Assembly line ADC background calibration method of ant colony optimization delay neural network

Abstract

The invention discloses an assembly line ADC background calibration method of an ant colony optimization time delay neural network, and relates to the technical field of ADC calibration, and the method comprises the steps: enabling a to-be-calibrated assembly line ADC and a high-precision reference ADC to be connected to the same signal source, constructing a time delay neural network through the combination of a time delay unit and a three-layer BP neural network, the output of the to-be-calibrated assembly line ADC is connected to a 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 assembly line ADC jointly form input data of neural network training; carrying out dimension optimization screening on the training data through an ant colony algorithm, and carrying out global optimization on the initial configuration of the time delay neural network weight and bias through the ant colony algorithm; and the time delay neural network outputs a calibrated result backwards.

Description

Technical Field

[0001] The present invention relates to the technical field of ADC calibration, and in particular to a pipeline ADC background calibration method of an ant colony optimized delay neural network. Background Art

[0002] Pipelined analog-to-digital converters (Pipelined ADCs) are widely used in communications, aerospace, and measuring instruments due to their high speed and high throughput. However, errors such as offset, mismatch, and thermal noise in the circuit will cause nonlinear output, and the errors in the front stage of the pipelined ADC will be transmitted to the back stage, limiting the performance of the pipeline structure. In order to pursue high-precision output while maintaining high speed, it is particularly necessary to use calibration technology. Compared with analog calibration circuits, which are more 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 method of ADC calibration.

[0003] Currently, the most widely used digital calibration methods include calibration based on dither, calibration based on PN injection, calibration based on split architecture, and calibration based on reference circuit, etc. The calibration methods based on dither and PN injection are simple to implement, but have poor effects on severe nonlinearity; the performance of the calibration method based on split architecture is limited to the best performance of the two channels; the calibration method based on reference circuit requires the construction of a reference circuit, and the performance of the filter also constrains the calibration effect.

[0004] In recent years, with the widespread 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 clipping technology, shared neuron technology, binary neural network, etc. However, the above optimization methods are more dependent on empirical adjustments, and it is very necessary to use optimization algorithms to optimize the neural network calibration model. Some researchers have used particle swarm optimization combined with SGD to optimize the neural network and obtained better calibration results. However, the particle swarm algorithm is easily affected by the historical optimal solution and is sensitive to the initial value. Improper initial values ​​may lead to slow convergence or failure to converge, and current research rarely involves the feature optimization problems that may arise in the application of neural networks in ADC error calibration. Summary of the invention

[0005] The purpose of the present invention is to provide 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 in order to solve the problem that the error calibration effect of the existing pipeline ADC is poor.

[0006] The technical solution of the present invention is to provide a pipeline ADC background calibration method of an ant colony optimized delay neural network, the method comprising:

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

[0008] There are three modes for calibrating the pipeline ADC to be calibrated using a time-delay neural network:

[0009] Preparation mode: All delay dimensions output by the delay unit are optimized and screened through the ant colony algorithm to obtain the optimal global delay dimension combination. The number of nodes in the input layer of the delay neural network is determined according to the optimal global delay dimension combination, thereby determining the structure of the delay neural network; the initial configuration of the weights and bias of the 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: 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;

[0011] Output mode: lock the weights and biases and do not adjust them; when the loss value calculated by the loss function of the delay neural network is less than the preset calibration target value, enter the training mode;

[0012] No matter the time delay neural network is in training mode or output mode, it outputs backward the transformation result of the pipeline ADC to be calibrated after being processed by the time delay neural network. The time delay neural network training does not affect the normal operation of the ADC.

