Ant colony optimization method and device for parameter of time delay neural network based on calibration of pipeline analog-digital converter

Abstract

The application provides a time delay neural network parameter ant colony optimization method and device based on pipeline analog-to-digital converter calibration, relates to the technical field of analog-to-digital converter calibration, and comprises the following steps: first, configuring ant colony algorithm basic parameters and time delay neural network initial parameters, then constructing a joint solution space containing initial weight configuration and initial bias configuration based on the initial parameters, controlling ants to traverse nodes according to a greedy strategy and a random strategy to form paths (each path corresponds to a complete solution), then calculating fitness function values based on the minimum mean square error of each frequency multiple times of training under a multi-frequency verification set, and finally iteratively optimizing through a pheromone updating mechanism until a preset iteration number is reached and an optimal solution is output, so as to solve the problems of unstable performance in the frequency domain and limited application bandwidth.

Description

Technical Field

[0001] This invention relates to the technical field of analog-to-digital converter calibration, and in particular to a method and apparatus for ant colony optimization of time-delay neural network parameters based on pipelined analog-to-digital converter calibration. Background Technology

[0002] As a core component of high-speed, high-precision data acquisition systems, the performance of pipelined analog-to-digital converters (ADCs) directly affects the signal processing capabilities of the entire system. With advancements in semiconductor technology and increasingly demanding requirements for ADC performance, back-end digital calibration technology has become a crucial means to improve the performance of pipelined ADCs.

[0003] Due to their advantages in processing time-series signals, time-delay neural networks are widely used in the background calibration of pipelined ADCs. In this technical solution, the pipelined ADC to be calibrated and a high-precision reference ADC are connected to the same signal source. A structure is constructed containing a time-delay layer and a three-layer backpropagation (BP) neural network. The time-delay layer has an adjustable time-delay order, and its output, together with the current ADC output, constitutes the neural network input features. The output of the reference ADC serves as the label data for supervised learning. The number of neurons in the hidden layers of the neural network is adjustable, and the output layer is a single node that directly outputs the calibrated conversion result.

[0004] However, this technical solution reveals three main problems in practical applications: First, the neural network training process is prone to getting trapped in local optima, leading to limited calibration accuracy; second, redundancy exists in time delay feature extraction, increasing computational complexity; and finally, and most critically, the calibration method suffers from insufficient stability in the frequency domain and limited application bandwidth. Although existing patented technologies have solved the problems of local optima and feature redundancy using ant colony optimization algorithms, current solutions have not yet provided effective solutions to the technical challenges of unstable frequency domain performance and limited bandwidth. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for ant colony optimization of time-delay neural network parameters based on pipelined analog-to-digital converter calibration. The ant colony optimization algorithm is introduced to globally optimize the initial weights and biases of the time-delay neural network, avoiding the problem of difficult convergence or large performance fluctuations in the high-frequency range caused by improper initial parameter settings in the traditional training process, thereby significantly improving the stability of calibration applications over a wide frequency range.

[0006] In a first aspect, the present invention provides an ant colony optimization method for time-delay neural network parameters based on pipelined analog-to-digital converter calibration, comprising: Configure the basic parameters of the ant colony algorithm and the initial parameters of the time-delay neural network; Based on the initial parameters of the time-delay neural network, a joint solution space including initial weight configuration and initial bias configuration is constructed, and ants are controlled to start from random nodes and traverse independent nodes in the joint solution space according to greedy selection strategy and random selection strategy to form ant paths; wherein, each ant path corresponds to a complete solution in the joint solution space. Based on the complete solution, configure the initial parameters of the time-delay neural network, train the time-delay neural network multiple times on a validation set consisting of multiple frequencies, obtain the minimum mean square error of multiple training sessions at each frequency, and calculate the fitness function value of each ant path. Update the pheromone concentration on each ant path according to the fitness function value, and repeat the steps of calculating the complete solution and fitness function value until the number of iterations reaches the preset iteration threshold, and then output the optimal solution of the joint solution space at this time.

[0007] In an optional implementation, the step of constructing a joint solution space including initial weight configuration and initial bias configuration based on the initial parameters of the time-delay neural network, and controlling ants to start from random nodes and traverse the independent nodes in the joint solution space according to a greedy selection strategy and a random selection strategy to form an ant path, includes: Based on a preset number of discrete values ​​evenly distributed within the range of weight values, the independent nodes of the ant colony's foraging path are determined; the ants start from a randomly selected starting point and traverse all the independent nodes according to a greedy selection strategy and a random selection strategy based on the current pheromone concentration, forming an ant path that matches the number of network connections, and determining the initial weight configuration. Based on the fact that the bias value range is divided into a preset number of discrete values, the independent nodes of the ant colony's foraging path are determined; the ants start from a randomly selected starting point and traverse all the independent nodes according to the current pheromone concentration using a greedy selection strategy and a random selection strategy to form an ant path that matches the total number of biases, thus determining the initial bias configuration. Each complete solution in the joint solution space is determined based on the initial weight configuration and the initial bias configuration.

