Indoor positioning method fusing stacked auto-encoder and residual network

Abstract

The invention discloses an indoor positioning method fusing a stacked auto-encoder and a residual network, and aims to solve the problems of low positioning precision, insufficient feature extraction and difficult model optimization in a complex multi-floor indoor environment in the prior art. The method comprises the following steps: firstly, normalizing original indoor signal data through data preprocessing; afterwards, an improved SAE is adopted to carry out feature dimension reduction and preliminary feature extraction, the SAE integrates a convolution branch to enhance the local feature learning ability, and important information is reserved through residual connection; next, low-dimensional features output by the SAE are sent to an improved Resnet for deep feature learning and position regression, the Resnet adopts one-dimensional convolution to adapt to signal features, the gradient disappearance problem is effectively relieved through an adaptive residual block, and a global average pooling layer is used at the tail end to aggregate the features; and finally, an improved genetic algorithm is adopted to carry out intelligent optimization on the key hyper-parameters of the ResNet, and the GA improves the optimization efficiency and effect through a non-uniform variation and elitism strategy. Through cooperative work of all the modules, high-dimensional RSSI data can be effectively processed, a deep network can be stably trained, model parameters can be intelligently optimized, and the positioning precision and robustness in a complex indoor environment are remarkably improved.

Description

[0001] The present invention relates to the field of indoor positioning technology, and specifically to an indoor positioning method that integrates stacked autoencoders and residual networks. The method improves the accuracy and robustness of positioning through the integration of deep learning models and intelligent optimization algorithms. Background Art

[0002] With the rapid development of wireless communications and the Internet of Things (IoT), indoor positioning technology has attracted widespread attention due to its enormous potential for various location-based services. While the Global Positioning System (GPS) performs well in outdoor environments, its accuracy and reliability are significantly compromised indoors due to poor signal penetration, occlusion, and reflection. Therefore, research into high-precision indoor positioning technology is of great significance.

[0003] Existing indoor positioning technologies vary, including those based on Wi-Fi, Bluetooth, RFID, and ultrasound. Among them, Wi-Fi fingerprint-based positioning methods have been widely studied due to their advantages of requiring no additional hardware deployment and relatively low cost. Traditional Wi-Fi fingerprint positioning methods typically use machine learning algorithms such as K-nearest neighbor (KNN) and support vector machines (SVM). However, these traditional machine learning algorithms struggle to fully extract the deep and complex features in the signal when dealing with complex and changing indoor environments, resulting in limited positioning accuracy.

[0004] In recent years, deep learning, with its powerful feature extraction and nonlinear fitting capabilities, has demonstrated tremendous potential in the field of indoor positioning. Researchers have attempted to improve positioning performance using models such as deep neural networks (DNNs), convolutional neural networks (CNNs), recurrent neural networks (RNNs), and stacked autoencoders (SAEs). While these methods have achieved some progress, they still face challenges, such as the impact of RSSI signal volatility, the vanishing or exploding gradients caused by high-dimensional data, and the significant impact of model hyperparameter selection on performance. Further improving positioning accuracy and robustness, particularly in multi-story, complex indoor environments, remains a pressing technical challenge. To address these issues, this paper proposes an indoor positioning algorithm that integrates a stacked autoencoder with a residual network. Summary of the Invention

