An ant colony optimization time delay neural network pipelined ADC background calibration method

Abstract

The application discloses a pipeline ADC background calibration method of an ant colony optimization time delay neural network, relates to the technical field of ADC calibration, and comprises the following steps: connecting a to-be-calibrated pipeline ADC and a high-precision reference ADC to a same signal source, combining a time delay unit and a three-layer BP neural network to construct a time delay neural network, connecting the output of the to-be-calibrated pipeline ADC to the time delay unit of the time delay neural network, and connecting the output of the time delay unit and the output of the to-be-calibrated pipeline ADC to input data of neural network training; performing dimension optimization filtering on training data through an ant colony algorithm; performing global optimization on initial configuration of time delay neural network weights and biases through the ant colony algorithm; and outputting a calibrated result by the time delay neural network.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of ADC calibration, in particular to a pipeline ADC background calibration method based on an ant colony optimization time delay neural network. BACKGROUND

[0002] Pipelined ADCs are widely used in communication, aerospace and measurement instruments due to their high speed and high throughput. However, errors such as offset, mismatch, thermal noise in the circuit will cause nonlinearity of the output, and the errors in the front stage of the pipelined ADC will be transmitted to the rear stage, limiting the performance of the pipelined structure. In order to maintain high speed while pursuing high precision output, it is particularly necessary to use calibration technology. Compared with analog calibration circuits, which are sensitive to process changes and environmental influences and lack flexibility, digital calibration technology has the advantages of strong noise resistance, good process compatibility, low power consumption, small area and programmable flexibility, and has become the mainstream of ADC calibration.

[0003] At present, the widely used digital calibration methods mainly include dither-based calibration, PN injection-based calibration, split architecture-based calibration and reference circuit-based calibration. Among them, the calibration methods based on dither and PN injection are simple to implement, but they are poor for serious nonlinearity; the performance of the split architecture-based calibration method is limited by the best performance of the two channels; the reference circuit-based calibration method needs to construct a reference circuit, and the performance of the filter also constrains the calibration effect.

[0004] In recent years, with the wide application of artificial intelligence, scholars have begun to try to apply neural network technology to ADC calibration and have achieved certain research results, such as using BP neural network, convolutional neural network and recurrent neural network to calibrate ADC errors. There are also many achievements in the optimization of neural network calibration models, such as neuron pruning technology, shared neuron technology and binary neural network. However, the above optimization methods are more dependent on experience adjustment, and it is necessary to optimize the neural network calibration model using an optimization algorithm. Some researchers use particle swarm optimization combined with SGD to optimize the neural network and obtain better calibration effect. However, the particle swarm optimization is easily affected by the historical optimal solution and is sensitive to the initial value. Improper initial value may lead to slow convergence or failure to converge, and current researches rarely involve the feature optimization problem that may occur when neural networks are applied to ADC error calibration. SUMMARY

[0005] The application aims at the problem of poor error calibration effect of the existing pipeline ADC, and provides a method for optimizing a time delay neural network by using an ant colony algorithm and performing background calibration on a pipeline ADC based on the time delay neural network.

[0006] The technical solution of the application is to provide a pipeline ADC background calibration method for optimizing a time delay neural network by using an ant colony algorithm, which comprises the following steps:

[0007] The to-be-calibrated pipeline ADC and a high-precision reference ADC are connected to the same signal source, a time delay unit is combined with a three-layer BP neural network to construct a time delay neural network, the output of the to-be-calibrated pipeline ADC is connected to the time delay unit of the time delay neural network, and the output of the time delay unit and the output of the to-be-calibrated pipeline ADC together constitute the input data of the neural network training.

[0008] The calibration of the to-be-calibrated pipeline ADC by using the time delay neural network is divided into three modes:

[0009] Preparation mode: the optimal global time delay dimension combination is obtained by optimizing and screening all time delay dimensions output by the time delay unit through the ant colony algorithm, the number of nodes of the input layer of the time delay neural network is determined according to the optimal global time delay dimension combination, so as to determine the structure of the time delay neural network, and the initial configuration of the weight and bias of the time delay neural network is globally optimized through the ant colony algorithm; after the above steps are completed, the training mode is entered;

[0010] Training mode: the input data of the neural network training is input into the time delay neural network for training, the weight and bias of the time delay neural network are adjusted, and the neural network is iteratively optimized; when the loss value calculated by the loss function of the time delay neural network is greater than the preset calibration target value, the output mode is entered;

[0011] Output mode: the weight and bias are locked and are not adjusted; when the loss value calculated by the loss function of the time delay neural network is less than the preset calibration target value, the training mode is entered;

[0012] Regardless of whether the time delay neural network is in the training mode or the output mode, the time delay neural network outputs the conversion result of the to-be-calibrated pipeline ADC processed by the time delay neural network, and the training of the time delay neural network does not affect the normal operation of the ADC.