[0008] In an optional implementation, the step of calculating the fitness function value for each ant path includes: Based on the minimum mean square error of multiple training sessions at each frequency, the average of the minimum mean square error across all frequencies is calculated, and this average is used as the fitness function value for the current ant path.

[0009] In an optional implementation, the step of calculating the fitness function value for each ant path includes: Based on the minimum mean square error of multiple training sessions at each frequency and the preset weight coefficients corresponding to each frequency, the minimum mean square error of all frequencies is calculated and weighted averaged. The weighted average result is used as the fitness function value of the current ant path.

[0010] In an optional implementation, the step of updating the pheromone concentration on each ant path according to the fitness function value, repeating the aforementioned steps of calculating the complete solution and fitness function value until the number of iterations reaches a preset iteration threshold, and then outputting the optimal solution of the joint solution space at this point includes: Based on the fitness function value, the pheromone concentration of each ant path in the joint solution space is updated through the pheromone positive feedback mechanism of the ant colony algorithm. The process involves re-executing the initial parameters based on the time-delay neural network to construct a joint solution space including initial weight configuration and initial bias configuration. Ants are then controlled to start from random nodes and traverse independent nodes in the joint solution space using both greedy and random selection strategies to form ant paths. Based on the complete solution, the initial parameters of the time-delay neural network are configured. The network is then trained multiple times on a validation set consisting of multiple frequencies. The minimum mean square error of the training at each frequency is obtained, and the fitness function value of each ant path is calculated. This process continues until the number of iterations reaches a preset iteration threshold, at which point the complete solution with the smallest fitness function value is taken as the optimal solution. The optimal solution includes the optimal initial weight configuration and the optimal initial bias configuration.

[0011] In an optional implementation, the method further includes: The optimal solution is assigned to the time-delay neural network, and the output data of the pipeline analog-to-digital converter to be calibrated and the high-precision reference signal are used as inputs and labels for training to complete the construction of the calibration model. The trained time-delay neural network is used to calibrate the output of the analog-to-digital converter, resulting in a high-precision digital signal.

[0012] In an optional implementation, the basic parameters of the ant colony algorithm include the number of ants, the pheromone concentration value, and the iteration number threshold; the initial parameters of the time-delay neural network include the total dimension of the time-delay unit, the weight value range, the bias value range, and the range of neurons in the hidden layer of the three-layer BP neural network, as well as a validation set composed of multiple frequencies for verification.

[0013] Secondly, the present invention provides an ant colony optimization device for delay neural network parameters based on pipelined analog-to-digital converter calibration, comprising: The configuration module configures the basic parameters of the ant colony algorithm and the initial parameters of the time-delay neural network. The solution module constructs a joint solution space containing initial weight configuration and initial bias configuration based on the initial parameters of the time-delay neural network, and controls ants to start from random nodes and traverse independent nodes in the joint solution space according to greedy selection strategy and random selection strategy to form ant paths; wherein, each ant path corresponds to a complete solution in the joint solution space. The training module configures the initial parameters of the time-delay neural network based on the complete solution, trains the time-delay neural network multiple times on a validation set consisting of multiple frequencies, obtains the minimum mean square error of multiple training sessions at each frequency, and calculates the fitness function value of each ant path. The iteration module updates the pheromone concentration on each ant path according to the fitness function value, and repeats the steps of calculating the complete solution and fitness function value until the number of iterations reaches the preset iteration threshold, and then outputs the optimal solution of the joint solution space at this time.

[0014] Thirdly, the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the method as described in any of the foregoing embodiments.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed, implements the method described in any of the foregoing embodiments.