[0005] The main purpose of the present invention is to overcome the shortcomings of existing indoor positioning technology and provide an indoor positioning method that integrates an improved stacked autoencoder, an improved residual neural network and an improved genetic algorithm to significantly improve the positioning accuracy and robustness in complex multi-story environments.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows: Step 1: Normalize the input raw indoor environmental signal data, including uniformly representing missing or undetected signal values ​​and performing minimum-maximum normalization on signal strength values, mapping them to a unified numerical range to facilitate subsequent processing by the neural network model; Step 2: Feature dimension reduction and preliminary feature extraction based on the improved stacked autoencoder, which specifically includes the following sub-steps: (1) By stacking multiple autoencoders, the nonlinear feature transformation and compression of input data from high dimension to low dimension can be realized layer by layer; (2) A convolution branch consisting of multiple layers of one-dimensional convolution (Conv1D) operations, batch normalization (Batch Norm), and ELU activation function is designed and integrated in the encoder part of the SAE. This convolution branch is used to extract local correlation features and patterns from the input signal sequence; (3) By introducing the residual connection mechanism, part of the original feature information or shallow feature information learned during the SAE encoding process is directly transferred to the subsequent layers to reduce information loss and enhance feature expression; Step 3: The low-dimensional feature vector output by the SAE module is fed into an improved residual neural network as input. This includes the following sub-steps: (1) All convolution operations in the network use one-dimensional convolution (Conv1D) to adapt to the one-dimensional sequence characteristics of the input feature vector; (2) The first convolutional layer of the network uses a one-dimensional convolution operation with a large kernel size and a specific step size to effectively capture the initial global pattern of the input features; (3) The skip connection of each residual block can automatically decide whether to adjust the dimension of the features on the skip connection through a 1×1 one-dimensional convolution layer according to whether the dimension of the input feature map matches the dimension of the output feature map of the main path (convolutional layer sequence), thereby ensuring the effectiveness of residual learning and smooth propagation of gradients; (4) Global feature aggregation: After all convolution and residual blocks, a global average pooling layer (GAP) is used to aggregate the final extracted deep feature map and convert the high-dimensional feature map into a one-dimensional feature vector; (5) Position coordinate output: the feature vector output by the GAP layer is nonlinearly mapped through one or more fully connected layers, and finally the predicted indoor position coordinates are regressed; Step 4: To achieve the best performance of the ResNet model, an improved genetic algorithm is used to automatically optimize its key hyperparameters (including the number of training rounds, batch size, and the number of neurons in the fully connected layer). The specific steps include the following: (1) In the mutation operation of the genetic algorithm, a non-uniform mutation mechanism is adopted so that the intensity of the mutation can be dynamically adjusted according to the evolutionary stage, balancing global exploration and local fine search; (2) In each generation of evolution, several individuals (elites) with the highest fitness in the current population are not involved in crossover and mutation and are directly retained in the next generation of population to prevent the loss of the optimal solution and accelerate convergence; (3) Fitness function for positioning accuracy: The fitness function of the genetic algorithm is designed directly based on the mean positioning error (MPE) of the model on the validation set. The smaller the MPE, the higher the fitness.

[0007] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: The present invention uses an improved stacked autoencoder, whose integrated convolutional branches can extract more representative local features from the original high-dimensional RSSI data. At the same time, the residual connection helps to retain important information. Compared with the traditional SAE, it effectively reduces the dimension while improving the quality of the initial features. The improved residual neural network of the present invention adopts an adaptive residual block design, which can intelligently adjust the skip connection according to the dimensionality changes of the network layer, effectively alleviating the gradient vanishing problem in deep network training, allowing the network to be built deeper and thus learn more complex positioning features; The present invention uses an improved genetic algorithm to optimize the key hyperparameters of the neural network. Its non-uniform mutation and elite retention strategies improve the search efficiency and the probability of finding the global optimal (or near-global optimal) solution, avoiding the inefficiency and subjectivity of traditional manual parameter adjustment, and fully utilizing the model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is the overall architecture diagram of the present invention; Figure 2 This is a diagram of the improved stacked autoencoder structure of the present invention; Figure 3 This is the improved ResNet34 network structure diagram of the present invention; Figure 4 It is an optimization diagram of the algorithm steps of the present invention. DETAILED DESCRIPTION

[0009] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings. The technical solution of the present invention will be further described below in conjunction with the drawings and embodiments.

[0010] Figure 1 The overall architecture of the indoor positioning model that stacks autoencoders and residual neural networks, reflecting the data processing flow in the model; Figure 2 This is a diagram of the improved stacked autoencoder structure of the present invention, showing its internal structure and connection method; Figure 3 The overall architecture of the improved ResNet34 network model of the present invention shows the characteristics of the improved network structure; Figure 4 The improved genetic algorithm of the present invention optimizes the algorithm steps of the residual neural network and clearly demonstrates the optimization process.

