A MIMO Detection System Based on MCMC
By using random circuit units in the MIMO detection system to replace the nonlinear computing units in the traditional MCMC algorithm, parallel sampling is realized, solving the problems of delay and high complexity during calculation of the traditional MIMO detection algorithm, and improving the energy efficiency ratio of the system Compare.
Patent Information
- Application Number
- CN202211342181.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-31
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-10-31
AI Technical Summary
The traditional MIMO detection algorithm based on MCMC algorithm has a prolonged calculation time and is highly complex, so complex hardware is required to design the calculation.
A MIMO detection system based on random circuit units is adopted, and the Exp function, divider and squarer in traditional MCMC is replaced by the LLR operation unit, and non-linear calculation is performed, and it is changed to parallel sampling to reduce the calculation delay and complexity.
While ensuring detection accuracy, the complexity of the algorithm and hardware power consumption are greatly reduced, and the energy efficiency ratio is improved.
Smart Images

Figure CN115733525B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of communication technology, and in particular to a MIMO detection system based on MCMC. Background Art
[0002] Existing research results show that random computing can achieve very high efficiency in many applications of solving linear calculations, including digital signal processing, image processing, and artificial intelligence. Gaines proposed logical calculations of random bit streams. Based on Gaines' idea, some studies proposed the implementation of specific arithmetic functions based on random computing, including multiplication, addition, and linear polynomials, to achieve ultra-low complexity and shorter critical paths in terms of full adders and multiplexers. However, processing nonlinear functions is more challenging. In 2001, Brown and Card proposed a scheme based on random computing with a wired state machine for solving nonlinear functions such as Tanh and exponentials.
[0003] Multiple-input multiple-output (MIMO) technology is a technology that uses the degrees of freedom of transmission space to multiply the data transmission rate and reliability. Compared with the single-input single-output (SISO) system, the MIMO system has two major performance gains: diversity gain and multiplexing gain, which are used to measure the reliability of the system and the spatial channel capacity respectively. The research on detection algorithms in MIMO technology has been a hot topic in recent decades, because the performance and complexity of the signal processing at the receiving end directly determine the quality and application prospects of the entire communication system. The Markov chain Monte Carlo method is a modern random simulation method that can obtain samples that obey the required distribution in a convenient way, reduce the cost of random number generation and promote the application of Bayesian methods. The core idea of this algorithm is to use the decoder feedback information to calculate the log-likelihood ratio of each bit, which was first proposed by Wang and Poor. This model can approach the Shannon limit of the MIMO channel through multiple joint iterative detection and decoding. Since the complexity of a Markov chain Monte Carlo (MCMC) algorithm based on Bayesian reasoning grows linearly or polynomially with the number of transmitting antennas and the modulation order, it has received widespread attention in the field of MIMO detection. However, there are still the following problems in the MCMC-based soft output MIMO detection algorithm: 1. The traditional MCMC-based MIMO detection calculation delay is very long, and a considerable iteration depth is required to make the algorithm converge; 2. The traditional MCMC-based MIMO detection calculation complexity is relatively high. In the traditional algorithm, it is often necessary to calculate the square of the Euclidean distance between constellation points, division, addition and sorting, etc. Among them, the square and division as non-linear calculations require very high complexity, so it is also necessary to design more complex hardware to complete the calculation. Summary of the invention
[0004] The purpose of the present invention is to overcome the shortcomings of the traditional MIMO detection algorithm based on the MCMC algorithm in the prior art, such as long delay and high computational complexity, and to provide a MIMO detection system and method based on MCMC.
[0005] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions:
[0006] A MIMO detection system based on MCMC includes an LLR (log likelihood ratio) operation unit, wherein the LLR operation unit includes a multiplication and addition operation, a comparator and a random circuit unit;
[0007] The random circuit unit is used to replace the Exp function, divider and squarer in the traditional MCMC to complete nonlinear calculation.
