One-step calculation circuit for signal blind detection based on complex-valued hopfield neural network

Abstract

The application will break through the traditional signal blind detection implementation mode, explore the high-energy efficient hardware implementation method of analog signal blind detection, and innovatively propose an efficient signal blind detection circuit from the analog circuit angle, belonging to the field of signal processing. First, how to realize the activation function circuit is studied from the analog circuit angle; secondly, on this basis, a complete signal blind detection one-step calculation circuit based on complex-valued Hopfield neural network is designed. The application has very important scientific significance and potential economic benefits for signal blind detection technology; the research results can be widely applied in the field of signal processing, and help to form core intellectual property rights in this field.

Description

TECHNICAL FIELD

[0001] The present application relates to a signal blind detection one-step calculation circuit based on a complex-valued Hopfield neural network, and belongs to the field of signal processing. BACKGROUND

[0002] With the development of wireless communication technology, signal processing calculation becomes more and more important. Signal blind detection technology has been widely used in communication systems. The traditional method needs to spend a long time for large-scale signal processing, and how to improve the speed of blind detection is the research focus of current researchers. But current algorithms mostly rely on software calculation, limited by traditional computer architecture, unable to balance the contradiction between speed and power consumption.

[0003] (1) Traditional algorithm: Sato first proposed a self-recovery algorithm that can complete equalization without training sequence, but the convergence speed is slow. Li et al. developed a fast serial blind equalization algorithm for MIMO QAM system, which significantly improved the performance of MSE, and the equalization performance is good and the convergence speed is fast. Mashimo et al. proposed an adaptive blind equalizer with noise reduction function, which improves the equalization accuracy. Komatsu et al. proposed an adaptive direct blind equalization in noise environment. Zhou proposed a blind equalization algorithm based on mldpc code, which can suppress inter-symbol interference, reduce bit error rate, improve signal robustness and improve communication quality. Kolosovs proposed a novel blind equalization method based on EDAMA algorithm. Lv et al. proposed an improved dual-mode blind equalization algorithm, which improves the blind equalization performance of QAM signal, and has smaller bit error rate and residual inter-symbol interference than the original algorithm. Yuan et al. proposed a new diffusion constant modulus algorithm (diffusion constant modulus algorithm, DCMA), which performs blind equalization on the communication channel composed of sensors located in the receiver array, greatly improving the bit error rate performance. But it needs to be noted that these methods mostly need a large amount of calculation and data support, and the recovery effect is not ideal in the case of small amount of data.

[0004] (2) Neural Network Algorithms: Neural network algorithms are mainly divided into cost function-based algorithms and energy function-based algorithms. Ying et al. proposed a neural network blind equalization algorithm based on gradient iteration, which has good performance under nonlinear communication channel conditions. Zhang et al. proposed a complex-valued Hopfield neural network based on energy function for blind signal detection. Wu et al. proposed a signal blind detection algorithm based on MC-CHNN (Multi-Clustering complex-valued Hopfield Neural Network), which improved the convergence speed of blind detection, resulting in faster convergence and stronger noise resistance. Mei et al. proposed a new blind equalization algorithm based on convolutional neural network (CNN) to improve the bit error rate (BER) performance of the equalizer against multipath fading effects and nonlinear distortion. Xie et al. proposed a variable step-size multilayer neural network blind equalization algorithm, which not only improved the convergence accuracy and reduced the error, but also effectively accelerated the convergence speed. It is worth noting that some of these neural network algorithms use real-valued neural network algorithms, while others use complex-valued neural network algorithms. When faced with complex signal and channel characteristics, the training efficiency of real-valued neural networks is low, and the results are not satisfactory.

[0005] The analysis of the above literature shows that current blind detection technology is still implemented in software, which suffers from problems such as separation of storage and computation and low real-time performance. Furthermore, few researchers have designed solutions to the blind detection problem from a hardware perspective. Summary of the Invention

[0006] The purpose of this invention is to break through the traditional methods of blind signal detection and to innovatively propose a solution circuit for blind signal detection from the perspective of analog circuits, thereby overcoming the shortcomings of existing technologies.

