Signal processing algorithm quantification method and system based on DIGWO

Abstract

The invention discloses a signal processing algorithm quantification method and system based on a discrete grey wolf optimizer (DIGWO). The method comprises the following steps: acquiring a to-be-quantized parameter in a signal processing process; setting a quantization bit width for each parameter to be quantized, wherein the quantization bit width comprises a decimal bit width; optimizing the decimal bit width of each parameter by adopting a discrete grey wolf optimizer DIGWO to obtain a final small-quantity bit width; wherein the feedback function of the DIGWO is established based on obtaining relative balance between two targets of reducing bit width and improving signal processing performance. Experiments show that when the method is applied to a channel hardening message passing (CHEMP) MIMO detector, the DIGWO quantization provided by the invention realizes that the average decimal bit width is reduced by 41.8% compared with the unified quantization (UQ) and is reduced by 24.9% compared with the automatic hybrid precision quantization (AHPQ) of the existing algorithm, and basically no performance loss exists. The method is also suitable for quantifying other communication system algorithms.

Description

Technical Field

[0001] The present invention relates to a signal processing algorithm parameter quantization technology, and particularly to a MIMO detection quantization technology, belonging to the field of quantization technology. Background Art

[0002] With the development of communication systems, multiple-input multiple-output (MIMO) technology has continuously demonstrated its superior performance in terms of high spectral efficiency and energy efficiency. To reduce the complexity of MIMO detectors, various MIMO detection algorithms have been proposed, especially the Bayesian message passing (BMP) algorithms such as belief propagation (BP), expectation propagation (EP), approximate message passing (AMP), and message passing using channel hardening (CHEMP). However, MIMO detection algorithms need to quantify multiple parameters to reduce system redundancy and improve system efficiency. For different algorithms, many parameters need to be quantified for hardware implementation.

[0003] However, the research on quantization optimization of MIMO detection algorithms is still relatively limited. However, quantization is a key step that directly affects hardware efficiency and detection accuracy. For MIMO detectors, uniform quantization (UQ) is usually used, and a previous work proposed the AHPQ algorithm optimized using deep reinforcement learning (DRL) (CN114897098A). However, there is still room for improvement in the bit-width compression obtained by this algorithm. Summary of the Invention

[0004] Object of the Invention: Aiming at the quantization problem of signal processing systems, the object of the present invention is to provide a signal processing quantization method and system based on a discrete grey wolf optimizer (DIGWO) to obtain a lower average bit width.

[0005] Technical Solution: To achieve the above object of the invention, the present invention adopts the following technical solutions:

[0006] A signal processing algorithm quantization method based on DIGWO, comprising the following steps:

[0007] Obtain the parameters to be quantified during the signal processing process;

[0008] Set the quantization bit width for each parameter to be quantified, including the fractional bit width;

[0009] Optimize the fractional bit width of each parameter using the discrete grey wolf optimizer DIGWO to obtain the final fractional quantization bit width; the feedback function of the DIGWO is established based on achieving a relative balance between the two objectives of reducing the bit width and improving signal processing performance.

[0010] Preferably, in the feedback function, the relative error between the unquantized performance and the quantized performance is defined, and the value-taking method of the feedback function is determined based on the relationship between the relative error and a preset threshold; when the relative error is greater than the preset threshold, a penalty term is added; the penalty term takes into account the incorrect data caused by quantization.

[0011] Preferably, when the relative error is less than the preset threshold, the feedback function is determined according to the exponential function of the average small quantization bit width.

[0012] Preferably, when the relative error is less than the preset threshold, the value of the feedback function F is Ae Bq , otherwise the value of the feedback function F is increased by a penalty term on the basis of Ae Bq ; where A and B are empirical parameters, and q is the average small quantization bit width.

[0013] Preferably, in the DIGWO algorithm, the fractional bit width of each parameter to be quantized is used as the position of the solution required by the algorithm. First, random initialization is performed to obtain the initial solution group before iteration; in each iteration, after calculating the feedback function values of all solutions, the optimal three solutions are selected, and then iteration is performed according to GWO, and the values obtained in each iteration are rounded down; finally, after the iteration is completed, the obtained bit width is the final quantized fractional bit width.

