TW309675B - Method and apparatus for complex fuzzy signal processing - Google Patents

Abstract

A method for complex fuzzy signal processing includes the following steps:1) initialization mode: setting all the adjustable parameters to zero;2) training mode: utilizing the CANFIS algorithm to adjust the forward and the backward parameters according to the pre-stored training reference signal. When the error converges, the mode goes to the decision guiding mode; 3) decision guiding mode: obtaining the inference output signal from fuzzy inference and quantilizing it to obtain the decision output signal. Utilize the decision signal as reference signal and continuing to adjust the forward and the backward parameters by CANFIS algorithm. The CANFIS algorithm includes two steps: the fuzzy inference step and the parameter adjustment mode.

Description

309675 A7 B7 V. Description of the invention (/) Printed by the Employee Consumer Cooperative of the Central Bureau of Standards of the Ministry of Economy Field of the Invention The present invention relates to a method and a device for processing complex fuzzy signals, in particular, it is possible to realize complex signals in a fuzzy way Adaptive nonlinear signal processing. Its main function is to perform nonlinear filtering, estimation, detection and other functions in a system using complex signals-such as sonar, radar, and communication-to reduce nonlinear support and noise And interference, improve signal quality. Description of conventional skills Existing fuzzy technology: real type, non-linear In recent years, research and application in the field of fuzzy are very prosperous. Household appliances (especially made in Japan) such as washing machines and air conditioners already have fuzzy intelligent control functions. Academic circles have been loyal and can follow the system. Tong Te _ strange broken-韞 矣 4 Yu Ju. Ying. 4 weeks-the whole sexual fuzzy system, such as Jang ’s adaptive fuzzy network. Lutui 1 ^. (See 1993 May, IEEE Transaction on System, Men and Cybernetics, pages 665 to 685). These systems can realize nonlinear adaptive control, dynamic system identification, adaptive signal processing (such as equalization) and other functions; however, all the processing is in sonar, radar, communication systems (analog or digital and quadrature amplitude modulation QAM modulator), the signal is Suiguilong1, so the existing blur method fails to perform the ft signal processing function in these systems. The existing adaptive signal processing technology: complex, linear, there are many existing adaptive signal processing methods. The signals are combined in a linear combination, and then their weights are adjusted. The signal type can be real or complex. The main weight adjustment rules are least mean square (LMS) and recursive least square. , RLS), RLS is a special case of the Kalman filtering method (^ 丨 《1 & 11 £ ^ 1 '). It should be applied to a bidder ’s country, a country, a middle school, and a suitable X ruler | sheet Μ S Ν I -Public 7 9 2 (please read the precautions on the back before filling in this page), τ 809675 A7 B7 Printed by the Employee Consumer Cooperative of the Central Standards Bureau of the Ministry of Economy V. Invention Instructions (2) Use noise elimination (no isecancellation), equalization (adaptiveequalization), signal combining of array antennas, etc. Taking the signal combining of array antennas as an example (complex signals), it is traditionally called beamforming; adjusting the weight of signals received by each antenna can adjust the incident signal in different directions The gain can be regarded as a kind of spatial filtering. If it is adjusted to make the gain in the direction of the interference smaller, and the gain in the direction of the received signal is larger, the interference can be suppressed and the signal quality can be improved. The effect of processing the signals in a linear manner, It is not ideal when the signal suffers from nonlinear distortion. The British two-year smart antenna project closing report (see Barret and Arnold, Electronics and Communication Engineering Journal, August 1994, pages 203 to 214) suggests that the nonlinear signal merger ( Non-linear beamforming) is an important research topic in the future, which is also one of the motivations for the creation of the present invention. What the present invention intends to solve is the signal processing performance when the signal environment or the transmission and reception devices are not ideal, and there may be nonlinear distortion The problem of recession. Summary of the invention to solve the above problems The present invention can process complex driving numbers in a non-linear fuzzy inference way. One object of the present invention is to provide a method for processing complex fuzzy signals, which combines the signals of the complex types with a linearly linear vibrating inference method, and can be realized by a parameter adjustment method. The learning function of adaptive signal processing, including initial mode, training mode and decision guidance mode; the initial mode first sets all adjustable parameters to zero 'training mode uses pre-stored training data as training reference signals, using the parameter adjustment method in CANFIS To adjust the parameters before and after to reduce the signal processing error. When the reference signal is exhausted or the error is received, switch to the decision guidance mode. Use the fuzzy inference method in CANFIS to combine the input signals of the complex type to obtain the output signal. The output signal is quantized to obtain the decision signal, and the decision signal is used as the decision reference signal to continue to adjust the front and rear parameters according to the parameter adjustment method in CANHS. pANFISj 宗 簾 氺 句 Today's next fuzzy inference steps and parameter adjustment steps, see Figure 1 (3); the fuzzy inference steps also include the paper size applicable to the Chinese National Standard (CNS) A4 specifications (210X297 mm) (please first Read the precautions on the back and fill in this page) -5 A7 B7 Printed by the Central Provincial Bureau of the Ministry of Economic