CN-121997108-A - Visual distance and non-visual distance signal classification device based on sand cat group optimization algorithm

Abstract

The invention relates to a line of sight (LOS) and non-line of sight (NLOS) signal classification device based on a sand cat swarm optimization algorithm, belonging to the crossing field of an intelligent optimization algorithm and a neural network. The device combines LSTM and CNN through a deep learning model to extract and classify the characteristics of the channel impulse response sequence of the communication signal. Wherein LSTM captures global temporal features and CNN extracts local patterns. According to the invention, the super parameters of the CNN convolution layer are dynamically adjusted by utilizing the sand cat swarm optimization algorithm, so that local optimization is avoided, and global searching capability is improved. The device realizes accurate identification of NLOS propagation, improves the accuracy and robustness of indoor positioning, and solves the problem of low efficiency of traditional manual parameter adjustment. The invention has excellent performance in complex environment and obvious technical advantages and application prospect.

Inventors

• DENG ZHONGLIANG
• Tian Zidu
• LIU YANG

Assignees

• 北京邮电大学

Dates

Publication Date

20260508

Application Date

20241104

Claims (8)

1. 1. The utility model provides a sight distance and non-sight distance signal classification device based on sand cat crowd optimization algorithm which characterized in that includes: a) The data preprocessing module is used for receiving Channel Impulse Response (CIR) sequences of communication signals, and carrying out standardization processing on the sequences to ensure that all CIR sequences have the same length so as to adapt to the input requirement of a deep learning model; b) A deep learning model combining a long short term memory network (LSTM) and a Convolutional Neural Network (CNN) for feature extraction and classification of the normalized CIR sequence; An LSTM layer for capturing global temporal features in the CIR sequence; The CNN layer is arranged behind the LSTM layer and is used for further extracting local features in the CIR sequence, wherein the convolution kernel size and the convolution quantity of the CNN layer are optimized and adjusted through a satay group optimization algorithm (SCSO); c) Sha Mao group optimization algorithm (SCSO) module for automatically adjusting the number of convolution kernels in CNN layer to optimize the performance of deep learning model, wherein the SCSO module comprises a hunting stage and an attack hunting stage, and the positions of particles are updated by different strategies to adjust the super parameters of CNN; d) The global pooling layer is arranged behind the CNN layer and is used for pooling the output of the CNN layer and uniformly representing CIR sequences with different lengths into a characteristic vector with fixed length; e) The full-connection layer and the classification module are used for receiving the output of the global pooling layer, mapping the output to a scalar, representing the probability that the sample belongs to a non-line-of-sight (NLOS) or line-of-sight (LOS) signal, and converting the probability value into a classification label through a Sigmoid activation function and a preset threshold value; Wherein, the optimization algorithm (SCSO) module of the sand cat group is specifically realized as follows: initializing a population of particles, each particle representing one possible CNN convolution kernel configuration; Training a CNN model according to the convolution kernel configuration represented by the current particle, and calculating the F1 fraction of the model on the verification set as the fitness value of the particle; in the hunting stage, the random expansion of the hunting range is introduced to avoid sinking into local optimum; in the stage of attacking prey, the method is focused on the current optimal solution, and local searching is carried out by adopting a more targeted strategy; and iteratively updating the positions of the particles until the preset termination condition is met, so as to obtain the optimal convolution kernel configuration.
2. 2. The apparatus of claim 1, wherein the data preprocessing module unifies all CIR sequences to the same length by padding zeros.
3. 3. The apparatus of claim 1, wherein the LSTM layer is a bi-directional LSTM layer for processing CIR sequences simultaneously in both forward and backward directions to more fully capture dependencies in the sequences.
4. 4. The apparatus of claim 1, wherein the CNN layer comprises three convolutional layers, each of which is followed by a ReLU activation function to increase the nonlinear representation capability of the network.
5. 5. The apparatus of claim 1, wherein the global pooling layer employs an averaging pooling operation to average the output of the convolutional layer in a time dimension to eliminate the effect of input sequence length on the model.
6. 6. The apparatus of claim 1, wherein the full-join layer maps the high-dimensional feature vector onto a scalar and limits the output to between 0 and 1 through a Sigmoid activation function, representing the probability value of the classification, and wherein the final classification result classifies samples with probability values greater than 0.5 as NLOS and less than 0.5 as LOS through a threshold determination.
7. 7. The apparatus of claim 1 wherein the SCSO module introduces a random factor and a random angle during a search for prey stage to increase diversity of searches, and employs a learning factor to increase convergence of particles to an optimal location during an attack for prey stage.
8. 8. The apparatus of any one of claims 1 to 7, further comprising a computer readable storage medium having computer readable instructions stored therein which, when executed by a processor, implement the line-of-sight and non-line-of-sight signal classification method based on a salsa group optimization algorithm.

