Chassis electric control system parameter sensitivity analysis method based on deep agent model

Abstract

The present application relates to the field of automobile chassis electric control, and particularly relates to a chassis electric control system parameter sensitivity analysis method based on a deep agent model, adopting a three-level progressive strategy of'single factor preliminary screening-modeling-Sobol analysis', first screening by single factor, adjusting each parameter value one by one and observing the system response, eliminating the parameters with less influence on performance, and reducing the range of parameters to be calibrated; then constructing a system response agent model based on BP neural network to replace the simulation model for fast calculation, ensuring the accuracy of the analysis while greatly reducing the calculation cost; finally, through Sobol global sensitivity analysis based on the agent model, the key parameters under each working condition are further extracted. The present application solves the contradiction between calculation efficiency and accuracy in high-dimensional parameter space, and takes into account efficiency and accuracy, providing standardized method support for virtual calibration under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of automotive chassis electronic control, and in particular to a method for parameter sensitivity analysis of chassis electronic control systems based on a deep proxy model. Background Technology

[0002] The calibration of the chassis electronic control system is one of the most crucial steps in the virtual calibration process of intelligent vehicle chassis electronic control systems. Traditional chassis electronic control system calibration methods, based on real vehicles and real-world scenarios and relying on expert experience, are inefficient and cannot perform comprehensive electromechanical performance testing of multiple intelligent chassis systems before real-world road tests. This results in long vehicle development cycles and high calibration costs, contradicting the market demand for shorter vehicle development cycles. Therefore, developing virtual calibration technology for chassis electronic control systems is essential. Virtual calibration is a model-based calibration technology that allows calibration work to begin before vehicle road tests. Furthermore, by employing efficient automatic calibration algorithms, it can significantly shorten development cycles and reduce calibration costs.

[0003] The automotive chassis electronic control system (ECU) has a large number of parameters, each with varying degrees of impact on vehicle handling stability, comfort, and braking performance. Directly calibrating all parameters would lead to excessive computational load, low efficiency, and potentially compromised reliability. Therefore, reducing the number of parameters to be calibrated in the automotive chassis ECU is a key issue in virtual calibration technology. Selecting effective and reasonably sized calibration parameters, i.e., conducting parameter sensitivity analysis, is crucial for automotive chassis ECU calibration. Current parameter sensitivity analysis methods for automotive chassis ECU calibration often employ a brute-force approach, which is time-consuming. Therefore, a scientific and reasonable parameter selection method is urgently needed to quickly and accurately identify and focus on calibrating key parameters that are critical to the control effect of the automotive chassis ECU. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention provides a parameter sensitivity analysis method for chassis electronic control systems based on a deep surrogate model, utilizing the Sobol global sensitivity analysis method, BP neural network, and single-factor analysis. In the single-factor screening stage, each parameter value is adjusted sequentially, and the system response is observed to initially eliminate parameters with minimal impact on performance, thereby narrowing the range of parameters to be calibrated and improving the efficiency of subsequent analysis. Based on this, a system response surrogate model based on a BP neural network is constructed to replace the simulation model for rapid calculation, ensuring analytical accuracy while significantly reducing computational costs. Finally, through Sobol global sensitivity analysis based on the surrogate model, key parameters under various operating conditions are further extracted.

[0005] This invention provides a method for parameter sensitivity analysis of a chassis electronic control system based on a deep proxy model, comprising the following steps:

[0006] Step 1: Single-factor screening:

[0007] For each typical working condition, the single-factor method was used to adjust the parameter values ​​one by one and observe their impact on the performance evaluation index of the chassis electronic control system. The importance of each parameter was quantified, redundant parameters with sensitivity below the threshold were eliminated, and the parameter range for subsequent analysis was narrowed.

