Indoor integrated positioning system based on distributed extended Kalman filtering deep learning optimization algorithm

Abstract

The invention relates to an indoor integrated positioning system based on a distributed extended Kalman filtering deep learning optimization algorithm, and belongs to the field of indoor positioning. According to the system, a neural network is introduced, a deep learning method CNN-LSTM is applied to UWB NLOS / LOS signal classification, and then distributed extended Kalman filtering DEKF is introduced to fuse data, optimized by the CNN-LSTM, of the UWB with data of an inertial measurement unit IMU, so that the UWB NLOS / LOS signal classification is realized. Accumulative errors caused by time drift of the IMU are inhibited, so that the problems of accumulative errors and time synchronization in data fusion of the sensors are avoided while the ranging data precision is ensured. The deep learning optimization algorithm CNN-LSTM-DEKF based on distributed extended Kalman filtering provided by the invention can automatically adapt to different sensor characteristics and environment changes, and provides a more accurate and robust positioning result in a complex indoor environment. Experimental results show that in an indoor environment with a remarkable multipath effect and signal shielding gaps, the method disclosed by the invention obviously improves the positioning precision.

Description

Technical Field

[0001] The present invention belongs to the field of indoor positioning, and specifically relates to an indoor comprehensive positioning system based on a distributed extended Kalman filter deep learning optimization algorithm. Background Art

[0002] With the development of the Internet of Things (IoT) and smart spaces, indoor positioning has become an urgent need in industrial production, healthcare, connected vehicle logistics, and smart homes. The challenges faced by indoor positioning are far more complex than those in outdoor environments, including signal multipath effects, non-line-of-sight (NLOS) propagation, and environmental interference [1]. Traditional single positioning technologies, such as ultra-wideband (UWB) [2], inertial measurement unit (IMU) [3], Wi-Fi [4], and Bluetooth (BLE) [5], cannot provide sufficiently accurate and stable positioning effects in complex indoor environments. Therefore, multi-sensor fusion has gradually become an effective method to achieve indoor positioning, which compensates for the shortcomings of single technology by fusing multiple sensor data. The UWB / IMU combined positioning method has attracted widespread attention due to its robustness and high accuracy in indoor environments. The combination of UWB and IMU has obvious advantages in multi-sensor fusion. In ultra-wideband (UWB) positioning systems, commonly used ranging methods include time of arrival (TOA), time difference of arrival (TDOA), angle of arrival (AOA), etc. [8], which can provide centimeter-level positioning accuracy. However, due to the non-line-of-sight (NLoS) ( Figure 1) and multipath interference, large distance positioning errors will appear in UWB sensor data from time to time, and its ranging accuracy may drop significantly, affecting the positioning accuracy of the system. In recent years, with the rapid development of deep neural networks, multi-layer neural networks in deep learning have been widely used in the identification of LOS and NLOS propagation conditions. Liu et al. proposed a method of using reversible transform denoising combined with convolutional neural network (CNN) to identify NLOS signals, with a recognition accuracy of up to 81.68% [9]. Fang et al. used the original CIR as input, trained and tested the long short-term memory (LSTM) model, and achieved a recognition accuracy of up to 82.14%

[10] . Inertial measurement unit (IMU) has been widely studied and applied because of its low cost and immunity to environmental influence. IMU uses three-axis gyroscope and three-axis accelerometer as basic sensitive elements to provide real-time angular velocity and acceleration measurement, thereby obtaining the motion characteristics and attitude information of the carrier (the attitude, velocity and position information of the target are calculated by integrating the measurement data). However, due to the long-term drift of inertial sensors, the errors obtained by calculating the navigation state (position, velocity and attitude) will accumulate over time, and therefore cannot provide reliable long-term positioning information. UWB provides distance measurement of the position during positioning, is not easily affected by the indoor environment, and is suitable for solving the precise positioning of most static scenes; while IMU can provide movement direction and acceleration information, and is suitable for real-time tracking of the user's movement trajectory. The complementarity of these two sensors makes the positioning fusion solution based on UWB and IMU an ideal solution in complex environments. On this basis, [6] proposed a UWB / IMU combined positioning algorithm based on machine learning, combining the separate UWB and IMU to give full play to their respective advantages and improve the system positioning accuracy. [7] proposed a fusion framework based on a graph optimization model that tightly integrates IMU with multiple UWB tags, which can provide the system with high-precision position and attitude information. Nevertheless, the fusion of UWB and IMU also faces some challenges. IMU is easily affected by cumulative accuracy (such as bias and offset), which eventually leads to accumulation of positioning accuracy over time, while the data accuracy of UWB is affected by non-line-of-sight (NLoS) and multipath effects. To address these issues, Kalman Filter (KF) and its extended and optimized version Distributed Extended Kalman Filter (DEKF) have been widely evaluated in multi-sensor fusion, especially in dynamic positioning systems. Kalman filter is a cyclic algorithm that gradually updates the state estimate of the system by combining measurement data and prediction models, thereby reducing the impact of measurement. Shi Jiantao et al. proposed a new motion estimation algorithm that can reduce the drift of IMU sensors. The algorithm uses Kalman filter to estimate and fuse UWB measurement data, and fuses the drift-free position obtained by the UWB positioning Kalman filter with the position calculated by the IMU

