Line-of-sight and non-line-of-sight signal classification device based on salat swarm optimization algorithm

Abstract

The invention relates to a LOS (Line Of Sight) and NLOS (Non-Line Of Sight) signal classification device based on a salat swarm optimization algorithm, and belongs to the crossing field of an intelligent optimization algorithm and a neural network. The device performs feature extraction and classification on a channel impulse response sequence of a communication signal through a deep learning model in combination with LSTM and CNN. Wherein the LSTM captures global time features, and the CNN extracts local modes. According to the method, the hyper-parameters of the CNN convolutional layer are dynamically adjusted by using the samson swarm optimization algorithm, so that local optimum is avoided, and the global search capability is improved. According to the device, accurate identification of NLOS propagation is realized, the precision and robustness of indoor positioning are improved, and the problem of low efficiency of traditional manual parameter adjustment is solved. The method is excellent in performance in a complex environment, and has remarkable technical advantages and application prospects.

Description

Technical Field

[0001] This invention relates to the intersection of intelligent optimization algorithms and neural networks, particularly combining the Sand Cat Group Optimization (SCSO) algorithm with neural networks for complex data analysis and classification tasks. By optimizing the hyperparameter configuration in the neural network structure, 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 Background

[0002] Since NLOS propagation has a significant negative impact on the positioning performance of UWB indoor positioning systems, its identification is crucial. Many studies have focused on developing NLOS identification methods, with common techniques including signal feature analysis, machine learning, and deep learning. Signal feature analysis typically relies on the signal propagation characteristics, such as signal strength, phase, and time delay. By analyzing these characteristics, it is possible to determine whether NLOS occurs in the signal propagation path.

[0003] Many deep learning and machine learning methods rely on extracting features from channel impulse response (CIR) data and training models using labeled datasets to distinguish between Loss of System (LOS) and Non-Standard Operating System (NLOS) propagation. Feature extraction schemes sometimes employ Gaussian and generalized Gaussian distributions, which show higher classification accuracy when dealing with imbalanced datasets. Furthermore, researchers have compared the performance of various machine learning classifiers and explored the effectiveness of different feature selection strategies. To improve the accuracy of NLOS identification, researchers have also experimented with methods combining the analysis of multiple signal features, and even introduced complex deep learning models such as CNNs and transfer learning.

[0004] However, manually extracted features have certain limitations. Manually designed features are often based on empirical analysis of channel impulse responses, which may exhibit significant differences in different environments and are difficult to adapt to complex propagation conditions. Traditional feature extraction methods typically focus only on single aspects of signal characteristics, making it difficult to comprehensively characterize the complexity of UWB signals under NLOS conditions. Therefore, these manually extracted features may lead to a significant decrease in classification accuracy in variable real-world scenarios.

[0005] To address these issues, many researchers have begun to explore the characteristics of UWB signals globally, employing novel methods such as fully convolutional networks and self-attention mechanisms to improve feature extraction capabilities and classification accuracy. Other researchers have proposed Transformer-based signal denoising networks capable of recovering clean CIRs from noisy signals, reducing errors introduced by NLOS and multipath signals. Furthermore, methods utilizing high-level abstract feature transformations and parallel deep learning models have also provided new insights into NLOS recognition.

[0006] This study focuses on a global analysis of the characteristics of UWB signals, aiming to achieve identification of NLOS propagation without relying on manual feature extraction, thereby improving identification accuracy in complex environments. Furthermore, the performance of the designed device is improved through algorithm optimization. Summary of the Invention

[0007] In view of this, this invention proposes a non-line-of-sight signal classification device that combines a sand cat search optimization algorithm with a deep learning model for feature extraction and classification tasks of complex data. SCSO simulates the hunting behavior of sand cats and effectively avoids the problem of getting trapped in local optima in traditional intelligent optimization algorithms by dynamically adjusting the jumping and fine-grained search strategies during the search process, thereby improving the global search capability.

[0008] One aspect of the present invention provides a neural network for line-of-sight and non-line-of-sight signal classification, comprising:

[0009] The channel impulse response sequence of the received communication signal is used to reflect the propagation path information of the signal at different points in time.