[0007] This invention is achieved through the following technical solution: based on the blind detection technology of energy function, a high-speed signal blind detection circuit is proposed from the perspective of analog circuits, the rules of blind detection technology are analyzed, and the corresponding circuit model is established;

[0008] Based on the circuit model, design and build the memristor array circuit and the complex-valued activation function circuit for the complex-valued weight matrix; including designing the complex-valued Hopfield weight matrix operation circuit using the memristor array, designing the corresponding processing modules according to the activation function design rules, and establishing circuit models for each module.

[0009] According to the established circuit model, the one-step calculation circuit design of signal blind detection based on complex-valued Hopfield neural network is researched, the setting relationship of the activation function in the signal blind detection technology is researched, the circuit model is constructed, and then the one-step calculation circuit of signal blind detection based on complex-valued Hopfield neural network is realized, and is used for processing 8PSK constellation signals.

[0010] The present application comprises the following steps:

[0011] Step 1, complex-valued activation function circuit design;

[0012] The designed activation function is:

[0013]

[0014] The activation function circuit is composed of an absolute value judgment basic module, an imaginary part basic module and a real part basic module.

[0015] S1.1 absolute value judgment basic module

[0016] The absolute value judgment module is composed of an absolute value circuit and a division circuit, and the basic modules are resistors, operational amplifiers and diodes. Voltage inputs V iI and V iR are applied. For the second operational amplifier, there is:

[0017]

[0018] wherein R1=2R, and V

[0019]

[0020] When V iR >0, according to the "virtual short" and "virtual break" characteristics of the operational amplifier, the output V o1 of the first operational amplifier is -V iR <0, and the diode is cut off, so that V o2 =V iR . When V iR <0, the output V o1 of the first operational amplifier is -V iR >0, the diode is turned on, and V o1 =0 due to "virtual ground". At this time, V o2 =-V iR . Therefore, V o2 =|V iR |. For the division circuit, it is mainly composed of a multiplier and an operational amplifier. The multiplier outputs V o3 =V o2 V o . According to the "virtual short" and "virtual break" characteristics of the operational amplifier, it can be obtained:

[0021]

[0022]

[0023] V o2 =|V iR |Substituting into the equation yields... S1.2 Basic Module of Imaginary Part

[0024] Based on the output of the defined piecewise function, a selection circuit was designed. The imaginary part of the output has only four cases: (-0.9239, -0.3827, 0.3827, 0.9239). The corresponding inflection points should be (-1, 0, 1). The voltage E is then... i Let's set this as a turning point. Based on Kirchhoff's laws and Ohm's law, we can conclude that:

[0025]

[0026] V out =-f(V in -E j (6)

[0027]

[0028]

[0029] Where R zj ,R xj These are the relevant parameters for the operational amplifier.

[0030] At this point, a current output i is obtained, which needs to be converted to voltage. A transimpedance amplifier (TIA) is connected after the imaginary part basic module. The current flows through the TIA and is converted into voltage. This yields the output V. oI :

[0031] V oI =-iR img (9)

[0032] S1.3 Real Part Basic Module

[0033] Unlike the imaginary part module, the real part output exhibits symmetry about the y-axis, thus requiring adjustments to the circuit structure.

[0034]

[0035]

[0036] The circuit consists of a basic module and a subtraction circuit. The subtraction circuit calculates the voltage difference between the two input terminals, and then the final real part output voltage Vo is obtained through the voltage divider circuit.R ;

[0037] Step 2, the signal blind detection one-step calculation circuit based on the complex-valued Hopfield neural network is designed based on the complex-valued activation function circuit, which is composed of an activation function circuit and a weight matrix circuit, and the weight matrix circuit is composed of a memristor array and an operational amplifier;

[0038] S2.1 Memristor array configuration

[0039] The basic unit of the memristor array is a memristor, which corresponds to the weight value in the weight matrix. The memristor is connected to a PMOS tube and an NMOS tube to form a mode selection switch, which can program and calculate the memristor. V tn and V tp are the threshold voltages of the NMOS tube and the PMOS tube, respectively. When V c <V tp , the PMOS tube is turned on and the NMOS tube is turned off, and the memristor is in programming mode. At this time, we can adjust the value of the memristor by applying a voltage of different amplitude and width. V c >V tn , the NMOS tube is turned on and the PMOS tube is turned off, and the memristor is in calculation mode. The conductance value of the memristor is denoted as G. When a voltage V is applied to the memristor, the current I flowing through the memristor is I=G·V. The memductance value of the memristor cannot be negative, but the weight value can be negative. We use a double-column memristor array to subtract to solve the problem of negative weight coefficients. Each weight value is represented by two memristors, G and G'. When the weight value A is negative, let G=A and G'=2A. The subtraction of the two memristors can obtain a negative weight value. Conversely, let G=2A and G'=A to obtain a positive memductance value.