[0013] In any of the above technical solutions, further, the process of obtaining the optimal global time delay dimension combination by optimizing and screening all time delay dimensions output by the time delay unit through the ant colony algorithm comprises the following steps:

[0014] S11, in the initialization stage, the number of ant colony foraging path nodes is the total number of time delay module dimensions, the initial pheromone is set for the interconnected node path, and the number of ants and the maximum number of iterations are set;

[0015] S12, construct and evaluate the solution space, ants start from a randomly selected starting point, based on the current pheromone concentration according to the greedy selection strategy and the random selection strategy to determine the path node, until the path reaches all nodes after the end, its transition probability formula is as follows:

[0016]

[0017] In the formula, η ij (n) is the heuristic factor from node i to node j at the nth iteration, τ ij (n) is the pheromone from node i to node j at the nth iteration, alpha represents the importance of pheromone, beta represents the importance of heuristic factor, tabu k is the tabu list;

[0018] All the paths formed above constitute the delay dimension combination as the input of neural network training, and the minimum mean square error is taken as the fitness function value of the ant path after multiple training, and its formula is:

[0019] F k = MMSE = min (mse1, mse2, … mse p );

[0020]

[0021] Where p represents the number of neural network training, mse p mse loss function of the pth iteration neural network training, m represents the number of samples in a training process;

[0022] Then select the ant with the minimum fitness value as the optimal solution of this iteration;

[0023] S13, after the end of one iteration, update the pheromone on the path of the optimal ant, and the update formula is:

[0024]

[0025] In the formula, p is the evaporation factor of pheromone, Δτ ij is the pheromone increment, which is calculated by the fitness value of the optimal ant;

[0026] S14, repeat steps S12 and S13 until the maximum number of iterations is reached, and output the pheromone of the optimal ant path as the optimal global delay dimension combination.

[0027] In any of the above technical solutions, further, the specific steps of globally optimizing the initial configuration of the delay neural network weight and bias by the ant colony algorithm include:

[0028] S21, in the initialization stage, the weight and bias are divided into numerical sets within the predetermined upper and lower limit range, these sets are regarded as the nodes of the ant foraging path, the number of nodes of the ant foraging path is set, the initial pheromone of the node path connected in front and back is set, and the number of ants and the maximum iteration number are set;

[0029] S22, constructing and evaluating the solution space, the weight and bias are divided into numerical sets within the predetermined upper and lower limit range, these sets are regarded as the nodes of the ant foraging path, the ant starts from the randomly selected starting point, and the subsequent path nodes are determined based on the current pheromone concentration according to the greedy selection strategy and the random selection strategy, until the path reaches all the nodes and ends, and the transition probability formula is as follows:

[0030]

[0031] In the formula, η ij (n) is the heuristic factor from node i to node j in the nth iteration, τ ij (n) is the pheromone from node i to node j in the nth iteration, α represents the importance of pheromone, β represents the importance of heuristic factor, tabu k is the tabu list;

[0032] The numerical set composed of the above path is used as the initial parameter configuration of neural network training, and the minimum mean square error is taken as the fitness function value of the ant path after multiple training, and the formula is:

[0033] F k = MMSE = min (mse1, mse2, … mse p );

[0034]

[0035] Where p represents the number of neural network training, mse p represents the mse loss function of the pth iteration neural network training, and m represents the number of samples in a training process; then, the ant with the minimum fitness value is selected as the optimal solution of this iteration;

[0036] S23, updating pheromone, after the end of one iteration, the pheromone on the path of the optimal ant is updated, and the update formula is:

[0037]

[0038] In the formula, ρ is the volatilization factor of pheromone, Δτ ij is the pheromone increment, Δτ ij is calculated by the fitness value of the optimal ant;

[0039] S24, after iteration, terminate, repeat steps S22 and S23 until the set maximum number of iterations is reached, and the path set composed of the paths of the ants is assigned to the three-layer BP neural network as the initial weights and biases.