Affairs Employee Consumer Cooperative V. Description of the invention (3 fuzzy steps, inference steps, and combination steps, see more-Figure 1) (1 >); The fuzzification step first obtains the corresponding attribution function value for individual input signals, then takes the product of the appropriate combination of individual attribution functions, and then normalizes the obtained product to obtain the normalized excitation intensity, the attribution function is Gaussian Function; the inference step uses the input signal to find the output function value of the corresponding fuzzy rule. The output function contains the adjustable post-parameter, and then the output function value and the normalized excitation intensity are multiplied to obtain the output of each rule being excited: The combination step adds the excited outputs of the rules obtained in the inference step to generate an output; the parameter adjustment step includes the previous parameter adjustment step Stochastic gradient method is used to adjust the antecedent parameter; the latter parameter adjustment step is to adjust the latter parameter by the minimum recursive variance method. Another object of the present invention is to provide a complex surface editing 'sound benefit method, based on non-Zhu Di's fuzzy inference Methods Combine the complex type tfL numbers, and use the parameter adjustment method to realize the learning function of the passability signal processing, including the initial mode, training mode and decision guidance mode; the initial mode first sets all adjustable parameters to zero 'training mode Use the pre-stored training data as the training reference signal, and use the parameter adjustment steps in CFBFN to adjust the front and back parameters to reduce the signal processing error. When the reference signal is exhausted or the error converges, switch to the decision guidance mode. The fuzzy inference method combines the input signal of the complex type to obtain the output signal, and then quantizes the output signal to obtain the decision signal, and uses the decision signal as the decision reference signal to continue to adjust the parameters before and after the parameter adjustment step in CFBFN. CFBFN-brook The deduction method contains the following fuzzy search for Aiman and the inspection of the cover. A); the steps of fuzzy inference include fuzzy Steps, inference steps, and combination steps, see Figure 2 (1)); The fuzzification step first obtains the corresponding high-dimensional attribution function value for the entire input signal vector, and then normalizes the attribution function value, which is normalized Excitation intensity 'The attribution function is a Gaussian function; the inference step uses the input signal to find the output function value of the corresponding fuzzy rule, and the output function contains an adjustable post-term parameter' and then multiplies the output function value by the normalized excitation intensity and Obtain the excited output of each rule; the combination step adds the excited output of each rule obtained in the inference step to generate an output; the parameter adjustment step includes the previous parameter adjustment step, and the previous parameter is adjusted by the stochastic gradient method; the latter Parameter adjustment steps, adjusted by the minimum recursive variance method. The paper size of this item is applicable to the Chinese National Standard (CNS) Μ specification (210X297 public reading :) ~ '~' II II Pack — II order 1! Line (please first Read the precautions on the back and fill in this page) 309675 V. Description of the invention (Xishu according to the invention-the purpose is to provide-a kind of chess signal processing equipment , Can be processed by fuzzy inference method of nonlinear signal, see Figure 3; including input mechanism for input «, operation processing touch ', first set the parameters of the village to the vocational mode, decision guide mode operation # 'Using the CANFIS deduction method to combine the signals of the plural types by fuzzy inference method' and implement the calculation of the new item parameters by the parameter hybrid method; remember the storage mechanism: store the calculation processor to ride the data; the output mechanism, the operational calculation The final output signal obtained by the processing mechanism is output. According to the present mosquito XK Wei supply-batter Eqin received a non-linear signal fuzzy inference method to process the call, including the signal of the input agency L people; the calculation processing agency's first set the financial miscellaneous parameters to zero, and then execute the Lin Fu, Decision-guided mode operation 'based on the CTBFN scale method, it is difficult to infer the prostitute merger, miscellaneous search for numbers, and the parameter adjustment method to realize the calculation of the tank secret item parameters. The domain storage complex structure stores the social data of the operation processing agency; output (iv) , The final output signal obtained by arithmetic processing is output. »In _ I ^ — (please read the precautions on the back before filling in this page) Order the detailed description of the printed example of the employee consumer cooperative of the Central Bureau of Standards of the Ministry of Economic Affairs • —. — The above and other features and advantages of the present invention will be available It is clearly presented from below with reference to the drawings. The first embodiment (signaling with CANFIS Acting Method) This embodiment is to successively use CANHS for adaptive signal processing, which can be started by & _ | _ fine-grained examination, adjusted by the parameter adjustment method in dNFis Before and after the parameters, adjust the adjustable parameters to a better value to find the field; Μ · Each Fen "lacks from the food A secret ¥. 丨 L. 1-**? To the ultimate mode, in a paper size Applicable to the Chinese National Standard (CNS) A4 specification (210X29Ti ^ y line A7 B7 printed by the Employee Consumer Cooperative of the Central Standard Falcon Bureau of the Ministry of Economy V. Description of the invention (r)