[0040] S2.2 Signal blind detection one-step calculation circuit design

[0041] The calculation of the weight matrix depends on complex-valued calculation. There are two complex-valued vectors x=a+bi and y=c+di. The result of complex-valued multiplication is z=x·y=(ac-bd)+(ad+bc)i. Using matrix-vector calculation, it can be represented as:

[0042]

[0043] The weight value is complex-valued, and the output is the real part and the imaginary part. Using the memristor array to map the weight matrix, W ire and W jim can be represented as:

[0044]

[0045] Let G 11 , G 22 = 2a, G 21 = 2b, G 12 = b, substituting formula (13) into formula (13) obtains W 1re = a, W 1im = b, substituting the obtained value into formula (12) obtains:

[0046]

[0047] Simplifying formula (14) obtains the same result as formula (12);

[0048] The specific principle of the one-step calculation circuit based on the complex-valued Hopfield neural network is as follows:

[0049]

[0050]

[0051] V Yi = F(V Ii )

[0052] = F(-I xi ·R) (17)

[0053] Thus, the output voltage V Yi is obtained, which is also used as the input of the next round of circuit;

[0054] The main process of the blind detection circuit is as follows: V Xi and W are obtained through preprocessing, wherein V Xj is used as the input voltage, the weight matrix W is configured as the memristor array, the input voltage V Xj flows through the memristor array, is transformed into the current I xi according to the Kirchhoff's law, is transformed into the voltage V Ii by means of the TIA, is activated by the activation module to obtain the final output V Yi , and then it is used as the input of the new round; when the neural network reaches a stable state, V Yi is the final output of the circuit.

[0055] The present application has the beneficial effects of breaking through the traditional signal blind detection implementation mode, exploring the high-energy-efficiency hardware implementation method of analog signal blind detection, and having very important scientific significance and potential economic benefits for the signal blind detection technology, and the research results can be widely applied in the signal processing field and help to form core intellectual property rights in the field. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1is a constellation signal phase diagram.

[0057] Figure 2 is an absolute value decision module diagram.

[0058] Figure 3 is an imaginary part basic module diagram.

[0059] Figure 4 is a real part basic module diagram

[0060] Figure 5 is a memristor unit diagram

[0061] Figure 6 is a complex value multiplication circuit diagram

[0062] Figure 7 is a signal blind detection one-step calculation circuit diagram based on a complex value Hopfield neural network

[0063] Figure 8 is an overall processing block diagram DETAILED DESCRIPTION

[0064] In order to make the objects, technical solutions and advantages of the present application clearer and more comprehensible, the following will combine specific embodiments and refer to the accompanying drawings to further specifically describe the present application. Figures 1 to 8 It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present application. In addition, in the following description, the description of known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present application; in addition, the technical features involved in the different embodiments of the present application described below can be combined with each other as long as they do not conflict with each other. The present application will be described in more detail below with reference to the accompanying drawings. In each of the drawings, the same elements are represented by similar reference numerals. For the sake of clarity, each part of the drawings is not drawn to scale; the following will describe the embodiments of the present application based on the accompanying drawings. Figures 1 to 8 The one-step calculation circuit of the signal blind detection based on the complex value Hopfield neural network of the embodiment of the present application is described below with reference to the accompanying drawings.

[0065] Based on the energy function blind detection technology, a high-speed signal blind detection circuit is proposed from the perspective of analog circuit, the blind detection technology rule is analyzed, and the corresponding circuit model is established;

[0066] According to the circuit model, the memristor array circuit of the complex value weight matrix and the complex value activation function circuit are designed and built; including using the memristor array to design the complex value Hopfield weight matrix operation circuit, according to the design rule of the activation function, the corresponding processing module is designed, and each module is established into a circuit model;

[0067] According to the established circuit model, the one-step calculation circuit design of signal blind detection based on complex Hopfield neural network is researched, the setting relationship of the activation function in the signal blind detection technology is researched, the circuit model is constructed, and then the one-step calculation circuit of signal blind detection based on complex Hopfield neural network is realized, and is used for processing 8PSK constellation signals.