[0008] By adopting the above technical solution, based on the random circuit unit, by considering the values entering the MIMO detection system as different independent variables of the calculation function, and the output values as the dependent variables of the calculation function, the input bit stream is finally mapped to the expected output bit stream. This algorithm can enhance the configurability of the entire system, thereby configuring more nonlinear functions, and changing the sampling method of the traditional MCMC algorithm to parallel sampling, which greatly reduces the calculation delay of the detection algorithm. Ultimately, while ensuring the detection accuracy, the complexity of the algorithm and the required hardware power consumption are greatly reduced, and ultimately the energy efficiency ratio and surface area efficiency ratio are greatly improved.
[0009] As a preferred solution of the present invention, the random circuit unit includes: N θ gates, a FSM and a CPT gate;
[0010] The output ends of the N θ gates are connected to the input end of the FSM, and the output end of the FSM is connected to the input end of the CPT gate.
[0011] The random circuit unit can calculate many simple one-dimensional nonlinear functions, such as EXP, SQRT, Log function, etc.
[0012] As a preferred solution of the present invention, the FSM is a two-dimensional state machine including 8 states.
[0013] As a preferred solution of the present invention, the CPT gate is used to sample from an input distribution with a K value.
[0014] As a preferred solution of the present invention, the CPT gate includes M θ gates, wherein N and M are positive integers, and N=M.
[0015] As a preferred solution of the present invention, the gate is used to obtain output sample values, namely 0 and 1, by comparing the relationship between the input probability value and the size of the random entropy source.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: based on a random circuit unit, by regarding the values entering the MIMO detection system as different independent variables of the calculation function, regarding the output values as the dependent variables of the calculation function, adding a θ gate and a CPT gate, the specific numerical values are converted into a random bit stream, thereby greatly reducing the complexity of the operation and reducing the operation cycle, and the random circuit unit can calculate many simple one-dimensional nonlinear functions, such as EXP, SQRT, Log function, etc. This algorithm can enhance the configurability of the entire system, thereby configuring more nonlinear functions, and changing the sampling method of the traditional MCMC algorithm to parallel sampling, which greatly reduces the calculation delay of the detection algorithm. Finally, while ensuring the detection accuracy, the complexity of the algorithm and the required hardware power consumption are greatly reduced, and ultimately the energy efficiency ratio and surface area efficiency ratio are greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 A structural diagram of a MCMC-based MIMO detection system according to embodiments 1 and 2 of the present invention;
[0018] Figure 2 A state transition diagram of a 2×4 FSM of a MIMO detection system based on MCMC as described in Embodiments 1 and 2 of the present invention;
[0019] Figure 3 Two circuit implementations of the θ gate of a MIMO detection system based on MCMC as described in Example 1 of the present invention and corresponding probability truth tables;
[0020] Figure 4 A circuit implementation of a conditional probability sampling gate (CPT gate) of a MIMO detection system based on MCMC as described in Example 1 of the present invention;
[0021] Figure 5 This is a structural block diagram of an MCMC-MIMO detection algorithm of an MCMC-based MIMO detection system according to Embodiment 1 of the present invention;
[0022] Figure 6 A graph of average relative error of LLR calculated by using MM-FSM in a MIMO detection system based on MCMC according to Embodiment 3 of the present invention;
[0023] Figure 7 BER diagram of each method under the condition of 2×2 antenna number of a MCMC-based MIMO detection system described in Example 3 of the present invention;
[0024] Figure 8BER diagram of each method under the condition of 4×4 antenna number of a MIMO detection system based on MCMC as described in Example 3 of the present invention;
[0025] Fig. 9 BER diagram of each method under the condition of 2×4 antennas in a MIMO detection system based on MCMC as described in Example 3 of the present invention;
[0026] Fig.10 A BER (2×2 number of antennas) diagram of MM-FSM and MMSE methods under different channel estimation error conditions for a MCMC-based MIMO detection system described in Example 3 of the present invention;
[0027] Fig.11 A BER (4×4 antenna number) diagram of the MM-FSM method under different channel estimation error conditions of a MCMC-based MIMO detection system described in Example 3 of the present invention;