[0008] Further steps include:

[0009] First, organize the set of parameters to be optimized, P = {p1, p2, ..., p...} n} and the set of performance evaluation metrics M = {m1, m2, ... m} used to evaluate system performance k}, where p i m is the parameter to be optimized. j As a performance evaluation metric;

[0010] Next, determine the initial or preset values ​​of the parameter set as a benchmark for subsequent comparison and evaluation;

[0011] Then, parameter-by-parameter adjustment is performed: for each parameter p i While keeping other parameters constant, only adjust p i The value of the parameter is calculated, and the adjusted performance evaluation index is determined; it is then determined whether the predetermined number of repetitions has been reached. If not, parameter adjustment continues; if so, the step of calculating the impact factor is initiated. The formula for calculating the impact factor is as follows:

[0012]

[0013] in, Indicates parameter p i The three possible values; Indicates parameter p under operating condition C. i The value of the j-th performance evaluation index is s; m C,basej This represents the baseline value of the j-th performance evaluation index under operating condition C; for each parameter p i Calculate the relative changes of the performance evaluation index for all values ​​under operating condition C, and select the maximum value as the parameter p under that operating condition. i Influence factors;

[0014]

[0015] By quantifying the impact of each parameter on the performance evaluation index using formulas (1) and (2), the key parameters with the greatest impact on the performance evaluation index are selected.

[0016] The influencing factors for each operating condition were normalized:

[0017]

[0018] in, For parameter p under operating condition C i Influence factors; I C,max and I C,min These represent the maximum and minimum values ​​of the influence factors of all parameters under operating condition C, respectively;

[0019] Finally, values ​​with large impact factors are selected and included in the set of key parameters.

[0020] Step 2: Construct a BP neural network proxy model:

[0021] Building upon single-factor screening, a surrogate model specific to each operating condition is constructed using a backpropagation neural network to replace the high-fidelity simulation model, providing data for subsequent Sobol global sensitivity analysis. By reducing the direct use of the high-precision simulation model during Sobol analysis, the surrogate model significantly improves computational efficiency while maintaining high computational accuracy.

[0022] Further steps include:

[0023] First, when constructing the BP neural network surrogate model, a uniform sampling method is used as the sampling method for training the surrogate model to generate high-quality input samples to cover the parameter space and provide sufficient data support for model training.

[0024] Secondly, the data is normalized using the following formula:

[0025]

[0026] Where z is the normalization result, and x is the calculated value. min x is the minimum value in the data. max This represents the maximum value in the data.

[0027] Finally, the relative root mean square error is chosen as the evaluation metric for the model's predictive performance to measure the error between the predicted and actual values. The calculation formula is as follows:

[0028]

[0029] Where RRMSE is the relative root mean square error, t i For the actual value, y i This is an estimated value.

[0030] The number of input layer nodes in the BP neural network structure is the number of parameters obtained from single-factor screening; the output of the surrogate model is the performance evaluation index value of the chassis electronic control system; the number of intermediate layers is selected as 2-4 according to the number of indicators, and the number of neurons is set to 20-80.

[0031] Step 3: Sobol Global Sensitivity Analysis:

[0032] Based on the surrogate model, a systematic global sensitivity analysis is adopted. The first-order and total sensitivity indices are calculated using the Sobol method to quantify the main effects and interaction effects of parameters, thereby achieving efficient and accurate assessment of the sensitivity of chassis electronic control system parameters. This guides the efficient and accurate optimization of the chassis electronic control system and provides reliable technical support for virtual calibration under complex working conditions.

[0033] Further steps include:

[0034] First, the input parameters and their feasible domain range are determined based on the single-factor screening results, and the physical constraints of each parameter are defined. A uniform sampling strategy is adopted to generate a set of sample points with uniform distribution characteristics in the high-dimensional parameter space to ensure full coverage of the parameter space.

[0035] Then, the BP neural network surrogate model trained in step two is used to perform efficient performance prediction, calculate the sensitivity index, and establish a sensitivity feedback mechanism to guide closed-loop calibration optimization.

[0036] The steps for calculating the sensitivity index using Sobol include: First, defining the mathematical model, i.e., first specifying a mathematical model Y = f(X), where X = (x1, x2, ... x...). n The first step involves decomposing the input variable set to establish the relationship between the output and these inputs. The second step is to decompose the function f(X) into a polynomial form that includes constant terms, first-order effect terms, second-order effect terms, and so on up to the total effect term, so that the influence of each input variable and its interaction on the model output can be quantified. The third step is to determine the total variance and the contribution of each effect term to the variance by calculating the integral of each effect term. Then, based on these contributions, the corresponding sensitivity indices are calculated, including the main effect index (first-order sensitivity index), the second-order interaction effect index (second-order sensitivity index), and the total effect index (total sensitivity index), so as to assess the importance of each input variable and the degree of influence of its interaction effect on the output.