[0016] This invention provides a method and apparatus for ant colony optimization of time-delay neural network parameters based on pipelined analog-to-digital converter calibration. The core logic is "parameter configuration - solution space construction and path generation - fitness calculation - iterative optimization": First, configure the basic parameters of the ant colony algorithm (such as the number of ants, pheromone concentration, etc.) and the initial parameters of the time-delay neural network (such as the total dimension of the time-delay units, the range of weight / bias values, multi-frequency validation sets, etc.). Then, construct a joint solution space containing initial weight and bias configurations based on the initial parameters, and control the ants to traverse nodes and form paths using greedy and random strategies (each path corresponds to a complete solution). Subsequently, calculate the fitness function value based on the minimum mean square error of multiple training iterations at each frequency under the multi-frequency validation set. Finally, iteratively optimize through a pheromone update mechanism until a preset number of iterations is reached and the optimal solution is output. In terms of effectiveness, this scheme solves the problems of traditional parameter configuration being prone to local optima and low efficiency by leveraging the global search characteristics of the ant colony algorithm, thereby improving the reliability and efficiency of parameter optimization. Combined with fitness calculation of multi-frequency validation sets, it ensures the calibration accuracy of ADCs across a wide frequency band and avoids the problem of errors exceeding the standard in some frequency bands caused by single-frequency optimization. Finally, the optimal solution output can support the construction of a high-precision calibration model, effectively reducing ADC quantization errors and nonlinear errors, and providing technical support for pipelined ADCs to output high-precision digital signals.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 A flowchart of an ant colony optimization method for delay neural network parameters based on pipelined analog-to-digital converter calibration is provided for an embodiment of the present invention. Figure 2 Flowchart of another delay neural network parameter ant colony optimization method based on pipelined analog-to-digital converter calibration provided in this embodiment of the invention; Figure 3(a) is the ADC output spectrum of the ADC output signal according to an embodiment of the present invention; Figure 3(b) is the ADC output spectrum of the high-precision fitted signal according to an embodiment of the present invention; Figure 4(a) is a schematic diagram of the spectrum of the time-delay neural network before optimization after ant colony optimization in an embodiment of the present invention; Figure 4(b) is a schematic diagram of the spectrum of the time-delay neural network after ant colony optimization according to an embodiment of the present invention; Figure 5 A schematic diagram of the functional modules of an ant colony optimization device for delay neural network parameters based on pipelined analog-to-digital converter calibration provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the hardware architecture of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Currently, the problems of unstable frequency domain performance and narrow frequency domain range remain unsolved. Based on this, the present invention provides a method and apparatus for ant colony optimization of time delay neural network parameters based on pipelined analog-to-digital converter calibration, which can improve frequency domain performance stability and broaden the frequency domain range.

[0023] To facilitate understanding of this embodiment, a detailed description of the ant colony optimization method for time-delay neural network parameters based on pipelined analog-to-digital converter calibration disclosed in this embodiment of the invention will be provided first. This method can be applied to intelligent control devices such as host computers and servers.

[0024] Figure 1 The flowchart illustrates an ant colony optimization method for delay neural network parameters based on pipelined analog-to-digital converter calibration, as provided in this embodiment of the invention.

[0025] like Figure 1 As shown, the method includes the following steps: Step S102: Configure the basic parameters of the ant colony algorithm and the initial parameters of the time-delay neural network.

[0026] The basic parameters of the ant colony algorithm include the number of ants, the pheromone concentration value, and the threshold for the number of iterations; the initial parameters of the time-delay neural network include the total dimension of the time-delay unit, the range of weight values, the range of bias values, the range of the number of neurons in the hidden layer of the three-layer BP neural network, and a validation set composed of multiple frequencies for validation.

[0027] For example, this initial configuration can be understood as dividing the weights and biases into several sets of values ​​within a preset range, and using these sets as nodes on the ant foraging path. Initial pheromones are set for the interconnected node paths, along with the number of ants and the maximum number of iterations.

[0028] For example, first, set the search range for weights and biases to [-1, 1], and then divide them evenly into 100 discrete numerical intervals, with each interval serving as a path node. Set the total number of ants to 50, the maximum number of iterations to 100, and the initial pheromone... The residual information coefficient ρ=0.2, the pheromone constant Q=10, the pheromone factor α=2, and the heuristic factor β=2.5.

[0029] Step S104: Based on the initial parameters of the time-delay neural network, construct a joint solution space containing initial weight configuration and initial bias configuration, and control the ants to start from random nodes and traverse the independent nodes in the joint solution space according to the greedy selection strategy and the random selection strategy to form an ant path.

[0030] Each ant path corresponds to a complete solution in the joint solution space. If a greedy selection strategy is used, nodes with higher pheromone concentrations are selected first; if a random selection strategy is used, nodes with lower pheromone concentrations are selected with a preset probability. In some embodiments, step S104 may be implemented by the following steps: Step 1.1: Based on a preset number of discrete values ​​evenly distributed within the weight range, determine the independent nodes of the ant colony's foraging path; the ants start from a randomly selected starting point and traverse all independent nodes according to a greedy selection strategy and a random selection strategy based on the current pheromone concentration, forming an ant path that matches the number of network connections, and determining the initial weight configuration.

[0031] The number of network connections is calculated as follows: (Number of hidden layer neurons in the initial configuration) × (Number of combinations of latency dimensions) + (Number of hidden layer neurons in the initial configuration) × (Number of output layer nodes). The number of output layer nodes can be 1.

[0032] Step 1.2: Based on the bias value range being evenly divided into a preset number of discrete values, the independent nodes of the ant colony's foraging path are determined; the ants start from a randomly selected starting point and traverse all independent nodes according to the current pheromone concentration using a greedy selection strategy and a random selection strategy to form an ant path that matches the total number of biases, thus determining the initial bias configuration. Wherein, the total number of biases = the number of hidden layer neurons initially configured + the number of output layer nodes; the number of output layer nodes can be 1.

[0033] Step 1.3: Determine each complete solution in the joint solution space based on the initial weight configuration and initial bias configuration.