[11] .However, the performance of KF in nonlinear scenarios is limited, so the Extended Kalman Filter (EKF) was proposed to solve nonlinear problems. Yue Xiaodong et al. used an IMU / UWB loosely combined precise positioning system to solve the robot's precise position, velocity and posture information, in which the positioning coordinates (EKF) from the UWB base station corrected the UWB error in real time

[12] . The distributed extended Kalman filter (DEKF) enhances the noise resistance and robustness of the system by distributing the state estimates of different data sources to multiple nodes. Although DEKF can improve positioning accuracy in theory, in practical applications, the accuracy of data is still susceptible to factors such as environmental fluctuations and sensor ranging accuracy. The performance of the Kalman filter and its extended methods depends on the accuracy of the system model. In practical applications, noise inconsistency and dynamic environmental changes often lead to model deviations, thereby affecting simulation performance. To solve this problem, the adaptive Kalman filter (AKF) has been introduced in recent years to improve auxiliary robustness by adjusting the noise support matrix in real time. Liu et al. proposed an adaptive complementary Kalman filter to filter UWB and IMU data and track the errors of variables such as position, speed, and direction. It can not only eliminate the multipath effect caused by the shielding of UWB signals, but also correct the speed and acceleration deviations caused by IMU drift

[13] . In the above discussion, people have conducted a lot of research on the fusion of different sensors for indoor positioning. The commonly used method is to use the extended Kalman filter and the evolution of nonlinear filtering methods based on it to simplify the nonlinear system into a local linear system, thereby achieving recursive estimation. However, these nonlinear filtering methods do not consider the impact of sensor fusion on the performance of the sensor itself. The performance of sensor fusion is also crucial to improving positioning. Based on the analysis of the above related work, most UWB / IMU fusion methods usually ignore the data synchronization problem in the sensor fusion process. Summary of the invention

[0003] The purpose of the present invention is to overcome the defects of the prior art, further improve the robustness of the multi-sensor fusion positioning system in a complex environment, provide an indoor comprehensive positioning system based on a distributed extended Kalman filter deep learning optimization algorithm, and propose an optimization algorithm CNN-LSTM-DEKF that fuses a convolutional long short-term memory network CNN-LSTM and a distributed extended Kalman filter DEKF. The optimization algorithm ensures the positioning accuracy of the fusion positioning system by performing non-line-of-sight recognition on UWB ranging data and cumulative error suppression on IMU ranging data. At the same time, by dispersing the state estimation of multiple nodes, there will be no problems such as data synchronization between various data sources during the fusion process of sensor data, thereby further improving the robustness and positioning accuracy of the sensor fusion positioning system.

[0004] To achieve the above object, the technical solution of the present invention is: an indoor integrated positioning system based on a distributed extended Kalman filter deep learning optimization algorithm, comprising:

[0005] Ultra-wideband UWB module, collects ultra-wideband UWB ranging data, including ultra-wideband UWB channel impulse response CIR data;

[0006] Inertial measurement unit IMU module, collects IMU ranging data;

[0007] Based on the distributed extended Kalman filter deep learning optimization algorithm module, an optimization algorithm CNN-LSTM-DEKF that integrates the convolutional long short-term memory network CNN-LSTM and the distributed extended Kalman filter DEKF is proposed. The CNN-LSTM is used to extract features and classify the CIR data of the ultra-wideband UWB channel impulse response. The classified prediction data is then fused with the inertial measurement unit IMU ranging data through DEKF to suppress the cumulative error caused by the drift of the IMU over time, thereby ensuring the accuracy of the ranging data while also ensuring that the fusion of each sensor data will not have cumulative errors and time synchronization problems.

[0008] In one embodiment of the present invention, the network framework of CNN-LSTM includes two basic components: a convolutional neural network CNN and a long short-term memory LSTM. In CNN-LSTM, CIR data is used as an input vector of CNN. CNN uses the non-temporal structure of the input signal to extract features by learning the frequency characteristics of CIR data. The output of CNN is then input into LSTM to learn the dependency of time series data for classification and prediction of non-line-of-sight NLOS signals and line-of-sight LOS signals.

[0009] In one embodiment of the present invention, the CNN-LSTM is composed of a convolutional layer, a maximum pooling layer, a fully connected layer, an LSTM layer and a fully connected layer, and the generalization ability is improved through Dropout regularization.

