Large dynamic back demodulation method and device based on tdma system

Abstract

The application discloses a large dynamic reverse demodulation method and device based on a TDMA system, which comprises the following steps: maintaining a terminal image lookup table for storing statistical quantities of receiving power of each terminal; before a target time slot of a terminal arrives, predicting the receiving power distribution of the terminal based on the image, constructing a risk target function balancing overload risk and quantization loss, and determining the optimal pre-cut gain by optimizing the function; in a login time slot, using a coded power reference pilot, further estimating the absolute gain deviation and the nonlinear coefficient of the receiving link on the basis of the pre-cut gain, calculating the quadratic scaling gain and the linearization compensation parameter, and applying the quadratic scaling gain and the linearization compensation for demodulation in subsequent service time slots. After demodulation, the terminal image is updated by using the measured power closed loop. Through the predictive risk minimization gain decision and the nonlinear online compensation, the application realizes accurate and adaptive gain control, and improves the receiving performance and dynamic range of the system.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of satellite communication, and particularly relates to a large dynamic reverse demodulation method and device based on a TDMA system. BACKGROUND

[0002] In a modern communication system, a satellite communication system based on time division multiple access (TDMA) technology has important applications in emergency communication, ocean transportation and other fields due to its wide coverage and high reliability. In such a system, a central master station needs to simultaneously receive and process burst signals from a large number of remote stations with different geographical locations and different device capabilities. In the face of a large dynamic range scenario where the input signal power span is large, the master station needs to stably and efficiently complete demodulation to improve the communication quality and capacity of the entire system.

[0003] Currently, to cope with the problem of large dynamic range reception, the existing technology mainly adopts the following methods: setting a compromised fixed reception gain, which is optimized statically and can limitedly take into account the reception of strong signals and weak signals; and using a traditional automatic gain control (AGC) loop, which dynamically adjusts the gain within a time slot through a fast feedback mechanism after receiving the preamble part of the burst signal, so as to stabilize the signal amplitude within the ideal working area of an analog-to-digital converter (ADC). The above methods alleviate the challenges brought by the signal dynamic range to a certain extent and have been applied in many systems.

[0004] However, the existing technology still faces the problems of staticity and reactivity of its control strategy when coping with high-speed and high-dynamic TDMA systems, which jointly restrict the further improvement of the demodulation performance. Therefore, further research and innovation are needed to solve the above problems existing in the existing technology. SUMMARY

[0005] The application provides a large dynamic reverse demodulation method and device based on a TDMA system.

[0006] Technical solution: In a first aspect, a large dynamic reverse demodulation method based on a TDMA system is provided, comprising:

[0007] maintaining a terminal image lookup table storing a statistical quantity of reception power of at least one terminal;

[0008] reading the statistical quantity of reception power in the terminal image lookup table before the target time slot of the terminal arrives, and predicting the reception power distribution of the target time slot based on the statistical quantity of reception power;

[0009] constructing a risk objective function balancing overload risk and quantization loss according to the reception power distribution;

[0010] determining a pre-cut gain by optimizing the risk objective function;

[0011] In the target time slot, the received signal is demodulated using the receive gain determined based on the pre-cut gain, and the measured received power is estimated.

[0012] Update the received power statistics in the terminal profile lookup table using the measured received power.

[0013] In a second aspect, a reverse demodulation device is provided, comprising: a memory for storing computer program instructions and a terminal profile lookup table;

[0014] A processor configured to execute computer program instructions for use in some implementations of the first aspect described above.

[0015] Beneficial effects: This invention solves the problems of staticity and responsiveness of control strategies through predictive risk-minimizing gain decision-making and nonlinear online compensation, thereby improving the system's receiving performance and dynamic range. The related technical effects will be described in detail below with reference to specific embodiments. Attached Figure Description

[0016] Figure 1 A flowchart of a large dynamic reverse demodulation method based on a TDMA system provided in this application embodiment.

[0017] Figure 2 This is a flowchart illustrating an example of constructing a risk objective function, provided as an embodiment of this application.

[0018] Figure 3 This is another flowchart illustrating the construction of a risk objective function, provided as an embodiment of this application.

[0019] Figure 4 A flowchart illustrating the determination of an overload probability term provided in an embodiment of this application.

[0020] Figure 5 A flowchart illustrating the determination of another overload probability term provided in this application embodiment. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] It is to be understood that the terms first, second, etc. used in the description and the claims of the present application are used to differentiate similar objects, and are not necessarily used to describe a particular sequential or chronological order. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms include and have and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily limit to the clearly listed steps or units, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0023] To solve the above problems, the applicant has conducted in-depth search and analysis, and found that:

[0024] The existing method adopts static gain or simple AGC strategy, lacks adaptation to the channel characteristics of each terminal, and causes the gain decision to be unable to balance between overload clipping and quantization noise;

[0025] Further, the reactive time slot gain adjustment mechanism consumes the overhead of the preamble, is used for gain stabilization, and also has the inherent risk of gain adjustment lag and data pollution caused by signal power mutation;

[0026] On this basis, such methods ignore the nonlinear compression effect of the receiver front end under high power input, that is, even if hard clipping is successfully avoided, nonlinear distortion cannot eliminate the compression and distortion of signal constellation points, which limits the performance of high-order modulation.

[0027] In order to solve these problems, in combination with Figures 1 to 5 The application is specifically illustrated by the following embodiments.

[0028] Embodiment 1 provides a large dynamic reverse demodulation method based on a TDMA system. The method can be realized by a reverse demodulation device, which includes but is not limited to a processor, a memory, and a radio frequency receiving link including a variable gain amplifier (VGA) and an analog-to-digital converter (ADC). The memory is used to store computer program instructions and a terminal portrait lookup table; the processor is configured to execute the computer program instructions to realize the method of the application.

[0029] In the context of the present application, in order to keep the description consistent, the technical terms and data items in some embodiments are defined:

[0030] The coded power reference pilot CRP-UW: the pilot introducing the amplitude step ratio set in the preamble of the login time slot, used for absolute power scaling and front-end nonlinear parameter estimation, does not carry service information.

[0031] Terminal profile lookup table (LUT): A lookup table maintained by terminal identity and time slot index, recording and predicting the terminal's received power statistics, receive gain settings, nonlinearity parameters, etc.

[0032] Risk-minimized pre-cut control R2AGC: A target function that minimizes the risk of overload + quantization loss based on received power statistics and hardware boundaries to obtain the optimal pre-cut gain before the time slot arrives.

[0033] Variable gain amplifier VGA, analog-to-digital converter ADC, peak-to-average power ratio PAPR: Used according to their conventional definitions.

[0034] Received power P: Refers to the average power entering the input end of the variable gain amplifier VGA, with units in linear power scale. Its statistics are mean μ and standard deviation σ.

[0035] Pre-cut gain G _1 , secondary scaling gain G _2 , safety gain G _safe : Refers to the scalar (linear multiple) of the equivalent voltage gain of the VGA.

[0036] ADC full scale A _max : Refers to the peak input amplitude allowed by the analog-to-digital converter.

[0037] Peak-to-average power ratio PAPR _max : Refers to the upper bound of the ratio of peak power to average power within the same burst.

[0038] CRP-UW parameter set: Contains amplitude step ratio set, amplitude step ratio range, step number K, number of symbols per step, phase skeleton sequence, nominal amplitude set {a _k}.

[0039] Linearization compensation parameter Lin _coef : Digital pre-linearization or VGA setting correction parameter generated according to α and β.

[0040] Fallback command Fallback_cmd: Control command that triggers reconfiguration with safety gain.

[0041] Pre-cut timing t _pre : Gain switching advance time set to ensure VGA stability and protection interval.

[0042] VGA step ΔG: Discrete gain step supported by VGA hardware.

[0043] Hardware boundary parameter set: Includes but is not limited to ADC full scale A _max , PAPR upper limit PAPR _max , VGA step ΔG, allowed processing delay budget, ADC bit number N_bit full-scale voltage V _fs sampling rate, baseband normalization constant, VGA settling time t _settle guard interval t _guard .

[0044] Profile: a multi-dimensional feature model established for a specific object (such as a user, a device, a terminal, a risk, etc.) by collecting and analyzing data; for example, a terminal profile and a risk profile, which represent a multi-dimensional data description of a specific object (here, a terminal and a risk, respectively).

[0045] In some embodiments, the application is applied to a time division multiple access (TDMA) communication system, especially in a satellite communication scenario. In a TDMA satellite communication network, a master station needs to receive burst uplink signals from multiple ground sub-stations (also referred to as terminals). Due to differences in their own transmission capabilities (such as antenna size, power amplifier power), the positions of the satellite beams they are in (beam center or edge), and the instantaneous channel conditions (such as weather changes), the power of the signals from the sub-stations to the master station receiver exhibits a large dynamic range.

