segment interval, divided by the number of sampling points in that segment interval, reflecting the overall trend of drift along the frequency direction within that interval. The mean deviation feature and the deviation slope feature characterize the drift state within the segment interval from two orthogonal dimensions: the mean deviation feature represents the "magnitude" of the drift, and the deviation slope feature represents the "direction" of the drift. The combined use of these two features allows for the differentiation of drift conditions with similar average deviations but different trends when selecting the Volterra inverse model kernel parameter set, thereby improving the targeting of model selection.

[0058] refer to Figure 2 , Figure 2The input on the left is the transmitted digital baseband signal segment to be compensated. This signal segment corresponds to the segment of digital baseband signal covered by a certain segment interval determined in step 2 on the transmitted frequency coordinates. After the transmitted digital baseband signal segment enters the filter, it is simultaneously distributed to three parallel processing branches at the branch point. The first branch is a first-order linear convolution branch, which performs a linear discrete convolution operation on the transmitted digital baseband signal segment using a first-order linear kernel coefficient vector read from the drift feature classification table. This branch functions equivalently to a finite impulse response filter, capturing the linear transfer component in the closed-loop equivalent nonlinear system and outputting a first-order output sequence. The second branch is a second-order Volterra convolution branch, which performs a second-order discrete Volterra convolution operation on the transmitted digital baseband signal segment using a second-order pruned nonlinear kernel coefficient matrix. This branch takes two sample values ​​at different time delay positions within a finite memory depth range for each sampling point in the input signal, multiplies them, and weights them with the corresponding second-order kernel coefficients to capture the second-order nonlinear effect in the closed-loop equivalent nonlinear system and outputs a second-order output sequence. The third branch is a 3rd-order Volterra convolution branch. It uses a 3rd-order pruned nonlinear kernel coefficient tensor to perform a 3rd-order discrete Volterra convolution operation on the transmitted digital baseband signal segment. This branch multiplies the sampled values ​​at three different time delay positions within the memory depth range and weights them with the corresponding 3rd-order kernel coefficients to capture 3rd-order nonlinear effects, outputting a 3rd-order output sequence. The outputs of the three branches are... Figure 2 The right side first converges to the point-by-point summation node, where the 1st, 2nd, and 3rd order output sequences are added point-by-point according to the sampling points, resulting in the total output sequence of the Volterra inverse filter. The total output sequence then enters the point-by-point subtraction node, where it is subtracted point-by-point from the original transmitted digital baseband signal segment derived from the branch point to obtain the compensation segment. Figure 2 The original signal is led downwards from the branch point and then loops to the bottom input of the point-by-point subtraction node, distinguishing it from the total output sequence fed forward from the summation node. The purpose of point-by-point subtraction is to strip the linear pass-through components belonging to the original signal from the total output of the Volterra inverse filter, retaining only the nonlinear correction increment applied by the inverse model as the compensation amount. The compensation amount segment is led out from the output of the point-by-point subtraction node as the final compensation result for that segment interval. Subsequently, it is spliced ​​with the compensation amount segments of adjacent segment intervals through a linear gradient transition and then superimposed onto the transmitted digital baseband signal. The three branches adopt a parallel structure rather than a cascaded structure, making the calculation of the 1st, 2nd, and 3rd order components independent of each other. They can be executed synchronously in the hardware implementation to reduce processing latency, which is particularly important for real-time vehicle applications that require rapid compensation calculation within the calibration time slot.

[0059] After completing step 2, each segment interval is assigned its own drift feature vector. Step 3 uses these drift feature vectors to select the Volterra inverse model kernel parameter set that best matches the current drift state from the pre-established drift feature classification table, and applies predistortion compensation to the transmitted digital baseband signal so that the instantaneous frequency deviation of the loop hysteresis generated by the compensated transmitted chirp after passing through the transmit link approaches zero.