[0014] Furthermore, the quantization method of the signal processing algorithm based on DIGWO further includes quantizing the integer bit width of each parameter to be quantized, including:

[0015] Obtaining the accurate data of each parameter from the Monte Carlo simulation sample set D;

[0016] Setting the integer bit width, and recording the parameter samples not within the quantization range to obtain the set E;

[0017] Taking the ratio of the number of the set E to the set D being less than the preset threshold as the goal, and determining the minimum integer bit width.

[0018] In some embodiments, the signal processing process is MIMO detection, and the feedback function is expressed as:

[0019]

[0020] where P F is the bit error rate BER without any quantization strategy, P Q is the BER after quantization, ε is the set threshold, C is an empirical parameter, and b e is the number of error bits under quantization.

[0021] The present invention also provides a quantization system of a signal processing algorithm based on DIGWO, including:

[0022] A parameter acquisition module that acquires parameters to be quantized during the signal processing process;

[0023] A bit width setting module for setting the quantization bit width for each parameter to be quantized, including the fractional bit width;

[0024] And a fractional quantization bit width optimization module for optimizing the fractional bit widths of various parameters using the DIGWO algorithm to obtain the final fractional quantization bit width; the feedback function of the DIGWO algorithm is established based on achieving a relative balance between the two goals of reducing the bit width and improving the signal processing performance.

[0025] Furthermore, the system further includes an integer bit width quantization module for quantizing the integer bit width of each parameter to be quantized, including: obtaining the exact data of each parameter from the Monte Carlo simulation sample set D; setting the integer bit width, recording the parameter samples not within the quantization range to obtain the set E; and determining the minimum integer bit width with the goal that the ratio of the number of elements in set E to set D is less than a preset threshold.

[0026] The present invention also provides a computer program product, including a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the signal processing algorithm quantization method based on DIGWO are implemented.

[0027] Beneficial effects: The signal processing quantization method provided by the present invention quantizes the fractional bits of the parameters in the signal processing process through the relatively simple and effective Discrete Grey Wolf Optimizer (DIGWO) in the meta-heuristic algorithm. Experiments show that when applied to the Channel Hardening Enabled Message Passing (CHEMP) MIMO detector, DIGWO quantization achieves an average fractional bit width reduction of 41.8% compared to Uniform Quantization (UQ) and 24.9% lower than the existing algorithm Automatic Mixed Precision Quantization (AHPQ), and there is basically no performance loss. In addition, the quantization method based on DIGWO proposed by the present invention has universality and can be applied to quantize the signal processing algorithms of other communication systems. Description of the Drawings

[0028] Figure 1 It is a flowchart of the method according to an embodiment of the present invention.

[0029] Figure 2 It is a schematic diagram of the Grey Wolf Optimizer.

[0030] Figure 3 It is a BER comparison diagram of different algorithms and quantization methods in an embodiment of the present invention.

[0031] Figure 4 It is a comparison diagram of the fractional bit widths obtained by the existing AHPQ and DIGWO in an embodiment of the present invention. Detailed Embodiments

[0032] To make the objectives, technical solutions, and effects of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts belong to the scope of protection of the present invention.

[0033] As Figure 1 shown, a quantization method for a signal processing algorithm based on DIGWO disclosed in an embodiment of the present invention mainly includes the following steps: obtaining parameters to be quantized during the signal processing; setting quantization bit widths for each parameter to be quantized, including fractional bit widths; and optimizing the fractional bit widths of each parameter using a discrete grey wolf optimizer (DIGWO) to obtain the final fractional quantization bit widths.

[0034] In some specific embodiments, the signal processing process may be channel estimation, signal detection, signal decoding, etc. In this embodiment, MIMO detection is taken as an example for illustrative explanation, and the quantization method for MIMO detection based on DIGWO is specifically described.