[0010] In some embodiments of the present invention, since the channel impulse response sequence lengths of different communication signals may be inconsistent, it is first necessary to standardize these data. Specifically, this is achieved by padding all channel impulse response sequences to the same length. This process ensures that the sequences have the same dimension when input to the deep learning model, avoiding network instability or computational errors caused by irregular input data.

[0011] The line-of-sight (NLOS) and non-line-of-sight (LOS) signal classification device proposed in this invention incorporates a deep learning model. This network model combines a Long Short-Term Memory (LSTM) network and a Convolutional Neural Network (CNN) to perform in-depth feature extraction and achieve high-precision classification of channel impulse response (LOS) sequences of communication signals. LSTM is a neural network structure specifically designed for processing sequential data, capable of capturing long-term temporal dependencies within sequences. Channel impulse response sequences are essentially time-series data, containing important temporal information. Through the LSTM network, the model can learn the evolution trend of the signal at different time steps, thereby extracting features useful for NLOS and LOS classification, such as... Figure 1The model addresses the differences in impulse responses between NLOS and LOS signals. LSTM, with its memory gating mechanism, effectively preserves long-term dependency information, preventing information from fading over time. The model uses a bidirectional LSTM, meaning it captures dependencies not only forward but also backward, enhancing the understanding of global signal features. CNN further extracts local features from the sequence; after the LSTM layer, the model processes the LSTM output using a one-dimensional convolutional neural network (CNN). CNN, through sliding convolutions over time, captures local patterns and features in the channel impulse response sequence. For communication signals, CNN can identify patterns in specific time segments caused by reflection and scattering, helping the model better distinguish between NLOS and LOS scenarios. The model employs three convolutional layers, with the kernel size and number adjusted using a sandcat swarm optimization algorithm. After each convolutional layer, a ReLU activation function is applied to increase the network's non-linear representation capability. Through layer-by-layer convolution, the network extracts higher-level features and gradually reduces information over time, refining a more compact feature representation. Combining these two approaches can effectively improve model performance, especially in complex NLOS environments, such as... Figure 2 The processing flow of the signal input device.

[0012] In some embodiments of this invention, after the CNN extracts local features, the model performs pooling on the output of the convolutional layer. Global pooling is used here, which averages the features over time. This operation aggregates the local features of each time series into a global feature vector. The advantage of global pooling is that it eliminates the influence of the input sequence length on the model, allowing CIR sequences of different lengths to be uniformly represented as a fixed-length feature vector.

[0013] In some embodiments of this invention, the globally pooled feature vector is input into a fully connected layer, a core component of this binary classification task. The fully connected layer maps the high-dimensional feature vector to the output space, outputting a scalar representing the probability that a sample belongs to NLOS or LOS. To convert the output into a classification label, a sigmoid activation function is used, restricting the network output to between 0 and 1, representing the classification probability value. The final classification result is determined by a threshold: samples with a probability value greater than 0.5 are classified as NLOS, and those less than 0.5 are classified as LOS. This classification layer design is simple yet effective, enabling accurate binary classification prediction based on features extracted from preceding networks.

[0014] In some embodiments of the present invention, a Dropout layer is introduced into the model to avoid overfitting during training. Dropout forces the model to learn broader and more robust features by randomly discarding a portion of the outputs of neurons. This regularization strategy can effectively improve the model's performance on test data, ensuring that the model not only achieves high accuracy on training data but also maintains good generalization ability when faced with unseen data.

[0015] In some embodiments of this invention, the entire network is trained using the binary cross-entropy loss function (BCELoss), a common loss function for binary classification tasks, suitable for measuring the difference between the model output and the true label. During training, the model calculates gradients through backpropagation and uses the Adam optimizer to continuously adjust the model's weight parameters, gradually improving classification accuracy. After each training cycle, the model's performance on the test set is evaluated, and hyperparameters are tuned based on the test set results to further enhance the network's classification ability.

[0016] In this invention, the size and number of convolutional kernels in each of the three convolutional layers of the device model can be adjusted using a sand cat swarm optimization algorithm. This algorithm successfully optimizes the hyperparameters of the convolutional layers by simulating the hunting behavior of a sand cat swarm. The performance of a convolutional neural network largely depends on the configuration of key parameters such as the size and number of convolutional kernels, which directly affect the model's feature extraction ability and final classification results. To avoid the uncertainty and inefficiency of manually adjusting these hyperparameters, the sand cat swarm optimization algorithm plays a crucial role in searching for the optimal combination of hyperparameters.

