TIADC system time sequence deviation mismatch detection method based on GA-BP neural network

Abstract

The invention discloses a TIADC system time sequence deviation mismatch detection method based on an GA-BP neural network, and provides a BP neural network time sequence deviation mismatch calibration method based on AGA algorithm optimization in order to solve the time sequence deviation mismatch problem existing in a TIADC system. According to the invention, the trained BP neural network is used to detect the time sequence deviation mismatch in the TIADC system, and the variable delay line is used to calibrate the TIADC system. According to the method, the BP neural network is optimized by adopting the AGA, so that the training speed of the BP neural network is increased, and the detection precision of the time sequence deviation mismatch in the system is improved. Compared with other methods, the method provided by the invention has excellent detection speed and precision. According to the invention, a 18-bit 1GS / S 4-channel TIADC system is simulated, and time sequence deviation mismatch in the system is calibrated.

Description

Technical Field

[0001] The invention relates to the technical field of digital signal processing and high-speed and high-precision TIADC, and in particular to a TIADC system timing deviation mismatch detection method based on an AGA-BP neural network. Background Art

[0002] With the rapid development of information technology, digitization has become an irreversible trend. In the digitization process, ADC, which converts analog signals into digital signals, has become a key component. However, it is not easy to strike a balance between high conversion accuracy and high sampling rate. In order to overcome this challenge, time-interleaved ADC (TIADC) has emerged, providing a more flexible and efficient solution for digital signal processing, and also promoting the wider application of digital systems in various fields.

[0003] The channel mismatch problem in the TIADC system is caused by several factors, including the offset voltage of the internal sub-ADC front-end holding circuit, the delay of the clock layout, and the input signal. The mismatch problem in the TIADC system can be divided into three categories: offset mismatch, gain mismatch, and timing skew mismatch. The calibration of offset mismatch and gain mismatch is relatively mature, but the calibration of timing offset mismatch is more complicated because its causes are different from the first two mismatches. Timing skew mismatch calibration technology also has many disadvantages and limitations.

[0004] The prior art includes the following methods to calibrate the offset, gain and timing tilt mismatch generated in the TIADC system: 1. Select the first channel as the reference channel and add digital offsets o d,i , g d,i and t d,i To detect the offset, gain and timing skew mismatch in the system. At the same time, the method realizes the mismatch calibration of the system through the least mean square (LMS) algorithm and variable delay line; 2. The peak detection algorithm is adopted to simplify the algorithm and slightly improve the speed; 3. The mismatch estimation technology based on fast Fourier transform is used to estimate the mismatch; 4. The binary search algorithm is used to replace the LMS algorithm, which significantly improves the convergence speed and accuracy of timing skew mismatch detection; 5. The two-stage analog-to-digital hybrid compensation technology is used to calculate the timing skew mismatch; 6. The interpolation filter, Hilbert filter and M-1 adaptive filter are used to calibrate the gain, timing offset and bandwidth mismatch in the TIADC system; 7. The compensation method based on the back propagation (BP) neural network is adopted. The algorithm can adaptively track the mismatch changes in the TIADC system without user intervention and achieve good calibration results.

[0005] However, the prior art still has the following defects:

[0006] 1. During the mismatch calibration process, the calculations for offset, gain mismatch, and timing tilt mismatch are performed separately, so these calculations must be performed separately, making the calibration process more complicated;

[0007] 2. The bias and gain mismatch parameters must be optimized using the LMS algorithm, which converges slowly;

[0008] 3. Reference channel detection is still required, and the problem of complex algorithm structure is not solved;

[0009] 4. The resource consumption of the binary search algorithm is much higher than that of the LMS algorithm, and it has a higher computational complexity;

[0010] 5. Currently, most calibration methods have problems such as poor calibration effectiveness, high resource consumption, and slow convergence speed. Summary of the invention

[0011] The purpose of the present invention is to provide a TIADC system timing deviation mismatch detection method based on AGA-BP neural network, which has obvious improvements in resource consumption and convergence speed. At the same time, the method also solves the problem that the traditional BP neural network has a slow convergence speed and cannot converge to the global optimum during the training process.