[0046] In the prior art, the master station receiver usually adopts a fixed reception gain, which is set as a compromise for signals of different powers. However, the fixed gain strategy has inherent problems; when a strong signal from a high-power terminal is received, the fixed gain may be too high, causing the front end of the reception chain (such as a low-noise amplifier LNA or a variable gain amplifier VGA) to saturate, or the input of an analog-to-digital converter (ADC) to exceed its full scale, resulting in clipping distortion, which can destroy the constellation structure of the signal and cause demodulation performance to deteriorate. Conversely, when a weak signal from a low-power terminal is received, the fixed gain may be too low, causing the amplitude of the signal after ADC sampling to be too small, and the number of quantization bits to be insufficient, resulting in a high proportion of quantization noise in the signal, which also reduces the demodulation signal-to-noise ratio and affects communication quality.

[0047] To solve the problem of demodulation performance degradation due to a large dynamic range of received signals, the application provides an adaptive reverse demodulation method. It discards the fixed gain strategy and performs individualized and predictive gain control for each terminal in each time slot allocated to it. The profile of each terminal is established and maintained, the power of the signal is predicted before it arrives, and the optimal reception gain is dynamically decided based on a risk model that balances overload and quantization loss, so that the amplitude of the received signal is always within the adaptive working interval of the ADC, and high-performance demodulation is achieved under a large dynamic range.

[0048] In some embodiments, the method comprises the following steps:

[0049] Step S1.1, maintain a terminal profile lookup table storing received power statistics of at least one terminal.

[0050] In this embodiment, the terminal profile lookup table (LUT) is a data structure stored in the device memory, which establishes a profile for each terminal in the network. The lookup table can take the unique identifier of the terminal (such as terminal ID) and the time slot index it occupies as the primary key. In each entry, the core information is the received power statistics, which preferably includes the statistical mean μ and statistical standard deviation σ of the average power of the terminal's historical signal arriving at the receiver VGA input. Such statistics are obtained by long-term smoothing calculation of the measured power after each successful demodulation of the terminal, reflecting the long-term stability and short-term fluctuation range of the terminal's signal power.

[0051] Step S1.2, read the received power statistics in the terminal profile lookup table before the target time slot of the terminal arrives, and predict the received power distribution of the target time slot based on it.

[0052] Wherein, the target time slot refers to the time slot allocated to the specific terminal for uplink transmission according to the time slot allocation plan of the TDMA system. Within a short time window (for example, a few hundred microseconds to a few milliseconds) before the actual arrival of the time slot, the processor accesses the terminal profile lookup table to extract the power statistics μ and σ corresponding to the terminal. Next, the processor calls the prediction model to generate a predicted received power distribution P _pred of the target time slot based on the historical statistics. Optionally, the prediction model can be an exponential weighting model, giving higher weight to recent observations; or, in a more preferred embodiment, a Kalman filter model can be used, which can better handle process noise and observation noise, providing smoother and more accurate predictions. The prediction distribution contains a probability distribution of mean and variance (such as lognormal distribution or normal distribution), providing a basis for subsequent risk assessment.

[0053] Step S1.3, according to the received power distribution, construct a risk objective function balancing overload risk and quantization loss. Used to determine the adaptive receive gain.

[0054] Wherein, the overload risk is the risk of signal clipping caused by excessive gain; the quantization loss is the risk of signal-to-noise ratio drop caused by insufficient gain; the two are mutually restrictive. This step constructs a risk objective function J(G) to describe both risks. The function takes the receive gain as the independent variable, and its function value represents the overall risk level at a given gain. The construction of the function is based on the predicted received power distribution, which determines the probability of signal power falling into the overload or under-quantization area at different preset gains.

[0055] Step S1.4, determine the pre-cut gain by optimizing the risk objective function.

[0056] Accordingly, after the risk objective function J(G) is constructed, the gain G that minimizes the value of J(G) is found, which is called optimization. The processor executes a numerical optimization algorithm to systematically search for the minimum point of J(G) within a preset gain search range. Since the decision must be made in a very short time before the time slot arrives, an algorithm with fast convergence speed is preferred, such as the three-part search method (for unimodal functions) or the Newton iteration method. The output result of the optimization algorithm, i.e., the gain that minimizes the total risk, is defined as the pre-cut gain G _1 .

[0057] Step S1.5, in the target time slot, demodulate the received signal using the receive gain determined based on the pre-cut gain, and estimate the measured received power.

[0058] Specifically, when the target time slot arrives at the predetermined time, the controller of the reverse demodulation device sets the VGA gain in the radio frequency link to the pre-cut gain. Next, the ADC samples and quantizes the analog signal after gain adjustment. The digital signal processor (DSP) then performs a series of demodulation operations such as synchronization, equalization, and decoding on the sampled data. During or after the demodulation process, the system calculates the average power of the received signal samples to obtain the measured received power of this transmission; it is a true reflection of the channel and terminal transmission state this time.

[0059] Step S1.6, use the measured received power to update the received power statistics in the terminal portrait lookup table. Used to constitute the closed-loop feedback of the entire adaptive system.

[0060] Specifically, the processor uses the measured received power as a new observation sample to update the power statistics μ and σ stored in the terminal portrait. The update process can also use exponential weighted average or Kalman filter correction. Based on this, the terminal portrait can continuously learn and evolve, and can dynamically track the receive power drift caused by terminal movement, power adjustment or environmental changes, so that subsequent predictions can be based on the latest information, and the entire adaptive gain control system can remain accurate and efficient.

[0061] Embodiment 2, provides an optional implementation of the construction and solution of the risk objective function; describes the mathematical model for deciding the pre-cut gain.

[0062] The method of this embodiment is based on the predicted received power distribution and further includes:

[0063] Step S2.1, optionally, define the risk objective function as the weighted sum of the overload probability term and the quantization noise power term; wherein the overload probability term represents the risk of signal overload caused by the pre-cut gain, and the quantization noise power term represents the quantization loss caused by the pre-cut gain.

[0064] In some embodiments, the risk objective function J(G) is defined as: J(G) = λ _clip • Pr _clip (G) + λ _q • QuantNoise(G); where J(G) is an overall risk function with respect to the receive gain G, i.e., the pre-cut gain value to be optimized. Pr _clip (G) is an overload probability term corresponding to the overload risk, representing the probability of signal clipping distortion when the gain is set to G. QuantNoise(G) is a quantization noise power term corresponding to the quantization loss, representing the equivalent impact of quantization noise on signal quality when the gain is set to G. λ _clip is a risk coefficient for balancing the overload risk and the quantization loss; both are configurable policy parameters, for example, in a scenario where the signal fidelity requirement is extremely high and clipping is never allowed, the value of λ _q may be increased; while in a weak signal scenario where the signal-to-noise ratio is more critical, the value of λ _clip may be increased. _q

[0065] Step S2.2, according to the pre-cut gain, the full-scale range of the analog-to-digital converter ADC, and the signal peak-to-average ratio, the power threshold is defined; combined with the received power distribution, the probability of the power threshold being exceeded is evaluated to obtain the overload probability term.

[0066] In some embodiments, the determination of the overload probability term Pr _clip (G) includes: calculating the power threshold P _th (G) = {A _max 2} / {G _2 • PAPR _max}. Where P _th (G) represents the maximum average power threshold that the VGA input can withstand without causing ADC clipping at a given receive gain G. A _max is the full-scale amplitude of the analog-to-digital converter (ADC), which is a fixed hardware parameter, for example, 1.0V, representing the maximum voltage peak value that the ADC can handle. PAPR _max is the upper limit value of the signal peak-to-average ratio, reflecting the peak value characteristics of the signal waveform, for example, for a burst signal using QPSK modulation, the value can be set to 6dB (i.e., the linear value is 4.0).

[0067] Correspondingly, for a signal with an average power of P, its peak power is approximately P•PAPR _max , and the corresponding peak amplitude is P•PAPR _max ; after being amplified by the gain G, the peak amplitude becomes G•P•PAPR _max ; let its value equal the full-scale range A​_max , the inverse P can be obtained power threshold P _th (G).

[0068] Further, based on the received power distribution, the overload probability term Pr _clip (G) is calculated. _clip (G) = 1 - F _P (P _th (G)); where F _P is the cumulative distribution function (CDF) corresponding to the received power distribution. Pr _clip (G) is the probability that the predicted received power P exceeds the threshold P _th (G). In the preferred embodiment, since the received power is closer to Gaussian distribution in the logarithmic domain (dBm), P _pred is modeled as a lognormal distribution. At this time, F _P is the CDF of the lognormal distribution. Alternatively, if the lognormal model does not fit well, a normal distribution model can also be used.

[0069] Step S2.3, model the quantization noise power term as inversely proportional to the square of the pre-amplification gain. In other embodiments, the quantization noise power term QuantNoise(G) is modeled as: QuantNoise(G) = κ / G _2 ; where κ is a coefficient determined by the ADC quantization step and the baseband normalization constant.

