Recently, there has been great interest in estimating the mean of the maximum field strength in a nested reverberation chamber in such conditions that the field coupled inside the equipment under test (EUT) deviates from the parent distribution, thus generating a non-Rayleigh distribution. For this purpose, a new regression model with the deep feedforward neural network is proposed to predict the mean of the maximum field inside a nested reverberation chamber configuration. In our proposed method, a frequency range that comprises the EUT in the overmoded regime is treated as an input of the network, and the mean of the maximum field is treated as the output of the network. Several networks with different numbers of hidden layers are trained, while adaptive learning rates and early stopping techniques are used to improve the network training process, subsequently reducing the uncertainties. After training, the networks are verified using a test set that is not implicitly employed during the training process. The testing and training mean-squared errors (625e-5 and 325e-5) with the network with five layers have a good agreement for a considered configuration, demonstrating a novel regression model that is able to rigorously extrapolate the mean of the maximum field in the other frequency steps that are not used in the training set.

中文翻译：

---

使用深度神经网络估计混响室内最大场的平均值

最近，在被测设备 (EUT) 内耦合的场偏离母分布的情况下，估计嵌套混响室中最大场强的平均值引起了极大的兴趣，从而产生了非瑞利分布。为此，提出了一种具有深度前馈神经网络的新回归模型来预测嵌套混响室配置内最大场的平均值。在我们提出的方法中，包含处于过模态的 EUT 的频率范围被视为网络的输入，最大场的平均值被视为网络的输出。训练了几个具有不同隐藏层数的网络，而自适应学习率和早期停止技术用于改进网络训练过程，从而减少不确定性。训练后，使用训练过程中未隐式使用的测试集验证网络。具有五层网络的测试和训练均方误差（625e-5 和 325e-5）与所考虑的配置具有良好的一致性，展示了一种新颖的回归模型，该模型能够严格外推最大场的平均值训练集中未使用的其他频率步长。