[0035] Consider an N t × N r MIMO system, where N t is the number of transmit antennas and N r is the number of receive antennas. The input data stream is replicated into N t copies, modulated by M-QAM with a constellation diagram of Ω, and then transmitted by the antennas respectively. This model can be represented by the following formula:

[0036] y = Hx + n, (1)

[0037] where is the received vector, is the transmitted vector, is the channel matrix, and n is the noise. Assume that the channel is modeled as an independent and identically distributed (i.i.d.) Rayleigh fading channel, and the noise n is additive white Gaussian noise (AWGN) with a mean of zero and a variance of σ 2 . Assume that the receiver has perfect channel state information. Converting the above complex-valued model to the real domain, the following model can be obtained:

[0038]

[0039] where is the real or imaginary part of the M-QAM constellation diagram,

[0040] For the quantization of MIMO system detection, the quantization result can be divided into three parts: the sign part, the integer part, and the fractional part, occupying 1, m, and n bits respectively. Therefore, the quantization bit width of any parameter can be expressed as 1 - m - n. The quantization result will be within Q max and Q min , and is a multiple of the smallest quantization unit S closest to the original data, where Q max = 2 m - 2 -n , Q min = -2 m + 2 -n , S = 2 -n . The present invention mainly relates to the fractional part.

[0041] For the fixed-point quantization of any algorithm, we can extract the various parameters to be quantized, which is the basis of our quantization. For each parameter, the quantization bit width can be set as described above. We can set the integer bit width or fractional bit width sequence containing k parameters as a = [a 1 ,... a i ,... a k T , where each element in the sequence represents the quantization integer or fractional bit width of each parameter. Assuming that the quantization bit width can be infinite, in order to find the optimal quantization bit width, we need to find the optimal point in this k-dimensional space, and this point is the quantization bit width sequence to be obtained.

[0042] In an embodiment of the present invention, the integer part can use the method of Monte Carlo simulation to obtain the bit width, and the fractional part uses DIGWO for optimization to find the final quantization bit width. The following will be described in these two parts.

[0043] For the quantization of the integer part, the solution can be obtained from the perspective of probability statistics. We can assume a Monte Carlo simulation set D containing a sufficiently large number N of samples, and obtain the exact data u of each parameter from each simulation sample. Since the quantization result of each parameter is 1 - m - n, its data range is between Q max and Q min , and when estimating the integer bit width without knowing the quantization situation of the fractional part, we can assume the worst case n = 0. Under this assumption, the samples not within the range are recorded as set E, and our goal is to find the smallest m that satisfies , where ε 0 is the threshold, and card represents the number of samples in the set. Since the integer bit width bit width has a greater impact on the result, the threshold should not be set too large. After setting the threshold, the integer bit width size can be obtained through Monte Carlo simulation.

[0044] ​For the quantization of the fractional part, the DIGWO algorithm is adopted in the present invention. First, the Grey Wolf Optimizer (GWO) is introduced.

[0045] GWO is an optimization search method inspired by the hunting behavior of grey wolves. It has good convergence performance, fewer parameters, and is easy to implement. GWO follows the laws of nature. First, a group of "grey wolves" needs to be generated as the initial solution before the iteration based on this algorithm. Then, the initial solution needs to be iteratively updated. Specifically, as Figure 2 shown, the population members are ranked, and the top three optimal solutions are set as α, β, and γ respectively, while the other wolves are set as ω. The other wolves need to update their positions by surrounding these three optimal wolves.

[0046] Set the positions of the three optimal grey wolves as X α , X β and X γ , and let the original position of a certain other wolf be Then the first equation for algorithm update is

[0047]

[0048] where i = α, β, γ, L α , L β and L γ represent the distances between this wolf and the three optimal wolves, and W i is a random number in the range [0, 2]. The second equation is

[0049] X j = X i - R j ·D i (4)

[0050] where i = α, β, γ, j = 1, 2, 3, X 1 , X 2 and X 3 represent the degree to which this wolf is affected by the positions of the three optimal wolves, and R j can be calculated by the following formula

[0051] R j = 2a·r j - a (5)

[0052] where r j is a random number in the range [0, 1], and a is the convergence factor, which linearly decreases from 2 to 0 as the number of iterations increases.