[0028] Fig.12 A BER (2×4 antenna number) diagram of the MM-FSM method under different channel estimation error conditions of a MCMC-based MIMO detection system described in Example 3 of the present invention;
[0029] Fig.13 This is a performance comparison diagram between MM-FSM and MMSE under the condition of 2×2 antennas for a MIMO detection system based on MCMC according to Embodiment 3 of the present invention;
[0030] Fig.14 This is a performance comparison diagram of MM-FSM and MMSE under the condition of 2×4 antennas for a MCMC-based MIMO detection system according to Embodiment 3 of the present invention;
[0031] Fig.15 This is a performance comparison diagram of MM-FSM and MMSE under BPSK modulation of a MIMO detection system based on MCMC described in Example 3 of the present invention;
[0032] Fig.16 This is a performance comparison diagram of MM-FSM and MMSE under 16QAM modulation of a MCMC-based MIMO detection system described in Example 3 of the present invention;
[0033] Fig.17 The performance diagram of the MMSE algorithm under BPSK modulation and the MM-FSM plus the channel condition number of the MCMC-based MIMO detection system described in Example 3 of the present invention;
[0034] Fig.18This is a performance diagram of MM-FSM and MMSE plus a channel condition number under 16QAM modulation of a MCMC-based MIMO detection system described in Example 3 of the present invention;
[0035] Fig.19 This is a performance diagram of MM-FSM under BPSK modulation and MMSE plus channel estimation error of a MCMC-based MIMO detection system described in Example 3 of the present invention;
[0036] Fig. 20 This is a performance diagram of MM-FSM under 16QAM modulation and MMSE plus channel estimation error of a MCMC-based MIMO detection system described in Example 3 of the present invention; DETAILED DESCRIPTION
[0037] The present invention is further described in detail below in conjunction with test examples and specific implementation methods. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.
[0038] Example 1
[0039] A MIMO detection system based on MCMC, Figure 1 As shown, it includes an LLR operation unit, and the LLR operation unit includes a multiplication and addition operation, a comparator and a random circuit unit;
[0040] The random circuit unit is used to replace the Exp function, divider and squarer in the traditional MCMC to complete nonlinear calculation.
[0041] The random circuit unit includes: N θ gates, a FSM and a CPT gate;
[0042] The output ends of the N θ gates are connected to the input end of the FSM, and the output end of the FSM is connected to the input end of the CPT gate.
[0043] The input before the θ gate is a random number on (0,1), which is converted into a random bit stream after passing through the θ gate.
[0044] like Figure 2 As shown, the FSM is a two-dimensional state machine including 8 states.
[0045] The CPT gate is used to sample from a distribution with K values of the input.
[0046] like Figure 4 As shown, the CPT gate includes M θ gates, wherein N and M are positive integers, and N=M.
[0047] The gate is used to obtain output sample values, namely 0 and 1, by comparing the relationship between the input probability value and the size of the random entropy source.
[0048] Specifically, the sampling process of the θ gate can be implemented by a single comparator, and the threshold of the comparator is determined by the weight θ value of the input port. By comparing the relationship between the input probability value and the size of the random entropy source, the output sample value, that is, 0 and 1, can be obtained, indicating whether the event occurs or not.
[0049] like Figure 3 As shown in the figure, a and b respectively represent the θ gates implemented using two types of comparators. From the probability truth table, we can see from c and d that the difference between the two structures lies in whether the probability values 0 and 1 at both ends can be taken. The probability of the θ gate using the < comparator outputting 0 when the input is the minimum value is 1, while the probability of the θ gate using the ≤ comparator outputting 1 when the input is the maximum value is 1.
[0050] The conditional probability gate (CPT gate) samples from a distribution of K values of the input. This is equivalent to throwing a set of biased k-sided dice and then choosing which dice value to use based on the input.
[0051] like Figure 1 In the example shown, a random number generator (θ gate), an M-1 multiplexer (FSM), and M θ gates are used to form a conditional sampling gate (CPT gate). During the operation of the FSM, the state machine transmits the generated state number t to the CPT gate and selects the corresponding Y bit as the output.
[0052] By using the CPT gate, the input LLR is converted into a probability bit stream through a simple structure, so as to perform sampling, thereby greatly reducing the complexity of the sampling process.