[0013] In any of the above technical solutions, further, the process of optimizing and screening all the delay dimensions output by the delay unit by using an ant colony algorithm to obtain the optimal global delay dimension combination includes:

[0014] S11, initialization stage, the number of nodes in the ant colony foraging path is the total number of dimensions of the delay module, 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. The ants start from the randomly selected starting point and determine 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. The transition probability formula is as follows;

[0016]

[0017] Where 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, α represents the importance of pheromone, β represents the importance of heuristic factor, tabu k For the taboo table;

[0018] The delay dimension combination of all the paths formed above is used as the input of the neural network training. After multiple trainings, the minimum mean square error is taken as the fitness function value of the ant path. The formula is:

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

[0020]

[0021] Where p represents the number of neural network training times, mse p represents the mse loss function of the p-th iteration of neural network training, and m represents the number of samples in a certain training process;

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

[0023] S13. After one iteration, the pheromone on the path of the optimal ant is updated. The update formula is:

[0024]

[0025] Where ρ is the volatility 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 set 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 weights and biases of the time-delay neural network by using the ant colony algorithm include:

[0028] S21, initialization stage, weights and biases are divided into numerical sets within the predetermined upper and lower limits, these sets are regarded as nodes of the ant foraging route, the number of nodes in the ant colony foraging path, initial pheromones are set for the node paths connected before and after, and the number of ants and the maximum number of iterations are set;

[0029] S22. Construct and evaluate the solution space. The weights and biases are evenly divided into numerical sets within the predetermined upper and lower limits. These sets are regarded as nodes of the ant foraging route. The ants start from a randomly selected starting point and determine 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. The transition probability formula is as follows:

[0030]

[0031] Where 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, α represents the importance of pheromone, β represents the importance of heuristic factor, tabu k For the taboo table;

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

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

[0034]

[0035] 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 smallest fitness value is selected as the optimal solution for this iteration;

[0036] S23, update pheromone. After one iteration, update the pheromone on the path of the optimal ant. The update formula is:

[0037]

[0038] Where ρ is the volatility factor of pheromone, Δτ ij is the pheromone increment, Δτ ij Calculated by the fitness value of the best ant;

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

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

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

[0042]

[0043] In any of the above technical solutions, further, the time-delay neural network adopts the loss value stage change rate index K ΔMSE As the loss function, the output of the ADC in the pipeline to be calibrated after passing through the time-delay neural network and the output of the high-precision reference ADC are used as the calculation elements of the loss function of the time-delay neural network, and the loss value is calculated. 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 is completed, the MSE value after training is obtained and saved, and the MSE value is the index K of the rate of change of the calculated loss value stage ΔMSE The intermediate result.

[0044] In any of the above technical solutions, further, the loss value stage change rate index K ΔMSE The calculation method is as follows:

[0045] Calculate the mean square error MSE:

[0046]

[0047] where Y i is the output result of the time-delay neural network, It is the result of high-precision reference ADC output;

[0048] Calculate the loss value change rate ΔMSE t :

[0049]

[0050] Among them, t represents the current moment, t-1 represents the previous moment, and the loss value change rate ΔMSE tIndicates the difference between the loss value at the current moment and the previous moment;

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

[0052]

[0053] where [ΔMSE t ≥threshold] is an indicative function, threshold is the preset threshold, when ΔMSE t ≥threshold [ΔMSE t ≥threshold] takes 1, otherwise takes 0, and the loss value stage change rate index K ΔMSE Indicates the percentage of loss value change rates exceeding the preset threshold in the past T moments.

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

[0055] The technical solution of the present invention uses a time delay neural network formed by combining a time delay unit with a three-layer BP neural network for error calibration of pipeline ADC, which can improve the calibration effect of the neural network on nonlinear errors with memory effects;

[0056] The ant colony algorithm is used to perform dual optimization of the global delay dimension and the initial parameter configuration of the neural network, which improves the calibration performance while reducing the scale of the calibration model, and effectively solves the problem of the calibration process falling into local optimality. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The advantages of the above and additional aspects of the present invention will become apparent and easily understood in the description of the embodiments in conjunction with the following drawings, in which:

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

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

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

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

[0062] Figure 5 Schematic diagram of a test device for a pipeline ADC background calibration method of an ant colony optimized delay neural network according to an embodiment of the present invention;

[0063] Figure 6 It is a schematic diagram comparing the calibration effects of different delay dimension combinations under different neural network scales before and after optimization of the pipeline ADC background calibration method of the ant colony optimized delay neural network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0065] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0066] like Figure 1 and Figure 2 As shown, this embodiment provides a pipeline ADC background calibration method for an ant colony optimized delay neural network, the method comprising:

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

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

[0069] There are three modes for calibrating the pipeline ADC to be calibrated using a time-delay neural network:

[0070] Preparation mode: All delay dimensions output by the delay unit are optimized and screened through the ant colony algorithm to obtain the optimal global delay dimension combination. The number of nodes in the input layer of the delay neural network is determined based on the optimal global delay dimension combination, thereby determining the structure of the delay neural network. The initial configuration of the weights and bias of the delay neural network is globally optimized through the ant colony algorithm. After completing the above steps, enter the training mode.