[0037] Where Y = f(X) represents the model being analyzed, which has an n-dimensional vector input X = (x1, x2, ... x... n );x i V represents the input to the model; V represents the total variance of the model's output f(X) when all input values ​​fluctuate; V i This represents a portion of the variance in the model output when the value of a single input fluctuates. This indicates that the interaction between multiple input variables contributes a portion of the variance to the model output.

[0038] The formula for calculating the main effect index is as follows:

[0039]

[0040] Among them, S i x represents i The main effects index, or first-order sensitivity index, is used to measure the effects of a single input variable x. i The fluctuations of the output have an independent effect on the output; E represents the expected value;

[0041] The formula for calculating the second-order interaction effect index is as follows:

[0042]

[0043] Among them, S ij x represents i and x j The second-order interaction effect index, or second-order sensitivity index, is used to measure the input variable x. i and x j The degree to which the interaction affects the output;

[0044] The formula for calculating the total effect index is as follows:

[0045]

[0046] in, This represents the total effect index or total sensitivity index, used to measure the overall sensitivity when x... i When the value of fluctuates, the degree of contribution of this input to the total output variance is... The larger the value, the greater the impact of this parameter across all scenarios.

[0047] Finally, a screening threshold for the total effect index is set. For a certain performance indicator, if the total effect index of a certain parameter is... If the value is greater than the screening threshold, it is included in the range of parameters to be calibrated, i.e., the set of key parameters; if If the value is less than the screening threshold, it will not be included in the range of parameters to be calibrated.

[0048] The beneficial effects of this invention are:

[0049] This invention employs a three-tiered progressive strategy: "single-factor preliminary screening - modeling - Sobol analysis." First, single-factor screening eliminates parameters with minimal impact on system performance, avoiding redundancy in global analysis. Then, a surrogate model replaces high-precision simulation to accelerate computation, dynamically optimizing sample distribution and reducing the computational complexity of high-dimensional sensitivity analysis. Finally, Sobol global sensitivity analysis quantifies main effects and interaction effects, comprehensively evaluating the impact of parameters on system performance under different operating conditions, extracting key parameters from numerous parameters, and laying the foundation for subsequent parameter calibration. This approach resolves the contradiction between computational efficiency and accuracy in high-dimensional parameter spaces, balancing efficiency and accuracy, and providing standardized methodological support for virtual calibration under complex operating conditions. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the overall process of the chassis electronic control system parameter sensitivity analysis method based on the deep proxy model of the present invention;

[0051] Figure 2 This is a schematic diagram of the single-factor screening process for the parameters of this invention;

[0052] Figure 3 This is a schematic diagram of the BP neural network structure of the present invention;

[0053] Figure 4 This is a schematic diagram of the Sobol global sensitivity analysis process of the present invention;

[0054] Figure 5 This is a schematic diagram illustrating the steps for calculating the Sobol sensitivity index in this invention. Detailed Implementation

[0055] like Figure 1 As shown, this invention provides a method for sensitivity analysis of chassis electronic control system parameters based on a deep proxy model. This embodiment takes an ABS system as an example, and the analysis method includes the following steps:

[0056] Step 1: Single-factor screening:

[0057] Screening process as follows Figure 2 As shown, for each typical working condition, the single-factor method was used to adjust the parameter values ​​one by one and observe their impact on the ABS performance evaluation index, thereby quantifying the importance of each parameter, quickly eliminating redundant parameters with sensitivity below the threshold, narrowing the parameter range for subsequent analysis, and significantly reducing the parameter dimension of subsequent analysis.

[0058] Further steps include:

[0059] First, organize the set of parameters to be optimized in the ABS system, P = {p1, p2, ..., p...} n} and the set of performance evaluation metrics M = {m1, m2, ... m} used to evaluate system performancek}, where p i For the parameters to be optimized in the ABS system, m j As a performance evaluation metric;

[0060] Next, determine the initial or preset values ​​of the parameter set as a benchmark for subsequent comparison and evaluation;

[0061] Then, parameter-by-parameter adjustment is performed: for each parameter p i While keeping other parameters constant, only adjust p i The value of the parameter is calculated, and the adjusted performance evaluation index is determined; it is then determined whether the predetermined number of repetitions has been reached. If not, parameter adjustment continues; if so, the step of calculating the impact factor is initiated. The formula for calculating the impact factor is as follows:

[0062]

[0063] in, Indicates parameter p i The three possible values; Indicates parameter p under operating condition C. i The value of the j-th performance evaluation index is s; m C,basej This represents the baseline value of the j-th performance evaluation index under operating condition C; for each parameter p i Calculate the relative changes of the performance evaluation index for all values ​​under operating condition C, and select the maximum value as the parameter p under that operating condition. i Influence factors;

[0064]

[0065] By quantifying the impact of each parameter on the performance evaluation index using formulas (1) and (2), the key parameters with the greatest impact on the performance evaluation index are selected.

[0066] Since the order of magnitude of the influencing factors of parameters may vary significantly under different operating conditions, this invention normalizes the influencing factors for each operating condition to ensure comparability.

[0067]

[0068] in, For parameter p under operating condition C i Influence factors; I C,max and I C,min These represent the maximum and minimum values ​​of the influence factors of all parameters under operating condition C, respectively;

[0069] Finally, values ​​with large impact factors are selected and included in the set of key parameters.

[0070] Step 2: Construct a BP neural network proxy model:

[0071] Based on single-factor screening, a proxy model for each working condition is constructed using a BP neural network to replace the high-fidelity simulation model. In this embodiment, it is used to replace the complex ABS simulation model, providing data for subsequent Sobol global sensitivity analysis. By reducing the direct calls to the high-precision simulation model during Sobol analysis, the proxy model reduces computation time by more than 90% while maintaining accuracy, thus significantly improving computational efficiency while ensuring high computational accuracy.

[0072] Further steps include:

[0073] First, when constructing the BP neural network surrogate model, high-quality input samples need to be generated to cover the parameter space and provide sufficient data support for model training. Due to the complex characteristics of the ABS system with multiple parameters and multiple outputs, efficient sampling should be achieved with limited computational resources. Therefore, uniform sampling is used as the sampling method for training the surrogate model.

[0074] Secondly, since the various control parameters of the ABS system are index values ​​with different dimensions and large differences in their value ranges, in order to avoid certain variables having an excessive impact on the results during network training, accelerate model convergence, and improve prediction accuracy, this invention performs data normalization processing, and the calculation formula is as follows:

[0075]

[0076] Where z is the normalization result, and x is the calculated value. min x is the minimum value in the data. max This represents the maximum value in the data.

[0077] Finally, the relative root mean square error is chosen as the evaluation metric for the model's predictive performance to measure the error between the predicted and actual values. The calculation formula is as follows:

[0078]

[0079] Where RRMSE is the relative root mean square error, t i For the actual value, y i This is an estimated value.

[0080] The BP neural network structure diagram is as follows: Figure 3As shown, the input data of the proxy model for each working condition are parameters obtained through single-factor screening. Therefore, the number of nodes in the input layer should be the number of parameters obtained through single-factor screening. The output of the proxy model is the ABS performance evaluation index value. The specific number of output nodes is determined by the number of evaluation indicators used in the working condition. The number of intermediate layers is selected as 2-4 according to the number of indicators, and the number of neurons is set to 20-80. The specific values ​​are adjusted through experiments to find the optimal configuration.

[0081] Step 3: Sobol Global Sensitivity Analysis:

[0082] Based on the surrogate model, a systematic global sensitivity analysis is adopted, and the first-order and total effect sensitivity indices are calculated using the Sobol method to quantify the main effects and interaction effects of parameters, thereby guiding the efficient and accurate optimization of the ABS system.

[0083] Further steps include:

[0084] First, the input parameters and their feasible domain range are determined based on the single-factor screening results, and the physical constraints of each parameter are defined. A uniform sampling strategy is adopted to generate a set of sample points with uniform distribution characteristics in the high-dimensional parameter space to ensure full coverage of the parameter space.

[0085] Then, the BP neural network surrogate model trained in step two is used to perform efficient performance prediction, calculate the sensitivity index, and establish a sensitivity feedback mechanism to guide closed-loop calibration optimization.