[0060] Before proceeding with the specific compensation process, it is necessary to define the compensation object. The complete signal link from the input of the transmitted digital baseband signal to the output of the instantaneous frequency deviation of the loop hysteresis signal is defined as a closed-loop equivalent nonlinear system. This link sequentially includes a digital-to-analog converter, an up-converter, a power amplifier, a directional coupler, a known dielectric delay line, a receiver channel mixer, a low-pass anti-aliasing filter, an analog-to-digital converter, and the Hilbert transform and instantaneous frequency extraction operation in step 2. The reason for considering such a long link as a whole as an equivalent nonlinear system rather than compensating only a single device in the transmit link is that the instantaneous frequency deviation observed in steps 1 and 2 is the combined result of the combined effects of all the above links. If only an inverse model is established for the transmit link while ignoring the nonlinear contributions introduced by the receiver link and digital processing, the compensation object will be inconsistent with the observed object, and the compensation accuracy cannot be guaranteed. After defining the entire link as a closed-loop equivalent nonlinear system, the object identified by the Volterra inverse model is completely aligned with the source of deviation measured in step 2. The predistortion compensation sequence is applied to the input of the transmit digital baseband, and its effect is observed and verified at the output of the instantaneous frequency deviation after propagating through the entire closed-loop link, forming a complete closed loop of "compensation applied at the input and effect observed at the output".

[0061] refer to Figure 3 , Figure 3 The horizontal axis represents frequency, in kHz, covering a symmetrical range of positive and negative frequencies. The vertical axis represents the amplitude spectrum, expressed in normalized form. Figure 3 It contains two spectral curves. The first curve is the digitally sampled sequence of the loopback beat signal. The original two-sided spectrum obtained after the discrete Fourier transform of the zero frequency exhibits a conjugate symmetric distribution, with equal amplitude spectral peaks appearing at symmetrical positions on both the positive and negative frequency axes. This symmetry is an inherent property of the spectrum of real-valued signals. The second curve is the spectrum after single-sided spectrum construction. The single-sided spectrum construction operation involves taking the second to the third frequency complex number sequence in the frequency domain... Multiply the complex values ​​of each frequency point by 2, and then... Add 1 to the number The complex values ​​of each frequency point are set to 0, and the first frequency point and the second frequency point are... Adding one frequency point keeps the complex value unchanged. After the above processing, the spectral components in the negative frequency region are completely eliminated, and the peak amplitude of the spectral components in the positive frequency region is correspondingly increased to twice its original value. Figure 3 The arrows and labels indicate the effects of multiplying the positive frequency region by 2 and setting the negative frequency region to zero, respectively. Multiplying the positive frequency component by 2 compensates for the half-energy loss due to the removal of the negative frequency component, ensuring that the amplitude of the analytic signal sequence obtained by performing an inverse discrete Fourier transform on the processed frequency domain sequence remains consistent with the original signal, thus avoiding a systematic underestimation of the instantaneous amplitude in subsequent calculations. The analytic signal sequence obtained after single-sided spectrum construction and inverse transform is a complex-valued sequence. Its real part is equal to the original loop hysteresis signal, and its imaginary part is the Hilbert transform result of the original signal. These two parts form an orthogonal component pair, allowing the instantaneous phase value to be directly extracted at each sampling point using a four-quadrant arctangent operation, and then the instantaneous frequency value to be obtained through point-by-point difference. Single-sided spectrum construction is a crucial intermediate step in obtaining the analytic signal from the real-valued time-domain signal. Its frequency-domain operation is equivalent to applying a Hilbert transform to the original signal in the time domain and combining them into a complex signal. However, it can be efficiently performed in the frequency domain using a fast Fourier transform algorithm, making it suitable for real-time execution on automotive embedded platforms.

[0062] The drift feature classification table is pre-stored in the radar's on-chip non-volatile memory and is created offline during the factory calibration phase. The table divides the combined range of the mean deviation feature and the slope deviation feature into multiple non-overlapping and continuously covering drift feature regions. In one typical implementation, the mean deviation feature range is divided into eight equally spaced intervals from 0 to its estimated maximum value, and the slope deviation feature range is divided into four equally spaced intervals from its estimated minimum to its maximum value. These intervals are combined to form 32 drift feature regions. Each drift feature region corresponds to one set of 3rd-order Volterra inverse model kernel parameters with finite memory depth. In another optional implementation, the number of partitions for the mean deviation feature and the slope deviation feature can be non-uniformly divided according to the distribution density of the measured drift data. Denseer partitions are used in regions where drift states occur frequently to improve matching accuracy, while coarser partitions are used in regions where drift states occur infrequently to save storage space.