[0040] In any of the technical solutions above, further, the three-layer BP neural network comprises an input layer, a hidden layer and an output layer, nodes of the front and rear layers are connected to each other, the number of input layer nodes is consistent with the number of global time delay dimensions filtered by the ant colony algorithm, the number of hidden layer neurons is adjustable, the number of output layer nodes is 1, and the output layer directly outputs the calibrated result.

[0041] The linear Satlins function is selected as the hidden layer activation function, and the formula is:

[0042]

[0043] In any of the technical solutions above, further, the loss value phase change rate index K ΔMSE As a loss function, the output of the to-be-calibrated pipeline ADC after passing through the time delay neural network and the output of the high-precision reference ADC are taken as calculation elements of the loss function of the time delay neural network, the loss value is calculated, the preset calibration target value is compared with the loss value, in the case that the loss value is greater than the calibration target value, the time delay neural network enters a training mode, the weights and biases of the time delay neural network are adjusted, and the neural network is iteratively optimized; in the case that the loss value is less than or equal to the calibration target value, the time delay neural network enters an output mode, and the weights and biases are not adjusted; after each training is completed, a trained MSE value is obtained once, and the MSE value is saved as a record, and the MSE value is an intermediate result of the calculation of the loss value phase change rate index K ΔMSE .

[0044] In any of the technical solutions above, further, the loss value phase change rate index K ΔMSE is calculated as follows:

[0045] The mean square error MSE is calculated as follows:

[0046]

[0047] Where Y g is the output result of the time delay neural network, is the output result of the high-precision reference ADC.

[0048] The loss value change rate AMSE t is calculated as follows:

[0049]

[0050] Where t represents the current time, t-1 represents the previous time, and the loss value change rate AMSE tloss value difference between the current time and the last time;

[0051] loss value stage change rate index K ΔMSE :

[0052]

[0053] where [Delta MSE t >=threshold] is an indicator function, threshold is a preset threshold, when Delta MSE t >=threshold, [Delta MSE t >=threshold] takes 1, otherwise takes 0, the loss value stage change rate index K ΔMSE represents the proportion of the number of loss value change rates exceeding the preset threshold in the past T time.

[0054] The beneficial effects of the present application are:

[0055] The technical solution in the present application adopts a time delay unit and a three-layer BP neural network to form a time delay neural network for error calibration of a pipeline ADC, which can improve the calibration effect of the neural network on the nonlinear error with memory effect.

[0056] The ant colony algorithm is adopted to perform double optimization on the global time delay dimension and the initial parameter configuration of the neural network, which improves the calibration performance while reducing the size of the calibration model, and effectively solves the problem of falling into local optimization in the calibration process. BRIEF DESCRIPTION OF DRAWINGS

[0057] The advantages of the above and additional aspects of the present application will become apparent and easy to understand in connection with the following description of embodiments, in conjunction with the accompanying drawings, in which:

[0058] Figure 1 is a schematic flow chart of a pipeline ADC background calibration method of an ant colony optimization time delay neural network according to an embodiment of the present application;

[0059] Figure 2 is a structural schematic diagram of a pipeline ADC background calibration method of an ant colony optimization time delay neural network according to an embodiment of the present application;

[0060] Figure 3 is a schematic diagram of the principle of an ant colony algorithm of a pipeline ADC background calibration method of an ant colony optimization time delay neural network according to an embodiment of the present application;

[0061] Figure 4 is a principle diagram of initial parameter configuration optimization based on an ant colony algorithm of a pipeline ADC background calibration method of an ant colony optimization time delay neural network according to an embodiment of the present application;

[0062] Figure 5 is a test equipment schematic diagram of an ant colony optimization time delay neural network pipeline ADC background calibration method according to an embodiment of the present application;

[0063] Figure 6 is a calibration effect comparison schematic diagram of different time delay dimension combinations before and after optimization in different neural network scales according to an embodiment of the present application. DETAILED DESCRIPTION

[0064] In order to enable a more clear understanding of the above-mentioned purposes, features and advantages of the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0065] In the following description, a lot of specific details are set forth in order to facilitate a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, and therefore, the protection scope of the present application is not limited by the specific embodiments disclosed below.