The Knitting Machine Vibration Theory Legislation in CANFIS ϋ Fulong: SL State's Round-up Information ^ | ^ Infer the jUjfL number, and then output the information. ¾ Set the Fushi match to hide it. Failed to make the reference signal continue to adjust the front and back parameters according to the parameter adjustment method in CANFIS, R adapts to changes in the characteristics of the name system, and reduces the signal processing error. According to the first embodiment of the present invention, the CANF clever deduction method can be divided into object selection and parameter week. The whole step-as shown in the first, 1 胤 ⑻ 'fuzzy inference step can also be divided into fuzzy north ^ inference, 4 combinations Sanhe Township, as shown in Figure 1 (b). The step of constructing the inference will be based on the input signal ·· '* ..: ··· ........ The output_signal will be obtained; the parameter adjustment step will be adjusted according to the law of adaptability. Parameters to reduce errors and achieve learning capabilities. In the fuzzy inference step, if the input signal $ of complex type should be regarded as a "multiply -...................._____ two / complex internal quantity, the output is regarded as a complex number If there is less, the relationship between input and output of ί and y can be described by a set of fuzzy rules as follows: If the input is not in the area / then the output is less /, (none), j_ = 〇 ^ .. w .......-. .... ^ — τ — »—............... The principle is to first select the space where the input Ϊ is located as the input fuzzy sales flexible cut and serve as the service area and book it in The output corresponding to each area is less = 乂 (3?), For example, 乂 (f) = #, «i + c, where the visit is the weight vector of the rule i output function (/, (none)) , C, is the rule, the constant of the output function (/ ((magic). When the actual input f appears, input 4¾. 踵 屋 屋% ^ degrees, and then the corresponding output of each area is weighted according to these degrees. weighted sum), interpolated (interp〇iati〇n). For example, the degree of not belonging to the first zone is ο.〗, the degree of belonging to the second zone is 0.7, (Ma (3〇 = 0.3, 4 (元) = 〇.7), the output output is 03 / ι⑺ + 0.7Λ⑺. To sum up, the fuzzy inference method with a regular input and output relationship Bian table is y " 'Ll = yMx) f, (x) ............................................. . ⑴ Here is a non-linear function. M is the number of rules. The size of the paper _ t S ϋ grid (21GX 297 Gongdong): --------- installed ------ order- ------ Line (please read the precautions on the back before filling in this page) 309675 A7 B7 Printed by the Employee Consumer Cooperative of the Central Bureau of Standards of the Ministry of Economy V. Invention Instructions (G) ⑴In the formula and ⑺ is the first formal The value of the attribution function of the normalized input module is ^ 绱, which is the degree of depletion in the first __ area, also known as the rule, the normalized excitation intensity, '⑺ = 1. Commonly attributed to [letter! Fully regulated Gaussian function, etc. A variety of choices can be selected according to the needs of the user. The adjustable parameters in Bing⑺ are called antecedent parameters, such as the center and standard deviation of the Gaussian function. The antecedent parameters will affect the way the input space is divided (the number of rules W is divided into regions) Number) ^ /, (none) is the output corresponding to the input f in the / area (the / rule), the common type is a constant, /, (none) = c, linear function ⑺ = #, + c, etc. where _ is a complex number vector multiplied by 7, and C is a complex number constant. The latter parameters include v. I and C. In short, the process of fuzzy inference can be divided into three steps: the sandalwood and the combination. The purpose of the shed is to generate positively normalized excitation intensity (none), ..., watt (f ). Referral: The purpose is to generate the rules .. The output is sent out by the public. Μχ) Α (χ), ..., Μχ) Μχ) Combination: The purpose is to produce | final round out when the fuzzy inference step is completed , Then proceed to the parameter adjustment step. This step is based on the difference between a reference signal r and the output is small, and the adjustable parameters of the previous term and the latter term are adjusted by the law of adaptability. The reference signal is an ideal output with a small value, and its source may be training data (training data) stored in advance or a decision value (decisi〇n) presumed from the inference results. Since Bing ⑺ is generally # item parameter, the non-witch H-number meets the whole 3 stitches ^ the main bee color inverse gradient method (stoch ^ stic gradient, SQ) iUt he and the travel method, so the weekly adjustment method, such as the base ja law. This embodiment uses random gradient colors (§Q). Since /, (¾ is a linear function, the parameter adjustment can be used to return the smallest square paper size suitable for the financial standard (CNS) A regulation iM 21GX 297 mm) 'I ϋ III Pack III binding II line (please read the back side first (Notes to fill out this page) Printed Α7 Β7 by the employee consumer cooperative of the Central Bureau of Standards of the Ministry of Economic Affairs 5. Description of the invention (Xi) Difference method (RLS) or other linear parameter adjustment methods such as the Conjugate Gradient Method (CGM), Least Mean Square Method (LMS), etc. In this embodiment, the recursive least variance method (RLS) is used. The parameter adjustment methods used in this embodiment all have the ability to process multiple signals and parameters. The processing steps according to this embodiment will be substantially described by way of example in the following. No. 戂 本 登 i 的 __ 色 色 體 理 立 产 為 ^ "Processing process. As shown in Figure 1 ⑻⑼, it represents a complex type input signal at a certain time. For the convenience of explanation, only three input signals X, X2, X3 are shown in the first picture. In fact, the number of input signals can be any number, not limited to 3. As shown in the first 10 (b), according to this embodiment, the fuzzy inference steps can be divided into five layers. The first three layers perform fuzzification, the fourth layer performs inference, and the fifth layer performs inference. After the input signal is processed through these five layers, the resulting output is ^, which can be expressed as (1), that is, the functions and processing procedures of each layer and the process of obtaining the output y will be described in detail below. Second; i "input the signal a, heart, and heart for complex numbers, and find the corresponding fuzzy occupancy values respectively. For example, when defining a quasi-Gaussian chess function for each input, ~ (ΉχΡ (-| \-~ || 7β), * = 1, 2, 3, / = 1, ··· , 4. The A1 (Α), gas (A), Wl3 (X,), / η | 4 (X,), w2l (χ2), m22), (ι2), w24 (χ2), % (Especially 3),% (Χι),% (heart),% (a) and other twelve normalized H values. Which is called the previous parameter, according to this implementation, M... ___,, I'J,; adjustment method will be detailed. One ~ second layer: multiply the attribution function value obtained in the first layer, and the product should be as follows: This paper scale is applicable to the Chinese National Standard (⑽) 84 specifications (training x297 public director) I- --------- 1 ------ 、 玎 --------- line (please read the precautions on the back before filling in this page) A7 B7 5. Description of the invention (5 = WllJXl) m2] (-^ 2) ^ 32 (-^ 3) μ, (χ) = ηιλλ (χλ) m2x (x2) m ^ x ^ where ^ = 0,, x2, x3) A (f) = mu (χι) w21 (jc,) mJ4 (x \) A (^) = w11 (x,) w22 (x2) W3i (X3), where 4 (none) = W14 (XI) cut 24 (h) 0¾ (丨3) Take the second layer. Take the multiplied screen. Accumulate. &Amp; Normalization (normalization) 'to obtain the normalized excitation intensity of each rule_⑺ as follows: 兵 ⑴ = ⑺ / Σ: ⑺ i = l, ..., 64 5 The first layer, the first layer, and the second layer perform fuzzification. The fourth layer: find out the output function value of each rule (none) = g ^ + c, and multiply it with the normalized excitation collapse and (magic) to get the output of each rule being excited and (Magic / '(magic, i = l, · .64. This step is also called inference). Among them, c; is called the latter parameter, adjusted using the recursive least variance method (RLS). Level 5 .Add the stimulated outputs of the inferred rules, and get the final total output less =) / (f). This step is also called Composition. ° The Central Committee of Bureau of Economics and Staff Employee's Consumer Cooperative printed the yuan into a blur After the inference step, the next step is to teach the teacher. As shown in section 1__, this step is based on the difference between the reference signal r and the output y, using the machine gradient method (SG) to adjust the previous parameters (the center of the Gaussian function and Standard deviation) and adjust the parameters of the latter term by the recursive least variance method (RLS) (weight # 'of the output function and the constant term ς). The method of adjusting the latter term parameters can be performed according to the existing complex number method. The parameter adjustment method (SG) in the plural form is as follows: CNS) A4 size (21〇Χ297 mm) 309 675 V. invention is described in (,