[0053] The calculation method of LLR is as follows: First, before the MCMC detection algorithm starts, use L A1 =L A1 ×σ 2 Normalize the prior information, and rewrite the formula as follows:
[0054] λ1(b i )=-(||y-Hs(b i =+1)|| 2 -||y-Hs(b i =-1)|| 2 )+L A1 (b i )
[0055] The bit update process is directly performed in the logarithmic domain, and the update rule is
[0056]
[0057] Accordingly, in the symbol-by-symbol MCMC algorithm, taking state θ1 as the comparison object, the conditional log-likelihood ratio calculation formula can be rewritten as:
[0058]
[0059] After Gibbs sampling is completed, substitute the bit vector or symbol vector into the external information calculation formula:
[0060]
[0061] Finally, the external information is denormalized, that is, L E1 (b i )=L E1 (b i ) / σ 2 .
[0062] When performing detection calculations, it is inevitable to preprocess dist = || y-Hs (b 0 )|| 2 When the LLR is 1 / N0, it still involves complex nonlinear operations. Therefore, this paper combines the MM-FSM nonlinear function calculation unit introduced above to reduce the complexity of this part. This paper has done a lot of simulations to verify that the MCMC-MIMO detection based on the MM-FSM probability calculation 65 can further reduce the complexity without losing performance. After using the MM-FSM operator to replace the calculation unit in the LLR, the following is obtained. Figure 5 The algorithm structure block diagram is shown.
[0063] As shown in the figure, the Euclidean distance and LLR in the constellation diagram are calculated by MM-FSM probability. The LLR obtained by probability calculation is directly fed back to the Gibbs sampling unit and is no longer updated by the Maxlog criterion. As mentioned in the previous chapter, the MM-FSM algorithm requires a certain length of bit stream to achieve proficiency. When performing MCMC-MIMO detection, the calculation can be converged by algorithm iteration without increasing the bit stream length.
[0064] By adopting the above technical solution, based on the random circuit unit, by considering the values entering the MIMO detection system as different independent variables of the calculation function and the output values as the dependent variables of the calculation function, the input bit stream is finally mapped to the expected output bit stream. In addition, the random circuit unit can calculate many simple one-dimensional nonlinear functions, such as EXP, SQRT, Log function, etc. This algorithm can enhance the configurability of the entire system, thereby configuring more nonlinear functions, and changing the sampling method of the traditional MCMC algorithm to parallel sampling, which greatly reduces the calculation delay of the detection algorithm. Finally, while ensuring the detection accuracy, the complexity of the algorithm and the required hardware power consumption are greatly reduced, and the energy efficiency ratio and surface area efficiency ratio are finally greatly improved.
[0065] Example 2
[0066] This embodiment is a specific derivation process of embodiment 1, using a random sampling gate (θ gate) to convert the probability value into a random bit stream. When the number of samples is large enough, the theoretical result can be approximated. However, it is impossible to obtain the expected nonlinear distribution from the uniform distribution through the random sampling gate circuit, so an additional structure needs to be added to the output bit stream. Complex calculations are performed on the random bit stream using fairly simple sequential logic, in which the single-drive FSM proposed by Brown and Card is replaced with a multi-drive FSM and an additional CPT gate to implement more nonlinear functions and Tanh functions in addition to Exp.
[0067] Random circuit unit structure and state transitions such as Figure 1 As shown. The structure consists of two θ gates, a FSM and a CPT gate. There are 8 states in the two-dimensional FSM. The input of this FSM is two independent random bit streams X and K, where X is the independent variable bit stream and K is the activation bit stream. The driving vector of this FSM is (X, K). Each bit in X and K will be 1, with probabilities Px and PK respectively. Since the probabilities Px and PK are fixed, the state transition process in the FSM can be considered as a time-homogeneous two-dimensional Markov chain. The state transition of this FSM is performed according to the driving bits of X and K, as shown Figure 2 shown.