[0010] In one embodiment of the present invention, CNN-LSTM performs feature extraction and classification prediction on the channel impulse response CIR data of ultra-wideband UWB, which specifically includes the following stages:

[0011] CNN components:

[0012] CNN processes the raw CIR data and extracts spatial features that capture the signal pattern changes caused by NLOS conditions. Multiple convolutional layers are used, and each convolutional layer uses the same activation function to detect the input data for features. To ensure that the activation function is nonlinear, ReLU is used, and its equation is as follows:

[0013]

[0014] x represents the parameters of the ReLU function;

[0015] After the convolutional layer outputs to the pooling layer, the pooling layer reduces the dimension and prevents overfitting by downsampling; then, the pooling layer output is input to the fully connected layer, and the fully connected layer selects the classifier as the output through the SoftMax function;

[0016] LSTM components:

[0017] The output of the CNN is passed to the LSTM to capture the temporal dependencies in the CIR data sequence. The gating architecture of the LSTM enables it to model long-term relationships in the data.

[0018] After the data extracted by CNN features is input into LSTM, LSTM is trained using a labeled dataset with a cross entropy loss function to ensure that the UWB signal is accurately classified into LOS or NLOS.

[0019] In one embodiment of the present invention, the gating architecture of LSTM includes input, forgetting and output gates.

[0020] In one embodiment of the present invention, the input gate and output gate functions in the LSTM gating architecture are described as follows:

[0021] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0022]

[0023] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0024] h t =o t tanh(C t )

[0025] Where i t is the input gate, which determines whether the spatial features extracted by CNN, i.e., the classification-related information including NLOS and LOS, are used to update the time series state; h t-1 represents the hidden state of the previous time step, which contains the dynamic information of the feature sequence of the past time step; x t Represents the feature vector of the CNN output at the current time step, including the UWB CIR spatial features of the current time step; W i, b i represents the weight and bias parameters of the input gate, W o , b o Represents the weight and bias parameters of the output gate, which are optimized by gradient descent during training to adapt to the influence of different input features; is a candidate state, which is used to generate a new candidate state for the current time step. It captures the potential features of the CNN features of the current time step and the historical hidden state, and retains important information related to LOS / NLOS. Tanh represents the hyperbolic tangent activation function, which compresses the candidate state to the range of (-1,1) to represent possible state updates. W C , b C The weights and bias parameters representing the candidate states can effectively characterize the nonlinear relationship of the time series after optimization; o t is the output gate, which extracts the most relevant information for LOS / NLOS classification from the candidate states and outputs it to h t ;h t Represents the hidden state of the current time step, which is the output of LSTM and one of the inputs of the next time step; tanh(C t ) means to perform a nonlinear transformation on the current time series state so that its range remains in (-1,1).

[0026] In one embodiment of the present invention, the DEKF calculation process is as follows:

[0027] Step 1: Local prediction stage

[0028] Each node i runs the prediction step independently, updating its state estimate based on the local model:

[0029]

[0030] is the predicted state estimate of node i at the current time k compared to the previous time k-1; represents the predicted value of the state of node i at time k-1; u k-1,i represents the control input of node i at time k-1; P k∣k-1,i is the prediction error covariance matrix of node i compared to the previous moment k-1 and the current moment k, F k,i represents the state transition function of node i at time k; P k-1∣k-1,i represents the prediction error covariance matrix of node i at time k-1; represents the state transfer matrix obtained by derivation of the state transfer function; Q k,i Represents the system noise covariance matrix, describing the uncertainty of state transition;

[0031] Step 2: Local update phase

[0032] Each node updates its local state based on the observed value:

[0033]

[0034] P k∣k,i =(IK k,i H k,i ) k∣k-1,i

[0035] K k,i is the Kalman gain of node i, which is used to weigh the influence of prediction and observation; H k,i The Jacobian matrix of the observation function; z k,i represents the measurement value of node i; h i Represents a nonlinear observation function, describing the mapping relationship from state to observation value; R k,i represents the observation noise covariance, which describes the uncertainty of sensor measurement.

[0036] Step 3: Information sharing stage: each node shares local state estimation and covariance information through communication to form a global information set. The shared content includes P k∣k,i , K k,i ;

[0037] Step 4: Global fusion stage: obtain the optimal state estimate of the system as a whole through global information fusion, and perform weighted summation of the estimates of each node according to the covariance matrix weights through the weighted average method:

[0038]

[0039] represents the optimal state estimation after global fusion, W k,i represents the weight of the i-th node, represents the local state estimate of the i-th node;

[0040] Step 5: In the global update feedback phase, the fusion results are fed back to each node and the local state is adjusted to maintain global consistency.

[0041] In one embodiment of the present invention, DEKF is verified by MAE and RMSE.

[0042] In one embodiment of the present invention, the ultra-wideband (UWB) module collects ranging data of the anchor point through an ultra-wideband (UWB) tag.

[0043] In one embodiment of the present invention, the inertial measurement unit IMU module collects ranging data through a three-axis acceleration sensor.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. The present invention proposes an optimization algorithm integrating CNN-LSTM and DEKF. The optimization algorithm extracts features and classifies UWB CIR data through CNN-LSTM, and then fuses it with IMU data through DEKF, thereby reducing the influence of non-line-of-sight and multipath effects on UWB and suppressing the IMU time accumulation error, thus significantly improving the accuracy of indoor positioning.