[0086] The steps for calculating the sensitivity index using Sobol include: First, defining the mathematical model, i.e., first specifying a mathematical model Y = f(X), where X = (x1, x2, ... x...). n The first step involves decomposing the input variable set to establish the relationship between the output and these inputs. The second step is to perform function decomposition, breaking down the function f(X) into a polynomial form that includes constant terms, first-order effect terms, second-order effect terms, and finally the total effect term. This allows for the quantification of the impact of each input variable and its interactions on the model output. The third step involves calculating the integral of each effect term to determine the total variance and the contribution of each effect term to the variance. Based on these contributions, corresponding sensitivity indices are calculated, including the main effect index (first-order sensitivity index), the second-order interaction effect index (second-order sensitivity index), and the total effect index (total sensitivity index). This allows for the assessment of the importance of each input variable and the extent to which its interaction effects affect the output. This series of operations ultimately helps identify key influencing factors and optimize model parameters.

[0087] Where Y = f(X) represents the model being analyzed, which has an n-dimensional vector input X = (x1, x2, ... x... n );x iV represents the input to the model; V represents the total variance of the model's output f(X) when all input values ​​fluctuate; V i This represents a portion of the variance in the model output when the value of a single input fluctuates. This indicates that the interaction between multiple input variables contributes a portion of the variance to the model output.

[0088] The formula for calculating the main effect index is as follows:

[0089]

[0090] Among them, S i x represents i The main effects index, or first-order sensitivity index, is used to measure the effects of a single input variable x. i The fluctuations of the output have an independent effect on the output; E represents the expected value;

[0091] The formula for calculating the second-order interaction effect index is as follows:

[0092]

[0093] Among them, S ij x represents i and x j The second-order interaction effect index, or second-order sensitivity index, is used to measure the input variable x. i and x j The degree to which the interaction affects the output;

[0094] The formula for calculating the total effect index is as follows:

[0095]

[0096] in, This represents the total effect index or total sensitivity index, used to measure the overall sensitivity when x... i When the value of fluctuates, the degree of contribution of this input to the total output variance is... The larger the value, the greater the impact of this parameter across all scenarios.

[0097] Set a screening threshold for the total effect index. For a certain performance indicator, if the total effect index of a certain parameter... If the value is greater than the screening threshold, it is included in the range of parameters to be calibrated, i.e., the set of key parameters; if If the value is less than the screening threshold, it will not be included in the range of parameters to be calibrated, so as to improve the calculation efficiency and ensure the relevance of the calibration work.

[0098] Finally, by combining sampling strategy optimization, surrogate model integration, and statistical verification, we can achieve efficient and accurate assessment of the sensitivity of ABS system parameters, providing reliable technical support for virtual calibration under complex working conditions.

Claims

1. A method for parameter sensitivity analysis of a chassis electronic control system based on a deep surrogate model, characterized in that: Includes the following steps: Step 1: Single-factor screening: For each typical working condition, the single-factor method was used to adjust the parameter values ​​one by one and observe their impact on the performance evaluation index of the chassis electronic control system. The importance of each parameter was quantified, redundant parameters with sensitivity below the threshold were eliminated, and the parameter range was narrowed. Step 2: Construct a BP neural network proxy model: Based on single-factor screening, a proxy model for each working condition is constructed using a BP neural network to replace the high-fidelity simulation model. First, when constructing the BP neural network proxy model, a uniform sampling method is used as the sampling method for training the proxy model. Second, the data is normalized. Finally, the relative root mean square error is selected as the evaluation index for the model's predictive performance. Step 3: Sobol Global Sensitivity Analysis: First, based on the results of single-factor screening, the input parameters and their feasible domain range are determined, the physical constraints of each parameter are defined, and a uniform sampling strategy is adopted to generate a set of sample points with uniform distribution characteristics in the high-dimensional parameter space to ensure full coverage of the parameter space. Then, the BP neural network surrogate model trained in step two is used for efficient performance prediction. A systematic global sensitivity analysis is adopted, and the sensitivity index is calculated by the Sobol method to quantify the main effect and interaction effect of the parameters. A sensitivity feedback mechanism is established to guide the closed-loop calibration optimization. The steps for calculating the sensitivity index using the Sobol method include: First, defining the mathematical model, i.e., first specifying a mathematical model Y = f(X), where X = (x1, x2, ... x...). n The first step involves decomposing the input variable set to establish the relationship between the output and these inputs. The second step is to decompose the function f(X) into a polynomial form that includes constant terms, first-order effect terms, second-order effect terms, and so on up to the total effect term, so that the impact of each input variable and its interaction on the model output can be quantified. The third step is to determine the total variance and the contribution of each effect term to the variance by calculating the integral of each effect term. Then, based on these contributions, the corresponding sensitivity indices are calculated, including the main effect index, the second-order interaction effect index, and the total effect index, so as to assess the importance of each input variable and the degree of influence of its interaction effect on the output. Where Y = f(X) represents the model being analyzed, which has an n-dimensional vector input X = (x1, x2, ... x... n );x i V represents the input to the model; V represents the total variance of the model's output f(X) when all input values ​​fluctuate; V i This represents a portion of the variance in the model output when the value of a single input fluctuates. This indicates that the interaction between multiple input variables contributes a portion of the variance to the model output. The formula for calculating the main effect index is as follows: Among them, S i x represents i The main effect index or first-order sensitivity index; E represents the expected value; The formula for calculating the second-order interaction effect index is as follows: Among them, S ij x represents i and x j The second-order interaction effect index or the second-order sensitivity index; The formula for calculating the total effect index is as follows: in, This represents the total effect index or the overall sensitivity index. Finally, a screening threshold for the total effect index is set. For a certain performance indicator, if the total effect index of a certain parameter is... If the value is greater than the screening threshold, it is included in the range of parameters to be calibrated, i.e., the set of key parameters; if If the value is less than the screening threshold, it will not be included in the range of parameters to be calibrated; thus achieving efficient and accurate evaluation of the sensitivity of chassis electronic control system parameters.