[0034] It should be noted that each complete solution in the joint solution space is a combination of the above-mentioned initial weight configuration sub-solution and initial bias configuration sub-solution, and each solution must satisfy the adaptability to the time-delay neural network structure to ensure that it can be used for neural network initialization; for example, each ant starts from a random starting point and selects the next node according to probability based on pheromone concentration and heuristic information (here, the preliminary error estimate in reciprocal form), eventually forming a path that runs through all dimensions, representing a complete set of initial weight and bias combinations.

[0035] Step S106: Based on the complete solution, configure the initial parameters of the time-delay neural network, train the time-delay neural network multiple times on a validation set consisting of multiple frequencies, obtain the minimum mean square error of multiple training sessions at each frequency, and calculate the fitness function value of each ant path.

[0036] It is understood that in the embodiments of the present invention, the frequency refers to the signal frequency input to the pipeline ADC, that is, the frequency of the analog input signal itself acquired by the ADC to be calibrated; the multiple frequencies in the verification set refer to the frequencies of such input signals covering different frequency bands (e.g., covering 10~200MHz).

[0037] As an optional embodiment, the step of calculating the fitness function value for each ant path in step S106 of the aforementioned embodiment may include: Based on the minimum mean square error of multiple training sessions at each frequency, the average of the minimum mean square error across all frequencies is calculated, and this average is used as the fitness function value for the current ant path.

[0038] This invention constructs a fitness function for multi-frequency validation sets, using the average of the minimum mean square error (MSE) obtained from multiple training sessions at multiple frequencies as the path evaluation index. This allows the ant colony optimization process to focus not only on the optimal solution at a single frequency, but also on the model's generalization ability and robustness across frequency domains, effectively broadening the effective bandwidth of the calibration method.

[0039] The numerical set of paths traversed by ants is used as the initial parameter configuration for training the neural network. The average of the minimum mean squared errors after multiple training runs at each frequency of the validation set (composed of multiple frequencies) is used as the fitness function value for that ant path. Its formula is:

[0040] Where p represents the number of times the neural network has been trained. Let represent the mean squared loss function for the p-th iteration of neural network training, and m represent the number of samples in a certain training process. This represents the neural network output of the j-th sample. This indicates that the j-th sample corresponds to the output of the high-precision reference ADC, k represents the ant number, and n represents the number of frequencies contained in the training set. This represents the frequency of the i-th frequency point in the training set. This represents the fitness function value. The ant with the smallest fitness value is then selected as the optimal solution for this iteration.

[0041] For example, load this set of parameters into a time-delay neural network model, and run the trainlm optimization algorithm 10 times independently on a validation set (containing five frequency points: 10.3MHz, 50MHz, 100MHz, 150MHz, and 200MHz). Record the minimum MSE value after convergence in each training iteration, and calculate their arithmetic mean as the fitness value of the ant path. The lower the fitness, the better the overall performance of the initial parameter configuration in multi-frequency scenarios.

[0042] Here, the average MSE value from multiple training sessions at different signal frequencies is used as the fitness function value. This is essentially to avoid frequency domain performance instability caused by optimizing only a single signal frequency. If only a single frequency is used for verification, the optimized parameters may become invalid in other frequency bands. That is, this embodiment of the invention verifies the calibration stability of the time-delay neural network optimized by the ant colony at different input signal frequencies. It ensures that the initial parameter configuration optimized by the ant colony can minimize the calibration error of the time-delay neural network over a wide signal frequency range, ultimately achieving the technical goal of broadening the frequency domain range (e.g., increasing the effective bandwidth from 30MHz to 40MHz).

[0043] As another optional embodiment, the step of calculating the fitness function value of each ant path in step S106 of the aforementioned embodiment may further include: Based on the minimum mean square error of multiple training sessions at each frequency and the preset weight coefficients corresponding to each frequency, the minimum mean square error of all frequencies is calculated and weighted averaged. The weighted average result is used as the fitness function value of the current ant path.

[0044] Here, weighting coefficients can be pre-set according to the business conditions of each frequency, and then the minimum mean square error corresponding to each frequency can be weighted and averaged to determine the fitness function value of each ant path.

[0045] Step S108: Update the pheromone concentration on each ant path according to the fitness function value, and repeat steps S104-S106 above for calculating the complete solution and fitness function value until the number of iterations reaches the preset iteration threshold, and output the optimal solution in the joint solution space at this time.

[0046] In some embodiments, step S108 outputs the optimal solution through the following steps: Step 2.1: Based on the fitness function value, update the pheromone concentration of each ant path in the joint solution space through the pheromone positive feedback mechanism of the ant colony algorithm.

[0047] After each iteration, the pheromone levels along the path taken by the optimal ant are updated using the following formula:

[0048] in, This represents the pheromone level from node i to node j in the nth iteration of ant k. This indicates the node at the (n+1)th iteration. i To the node j pheromones, It is the volatile factor of pheromones. For nodes i To the node j The pheromone increment is calculated using the fitness value of the optimal ant.