[0046] 2. Compared with previous studies, the CNN-LSTM-DEKF algorithm proposed in this paper not only takes into account the ranging error caused by the sensor being affected by the complex environment, but also takes into account the error accumulation and data synchronization between various data sources in the sensor fusion process, effectively improving the noise resistance and robustness of the sensor fusion positioning system. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 An example of NLoS and LoS propagation in UWB IPS.

[0048] Figure 2 It is the system structure of CNN-LSTM-DEKF optimization algorithm based on UWB and IMU fusion positioning system.

[0049] Figure 3 Three deep learning architectures for UWB channel identification, where (a) is CNN, (b) is LSTM, and (c) is CNN-LSTM.

[0050] Figure 4 Confusion matrix of the classification performance for various line-of-sight obstacles (Sce.1 to Sce.5 correspond to table, human body, glass, door and wall, respectively, among which (a) is DT, (b) is SVM, (c) is MLP, (d) is CNN, (e) is LSTM, and (f) is CNN-LSTM.

[0051] Figure 5 Figure 2. Motion trajectory diagram for the true path and the original data with 4 different methods.

[0052] Figure 6 Error box plots of five different optimization algorithms, where (a) is the root mean square error and (b) is the absolute mean error. DETAILED DESCRIPTION

[0053] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0054] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0055] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0056] The present invention provides an indoor integrated positioning system based on a distributed extended Kalman filter deep learning optimization algorithm, comprising:

[0057] Ultra-wideband UWB module, collects ultra-wideband UWB ranging data, including ultra-wideband UWB channel impulse response CIR data;

[0058] Inertial measurement unit IMU module, collects IMU ranging data;

[0059] Based on the distributed extended Kalman filter deep learning optimization algorithm module, an optimization algorithm CNN-LSTM-DEKF that integrates the convolutional long short-term memory network CNN-LSTM and the distributed extended Kalman filter DEKF is proposed. The CNN-LSTM is used to extract features and classify the CIR data of the ultra-wideband UWB channel impulse response. The classified prediction data is then fused with the inertial measurement unit IMU ranging data through DEKF to suppress the cumulative error caused by the drift of the IMU over time, thereby ensuring the accuracy of the ranging data while also ensuring that the fusion of each sensor data will not have cumulative errors and time synchronization problems.

[0060] The following is the specific implementation process of the present invention.

[0061] like Figure 2 As shown, this embodiment provides an indoor integrated positioning system based on a distributed extended Kalman filter deep learning optimization algorithm, including:

[0062] The ultra-wideband UWB module collects ultra-wideband UWB ranging data, including the channel impulse response CIR data of the ultra-wideband UWB; Figure 2 As shown, in this example, the ultra-wideband UWB module collects ranging data of anchor points 1, 2, 3, and 4 based on the ultra-wideband tag;

[0063] Inertial measurement unit IMU module, collects IMU ranging data; such as Figure 2 As shown, in this example, the inertial measurement unit IMU module collects ranging data through a three-axis acceleration sensor;

[0064] Based on the distributed extended Kalman filter deep learning optimization algorithm module, an optimization algorithm CNN-LSTM-DEKF that integrates convolutional long short-term memory network CNN-LSTM and distributed extended Kalman filter DEKF is proposed. CNN-LSTM is used to extract features and classify and predict the CIR data of ultra-wideband UWB channel impulse response (such as Figure 2 As shown in the figure, CNN-LSTM performs UWB NLOS / LOS signal classification based on the channel impulse response CIR data of ultra-wideband UWB), and then fuses the classified predicted data with the inertial measurement unit IMU ranging data through DEKF to suppress the cumulative error caused by IMU drift over time, thereby ensuring the accuracy of ranging data while also ensuring that the fusion of each sensor data will not have cumulative errors and time synchronization problems, and ultimately achieve indoor positioning; the proposed CNN-LSTM-DEKF algorithm can automatically adapt to different sensor characteristics and environmental changes, and provide more accurate and robust positioning results in complex indoor environments. Experimental results show that this method significantly improves positioning accuracy in indoor environments with significant multipath effects and signal shielding gaps.

[0065] 1. CNN-LSTM optimization model

[0066] Figure 3 There are three deep learning architectures used for UWB channel identification. The optimization algorithm framework used in this invention is as follows Figure 3 As shown in (c), it contains two basic components, CNN and LSTM. In CNN-LSTM, the original data of CIR is used as the input vector of CNN. CNN uses the non-temporal structure of the input signal to extract features by learning the frequency characteristics of CIR. The output of CNN is then input into LSTM to learn the dependency of time series data for NLOS and LOS signal classification and prediction.