[0063] The structure of the kernel parameter set for each 3rd-order Volterra inverse model with finite memory depth is as follows. A Volterra series is a polynomial expansion describing the input-output relationship of a nonlinear system, similar to a nonlinear generalization of convolution in a linear system. The 3rd-order Volterra model represents the system's output as the sum of the 1st (linear), 2nd, and 3rd-order contributions of the input signal. Each contribution is obtained by convolving the corresponding order Volterra kernel with the corresponding order product of the input signal. The 1st-order linear kernel coefficient vector describes the system's linear transfer characteristics, equivalent to the coefficients of a traditional finite impulse response filter; the 2nd-order nonlinear kernel coefficient matrix describes the system's 2nd-order nonlinear effects, such as second harmonics and intermodulation distortion; the 3rd-order nonlinear kernel coefficient tensor describes the system's 3rd-order nonlinear effects, such as third harmonics and third-order intermodulation products.

[0064] The meaning of finite memory depth is that the support range of the Volterra core is not infinitely extended but truncated to a finite time window. Memory depth This represents the number of coefficients in each dimension of the Volterra kernel, in a typical implementation. Taking 8, the first-order linear kernel coefficient vector contains 8 coefficients. Without any simplification of the second- and third-order kernels, the size of the second-order kernel coefficient matrix is... That is, 64 coefficients, the size of the 3rd order kernel coefficient tensor is The total number of parameters, i.e., 512 coefficients, increases exponentially with the order, placing a heavy burden on the storage and computing resources of the automotive embedded processing platform. To address this issue, pruning is performed on the 2nd and 3rd order kernels. The 2nd-order pruned nonlinear kernel coefficient matrix retains only elements within the order range of its main diagonal and its two adjacent preset neighborhoods. In a typical implementation, the preset neighborhood order range is set to 2, meaning that only elements in the 2nd-order kernel coefficient matrix that satisfy the condition that the absolute value of the difference between the delay indices of the two dimensions does not exceed 2 are retained, and the remaining elements are set to 0. After pruning, the effective number of parameters in the 2nd-order kernel is reduced from 64 to approximately 40. The 3rd-order pruned nonlinear kernel coefficient tensor retains only elements within the order range of its main diagonal and its two adjacent preset neighborhoods. Similarly, only elements whose absolute values ​​of the pairwise differences between the delay indices of the three dimensions do not exceed the preset neighborhood order range are retained. After pruning, the effective number of parameters in the 3rd-order kernel is reduced from 512 to approximately 120. The physical basis of the pruning strategy is that the nonlinear memory effect of actual radio frequency devices is mainly concentrated between neighboring samples with small time delay differences. The kernel coefficients far from the main diagonal correspond to cross-coupling terms with large time delay differences, and their physical contribution can usually be ignored.

[0065] The acquisition of the kernel parameter sets of the 3rd-order Volterra inverse model with finite memory depth for each group is completed during the factory calibration stage. The specific process is as follows: The radar front-end packaging module is installed in a programmable temperature chamber. For each drift feature region in the drift feature classification table, the chamber temperature is adjusted to a temperature value that can generate the range of the mean deviation feature and the deviation slope feature represented by that drift feature region. After the chamber temperature stabilizes, a known broadband excitation signal is injected into the transmit digital baseband port. This broadband excitation signal typically uses a pseudo-random binary sequence or a multi-sine composite signal, whose power spectral density is approximately flat across the entire chirped baseband bandwidth to ensure that the excitation can fully cover the nonlinear response characteristics of the closed-loop equivalent nonlinear system at all operating frequencies. After entering from the transmit digital baseband port, the broadband excitation signal passes through the entire closed-loop equivalent nonlinear system, and finally, at the output end, the loop hysteresis signal is acquired through the FMCW chirped loop homing calibration channel, and the instantaneous frequency deviation is extracted according to the method in step 2, serving as the output response of the closed-loop equivalent nonlinear system under this excitation.