[0017] In some embodiments of this invention, the sand cat swarm optimization algorithm updates the parameters of each individual sand cat based on the model's performance on the validation set after each iteration, i.e., evaluation metrics such as classification accuracy and F1 score. When a sand cat finds a convolutional layer parameter combination that outperforms other individuals, the algorithm guides other sand cats to move closer to that parameter combination. This "tracking" behavior ensures that the swarm gradually converges to the global optimum. Through this dynamic adjustment, the sand cat swarm can not only quickly find efficient parameter combinations but also avoid the problem of getting trapped in local optima in traditional optimization methods.

[0018] Beneficial effects of the present invention:

[0019] This invention proposes an efficient device for solving the classification problem of line-of-sight and non-line-of-sight signals by combining the channel impulse response sequence of communication signals and the deep learning network structure, and optimizes the hyperparameters of the convolutional layers in the network using the sand cat swarm optimization algorithm. Compared with related technologies, this invention achieves significant benefits in terms of positioning accuracy, robustness, and computational efficiency. Furthermore, it solves the problems of low efficiency and difficulty in obtaining the global optimum in traditional manual parameter tuning. Through global search and local fine-tuning simulating sand cat hunting behavior, the sand cat swarm optimization algorithm can find the optimal convolutional layer configuration in a short time, enabling… Figure 3 The network shown performs better in feature extraction. Attached Figure Description

[0020] Figure 1 The graphs show the amplitude curves of the signal impulse response for both line-of-sight and non-line-of-sight signals.

[0021] Figure 2 This is a flowchart illustrating the classification of non-line-of-sight and line-of-sight signals by the device in one embodiment of the present invention.

[0022] Figure 3 This is a structural diagram of a classification device for non-line-of-sight and line-of-sight signals in an embodiment of the present invention. Specific implementation methods

[0023] To more clearly understand the purpose, features, and advantages of this application, a detailed description is now provided in conjunction with the accompanying drawings and specific embodiments. Without causing conflict, the various embodiments and features of this application can be combined with each other. In the following description, numerous specific details are set forth to provide a comprehensive understanding of this application; these descriptions relate only to some embodiments and do not cover all possible implementations of this application.

[0024] Unless otherwise defined, the technical and scientific terms used in this specification should be interpreted in accordance with their meanings as commonly understood by those skilled in the art. The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. Furthermore, the term "and / or" indicates that the listed items may exist individually or in any combination.

[0025] In the following description, examples of the invention will be illustrated with accompanying drawings, and similar figures represent similar components.

[0026] This example provides a line-of-sight and non-line-of-sight signal classification device based on the sand cat swarm optimization algorithm. Figure 1A flowchart illustrating the classification device's implementation of non-line-of-sight propagation is presented. To achieve signal classification, the deep learning model in this invention needs to process the channel impulse response (CIR) sequence of the communication signal. Specifically, the CIR sequence reflects the propagation path information of the signal at different time points. However, since the length of the CIR sequence may vary depending on different communication conditions, this invention first standardizes the CIR sequence. To ensure consistent sequence length, the system uses zero-padding to unify all CIR sequences to the same length. This ensures that training instability or computational errors will not occur during the model's input phase due to inconsistent data lengths.

[0027] In one or more instances of this application, during data processing, the CIR sequences are first standardized into vectors of uniform length, and then input into the model for further feature extraction and classification. To improve processing efficiency, the standardized CIR data is input into the model in batches, and batch processing is used to improve computational speed, i.e., converting the string-formatted CIR sequences into floating-point arrays. The formula for CIR sequence data preprocessing is:

[0028] CIR i =[float(c1),float(c2),...,float(c n )]

[0029] In one or more instances of this application, since the CIR sequence lengths of each sample are different, the function uniformly pads to the maximum length L. The padding value is 0, so that each CIR sequence is either padded or truncated to length L, ensuring a uniform input size. The specific formula is:

[0030]

[0031] Long Short-Term Memory (LSTM) networks are a type of neural network architecture particularly well-suited for processing time-series data, capable of capturing the dependencies and dynamic changes of signals over time. This invention uses LSTM to perform preliminary processing on the input CIR sequence to extract global temporal features from the sequence.