[0066] As shown in Figure 1 and Figure 2 , the present embodiment provides an ant colony optimization time delay neural network pipeline ADC background calibration method, which comprises:

[0067] The to-be-calibrated pipeline ADC and the high-precision reference ADC are connected to the same signal source, the time delay unit is combined with the three-layer BP neural network to construct the time delay neural network, the input layer of the time delay neural network is connected to the output of the to-be-calibrated pipeline ADC, the output of the to-be-calibrated pipeline ADC is connected to the time delay unit of the time delay neural network, the number of the time delay units is consistent with the output precision of the to-be-calibrated pipeline ADC, and the output bit number of the ADC is binary.

[0068] Specifically, the output of the time delay unit and the output of the to-be-calibrated pipeline ADC jointly constitute the input data of the neural network training, and the output result of the high-precision reference ADC is used as the label data of the neural network training.

[0069] The calibration of the to-be-calibrated pipeline ADC using the time delay neural network is divided into three modes:

[0070] Preparation mode: all time delay dimensions output by the time delay unit are optimized and screened by the ant colony algorithm to obtain an optimal global time delay dimension combination, the number of nodes of the input layer of the time delay neural network is determined according to the optimal global time delay dimension combination, so as to determine the structure of the time delay neural network; the initial configuration of the weight and bias of the time delay neural network is globally optimized by the ant colony algorithm; after the above steps are completed, the training mode is entered.

[0071] Training mode: input the input data of neural network training into the time delay neural network for training, adjust the weights and biases of the time delay neural network, and iteratively optimize the neural network; when the loss value calculated by the loss function of the time delay neural network is greater than the preset calibration target value, enter the output mode.

[0072] Output mode: lock the weights and biases and do not adjust them; when the loss value calculated by the loss function of the time delay neural network is less than the preset calibration target value, enter the training mode.

[0073] The pre-optimization by the ant colony algorithm in the preparation mode is discussed later in this paper. First, the basic framework of the neural network model is constructed. In the training mode and the output mode, the time delay neural network uses the loss value stage change rate index K ΔMSE As the loss function, the output of the time delay neural network after the calibration pipeline ADC is calculated as the calculation element of the loss function of the time delay neural network, and the preset calibration target value is compared with the loss value. When the loss value is greater than the calibration target value, the time delay neural network enters the training mode, adjusts the weights and biases of the time delay neural network, and iteratively optimizes the neural network. When the loss value is less than or equal to the calibration target value, the time delay neural network enters the output mode and does not adjust the weights and biases. After each training, a trained MSE value is obtained and recorded, and the MSE value is the intermediate result of calculating the loss value stage change rate index K ΔMSE .

[0074] The calculation method of the loss value change rate AMSE t is as follows:

[0075] Calculate the mean square error MSE:

[0076]

[0077] Where Y g is the output result of the time delay neural network, is the result of the high-precision reference ADC output.

[0078] Calculate the loss value change rate AMSE t :

[0079]

[0080] Where t represents the current time, t-1 is the previous time, and the loss value change rate AMSE t represents the difference between the loss values at the current time and the previous time.

[0081] Calculate the loss value stage change rate index K ΔMSE :

[0082]

[0083] where [DeltaMSE t is an indicator function, threshold is a pre-set threshold, when DeltaMSE t is greater than threshold, [DeltaMSE t is 1, otherwise 0, the loss value phase change rate index K ΔMSE represents the proportion of the number of times that the loss value change rate exceeds the pre-set threshold in the past T time.

[0084] The delay neural network outputs the conversion result of the time-delay neural network processing of the to-be-calibrated pipeline ADC in the training mode or the output mode, and the method is a background calibration method.

[0085] The three-layer BP neural network includes an input layer, a hidden layer and an output layer, and the nodes of the front and rear layers are connected to each other, wherein the number of input layer nodes before optimization is equal to the sum of the bit number of the ADC quantization output and the delay amount, the number of input layer nodes after optimization is consistent with the global delay dimension number screened by the ant colony algorithm, the number of input layer nodes after optimization is less, the number of hidden layer neurons is adjustable, the number of output layer nodes is 1, and the output layer directly outputs the calibrated result.