[0012] To achieve the above functions, the present invention designs a TIADC system timing deviation mismatch detection method based on AGA-BP neural network, and performs the following steps S1 to S4 to complete the detection and calibration of the timing deviation mismatch of the TIADC system:

[0013] Step S1: Based on the AGA algorithm, the initial weight parameters and bias parameters of the BP neural network are used as genes in the AGA algorithm, and the chromosomes in the AGA algorithm are formed accordingly; through the crossover and mutation of chromosomes, the chromosome with the highest fitness is iteratively searched as the optimal chromosome;

[0014] Step S2: Based on the initial weight parameters and bias parameters corresponding to the optimal chromosome, a BP neural network for timing deviation mismatch detection is established, the non-uniform samples from the sub-ADC are input into the BP neural network, the BP neural network is trained and verified, and a trained BP neural network is obtained;

[0015] Step S3: using the trained BP neural network to perform timing deviation mismatch detection on the TIADC system input signal;

[0016] Step S4: with respect to the timing deviation mismatch of the TIADC system, a variable delay line is used to correct the timing deviation mismatch of the TIADC system, thereby completing the calibration of the timing deviation mismatch of the TIADC system.

[0017] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0018] 1. The AGA-BP neural network algorithm requires fewer training rounds and converges significantly faster than the traditional BP neural network algorithm;

[0019] 2. The resource consumption of AGA-BP neural network in physical memory, swap space and virtual memory is much smaller than that of traditional BP neural network, which solves the problem of high resource consumption of traditional BP neural network;

[0020] 3. From the perspective of timing tilt mismatch detection, compared with other algorithms, the timing tilt mismatch detected by AGA-BP neural network is closer to the true value and has better detection effect. From the perspective of algorithm running time, AGA-BP neural network only needs 0.0094s, which is much faster than other algorithms;

[0021] 4. The input signals at different normalized frequencies were simulated, and the results showed that this method can effectively calibrate the input signals within the entire Nyquist frequency range;

[0022] 5. After system calibration, the introduced timing tilt mismatch was corrected. The TIADC system effective number of bits (ENOB) was improved from 7.8 bits to 17.3 bits, and the spurious-free dynamic range (SFDR) was improved from 48.8 to 108.7 dB. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a schematic diagram of a TIADC system provided according to an embodiment of the present invention;

[0024] Figure 2 is a schematic diagram of the detection and calibration principle of the timing deviation mismatch of the TIADC system provided in accordance with an embodiment of the present invention;

[0025] Figure 3 It is an overall flow chart of a TIADC system timing deviation mismatch detection method based on an AGA-BP neural network provided according to an embodiment of the present invention;

[0026] Figure 4 is a schematic diagram of a chromosome provided according to an embodiment of the present invention;

[0027] Figure 5 is a schematic diagram of selection probabilities of different chromosomes arranged according to fitness values ​​provided in an embodiment of the present invention;

[0028] Figure 6 is a schematic diagram of the structure of a BP neural network provided according to an embodiment of the present invention;

[0029] Figure 7is a schematic diagram of the structure of a single hidden layer neuron provided according to an embodiment of the present invention;

[0030] Figure 8 is a schematic diagram of the structure of an output layer neuron provided according to an embodiment of the present invention;

[0031] Fig. 9 It is a simulation diagram of timing deviation mismatch before and after BP neural network training provided by an embodiment of the present invention;

[0032] Fig.10 is a simulation diagram of the mean square error of the BP neural network in different rounds of training according to an embodiment of the present invention;

[0033] Fig.11 is a simulation diagram of the mean square error of the BP neural network in different rounds of verification according to an embodiment of the present invention;

[0034] Fig.12 is a spectrum output simulation diagram of the TIADC system before and after calibration according to an embodiment of the present invention;

[0035] Fig.13 is a graph showing a relationship between a signal-to-noise ratio and an input signal frequency according to an embodiment of the present invention;

[0036] Fig.14 is a graph showing a relationship between a spurious free dynamic range and an input signal frequency according to an embodiment of the present invention;