[0053] Finally, the formula for updating the position of a grey wolf individual is

[0054] X = (X 1 + X2 +X 3 ) / 3 (6)

[0055] The above describes an iterative process for updating the positions of grey wolves. After the iteration is completed, the top three optimal wolves will be reselected for the next round of iteration until the preset number of iterations is reached.

[0056] Although the original GWO algorithm can find the optimal solution in the search space and converge to the result, there are problems when directly applying it to the parameter quantization of the MIMO detection algorithm. The original GWO algorithm is an algorithm in the continuous domain, while quantization needs to be carried out in the discrete domain. Therefore, the algorithm needs to be adapted from the continuous domain to the discrete domain.

[0057] The steps of the algorithm mainly adopt the following scheme:

[0058] First is the initialization. We can take the sizes of each fractional bitwidth as the positions of the solutions required by the GWO algorithm, X = [a 1 ,…a i ,…a k T , and initialize these solutions within a range. The initialization should be fully random and the number of solutions should be ensured. These initial solutions form the solution group before the first iteration.

[0059] Next, the most important function for the iterative requirement needs to be set - the feedback function. The purpose of designing this optimization algorithm is to achieve the smallest possible bitwidth for each parameter of the MIMO algorithm while obtaining the best possible performance. However, reducing the bitwidth and improving the performance are conflicting goals. Our feedback function is established based on achieving a relative balance between these two factors.

[0060] In the embodiment of the present invention, the relative error between the unquantized performance and the quantized performance is defined in the feedback function, and the value-taking manner of the feedback function is determined based on the relationship between the relative error and the preset threshold; when the relative error is greater than the preset threshold, a penalty term is increased; wherein, the penalty term considers the incorrect data caused by quantization. Optionally, when the relative error is less than the preset threshold, the feedback function is determined according to the exponential function of the average fractional quantization bitwidth. For example, when the relative error is less than the preset threshold, the value of the feedback function F is Ae Bq , otherwise the value of the feedback function F is increased by a penalty term on the basis of ; where A and B are empirical parameters, is the average fractional quantization bitwidth.

[0061] ​Taking the MIMO detection algorithm as an example, considering the performance of the MIMO detection algorithm, ensuring the performance of the original algorithm is a prerequisite for quantitative qualification. The most commonly used metric for evaluating the performance of the MIMO detection algorithm is the bit error rate (BER), which we use as our criterion. We define the BER obtained without any quantization strategy as P F , which serves as the criterion for judging quantization performance. The quantized BER value is defined as P Q , so the relative error E can be defined as (P Q - P F ) / P Q . When the relative error is less than the threshold ε, we set the function as where A and B are empirical parameters, is the average small quantization bit width. When the error is greater than the threshold, the function becomes where Cb e is the penalty term, C is also an empirical parameter, and b e is the number of error bits under this quantization. It can be seen that if the obtained BER is smaller, even if the relative error is greater than the threshold, this value will decrease relatively, thus ensuring the convergence of the function.

[0062]

[0063] For the application of the algorithm in other signal processing algorithms, the quantization and non - quantization performance errors such as bit error rate and frame error rate can also be used as the basis for designing the feedback function. What needs to be adjusted are the empirical parameters A and B and the penalty term setting.

[0064] After calculating the sizes of the feedback functions for all solutions, the optimal three solutions need to be selected, and then iterative calculations are performed according to GWO. Since the values obtained in each iteration are usually not integers, and quantization requires an integer bit width, a rounding operation is needed. We perform rounding down, which can obtain a lower bit width and is more beneficial for the algorithm iteration.

[0065] Finally, after the iteration is completed, the obtained bit width is the final quantized decimal bit width.