[0086] The linear Satlins function is selected as the hidden layer activation function, and the formula is:

[0087]

[0088] The delay neural network brings key information in the time dimension, which can improve the expression ability of the calibration model. However, the combination of the delay dimensions of adjacent times is not necessarily the optimal combination. In addition, due to the influence of multiple factors such as initialization parameter setting, training data quality and network structure, the neural network often falls into the dilemma of local optimum in the training process, resulting in unsatisfactory calibration effect or even calibration failure. The size of the delay dimension after global optimization is still very large, which means that the size of the input layer of the neural network is huge, which will lead to a large overall neural network size and unacceptable on-chip implementation resource consumption. In view of the above problems, the ant colony algorithm is adopted to optimize the delay dimension and the initialization parameter configuration of the neural network.

[0089] As Figure 3As shown, the ant colony algorithm is a heuristic algorithm, which is often used to solve the traveling salesman TSP, path planning and other problems, and has been widely used in the field of neural networks. The inspiration of the ant colony algorithm comes from the foraging behavior of ants in nature, ants will release pheromone in the process of foraging, and tend to choose the path with high pheromone content to move, and the pheromone content of the path which is not updated for a long time will gradually decrease and eventually disappear. For the time delay dimension optimization task of the present application, the process is regarded as the process of finding the optimal path in the foraging of ants. Each time delay is regarded as a node on the foraging route, and all these nodes together form a fully connected undirected graph.

[0090] Because the scale of the input layer of the neural network is huge, the network scale needs to be limited, so the ant colony algorithm is used to select the optimal dimension combination under the condition of limiting the feature dimension, and the specific steps are as follows:

[0091] S11, initialization stage, the number of ant colony foraging path nodes is the total dimension number of the time delay module, the initial pheromone of the interconnected node path is set, and the number of ants and the maximum iteration number are set.

[0092] S12, construct and evaluate the solution space, the ant starts from a randomly selected starting point, determines the subsequent path nodes based on the current pheromone concentration according to the greedy selection strategy and the random selection strategy, and ends until the path reaches all nodes, and the transition probability formula is as follows:

[0093]

[0094] In the formula, η ij (n) is the heuristic factor from node i to node j at the nth iteration, τ ij (n) is the pheromone from node i to node j at the nth iteration, alpha represents the importance of pheromone, beta represents the importance of heuristic factor, and tabu k is the tabu list.

[0095] The time delay dimension combination composed of the above path is used as the input of neural network training, and the minimum mean square error is taken as the fitness function value of the ant path after multiple training, and the formula is:

[0096] F k = MMSE = min (mse1, mse2, … mse p );

[0097]

[0098] Where p represents the number of neural network training, mse pMSE represents the mse loss function of the pth iteration of neural network training, and m represents the number of samples in a certain training process. Then, the ant with the minimum fitness value is selected as the optimal solution of this iteration.

[0099] S13, pheromone is updated. After one iteration, the pheromone on the path of the optimal ant is updated, and the update formula is:

[0100]

[0101] In the formula, p is the evaporation factor of pheromone, Δτ ij is the pheromone increment, which is calculated by the fitness value of the optimal ant.

[0102] S14, repeat steps S12 and S13 until the maximum number of iterations is reached, and output the pheromone of the optimal ant path as the optimal global latency dimension combination.

[0103] The initial configuration of the latency neural network weight and bias is globally optimized by using the ant colony algorithm, and the specific steps are as follows:

[0104] S21, initialization stage, divide the weight and bias into numerical sets within the predetermined upper and lower limit range, these sets are regarded as the nodes of the ant foraging path, set the number of ants and the maximum number of iterations.

[0105] S22, construct and evaluate the solution space, as shown in Figure 4 The weight and bias are divided into numerical sets within the predetermined upper and lower limit range, these sets are regarded as the nodes of the ant foraging path, the ant starts from a randomly selected starting point, and determines the subsequent path nodes based on the current pheromone concentration according to the greedy selection strategy and the random selection strategy, until the path reaches all nodes and ends, and the transition probability formula is as follows:

[0106]

[0107] In the formula, η ij (n) is the heuristic factor from node i to node j in the nth iteration, τ ij (n) is the pheromone from node i to node j in the nth iteration, α represents the importance of pheromone, β represents the importance of heuristic factor, and tabu k is the tabu list.