, Is the first reference signal, _ is the fuzzy deduced squared difference can be regarded as-cost function (-η): output 'then ♦ ΗΚβϋ2. W is an adjustable parameter, random gradient _ to do) Move in the opposite direction of the gradient , In order to reduce the cost of the tangential gradient, then ❸ Ρ (η + 1) = ρ {η) -ην pj (n) where V represents the gradient; "representing time; 9 is the actual step size ^ 4 ^ Regular editing I was distinguished before, "The center and standard deviation of the / th subtraction function defined above, h is a complex number) can be rushed to embrace: C ^ n + 〇--^ (2 ^«) ίΤ; Σ Re (/, .) i-

C

kl Cz i kl Ί) () | 2, Xk_ckl, σΜ J

akAn + l) = ^ kl {n) + ^ | 2Re [e («)] ΙΣRe (/.)^ L Φ; · 3η kl 2 · ^ mkr '° Λ σ \ ι 丨 binding ^-line f please Read the precautions on the back first and then fill out this page) Printed by the Ministry of Economic Affairs Central Enforcement Bureau Employee Consumer Cooperative and A = XMi Re ^ ·), B =% Mi Im (y.), C = Σ. According to this example As the number of inputs increases, the number of rules will grow exponentially, so the adjustable parameters and the amount of calculation also grow exponentially. Second Embodiment (Processing Signals with CFBFN) The size of this paper is applicable to the Chinese Standard (CNS) M Specification (210X297). The A7 B7 is printed by the Consumer Cooperative of the Central Standard Falconry Bureau of the Ministry of Economy. V. Invention Description (/ 〇) in In the first embodiment, the number of rules will grow exponentially due to the number of inputs. In this embodiment, the number of rules may be arbitrarily set, and will not grow exponentially due to the increase in the number of inputs. This embodiment differs from the first embodiment only in the blurring step, and the remaining steps are the same, see FIG. 2 (a). This embodiment defines a high-dimensional fuzzy attribution function for the entire complex-type input vector. The number of rules is the number of high-dimensional fuzzy attribution functions, which can be arbitrarily set. Therefore, when there are many input numbers, this embodiment can Significantly reduce the number of rules and calculations. As shown in the second sample, according to the example of this real estate, the step of pushing and pushing the ancestor of ρρΒ 祖 can be divided into 4 layers. The first two layers perform blurring, the first name is selected by Shu Zhe, and the fourth layer performs combination. After the input signal is processed through these four layers, the resulting output is less, which can still be expressed as (1), that is, less (magic. /; (Magic about the functions and processing procedures of each layer and the process of obtaining output y, will It is described in detail below. The first layer: For complex input signal vector ί = (ΧρΧ2, Χ3), find the corresponding fuzzy dimension function value of the post dimension. For example, if the rule number is set to m = 20, then the input The vector defines 20 Gaussian fuzzy membership functions as follows: μ, (χ) = exp (-|| χ-〇, \ [/ af), / = 1, .. 20. This layer obtains the 20 membership functions Value, where σ, called the previous parameter, is adjusted using the stochastic gradient rule. Second layer: Perform normalization on the attribution function values obtained in the first layer to obtain the regularization of each rule The excitation intensity is as follows: Μχ) = μι (χ) / Υ ^ = ιμ. (Χ) i = lt ... 7〇 paper ruler's money_Home Standard (CNS) A4 · mm) ----- ---- approved clothing ------, order ------ ^ (please read the notes on the back before filling this page)