[0032] In LSTM processing of CIR sequences, the input sequence is first fed into the LSTM unit step by step over time. LSTM's memory gating mechanism effectively preserves long-term dependency information, preventing the loss of important information over time. Therefore, LSTM can extract the signal's changing trends at different time steps from the CIR sequence, which is crucial for distinguishing between LOS and NLOS signals. For example, non-line-of-sight signals may exhibit specific delays and attenuations in time due to multipath effects, and LSTM can help with classification by capturing these temporal features.

[0033] In one or more embodiments of this application, to further enhance the feature extraction capability of the model, a bidirectional LSTM (Bi-LSTM) structure is used. Bi-LSTM can process sequence data simultaneously in both forward and backward directions, thereby capturing the dependencies in the sequence more comprehensively. The forward LSTM captures the forward temporal features of the signal, while the backward LSTM captures the reverse temporal information. The combination of the two enables the model to have a deeper understanding of the global temporal features of the CIR sequence, thereby improving the accuracy of classifying NLOS and LOS signals.

[0034] In one or more instances of this application, the CIR sequence is input into an LSTM layer. The LSTM layer extracts temporal features from the sequence data. The input dimension of the LSTM is L, and the output is a hidden state sequence of the same length as the input.

[0035] H = [h1, h2, ..., h L ]

[0036] Each h t It is the hidden state vector at time t. The output of LSTM is still a sequence containing features extracted from the input CIR sequence.

[0037] In one or more instances of this application, while LSTM performs well in extracting global features from time series, its ability to capture local features is relatively weak. To compensate for this deficiency, this invention uses a convolutional neural network (CNN) immediately after LSTM processing to further process the LSTM output in order to extract local patterns and features in the CIR sequence.

[0038] In one or more instances of this application, in the CNN part, the features output by the LSTM are used as input and processed through a one-dimensional convolution operation. The CNN's convolution kernel slides along the time series dimension to extract local patterns in the sequence. For communication signals, local patterns may represent short-term features caused by phenomena such as reflection and diffraction, which are crucial for distinguishing between LOS and NLOS signals. The CNN can identify and extract these patterns, further enhancing the model's ability to classify NLOS and LOS signals.

[0039] Specifically, the hidden state H after passing through the LSTM layer is passed to the convolutional layer. The convolutional layer extracts local features through convolutional kernels. The first convolutional layer:

[0040] C1 = ReLU(W convl *H+b convl )

[0041] Where * denotes a one-dimensional convolution operation, and the kernel size is 3 (i.e., convolution is performed over 3 consecutive time steps). ReLU is the activation function that ensures the output is non-negative.

[0042] Second convolutional layer:

[0043] C2 = ReLU(W conv2 *C1+b conv2 )

[0044] Third convolutional layer:

[0045] C3 = ReLU(W conv3 *C2+b conv3 )

[0046] In one or more instances of this application, a CNN hyperparameter optimization method based on the Sand Cat Group Optimization (SCSO) algorithm is used. This optimization process avoids the inefficiency of manual parameter tuning by automatically adjusting the number of convolutional kernels in each layer of the CNN. The Sand Cat Group Optimization algorithm utilizes the natural behavior of sand cats to dynamically optimize the hyperparameters of the CNN through two stages: exploration and development.

[0047] In this invention, the main parameters of the convolutional layer in a convolutional neural network include the number and size of the convolutional kernels. Assuming a CNN contains three convolutional layers, the parameters of each convolutional layer can be represented by the following vector:

[0048] P i =[p1,p2,p3]

[0049] Where p1, p2, and p3 represent the number of convolutional kernels in the first, second, and third layers, respectively. Each particle P... i This represents a possible convolution kernel configuration, where each dimension of the particle represents the number of convolution kernels in the corresponding layer. The number of convolution kernels is typically searched within a certain range (e.g., [32, 256]).