[0108] The numerical set composed of the above path is used as the initial parameter configuration of neural network training, and the minimum mean square error is taken as the fitness function value of the ant path after multiple training, and the formula is:

[0109] F k= MMSE = min(mse1, mse2, … mse p );

[0110]

[0111] wherein p represents the number of neural network training, mse p p represents the mse loss function of the pth iteration neural network training, m represents the number of samples in a certain training process; and then, the ant with the minimum fitness value is selected as the optimal solution of this iteration.

[0112] S23, pheromone is updated, after one iteration, the pheromone on the path of the optimal ant is updated, and the update formula is:

[0113]

[0114] In the formula, ρ is the evaporation factor of pheromone, Δτ ij is the pheromone increment, which is calculated by the fitness value of the optimal ant.

[0115] S24, iteration is terminated, steps S22 and S23 are repeated until the maximum number of iterations is reached, and the numerical set composed of the path of the ant is assigned to the three-layer BP neural network as the initial weight and bias.

[0116] As Figure 5 shown, another embodiment of the application uses a commercial 14-bit, 1Gsps pipeline ADC to verify the effectiveness of the proposed calibration technology, and the test equipment includes a signal generator, a clock board, an ADC test board, an FPGA development board and a computer for data processing. In the experiment, the sampling frequency of the ADC is set to 1GHz, the input frequency is a 10.3MHz single-frequency sinusoidal signal (-1dBFS), and the pipeline ADC output data is collected. In order to replace the high-precision reference ADC circuit, the curve fitting method is used to fit the collected data with high precision, and the fitting result is used as the label data for neural network training. The data fitting is realized by using the function fitting tool of MATLAB. After testing, the output performance indicators of the pipeline ADC chip to be calibrated are as follows: SNDR is 63.80dB, SFDR is 83.63dB, and ENOB is 10.30 bits. After fitting, the effective precision of the data can reach 25.95 bits, which meets the precision requirement of the reference circuit for the calibration task.

[0117] As Figure 6 shown, the calibration effect comparison of different delay dimension combinations in different neural network scales before and after the optimization of the ant colony algorithm is shown. Generally, with the increase of the network scale, the calibration effect of the delay neural network will be improved, and will tend to be stable (reach the saturation point of the calibration effect) after reaching a certain scale.

[0118] The following Table 1 shows the probability statistics of different scale neural networks achieving optimal calibration effect in 45 time delay dimensions:

[0119] Table 1 Comparison of probabilities of neural networks achieving optimization before and after ant colony optimization

[0120] Number of neurons 2 5 10 20 30 Optimized probability (before optimization) 20％ 20％ 50％ 20％ 30％ Optimized probability (after optimization) 100％ 100％ 100％ 100％ 100％

[0121] The above data shows that there is a serious phenomenon of falling into local optimum regardless of the scale of the neural network calibration model. To solve this problem, the initial configuration of the neural network is optimized by using the ant colony algorithm. The experimental results show that after the optimization of the ant colony algorithm, the problem of falling into local optimum is effectively solved, and almost all scales of neural networks can achieve 100% optimization probability.

[0122] The following Table 2 compares the performance of the calibration method provided by the present application with other calibration methods:

[0123] Table 2 Comparison with other calibration techniques

[0124]

[0125] As can be seen from Table 2 above, compared with other calibration methods, the calibration method provided by the present application achieves the greatest performance improvement. In the chip-level test of the calibration method in this paper, the ENOB reaches 13.75, which is very close to the nominal quantization accuracy of 14 bits of the calibration target, and the calibration task is completed excellently.

[0126] In summary, the present application proposes a pipeline ADC background calibration method for time delay neural network optimized by ant colony algorithm, comprising:

[0127] The to-be-calibrated pipeline ADC and the high-precision reference ADC are connected to the same signal source, the time delay unit and the three-layer BP neural network are combined to construct a time delay neural network, the output of the to-be-calibrated pipeline ADC is connected to the time delay unit of the time delay neural network, and the output of the time delay unit and the output of the to-be-calibrated pipeline ADC together constitute the input data of the neural network training.