The first layer and the second layer perform fuzzification. The third layer: find the output function value of each rule (none) =% null + £; and multiply it with the normalized excitation intensity and 〇〇 obtained in the second layer, and get the output of each rule excited (f) /; (i), i = 1, ..2〇. This step is also called inference. Among them, # ,, c, are called the latter parameters, adjusted using the recursive least variance method (RLS). The fourth layer: add the excited outputs of the inference rules to obtain the final total output less = Σ two and (none) /, (none). This step is also called composition. The parameter adjustment in this embodiment is performed according to the reference signal / -⑽. The stochastic gradient method (SG) is used to adjust the antecedent parameters (the center and standard deviation of the Gaussian function), and the recursive least squares method (RLS) is used to adjust the latter parameters (the weight of the round function and the constant term c,). According to the previous parameters ί ,, σ, according to this embodiment, the stochastic gradient method (SG) according to the complex pattern is adjusted as follows: --------- batch apparel! (Please read the notes on the back before filling this page) c («+1) = c, · («) + j 2 Re [e (n)] j Re (/,-)

D, · (»+1) = CT / («) + J 2 Re [e («)]-Re (/ ;.)-~ {Fl 2Im [e (rt)]

F fly

F

E X-C, .2. Σ) Line Printed bags of the Employee Consumer Cooperative of the Central Standard Falcon Bureau of the Ministry of Economic Affairs m /) = Σ Re here / = 1 ft

Mi £ = Σ Im fi

Mi m F · This application example uses the same method as the first embodiment (CANFIS) and the same method as the second embodiment (CFBFN) to realize the non-prosperous signal combination of the digital ligJk4 system t block strong jy ^ (Podong Also · Cheng). As shown in Figure 4. This paper scale is applicable to the Chinese National Standard (CNS) Α4 specification (21〇297 mm) Printed by the Consumer Standardization Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs Α7 Β7 i. Invention description (/ no) The digital signal in the communication system is passed (Quadrature Phase Shift .. Kfy 丨 ng, qPSK) method to modulate the QP signal from birth to patient. In the communication system, the equivalent fundamental frequency signal transmitted by a user at the time "can be regarded as one of ± 1 ± _ / four points in the complex signal space. During the transmission process, due to multipath delay and co-channel interference, the received signal received by the receiving end not only contains items, but also the interference items such as the signal item previously transmitted by the user and the transmission signal of another user, and the noise item V »Receive At the end there is a linear equidistant array antenna composed of three antennas. When the incident direction of the interference item ^^ is different from the desired signal c (㈨, the signal received by combining the three antennas properly (")": *: 2 («), 13 («) can suppress interference and reduce Error. The following describes the effectiveness of this application example in the implementation of linear and nonlinear signal environments. The judgment of effectiveness in digital communication systems is based on the sign error (sj | ji ^ 〇l err〇rj ·%. The more the sign error rate The smaller the better, the linear signal environment adopts Barrett and Amott (British smart antenna thorium painting) holiday blanket type as shown in Figure 5. The QPSK signal a (n) to be received is incident at 0 degrees: coffee … Is a resistance ⑻ a reflection symbol delayed by an obstacle, the intensity is loyal and private n) 5 dB lower, the incident fish is 20 degrees; b ⑻ represents the co-channel interference, the incidence angle is 5 degrees, and the intensity ratio is 4 ° degrees a ( n) 5dB lower. Assuming the distance between the antennas and the artificial carrier wavelength, the phase delay of the signals received by each antenna is 9 = = = π «> / (θ). Therefore, the < part of the female antenna received phase Extend the test to 1) φ, which can be expressed as (?) — 一一 — xk (n) = a {ri) + 0.5623α (η-l) ej (k ι) πΗ20) + 〇5523 ^ (^-• / (AI) tc blade>? (5.) + 0.5623 + Κ) (Α —,) π44 (Γ) + ν »(2) where ~ (《) is a complex white south For Si noise, the real and imaginary parts are #.