[0070] Further, this model reflects that the quantization noise of the ADC is mainly determined by the step of its least significant bit (LSB), and its total power at the input end of the ADC can be approximately constant. When the constant noise power is converted back to the input end of the VGA, its equivalent power is inversely proportional to the square of the applied gain. Therefore, when G is very small, the equivalent quantization noise QuantNoise(G) will increase sharply, resulting in a deterioration of the signal-to-noise ratio. κ is a coefficient, whose value is determined by the quantization step of the ADC (itself determined by A _max and the number of effective bits Nbit of the ADC) and the baseband normalization constant of the system. After system calibration, κ can be regarded as a constant.

[0071] Step S2.4, optimize the risk objective function J(G). After J(G) is constructed, it can be seen that the Pr _clip (G) term is monotonically non-decreasing with G (the larger the gain, the easier the overload); while the QuantNoise(G) term is monotonically decreasing with G (the larger the gain, the smaller the quantization noise effect). The weighted sum J(G) of the two terms forms a unimodal function with a unique minimum point, ensuring the stability and convergence of the optimization solution. To find the G _1The optimization must be done within the processing latency budget. Therefore, it is preferred to use a numerical algorithm that converges fast, for example, if the gradient of J(G) is not easy to compute, a ternary search method can be used; if the gradient is computable, a Newton-Raphson hybrid method can be used. The algorithm will have an early stopping mechanism, for example, when the maximum number of iterations is reached or a percentage (e.g., 90%) of the latency budget is used, the algorithm will be forced to stop and output the current optimal solution.

[0072] Further, the continuous solution G * obtained by the above optimization needs to be quantized to a discrete gain step supported by the VGA hardware (e.g., 0.5dB). The quantization process includes: finding G * and G _floor , which are the two adjacent discrete gain points of G _ceil . By computing the values of J(G _floor ) and J(G _ceil ), the one with smaller risk function value is chosen as the final pre-cut gain G _1 . As a preferred protection strategy, if the values of J(G _floor ) and J(G _ceil ) are very close, the gain point corresponding to the lower Pr _clip (G) value (i.e., the smaller overload risk) is preferred.

[0073] Embodiment 3 provides a preferred technical solution for the secondary scaling and non-linear compensation of the login time slot, which solves the technical problem that the pre-cut gain G _1 has a large deviation due to the absence or obsolescence of the terminal profile when the terminal accesses for the first time (i.e., the login time slot). Exemplarily, the embodiment can be implemented by the following steps:

[0074] Step S3.1, when the target time slot is a login time slot, the method further comprises: applying a pre-cut gain to receive an encoded power reference pilot containing a plurality of pilot segments, the pilot segments corresponding to a known set of nominal amplitudes. That is, when the target time slot is a login time slot, the gain is set to the pre-cut gain, and the encoded power reference pilot containing a plurality of pilot segments is received, the pilot segments corresponding to a known set of nominal amplitudes.

[0075] Wherein, the login time slot is a special time slot used for sending a login request when the terminal accesses the network. At this time, the pre-cut gain is only a rough estimate based on limited information. The encoded power reference pilot (CRP-UW) is a special preamble signal designed for channel sounding and link calibration, which does not carry service data; it is composed of K different pilot segments, each of which has a predetermined and known nominal amplitude a _k , all a _k for k = 1,..., K together form a set of nominal amplitudes {a _k}. For example, {a _k}may be a step signal with increasing amplitudes. At the login time slot, the receiver sets the gain to the pre-cut gain according to the decision made in the previous step, and captures the login burst signal. Next, by segmenting the login burst signal, the actual observed amplitudes of each pilot segment k are measured, and a set of observed amplitudes {A _obs}is obtained.

[0076] Step S3.2, jointly estimate the absolute gain deviation and the non-linear coefficients based on the set of observed amplitudes and the set of nominal amplitudes. Accordingly, by comparing the ideal set of nominal amplitudes with the actual set of observed amplitudes, the absolute gain deviation and the non-linear coefficients can be estimated simultaneously; where the absolute gain deviation reflects the error between the pre-cut gain and the actual required gain, i.e. the inaccuracy of the pre-cut gain; the non-linear coefficients describe the distortion characteristics of the receiver's RF front-end (especially the LNA and VGA) under strong signals.

[0077] Step S3.3, the non-linear coefficients a, b are used to fit a cubic soft-compression model of the receiver's front-end: y ~ a • x + b • x 3 ; where y is the observed amplitude, and x is the nominal amplitude. Or, the non-linear coefficients a, b are used to fit a cubic soft-compression model of the receiver's front-end: y ~ a • x + b • x 3 ; to relate the observed amplitude y to the nominal amplitude x.

[0078] Accordingly, the model is a standard model to describe non-linear distortion, where x represents the ideal input amplitude (e.g. a _k ), and y represents the observed output amplitude (e.g. A _k ). a is the linear gain coefficient of the model, and b is the cubic non-linear coefficient, usually negative, representing the larger the signal amplitude, the lower the gain (slope), i.e. soft-compression characteristics. The absolute gain deviation (or a) and the non-linear coefficients (b) can be jointly estimated, preferably by a robust regression algorithm. For example, in estimating the absolute gain deviation AG (or a), the weighted least squares method can be used, using Huber weights to calculate the weights. This method can reduce the adverse effects of individual pilot segments (outliers) affected by strong noise on the estimation results. In estimating b, to prevent unrealistic b values (i.e. overfitting) when there are few data points or high noise, a regularization term can be introduced into the cost function of the regression, for example, L2 regularization _beta • b 2 ; where l _beta is a tunable regularization parameter to penalize excessively large b values, making the model more stable.

[0079] Step S3.4, based on the non-linear coefficients a, b, construct the inverse mapping of the cubic soft-compression model, which is used to correct the observed amplitude y back to the nominal amplitude x: x ~ c _1 • y + c _3 • y 3; where c _1 , c _3 are coefficients of the inverse mapping.

[0080] Correspondingly, the inverse mapping is calculated; given a distorted observation y, the ideal value x that it should be is back-calculated. This inverse mapping can also be approximated by a cubic polynomial. c _1 and c _3 are coefficients of the inverse mapping, which can be derived from a and b (e.g., under small-signal approximation, c _1 ~ 1 / a and c _3 ~ -b / a 4 ). It should be understood that this step is used to compensate for the modeled non-linear distortion.

[0081] Step S3.5, based on the non-linear coefficients, generate linearization compensation parameters. In some embodiments, the linearization compensation parameters can also be generated based on the coefficients c _1 , c _3 of the inverse mapping.

[0082] Further, the linearization compensation parameters are the specific implementation of the inverse mapping to be performed in actual demodulation. As a preferred implementation, the linearization compensation parameters can be a digital compensation lookup table; this table stores a series of (y, x) mapping pairs (or (y, Ax) correction pairs). After ADC sampling, a digital signal processor (DSP) performs a non-linear correction on each digital sampling value by table lookup before demodulation, recovering the linearity of the signal in the digital domain.

[0083] As another alternative implementation, the linearization compensation parameters can also be a VGA setting correction table. This table guides the VGA controller to make small, dynamic adjustments to the gain of the VGA according to the current signal power level, introducing an inverse non-linear characteristic in the analog domain to counteract the compression effect in the front end. Further, to prevent the compensation process itself from introducing instability (e.g., a spike on the compensation curve can lead to spectral regrowth), when generating the linearization compensation parameters, it is preferred to impose a smoothing constraint on them, such as limiting the first derivative (slope) of the compensation curve.

[0084] Step S3.6, determine the gain synthesis function g(•); where g(•) synthesizes the absolute gain deviation AG, the effective slope of the inverse mapping (derived from c _1 , and the pre-set target ADC utilization.

[0085] Further, the system needs to calculate the secondary scaling gain, which is not only the correction of the pre-scaling gain plus the absolute gain deviation, but also must take into account the gain variation brought about by linearization compensation (mainly c _1The embodiment) and needs to adjust the final signal amplitude to the optimal working point of the ADC (i.e. the target ADC utilization, for example, to make the average power at -10dBfs of the full scale, i.e. about 70-85% of the amplitude utilization). The gain synthesis function g(•) is the function used to calculate the synthesis correction coefficient.

[0086] In another possible embodiment, g(•) also takes into account the reliability of the absolute gain deviation estimate. Specifically, the regression algorithm, when estimating the absolute gain deviation, also outputs its estimation variance Var(ΔG). A conservative correction coefficient ζ is included in g(•), whose value is positively correlated with Var(ΔG). When Var(ΔG) is large (indicating that the estimate of ΔG is unreliable), the conservative correction coefficient is increased, so that the adjustment range of the secondary scaling gain is more conservative, preventing system oscillation caused by false estimates.

[0087] Step S3.7, based on the pre-cut gain, the absolute gain deviation and the non-linear coefficient, calculate the secondary scaling gain. As another alternative implementation, the secondary scaling gain G _2 : G _2 =G _1 • g(•); where G _1 is the pre-cut gain.