[0037] Fig.15 is a diagram showing the relationship between the effective number of bits and the input signal frequency provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0039] The embodiment of the present invention provides a TIADC system timing deviation mismatch detection method based on AGA-BP neural network, which performs the following steps S1 to S4 to complete the detection and calibration of the timing deviation mismatch of the TIADC (time interleaved analog-to-digital converter) system, wherein the TIADC system schematic diagram refers to Figure 1 , the detection and calibration schematic diagram of the timing deviation mismatch of the TIADC system refers to Figure 2 ;

[0040] Step S1: Based on the AGA algorithm, the initial weight parameters and bias parameters of the BP (back propagation) neural network are used as genes in the AGA algorithm, and the chromosomes in the AGA algorithm are correspondingly formed; through the crossover and mutation of chromosomes, the chromosome with the highest fitness is iteratively searched as the optimal chromosome;

[0041] The AGA algorithm is a search algorithm for solving optimization problems. By referring to biological evolution, the AGA algorithm simulates the problem to be solved as a biological evolution problem. The next generation of solutions is generated through selection, hybridization and mutation. In successive generations, individuals with lower fitness are gradually eliminated, while individuals with higher fitness are selected. Through this process, after N generations, it is very likely that individuals with higher fitness values ​​will evolve, ultimately leading to the optimal result of the objective function. The present invention uses the AGA algorithm to optimize the initial weights and biases of the BP neural network, accelerate its convergence speed, and improve its performance.

[0042] The steps of using AGA to optimize BP neural network are as follows:

[0043] (1) Establish the initial population; (2) Calculate the adaptive value; (3) Select chromosomes; (4) Adaptive crossover; (5) Adaptive mutation; (6) Find the optimal chromosome; (7) Establish and train the BP neural network; (8) Detect timing deviation. For the overall flow chart, refer to Figure 3 .

[0044] The specific steps of step S1 are as follows:

[0045] Step S1.1: Refer to Figure 4 , respectively, with the initial weight parameters w of the BP neural network 1 and w 2 , bias parameter b 1 and b 2 As a pair of genes, genes form chromosomes; a set of chromosomes is randomly generated to form the initial population;

[0046] Step S1.2: Calculate the fitness value of each chromosome in the initial population as follows:

[0047] F i =|O e -O p |

[0048] In the formula, F i represents the fitness value of the i-th chromosome, O e Represents the predicted output after BP neural network training, O p represents the expected output of the BP neural network; the difference between the two represents the fitness value function in the algorithm, which measures the quality of the generated chromosome and helps in subsequent chromosome selection, crossover and mutation. The lower the fitness value, the closer the predicted value is to the expected value of the neural network, indicating that the chromosome quality is better.

[0049] Step S1.3: Select chromosomes; Chromosome selection in genetic algorithms refers to selecting chromosomes with high fitness from the current population as parents of the next generation to retain excellent gene combinations. The exponential selection sorting method assigns a selection probability to each individual, solving the problem of zero probability of individual selection. Therefore, the exponential selection method is used to select the required chromosomes.

[0050] Since the purpose of the AGA algorithm in the present invention is to optimize the weights and biases in the BP neural network, the smaller the fitness value, the better the chromosome. In chromosome sequencing, N represents the best chromosome, and 1 represents the worst chromosome.

[0051] The specific method of index sorting selection is as follows:

[0052] Sort the fitness values ​​of each chromosome in the initial population in descending order, referring to Figure 5 , the chromosome with the highest current fitness value is taken as the optimal chromosome, and the chromosome with the lowest fitness value is taken as the worst chromosome; the selection probability of each chromosome is calculated as follows:

[0053]

[0054] In the formula, N represents the best chromosome, 1 represents the worst chromosome, and p i represents the selection probability of the i-th chromosome, c represents the parameter of the exponential sorting selection method, and its value is between 0 and 1;

[0055] Generate a random number. If the random number is in sum i-1 and sum i If , then the i-th chromosome is selected.