[0068] For an M-state FSM, any state can be selected as the initial state. After a certain number of state transitions starting from the real-time state St, the output of the FSM can be obtained. The output of the FSM is the state number T at this time, t∈{0,...,M-1}. According to the steady-state principle of the Markov chain, the probability of each state in the FSM transitioning to another state is equal to the probability of the adjacent state transitioning to it at this time. The state number T is used as the output of the FSM and input into the CPT gate, and then an input end is selected with probability Wt. When a state transition occurs in the FSM, the real-time state number t is input into the multiplexer, and then the output random bit stream is selected according to the input number t. This output is the output of the entire system, which is called Py. According to the previous theory, if the parameter K and the conditional parameter W are set according to a preset probability and configured into the θ gate and the CPT gate, various nonlinear functions can be realized. Use the random bit stream sequence Y with probability PY as the output of the entire system. Here, PY=F(Px), where F(Px) is the target function to be calculated. Since PY is an equation about Px, PK, and PW, it is very important to configure PK and PWt in the computing architecture. The activation parameter K and the condition parameter W are both referred to as the parameters of the objective function in this paper. Y The expression is used to approximate the objective function F(P x ). Define the estimated error as ε:
[0069]
[0070] Among them, F(P x ) is the objective function, P Y is the output value of the system.
[0071] The goal is to reduce the error in the above formula to the minimum. In this way, when the optimal values of parameters PK and Pwt are obtained, the corresponding nonlinear function can be effectively configured.
[0072] Configuration method of parameters used by MM-FSM:
[0073] Assume that the system reaches the stable state St with probability Pt (or the probability that the output of the whole system is t is St), where Pt is an equation about PX and PK. The equality between Pt, PX and PK has been proved. The objective function F(Px) is expressed by PX as the independent variable function. Let the number of rows of FSM be M, the number of columns be N, and the quotient and remainder of M divided by N be i and j respectively. So we get an equation about M and N:
[0074] M=i×N+j
[0075] i and j are two constants used to limit the numerical relationship between rows and columns. These two constants are needed in the subsequent parameter calculation process. Since j is the remainder of a constant divided by 2, j can only be 0 or 1, which means that j represents the parity of the column number N. When the system reaches a steady state, we have:
[0076]
[0077] in
[0078]
[0079] as well as
[0080]
[0081] Where N is the number of states of the Markov chain, and when the system reaches a steady state, the transition probability between the two states is equal. That is:
[0082] P i×N+j ·(1-P X )·(1-P K )=P i×N+j-1 ·P X ·P K .
[0083] P i×N+j ·(1-P X )·P K =P (i-1)×N+j ·P X ·(1-P K ).
[0084] Because the sum of all individual state probabilities Pt must be 1, we have:
[0085]
[0086] After using X and K as the driving signals of the MM-FSM, the state number is used as the selection signal of the CPT gate. Assuming that the specific state of the MM-FSM is St, the channel connecting Wt and Y will be selected in the CPT gate, which selects the bit determined by the output t of the MM-FSM from Wt. The output Py is given by:
[0087]
[0088] The output value P YSubtract F(PX) to get the error value φ, which is the object of optimization. The values of Pwt and Pk can be adjusted to minimize the value of φ. It is found that this is a multivariate convex optimization problem. As long as the optimization objective is written in a certain form, the optimization method in Matlab can be used to minimize φ. The estimated error is defined in equation (1). Expand equation (1) and rewrite ε as
[0089]
[0090] Since the value of the first term ε depends entirely on T(PX), the first term is a constant. Therefore, the optimization object of becomes the following objective function φ:
[0091]
[0092] Next, we will describe how to configure PK and PWt. Define vector b, vector c, and vector H as:
[0093]
[0094]
[0095] H=[H0,H1,...,H MN-1 ] T ,
[0096] Ht is a column vector in H:
[0097]
[0098] Therefore, φ can be rewritten as
[0099] φ=b T Hb+2c T b
[0100] It can be proved that the above formula is a convex function. Therefore, the values of Pk and PWt can be selected to achieve the minimum configuration objective function parameters:
[0101]
[0102] Take a square root function as an example:
[0103]
[0104] In this case, the random variable is set to PX, so the equation can be rewritten as:
[0105]
[0106] PK and PWt were calculated using the above method. The final results are listed in Table 1.