[0067] Figure 3 The operation process of the CNN-LSTM model framework in (c) for processing UWB CIR data includes the following stages:

[0068] CNN components:

[0069] A convolutional neural network (CNN) processes the raw UWB CIR data and extracts spatial features that capture the signal pattern changes caused by NLOS conditions. Multiple convolutional layers are used, each of which uses the same activation function to detect features in the input data. To ensure that the activation function is nonlinear, ReLU (rectified linear unit)

[16] is used, and its equation is as follows:

[0070]

[0071] Here x represents the parameters of the ReLU function.

[0072] After the convolutional layer outputs to the pooling layer, the pooling layer reduces the dimension and prevents overfitting by downsampling. Then, the pooling layer output is input to the fully connected layer, and the fully connected layer selects the classifier as the output through the SoftMax function.

[0073] LSTM components:

[0074] The output of the CNN is passed to a long short-term memory (LSTM) network, which captures the temporal dependencies in the CIR sequence. The gating architecture of the LSTM (input, forget, and output gates) enables it to model long-term relationships in the data. The “input gate” and “output gate” functions

[17]

[18] are described as follows:

[0075] i t =σ(W i ·[h t-1 ,x t ]+b i ) (3-2)

[0076]

[0077] o t =σ(W o ·[h t-1 ,x t ]+b o ) (3-4)

[0078] h t =o t tanh(C t ) (3-5)

[0079] Where i t is the input gate, which determines whether the spatial features extracted by CNN, i.e., the classification-related information including NLOS and LOS, are used to update the time series state; h t-1 represents the hidden state of the previous time step, which contains the dynamic information of the feature sequence of the past time step; x t Represents the feature vector of the CNN output at the current time step, including the UWB CIR spatial features of the current time step; W i , b i represents the weight and bias parameters of the input gate, W o , b o Represents the weight and bias parameters of the output gate, which are optimized by gradient descent during training to adapt to the influence of different input features; is a candidate state, which is used to generate a new candidate state for the current time step. It captures the potential features of the CNN features of the current time step and the historical hidden state, and retains important information related to LOS / NLOS. Tanh represents the hyperbolic tangent activation function, which compresses the candidate state to the range of (-1,1) to represent possible state updates. W C , b C The weights and bias parameters representing the candidate states can effectively characterize the nonlinear relationship of the time series after optimization; o t is the output gate, which extracts the most relevant information for LOS / NLOS classification from the candidate states and outputs it to h t ;h t Represents the hidden state of the current time step, which is the output of LSTM and one of the inputs of the next time step; tanh(C t ) means to perform a nonlinear transformation on the current time series state so that its range remains in (-1,1).

[0080] After the data extracted by CNN features is input into the LSTM layer, the LSTM layer is trained using a labeled dataset with a cross entropy loss function to ensure that the UWB signal is accurately classified into LOS or NLOS. The optimized CNN-LSTM model significantly reduces the UWB positioning error by improving the signal classification accuracy, providing a robust preprocessing step for subsequent sensor fusion.

[0081] The basic technology used in collecting training samples in this section is to use sensor data for position calculation.

[0082] 2. DEKF optimization based on CNN-LSTM

[0083] EKF is a classic filtering method used to deal with the state estimation problem of linear and nonlinear dynamic systems

[19] . It linearizes the nonlinear system by first-order Taylor expansion, combines the system's prediction model with the sensor measurement value, and recursively calculates the state estimate. DEKF is a distributed version of EKF. It distributes the state estimation task to multiple nodes and optimizes the state estimation through distributed fusion. DEKF can not only handle the fusion problem of multiple sensors, but also significantly improve the robustness and noise resistance of the system. Therefore, DEKF can be used to reasonably correct and set the fusion of trained UWB ranging data and IMU data to avoid the occurrence of ranging outliers. It is verified by MAE and RMSE. Distributed Extended Kalman Filter (DEKF) calculation process.

[0084] Step 1: Local Prediction Phase

[0085] Each node i runs the prediction step independently, updating its state estimate based on the local model:

[0086]

[0087] Here is the predicted state estimate of node i at the current time k compared to the previous time k-1; represents the predicted value of the state of node i at time k-1; u k-1,i represents the control input of node i at time k-1; P k∣k-1,i is the prediction error covariance matrix of node i compared to the previous moment k-1 and the current moment k, F k,i represents the state transfer function (Jacobian matrix) of node i at time k; P k-1∣k-1,i represents the prediction error covariance matrix of node i at time k-1; represents the state transfer matrix (Jacobian matrix) obtained by derivation of the state transfer function; Q k,i Represents the system noise covariance matrix, which describes the uncertainty of state transition.

[0088] Step 2: Local Update Phase

[0089] Each node updates its local state based on the observed value:

[0090]

[0091] P k∣k,i =(IK k,i H k,i ) k∣k-1,i

[0092] Here K k,i is the Kalman gain of node i, which is used to weigh the influence of prediction and observation; H k,i The Jacobian matrix of the observation function; z k,i represents the measurement value of node i; h i Represents a nonlinear observation function, describing the mapping relationship from state to observation value; R k,i represents the observation noise covariance, which describes the uncertainty of sensor measurement.