[0049] Based on this, Based on the characteristics of the task, it is set as follows:

[0050] In the formula, Let Q be the fitness function value of the path traversed by the ant, and let Q be the pheromone constant. For nodes i To the node j The path.

[0051] Step 2.2, repeat steps S104-S106 until the number of iterations reaches the preset iteration threshold, and then terminate the process. The complete solution with the smallest current fitness function value is taken as the optimal solution. The optimal solution includes the optimal initial weight configuration and the optimal initial bias configuration.

[0052] In each iteration, the ant path with the lowest fitness is selected, and its pheromone intensity is updated according to the fitness function formula. The pheromone intensity on other paths is reduced by a factor of ρ. This process is repeated until a preset iteration threshold of 100 iterations is reached, ultimately outputting the globally optimal initial parameter configuration. This embodiment of the invention uses the loss value change rate exponent as a dynamic training control mechanism. By monitoring the trend of loss value changes during continuous iterations, it determines whether to continue training, preventing overfitting or ineffective iterations, thus improving training efficiency and output stability.

[0053] In practical applications, the global search characteristics of the ant colony algorithm solve the problems of traditional parameter configuration being prone to local optima and low efficiency, thus improving the reliability and efficiency of parameter optimization. Combined with the fitness calculation of multi-frequency validation sets, the calibration accuracy of ADCs under wide frequency bands is guaranteed, avoiding the problem of excessive error in some frequency bands caused by single-frequency optimization. Finally, the optimal solution output can support the construction of a high-precision calibration model, effectively reducing ADC quantization error and nonlinear error, and providing technical support for pipelined ADCs to output high-precision digital signals.

[0054] In some embodiments, the method of the present invention can also be used for calibration applications of pipelined analog-to-digital converters, including: Step 3.1: Assign the optimal solution to the time-delay neural network, and use the output data of the pipeline analog-to-digital converter to be calibrated and the high-precision reference signal as input and label for training to complete the construction of the calibration model.

[0055] Among them, high-precision reference signals are generated by combining numerical fitting techniques, which enables low-cost and high-fidelity training label generation in the absence of ultra-high-precision physical reference ADCs, thus ensuring the practical feasibility of deep learning calibration methods.

[0056] First, the pipelined analog-to-digital converter to be calibrated and the high-precision reference analog-to-digital converter can be connected to the same analog signal source to construct a time-delay neural network. The time-delay neural network includes a time-delay unit and a three-layer BP neural network. The output signal of the pipelined analog-to-digital converter to be calibrated is connected to the time-delay unit. The output of the time-delay unit and the output of the pipelined analog-to-digital converter to be calibrated together form a training dataset. It should be noted that the calibration pipeline ADC and the high-precision reference ADC are connected to the same signal source. The delay layer is combined with a 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 layer of the delay neural network. The delay order is adjustable. The output of the delay unit and the output of the pipeline ADC to be calibrated at the current moment together form the input data for training the neural network. The output result of the high-precision reference ADC is used as the label data for training the neural network. 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.

[0057] The output of the pipeline ADC to be calibrated, after passing through a time-delay neural network, and the output of a high-precision reference ADC are used as elements in the loss function calculation of the time-delay neural network to calculate the loss value. The loss value stage uses the rate of change exponent. As the loss function, a preset calibration target value is compared with the loss value. If the loss value is greater than the calibration target value, the time-delay neural network enters training mode, adjusting the weights and biases of the network and iteratively optimizing it. If the loss value is less than or equal to the calibration target value, the network enters output mode without adjusting the weights and biases. After each training iteration, the MSE value is recorded and stored. The MSE value is the exponential rate of change of the loss value during the calculation phase. Intermediate results.

[0058] Loss value stage change rate index The calculation method is as follows:

[0059] in, The output of the time-delay neural network, The result is used as a high-precision reference to the ADC output, where n represents the number of samples in a training process. The neural network loss value uses the mean squared error (MSE), where t represents the current time step, t-1 represents the previous time step, and the rate of change of the loss value is... This represents the rate of change of the difference between the current and previous loss values ​​relative to the previous time step, where... It is an indicator function (taking 1 when the condition in parentheses is met, and 0 otherwise), and threshold is a preset threshold value. hour Select 1 otherwise select 0, where T is a preset number of time points, and the loss value represents the stage change rate exponent. This represents the percentage of instances where the rate of change of the loss value exceeds a preset threshold within the past T time periods.