[0107]
[0108] Table 1 Parameters used to calculate the square root function
[0109] Therefore, when the parameters are set according to the results of Table 1, a square root function can be implemented.
[0110] Example 3
[0111] The MIMO detection method based on MCMC requires the calculated and configured parameters k and w before the operation. The method adopted is to traverse k on (0,1) with a step size of 0.01, and calculate the error between the output Py and the objective function T(PX), and take the smallest k and the corresponding wt. The final output Py is calculated by two criteria:
[0112] When the Markov process reaches a steady state, the probability of transitioning out of each state in the state machine is equal to the probability of transitioning into each state.
[0113] When the Markov process reaches a steady state, the sum of the transition probabilities between states in the state machine is 1.
[0114] Finally, the optimal parameters are obtained through matrix operations and multivariate function optimization methods. The parameters of the operators required in the MCMC detection algorithm and their average calculation relative errors are given (the input bit stream length is 128 bits):
[0115]
[0116] Table 2 Calculation parameters and average relative errors of several typical functions
[0117] Note: When calculating the 1 / x function, the input data Px value needs to be scaled to around 0.3 using the shift method. The calculation process has a large error value in other ranges.
[0118] 128 x bits and k bits are input to the FSM system respectively. When the input bit stream length is 128, the average relative error is calculated to be within 10%. Although the calculated error has not reached convergence, this relative error is sufficient for MCMC-MIMO detection.
[0119] In the calculation of LLR in the algorithm, several nonlinear functions are combined and the average relative error is analyzed. The results are as follows: Figure 6 As shown:
[0120] from Figure 6In the figure, we can see that when the input bit stream length reaches about 400, the calculated average relative error drops below -25dB, which is good enough for MCMC-MIMO detection. In fact, when performing MIMO detection of low-order modulation, only 200, 100, or even 50 iterations are needed to achieve good detection performance.
[0121] Detection performance and complexity analysis:
[0122] The simulation conditions for MCMC-MIMO detection are mainly small-scale antenna number (2×2, 2×4 and 4×4) MIMO detection under BPSK, QPSK and 16QAM modulation orders. This is because the search space is too large under high-order modulation, resulting in the inability to achieve convergence. The problem of too large search space will be solved in future work, so that the method can be applied to MCMC-MIMO detection under high-order modulation.
[0123] 1. Detection performance analysis, such as Figure 7 shown.
[0124] Using the sampling gate and MM-FSM operation module introduced above, simulation detection operations are performed on models with different channel estimation errors, different antenna numbers, and different modulation orders. This method is referred to as MM-FSM detection. The performance and complexity of the MM-FSM algorithm are compared with the traditional MCMC algorithm and the MMSE detection algorithm.
[0125] Figure 7 This is a comparison chart of MIMO detection performance under QPSK modulation when the number of transmitting and receiving antennas is 2×2. It can be seen from the figure that as the signal-to-noise ratio increases, the bit error rates of the MM-FSM algorithm, floating-point MCMC algorithm, and MMSE algorithm all show a downward trend. The BER of the MMSE algorithm reaches 2×10 -3 , which is better than the MM-FSM algorithm with an iteration depth of 10. The MM-FSM algorithm with an iteration depth of 10 converges faster, but reaches the error level very quickly, and the overall effect is not good. However, when the iteration depth is increased to 20, the detection effect is better. When SNR = 24, the BER of this method can reach 9×10 -4 , which is much better than the MMSE algorithm. When the iteration depth is further increased, the bit error rate is further reduced. When the iteration depth it=50, the BER of the MM-FSM algorithm is even lower than that of the floating-point based MCMC algorithm.