[0050] The fitness value of each particle is determined by the performance of the CNN model on the validation set. Specifically, the fitness value of a particle can be expressed as:

[0051] f(P i )=F1 score (P i )

[0052] Among them, P i The CNN model is trained based on the convolution kernel configuration represented by the current particle, and the F1 score of the model on the validation set, f(P), is calculated. i ) represents particle P i For classification performance under the current hyperparameter combinations, the F1 score is an important evaluation metric for model performance.

[0053] In this invention, the sand cat swarm optimization algorithm consists of two stages: a prey search stage and a prey attack stage. Each stage uses a different strategy to update the particle positions, thereby adjusting the hyperparameters of the CNN.

[0054] During the search phase, the sand cat expands the search range by introducing randomness, thus avoiding getting trapped in local optima. The sand cat's current position is determined based on the globally optimal particle.

[0055]

[0056] Here, P best This represents the optimal particle position found within the swarm. The random factor r varies within the range [0,1], and the random angle α also helps increase the diversity of the search. In this way, the sand cat can explore in various dimensions to discover new possible convolutional kernel configurations, increasing the probability of finding the global optimum.

[0057] In this invention, this stage emphasizes extensive search, allowing the sand cat population to discover new regions through diverse search paths, avoiding stagnation on a specific combination of hyperparameters. This random exploration characteristic of the algorithm makes it highly adaptable, capable of handling high-dimensional and complex hyperparameter spaces.

[0058] In one or more instances of this application, during the prey-attacking phase, the sand cat will focus more on the current optimal solution and employ a more targeted strategy for local search. The goal of this phase is to fine-tune the previously found optimal convolutional kernel configuration to more accurately optimize model performance. The formula for position update is:

[0059]

[0060] Here, c is a learning factor, usually set to a constant greater than 1, which can accelerate the convergence speed of particles towards the optimal position. In this stage, the sand cat uses the known optimal solution to perform a more detailed local search, further improving the classification performance of the CNN by slightly adjusting the position of the particles.

[0061] In this invention, this stage emphasizes optimizing the current best solution, utilizing existing information to reduce the search space, thereby improving efficiency. This combination of strategies enables the Sand Cat Swarm Optimization algorithm to achieve greater performance improvements in a shorter time.

[0062] In one or more instances of this application, the sand cat swarm optimization algorithm is an iterative process. In each iteration, the positions of all particles are adjusted according to the update rules described above, and the fitness values ​​are recalculated. The entire optimization process continues until a preset termination condition is met. For example, a maximum number of iterations or a predetermined threshold for the fitness value can be set. Through this iterative update mechanism, the sand cat swarm optimization algorithm can gradually converge to the optimal convolutional kernel configuration, improving the performance of CNNs on specific tasks.

[0063] In this invention, after the CNN completes local feature extraction, the model uses a global pooling layer to pool the output of the convolutional layers. The purpose of global pooling is to uniformly represent CIR sequences of different lengths into a single feature vector of fixed length. Specifically, global pooling performs average pooling on the convolutional output, aggregating local features across all time steps into a single global feature vector. This approach allows for the overall integration of local features while avoiding the impact of inconsistent input sequence lengths on the model.

[0064] In this invention, after global pooling, the local features of the CIR sequence are aggregated into a global feature vector. This feature vector is then used as input to a fully connected layer for classification. In this implementation, the fully connected layer maps the high-dimensional feature vector to a scalar, which represents the probability that the sample belongs to LOS or NLOS.

[0065] To convert the probability values ​​output by the network into explicit classification labels, this invention uses the Sigmoid activation function. The Sigmoid function restricts the network output values ​​to between 0 and 1. The closer the output value is to 1, the greater the probability that the signal belongs to NLOS; the closer the output value is to 0, the more likely the signal is to be LOS. Finally, by setting a threshold (usually 0.5), the output results are converted into binary classification labels: samples with a value greater than 0.5 are classified as NLOS, and samples with a value less than 0.5 are classified as LOS.

[0066] In one embodiment, by utilizing the joint processing of LSTM and CNN described above, the model can fully leverage the temporal and local features in the CIR sequence, thereby achieving efficient classification of LOS and NLOS signals. This device can effectively classify NLOS and LOS propagation in complex indoor environments, thereby enabling the identification of NLOS propagation, which is of great significance for improving the accuracy of indoor positioning.