[0088] Correspondingly, G _2 is called the secondary scaling gain, which is the accurate gain value synthesized from the pre-cut gain, the gain deviation correction (ΔG), the non-linear compensation slope (c _1 ) and the target utilization. The calculated secondary scaling gain and the linearization compensation parameters are stored (for example, written into the terminal's profile) for the terminal's subsequent demodulation of all regular service time slots.

[0089] Embodiment 4, provides a specific implementation process of the demodulation operation of the service time slot, for describing how the system processes the subsequent regular service or control time slot of the terminal after successfully completing the secondary scaling of the login time slot.

[0090] In this embodiment, in the service or control time slot after the login time slot, the received signal is demodulated using the receive gain determined based on the pre-cut gain, exemplarily including: demodulating the received signal of the service or control time slot using the secondary scaling gain and the linearization compensation parameters.

[0091] Wherein, the service or control time slot refers to the regular time slot used to transmit actual user data or control signaling after the terminal is logged in. In this embodiment, when the service / control time slot allocated to a certain terminal is about to arrive, the processor reads the secondary scaling gain and the linearization compensation parameters calculated for it from the profile lookup table (LUT) or the temporary context of the terminal.

[0092] Specifically, at a pre-cut timing point before the arrival of the time slot, the processor instructs the VGA to set the receive gain to a quadratic scaling gain. The ADC samples the analog signal adjusted by the quadratic scaling gain to obtain a digitized sample sequence. The digital signal processor (DSP) applies a linearization compensation parameter to the digital sample sequence for non-linear compensation. Preferably, the DSP utilizes the linearization compensation parameter (which is now a digital look-up table) to perform the inverse mapping modification of each sample value y to obtain the linearized sample value x. Alternatively, the linearization compensation parameter (which is now a VGA correction table) has already cooperated with the quadratic scaling gain to dynamically fine-tune the VGA gain to achieve linearization in the analog domain, and the digital domain no longer needs additional compensation. _1 • y + c _3 • y 3 The linearized sample value x. Optionally, the linearization compensation parameter (which is now a VGA correction table) has already cooperated with the quadratic scaling gain to dynamically fine-tune the VGA gain to achieve linearization in the analog domain, and the digital domain no longer needs additional compensation.

[0093] The linearized sample sequence is sent to the subsequent conventional demodulation link to perform operations such as burst synchronization, carrier recovery, clock recovery, equalization, and decoding, and output the demodulation result of the service data frame.

[0094] By applying the calibrated quadratic scaling gain and linearization compensation parameter, the amplitude and linearity of the signal are in an adaptive state before entering the demodulator, which improves the demodulation performance and link margin of the conventional service time slot.

[0095] Embodiment 5 describes an optional implementation of a safety gain and abnormal fallback mechanism for robustness protection; solves the problem of communication interruption caused by setting an incorrect gain (such as a too high quadratic scaling gain) when the prediction model fails, the channel mutates, or the calibration process of the login time slot fails. In specific embodiments, including:

[0096] Step S5.1, determine a conservative power upper bound based on the received power statistics; calculate the safety gain according to the conservative power upper bound, the full scale of the analog-to-digital converter ADC, and the signal peak-to-average ratio.

[0097] In other embodiments, the conservative power upper bound P _max : P _max = μ + η • σ; where η is a conservative coefficient set by the strategy.

[0098] Correspondingly, this step is calculated in parallel with the risk optimization (R2AGC) as a preparation for the worst case. μ and σ are the power statistics read from the terminal profile look-up table (LUT). η is a conservative coefficient set by the strategy to control the degree of conservatism of the upper bound. The value of η reflects the tolerance to risk, for example, η = 3 corresponds to a 3-sigma upper bound in statistics, and η = 5 corresponds to a more conservative 5-sigma upper bound. The value of η can be dynamically adjusted by the (strategy adjustment) module.

[0099] Furthermore, calculate the security gain G. _safe :G _safe ={A _max} / {sqrt{PAPR _max •P _max}}; where A _max PAPR is the full-scale range of the analog-to-digital converter (ADC). _max This represents the signal peak-to-average power ratio (PAR). Wherein, the safety gain G... _safe It is a calculated, conservative gain value; even if the terminal's received power P spikes to a conservative upper bound P with an extremely low probability. _max , application G _safe After gaining, the peak amplitude of the signal will just not exceed the full-scale A of the ADC. _max The safety gain is calculated and cached as a contingency plan.

[0100] Step S5.2: Estimate the clipping rate based on the second-order scaling gain and the received power distribution. In some embodiments, the second-order scaling gain G is used. _2 Replace the pre-cut gain G and calculate the power threshold P. _th (G _2 Based on the cumulative distribution function F corresponding to the received power distribution. _P and P _th (G _2 Estimate the clipping rate (ClipRate): ClipRate = 1 - F _P (P _th (G _2 )).

[0101] Specifically, this step is a risk posterior check performed immediately after the second-order calibration gain is calculated. The system uses the predicted power distribution, but the gain substituted is the second-order calibration gain that will be applied, used to estimate the actual overload risk (ClipRate) of the second-order calibration gain under the current prediction.

[0102] Step S5.3: When the clipping rate exceeds a preset threshold, an abnormal state is determined and a rollback instruction is generated. The rollback instruction is used to instruct the application of a safety gain. In some embodiments, when ClipRate exceeds a preset clipping rate threshold, an abnormal flag is set; in response to the abnormal flag, a reference safety gain G is generated. _safe The rollback command.

[0103] The preset clipping rate threshold is an engineering parameter, such as 0.5% or 1%. If ClipRate(G _2 If the value exceeds this threshold (indicating that the secondary scaling gain may be too high and there is a significant risk of clipping), the system will immediately set the exception flag Abn_flag.

[0104] Optionally, in the login time slot, if the CRP-UW pilot detection fails, or the residual of the non-linear fitting is too large, Abn_flag will also be triggered, all of which indicate that the secondary scaling process has failed, and the secondary scaling gain is not reliable. Once Abn_flag is set, the system will immediately generate a fallback instruction Fallback_cmd, the content of which explicitly points to or references the calculated safe gain.

[0105] Step S5.4, applying the secondary scaling gain and the linearization compensation parameter for demodulation, specifically including: responding to the fallback instruction, applying the safe gain instead of the secondary scaling gain for demodulation.

[0106] This step describes the execution of the fallback instruction. Before demodulation is executed, the processor will check whether there is a valid Fallback_cmd. If Fallback_cmd is invalid, demodulation is executed, and the secondary scaling gain and the linearization compensation parameter are applied. If Fallback_cmd is valid, the processor will discard the secondary scaling gain and the linearization compensation parameter (even if they have been calculated), and instead instruct the VGA to apply the safe gain as the receive gain. The valid determination condition can be determined by existing methods (such as verifying the integrity by the instruction check bit, checking whether the instruction timestamp is within the valid period, etc.).

[0107] As a preferred embodiment, while applying the safe gain, the system will also disable the high-order linearization compensation (i.e. the linearization compensation parameter). Since the safe gain is designed to make the signal work in the low amplitude region of the ADC, the non-linear effect in this region is already weak, and forcibly compensating may introduce noise; at the same time, the safe gain itself is a conservative estimate, and its accuracy is much lower than that of the secondary scaling gain, so it is meaningless to cooperate with the linearization compensation parameter.

[0108] In this embodiment, the present application constructs a complete prediction-calibration-verification-fallback robust control loop, so that even in the extreme case of prediction or calibration failure, the receive link can avoid clipping distortion by falling back to the safe gain, and the minimum availability of the communication link is guaranteed.

[0109] Embodiment 6, provides specific, reproducible numerical calculation cases, especially the optional implementation process of R2AGC pre-cut gain and safe gain.

[0110] Assume that the full-scale A of the ADC _max =1.0 (linear amplitude unit). The upper limit of the signal peak-to-average ratio PAPR _max =4.0 (i.e. 6dB). The VGA gain step ΔG=0.5dB. The overload risk coefficient λ _clip =10.0. The quantization risk coefficient λ _q =1.0×10 -12(the value already includes the effect of the kappa factor). The conservative factor η = 3.0 (for safety gain calculation). Read the predicted received power distribution of a certain terminal from the LUT. The distribution is preferably modeled as a lognormal distribution. Its statistical parameters in the log domain (natural logarithm) are: mean μ _Z = -10.0, standard deviation σ _Z = 0.5. (Note: μ _Z = -10.0 corresponds to the linear power median P _median = exp(-10.0) ≈ 4.54 x 10 -5 ).

[0111] Further, determine the conservative power upper bound P _max : for a lognormal distribution, the conservative upper bound at 3-sigma (in the log domain) is μ _Z + η • σ _Z . P _max = exp(μ _Z + η • σ _Z ) = exp(-10.0 + 3.0 • 0.5) = exp(-8.5) ≈ 2.03 x 10 -4 . Calculate the safety gain G _safe : G _safe = A _max / sqrt{PAPR _max • P _max}; G _safe = 1.0 / sqrt{4.0 • 2.03 x 10 -4} = 1.0 / sqrt{8.12 x 10 -4} ≈ 1.0 / 0.0285; G _safe ≈ 35.09 (linear gain factor).