[0068] The application comprises the following steps:

[0069] Step 1, complex activation function circuit design;

[0070] The designed activation function is:

[0071]

[0072] The phase diagram of the constellation signal is as shown in Figure 1 The activation function circuit is composed of an absolute value decision basic module, an imaginary part basic module and a real part basic module.

[0073] S1.1 absolute value decision basic module

[0074] The absolute value decision module is composed of an absolute value circuit and a division circuit, and the basic modules are resistors, operational amplifiers and diodes. Figure 2 As shown in the figure, voltage inputs V iI and V iR are applied. For the second operational amplifier, there is:

[0075]

[0076] Wherein R1=2R, and

[0077]

[0078] When V iR > 0, according to the "virtual short" and "virtual break" characteristics of the operational amplifier, the output V o1 of the first operational amplifier can be obtained as -V iR < 0, and the diode is cut off, at this time V o2 = V iR . When V iR < 0, the output V o1 of the first operational amplifier is -V iR > 0, the diode is turned on, at this time, due to "virtual ground", V o1 = 0. At this time, V o2 = -V iR . Therefore, V o2 = |V iR |. For the division circuit, it is mainly composed of a multiplier and an operational amplifier. The multiplier outputs V o3 = V o2V o Based on the "virtual short" and "virtual open" characteristics of operational amplifiers, we can conclude that:

[0079]

[0080]

[0081] V o2 =|V iR |Substituting into the equation yields...

[0082] S1.2 Basic Module of Imaginary Part

[0083] Based on the output of the defined piecewise function, a selection circuit was designed. Its imaginary piecewise functional circuit and characteristics are as follows: Figure 3 As shown, this constitutes the basic unit of the selection circuit. The imaginary part of the output has only four possible values: (-0.9239, -0.3827) ，0.3827， 0.9239). The corresponding inflection point should be (-1). ，0， 1). And the voltage E i Let's set this as a turning point. Based on Kirchhoff's laws and Ohm's law, we can conclude that:

[0084]

[0085] V out =-f(V in -E j (6)

[0086]

[0087]

[0088] Where R zj R xj These are the relevant parameters for the operational amplifier.

[0089] At this point, a current output i is obtained, which needs to be converted to voltage. A transimpedance amplifier (TIA) is connected after the imaginary part basic module. The current flows through the TIA and is converted into voltage. This yields the output V. oI :

[0090] V oI =-iR img (9)

[0091] S1.3 Real Part Basic Module

[0092] Unlike the imaginary part module, the real part output exhibits symmetry about the y-axis, thus requiring adjustments to the circuit structure. The adjusted real part circuit is as follows: Figure 4 As shown;

[0093]

[0094]

[0095] The circuit consists of a basic module and a subtraction circuit. The subtraction circuit calculates the voltage difference between the two input terminals, and then the final real output voltage V is obtained through the voltage divider circuit. oR ;

[0096] Step 2: Design of a one-step computational circuit for blind signal detection based on a complex-valued Hopfield neural network.

[0097] Based on the complex-valued activation function circuit, a one-step calculation circuit for blind signal detection based on the complex-valued Hopfield neural network was designed. It consists of two parts: the activation function circuit and the weight matrix circuit. The weight matrix circuit consists of a memristor array and an operational amplifier.

[0098] S2.1 Memristor Array Configuration

[0099] The basic building block of a memristor array is the memristor, which corresponds to the weight value in the weight matrix. A memristor is like... Figure 5 As shown, a mode selection switch is formed by connecting a PMOS transistor and an NMOS transistor to the two ends of the memristor, which enables programming and calculation of the memristor. V tn and V tp These are the threshold voltages of the NMOS and PMOS transistors, respectively. When V c <V tp When the PMOS transistor is turned on and the NMOS transistor is turned off, the memristor is in programming mode. At this time, we can adjust the value of the memristor by applying voltages of different amplitudes and widths. V c >V tn When the NMOS transistor is turned on and the PMOS transistor is turned off, the memristor is in calculation mode. The conductance of the memristor is denoted as G. When a voltage V is applied to the memristor, the current flowing through the memristor is I = G·V. Since the memristor conductance cannot be negative, but the weight value may be negative, we use the subtraction of a dual-array memristor array to solve the problem of negative weight coefficients. Each weight value is represented by two memristors, namely G and G′. When the weight value A is negative, let G = A and G′ = 2A. Subtracting the two memristors will give a negative weight value. Conversely, let G = 2A and G′ = A to get a positive memristor conductance value.