[0067] This application also provides a computer-readable storage medium storing computer-readable instructions, which a processor executes to implement the methods described in this application. It is worth noting that the embodiments of this application are not limited to specific implementation methods. The division of modules is a logical functional division, and in actual implementation, there can be various forms. Modules can be physically separate or integrated. Furthermore, the term "comprising" in this application does not exclude other units or steps, and the singular form does not exclude the plural; multiple units can be implemented by a single unit through software or hardware. All embodiments are illustrative only and do not limit the technical solutions of this application. Those skilled in the art can modify or make equivalent substitutions without departing from the technical spirit and scope of this application.

Claims

1. A line-of-sight and non-line-of-sight signal classification device based on a sand cat swarm optimization algorithm, characterized in that, include: a) A data preprocessing module is used to receive the channel impulse response (CIR) sequences of communication signals and to standardize these sequences to ensure that all CIR sequences have the same length to meet the input requirements of deep learning models; b) A deep learning model that combines a long short-term memory network (LSTM) and a convolutional neural network (CNN) for feature extraction and classification of standardized CIR sequences; LSTM layers are used to capture global temporal features in CIR sequences; The CNN layer, placed after the LSTM layer, is used to further extract local features from the CIR sequence; the size and number of convolutional kernels in the CNN layer are optimized and adjusted using the Sand Cat Group Optimization (SCSO) algorithm. c) The Sand Cat Swarm Optimization (SCSO) module is used to automatically adjust the number of convolutional kernels in CNN layers to optimize the performance of deep learning models. The SCSO module includes a prey search phase and a prey attack phase, which update the position of particles through different strategies to adjust the hyperparameters of the CNN. d) Global pooling layer, placed after the CNN layer, is used to pool the output of the CNN layer, and represent CIR sequences of different lengths into a fixed-length feature vector; e) Fully connected layer and classification module: receive the output of the global pooling layer, map it to a scalar, representing the probability that the sample belongs to the non-line-of-sight (NLOS) or line-of-sight (LOS) signal, and convert the probability value into a classification label through the Sigmoid activation function and a preset threshold. The Sand Cat Group Optimization (SCSO) module is specifically implemented as follows: Initialize a group of particles, each particle representing a possible CNN convolutional kernel configuration; The CNN model is trained based on the convolution kernel configuration represented by the current particle, and the F1 score of the model on the validation set is calculated as the fitness value of the particle. During the prey-hunting phase, randomness is introduced to expand the search range and avoid getting trapped in local optima; During the prey-attacking phase, the focus is more on the current optimal solution, and more targeted strategies are used for local searches; The particle positions are iteratively updated until a preset termination condition is met, resulting in the optimal convolution kernel configuration.

2. The apparatus according to claim 1, characterized in that, The data preprocessing module standardizes all CIR sequences to the same length by padding with zeros.

3. The apparatus according to claim 1, characterized in that, The LSTM layer is a bidirectional LSTM layer, used to process CIR sequences simultaneously from both forward and backward directions, in order to capture the dependencies in the sequence more comprehensively.

4. The apparatus according to claim 1, characterized in that, The CNN layer consists of three convolutional layers, with a ReLU activation function applied after each convolutional layer to increase the network's non-linear representation capability.

5. The apparatus according to claim 1, characterized in that, The global pooling layer employs average pooling to average the output of the convolutional layer over time, thereby eliminating the influence of the input sequence length on the model.

6. The apparatus according to claim 1, characterized in that, The fully connected layer maps the high-dimensional feature vector to a scalar and restricts the output to between 0 and 1 through the Sigmoid activation function, representing the probability value of classification. The final classification result is determined by a threshold, classifying samples with a probability value greater than 0.5 as NLOS and those with a probability value less than 0.5 as LOS.

7. The apparatus according to claim 1, characterized in that, The SCSO module introduces random factors and random angles during the prey search phase to increase the diversity of the search; during the prey attack phase, a learning factor is used to accelerate the convergence speed of particles toward the optimal position.

8. The apparatus according to any one of claims 1 to 7, characterized in that, It also includes a computer-readable storage medium storing computer-readable instructions that, when executed by a processor, implement the above-described line-of-sight and non-line-of-sight signal classification method based on the sand cat swarm optimization algorithm.

Images