[0060] The input layer of the time-delay neural network structure constructed in this embodiment receives the current output and the delayed output of the previous 10 time steps (i.e., the delay order is 10) from the pipeline ADC to be calibrated, for a total of 11 dimensions of input; it is connected to a hidden layer with 10 neurons and the activation function is logsig; the output layer is a single node that directly outputs the calibrated value. During training, a high-precision reference signal is used as a label. Since there is no ultra-high-precision physical ADC available in the actual test environment, the MATLAB CurveFitting Toolbox is used to perform high-order polynomial fitting on the acquired ADC output sequence to reconstruct the ideal waveform, with an SNDR of over 157.30dB, as shown in Figures 3(a) and 3(b), which meets the training requirements. During training, a dynamic termination mechanism is adopted: the loss value MSE of each iteration is calculated, and the "loss value change rate exponent" over the past T=20 time steps is calculated according to the aforementioned formula. If the exponent is lower than a set threshold (e.g., 0.1), and the current MSE ≤ the calibration target value (e.g., 1e-6), the model is determined to have converged and enters the output mode.

[0061] In practical applications, this invention uses a commercially available 14-bit resolution, 1Gsps sampling rate pipelined ADC as the test object to verify the practicality of the calibration method. The experimental platform consists of a high-frequency signal source, a clock module, an ADC evaluation board, an FPGA development system, and a data processing host. During the test, the ADC sampling rate is configured to 1GHz, the input signal is a 10.3MHz single-tone sine wave (amplitude set to -1dBFS full scale), and the ADC output data is recorded. Due to the lack of an ultra-high precision reference ADC circuit, this scheme innovatively uses numerical fitting technology to perform precise curve reconstruction on the collected samples. The fitting result will be used as the training target value for the deep learning network. The data modeling work is completed using the curve fitting module of MATLAB. As shown in Figures 3(a) and 3(b), the measured data show that the main performance parameters of the ADC chip to be calibrated are SFDR of 74.88dB, THD of -73.56dB, and ENOB of 10.07. The fitted SNDR is 157.30dB, SFDR is 196.68dB, THD is -196.53dB, and ENOB is 25.83, which fully meet the design requirements of this calibration scheme for the accuracy of the reference source.

[0062] Multiple single-frequency sinusoidal input signals with an amplitude of 0.99V and a time delay order of 10 (11 feature dimensions in total) were used. The activation function for the hidden layer of the neural network was chosen as logsig, and the number of neurons in the hidden layer was set to 10. The neural network optimization algorithm was trainlm. The simulation was based on the same error environment and reference circuit (the effective accuracy of the reference ADC was 22 bits). The training set frequency was set to 10.3MHz, and the test set frequency covered the range of 10 to 200MHz. The simulation was based on 100 independent experiments, and the average of the best results was taken as the performance evaluation index. The initialization parameters of the ant colony algorithm were set as follows: the number of ants was 50, the residual information coefficient ρ was 0.2, the pheromone constant Q was 10, the pheromone factor α and the heuristic factor β were 2 and 2.5, respectively, and the maximum number of iterations was 100.

[0063] As shown in Figures 4(a) and 4(b), the calibration effect of the time-delay neural network after optimization by the ant colony algorithm is stable and effective within a certain frequency domain range, and the applicable frequency domain range is slightly increased. As shown in the figure, the effective application frequency band is approximately 10~40MHz, which is 10MHz higher than before optimization.

[0064] Step 3.2: Use the trained time-delay neural network to calibrate the output of the analog-to-digital converter, and output a high-precision digital signal.

[0065] The present invention represents a technological leap from local optimization to global frequency domain performance enhancement, breaking through the bottleneck of traditional time-delay neural network-based ADC calibration methods in high-frequency applications, and possesses good engineering practicality and promotional value.

[0066] In some embodiments, this invention also provides another method for ant colony optimization of time-delay neural network parameters based on pipelined analog-to-digital converter calibration, specifically, as follows: Figure 2 As shown, this ant colony optimization method treats the combination of weights and biases of the time-delay neural network model as nodes on the ant's foraging path. The optimization process for the initial optimal parameter configuration of the neural network model can be divided into the following four steps: The first step is initialization. The weights and biases are divided into several sets of values ​​within a preset range, and these sets are used as nodes on the ant foraging path. Initial pheromones are set for connected node paths, and the number of ants and the maximum number of iterations are also set.

[0067] The second step is to construct and evaluate the solution space. Ants start from a randomly selected starting point and determine subsequent path nodes based on the current pheromone concentration using both greedy and random selection strategies, until all nodes have been traversed. The numerical set of paths traversed by the ant is used as the initial parameter configuration for neural network training. The average of the minimum mean squared errors after multiple training runs at each frequency of the validation set (composed of multiple frequencies) is used as the fitness function value for that ant path, with the following formula:

[0068] Where p represents the number of times the neural network has been trained. Let represent the mean squared loss function for the p-th iteration of neural network training, and m represent 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.

[0069] The third step is to update the pheromones. After each iteration, the pheromones along the path traversed by the optimal ant are updated according to formula (1.14), where... Based on the characteristics of the task in this article, the following settings are made:

[0070]

[0071] In the formula, Let Q be the fitness function value of the path traversed by the ant, and let Q be the pheromone constant.