[0126] The simulation was also carried out under the condition of 4×4 antenna number. The simulation results are as follows Figure 8 As shown:
[0127] Depend on Figure 8It can be seen that the BER of the MMSE algorithm continues to decrease as the signal-to-noise ratio increases, and finally reaches 2×10 -3 The BER of the MCMC method based on MM-FSM probability calculation and the MCMC algorithm based on floating point both dropped rapidly at the beginning and reached a flat floor at SNR = 16dB. The BER of the MCMC method based on floating point was lower than that of the MMSE method at all signal-to-noise ratios, and the BER of the MM-FSM probability calculation method with an iteration depth of 512 was lower than that of the floating point MCMC method after 20 iterations, and was also lower than that of the MMSE. It can be seen that under the condition of 4×4 antennas, the MM-FSM probability calculation method can achieve good performance when the iteration depth is large. Although this method has a long delay, it has excellent performance and low complexity.
[0128] In addition, the simulation was also performed for the case where the number of transmitting and receiving antennas is asymmetric. The results are as follows: Fig. 9 As shown:
[0129] Depend on Fig. 9 It can be concluded that MMSE performs better when the number of antennas is 2×4, and can quickly reach a very low bit error rate. After 20 iterations of the floating-point MCMC method, the final BER converges to 2×10 -5 After 50 iterations of the proposed MM-FSM probability calculation method, the final BER converges to 3×10 -4 , and finally converged to 5×10 after 128 iterations. -5 Although the BER of the MM-FSM method is not as good as that of the MMSE and floating-point based MCMC methods, when the iteration depth is large enough, the MM-FSM can reach or even exceed the performance of the MMSE method.
[0130] The above simulations are performed under ideal channel conditions, while real channels have estimation errors. The performance of the MM-FSM algorithm under different channel error conditions is compared. The results are as follows: Fig.10 As shown:
[0131] In addition, the estimation error of the channel is changed under the condition of 2×2 antenna number, and the detection performance of the MMSE algorithm is obtained. Fig.12 It can be seen that when the channel estimation error increases, the BER increases significantly, and even when the estimation error is large, the MMSE algorithm cannot be leveled. Therefore, it is determined that under non-ideal channel conditions, the detection performance of MM-FSM is better than that of MMSE.
[0132] The simulation results when the number of antennas is 2×2 are as follows Figure 10 to 4 ×10 -5 When the channel estimation variance is 0.01, the BER can reach 10 at a large signal-to-noise ratio. -3When the channel estimation variance increases, the detection performance will be further degraded.
[0133] The simulation results when the number of antennas is 4×4 are as follows Fig.11 When the channel is an ideal channel with unbiased estimation, the BER of MM-FSM detection can reach 2×10 at a large signal-to-noise ratio. -2 , while the detection performance when the channel variance is 0.01 is equivalent to the unbiased estimation. In addition, when the channel estimation variance increases, the detection performance further decreases.
[0134] The channel condition number of the matrix is changed to observe the detection performance of MM-FSM. Traditional detection methods often degrade in performance as the channel condition number increases, so detection is performed under different channel condition numbers. The detection results are shown in Figure 2. Fig.13 and 14 As shown:
[0135] like Fig.12 and 13 As shown in the figure, the channel condition number is changed under the conditions of 2×2 and 2×4 antenna numbers, and the curve shown in the figure above is obtained after simulation. It can be observed that when the channel condition number increases, the correlation of the channel increases, so the detection performance inevitably degrades. When the number of MM-FSM iterations is 50, the detection performance is better than MMSE. The degree of degradation of the two is similar, but the overall BER of MM-FSM is low. When the number of iterations increases, the BER of MM-FSM will be further reduced.
[0136] In addition, the modulation mode of the system was changed to BPSK and 16QAM. The simulation results are as follows Fig.15 As shown:
[0137] Under BPSK modulation, both the floating-point MCMC detection algorithm and the probability-based calculation BER are lower than MMSE. Fig.16 The results are shown in Figure 10 for the floating-point MCMC detection algorithm and 20 and 50 for the MM-FSM detection algorithm. When the SNR reaches 24 dB, the BER of the MM-FSM detection algorithm can be reduced to 5×10 -5 The BER of MMSE can only be reduced to 2×10 -4 , the difference between the two is about 12dB.
[0138] Under 16QAM modulation, the computational performance of the MCMC detection algorithm is better than that of MMSE. When the SNR increases from 16 to 20, the computational performance of the MM-FSM algorithm with 128 iterations is on par with that of MMSE, but when the iterations are 256 and 512, the computational performance exceeds that of the MMSE algorithm, and the detection error is reduced by 4.86dB and 6.29dB respectively.