[0071] Training mode: Input the input data of neural network training into the time-delay neural network for training, adjust the weights 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.

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

[0073] The pre-optimization of the preparation mode through the ant colony algorithm is discussed in the subsequent discussion of this article. First, the basic framework of the neural network model is constructed. In the training mode and output mode, the time-delay neural network adopts the loss value stage change rate index K ΔMSE As the loss function, the output of the ADC in the pipeline to be calibrated after passing through the time-delay neural network and the output of the high-precision reference ADC are used as the calculation elements of the loss function of the time-delay neural network to calculate the loss value, 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 is completed, the MSE value after training is obtained and saved and recorded, and the MSE value is the index K of the rate of change of the calculated loss value stage ΔMSE The intermediate result.

[0074] Loss value change rate ΔMSE t The calculation method is as follows:

[0075] Calculate the mean square error MSE:

[0076]

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

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

[0079]

[0080] Among them, t represents the current moment, t-1 represents the previous moment, and the loss value change rate ΔMSE t Indicates the difference in loss value between the current moment and the previous moment.

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

[0082]

[0083] where [ΔMSE t ≥threshold] is an indicative function, threshold is the preset threshold, when ΔMSE t ≥threshold [ΔMSE t ≥threshold] takes 1, otherwise takes 0, and the loss value stage change rate index K ΔMSE Indicates the percentage of loss value change rates exceeding the preset threshold in the past T moments.

[0084] No matter the time delay neural network is in training mode or output mode, it outputs backward the conversion result of the pipeline ADC to be calibrated after being processed by the time delay neural network. This method is a background calibration method.

[0085] The three-layer BP neural network includes an input layer, a hidden layer and an output layer. The nodes of the front and back layers are interconnected. The number of input layer nodes before optimization (through the ant colony algorithm introduced later) is equal to the sum of the bits of the ADC quantization output and its delay. The number of input layer nodes after optimization is consistent with the number of global delay dimensions after screening by the ant colony algorithm. The number of input layer nodes after optimization is smaller, the number of hidden layer neurons is adjustable, the number of output layer nodes is 1, and the output layer directly outputs the calibrated results.

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

[0087]

[0088] The time-delay neural network brings key information in the time dimension, which can improve the expressiveness of the calibration model. However, the combination of delay dimensions of adjacent times is not necessarily the optimal combination. In addition, due to the influence of multiple factors such as initialization parameter settings, training data quality, and network structure, neural networks often fall into the dilemma of local optimality during training, resulting in unsatisfactory calibration effects or even calibration failure. The scale of the delay dimension after global optimization is still large, which means that the scale of the input layer of the neural network is huge, which will lead to the overall scale of the neural network being huge, and the on-chip resource consumption is huge and unacceptable. In response to the above problems, this paper uses the ant colony algorithm to perform dual optimization of the delay dimension and the initialization parameter configuration of the neural network.

[0089] like Figure 3As shown, the ant colony algorithm is a heuristic algorithm that is often used to solve traveling salesman TSP, path planning and other problems. It has also 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 release chemical pheromones during the foraging process and tend to choose paths with higher pheromone content to move. The pheromone content of paths that have not been updated for a long time will gradually decrease and eventually dissipate. For the delay dimension optimization task of the present invention, the present invention regards this process as the process of finding the optimal path when ants forage. Each delay is regarded as a node on the foraging route, and all these nodes together form a fully connected undirected graph.