[0112] Further, define the power threshold P _th (G): P _th (G) = A _max 2 / (G 2 • PAPR _max ) = 1.0 2 / (G 2 • 4.0) = 0.25 / G 2 . Define the overload probability term Pr _clip (G): Pr _clip (G) = P(P > P _th (G)). Let P be a lognormal random variable. Pr _clip (G) = P(ln(P) > ln(P _th (G))); Pr _clip (G) = 1 - Φ((ln(0.25 / G2 )-μ _Z ) / σ _Z ), where Φ is the CDF of the standard normal distribution. Pr _clip (G) = 1 - Φ((ln(0.25) - 2ln(G) - (-10.0)) / 0.5); Pr _clip (G) = 1 - Φ((8.614 - 2ln(G)) / 0.5); define the quantization noise term QuantNoise(G): QuantNoise(G) = κ / G 2 In this example, λ _q The combined impact of κ is set to 1.0 x 10 -12 . For simplicity of representation, let λ _q QuantNoise(G) = (1.0 x 10 -12 ) / G 2 .

[0113] The combined J(G): J(G) = 10.0 • [1 - Φ((8.614 - 2ln(G)) / 0.5)] + (1.0 x 10 -12 ) / G 2 .

[0114] Further, the minimum of J(G) is found within the latency budget by a numerical optimization algorithm (such as ternary search or Newton's method). Exemplarily, the function value trend is shown for several G points: try G = G _safe ≈ 35.1: 2ln(35.1) ≈ 7.12. Pr _clip the Φ kernel of the term ≈ (8.614 - 7.12) / 0.5 = 2.988. Pr _clip ≈ 1 - Φ(2.988) ≈ 0.0014. The QuantNoise term ≈ (1.0 x 10 -12 / 35.1 2 ≈ 8.12 x 10 -16 . J(35.1) ≈ 10.0 • 0.0014 + 8.12 x 10 -16 ≈ 0.014.

[0115] try G = 100.0 (a larger gain): 2ln(100) ≈ 9.21. Pr _clip the Φ kernel of the term ≈ (8.614 - 9.21) / 0.5 = -1.192. Pr _clip ≈ 1 - Φ(-1.192) ≈ 0.883. The QuantNoise term ≈ (1.0 x 10 -12 ) / 100 2 = 1.0 x 10 -16 . J(100) ≈ 10.0 • 0.883 + 1.0 x 10 -16≈8.83. (Very high risk)

[0116] Trial G = 80.0: 2ln(80) ≈ 8.77. Pr _clip Kernel of term Φ ≈ (8.614 - 8.77) / 0.5 = -0.312. Pr _clip ≈1 - Φ(-0.312) ≈ 0.622. QuantNoise term ≈ (1.0 x 10 -12 ) / 80 2 ≈1.56 x 10 -16 . J(80) ≈ 10.0 • 0.622 + 1.56 x 10 -16 ≈6.22. By numerical solver (e.g. search between G = 35 and G = 50), it can be found that J(G) reaches a minimum near G ≈ 45.0.

[0117] Assume the continuous optimal solution is 45.1. VGA step is ΔG = 0.5 dB (about 10 0.5 / 20 ≈1.059 times). Find the two closest discrete points G _floor and G _ceil . Calculate J(G _floor ) and J(G _ceil ). Choose the one that makes J(G) smaller. Assume J(G) = 0.0051 at G _floor = 44.8, and J(G) = 0.0050 at G _ceil = 45.3. The system chooses G _ceil = 45.3 as the final G _1 . If J(G _floor ) and J(G _ceil ) are close, but Pr _clip (G _floor ) < Pr _clip (G _ceil ), the system prefers the G _floor with lower overload probability. Output safe gain G _safe ≈ 35.1. Optimal pre-cut gain G _1 = 45.3 (after quantization).

[0118] It can be seen that G _1 (45.3) is slightly larger than G _safe (35.1), as expected; R2AGC, after weighing the quantization noise, chooses a slightly larger gain than the absolute safety, to get a better signal-to-noise ratio, while controlling the overload probability Pr _clip within an acceptable range.

[0119] Embodiment 7, optional implementation procedure of closed-loop backwriting and prediction generation of terminal profile, and how the system learns and evolves with real measurement data to make the profile continuously accurate. As an example, this embodiment can be implemented by the following steps:

[0120] Step S7.1, fuse multi-source real measurement power data. In this embodiment, the system preferably fuses power observations from different time slots to obtain more reliable power estimates before updating the profile.

[0121] Specifically, the processor obtains at least two power estimates: one is the login power estimate obtained by CRP-UW pilot estimation in the login time slot; the other is the service power estimate obtained by demodulation in the service time slot.

[0122] The fusion process adopts confidence weighted average. The weight of the login power is preferably determined based on the robustness evaluation and fitting residual statistics of the regression algorithm; the better the fitting effect, the higher the weight. The weight of the service power is determined based on whether it is marked as abnormal (such as Abn_flag) or its demodulation quality (such as error vector amplitude EVM). The obtained fused power P _fused and its fusion confidence Conf _fused are used for subsequent update.

[0123] In some scenarios, if the current time slot is not a login time slot, the system can skip fusion and directly use the service power estimate as the fused power.

[0124] Step S7.2, adaptively update the profile statistics. The processor uses the fused power to update the received power statistics (mean μ and standard deviation σ) stored in the terminal profile lookup table.

[0125] Correspondingly, this update can adopt an adaptive smoothing algorithm. For example, use exponential weighted moving average: μ _new =α _ema •P _fused +(1-α _ema )•μ _old . Wherein μ _new , μ _old are the profile mean before and after update respectively; α _ema is the smoothing coefficient, which is not fixed and is associated with the fusion confidence Conf _fused ; the higher Conf _fused , the larger the value of α _ema , so that this observation can correct the profile faster. In another possible implementation, Kalman filter can also be used for update. At this time, the fused power is used as the observation value, and the Kalman gain will be determined according to the prediction error and the observation noise (which can be Conf _fusedThe automatic adjustment is implemented to achieve statistically optimal estimation. At the same time, the standard deviation σ is also updated according to the residual error of μ _fused and the residual error of μ _old to reflect the fluctuation characteristics of the power.

[0126] Step S7.3, updating the calibration and aging information in the profile. Accordingly, the processor uses the fused power to update the received power statistics (mean μ and standard deviation σ) stored in the terminal profile lookup table. While updating μ and σ, the processor also writes the nonlinear coefficients α, β and the quadratic scaling gain specific to the terminal into the profile entry of the terminal, along with the current timestamp.

[0127] Further, the embodiment also includes a profile aging mechanism. Accordingly, the background process periodically scans the LUT, and for entries whose timestamp (i.e. last update time) has exceeded a preset aging threshold (e.g. a few minutes or hours), the profile entries are aging, and the system actively reduces their credibility or artificially increases their standard deviation σ.

[0128] Due to the aging of the profile entries, they may no longer reflect the current channel state. By actively increasing their σ _max , the upper bound of the power P _new will be increased accordingly, so that the calculated safety gain is more conservative; on this basis, a lower credibility will also trigger a larger pre-cut timing margin. So that the system automatically falls back to a safer and more robust working mode when facing unknown states (i.e. obsolete profiles).

[0129] Step S7.4, generating the next frame of prediction parameters. Specifically, after the profile is updated, the system will immediately perform a prediction to prepare for the next time slot of the terminal. Based on the updated μ _new and σ _new , the prediction model (such as the drift-white noise Kalman model or simple exponential weighting) is applied to generate the next frame of prediction parameters, which will be stored in the profile as direct input data for the next cycle.

[0130] Above, the embodiment constitutes a closed loop of observation-fusion-update-prediction, so that the terminal profile lookup table becomes a dynamic evolving live database, which is the basis for the invention to achieve long-term, stable and adaptive gain control.

[0131] Embodiment 8, preferred implementation of providing adaptive adjustment of risk coefficient. In some possible implementations, the risk coefficient can be manually configured by the operator according to experience. However, the embodiment provides an adaptive strategy to enable the system to self-tune. The value of the risk coefficient is automatically optimized according to the long-term running performance of the system, so that the working point of the entire system remains on the adaptive risk-performance balance point.

[0132] In particular, the method comprises the following steps:

[0133] Step S8.1, aggregate risk and abnormal information. Accordingly, the system sets a strategy adjustment module, which monitors and aggregates (e.g., by hour or by day) the performance statistics of the entire network or a specific terminal group.