[0128] The calibration of the to-be-calibrated pipeline ADC using the time delay neural network is divided into three modes:

[0129] Preparation mode: all time delay dimensions output by the time delay unit are optimized and screened by the ant colony algorithm to obtain an optimal global time delay dimension combination, the number of nodes of the input layer of the time delay neural network is determined according to the optimal global time delay dimension combination, and the structure of the time delay neural network is determined accordingly; the initial configuration of the weight and bias of the time delay neural network is globally optimized by the ant colony algorithm; after the above steps are completed, the training mode is entered.

[0130] Training mode: input the input data of neural network training into the time delay neural network for training, adjust the weight and bias of the time delay neural network, and iteratively optimize the neural network; when the loss value calculated by the loss function of the time delay neural network is greater than the preset calibration target value, enter the output mode.

[0131] Output mode: the weight and bias are locked and not adjusted; when the loss value calculated by the loss function of the time delay neural network is less than the preset calibration target value, enter the training mode.

[0132] The steps in the application can be adjusted, combined and deleted according to actual needs.

[0133] The units in the device can be combined, divided and deleted according to actual needs.

[0134] Although the application has been disclosed with reference to the drawings, it should be understood that these descriptions are merely exemplary and are not intended to limit the application. The scope of protection of the application is defined by the appended claims, and can include various modifications, improvements and equivalents of the application made without departing from the scope and spirit of the application.

Claims

1. A pipeline ADC background calibration method of an ant colony optimization latency neural network, characterized in that, The method comprises: The method comprises: The method comprises: The preparation mode: the optimal global time delay dimension combination is obtained by optimizing and screening all time delay dimensions output by the time delay unit through the ant colony algorithm, the node number of the input layer of the time delay neural network is determined according to the optimal global time delay dimension combination, and the structure of the time delay neural network is determined accordingly; the initial configuration of the weight and bias of the time delay neural network is globally optimized through the ant colony algorithm; after the above steps are completed, the training mode is entered; The training mode: the input data of the neural network training is input into the time delay neural network for training, the weight and bias of the time delay neural network are adjusted, and the neural network is iteratively optimized; when the loss value calculated by the loss function of the time delay neural network is greater than the preset calibration target value, the output mode is entered; The output mode: the weight and bias are locked and are not adjusted; when the loss value calculated by the loss function of the time delay neural network is less than the preset calibration target value, the training mode is entered; The time delay neural network outputs the conversion result of the pipeline ADC to be calibrated processed by the time delay neural network regardless of the training mode or the output mode, and the training of the time delay neural network does not affect the normal work of the ADC.

2. The pipeline ADC background calibration method of ant colony optimization time delay neural network according to claim 1, wherein, The process of obtaining the optimal global time delay dimension combination by optimizing and screening all time delay dimensions output by the time delay unit through the ant colony algorithm comprises: S11, initialization stage, the number of ant colony foraging path nodes is the total dimension number of the time delay module, the initial pheromone is set for the interconnected node path, and the number of ants and the maximum iteration number are set; S12, construct and evaluate the solution space, the ant starts from a randomly selected starting point, determines the path node after the starting point based on the current pheromone concentration according to the greedy selection strategy and the random selection strategy, and ends until the path reaches all nodes, and the transition probability formula is as follows; wherein η ij (n) is the heuristic factor from node i to node j at the n-th iteration, τ ij (n) is the pheromone from node i to node j at the n-th iteration, a indicates the importance of the pheromone, β indicates the importance of the heuristic factor, tabu k is the tabu list; All the paths formed above constitute the input of the neural network training, and the minimum mean square error is taken as the fitness function value of the ant path after multiple training, and the formula is as follows: F k = MMSE = min(mse1, mse2,... mse p ); where p represents the number of neural network training times, mse p represents the mse loss function of the pth iteration of neural network training, and m represents the number of samples in a certain training process. Then, the ant with the minimum fitness value is selected as the optimal solution of this iteration; S13, after one iteration, update the pheromone on the path of the optimal ant, and the update formula is as follows: where p is the pheromone evaporation factor, Δτ ij is the pheromone increment, which is calculated from the fitness value of the best ant. S14, repeat steps S12 and S13 until the maximum iteration number is reached, and output the pheromone of the optimal ant path as the optimal global time delay dimension combination.