[0100] S2.2 Signal Blind Detection One-Step Calculation Circuit Design

[0101] The calculation of the weight matrix relies on complex value computation. Given two complex value vectors x = a + bi and y = c + di, the result of complex value multiplication is z = x·y = (ac - bd) + (ad + bc)i, which can be expressed using matrix-vector computation as follows:

[0102]

[0103] The weights are complex values, and the outputs are the real and imaginary parts, respectively. A memristor array is used to map the weight matrix.

[0104] by Figure 6 Taking complex-valued matrix computation as an example, W ire and W jim It can be represented as:

[0105]

[0106] Let G 11 G 22 The value is 2a, G 21 The value is 2b, G 12 Taking the value as b, substituting it into formula (13) yields W. 1re =a,W 1im =b, and substituting the obtained value into formula (12) gives:

[0107]

[0108] Simplifying formula (14) yields the same result as formula (12);

[0109] The one-step calculation circuit for blind signal detection based on complex-valued Hopfield neural network is as follows: Figure 7 As shown. The specific principle of the circuit is as follows.

[0110]

[0111]

[0112] V Yi =F(V Ii )

[0113] =F(-I xi ·R) (17)

[0114] Therefore, the output voltage V is obtained. Yi It is also used as the input to the next round of circuitry;

[0115] The main process of the blind detection circuit is as follows: v is obtained through preprocessing. xj and W, where V xj As the input voltage, the weight matrix W is configured as a memristor array, and the input voltage V xj The current flowing through the memristor array is converted into a current I according to Kirchhoff's laws. xi Then, with the help of TIA, the output is converted into a voltage V. IiThe final output V is obtained after activation by the activation module. Yi Then it is used as the input for the next round; when the neural network reaches a steady state, V Yi It is the final output of the circuit;

[0116] The following section uses a 6-input constellation signal as an example to illustrate the circuit design process. First, the signal is preprocessed to obtain the weight matrix W and the complex-valued input signal X. j (j = 1, 2, ..., 6), using Figure 6 The memristor array maps and configures the weighting coefficients of W, and the input signal is a voltage V. Xj The input is in the form of a form, flows through the memristor array and then into the activation circuit, where it passes through... Figure 2 , 3 The activation circuit composed of 4 and 4 produces a voltage output V. Yi V Yi It then re-enters the circuit as a voltage input until the circuit reaches a steady state, at which point the final V... Yi The overall processing is in Figure 8 It has already been given in the text.

[0117] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

Claims

1. A signal blind detection one-step calculation circuit based on a complex-valued Hopfield neural network, comprising an activation function circuit and a complex-valued Hopfield neural network weight circuit; characterized in that: Based on the energy function blind detection technology, a high-speed signal blind detection circuit is proposed from the perspective of analog circuit, the blind detection technology rule is analyzed, and the corresponding circuit model 1 is established; According to the circuit model 1, the complex-valued weight matrix memristor array circuit and the complex-valued activation function circuit are designed and built; including using the memristor array to design the complex-valued Hopfield weight matrix operation circuit, according to the design rule of the activation function, the corresponding processing module is designed, and each module is established as a circuit model 2; According to the established circuit model 2, the signal blind detection one-step calculation circuit based on the complex-valued Hopfield neural network is designed, the setting relationship of the activation function in the signal blind detection technology is studied, and the circuit model 3 is constructed, and then the signal blind detection one-step calculation circuit based on the complex-valued Hopfield neural network is realized, and used to process 8PSK constellation signals.