[0056] Step S1.4: Adaptive crossover is performed on chromosomes. Adaptive crossover is an evolutionary strategy in the AGA algorithm. Its purpose is to dynamically adjust the crossover probability according to the fitness of individuals in the population to improve the performance of the algorithm. Select a pair of chromosomes, calculate the crossover probability of the chromosomes, and generate a random number; the random number represents the position of the crossover gene on the chromosome that needs to be crossed; if the random number is less than the crossover probability, the crossover operation is performed, and the crossover probability p c The formula is as follows:

[0057]

[0058] In the formula, f min is the minimum fitness value in the population, f′ is the minimum fitness value of the two chromosomes to be crossed, and f mean is the average fitness value of the population; to ensure that it is greater than f mean The fitness of chromosomes is crossed to avoid the search falling into the local optimum, and k1 and k3 are both set to 1;

[0059] The crossover value in the adaptive crossover operation is as follows. The crossover method simulates binary crossover to further speed up the population update:

[0060]

[0061] In the formula, c1 and c2 are the values ​​at the intersection after the crossover is completed, p1 and p2 are the values ​​at the original intersection, u is the generated random number, and n is the crossover distribution index; repeat the above formula to complete the adaptive crossover operation;

[0062] Step S1.5: Adaptive mutation is performed on chromosomes. Adaptive mutation is an important strategy of the AGA algorithm, and its purpose is to dynamically adjust the probability or mode of mutation according to individual fitness. The process of adaptive mutation is basically the same as the crossover process described above; the mutation probability is calculated as follows:

[0063]

[0064] In the formula, f min is the minimum fitness value in the population, f is the fitness value of the chromosome that needs to be mutated, and fmea n is the average fitness value of the population; to avoid the search falling into the local optimum, k2 and k4 in the formula are set to 0.5;

[0065] The mutation operation is to select the mutation point that needs to be modified on the selected chromosome and mutate the gene value; the crossover value in the adaptive mutation operation is as follows:

[0066] 9=p+6Δmax

[0067]

[0068] In the formula, q represents the value of the mutation point after mutation, p represents the value of the original mutation point on the chromosome, Δ max is the maximum disturbance value of the mutation point, u is the generated random number, and n is the cross-distribution index; repeat the above formula to complete the adaptive mutation operation;

[0069] Step S1.6: Repeat steps S1.2 to S1.5, and after a preset number of iterations, the optimal chromosome is obtained.

[0070] Step S2: Based on the initial weight parameters and bias parameters corresponding to the optimal chromosome, a BP neural network for timing deviation mismatch detection is established, the non-uniform samples from the sub-ADC are input into the BP neural network, the BP neural network is trained and verified, and a trained BP neural network is obtained;

[0071] BP neural network is one of the most classic artificial neural network models, also known as multilayer perceptron. It is a multilayer feedforward neural network trained by the back propagation algorithm. In practical applications, the learning performance of the neural network can be improved by adjusting the structure and training parameters of the neural network. The BP neural network consists of three parts: input layer, hidden layer and output layer. The number of neurons in the input layer is determined by the number of input data in the system, while the number of neurons in the hidden layer is determined by the specific network training requirements. Usually, the output layer has only one neuron. In this embodiment, the number of neurons in the input layer is 58, the number of neurons in the hidden layer is 40, and the number of neurons in the output layer is 1.

[0072] The structural diagram of a single hidden layer neuron is shown in Figure 2. Figure 7 ; The schematic diagram of the output layer neuron structure is shown in Figure 2. Figure 8 ;

[0073] The present invention uses an AGA-optimized BP neural network to estimate the timing deviation mismatch in the TIADC system. During the training process of the BP neural network, a sine wave signal with a frequency of 47MHz is selected as the input signal of the network training stage. There are multiple options for the input training signal, including a signal tone sine wave, a multi-tone sine wave, a low-frequency sine wave, a high-frequency sine wave, and other stable sine wave signals. It should be pointed out that in this system, a band-limited noise signal cannot be used as a training of the neural network. A band-limited noise signal is a noise signal with a specific range. It is different from Gaussian white noise. Gaussian white noise has equal power density at all frequencies, while band-limited noise has a higher power density in one frequency range and a lower power density in other frequency ranges. Band-limited noise signals are usually confined to a certain frequency band. During the training process of the BP neural network, the input signal has a great influence on the training process. From the perspective of the noise signal, the noise signal can increase the robustness and generalization ability of the network and help the network learn more general patterns. However, the noise signal may also interfere with the learning process, resulting in performance degradation. From the perspective of a stable sinusoidal signal, this signal makes it easier for the neural network to converge, especially when dealing with deterministic tasks. The present invention aims to solve the problems of slow training time and slow detection of time deviation mismatch in the BP neural network. Therefore, it is more appropriate to select a stable sinusoidal signal as the input signal of the BP neural network.