3. The pipeline ADC background calibration method of ant colony optimization time delay neural network according to claim 1, wherein, The specific steps of globally optimizing the initial configuration of the weight and bias of the time delay neural network through the ant colony algorithm comprise: S21, in the initialization stage, the weight and bias are divided into numerical sets within the predetermined upper and lower limit range, these sets are regarded as the nodes of the ant foraging path, the number of nodes of the ant foraging path is set, the initial pheromone of the node path is set, and the number of ants and the maximum iteration number are set; S22, construct and evaluate the solution space, the weight and bias are divided into numerical sets within the predetermined upper and lower limit range, these sets are regarded as the nodes of the ant foraging path, the ant starts from the randomly selected starting point, determines the subsequent path node based on the current pheromone concentration according to the greedy selection strategy and the random selection strategy, and ends until the path reaches all nodes, and the transition probability formula is as follows: wherein η ij (n) is the heuristic factor from node i to node j at the n-th iteration, τ ij (n) is the pheromone from node i to node j at the n-th iteration, a indicates the importance of the pheromone, β indicates the importance of the heuristic factor, tabu k is the tabu list; The numerical set composed of the above path is used as the initial parameter configuration of the neural network training, the minimum mean square error is taken as the fitness function value of the ant path after multiple training, and the formula is: F k = MMSE = min(mse1, mse2,... mse p ); where p represents the number of neural network training, mse p The mse loss function of the pth iteration of neural network training is represented, and m represents the number of samples in a certain training process. Then, the ant with the minimum fitness value is selected as the optimal solution of this iteration. S23, update the pheromone, after one iteration, update the pheromone on the path of the optimal ant, and the update formula is: where p is the pheromone evaporation factor, Δτ ij is the pheromone increment, Δτ ij is calculated by the fitness value of the optimal ant. S24, terminate after iteration, repeat steps S22 and S23 until the maximum iteration number is reached, and the numerical set composed of the path of the ant is assigned to the three-layer BP neural network as the initial weight and bias.

4. The pipeline ADC background calibration method of ant colony optimization time delay neural network according to claim 1, wherein, The three-layer BP neural network includes an input layer, a hidden layer and an output layer, the nodes between the front and rear layers are connected, the number of input layer nodes is consistent with the global time delay dimension number screened by the ant colony algorithm, the number of hidden layer neurons is adjustable, the number of output layer nodes is 1, and the output layer directly outputs the calibrated result; The linear Satlins function is selected as the hidden layer activation function, and the formula is:

5. The pipeline ADC background calibration method of ant colony optimization time delay neural network according to claim 1, wherein, The time delay neural network adopts a loss value stage change rate index K ΔMSE As a loss function, the output of the to-be-calibrated pipeline ADC after passing through the time delay neural network and the output of the high-precision reference ADC are taken as calculation elements of the loss function of the time delay neural network, the loss value is calculated, the preset calibration target value is compared with the loss value, in the case that the loss value is greater than the calibration target value, the time delay neural network enters a training mode, the weight and bias of the time delay neural network are adjusted, and the neural network is iteratively optimized. When the loss value is less than or equal to the calibration target value, the time delay neural network enters the output mode, and the weight and bias are not adjusted; The trained MSE value is obtained once after each training is completed, and the MSE value is recorded as a change rate index K of the loss value in the calculation stage ΔMSE of intermediate results.

6. The pipeline ADC background calibration method of ant colony optimization time delay neural network according to claim 5, wherein, The loss value phase change rate index K ΔMSE The calculation method is as follows: Calculate the mean square error MSE: where Y g is the output result of the latency neural network, is the output result of the high-precision reference ADC. calculating a loss value change rate AMSE t : wherein t represents the current time, t-1 represents the previous time, the loss value change rate AMSE t represents the loss value difference between the current time and the previous time; computing the loss value phase change rate index K ΔMSE : where [ΔMSE t ≥ threshold] is an indicator function, threshold is a preset threshold, when ΔMSE t ≥ threshold, [ΔMSE t ≥ threshold] takes 1, otherwise takes 0, the loss value phase change rate index K ΔMSE represents the proportion of the number of times that the loss value change rate exceeds the preset threshold in the past T time points.

Images