[0066] Using the input signal (i.e., the broadband excitation signal) from the transmitted digital baseband port and the output (i.e., the instantaneous frequency deviation) of the closed-loop equivalent nonlinear system, a third-order Volterra forward kernel identification with finite memory depth is performed. The identification method employs a least-squares algorithm: the input signal is constructed into a regression matrix containing first-, second-, and third-order components according to the rules of Volterra expansion; the output response is used as the observation vector; and the Volterra forward kernel coefficients that minimize the energy of the output prediction error are obtained through least-squares solving. The identified third-order Volterra forward kernel describes the forward transfer characteristics of the closed-loop equivalent nonlinear system.

[0067] After obtaining the forward kernel, its inverse kernel needs to be solved to construct the predistortion compensation model. The basic idea for solving the 3rd-order inverse kernel is to find a set of Volterra kernels such that after processing the original input signal through the inverse kernel and then through the forward system, the output is as close as possible to the ideal linear transfer result. The inverse kernel solution usually adopts an indirect learning architecture: the output of the forward system is used as the input of the inverse model, and the original input signal is used as the expected output of the inverse model. The Volterra kernel coefficients of the inverse model are identified by the least squares method. After the solution is completed, the above pruning process is performed on the 2nd-order and 3rd-order inverse kernels to obtain a set of 3rd-order Volterra inverse model kernel parameters with finite memory depth, which are stored in the corresponding drift feature region entry in the drift feature classification table.

[0068] When the radar is running online, the real-time compensation process in step 3 is as follows: For each segmented interval determined in step 2, the drift feature vector of that segmented interval is compared one by one with each drift feature region in the drift feature classification table. The comparison method is as follows: calculate the Euclidean distance between the drift feature vector and the center point of each drift feature region, select the drift feature region with the smallest Euclidean distance, and read the 3rd order Volterra inverse model kernel parameter set with the finite memory depth stored for that drift feature region. In another optional implementation, the drift feature region can also be determined by directly judging which partition the mean deviation feature and the slope deviation feature of the drift feature vector fall into, without calculating the Euclidean distance.

[0069] The transmitted digital baseband signal segment corresponding to the transmission frequency coordinates of the segmented interval is extracted as the input signal segment to be compensated. A third-order Volterra inverse filter is constructed using the kernel parameter set of the third-order Volterra inverse model with finite memory depth. The third-order Volterra inverse filter is configured with three parallel processing branches. The first branch is a first-order linear convolution branch, which performs linear discrete convolution operation on the input signal segment to be compensated using the first-order linear kernel coefficient vector. Its operation is exactly the same as the convolution operation of a conventional finite impulse response filter, performing linear discrete convolution operation on each sampling point and its preceding points in the input signal segment. The first branch is a second-order Volterra convolution branch. It uses a second-order pruned nonlinear kernel coefficient matrix to perform a second-order discrete Volterra convolution operation on the input signal segment to be compensated: for each sampling point in the input signal segment, it multiplies the sample values ​​at two different time delay positions within the memory depth range, then multiplies them by the coefficient at the corresponding position in the second-order pruned nonlinear kernel coefficient matrix, summing all valid time delay index combinations to output a second-order output sequence. The second branch is a third-order Volterra convolution branch. It uses a third-order pruned nonlinear kernel coefficient tensor to perform a third-order discrete Volterra convolution operation on the input signal segment to be compensated: for each sampling point in the input signal segment, it multiplies the sample values ​​at three different time delay positions within the memory depth range, then multiplies them by the coefficient at the corresponding position in the third-order pruned nonlinear kernel coefficient tensor, summing all valid time delay index combinations to output a third-order output sequence.

[0070] The first-order, second-order, and third-order output sequences are summed point-by-point to obtain the total output sequence of the Volterra inverse filter. This total output sequence represents the predistorted transmit segment after processing by the inverse model, not the compensation amount itself. The compensation segment is equal to the point-by-point subtraction of the total output sequence of the Volterra inverse filter and the input signal segment to be compensated; that is, the value of each sample point in the compensation segment is equal to the output value of the Volterra inverse filter at that sample point minus the value of the original transmitted digital baseband signal at that sample point. The compensation segment reflects the correction increment that the inverse model believes needs to be applied.