[0074] The training of BP neural network includes two processes: forward propagation and back propagation. The relationship between the parameters of the hidden layer neurons of BP neural network in forward propagation is as follows:

[0075]

[0076] In the formula, is the output of the hidden layer, Sigmoid is the S-shaped growth curve, representing the activation function, z i It is a node of a nerve cell. is the hidden layer weight, x i is the input signal, b i is the hidden layer bias; and b i From genes on the optimal chromosome;

[0077] The relationship between the parameters in the output layer neurons of the BP neural network in forward propagation is as follows:

[0078]

[0079] In the formula, is the output of the output layer neurons, Sigmoid is the activation function, is the node in the neuron, is the weight of the output layer, is the output of the hidden layer, b2 is the bias of the output layer;

[0080] The back propagation process of the BP neural network aims to obtain the error between the actual output of the neural network and the expected output, and feed the error back to the neural network. The gradient descent algorithm is used to update the parameters of each layer in the BP neural network, adjust the connection weights and bias of each layer, and make the output of the BP neural network gradually approach the expected output.

[0081] After the forward propagation of the neural network is completed, if the difference between the network output and the expected output value does not meet the training accuracy requirements, the error back propagation process will be entered. During the back propagation process, the mean square error (MSE) value L is calculated and returned to the previous node of the BP neural network. The mean square error value L is as follows:

[0082]

[0083] In the formula, is the final output of the output layer neuron, and o is the expected output value;

[0084] After calculating the mean square error value L of the BP neural network, the mean square error value L is passed to the input layer, hidden layer and output layer of the BP neural network; the gradient descent method is used to update the connection weights and biases between the layers to minimize the mean square error value L of the BP neural network; the calculation formula for error back propagation and parameter update of each layer is as follows:

[0085]

[0086] Among them, w l is the connection weight matrix between the lth layer and the l+1th layer in the BP neural network, bl It is the bias between layer l and layer l+1 in the BP neural network, μ is the learning rate, and its value is greater than 0 and less than 1;

[0087] After updating the parameters of the BP neural network, check whether the output of the BP neural network meets the preset training accuracy; if it meets the training accuracy requirements, the training of the BP neural network is completed; if it does not meet the preset training accuracy requirements, repeat the forward propagation, back propagation and training process until the preset training accuracy requirements are met.

[0088] Step S3: using the trained BP neural network to perform timing deviation mismatch detection on the TIADC system input signal;

[0089] In step S3, the trained BP neural network is used to input the actual ADC error data into the trained BP neural network to detect the timing deviation mismatch. After verification by the BP neural network, the final timing deviation mismatch τ is obtained. i .

[0090] Step S4: with respect to the timing deviation mismatch of the TIADC system, a variable delay line is used to correct the timing deviation mismatch of the TIADC system, thereby completing the calibration of the timing deviation mismatch of the TIADC system.

[0091] In step S4, a variable delay line is inserted into the signal path, and the delay is adjusted until each sub-ADC in the TIADC system works at the expected time, thereby completing the calibration of the timing offset mismatch of the TIADC system.

[0092] The following is an application example of the present invention:

[0093] In order to verify the calibration performance of the timing deviation mismatch detection method of the TIADC system based on the AGA-BP neural network designed by the present invention, an 18-bit, 1GS / s four-channel TIADC system was simulated. In order to simplify the analysis of the timing tilt mismatch in the TIADC system, it is assumed that the offset mismatch and gain mismatch have no effect on the system. In this embodiment, a sinusoidal signal with an input frequency of 47.71MHz is used as the input of the TIADC system, and the timing deviation mismatch of different channels in the system is configured as [0, 0.0335Ts, 0.017Ts, 0.023Ts]. The parameter c of the index sorting selection method of AGA is set to 0.9, and the distribution index n in the adaptive crossover and adaptive change is set to 1; the specific analysis is as follows:

[0094] Reference to the simulation diagram of timing deviation mismatch before and after BP neural network training Fig. 9 ;exist Fig. 9In the figure, the horizontal axis is the ratio of the timing tilt to the TIADC system sampling period, and the vertical axis error is the absolute value of the difference between the actual timing tilt and the timing tilt detected by the AGA-BP neural network. Fig. 9 In the example, before the AGA-BP neural network training process is completed, the timing deviation detected by the BP neural network always does not match the actual timing deviation in the system. However, after the AGA-BP neural network training process is completed, the difference between the timing deviation detected by the neural network and the actual timing deviation in the system is always 0. Fig. 9 The effectiveness of the trained AGA-BP neural network in detecting timing skew mismatch was verified.

[0095] Reference to the simulation diagram of mean square error of BP neural network in different rounds of training Fig.10 , the mean square error simulation diagram of BP neural network in different rounds of verification refers to Fig.11 , Fig.10 and 11 The number of rounds required for the training and validation of the traditional BP neural network and the AGA-BP neural network is shown. Fig.10 and 11 In the figure, the horizontal axis is the training cycle of the neural network, and the vertical axis is the mean square error. Fig.10 and 11 It can be seen that in the training and verification process of the neural network, the relationship between the training cycle and the mean square error is generally consistent. In contrast, the traditional BP neural network requires an average of 52 rounds of training to converge, while the AGA-BP neural network requires an average of about 14 rounds of training to converge. This shows that the AGA-BP neural network algorithm proposed in the present invention requires fewer training rounds and converges significantly faster than the traditional BP neural network algorithm.

[0096] Table 1 shows the sensitivity to various frequencies in the input signal:

[0097] Table 1: Sensitivity to various frequencies in the input signal

[0098]

[0099]

[0100] Taking the timing deviation mismatch of the second channel as an example, it is set to 0.0335Ts. In Table 1, Ts represents the sampling period of the TIADC system, and its value is 106. It can be seen from Table 1 that when the input signal during the neural network training process is a single-tone sine wave, the BP neural network can detect the timing deviation mismatch in the TIADC system, regardless of whether the input is a single-tone sine wave or a multi-tone sine wave. If the neural network is trained with a multi-tone sine wave, the neural network can use a single-tone sine wave to estimate the timing deviation. At the same time, when the neural network is trained with a high-frequency input signal, the system can successfully detect the timing deviation mismatch in the TIADC system under a high-frequency input signal. When the neural network is trained with a low-frequency input signal, the system can also detect the timing deviation mismatch in the TIADC system under a high-frequency input signal. However, compared with the previous case, the timing tilt mismatch error detected by the system is larger.

[0101] Table 2 compares the resource consumption of physical memory, swap space and virtual memory by the traditional BP neural network and the AGA-BP neural network proposed in this invention:

[0102] Table 2: Comparison of resource consumption between traditional BP neural network and AGA-BP neural network

[0103]

[0104] As can be seen from Table 2, the resource consumption of AGA-BP neural network in three aspects is much smaller than that of traditional BP neural network, which solves the problem of high resource consumption of traditional BP neural network.

[0105] Table 3 compares the AGA-BP neural network algorithm with other methods:

[0106] Table 3: Comparison with other methods

[0107]

[0108] It can be seen from Table 3 that although the mean square error of AGA-BP neural network and traditional BP neural network can converge to 10 -11 of magnitude, but the training time they require is very different. From the perspective of the number of training rounds, the traditional BP neural network requires 52 rounds to converge, while the AGA-BP neural network only needs 14 rounds to converge. From the perspective of training time, the traditional BP neural network requires 20 seconds of training, while the AGA-BP neural network only needs 1 second to complete training. From the perspective of timing skew mismatch detection, compared with other algorithms, the timing skew mismatch detected by the AGA-BP neural network is closer to the true value and has a better detection effect. From the perspective of the algorithm's running time, the AGA-BP neural network only needs 0.0094s, which is much faster than other algorithms.