[0134] The aggregated information preferably includes: a risk assessment summary, in particular, the optimized risk objective function value and the predicted value of the pre-quantization gain G _1 , the overload probability Pr _clip (G _1 ), the actual clipping rate and the frequency of abnormal flag Abn_flag; the error vector magnitude statistics, the bit error rate statistics and the acquisition time. The system forms a risk portrait according to such data, reflecting the recent risk trajectory and performance.

[0135] Among them, the bit error rate statistics refers to the ratio of the number of error code symbols to the total number of transmission code symbols. The acquisition time refers to the time taken by the demodulator to complete the preamble detection and synchronization. The error vector magnitude statistics refers to the statistical value of the error vector magnitude between the actual received signal and the ideal modulated signal, reflecting the distortion degree of the modulated signal.

[0136] Step S8.2, adaptively adjust the risk coefficient. Accordingly, the strategy adjustment module inversely optimizes the overload risk coefficient λ _clip and the quantization risk coefficient λ _q according to the risk portrait and the preset system-level operation target (e.g., target clipping rate ≤ 0.1%, target error vector magnitude ≤ 5%). Exemplarily, its adjustment logic includes: if the actual clipping rate statistics are continuously higher than the system target (e.g., ≥ 0.1%), indicating that the system is too tolerant to the overload risk; the strategy module will increase the value of λ _clip . In subsequent optimization, since the weight of λ _clip • Pr _clip (G) term in J(G) increases, the optimizer will tend to choose a smaller G _1 to avoid overload and make the actual clipping rate fall. If the actual clipping rate is much lower than the target (e.g., ≈ 0%), but the error vector magnitude statistics or the bit error rate statistics are continuously high (indicating that the quantization noise is too large and the signal quality is poor), indicating that the system is too conservative. The strategy module will increase the value of λ _q (or relatively reduce λ _clip ). In subsequent optimization, the weight of λ _q • QuantNoise(G) term increases, and the optimizer will tend to choose a larger G _1 to reduce the quantization noise at the cost of a little overload risk in exchange for the improvement of the error vector magnitude.

[0137] Step S8.3, adjust the abnormal criterion threshold. As an optional implementation, this module not only adjusts the coefficient λ, but also adjusts the conservative coefficient η or the clipping rate threshold. For example, if the fallback path is frequently triggered, resulting in a decline in link performance, the system can appropriately relax the threshold to reduce unnecessary fallback.

[0138] Step S8.4, issue a policy update. The policy adjustment module packages the calculated new {λ _clip , λ _q} and new threshold into a policy adjustment package, which is issued to the R2AGC decision module for execution.

[0139] Through this embodiment, the system is upgraded from a passive adaptation (adapt to power changes) system to an active evolution (actively optimize the decision model) system, improving the intelligence level of the scheme and the robustness of long-term operation.

[0140] Embodiment 9 provides an optional implementation of the calculation and adaptive adjustment of the pre-cut timing (t _pre ), which discloses the control of the gain switching timing. This embodiment solves the technical problem that VGA gain switching requires time and that improper switching timing can cause signal transient distortion. This embodiment introduces the calculation and management of the pre-cut timing t _1 after determining G _1 and before applying G _pre . An example is as follows:

[0141] Step S9.1, basic calculation and constraints of t _pre . Accordingly, the pre-cut timing t _pre is defined as: the gain of the VGA must be set to G _pre at least t _1 milliseconds before the theoretical arrival time of the target time slot. It must meet the physical constraints of the hardware. Specifically, t _pre must be greater than or equal to the sum of the VGA settling time t _settle and the guard interval t _guard .

[0142] Where t _settle is the physical time required for the VGA hardware to receive a new gain instruction and for its analog output to stabilize to the new gain value (e.g., 50 microseconds). t _guard is the additional protection time added by the system to cope with uncertainties such as clock synchronization errors (e.g., 10 microseconds). By making t _pre ≥ t _settle + t _guard , the application ensures that when the burst signal of the target time slot actually arrives at the VGA, the gain of the VGA has already stabilized at G _1The preamble part of the signal is thus avoided from being polluted by the gain transient of the VGA.

[0143] Step S9.2, t is adjusted based on the prediction confidence _pre Margin adjustment. As a preferred implementation, t _pre is not a fixed value, but will be dynamically adjusted according to the confidence of the current decision. Specifically, when making a power prediction, the system will get a prediction confidence (e.g. based on the aging degree of the profile entry). t _pre is calculated with a margin that is negatively correlated with the prediction confidence. In other words, when the prediction confidence is low (e.g. the profile used is old, G _1 accuracy is questionable), the system will proactively increase the margin of t _pre .

[0144] An inaccurate G _1 may cause the subsequent demodulator to take longer time to capture. By setting the gain in advance, a longer stable signal window is reserved for the demodulator (especially the preamble detection and synchronization module), improving the success rate of capture under uncertain conditions.

[0145] Step S9.3, t is adjusted based on the capture time _pre Closed-loop adaptive adjustment. Preferably implemented in a strategy adjustment module. This module will continuously monitor the average value and fluctuation of the statistical capture time. If it is monitored that the capture time is consistently long, or frequently close to the upper limit of the time slot window (indicating difficulty in capture), the strategy module will trigger timing adjustment, proactively increasing the base value of t _pre . If the capture time is consistently stable and short, indicating that capture is very easy, the current t _pre margin is too large. The strategy module will appropriately reduce t _pre , to reduce the protection overhead of the time slot, and improve the overall throughput of the system (i.e. time slot utilization rate). On this basis, a closed-loop adaptive control of t _pre is constituted.

[0146] In summary, through the base calculation of t _pre , predictive margin adjustment and closed-loop adaptive adjustment, the embodiment realizes accurate and intelligent management of the gain switching timing, ensuring the stable operation of the R2AGC method of the application in high-speed burst TDMA systems.

[0147] According to one aspect of the application, the large dynamic reverse demodulation method based on the TDMA system can also be implemented by the following steps:

[0148] Step S1, performing time slot pre-prediction and risk minimization pre-gain decision.

[0149] Specifically, read the terminal portrait lookup table and the hardware margin parameter set, combine the cross-frame prediction parameter to model the receive power distribution of the same terminal in the current time slot, and obtain the predicted receive power distribution. Based on the predicted receive power distribution and the hardware margin parameter set, construct the objective function J(G) = λ _clip • Pr{G• receive power > ADC full scale A _max} + λ _q • quantization noise power (G | ADC full scale A _max , PAPR upper limit PAPR _max , in the time budget before the time slot arrives, numerically optimize the candidate gain, obtain the pre-cut gain and the risk assessment summary; at the same time, calculate the safety gain used in the abnormal time and the pre-cut timing. In the subsequent step, it is used for CRP-UW reception, abnormal fallback and service demodulation.

[0150] Step S2, perform login time slot CRP-UW reception and secondary calibration.

[0151] Further, after the pre-cut timing triggers, the pre-cut gain is applied to receive the preamble of the login time slot, read the CRP-UW parameter set and the login burst sample. Perform CRP-UW arrival detection and segmented energy estimation on the login burst sample to obtain the observed amplitude set of each segment, perform joint regression based on the CRP-UW parameter set and the observed amplitude set to solve the absolute gain deviation ΔG and the nonlinear coefficients α, β (used to describe the front-end soft compression characteristic); according to this, calculate the secondary calibration gain and the linearization compensation parameter (digital pre-linearization or VGA setting correction), calculate the clipping rate and the abnormal flag.

[0152] In the subsequent step, it is used for service / control time slot demodulation and LUT backwrite; when the abnormal flag or the clipping rate exceeds the threshold, a fallback instruction referencing the safety gain is generated.

[0153] Step S3, service / control time slot demodulation and closed-loop backwrite.

[0154] Optionally, before the same terminal's subsequent service or control time slot arrives, read the input data secondary calibration gain and linearization compensation parameter, and select the gain scheme according to the state of the fallback instruction (normal path uses the secondary calibration gain, and abnormal path uses the safety gain). After the normal processing such as synchronization, carrier and clock recovery, and equalization on the uplink burst of the current time slot, complete the demodulation with the set gain and linearization configuration to obtain the service data frame demodulation result. In the demodulation process, the receive power estimation, the error vector amplitude statistics, the bit error rate statistics and the capture time are counted, the terminal portrait lookup table of the same terminal is updated based on the receive power estimation and the nonlinear coefficients α, β, the terminal portrait update package and the next frame prediction parameter are generated. It is used for the next round of risk minimization pre-cut and system-level statistics.

[0155] Step S4, execute exception handling and protection policy adjustment.

[0156] Read risk assessment summary, clipping rate, exception flag, received power estimate and capture time, evaluate current link state according to preset differential threshold and clipping threshold rule. When the trigger condition is met, generate or update fallback instruction, adjust pre-cut timing and risk coefficient set {λ _clip , λ _q}, form policy adjustment and protection action record; synchronize policy adjustment and participate in next risk minimization pre-cut calculation. Ensure that the above steps are executed in the updated control law and timing configuration in the subsequent time slot.