2. The one-step computation circuit for blind detection of signals based on complex-valued Hopfield neural network according to claim 1, characterized in that Including the following steps: Step 1, design of complex-valued activation function circuit; The designed activation function is: The activation function circuit is composed of absolute value judgment basic module, imaginary part basic module and real part basic module; S1.1 absolute value judgment basic module The absolute value decision module is composed of an absolute value circuit and a division circuit, and basic modules are resistors, operational amplifiers and diodes; a voltage input V iI and V iR For the second operational amplifier, there are: Where R1=2R, we can get When V iR > 0, according to the "virtual short" and "virtual break" characteristics of the operational amplifier, the first operational amplifier output V o1 = -V iR < 0, the diode is off, at this time V o2 = V iR ; and when V iR < 0, the first operational amplifier output V o1 = -V iR > 0, the diode is on, at this time due to "virtual ground", V o1 = 0, at this time V o2 = -V iR , thus V o2 = |V iR |; for the division circuit, mainly composed of a multiplier and an operational amplifier, the multiplier output V o3 = V o2 V o ; according to the "virtual short" and "virtual break" characteristics of the operational amplifier, it can be obtained: V o2 = |V iR | substitution gives S1.2 imaginary part basic module According to the output of the defined piecewise function, the selection circuit is designed; there are only four cases for the imaginary part output, which are (-0.9239, -0.3827, 0.3827, 0.9239), and the corresponding inflection points should be (-1, 0, 1), and the voltage E j is set as the turning point; according to the analysis of Kirchhoff's law and Ohm's law: V out = -f(V in -E j ) (6) wherein R zj • R xj are amplifier related parameters; At this time, the current output i is obtained, and a current-to-voltage operation needs to be performed. The current flows through a trans-impedance amplifier (TIA) connected after the imaginary part basic module, and the current is converted into a voltage, thereby obtaining the output V oI : V oI = -iR img (9S1.3 Real part basic module The difference between the imaginary part module and the real part module is that the real part output presents the characteristic of symmetry about the y-axis, so the circuit structure needs to be adjusted; The circuit is composed of basic modules and a subtraction circuit. The subtraction circuit calculates the difference between the input voltages at both ends, and then the final real part output voltage V is obtained through the final voltage dividing circuit oR ; Step 2, design of signal blind detection one-step calculation circuit based on complex-valued Hopfield neural network On the basis of the complex-valued activation function circuit, the signal blind detection one-step calculation circuit based on the complex-valued Hopfield neural network is designed, which is composed of activation function circuit and weight matrix circuit, and the weight matrix circuit is composed of memristor array and operational amplifier; S2.1 memristor array configuration The basic unit of the memristor array is the memristor, which corresponds to the weight value in the weight matrix. The memristor is connected to a PMOS transistor and an NMOS transistor to form a mode selection switch, which can program and calculate the memristor. V tn and V tp are the threshold voltages of the NMOS transistor and the PMOS transistor, respectively. When V c <V tp , the PMOS transistor is on and the NMOS transistor is off, and the memristor is in programming mode. At this time, we can adjust the value of the memristor by applying voltages of different amplitudes and widths. When V c >V tn , the NMOS transistor is on and the PMOS transistor is off, and the memristor is in calculation mode. The conductance value of the memristor is denoted as G. When a voltage V is applied to the memristor, the current I flowing through the memristor is I = G·V. The memductance value of the memristor cannot be negative, but the weight value can be negative. To solve the problem of negative weight coefficients, we use a double-column memristor array to subtract. Each weight value is represented by two memristors, G and G'. When the weight value A is negative, let G = A and G' = 2A. The subtraction of the two memristors gives a negative weight value. Conversely, let G = 2A and G' = A to get a positive memductance value. S2.2 signal blind detection one-step calculation circuit design The calculation of weight matrix depends on complex-valued calculation, there are two complex-valued vectors x=a+bi, y=c+di, then the result of complex-valued multiplication is z=x·y=(ac-bd)+(ad+bc)i, which can be represented by matrix-vector calculation: The weight value is complex-valued, and the output is real part and imaginary part respectively, and the weight matrix is mapped by using the memristor array For example, W jre and W jim may be expressed as: Let G 11 , G 22 = 2a, G 21 = 2b, G 12 = b, substituting into equation (13) gives W 1re = a, W 1im = b, and substituting the values obtained into equation (12) gives: Simplify formula (14), the result is the same as formula (12); The specific principle of the signal blind detection one-step calculation circuit based on the complex-valued Hopfield neural network is as follows: V Yi = F(V Ii ) = F(-I xi • R) (17) Thus, the output voltage V Yi is also used as the input to the next round of circuits; The main process of blind detection circuit is as follows: through preprocessing, V xj and W are obtained, wherein V xj is input voltage, W is weight matrix configured as a memristor array, input voltage V xj flows through the memristor array, and is transformed into current I xi according to Kirchhoff's law, then the output is transformed into voltage V Ii by means of a TIA, and the final output V Yi is obtained through activation of an activation module, then it is used as input of a new round; when the neural network reaches a steady state, V Yi is the final output of the circuit.

Images