[0093] Step 3: Information Sharing Phase: Each node shares local state estimation and covariance information through communication to form a global information set. The shared content includes P k∣k,i , K k,i wait.

[0094] Step 4: Global Fusion Phase: Obtain the optimal state estimate of the system as a whole through global information fusion, and perform weighted summation of the estimates of each node according to the covariance matrix weights using the weighted average method:

[0095]

[0096] Here represents the optimal state estimation after global fusion, W k,i represents the weight of the i-th node, represents the local state estimate of the ith node.

[0097] Step 5: Global Feedback Phase: Feedback the fusion results to each node and adjust the local state to maintain global consistency.

[0098] According to the above iterative process, DEKF reduces the computational burden of centralized methods through decentralized computing, while enhancing system robustness. A single node failure will not cause the entire system to fail. Local computing accelerates state updates and improves system response speed. It can adapt to heterogeneous sensor data and is particularly suitable for multi-sensor fusion systems such as UWB / IMU. DEKF can effectively utilize the data fusion of multiple UWB base stations to reduce the impact of NLOS and multipath effects. Through distributed nodes, the drift error of IMU is corrected in real time. DEKF can integrate multi-source data in complex indoor environments to improve positioning accuracy and system robustness. The contribution of DEKF in indoor positioning lies in its powerful fusion capability and distributed architecture, which can significantly improve system performance, especially in scenarios with complex environments and dense node distribution.

[0099] In summary, the CNN-LSTM-DEKF algorithm provides a comprehensive solution for UWB / IMU-based indoor positioning, addressing key challenges in sensor fusion while providing higher accuracy and robustness in complex indoor environments.

[0100] Experiment and data analysis

[0101] 1. Evaluation of non-line-of-sight recognition performance of CNN-LSTM optimization algorithm

[0102] In order to evaluate the performance of the CNN-LSTM algorithm in UWB CIR data signal classification, the present invention conducts comparative experiments through public data sets and benchmark algorithms to evaluate the accuracy of various models in UWB LOS / NLOS signal classification to verify the advantages of the CNN-LSTM algorithm.

[0103] The experiment used a publicly available UWB LOS / NLOS dataset

[20] . Specifically, the dataset covers seven indoor scenes (office 1, office 2, small apartment, small workshop, kitchen with living room, bedroom, and boiler room). The dataset can avoid the deviation between models caused by constructing specific locations. The dataset contains 42,000 samples (3,000 LOS and 3,000 NLOS), each of which contains 1,016 CIR raw data with a resolution of 1 nanosecond. The dataset is divided into a training dataset (33,600), a validation dataset (4,200), and a test dataset (4,200). The 42,000 sample datasets are randomly assigned, and the input vector dimension is (1×1016), which is directly input into the model for training. Five different training models are selected as benchmarks for comparison. The five training models are DT (decision tree), SVM (support vector machine), MLP (multi-layer perceptron network), CNN (convolutional neural network), and LSTM (long short-term memory network). The CNN-LSTM (convolutional long short-term memory network) model used in the present invention consists of a convolutional layer, a maximum pooling layer, a fully connected layer, an LSTM layer and a fully connected layer, and the generalization ability is improved by Dropout regularization. All experiments were processed on a computer equipped with an Intel Core i5-10500H CPU, 16GB RAM and an NVIDIA GeForce RTX 3050Ti Laptop GPU.

[0104] In order to evaluate the classification performance of various methods for various line-of-sight obstacles in complex indoor environments, Figure 4 The confusion matrix shown reflects the recognition and classification performance of various methods in five types of occlusion scenes (Sce.1 to Sce.5 correspond to table, human body, glass, door and wall respectively). Both CNN-LSTM and LSTM classification algorithms can realize NLOS recognition of UWB to a certain extent, but in some scenarios with vague differences in attenuation characteristics, the two algorithms perform poorly and it is difficult to distinguish the recognition scene of human body (Sce.2). Compared with other occlusions, human body occlusion has the greatest interference on NLOS recognition of UWB. The misclassification rate of human body occlusion recognition of the five methods exceeds 10%, and the classification effect is poor. For SVM and CNN algorithms, the errors are more obvious, and the misclassification error rate of Sce.2 exceeds 1 / 3. Overall, the CNN-LSTM algorithm has the greatest impact on NLOS recognition of UWB. Figure 4 (f) It can be seen that it shows good performance in the classification of various line-of-sight obstacles in UWB non-line-of-sight scenarios. Compared with several other algorithms with generally low recognition rates in Sce.2 scenarios, the CNN-LSTM algorithm still achieves a recognition accuracy rate of no less than 89%, and the overall average misclassification rate is only 4.38%.

[0105] The performance of various models in NLoS recognition is tested on public datasets, and the performance comparison for NLOS occluder recognition is shown in Table 1. In terms of algorithm recognition accuracy, precision, and recall, the CNN-LSTM algorithm has a higher advantage than other recognition algorithms.