This paper wave scale is applicable to the Chinese National Standard (CNS IIIII Order II ^ Chain (please read the precautions on the back before filling in this page) A7 _________ B7 printed by the Employee Consumer Cooperative of the Central Bureau of Standards of the Ministry of Economy V. Invention description () Simulation experiment results As shown in Figure 6, Figure 7, Figure 8. Figure 6 is the accumulation of error signals during the training phase, Figure 7 is the diagram of the equivalent fundamental frequency signal in the decision guidance mode, and Figure 7 (a) is the second branch. The antenna received the signal. Figure 7 (b), (c), (d) are the inferred output signals obtained by the linear RLS signal combination, CANFIS and CFBFN fuzzy inference. Figure 8 compares 30 independent experiments , 2⑻ signals per training mode, the average symbol error rate (SER) of 5000 signals in the decision guidance mode. The experimental results show that the linear signal environment has the best beamforming effect with linear RLS method and the least error signal accumulation , SER is also the smallest (Ι.ηχΚΓ4). However, the performance of CANFIS (SER 1.53χ1 (Γ4) and CFBFN (SER 1.8〇x10-4) is also very close. The nonlinear system environment is used by Cha and Kassam to study nonlinear equalization. Model, assuming that there is nonlinear distortion at the receiving end, and assuming that the main path a (n) undergoes large amplitude attenuation, as shown in the first figure. Then the signal received by the * th antenna can be expressed as follows: xk {n ) = ok (η) + 0.1〇1 (η) + 0.05〇1 (η) + vk (η) (3) (/ 7) = (0.34-_ / 0.27) α («) + (0.87 + y〇 .43) a (/ i-l) e—7 (* _ 1) π —2〇. ) + (0.34-y〇.21> («V Qin" (5.) + (0.25-—_4 °.) ⑷ ν * («) is the complex white Gaussian noise, both real and imaginary parts are tV (〇, 0.025). The simulation results are shown in the 40th tree and Fig. 12. Fig. 10 is the accumulation of error signals during the training phase, and Fig. 11 is the diagram of the equivalent fundamental frequency signal in the decision guidance mode. ()) Is the signal received by the second antenna. Figure 11 (b), (c), (d) are the inferred output signals obtained by the linear RLS signal combination, CANFIS and CFBFN fuzzy inference. Figure 12 Comparison 3 〇 independent experiments, 200 signals per training mode, and the average symbol error rate (Symb〇i en * 〇r rate, SER) obtained from 5000 signals in the decision guidance mode. Experimental Results ------ .. The results show that the beamforming effect of the linear RLS method in a nonlinear signal environment is greatly degraded. Not only the number of error symbols accumulates rapidly, but the average symbol error rate SER value is the largest (0.175), and it is as close as 5 times the SER value of CANFIS .CANFIS error signal accumulates the slowest, and the SER is also the smallest (0_0365) ° The SER value of CFBFN is 0.0908, 0.0473, 0.038 when the rule number is 20, 50, 64 respectively 0. The simulation results show that the traditional linear beam paper size is applicable to the Chinese National Standard (CNS) A4 specification (210X 297 mm) II Pack-III line (please read the precautions on the back before filling this page) 309675 A7 B7 5. 2. Description of the invention (with) The forming method cannot effectively overcome the non-linear distortion. The fuzzy beamformer can effectively combine the signals and overcome the non-linear distortion, and has a strong adaptability to the changes of the signal environment. The above is described by way of examples Inventions, those who are accustomed to this skill, can perform modifications and variations according to the above description without departing from the spirit and scope of the patent application scope. ≪ Please read the precautions on the back before filling out this page) Central Standard Falcon Bureau employee consumption cooperation Du printed paper standard is applicable to China National Standards (CNs) a4 specifications (210X297 mm) A7 B7 V. Invention description (/ 夂)