[0109] In summary, the TIADC system timing deviation mismatch detection method based on AGA-BP neural network proposed by the present invention has fewer training rounds, shorter training time, and faster convergence speed. The algorithm consumes fewer resources. The algorithm has a better detection effect on timing tilt mismatch.

[0110] After the timing deviation is detected, the timing tilt in the channel is calibrated through the variable delay line. The spectrum output simulation diagram of the TIADC system before and after calibration is shown in the figure below. Fig.12 As shown, the black line is the output spectrum of the TIADC system before calibration, and the red line is the output spectrum of the system after calibration. Fig.12 It can be seen that after the system calibration, the introduced timing tilt mismatch has been corrected. At the same time, the TIADC system effective number of bits (ENOB) has increased from 7.8 bits to 17.3 bits, and the spurious-free dynamic range (SFDR) has increased from 48.8 to 108.7dB.

[0111] The relationship between signal-to-noise ratio (SNR) and input signal frequency is shown in the figure below. Fig.13 ; For the relationship between spurious free dynamic range (SFDR) and input signal frequency, refer to Fig.14 ; Refer to the relationship between the effective number of bits (ENOB) and the input signal frequency Fig.15 ; Fig.13 , 14 Figures 1 and 15 show the parameter comparison of different normalized frequency (Fin / Fs) input signals before and after TIADC system calibration. Fig.13 , 14 As shown in Figure 15, after calibration, the signal-to-noise ratio (SNR), spurious-free dynamic range (SFDR), and effective number of bits (ENOB) of the TIADC system are greatly improved. The output of the system performs well over the entire Nyquist range, and the input signal near the edge of the Nyquist frequency is also well calibrated.

[0112] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A TIADC system timing deviation mismatch detection method based on AGA-BP neural network, characterized in that: Execute the following steps S1 to S4 to complete the detection and calibration of the timing deviation mismatch of the TIADC system: Step S1: Based on the AGA algorithm, the initial weight parameters and bias parameters of the BP neural network are used as genes in the AGA algorithm, and the chromosomes in the AGA algorithm are formed accordingly; through the crossover and mutation of chromosomes, the chromosome with the highest fitness is iteratively searched as the optimal chromosome; Step S2: Based on the initial weight parameters and bias parameters corresponding to the optimal chromosome, a BP neural network for timing deviation mismatch detection is established, the non-uniform samples from the sub-ADC are input into the BP neural network, the BP neural network is trained and verified, and a trained BP neural network is obtained; Step S3: using the trained BP neural network to perform timing deviation mismatch detection on the TIADC system input signal; Step S4: with respect to the timing deviation mismatch of the TIADC system, a variable delay line is used to correct the timing deviation mismatch of the TIADC system, thereby completing the calibration of the timing deviation mismatch of the TIADC system.

2. The method for detecting timing deviation mismatch of TIADC system based on AGA-BP neural network according to claim 1, characterized in that: The specific steps of step S1 are as follows: Step S1.1: Use the initial weight parameters w of the BP neural network respectively 1 and w 2 , bias parameter b 1 and b 2 As a pair of genes, genes form chromosomes; a set of chromosomes is randomly generated to form the initial population; Step S1.2: Calculate the fitness value of each chromosome in the initial population as follows: F i =|O e -Oh p | In the formula, F i represents the fitness value of the i-th chromosome, O e Represents the predicted output after BP neural network training, O p Represents the expected output of the BP neural network; Step S1.3: Sort the fitness values ​​of each chromosome in the initial population in descending order, take the chromosome with the highest current fitness value as the best chromosome, and take the chromosome with the lowest fitness value as the worst chromosome; calculate the selection probability of each chromosome as follows: In the formula, N represents the best chromosome, 1 represents the worst chromosome, and p i represents the selection probability of the i-th chromosome, c represents the parameter of the exponential sorting selection method, and its value is between 0 and 1; Step S1.4: Perform adaptive crossover on chromosomes, select a pair of chromosomes, calculate the crossover probability of chromosomes, and generate random numbers; The random number indicates the position of the crossover gene on the chromosome that needs to be crossed; if the random number is less than the crossover probability, the crossover operation is performed, and the crossover probability p c The formula is as follows: In the formula, f min is the minimum fitness value in the population, f′ is the minimum fitness value of the two chromosomes to be crossed, and f mean is the average fitness value of the population; k1 and k3 are both set to 1; The crossover value in the adaptive crossover operation is as follows: In the formula, c1 and c2 are the values ​​at the intersection after the crossover is completed, p1 and p2 are the values ​​at the original intersection, u is the generated random number, and n is the crossover distribution index; repeat the above formula to complete the adaptive crossover operation; Step S1.5: Perform adaptive mutation on the chromosome and calculate the mutation probability as follows: In the formula, f min is the minimum fitness value in the population, f is the fitness value of the chromosome that needs to mutate, and f mean is the average fitness value of the population; k2 and k4 in the formula are set to 0.5; The crossover value in the adaptive mutation operation is as follows: q=p+δΔ max In the formula, q represents the value of the mutation point after mutation, p represents the value of the original mutation point on the chromosome, Δ max is the maximum disturbance value of the mutation point, u is the generated random number, and n is the cross-distribution index; repeat the above formula to complete the adaptive mutation operation; Step S1.6: Repeat steps S1.2 to S1.5, and after a preset number of iterations, the optimal chromosome is obtained.