[0090] Since the input layer of the neural network is huge, the network scale needs to be limited. Therefore, the ant colony algorithm is used to select the optimal dimension combination under the condition of limited feature dimensions. The specific steps are as follows:

[0091] S11, initialization phase, the number of nodes in the ant colony foraging path is the total number of dimensions of the delay module, the initial pheromone is set for the interconnected node path, and the number of ants and the maximum number of iterations are set.

[0092] S12, construct and evaluate the solution space. The ants start from a randomly selected starting point and determine 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. The transition probability formula is as follows:

[0093]

[0094] Where 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, α represents the importance of pheromone, β represents the importance of heuristic factor, tabu k It is a taboo table.

[0095] The delay dimension combination of the above paths is used as the input of the neural network training. After multiple trainings, the minimum mean square error is taken as the fitness function value of the ant path. The formula is:

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

[0097]

[0098] Where p represents the number of neural network training times, mse prepresents the MSE loss function of the p-th iteration of neural network training, and m represents the number of samples in a certain training process; then, the ant with the smallest fitness value is selected as the optimal solution for this iteration.

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

[0100]

[0101] Where ρ is the volatility 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 set maximum number of iterations is reached, and output the pheromone of the optimal ant path as the optimal global delay dimension combination.

[0103] The ant colony algorithm is used to globally optimize the initial configuration of the weights and bias of the time-delay neural network. The specific steps are as follows:

[0104] S21, initialization stage, weights and biases are divided into numerical sets within the predetermined upper and lower limits. These sets are regarded as nodes of the ant foraging route. The number of nodes in the ant colony foraging path is set, and the initial pheromone is set for the node path connected to the front and back, and the number of ants and the maximum number of iterations are set.

[0105] S22, construct and evaluate the solution space, such as Figure 4 As shown in the figure, the weights and biases are evenly divided into numerical sets within the predetermined upper and lower limits. These sets are regarded as nodes of the ant foraging route. The ants start from a randomly selected starting point and determine 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. The transition probability formula is as follows:

[0106]

[0107] Where 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, α represents the importance of pheromone, β represents the importance of heuristic factor, tabu k It is a taboo table.

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

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

[0110]

[0111] Where p represents the number of neural network training times, mse p represents the MSE loss function of the p-th iteration of neural network training, and m represents the number of samples in a certain training process; then, the ant with the smallest fitness value is selected as the optimal solution for this iteration.

[0112] S23, update pheromone. After one iteration, update the pheromone on the path of the optimal ant. The update formula is:

[0113]

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

[0115] S24, terminate after iteration, repeat steps S22 and S23 until the set maximum number of iterations is reached, and assign the numerical set composed of the paths passed by the ants to the three-layer BP neural network as initial weights and biases.

[0116] like Figure 5 As shown, another embodiment of the present invention uses a commercial 14-bit, 1Gsps pipeline ADC to verify the effectiveness of the proposed calibration technology. 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 ADC sampling frequency 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, this embodiment uses a curve fitting method to perform high-precision fitting on the collected data, and uses the fitting result as the label data for neural network training. The data fitting is implemented using MATLAB's function fitting tool. 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 accuracy of the data can reach 25.95 bits, which meets the accuracy requirements of the reference circuit for the calibration task of the present invention.

[0117] like Figure 6 As shown in the figure, the calibration effect comparison of different delay dimension combinations under different neural network scales before and after the ant colony algorithm optimization is shown. In general, as the network scale increases, the calibration effect of the delay neural network will improve, and will tend to stabilize after reaching a certain scale (reaching the calibration effect saturation point).

[0118] Table 1 below shows the probability statistics of achieving optimal calibration effects for neural networks of different sizes under 45 delay dimensions:

[0119] Table 1 Comparison of the probability of the neural network reaching the optimal state before and after ant colony optimization

[0120]

[0121]

[0122] The above data show that no matter how large the neural network calibration model is, there is a serious phenomenon of falling into local optimum. To solve this problem, this paper uses the ant colony algorithm to optimize the initial configuration of the neural network. The experimental results show that after the ant colony algorithm is optimized, the problem of falling into local optimum is effectively solved, and almost all neural networks of all sizes can achieve 100% optimization probability.