[0071] After obtaining the compensation segments for each segmented interval according to the above process, they need to be spliced ​​together to form a compensation sequence covering the complete FMCW chirp period. Since the Volterra inverse model kernel parameter sets selected for adjacent segmented intervals are usually different, numerical jumps may exist between the compensation segments on both sides of the segment boundary. If directly hard-splitting is used, this jump will introduce artificial discontinuities into the compensated transmitted signal, generating additional spectral spurious signals. To eliminate splicing jumps, a linear gradient transition zone with a preset number of transition sampling points is set at the splicing boundary of the compensation segments between two adjacent segmented intervals. In a typical implementation, the preset number of transition sampling points is 10. Within the linear gradient transition zone, for each transition sampling point, let the position number of that sampling point within the transition zone be... (Incrementing from 0 to the preset number of transition sampling points minus 1), calculate the mixing ratio. ,in To preset the number of transition sampling points, The value increases linearly from 0 to 1. The compensation value at this transition sampling point is equal to the value of the compensation segment at the corresponding position in the previous segment interval multiplied by . Add the compensation value of the next segment interval at the corresponding position multiplied by .when When the value is 0, the output of the previous segment interval is used completely. When the value equals 1, the output of the last segment interval is fully utilized, with a smooth transition in the middle position using a linear ratio. This linear interpolation mixing strategy ensures the continuity of the compensation sequence at the segment boundaries, avoiding spectral spurious signals introduced by jumps. In another optional implementation, a transition curve shaped like a cosine gradient window or a Hanning window can be used instead of a linear gradient. The rate of change of the cosine gradient is zero at both ends of the transition region, which can simultaneously ensure the continuity of the first derivative of the compensation sequence at the splicing boundary, further reducing the spurious signal level.

[0072] After performing linear gradient transition processing on the splicing boundaries of all adjacent segmented intervals, the compensation segments of all segmented intervals and the interpolated compensation values ​​of the transition zone are spliced ​​together to form a compensation sequence covering the complete FMCW chirp period. This compensation sequence is then added point-by-point to the transmitted digital baseband signal—that is, for each sampling point in the transmitted digital baseband signal, the compensation value at the same position in the compensation sequence is added—to obtain the predistortion-corrected transmitted digital baseband signal. The predistortion-corrected transmitted digital baseband signal is converted into an analog signal by a digital-to-analog converter and output to the upconverter and power amplifier of the transmit channel, ultimately forming an FMCW chirped transmit signal compensated for nonlinear drift. After the compensated transmit signal passes through the transmit link and loopback channel, the instantaneous frequency deviation generated in the next calibration time slot will be significantly reduced, thereby completing the online self-calibration and compensation of chirp nonlinear drift of the vehicle-mounted small radar.

[0073] In an optional advanced implementation, an iterative convergence mechanism can be introduced between multiple consecutive calibration time slots: after the compensation is completed in the first calibration time slot, the loop hysteresis signal is reacquired in the second calibration time slot and the instantaneous frequency deviation sequence is extracted. If the deviation still exceeds the preset convergence threshold, steps 2 and 3 are executed again based on the residual deviation to generate an incremental compensation amount, which is then superimposed on the existing compensation amount. This process is iterated until the deviation converges below the preset convergence threshold or the maximum number of iterations is reached. This iterative method can further compensate for the lookup matching error caused by the discretization granularity limitation of the drift feature classification table.

[0074] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for online self-calibration and drift compensation of a vehicle-mounted small radar, characterized in that, The method includes the following steps: Step 1: In the front-end packaging module of the vehicle-mounted small radar, the output of the power amplifier of the transmitting channel is connected to the calibration input port of the mixer of the receiving channel via a directional coupler and a known dielectric delay line to form an FMCW chirped loopback calibration channel; during the calibration time slot, the FMCW chirped transmitting signal is coupled by the directional coupler and loops back to the mixer of the receiving channel via the known dielectric delay line for down-conversion, outputting the loopback beat signal and obtaining the digital sampling sequence of the loopback beat signal through analog-to-digital conversion; Step 2: Perform Hilbert transform on the digital sampling sequence of the loop hysteresis signal to obtain the analytical signal sequence. Extract the instantaneous frequency sequence from the analytical signal sequence. Subtract the instantaneous frequency sequence from the ideal instantaneous frequency reference sequence recorded during factory calibration point by point to obtain the instantaneous frequency deviation sequence. Perform adaptive segmentation on the instantaneous frequency deviation sequence and extract the drift feature vector of each segment. Step 3: Define the signal link from the input of the transmitted digital baseband signal to the output of the instantaneous frequency deviation as a closed-loop equivalent nonlinear system. Select the corresponding Volterra inverse model kernel parameter set from the drift feature classification table according to the drift feature vector of each segment interval. Perform Volterra inverse filtering operation on the transmitted digital baseband signal segment corresponding to each segment interval to obtain the compensation segment. Concatenate all the compensation segments into a compensation sequence, add it point by point with the transmitted digital baseband signal, and output it to the transmission channel through the digital-to-analog converter.