[0106] Table 1 Performance comparison of various algorithms for NLOS occluder recognition

[0107]

[0108] 2. Evaluation of positioning performance of fusion optimization algorithm CNN-LSTM-DEKF

[0109] In order to evaluate the performance of the CNN-LSTM-DEKF optimization algorithm proposed in the present invention, the positioning effects of the original data and four different optimization algorithms (EKF, DEKF, CNN-LSTM, CNN-STM-DEKF) are compared by performing simulation experiments. In the simulation experiment, it is assumed that the coordinates of the four base stations are (0, 0), (0, 50), (50, 0), (50, 50) m, respectively, and the target to be measured is uniformly accelerated in a two-dimensional plane. The initial position is (20, 20) m, the acceleration is 0.01 m / s, the initial speed is 0.5 m / s, the sampling period is 1s, and the motion path of the target to be measured is a broken line path and a square path. In order to simulate the interference of the external environment on the positioning accuracy of the UWB module, Gaussian white noise is added to the simulated trajectory data, and the mean of the Gaussian white noise is set to 0.06 and the variance is set to 0.1. At the same time, the IMU error model is set in the simulation framework, and the sampling data of the IMU module is integrated into the motion trajectory at a sampling frequency of 100 Hz. The original data under two different real paths and the four motion trajectories after algorithm optimization are shown in the figure. Figure 5 (a) and (b). The green solid line trajectory represents the true path, the red dots represent the unoptimized original data, the blue, pink, and cyan solid lines represent the motion trajectories after EKF, DEKF, and CNN-LSTM optimization, respectively, and the black dotted line trajectory is the motion trajectory after CNN-LSTM-DEKF optimization.

[0110] from Figure 5It can be seen from (a) and (b) that as the target to be measured accelerates, the oscillation amplitude of the four algorithms will increase, resulting in a larger deviation from the actual position. The oscillation of the DEKF and EKF algorithms is more obvious than that of CNN-LSTM and CNN-LSTM-DEKF, indicating that their positioning accuracy is more affected by the errors of IMU and UWB, and the trajectory accuracy of CNN-LSTM and CNN-LSTM-DEKF is higher and closer to the real coordinates. At the same time, DEKF and CNN-LSTM are fused into the CNN-LSTM-DEKF optimization algorithm proposed in the present invention, and the ranging anomalies are much less than DEKF. This is because the combination of the CNN-LSTM optimization algorithm and the DEKF filtering algorithm can ensure the reduction of the influence of NLOS propagation, multipath interference and accumulated error of IMU drift over time of UWB, and the distributed architecture of DEKF can prevent problems caused by asynchronous sensor data input, thereby maintaining a consistent system state. Overall, the proposed system can dynamically adapt to environmental changes, ensure stable operation in different indoor scenes, and achieve highly robust positioning.

[0111] Table 2 Comparison of root mean square error and absolute average error of five different optimization algorithms

[0112]

[0113] Table 2 is a comparison of the root mean square error and absolute average error of five optimization algorithms (no algorithm, EKF, DEKF, CNN-LSTM, CNN-STM-DEKF). The root mean square error (RMSE) values ​​of the CNN-STM-DEKFRMSE optimization algorithm proposed in the present invention are 0.242 and 0.205 (broken line path and square path), which are significantly lower than other algorithms; the absolute average error (MAE) values ​​are 0.208 and 0.192, respectively, which further proves its optimization effect. In the broken line path, the errors of each algorithm are large, but the error of CNN-LSTM-DEKF is significantly lower than that of other algorithms; in the square path, CNN-LSTM-DEKF also maintains the lowest error, showing high adaptability to regular trajectories. Comprehensive analysis shows that CNN-LSTM-DEKF significantly reduces the impact of NLOS interference and IMU drift through CNN-LSTM deep feature optimization of UWB signals, and combines the distributed filtering method of DEKF to achieve the complementarity of time and space information. In complex indoor scenes, the proposed algorithm exhibits excellent positioning accuracy and robustness, and is suitable for dynamic and changing environments.

[0114] In order to more clearly show the performance evaluation results of the algorithm proposed in this invention and other algorithms, Figure 6The box plots of ranging errors of five optimization algorithms (no algorithm, EKF, DEKF, CNN-LSTM, CNN-STM-DEKF) are shown. Figure 6 (a) is a box plot of the root mean square error (RMSE), and the data reflects the error distribution of each optimization algorithm; Figure 6 (b) is a box plot of mean absolute error (MAE), which intuitively compares the positioning accuracy performance of the five algorithms. Through the comparison, it can be clearly seen that the CNN-LSTM-DEKF algorithm proposed in the present invention shows better performance in both RMSE and MAE, and the error distribution range is the smallest.