[0123] Table 2 below compares the performance of the calibration method provided by the present invention with other calibration methods:

[0124] Table 2 Comparison with other calibration techniques

[0125]

[0126] According to Table 2 above, compared with other calibration methods, the calibration method provided by the present invention 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 14-bit nominal quantization accuracy of the calibration target, and the calibration task is completed excellently.

[0127] In summary, the present invention proposes a pipeline ADC background calibration method of an ant colony optimized delay neural network, comprising:

[0128] The pipeline ADC to be calibrated and the high-precision reference ADC are connected to the same signal source, the delay unit is combined with the three-layer BP neural network to construct a delay neural network, the output of the pipeline ADC to be calibrated is connected to the delay unit of the delay neural network, and the output of the delay unit and the output of the pipeline ADC to be calibrated together constitute the input data for neural network training;

[0129] There are three modes for calibrating the pipeline ADC to be calibrated using a time-delay neural network:

[0130] Preparation mode: All delay dimensions output by the delay unit are optimized and screened through the ant colony algorithm to obtain the optimal global delay dimension combination. The number of nodes in the input layer of the delay neural network is determined based on the optimal global delay dimension combination, thereby determining the structure of the delay neural network. The initial configuration of the weights and bias of the delay neural network is globally optimized through the ant colony algorithm. After completing the above steps, enter the training mode.

[0131] Training mode: Input the input data of neural network training into the time-delay neural network for training, adjust the weights 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.

[0132] Output mode: lock the weights and biases and do not adjust them; when the loss value calculated by the loss function of the delay neural network is less than the preset calibration target value, enter the training mode.

[0133] The steps in the present invention can be adjusted in order, combined or deleted according to actual needs.

[0134] The units in the device of the present invention can be combined, divided and deleted according to actual needs.

[0135] Although the present invention is disclosed in detail with reference to the accompanying drawings, it should be understood that these descriptions are merely exemplary and are not intended to limit the application of the present invention. The scope of protection of the present invention is defined by the appended claims and may include various modifications, alterations and equivalents made to the invention without departing from the scope and spirit of the present invention.

Claims

1. A pipeline ADC background calibration method for an ant colony optimized delay neural network, characterized in that: The method comprises: The pipeline ADC to be calibrated and the high-precision reference ADC are connected to the same signal source, the delay unit is combined with the three-layer BP neural network to construct a delay neural network, the output of the pipeline ADC to be calibrated is connected to the delay unit of the delay neural network, and the output of the delay unit and the output of the pipeline ADC to be calibrated together constitute the input data of the neural network training; There are three modes for calibrating the pipeline ADC to be calibrated using a time-delay neural network: Preparation mode: All delay dimensions output by the delay unit are optimized and screened through the ant colony algorithm to obtain the optimal global delay dimension combination. The number of nodes in the input layer of the delay neural network is determined according to the optimal global delay dimension combination, thereby determining the structure of the delay neural network; the initial configuration of the weights and bias of the delay neural network is globally optimized through the ant colony algorithm; after the above steps are completed, the training mode is entered; 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; Output mode: lock the weights and biases and do not adjust them; when the loss value calculated by the loss function of the delay neural network is less than the preset calibration target value, enter the training mode; The time-delay neural network outputs backward the conversion result of the pipeline ADC to be calibrated after being processed by the time-delay neural network, whether in training mode or output mode. The time-delay neural network training does not affect the normal operation of the ADC.

2. The pipeline ADC background calibration method of the ant colony optimized delay neural network according to claim 1, characterized in that: The process of optimizing and screening all the delay dimensions output by the delay unit through the ant colony algorithm to obtain the optimal global delay dimension combination includes: S11, initialization stage, the number of nodes in the ant colony foraging path is the total number of dimensions of the delay module, the initial pheromone is set for the interconnected node path, and the number of ants and the maximum number of iterations are set; S12, construct and evaluate the solution space. The ants start from the randomly selected starting point and determine 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. The transition probability formula is as follows; Where 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, α represents the importance of pheromone, β represents the importance of heuristic factor, tabu k For the taboo table; The delay dimension combination of all the paths formed above is used as the input of the neural network training. After multiple trainings, the minimum mean square error is taken as the fitness function value of the ant path. The formula is: 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 p-th iteration of neural network training, and m represents the number of samples in a certain training process; Then select the ant with the smallest fitness value as the optimal solution for this iteration; S13. After one iteration, the pheromone on the path of the optimal ant is updated. The update formula is: Where ρ is the volatility factor of pheromone, Δτ ij is the pheromone increment, which is calculated by the fitness value of the optimal ant; S14, repeat steps S12 and S13 until the set maximum number of iterations is reached, and output the pheromone of the optimal ant path as the optimal global delay dimension combination.