[0072] Fourth step, terminate after iteration. Repeat steps two and three until the set maximum number of iterations is reached, finally obtaining the optimal initial parameter configuration.

[0073] As shown in Figures 4(a) and 4(b), the performance before and after calibration was tested in the frequency range of 10–200MHz. The results show that the effective operating bandwidth of the original model without ant colony optimization is about 30MHz; while after optimization by the method of this invention, the effective bandwidth of the system is extended to about 40MHz while maintaining stable calibration accuracy, which is a 33% increase in bandwidth. Moreover, the fluctuation of key indicators such as SFDR, THD, and ENOB is less than ±0.5dB throughout the entire frequency band, demonstrating excellent frequency domain consistency and stability.

[0074] Furthermore, comparative experiments show that this method converges faster than standard BP training and genetic algorithm optimization, and can still maintain good calibration results in the high-frequency range (>80MHz).

[0075] This embodiment achieves the following technical effects by incorporating ant colony optimization into the initial parameter configuration process of a time-delay neural network: 1. Significantly improved the adaptability of the calibration model over a wide frequency range, with an effective bandwidth increase of more than 10MHz; 2. It reduces the risk of oscillation and divergence during training and improves convergence stability; 3. By using numerical fitting to replace the ultra-high precision hardware reference source, the system cost and complexity are reduced; 4. Introducing the rate of change index as a criterion for training state enhances the system's autonomous decision-making ability.

[0076] Therefore, this invention improves the practicality and engineering applicability of the pipeline ADC back-end calibration method while ensuring high precision.

[0077] In some embodiments, such as Figure 5 As shown, this embodiment of the invention provides an ant colony optimization device for time-delay neural network parameters based on pipelined analog-to-digital converter calibration, comprising: The configuration module configures the basic parameters of the ant colony algorithm and the initial parameters of the time-delay neural network. The solution module constructs a joint solution space containing initial weight configuration and initial bias configuration based on the initial parameters of the time-delay neural network, and controls ants to start from random nodes and traverse independent nodes in the joint solution space according to greedy selection strategy and random selection strategy to form ant paths; wherein, each ant path corresponds to a complete solution in the joint solution space. The training module configures the initial parameters of the time-delay neural network based on the complete solution, trains the time-delay neural network multiple times on a validation set consisting of multiple frequencies, obtains the minimum mean square error of multiple training sessions at each frequency, and calculates the fitness function value of each ant path. The iteration module updates the pheromone concentration on each ant path according to the fitness function value, and repeats the steps of calculating the complete solution and fitness function value until the number of iterations reaches the preset iteration threshold, and then outputs the optimal solution of the joint solution space at this time.

[0078] The present invention provides an embodiment of an electronic device. In this embodiment, the electronic device may be, but is not limited to, a personal computer (PC), a laptop computer, a monitoring device, a server, or other computer device with analysis and processing capabilities.

[0079] As an exemplary embodiment, see [reference]. Figure 6 The electronic device 110 includes a communication interface 111, a processor 112, a memory 113, and a bus 114. The processor 112, the communication interface 111, and the memory 113 are connected via the bus 114. The memory 113 is used to store a computer program that supports the processor 112 in executing the above-described method. The processor 112 is configured to execute the program stored in the memory 113.

[0080] The machine-readable storage medium mentioned in this article can be any electronic, magnetic, optical, or other physical storage device that can contain or store information such as executable instructions, data, etc. For example, machine-readable storage media can be: RAM (Random Access Memory), volatile memory, non-volatile memory, flash memory, storage drives (such as hard disk drives), any type of storage disk (such as optical discs, DVDs, etc.), or similar storage media, or combinations thereof.

[0081] Non-volatile media can be non-volatile memory, flash memory, storage drives (such as hard disk drives), any type of storage disk (such as optical discs, DVDs, etc.), or similar non-volatile storage media, or combinations thereof.

[0082] It is understood that the specific operation methods of each functional module in this embodiment can be referred to the detailed description of the corresponding steps in the above method embodiment, and will not be repeated here.

[0083] The computer-readable storage medium provided in the embodiments of the present invention stores a computer program. When the computer program code is executed, it can implement the method described in any of the above embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0084] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0085] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0086] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0087] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention.