[0011] The specific implementation steps are: Step 1.1: Missing Value Handling: In the original RSSI dataset, some wireless access point (WAP) signals may not be detected by the user device for various reasons. These undetected signals are typically marked with a specific value (for example, in the referenced paper, this value is set to -110 dBm). In this step, all such undetected signal values ​​are uniformly handled to ensure consistency in their numerical representation. Step 1.2 Data Normalization: Due to differences in WAP transmit power, device receive sensitivity, and other factors, the range of raw RSSI values ​​can be quite large. To eliminate the adverse effects of this range on model training (such as uneven gradient updates and slow convergence), a minimum-maximum normalization method is used. Step 2.1 Each AE consists of an encoder and a decoder. The encoder maps the input data to a low-dimensional latent representation, while the decoder attempts to reconstruct the original input from this latent representation. Through unsupervised layer-by-layer training, the SAE can learn an effective low-dimensional representation of the input data, gradually reducing the input dimension from 128 to 64 and then further to 32 dimensions to achieve feature compression. Step 2.2 introduces a convolutional branch. Unlike traditional fully connected SAEs, this SAE innovatively integrates a convolutional branch into its encoder. This convolutional branch specifically processes the input RSSI sequence features, using a multi-layer one-dimensional convolution (Conv1D) operation, consisting of several "Conv1D - Batch Norm - ELU - Max-Pool" combination units. The first convolution layer uses 128 filters, followed by a pooling layer and a 64-filter convolution layer. This convolution operation effectively captures local correlations and spatial patterns from the RSSI sequence, extracting more structured features than a fully connected layer. Step 2.3: To better preserve and transmit valuable feature information in deep networks and prevent information loss during layer-by-layer encoding, SAE introduces residual connections. These connections allow some feature information from shallow layers (even the original input after preliminary transformation) to "skip" several layers and directly merge with features from deeper layers. Step 2.4 In each layer of the SAE, use the exponential linear unit (ELU) as the activation function. ELU can alleviate the vanishing gradient problem, and its negative output helps push the average activation value of neurons towards zero, thereby accelerating the learning process and improving the robustness of the model. Step 3.1 The low-dimensional feature vector output by the SAE module will be used as input and sent to the improved residual neural network module for deeper feature learning and final position coordinate prediction: Step 3.2: Network adaptation for one-dimensional data. Considering that the input feature vector is one-dimensional sequence data, all convolution operations in ResNet use one-dimensional convolution. Step 3.3: After all convolutional layers and residual blocks in ResNet, a GAP layer is applied. The GAP layer calculates the average value of all elements in each feature map output by the last convolutional layer, thereby compressing the high-dimensional feature map into a one-dimensional feature vector. The GAP layer does not contain learnable parameters, which can effectively reduce the total number of model parameters, reduce the risk of overfitting, and enhance the model's robustness to small changes in the spatial position of the input features. Step 3.4 Fully Connected Layer and Position Output: The feature vector output by the GAP layer is ultimately fed into one or more fully connected layers (Dense Layers). These fully connected layers perform the final nonlinear feature transformation and information integration, and regress the predicted indoor position coordinates. The output can be two-dimensional coordinates (x, y) or, in multi-story scenarios, three-dimensional coordinates including floor information. Step 4.1: To enable the ResNet model to run under its optimal parameter configuration and thus achieve the best positioning performance, the hyperparameter combination to be optimized is regarded as an individual in the genetic algorithm. The value of each hyperparameter is usually encoded as an integer, and a certain number of individuals are randomly generated to form the initial population of the GA. Step 4.2: For each individual in the population, use this set of hyperparameters to configure and train the ResNet model. After training, evaluate the model's positioning performance on the reserved validation dataset. Use the mean positioning error (MPE) as a metric to measure model performance and convert it into a fitness value. The smaller the MPE, the higher the model's positioning accuracy and the higher the corresponding individual's fitness. The formula for calculating the mean positioning error is: ; in, , , are the coordinates of the predicted position, , is the coordinate of the real position; is the number of test samples; Step 4.3 Based on the individual fitness value, a selection strategy is adopted to select excellent individuals from the current population as parents to participate in subsequent reproduction. Individuals with high fitness are more likely to be selected; Step 4.4 Perform a crossover operation on the selected parent individuals, exchanging some of their genes with a certain crossover probability to produce new offspring individuals; Step 4.5 Perform non-uniform mutation on the offspring individuals with a certain mutation probability; Step 4.6 To prevent the loss of the discovered optimal solution during the evolution process, an elite retention strategy is adopted. That is, after each generation of evolution, the individuals with the highest fitness in the current population are directly copied to the next generation of population without participating in crossover and mutation operations. Step 4.7 Repeat steps 4.2 and 4.3 until the preset maximum number of evolutionary generations is reached, or if the fitness of the optimal individual does not significantly improve after several consecutive generations, the algorithm is considered to have converged and the optimization process terminates. At this point, the hyperparameter combination corresponding to the individual with the highest fitness in the population is the optimal hyperparameter configuration found by the GA.