3. The method for detecting timing deviation mismatch of TIADC system based on AGA-BP neural network according to claim 1, characterized in that: The BP neural network consists of three parts: input layer, hidden layer and output layer. The number of neurons in the input layer is 58, the number of neurons in the hidden layer is 40, and the number of neurons in the output layer is 1.

4. The method for detecting timing deviation mismatch of TIADC system based on AGA-BP neural network according to claim 1, characterized in that: A sine wave signal with a frequency of 47 MHz is used as the input signal in the BP neural network training stage.

5. The method for detecting timing deviation mismatch of TIADC system based on AGA-BP neural network according to claim 1, characterized in that: The training of the BP neural network in step S2 includes two processes: forward propagation and back propagation. The relationship between the parameters of the hidden layer neurons of the BP neural network in the forward propagation is as follows: In the formula, is the output of the hidden layer, Sigmoid is the activation function, z i It is a node of a nerve cell. is the hidden layer weight, x i is the input signal, b i is the hidden layer bias; and b i From genes on the optimal chromosome; The relationship between the parameters in the output layer neurons of the BP neural network in forward propagation is as follows: In the formula, is the output of the output layer neurons, Sigmoid is the activation function, is the node in the neuron, is the weight of the output layer, is the output of the hidden layer, b2 is the bias of the output layer; During the back propagation process, the mean square error value L is calculated and returned to the previous node of the BP neural network. The mean square error value L is as follows: In the formula, is the final output of the output layer neuron, and o is the expected output value; After calculating the mean square error value L of the BP neural network, the mean square error value L is transmitted to the input layer, hidden layer and output layer of the BP neural network; The gradient descent method is used to update the connection weights and biases between layers to minimize the mean square error value L of the BP neural network; the calculation formula for error back propagation and parameter update of each layer is as follows: Among them, w l is the connection weight matrix between the lth layer and the l+1th layer in the BP neural network, b l It is the bias between layer l and layer l+1 in the BP neural network, μ is the learning rate, and its value is greater than 0 and less than 1; After updating the parameters of the BP neural network, check whether the output of the BP neural network meets the preset training accuracy; if it meets the training accuracy requirements, the training of the BP neural network is completed; if it does not meet the preset training accuracy requirements, repeat the forward propagation, back propagation and training process until the preset training accuracy requirements are met.

6. The method for detecting timing deviation mismatch of TIADC system based on AGA-BP neural network according to claim 1, characterized in that: In step S3, the trained BP neural network is used to input the actual ADC error data into the trained BP neural network to detect the timing deviation mismatch. After verification by the BP neural network, the final timing deviation mismatch τ is obtained. i .

7. The method for detecting timing deviation mismatch of TIADC system based on AGA-BP neural network according to claim 1, characterized in that: In step S4, a variable delay line is inserted into the signal path, and the delay is adjusted until each sub-ADC in the TIADC system works at the expected time, thereby completing the calibration of the timing offset mismatch of the TIADC system.

Images