3. The pipeline ADC background calibration method of the ant colony optimized delay neural network according to claim 1, characterized in that: The specific steps of globally optimizing the initial configuration of the weights and biases of the time-delay neural network through the ant colony algorithm include: S21, initialization stage, weights and biases are divided into numerical sets within the predetermined upper and lower limits, these sets are regarded as nodes of the ant foraging route, the number of nodes in the ant colony foraging path, initial pheromones are set for the node paths connected before and after, and the number of ants and the maximum number of iterations are set; S22. Construct and evaluate the solution space. The weights and biases are evenly divided into numerical sets within the predetermined upper and lower limits. These sets are regarded as nodes of the ant foraging route. The ants start from a randomly selected starting point and determine 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. The transition probability formula is as follows: Where 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, α represents the importance of pheromone, β represents the importance of heuristic factor, tabu k For the taboo table; The numerical set composed of the above paths is used as the initial parameter configuration for neural network training. After multiple trainings, the minimum mean square error is taken as the fitness function value of the ant path. The formula is: 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 smallest fitness value is selected as the optimal solution for this iteration; S23, update pheromone. After one iteration, update the pheromone on the path of the optimal ant. The update formula is: Where ρ is the volatility factor of pheromone, Δτ ij is the pheromone increment, Δτ ij Calculated by the fitness value of the best ant; S24, terminate after iteration, repeat steps S22 and S23 until the set maximum number of iterations is reached, and assign the numerical set composed of the paths passed by the ants to the three-layer BP neural network as initial weights and biases.

4. The pipeline ADC background calibration method of the ant colony optimized delay neural network according to claim 1, characterized in that: The three-layer BP neural network includes an input layer, a hidden layer and an output layer. The nodes of the front and back layers are connected to each other. The number of nodes in the input layer is consistent with the number of global delay dimensions after screening by the ant colony algorithm. The number of neurons in the hidden layer is adjustable. The number of nodes in the output layer is 1, and the output layer directly outputs the calibrated result. The hidden layer activation function uses the linear Satlins function, and its formula is:

5. The pipeline ADC background calibration method of the ant colony optimized delay neural network according to claim 1, characterized in that: The time-delay neural network adopts the loss value stage change rate index K ΔMSE As a loss function, the output of the pipeline ADC to be calibrated after passing through the time-delay neural network and the output of the high-precision reference ADC are used as calculation elements of the loss function of the time-delay neural network, the loss value is calculated, 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 without adjusting the weights and biases; After each training is completed, the MSE value after training is obtained and saved. The MSE value is the index K of the change rate of the loss value stage. ΔMSE The intermediate result.

6. The pipeline ADC background calibration method of the ant colony optimized delay neural network as claimed in claim 5, characterized in that: The loss value stage change rate index K ΔMSE The calculation method is as follows: Calculate the mean square error MSE: where Y i is the output result of the time-delay neural network, It is the result of high-precision reference ADC output; Calculate the loss value change rate ΔMSE t : Among them, t represents the current moment, t-1 represents the previous moment, and the loss value change rate ΔMSE t Indicates the difference between the loss value at the current moment and the previous moment; Calculate the loss value stage change rate index K ΔMSE : where [ΔMSE t ≥threshold] is an indicative function, threshold is the preset threshold, when ΔMSE t ≥threshold [ΔMSE t ≥threshold] takes 1, otherwise takes 0, and the loss value stage change rate index K ΔMSE Indicates the percentage of loss value change rates exceeding the preset threshold in the past T moments.

Images