[0066] Experimental example:

[0067] In this experiment, a MIMO system with an antenna ratio of 128×8 was considered, and 16 - QAM modulation was used. By analyzing the calculation process of the CHENP algorithm, 14 variables were selected for quantization, and these variables are listed in the following table.

[0068] Table 1: Variable and its corresponding number table

[0069]

[0070] In this experiment, the parameters of the feedback function were set as A = 1, B = 1, C = 20, and ε = 0.05. After obtaining the results, the results were analyzed as follows:

[0071] The first comparison was the bit error rate (BER) performance of the MIMO detection algorithm quantized using different quantization methods. In Figure 3 , the MMSE, CHEMP, and FLOAT curves represent the BER performance of the MMSE, CHEMP, and CHEMP algorithms with unquantized parameters, respectively, while the DIGWO, AHPQ, and UQ curves represent the BER performance of the CHEMP algorithms quantized using different methods. As can be seen from Figure 3 , the BER performance of the CHEMP algorithm is better than that of the MMSE algorithm, and the BER performance of the CHEMP algorithm is basically not affected by different quantization results. Therefore, it can be concluded that these three quantization methods can relatively well maintain the performance of the MIMO detection algorithm.

[0072] Table 2: Comparison of fractional bit widths obtained by different quantization methods

[0073] Parameter <![CDATA[u k > <![CDATA[σ k > <![CDATA[σ w > J <![CDATA[e x > <![CDATA[v x > z <![CDATA[σ x > <![CDATA[u x > α δ γ <![CDATA[sum e > ρ Average bit width Percentage of average bit width in unified quantization UQ 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 100％ AHPQ 5 9 9 8 5 2 6 9 6 1 5 1 5 1 5.14 64.2％ DIGWO 9 9 9 8 0 0 4 9 5 0 0 0 0 1 3.86 48.2％

[0074] The second comparison was to compare the fractional bit widths obtained by different quantization methods. As shown in Table 2 and Figure 4 , the experiment compared three different quantization methods: UQ, AHPQ, and DIGWO. It can be seen that the DIGWO algorithm achieved the smallest average quantization bit width of 3.86, while UQ and AHPQ were 8 and 5.14, respectively. The average bit width of DIGWO was 41.8% and 24.9% lower than that of UQ and AHPQ, respectively. This indicates that DIGWO can achieve effective bit width compression.

[0075] In summary, DIGWO maintained the BER performance of the algorithm after quantization, while the obtained quantization bit width was lower than that of AHPQ. Therefore, DIGWO has better quantization effect.

[0076] Based on the same inventive concept, an embodiment of the present invention also discloses a signal processing algorithm quantization system based on DIGWO, including: a parameter acquisition module for acquiring parameters to be quantized during the signal processing process; a bit width setting module for setting a quantization bit width for each parameter to be quantized, including a fractional bit width; and a fractional quantization bit width optimization module for optimizing the fractional bit width of each parameter using the DIGWO algorithm to obtain a final fractional quantization bit width; the feedback function of the DIGWO algorithm is established based on achieving a relative balance between the two goals of reducing the bit width and improving the signal processing performance.

[0077] In addition, the system further includes an integer bit-width quantization module for quantizing the integer bit-width of each parameter to be quantized, including: obtaining the exact data of each parameter from the Monte Carlo simulation sample set D; setting the integer bit-width, and recording the parameter samples not within the quantization range to obtain set E; aiming at the ratio of the number of set E to the number of set D being less than a preset threshold, determining the minimum integer bit-width.

[0078] For the detailed working processes of the above-described modules, reference may be made to the corresponding processes in the foregoing method embodiments, which will not be elaborated herein. The division of the modules is only a logical function division, and there may be other division methods in actual implementation. For example, multiple modules may be combined or integrated into another system.

[0079] The embodiment of the present invention also discloses a computer program product, including computer programs / instructions, and when the computer programs / instructions are executed by a processor, the steps of the signal processing algorithm quantization method based on DIGWO are implemented. The program codes for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to the processor or controller of a general computer, a special computer or other programmable data processing devices, so that when the program codes are executed by the processor or controller, the steps of the method of the present invention are implemented. The program codes can be executed entirely on the machine, partially on the machine, executed partially on the machine and partially on a remote machine as an independent software package, or executed entirely on a remote machine or server.