[0157] Another example, a large dynamic reverse demodulation method based on TDMA system, can also be implemented as follows,

[0158] Step S11, read the profile and policy to form the initial statistics.

[0159] Specifically, read the terminal profile lookup table, policy adjustment and hardware boundary parameter set. Extract the next frame prediction parameters and the last time quadratic calibration gain G _2 , nonlinear coefficients α, β from the terminal profile lookup table, combine the risk coefficients and thresholds in the policy adjustment to generate the initial statistics and risk coefficient set {λ _clip , λ _q}.

[0160] Step S12, update the received power distribution across frames.

[0161] Correspondingly, read the initial statistics and the observation sequence in the profile (historical received power). Predict the received power by exponential weighting or Kalman filtering: if exponential weighting is used, the mean μ = α _meas_last • P _ema + (1-α _meas_last ) • μ0; where P _ema is the latest received power, μ0 is the received power mean estimate before updating, α _clip is the smoothing coefficient; σ is updated by sliding variance or Mahalanobis distance weighting; if Kalman filtering is used: model with state vector [μ, σ], perform one prediction-correction according to process noise and observation noise covariance. Get the predicted received power distribution and prediction confidence.

[0162] Step S13, construct the risk target of overload probability and quantization noise.

[0163] Further, read the predicted received power distribution, hardware boundary parameter set and risk coefficient set {λ _q , λ _clip}. Calculate the overload probability at a given gain G moment: Pr_clip (G)=1-F _P (A _max 2 / (G 2 •PAPR _max )), where F _P It is the cumulative distribution function of the received power P (determined by μ and σ of the predicted received power distribution). The quantization noise power term is approximated using an equivalent baseband: QuantNoise(G) = κ / G 2 (κ is given by the ADC quantization step size and the baseband normalization constant). Construct the risk objective function J(G)=λ _clip •Pr _clip (G)+λ _q •QuantNoise(G). Used for subsequent optimization.

[0164] Step S14: Solve for the optimal pre-cut gain and generate the safe gain under VGA step constraints.

[0165] Optionally, the risk objective function J(G), VGA step ΔG, and processing delay budget are read. A ternary search or Newton-Raphson iteration is used to find the continuous optimal solution for G within the allowable time budget. Quantization is performed according to the nearest available VGA step ΔG to obtain the pre-cut gain G. _1 Based on the worst-case received power P _max =μ+η•σ (η is set by the strategy adjustment, for example, η=3) and peak-to-average ratio (PAPR) _max Calculate the security gain G _safe =A _max / (sqrt(PAPR _max )•sqrt(P _max Provide a risk assessment summary (including Pr) _clip (G _1 ), QuantNoise(G _1 ), J(G _1 (Optimize convergence steps and time usage).

[0166] Step S15: Calculate the pre-cut timing and issue the control plan.

[0167] Furthermore, read the hardware boundary parameter set (including VGA settling time t). _settle , protection interval t _guard ), processing latency budget and prediction reliability Conf _pred Press t _pre ≥t _settle +t _guard The principle of +2 symbol period, combined with prediction confidence Conf _pred Perform margin adjustment (the lower the confidence level, the better the pre-cut timing t) _prelarger, the more the pre-cut timing t _pre and control plan, and is sent to the radio link controller. The symbol period refers to the duration of a single modulation symbol.

[0168] Step S21, pre-cutting and capturing CRP-UW segments according to timing.

[0169] Optionally, read the pre-cutting gain and pre-cutting timing t _pre , and the CRP-UW parameter set. At t _pre , apply the pre-cutting gain, receive the login burst preamble, slice the login burst samples based on the segment boundaries of the CRP-UW parameter set, and calculate the short-time mean square energy and short-time mean square amplitude of each segment; wherein the observation amplitude set = the short-time mean square amplitude set. Output the arrival flag.

[0170] Step S22, estimate the absolute gain deviation and perform coarse correction.

[0171] Optionally, read the observation amplitude set and the CRP-UW parameter set (including the nominal amplitude set {a _k}). Estimate the absolute gain deviation ΔG using the ratio regression method: ΔG = arg min _Δ Σ _k (A _k -Δ•a _k ) 2 . To suppress the influence of abnormal segments, assign low weights or remove segments with residuals greater than the q quantile (robust regression). Obtain the absolute gain deviation ΔG and robustness evaluation.

[0172] Step S23, estimate the nonlinear coefficient and generate linearization compensation.

[0173] Optionally, read the observation amplitude set and the CRP-UW parameter set. Fit using a cubic soft compression model: A _k ≈ α•a _k +β•a k 3 . Jointly solve the nonlinear coefficients α, β by least squares or Huber regression, and generate linearization compensation parameters such as constructing a digital pre-linearization inverse function or a VGA correction table.

[0174] Step S24, calculate the secondary scaling gain and clipping rate and determine abnormalities.

[0175] Optionally, read the absolute gain deviation ΔG, the nonlinear coefficients α, β, the pre-cutting gain G _1 , and the hardware boundary parameter set. According to the target, make the average utilization rate of the ADC between 70-85%, and the peak value does not exceed 92%, calculate the secondary scaling gain G _2 =G _1• G(ΔG, a, β), where g(•) is determined by the nonlinear inverse compensation jointly with the target utilization. The predicted received power profile and the quadratic scaling gain G are utilized simultaneously _2 The clipping rate is estimated as ClipRate = Pr _clip (G _2 ). If the clipping rate ClipRate exceeds a threshold or reaches a flag abnormality, the abnormality flag Abn_flag is set to True.

[0176] Step S25, generate or clear the fallback command.

[0177] Optionally, the abnormality flag Abn_flag, the safety gain, and the risk assessment summary are read. When the abnormality flag Abn_flag is True or the Pr _clip (G _1 ) in the risk assessment summary is higher than a policy threshold, the fallback command is generated and directed to the safety gain; otherwise, the fallback command is cleared.

[0178] Step S26, form the instantaneous power observation of the login time slot.

[0179] Optionally, the observation amplitude set and the quadratic scaling gain are read. The segment mean square amplitude of the CRP-UW is restored to the received power estimate at the VGA input side, which is used for subsequent profile update.

[0180] Step S31, select the gain and compensation scheme and perform demodulation.

[0181] Further, the quadratic scaling gain, the linearization compensation parameters, and the fallback command are read. If the fallback command Fallback_cmd is valid, the safety gain is applied and the high-order terms of linearization are disabled; otherwise, the quadratic scaling gain and the linearization compensation parameters are applied. The uplink burst of the current service / control time slot is synchronized, carrier recovered, clock recovered, and equalized to obtain the service data frame demodulation result and the received power estimate (time slot level).

[0182] Step S32, statistics performance indicators and assess the capture quality.

[0183] Correspondingly, the input data service data frame demodulation result and the received power estimate are read. The error vector amplitude statistics, the bit error rate statistics, and the capture time are calculated, and compared with the policy threshold to form the quality assessment.

[0184] Step S33, write back the terminal profile and generate the next frame prediction parameters.

[0185] Accordingly, read the received power estimate, the received power estimate, the nonlinear coefficients a, b, the quality assessment and the quadratic scaling gain on the input side of the VGA. Fuse the power observations of the registration time slot and the service time slot in a Kalman or exponentially weighted manner, update the mean μ and the standard deviation σ in the terminal image lookup table, and write the nonlinear coefficients a, b and the quadratic scaling gain into the image entry; calculate the next frame prediction parameters and the image update package (including the image timestamp and the reliability) based on the updated statistics.

[0186] Step S34, submit the image and generate system statistics summary.

[0187] Accordingly, read the terminal image update package and the quality assessment. Submit the terminal image update package to the image database, generate a system statistics summary (including Pr _clip , error vector magnitude statistics, bit error rate statistics, and sliding statistics of capture time), for operation and maintenance and strategy learning. Output image submission receipt for strategy adjustment and retention.

[0188] Step S41, aggregate risk and abnormal information.

[0189] Specifically, read the risk assessment summary, the pruning rate and the abnormal flag, the capture time and the quality assessment, and the system statistics summary. Form a risk image (including the risk trajectory of the last N frames and the abnormal events).

[0190] Step S42, adjust the risk coefficient and the rollback threshold.

[0191] Specifically, read the risk image and the system statistics summary. According to the target constraint (for example, the pruning rate target ≤ 0.5%, the error vector magnitude target ≤ the preset threshold EVM _target ). Optimize the risk coefficient set {λ _clip , λ _q} and the abnormal criterion threshold (difference threshold, conservative coefficient), and generate strategy adjustment.

[0192] Step S43, adaptively adjust the pre-cut timing.

[0193] Specifically, read the capture time and the abnormal flag. When the capture time is close to the upper limit or the abnormality occurs frequently, increase the margin of the pre-cut timing; when the capture margin is sufficient, reduce the pre-cut timing, get the timing adjustment and include it in the strategy adjustment.

[0194] Step S44, specifically record the protection action and generate a complete block log.