2. The method according to claim 1, characterized in that, In step 1, the directional coupler is a Lange microstrip directional coupler; it is known that the dielectric delay line adopts a microstrip transmission line structure embedded in a low-temperature co-fired ceramic substrate, and has a fixed propagation delay value pre-calibrated at the factory.

3. The method according to claim 1, characterized in that, In step 1, a platinum thin-film resistance temperature sensor is set at both the input and output ends of the known dielectric delay line to collect the temperature values ​​at both ends of the known dielectric delay line in real time and take the arithmetic mean as the delay line temperature measurement value; the temperature-corrected loop propagation delay value is obtained by looking up the temperature-delay lookup table pre-calibrated at the factory based on the delay line temperature measurement value.

4. The method according to claim 1, characterized in that, In step 1, a calibration time slot is inserted between the normal operating time slots of the transmit channel and the receive channel. During the calibration time slot, the mixer input of the receive channel is switched from the antenna port to the calibration input port using an RF single-pole double-throw switch.

5. The method according to claim 1, characterized in that, In step 2, the method of performing Hilbert transform on the digital sampling sequence of the loop hysteresis signal to obtain the analytical signal sequence is as follows: perform N-point discrete Fourier transform on the digital sampling sequence of the loop hysteresis signal to obtain a frequency domain complex sequence of length N; perform single-sided spectrum construction processing on the frequency domain complex sequence, multiply the complex values ​​of the 2nd to N / 2nd frequency points by 2 respectively, set the complex values ​​of the N / 2+1 to Nth frequency points to 0, and keep the complex values ​​of the 1st frequency point and the N / 2+1th frequency point unchanged; An analytic signal sequence is obtained by performing an N-point discrete Fourier inverse transform on the frequency domain complex sequence after single-sided spectrum construction.

6. The method according to claim 1, characterized in that, In step 2, the instantaneous frequency sequence is extracted from the analytical signal sequence by performing a four-quadrant arctangent operation on the complex value of each sampling point in the analytical signal sequence with its real and imaginary parts as inputs to obtain the wrapped phase value. Along the direction of the sampling point number, when the absolute value of the difference between the wrapped phase values ​​of two adjacent sampling points exceeds pi, the phase values ​​of subsequent sampling points are accumulated by adding or subtracting one complete circumference to obtain a continuously unfolded instantaneous phase sequence; the difference between the instantaneous phase values ​​of two adjacent sampling points in the instantaneous phase sequence is calculated and divided by the sampling time interval to obtain the instantaneous frequency value, thus forming an instantaneous frequency sequence.

7. The method according to claim 1, characterized in that, In step 2, after extracting the instantaneous frequency sequence and before subtracting it point by point from the ideal instantaneous frequency reference sequence, the instantaneous frequency sequence is de-squeezed and smoothed: sampling points in the analytical signal sequence with amplitudes lower than a preset amplitude threshold are marked, and the instantaneous frequency values ​​of the marked sampling points are replaced with the arithmetic mean of the instantaneous frequency values ​​of the unmarked sampling points within each preset number of neighboring sampling points; the replaced instantaneous frequency sequence is then subjected to a sliding median filter with a preset window length to obtain a smoothed instantaneous frequency sequence.