Claims

1. A signal processing algorithm quantization method based on DIGWO, characterized in that: The steps include: Obtaining parameters to be quantified in the signal processing process; For each parameter to be quantized, set the quantization bit width, including the fractional bit width; The discrete grey wolf optimizer DIGWO is used to optimize the decimal bit width of each parameter to obtain the final decimal quantization bit width; the feedback function of the DIGWO is established based on achieving a relative balance between the two goals of reducing the bit width and improving the performance of the signal processing algorithm.

2. The DIGWO-based signal processing algorithm quantization method according to claim 1, characterized in that: In the feedback function, a relative error between the unquantified performance and the quantized performance is defined, and a value determination method of the feedback function is determined based on a relationship between the relative error and a preset threshold; When the relative error is greater than the preset threshold, a penalty term is added; The penalty term takes into account erroneous data caused by quantization.

3. The signal processing algorithm quantization method based on DIGWO according to claim 2, characterized in that: When the relative error is less than a preset threshold, the feedback function is determined according to an exponential function of the average decimal quantization bit width.

4. The signal processing algorithm quantization method based on DIGWO according to claim 3, characterized in that: When the relative error is less than the preset threshold, the value of the feedback function F is Otherwise, the value of the feedback function F is Add a penalty term on the basis of ; where A and B are empirical parameters, is the average decimal quantization bit width.

5. The DIGWO-based signal processing algorithm quantization method according to claim 1, characterized in that: The DIGWO algorithm uses the decimal width of each parameter to be quantized as the position of the solution required by the algorithm, and first performs random initialization to obtain an initial solution group before iteration; In each iteration, after calculating the feedback function values ​​of all solutions, the best three solutions are selected, and then iterated according to GWO, and the value obtained in each iteration is rounded down; finally, after the iteration is completed, the bit width obtained is the final quantization decimal bit width.

6. The DIGWO-based signal processing algorithm quantization method according to claim 1, characterized in that: It also includes quantizing the integer bit width of each parameter to be quantized, including: Obtain accurate data for each parameter from the Monte Carlo simulation sample set D; Set the integer bit width and record the parameter samples that are not within the quantization range to obtain a set E; With the goal of making the ratio of the number of set E to the number of set D less than a preset threshold, the minimum integer bit width is determined.

7. The DIGWO-based signal processing algorithm quantization method according to claim 1, characterized in that: The signal processing process is MIMO detection, and the feedback function is expressed as: Among them, P F is the bit error rate BER without any quantization strategy, P Q is the quantized BER, ε is the set threshold, A, B, C are empirical parameters, is the average fractional quantization bit width, b e is the number of bits of error under quantization.

8. A signal processing algorithm quantization system based on DIGWO, characterized in that: include: A parameter acquisition module, which acquires the parameters to be quantified during the signal processing process; A bit width setting module, used to set the quantization bit width for each parameter to be quantized, including the decimal bit width; And, a decimal quantization bit width optimization module is used to optimize the decimal bit width of each parameter using the DIGWO algorithm to obtain the final decimal quantization bit width; the feedback function of the DIGWO algorithm is established based on achieving a relative balance between the two goals of reducing the bit width and improving the signal processing performance.

9. The DIGWO-based signal processing algorithm quantization system according to claim 8, characterized in that: It also includes an integer bit width quantization module for quantizing the integer bit width of each parameter to be quantized, including: obtaining accurate data of each parameter from a Monte Carlo simulation sample set D; setting the integer bit width, recording parameter samples that are not within the quantization range to obtain a set E; and determining the minimum integer bit width with the goal of making the ratio of the number of set E to set D less than a preset threshold.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the DIGWO-based signal processing algorithm quantization method according to any one of claims 1 to 7 are implemented.

Images