Description

Visual distance and non-visual distance signal classification device based on sand cat group optimization algorithm Technical Field The invention relates to the crossing field of intelligent optimization algorithms and neural networks, in particular to complex data analysis and classification tasks by combining a satay cluster optimization algorithm (SCSO) with the neural networks. By optimizing the hyper-parametric configuration in the neural network architecture, a line-of-sight (LOS) and non-line-of-sight (NLOS) signal classification device with more efficient feature extraction and classification capabilities is achieved. Technical Field Since NLOS propagation has a significant negative impact on the positioning performance of UWB indoor positioning systems, identification of NLOS propagation appears to be critical. Much research is devoted to the development of NLOS identification methods, with common techniques including signal feature analysis, machine learning, and deep learning methods. Signal characteristic analysis typically relies on signal propagation characteristics such as signal strength, phase and delay, and by analyzing these characteristics it is possible to determine whether NLOS conditions are present in the signal propagation path. Many methods of deep learning and machine learning rely on extracting features from the Channel Impulse Response (CIR) data and training a model using a marker dataset to distinguish LOS from NLOS propagation. The scheme of feature extraction sometimes adopts a Gaussian distribution method and a generalized Gaussian distribution method, and higher classification accuracy is displayed when an unbalanced data set is processed. In addition, researchers have compared the performance of a variety of machine learning classifiers and explored the effectiveness of different feature selection strategies. In order to improve the accuracy of NLOS recognition, researchers have also tried to combine various signal feature analysis methods, even introducing complex deep learning models such as CNN and transfer learning. However, there are certain limitations to manually extracting features. The characteristics of manual designs are often based on empirical analysis of the channel impulse response, and may exhibit significant differences from environment to environment, making it difficult to accommodate complex propagation environments. Conventional feature extraction methods typically focus on signal features in only a single aspect, and it is difficult to fully characterize the complexity of UWB signals under NLOS conditions. Thus, these manual features may lead to a significant decrease in classification accuracy in a diverse actual scenario. To solve these problems, many researchers have come to discuss the characteristics of UWB signals globally, and new methods such as full convolution networks and self-attention mechanisms are adopted to improve feature extraction capability and classification accuracy. Also, researchers have proposed a transducer-based signal denoising network that can recover clean CIR from noise signals, reducing errors introduced by NLOS and multipath signals. In addition, the method of high-level abstract feature conversion and parallel deep learning model also provides a new thought for NLOS identification. The focus of this study is to globally study the characteristics of UWB signals, aimed at achieving recognition of NLOS propagation without relying on artificial feature extraction, in order to improve recognition accuracy in complex environments. And the performance of the design device is improved through an optimization algorithm. Disclosure of Invention In view of this, the invention provides a classification device for non-line-of-sight signals combining a search optimization algorithm of a sand cat with a deep learning model (the classification device is used for the feature extraction and classification tasks of complex data), SCSO simulates hunting behavior of the sand cat, effectively avoids the problem that the conventional intelligent optimization algorithm is easy to sink into local optimum by dynamically adjusting the jump and fine search strategies in the search process, and improves the global search capability. One aspect of the invention provides a neural network for line-of-sight and non-line-of-sight signal classification, comprising: The channel impulse response sequence of the communication signal to be received is used to reflect the propagation path information of the signal at different points in time. In some embodiments of the present invention, since the channel impulse response sequence lengths of different communication signals may not be uniform, it is first necessary to normalize these data. The specific operation is that all channel impulse response sequences are uniformly padded to the same length by means of a padding sequence. The processing mode ensures that the sequences have the same dimension when the