Claims

1. An indoor positioning method integrating stacked autoencoders and residual networks, characterized in that: The following steps are involved: Step 1: Data preprocessing: normalize the input raw indoor environment signal data (such as Wi-Fi RSSI data) and unify the undetected signal values. Step 2: Construct and apply an improved stacked autoencoder (SAE) to perform feature dimensionality reduction and preliminary feature extraction on the preprocessed data. The improved stacked autoencoder includes convolution branches to implement multi-layer convolution operations, retains the original feature information through residual connections, and uses the ELU activation function; Step 3: The dimensionality-reduced features output by the stacked autoencoders are fed into an improved residual neural network (ResNet) for deep feature extraction. The improved residual neural network uses a one-dimensional convolutional layer (Conv1D) to extract features, includes an adaptive residual block structure to automatically adjust skip connections based on the input and output dimensions, and uses a global average pooling layer at the end of the network to aggregate features. Finally, the position information is output through multiple fully connected layers. Step 4: An improved genetic algorithm (GA) is used to optimize the hyperparameters of the residual neural network, including the number of training rounds, batch size, and the number of neurons in the fully connected layer. The improved genetic algorithm adopts a non-uniform mutation strategy and an elite retention strategy, and uses a fitness function based on the mean positioning error (MPE).

2. The method according to claim 1, characterized in that The specific process in step 1 is as follows: Step 1.1 Map the original RSSI data to within the interval; Step 1.2 For the WAP signals that are not detected in the data (for example, expressed as -110dBm), a unified numerical representation is performed before normalization.

3. The method according to claim 1, characterized in that The specific process in step 2 is as follows: Step 2.1 The improved stacked autoencoder designs a convolution branch to implement multi-layer convolution operations to further extract deep features; Step 2.2: The original feature information of the SAE encoder is retained through residual connections, and the ELU activation function is used to alleviate the gradient vanishing problem and enhance the nonlinear expression ability of the network for layer-by-layer feature extraction and screening.

4. The method according to claim 1, wherein The specific process in step 3 is as follows: Step 3.1: The improved residual neural network replaces the Conv2D layer with the Conv1D layer. The initial feature extraction is achieved through a large-scale convolution layer. The convolution layer uses 64 filters, a convolution kernel size of 7, and a stride of 2 to effectively capture the initial features of the signal. Step 3.2 Based on this, an adaptive residual block structure is proposed, which can automatically adjust the skip connection according to the input and output dimensions. When the stride is not 1 or the number of input channels does not match the filter, a 1×1 convolution is automatically added to adjust the dimension of the skip connection; Step 3.3 uses an asymmetric residual network structure. The model contains four stages, and the number of filters is gradually increased in each stage, allowing the network to extract features at different depths and scales. Step 3.4 In the improved ResNet34 network, global average pooling is finally used to reduce the dimension of the features, which greatly reduces the number of parameters. The features output by global average pooling are then fused with multiple layers of fully connected layers to further extract features, allowing the model to perform more complex nonlinear transformations after extracting global features.

5. The method according to claim 1, wherein The non-uniform mutation strategy and elite retention strategy of the improved genetic algorithm in step 4, the mutation operation is expressed as: ； in, For the parameters, is the variation value, which obeys the uniform distribution. is the mutation probability, For interval Random numbers sampled from a uniform distribution; The elite retention strategy is expressed as: ; in, is the population size, for elite proportions; This is a rounding down operation; The fitness function of the genetic algorithm is based on the mean positioning error (MPE), which is calculated as: ； in, , , are the coordinates of the predicted position, , is the coordinate of the real position; is the number of test samples.

Images