2. The method for sensitivity analysis of chassis electronic control system parameters based on a deep proxy model according to claim 1, characterized in that: The steps of the single-factor screening described in step one include: First, organize the set of parameters to be optimized, P = {p1, p2, ..., p...} n } and the set of performance evaluation metrics M = {m1, m2, ... m} used to evaluate system performance k }, where p i m is the parameter to be optimized. j As a performance evaluation metric; Next, determine the initial or preset values ​​of the parameter set as a benchmark for subsequent comparison and evaluation; Then, parameter-by-parameter adjustment is performed: for each parameter p i While keeping other parameters constant, only adjust p i The value of the parameter is calculated, and the adjusted performance evaluation index is determined; it is then determined whether the predetermined number of repetitions has been reached. If not, parameter adjustment continues; if so, the step of calculating the impact factor is initiated. The formula for calculating the impact factor is as follows: in, Indicates parameter p i The three possible values; Indicates parameter p under operating condition C. i The value of the j-th performance evaluation index is s; m C,basej This represents the baseline value of the j-th performance evaluation index under operating condition C; for each parameter p i Calculate the relative changes of the performance evaluation index for all values ​​under operating condition C, and select the maximum value as the parameter p under that operating condition. i Influence factors; By quantifying the impact of each parameter on the performance evaluation index using formulas (1) and (2), the key parameters with the greatest impact on the performance evaluation index are selected. The influencing factors for each operating condition were normalized: in, For parameter p under operating condition C i Influence factors; I C,max and I C,min These represent the maximum and minimum values ​​of the influence factors of all parameters under operating condition C, respectively; Finally, values ​​with large impact factors are selected and included in the set of key parameters.

3. The method for sensitivity analysis of chassis electronic control system parameters based on a deep proxy model according to claim 1, characterized in that: The calculation formula for normalizing the data described in step two is as follows: Where z is the normalization result, and x is the calculated value. min x is the minimum value in the data. max This represents the maximum value in the data.

4. The method for sensitivity analysis of chassis electronic control system parameters based on a deep proxy model according to claim 1, characterized in that: The formula for calculating the relative root mean square error mentioned in step two is as follows: Where RRMSE is the relative root mean square error, t i For the actual value, y i This is an estimated value.

5. The method for sensitivity analysis of chassis electronic control system parameters based on a deep proxy model according to claim 1, characterized in that: The number of input layer nodes in the BP neural network structure described in step two is the number of parameters obtained from single-factor screening; the output of the surrogate model is the performance evaluation index value of the chassis electronic control system; the number of intermediate layers is selected as 2-4 according to the number of indicators, and the number of neurons is set to 20-80.

Images