Figure 1 ⑻ illustrates the architecture of the CANFIS theory system according to the _th embodiment of the present invention, which mainly divides money into inferences and refers to Section 1 (b) according to the first embodiment of the present invention, further analyzes the fuzzy inference part in CANFIS The structure of the share is mainly divided into three steps: fuzzification, inference and combination, which are realized by a five-layer network structure. Figure 2 (a) illustrates the structure of cFBFN paste base & number according to the second embodiment of the present invention, which mainly depends on paste reduction and parameter adjustment-part. Figure 2 (b) is based on the second embodiment of the present invention, and further analyzes the structure of the fuzzy inference part in CFBFN, which is mainly divided into three steps: fuzzy, inference and combination, which are realized by a four-layer network structure. Fig. 3 is an implementation diagram of an apparatus for implementing adaptive fuzzy signal processing according to the first embodiment and the second embodiment of the present invention. (Please read the precautions on the back and then fill out this page)-* Γ Printed by the Consumers ’Collaboration Department of the Central Bureau of Economics of the Ministry of Economic Affairs. Figure 4 is a digital wireless communication system using CANFIS and CFBFN according to the application example of the present invention In the figure, the structure of the received signal combining (fuzzy beam forming) of the array antenna. Figure 5 is the illustration of the linear signal environment used in the simulation experiment according to the application embodiment of the invention. Figure 6 is the application implementation according to the invention For example, the linear signal environment simulation experiment during the training mode stage linear (RLS) and chess (CANFIS, CFBFN) beamforming results error signal cumulative number graph. Figure 7 is an application embodiment of the present invention, the linear signal environment simulation experiment Fig. 7 (a) is the signal received by the second antenna, Fig. 7 (b), (c) and (d) are merged via linear RLS signal, CANFIS and The inference output signal obtained by the fuzzy inference of CFBFN. This paper scale is applicable to the Chinese National Standard (CNS) A4 specification (210X297mm). V. Description of the invention (/ 6) A7 B7 Figure 8 is the application according to the present invention Example, 30 independent experiments in a linear signal environment simulation experiment, 200 signals per training mode, and the average symbol error rate (Symb〇i error rate, SER) comparison of 5000 afl numbers in the decision guidance mode Table 9. Figure 9 is an illustration of the nonlinear signal environment used in the simulation experiment according to the application embodiment of the present invention. Figure 10 is a training mode stage linear according to the application embodiment of the present invention during the nonlinear signal environment simulation experiment ( RLS) and fuzzy (CANFIS, CFBFN) beamforming results. Accumulated number of erroneous signals. Figure 11 is a diagram of the equivalent fundamental frequency signal in the decision guidance mode during the simulation experiment of a nonlinear signal environment according to the application embodiment of the present invention. (a) is the signal received by the second antenna. Figure 11 (b), (c), (d) are the inferred output signals obtained by the linear RLS signal combination, CANFIS and CFBFN fuzzy inference. Figure 12 According to an application embodiment of the present invention, the linear signal environment simulation experiment is 30 independent experiments, each training mode has 200 signals, and the average symbol error obtained from 5000 signals in the decision guidance mode is Rate (Symb〇l error rate, SER) comparison table 0 (Please read the Notes on the back to fill out this page) • qi. Γ

I Printed by the National Bureau of Standards of the Ministry of Economic Affairs and the Consumer Cooperative Cooperative. This paper scale is applicable to the Chinese National Standard (CNS) A4 (210X297mm)

Claims (1)