Claims

1. A method for ant colony optimization of time-delay neural network parameters based on pipelined analog-to-digital converter calibration, characterized in that, include: Configure the basic parameters of the ant colony algorithm and the initial parameters of the time-delay neural network; Based on a preset number of discrete values ​​evenly distributed within the weight range, independent nodes of the ant colony's foraging path are determined; ants start from a randomly selected starting point and traverse all the independent nodes according to a greedy selection strategy and a random selection strategy based on the current pheromone concentration, forming an ant path that matches the number of network connections, and determining the initial weight configuration. Based on the bias value range being divided into a preset number of discrete values, the independent nodes of the ant colony's foraging path are determined; the ants start from a randomly selected starting point and traverse all the independent nodes according to the current pheromone concentration using a greedy selection strategy and a random selection strategy to form an ant path that matches the total number of biases, thus determining the initial bias configuration. Based on the initial weight configuration and the initial bias configuration, each complete solution in the joint solution space is determined; Based on the complete solution, configure the initial parameters of the time-delay neural network, train the time-delay neural network multiple times on a validation set consisting of multiple frequencies, obtain the minimum mean square error of multiple training sessions at each frequency, and calculate the fitness function value of each ant path. Update the pheromone concentration on each ant path according to the fitness function value, and repeat the steps of calculating the complete solution and fitness function value until the number of iterations reaches the preset iteration threshold, and then output the optimal solution of the joint solution space at this time.

2. The method according to claim 1, characterized in that, The step of calculating the fitness function value for each ant path includes: Based on the minimum mean square error of multiple training sessions at each frequency, the average of the minimum mean square error across all frequencies is calculated, and this average is used as the fitness function value for the current ant path.

3. The method according to claim 1, characterized in that, The step of calculating the fitness function value for each ant path includes: Based on the minimum mean square error of multiple training sessions at each frequency and the preset weight coefficients corresponding to each frequency, the minimum mean square error of all frequencies is calculated and weighted averaged. The weighted average result is used as the fitness function value of the current ant path.

4. The method according to claim 1, characterized in that, The steps of updating the pheromone concentration on each ant path based on the fitness function value, and repeating the aforementioned steps of calculating the complete solution and fitness function value until the number of iterations reaches a preset iteration threshold, and then outputting the optimal solution in the joint solution space at this point, include: Based on the fitness function value, the pheromone concentration of each ant path in the joint solution space is updated through the pheromone positive feedback mechanism of the ant colony algorithm. The process involves re-executing the initial parameters based on the time-delay neural network to construct a joint solution space including initial weight configuration and initial bias configuration. Ants are then controlled to start from random nodes and traverse independent nodes in the joint solution space using both greedy and random selection strategies to form ant paths. Based on the complete solution, the initial parameters of the time-delay neural network are configured. The network is then trained multiple times on a validation set consisting of multiple frequencies. The minimum mean square error of the training at each frequency is obtained, and the fitness function value of each ant path is calculated. This process continues until the number of iterations reaches a preset iteration threshold, at which point the complete solution with the smallest fitness function value is taken as the optimal solution. The optimal solution includes the optimal initial weight configuration and the optimal initial bias configuration.

5. The method according to claim 1 or 4, characterized in that, The method further includes: The optimal solution is assigned to the time-delay neural network, and the output data of the pipeline analog-to-digital converter to be calibrated and the high-precision reference signal are used as inputs and labels for training to complete the construction of the calibration model. The trained time-delay neural network is used to calibrate the output of the analog-to-digital converter, resulting in a high-precision digital signal.

6. The method according to claim 1, characterized in that, The basic parameters of the ant colony algorithm include the number of ants, the pheromone concentration value, and the iteration number threshold; the initial parameters of the time-delay neural network include the total number of dimensions of the time-delay units, the range of weight values, the range of bias values, the range of the number of neurons in the hidden layer of the three-layer BP neural network, and a validation set composed of multiple frequencies for verification.

7. A device for optimizing time-delay neural network parameters based on pipelined analog-to-digital converter calibration by ant colony optimization, characterized in that, include: The configuration module configures the basic parameters of the ant colony algorithm and the initial parameters of the time-delay neural network. The solution module determines the independent nodes of the ant colony's foraging path based on a preset number of discrete values ​​evenly distributed within the weight range. Ants start from a randomly selected starting point and traverse all the independent nodes using a greedy selection strategy and a random selection strategy based on the current pheromone concentration, forming an ant path matching the number of network connections, thus determining the initial weight configuration. Similarly, based on a preset number of discrete values ​​evenly distributed within the bias range, the module determines the independent nodes of the ant colony's foraging path. Ants start from a randomly selected starting point and traverse all the independent nodes using a greedy selection strategy and a random selection strategy based on the current pheromone concentration, forming an ant path matching the total number of biases, thus determining the initial bias configuration. Based on the initial weight configuration and the initial bias configuration, each complete solution in the joint solution space is determined. The training module configures the initial parameters of the time-delay neural network based on the complete solution, trains the time-delay neural network multiple times on a validation set consisting of multiple frequencies, obtains the minimum mean square error of multiple training sessions at each frequency, and calculates the fitness function value of each ant path. The iteration module updates the pheromone concentration on each ant path according to the fitness function value, and repeats the steps of calculating the complete solution and fitness function value until the number of iterations reaches the preset iteration threshold, and then outputs the optimal solution of the joint solution space at this time.

8. An electronic device, characterized in that, It includes a memory, a processor, and a program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed, implements the method described in any one of claims 1-6.

Images