[0139] The performance of MM-FSM and MMSE is verified by adding channel condition number and channel estimation variance under different modulation orders. Fig.17 As shown,
[0140] like Fig.19 As shown in Figure 1, under BPSK modulation, the MM-FSM algorithm is iterated 50 times. Its performance is roughly equivalent to that of the MMSE algorithm. It can be inferred that when the number of iterations increases, the performance of the BPSK algorithm will be better than that of the MMSE algorithm under different channel estimation errors.
[0141] like Fig. 20 As shown in the figure, the detection performance degradation degree of the MM-FSM algorithm and the MMSE algorithm is similar under 16-QAM modulation as the channel estimation variance increases, but the overall performance of MM-FSM is better than MMSE after 256 iterations.
[0142] 2. Complexity analysis:
[0143] When the throughput of traditional MCMC-MIMO detection and MM-FSM detection is the same, MM-FSM can be used to perform 4 times the number of computational iterations of traditional MCMC detection. The complexity of the traditional MCMC method and the proposed MCMC method based on probability calculation is evaluated using the number of basic gates equivalent to the number of operators based on FPGA area overhead. It is assumed that the comparator has unit complexity, the adder has twice the overhead of the comparator, the multiplier has 10 times the complexity of the adder, and the divider and square root have 4.5 times and 3.8 times the overhead of the multiplier, respectively. The computational complexity of the complex multiplier is the sum of the overheads of 4 real multipliers and 2 adders. The results are shown in Table 3.
[0144]
[0145] Table 3. Comparison of the complexity of the traditional MCMC algorithm and the probabilistic MCMC algorithm (the number of base gates in the critical path)
[0146] The overall complexity is the computational complexity overhead when M=N=2, Q=2, |L|=50. The MCMC algorithm in the table is a symbol-by-symbol MCMC detection algorithm. The results show that the MCMC detection algorithm based on probability calculation is 47.63% of the complexity of the traditional MCMC algorithm. In other words, the MCMC algorithm based on probability calculation can greatly reduce the complexity.
[0147] When it comes to small-scale MIMO systems, probability calculation has the characteristics of low complexity, low power consumption, and high fault tolerance. Therefore, this chapter applies probability calculation based on MM-FSM to MCMC-MIMO detection and achieves good results. First, this chapter proposes a combination of high-throughput, low-complexity MM-FSM probability calculation and MCMC-MIMO detection algorithm. At the same time, the pipeline technology is used to enable preprocessing and Gibbs sampling to be performed simultaneously, which speeds up convergence and improves throughput.
[0148] In future research, we will focus on large-scale MIMO situations. We can improve the efficiency of spatial search by reducing the search space, thereby increasing the convergence speed of the detection algorithm.
[0149] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A MIMO detection system based on MCMC, comprising an LLR operation unit, characterized in that: The LLR operation unit includes a multiplication and addition operation, a comparator and a random circuit unit; The random circuit unit is used to replace the Exp function, divider and squarer in the traditional MCMC to complete the nonlinear calculation; The random circuit unit includes: N θ gates, a FSM and a CPT gate; The output ends of the N θ gates are connected to the input end of the FSM, and the output end of the FSM is connected to the input end of the CPT gate; The CPT gate is used to sample from an input distribution with a K value; The K value is determined by traversing k on (0,1) with a step size of 0.01, calculating the error between the output of the detection system and the objective function T(PX), and taking the k that minimizes the error as the K value.
2. The MCMC-based MIMO detection system according to claim 1, characterized in that: The FSM is a two-dimensional state machine including 8 states.
3. The MCMC-based MIMO detection system according to claim 2, characterized in that: The CPT gate includes M θ gates, wherein N and M are positive integers, and N=M.
4. The MCMC-based MIMO detection system according to claim 1, characterized in that: The θ gate is used to obtain output sample values, namely 0 and 1, by comparing the relationship between the input probability value and the size of the random entropy source.