8. The method according to claim 3, characterized in that, In step 2, before subtracting the instantaneous frequency sequence from the ideal instantaneous frequency reference sequence point by point, the time axis coordinates of the instantaneous frequency sequence are converted to the transmit frequency coordinates point by point using the temperature-corrected loopback propagation delay and the FMCW chirped frequency modulation slope. The adaptive segmentation interval is divided as follows: an entry threshold and an exit threshold are set, with the exit threshold being less than the entry threshold. The absolute value of the difference between the instantaneous frequency deviation values ​​of adjacent sampling points in the instantaneous frequency deviation sequence is calculated point by point as the point-by-point deviation change. When the point-by-point deviation change rises from below the entry threshold to above the entry threshold for the first time... When the candidate segment boundary state is entered, the number of sampling points whose point-by-point deviation changes remain above the exit threshold is continuously counted. When the preset number of consecutive judgment points is reached, the position of the first sampling point that exceeds the limit is marked as the segment boundary point. When the point-by-point deviation changes fall below the exit threshold, the candidate segment boundary state is exited. All segment boundary points, together with the start and end points of the transmission frequency coordinates, define the entire segment interval. Segment intervals with fewer than the preset minimum segment length are merged with adjacent segment intervals with fewer than the preset minimum segment length into one segment interval. The drift feature vector consists of the mean deviation feature and the slope deviation feature. The mean deviation feature is the arithmetic mean of the absolute values ​​of the instantaneous frequency deviations of all sampling points within the segmented interval. The slope deviation feature is the value obtained by dividing the difference in instantaneous frequency deviations of the first and last sampling points within the segmented interval by the number of sampling points in the segmented interval.

9. The method according to claim 1, characterized in that, In step 3, the Volterra inverse model kernel parameter set is a 3rd-order Volterra inverse model kernel parameter set with finite memory depth, which includes a 1st-order linear kernel coefficient vector, a 2nd-order pruned nonlinear kernel coefficient matrix, and a 3rd-order pruned nonlinear kernel coefficient tensor. The 2nd-order pruned nonlinear kernel coefficient matrix retains only the elements within the order range of the main diagonal and its two sides, and the 3rd-order pruned nonlinear kernel coefficient tensor retains only the elements within the order range of the main diagonal and its two sides. The drift feature classification table divides the joint value range of the mean deviation feature and the slope deviation feature into multiple non-overlapping and continuously covered drift feature regions. Each drift feature region corresponds to storing one set of finite-memory-depth 3rd-order Volterra inverse model kernel parameter sets. During the factory calibration stage, the temperature of the front-end packaging module is adjusted by the temperature control platform, a broadband excitation signal is injected into the transmit digital baseband port, and the output of the closed-loop equivalent nonlinear system is acquired through the FMCW chirped loopback calibration channel. After performing finite-memory-depth 3rd-order Volterra forward kernel identification, the 3rd-order inverse kernel is solved, and the 2nd-order and 3rd-order inverse kernels are pruned. The Volterra inverse filtering operation is performed as follows: using the finite-memory-depth 3rd-order... A third-order Volterra inverse filter is constructed using the Volterra inverse model kernel parameter set. This third-order Volterra inverse filter has three parallel processing branches. The first branch performs a linear discrete convolution operation on the transmitted digital baseband signal segment using a first-order linear kernel coefficient vector to obtain a first-order output sequence. The second branch performs a second-order discrete Volterra convolution operation on the transmitted digital baseband signal segment using a second-order pruned nonlinear kernel coefficient matrix to obtain a second-order output sequence. The third branch performs a third-order discrete Volterra convolution operation on the transmitted digital baseband signal segment using a third-order pruned nonlinear kernel coefficient tensor to obtain a third-order output sequence. The first-order, second-order, and third-order output sequences are summed point-by-point, and then subtracted point-by-point from the transmitted digital baseband signal segment to obtain the compensation segment.

10. The method according to claim 1, characterized in that, In step 3, when splicing all compensation amount segments into a compensation amount sequence, a linear gradient transition zone with a preset number of transition sampling points is set at the splicing boundary of compensation amount segments between two adjacent segment intervals. Within the linear gradient transition zone, for each transition sampling point, the value of the compensation amount segment of the previous segment interval at the corresponding sampling point is linearly interpolated and mixed with the value of the compensation amount segment of the next segment interval at the corresponding sampling point. The interpolation ratio linearly transitions along the linear gradient transition zone from completely taking the output of the previous segment interval to completely taking the output of the next segment interval.

Images