S09675 8 8 88 ABCD π, patent application scope Central China Bureau of Economic Affairs Employee consumption cooperation Du Print 1 A method of complex fuzzy signal processing, which implements adaptive non-linear signal processing of complex type signals by fuzzy inference, including: The initial chess style 'set all the adjustable parameters to zero first; the training mode' uses pre-stored training data as training reference signals, and uses CANFIS deduction to adjust the parameters before and after to reduce the error of signal processing. When the reference signal When exhausted or error converges, switch to the decision-making chess game. The CANFIS deduction method includes the following inference steps and parameter adjustment steps; the fuzzy inference steps include: the fuzzification step. First, obtain the corresponding attribution function value for the individual signal , Then take the product of the appropriate combination of individual attribution functions, and then normalize the obtained product to obtain the normalized excitation intensity. The attribution function is a Gaussian function; the inference step is to find the output function corresponding to the fuzzy rule in the input signal Value, the output function contains the adjustable after parameter, and then the output function Multiply the normalized excitation intensity to obtain the output of each rule being excited; the combination step, the output of the _ inference step cap is added, and the output is generated; the parameter adjustment step includes: the previous parameter adjustment step, Adjust the previous parameter according to the following formula; here ▽ means to take the gradient; "represents the time ;; 7 is the step parameter% of the previous item, and the center and standard deviation of the / th Gaussian function defined on / are complex numbers) Stochastic gradient method adjustment: C 1 r /, \ Φ; Β Φ: 1 f \ ___V ^ 1 c2 i ^ ki. 1.2.¾. Xk ~ ckl σ \ ι > akAn + Ο = σΜ + 2 Re [e (M)] J binding ^ line · _ (please read the precautions on the back before filling in this page) 309675 A8 B8 C8 D8 Patent application scope d dm. Σ- • 2.j lkl ~ ckl \\ σ1ι The quasi-bureau-industrial and consumer cooperatives printed nkl c2 i ^ kl .. and h ^ ReU), c =: ^ ·· The adjustment step of the latter parameter 'is to adjust the latter parameter by the recursive minimum variance method; the decision guidance mode' is fuzzy Inference to obtain the inference output signal, after quantization to obtain the decision output signal 'and use it to make the decision output signal The signal is the reference signal, and the performance is adjusted by the CANFIS method before and after the parameters. 2 · —A kind of complex fuzzy signal processing method, which implements adaptive nonlinear signal processing for complex type signals by fuzzy inference, including: Initial mode, first set all adjustable parameters to zero; Training mode, use pre-storage The training data is used as the training reference signal. The CFBFN deduction method is used to adjust the parameters of the front and back items to reduce the error of signal processing. When the reference signal is exhausted or the error converges, switch to the decision-guided chess style. The CFBFN deduction method includes the following The fuzzy inference steps and parameter adjustment steps are described; the fuzzy inference steps include: the fuzzification step 'first' obtains the corresponding high-dimensional attribution function value for the entire input signal vector, and then normalizes the attribution function value to obtain the normalized excitation intensity, The attribution function is a Gaussian function; the inference step is to find the output function value of the corresponding fuzzy rule in the input signal, and the output function value contains an adjustable post-term parameter, and then the output function value is multiplied by the normalized excitation intensity And get the output of each rule being stimulated; combine the steps and get the results from the inference step The outputs that are excited by the rules are added to produce an output; the parameter adjustment steps include: the previous parameter adjustment step, which adjusts the previous parameter according to the following formula; the previous parameter ^, σ, adjusted according to the complex type random gradient method (SG) as follows: ( Please read the precautions on the back before filling out this page) This paper uses the national standards (CNS) Α4 specifications (210X297mm). Patent scope (w +1) = ci («) + I 2 Re [ e («)] Re | A8 B8 C8 D8 U)-F X-Ct < 7 / (« + l) = σ («) + I 2 Re [e («)] FD " T + 2 丨 m [e (”)] Shi lm (乃) — F xc; σ] m ^ > = Σ Re here mi; £ = Σ Im A Ui m Ministry of Economic Affairs Bureau of Central Standards Beigong Consumer Cooperative ’s policy adjustments Step, adjust the latter parameter by the minimum recursive variance method; Decision-making guide mode, obtain the inference output signal by fuzzy inference, obtain the silk output signal after quantization, and use the decision output signal signal as the reference signal, continue to adjust by the CFBFN method Before and after parameters 3. A kind of fuzzy signal processing device, which can be processed by nonlinear fuzzy inference Signals, including: Input mechanism for inputting signals; Operation processing mechanism, first set all adjustable parameters to zero, perform training mode, decision-guided mode operation, the training mode operation uses pre-stored training data as training reference The signal uses CANFIS deduction method to adjust the parameters of the front and back items to reduce the error at the signal. When the reference signal is exhausted or the error converges, switch to the decision-guided operation. The CANFIS deduction method includes the following fuzzy inference steps and parameter adjustment steps ; The fuzzy inference steps include: The fuzzification step. First, obtain the corresponding attribution function value for each input signal, and then take the product of the appropriate combination of individual attribution functions' and then normalize the obtained product to obtain the normalized excitation intensity. The attribution function is a Gaussian function; the inference step is to find the output function value of the corresponding fuzzy rule from the pre-stored training data, the output function value contains an adjustable parameter, and then the output function value and the normalized excitation intensity Multiply to get the output of each rule being excited; the combining step will The stimulated outputs of the rules obtained in the inference step are added to produce the output; the parameter adjustment steps include: This paper scale is applicable to the Chinese National Standard (CNS) A4 specification (210X297 mm) IIIIII Pack IIII Order III line * ( Please read the precautions on the back before filling this page) A8 B8 C8 D8 Patent application scope The Central Parameter Bureau of the Ministry of Economic Affairs Beigong Consumer Cooperative printed the first parameter adjustment step, based on the τ type mine previous parameter. The latter parameter is transferred, called the payment of bribery adjustment parameters; Decision guidance mode, obtained by fuzzy inference Infer the output signal, obtain it after quantization and use the decision output signal as the reference signal, __canfis method adjustment, and the number of subsequent items. The memory storage mechanism stores the training data of the arithmetic processing mechanism_; the output mechanism 'rotates the final decision-round signal obtained by the arithmetic processing mechanism. 4-A kind of difficult job processing device, which can deal with signals in ‘Miscellaneous _ on the way including: input mechanism 'for inputting signals; operation processing mechanism' first set all adjustable parameters to zero, perform training mode, decision guidance mode operation 'The training mode operation uses pre-stored training data as training reference signal = CFBFN extensive interpretation method to adjust the front and rear parameters to reduce the signal error, when the reference signal is exhausted or incorrectly converged, switch to the decision f 5丨 Derivative operation, the CFBFN deduction method includes the following fuzzy inference steps and parameter adjustment steps; fuzzy inference steps include: the fuzzification step 'first' ride the entire input job transfer to the high-dimensional attribution function value 'and then the value of the pass function Normalization, and the normalized excitation intensity is obtained. The attribution function is a Gaussian function; the inference step is to find the output function value of the corresponding fuzzy rule from the pre-stored training data. The output function value contains the adjustable after-term parameter, and then the The output function value is multiplied by the normalized excitation intensity to obtain the output of each rule being excited; the combination step is obtained from the inference step The output of each rule that is excited is added to produce an output; the parameter adjustment step includes: the previous parameter adjustment step, which adjusts the previous parameter according to the following formula; the latter parameter adjustment step, which adjusts the latter parameter by the least recursive variance method; decision "Guide mode" uses fuzzy inference to obtain the inference output signal and quantize the decision output signal, and uses the decision output signal signal as the reference signal, and continues to use the CFBFN method to adjust the paper size before and after. A4 specification (210X297 public shame) (please read the precautions on the back before filling in this page). Binding · Order * I--II II · A8 Bg C8 D8 々, patent application range parameters. Memory storage mechanism, store the calculation The training data required by the processing agency; The output agency outputs the final decision output signal obtained by the arithmetic processing agency. II Gutter (please read the precautions on the back before filling out this page) The paper printed by the cooperative is based on the Chinese National Standard (CNS) A4 (210X297 gong)