[0195] Specifically, read the input data risk image and the strategy adjustment. Generate a protection action record (including trigger conditions, action content, and effect summary) for operation and maintenance and audit, as an external terminal output and retention of the process.

[0196] As an optional implementation, step S13 can also be:

[0197] Further, read the predicted received power distribution, hardware boundary parameter set; based on the constraint that the ADC input peak value does not exceed the ADC full range A _max , under the condition that the average power is P, the instantaneous peak power is approximately PAPR _max ×P. Let the peak amplitude correspond to the ADC full range, and the power threshold P _th (G) = (A _max 2 ) / (G 2 •PAPR _max ).

[0198] According to the source of the predicted received power distribution and the goodness of fit, the lognormal model is preferentially used (the power is often taken as the logarithm, which is closer to Gaussian). When the Shapiro-Wilk test of the lognormal does not pass, the normal model is used. For the lognormal, let Z = ln(P). Z ~ N(μ _Z , σ Z 2 ). Pr _clip (G) = 1-Φ((ln(P _th (G))-μ _Z ) / σ _Z ). The normal Pr _clip (G) = 1-Φ((P _th (G) - μ) / σ) (0 when less than 0, 1 when greater than 1). The overload probability function Pr _clip (G) is obtained.

[0199] Further, read the hardware boundary parameter set (including ADC bit number N _bit , full range voltage V _fs , sampling rate, baseband normalization constant), predicted received power distribution. The quantization step size Δ _ADC = V _fs / 2 N_bit . In the unit normalized baseband, the quantization noise power is equivalent to Δ ADC 2 / 12. When the gain G is applied to the receive link, the average amplitude ratio at the ADC end rises to sqrt(G 2 •E[P]). The quantization noise is folded back to the VGA input side (or normalized baseband) to obtain the quantization noise power term QuantNoise(G) ∝ (Δ ADC 2 / 12) / G 2The constant term is determined by the normalization method, and is incorporated into the risk coefficient for calibration. To facilitate subsequent consistency with error vector amplitude statistics, the equivalent error vector amplitude contribution is calculated, which is used for log and policy learning (not directly into optimization, only into risk assessment summary).

[0200] Further, read the overload probability function Pr _clip (G), the quantization noise power term QuantNoise(G), and the risk coefficient set {λ _clip , λ _q}. The objective function is defined as: J(G) = λ _clip •Pr _clip (G) + λ _q •QuantNoise(G). Wherein, Pr _clip (G) is monotonically non-increasing with G (the greater the gain, the easier the overload, so the probability is monotonically increasing; after mapping with the inverse of G, it is monotonically); QuantNoise(G)≈constant / G 2 , monotonically decreasing with G; the sum of the two terms forms a hedge of rising and falling terms, and there is a single peak structure near the only minimum point. Combined with the VGA step ΔG, the time budget of the pre-cut timing, search in G ∈ [G _min , G _max ] (derived by historical G _1 / G _2 and hardware clipping) can be achieved. The output solvability conclusion (including the single peak interval and the initial search boundary) is output.

[0201] In one aspect of the application, step S14 can also be implemented as follows:

[0202] Optionally, read the risk objective function J(G) and the solvability conclusion, the processing delay budget, and the VGA step ΔG.

[0203] According to the solvability conclusion, determine [G _L , G _R ]. If there is no reliable prompt, generate a symmetric interval centered on the historical pre-cut gain G _1 . If the numerical gradient of J(G) can be calculated, use the Newton-tangent hybrid method and set an upper limit for iteration; otherwise, use the ternary search (ensure single peak convergence). Stop when the upper limit of iteration is reached or the time consumption exceeds the processing delay budget × α (α < 1 reserved amount), and take the current optimal value as a continuous domain candidate for quantization and neighborhood test. Output the optimization log (including the number of iterations, convergence flag, and time consumption), which is included in the risk assessment summary.

[0204] Optionally, read the continuous domain candidate, VGA step ΔG, and risk objective function J(G). Map the continuous domain candidate to the nearest two discrete points G _floor , G_ceil . Calculate J(G _floor ) and J(G _ceil ), select the smaller one as the pre-cut gain; if they are similar, prefer the one with lower overload probability. Record the final selected pre-cut gain and the alternative value, generate neighborhood record into risk assessment summary, facilitate hysteresis and hot start of next time slot.

[0205] Optionally, read the predicted received power distribution, hardware boundary parameter set, strategy parameters (differential threshold, conservative coefficient). Calculate the conservative power upper bound P _max = μ + η • σ (η is given by the strategy, default 3). Calculate the safety gain G _safe = A _max / (sqrt(PAPR _max )•sqrt(P _max )). If the safety gain makes QuantNoise(G _safe ) too large (exceeds the quality threshold), limit the lower limit and mark the quantization dominant state in the risk assessment summary, output the safety state marker; Facilitate subsequent parameter tuning.

[0206] The preferred embodiments of the application are described in detail above, but the application is not limited to the specific details of the above-described embodiments. Within the technical concept of the application, various equivalent transformations of the technical solutions of the application can be made, and these equivalent transformations all belong to the protection scope of the application.

Claims

1. A method of large dynamic reverse demodulation based on TDMA system, characterized in that, The method comprises: maintaining a terminal profile lookup table storing at least one terminal's receive power statistics; before a target time slot of the terminal arrives, reading the receive power statistics in the terminal profile lookup table, and predicting a receive power distribution of the target time slot based on the receive power statistics; constructing a risk objective function balancing overload risk and quantization loss according to the receive power distribution; determining a pre-cut gain by optimizing the risk objective function; applying a receive gain determined based on the pre-cut gain to demodulate a received signal in the target time slot, and estimating a measured receive power; updating the receive power statistics in the terminal profile lookup table using the measured receive power; wherein the risk objective function is constructed, specifically comprising: defining the risk objective function J(G) as: J(G) = λ _clip • Pr _clip (G) + λ _q • QuantNoise(G); wherein G is a pre-amplification gain, Pr _clip (G) is an overload probability term corresponding to an overload risk, QuantNoise(G) is a quantization noise power term corresponding to a quantization loss, λ _clip and λ _q are risk coefficients for balancing the overload risk and the quantization loss; wherein the overload probability term Pr _clip (G) is determined by calculating a power threshold P _th (G) : P _th (G) = {A _max 2} / {G 2 •PAPR _max} ; wherein A _max is a full scale of an analog-to-digital converter (ADC), and PAPR _max is an upper limit value of a peak-to-average ratio of a signal; and calculating an overload probability term Pr _clip (G) based on a received power distribution: Pr _clip (G) = 1 - F _P (P _th (G)) ; wherein F _P is a cumulative distribution function corresponding to the received power distribution. where the quantization noise power term QuantNoise(G) is modeled as: QuantNoise(G) = K / G 2 ; where K is a coefficient determined by the ADC quantization step size and the baseband normalization constant.

2. The method of claim 1, wherein, when the target time slot is a login time slot, the method further comprises: applying the pre-cut gain to receive a coded power reference pilot comprising a plurality of pilot segments, the pilot segments corresponding to a set of known nominal amplitudes; jointly estimating an absolute gain deviation and a non-linear coefficient based on observed amplitudes of the pilot segments and the set of nominal amplitudes.

3. The method of claim 2, wherein, The non-linear coefficients a, b are used to fit a cubic soft-compression model of the receive chain front-end: y ~ a • x + b • x 3 ; where y is the observed amplitude and x is the nominal amplitude.

4. The method of claim 3, wherein, The method further comprises: Based on the nonlinear coefficients a, b, the inverse mapping of the cubic soft compression model is constructed, which is used to correct the observed amplitude y back to the nominal amplitude x: x ~ c _1 • y + c _3 • y 3 ; wherein c _1 , c _3 are the coefficients of the inverse mapping; According to the inverse mapping coefficients c _1 , c _3 , the linearization compensation parameters are generated; determining a gain synthesis function g(•); wherein g(•) synthesizes an absolute gain deviation AG, an effective slope of the inverse mapping, and a preset target ADC utilization; the effective slope of the inverse mapping is derived from c _1 derivation; Computing the quadratic scaling gain G _2 : G _2 = G _1 • g(•); where G _1 is the pre-cut gain.

5. The method of claim 4, wherein, applying the receive gain determined based on the pre-cut gain to demodulate a received signal in a service or control time slot after the login time slot, specifically comprising: applying a quadratic scaling gain and a linearization compensation parameter to demodulate the received signal of the service or control time slot.

6. The method of claim 1, wherein, The method further comprises: determining a conservative power upper bound based on the receive power statistics; calculating a safety gain according to the conservative power upper bound, a full scale range of an analog-to-digital converter (ADC), and a signal peak-to-average ratio.

7. A reverse demodulation apparatus characterized by comprising: The apparatus comprises: a memory configured to store computer program instructions and a terminal profile lookup table; a processor configured to execute the computer program instructions, and configured to implement the method according to any one of claims 1 to 6.

Images