[0115] The present invention proposes an indoor integrated positioning system based on a distributed extended Kalman filter deep learning optimization algorithm, and proposes a CNN-LSTM DEKF optimization algorithm based on tight coupling of UWB and IMU. The indoor positioning method based on tight coupling of UWB and IMU obtains ranging information through IMU and UWB sensors, and then uses the CNN-LSTM-DEKF optimization algorithm to improve the positioning accuracy. The experimental results show that compared with the traditional UWB / IMU fusion method, the CNN-LSTM-DEKF method achieves significantly higher positioning accuracy. In order to evaluate the robustness of the proposed method, by comparing the RMSE and MAE of the optimization algorithm proposed in the present invention with other algorithms, it can be concluded that the CNN-LSTM-DEKF optimization algorithm proposed in the present invention shows better performance, and the RMSE values ​​and MAE values ​​under two different paths are 0.242, 0.208 and 0.205, 0.192 respectively. Therefore, the CNN-LSTM-DEKF optimization algorithm can be a better choice for improving the positioning accuracy of fusion sensors in indoor environments. In addition, future work will also require verifying the positioning accuracy of the optimization algorithm in complex environments and exploring the performance of integrating camera input into the system.

[0116] The present invention also provides an electronic device, comprising a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps involved in the system described above can be implemented.

[0117] The present invention also provides a computer-readable storage medium on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, the method steps involved in the system described above can be implemented.

[0118] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0119] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0120] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0121] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0122] References:

[0123] [1]Picard Joseph S,Weiss Anthony J,et al.Direct PositionDetermination Sensitivity to NLOS Propagation Effects on Doppler-Shift[J].IEEE Transactions on signal processing,2019,67(14),3870-3881.

[0124] [2]Hyunseob Oh,Seungbeom Seo,Jinwook Seo,et al.An Efficient Ultra-Wideband Infrastructure Planning for Indoor Positioning Systems Using DL-TDOA[J].IEEE conference 0n global communications,GLOBECOM,2023,201-206.

[0125] [3]Kim,Moon-Sik,Yu,Si-Bok,Lee,Kwang-Soo,et al.Development of a High-Precision Calibration Method for Inertial Measurement Unit[J].

[0126] Internationaljournal ofprecision engineering and manufacturing,2014,15(3),567-575.

[0127] [4]Gentner Christian,Avram Diana,et al.WiFi-RTT-SLAM:Simultaneouslyestimating the positions ofmobiledevices and WiFi-RTT access points[J].IONGNSS+,2021,3142-3148.

[0128] [5]Zhang Xiaotian,Jia Guanghui,Huang Hai,et al.An approach forconstituting double / multi wall BLE by single wall BLE of spacecraft shield[J].International journal of impact engineering,2014,69,114-121.

[0129] [6]Hao Pan,Yu Chen,Xiaogang Qi,Meili Liu,et al.A UWB / IMU-AssistedFingerprinting Localization Framework with Low Human Efforts[J].

[0130] China communications,2024,21(6),40-52.

[0131] [7]Zheng Shuaikang,Li Zhitian,Liu Yunfei,et al.An Optimization-BasedUWB-IMU Fusion Framework for UGV[J].IEEE sensorsjournal,2022,22(5),4369-4377.

[0132] [8]Zhu Si,Sun Fei,Chen Xiaohui,et al.Direct Position DeterminationSensitivity to NLOS Propagation Effects on Doppler-Shift[J].IEEEInternational Conference on Communications in China,2013,321-326.

[0133] [9]Jiang Changhui,Chen Shuai,Chen Yuwei,et al.An UWB Channel ImpulseResponse De-Noising Method for NLOS / LOS Classification Boosting[J].IEEEcommunications letters,2020,24(11),2513-2517.

[0134]

[10] Kim Dae-Ho,FarhadArshad,Pyun Jae-Young,et al.UWB PositioningSystem Based on LSTM Classification With MitigatedNLOS Effects[J].IEEEinternet ofthingsjournal,2023,10(2),1822-1835.

[0135]

[11] Lee Gang Toe,Seo Seung Beom,Jeon Wha Sook,et al.IndoorLocalization by Kalman Filter based Combining of UWB-Positioning and PDR[J].IEEE consumer communications and networking conference,2021.

[0136]

[12] Wang Yan,Li Xin,et al.An improved robust EKF algorithm based onsigma points for UWB and foot-mounted IMU fusion positioning[J].IEEETransactions on signal processing,2021,66(2),329-350.

[0137]

[13] Liu Fei,Li Xin,Wang Jian,et al.An Adaptive UWB / MEMS-IMUComplementary Kalman Filter for Indoor Location in NLOS Environment[J].Remote sensing,2019,11(22),2628.

[0138]

[14] J Zeng,J Liu,T Zou,et al.Distributed extended Kalman filteringfor wastewater treatment processes[J].Industrial and engineering chemistryresearch,2016,55(28),7720-7729.

[0139]

[15] Li Xiaojie,Yin Xunyuan,et al.Exploring the linkage betweenpartition-based distributed extended Kalman filter and distributed full-information estimation[J].IEEE 7th international conference on industrialcyber-physical systems,2024.

[0140]

[16] de Hidenori,Kurita Takio,et al.Improvement of Learning for CNNwith ReLU Activation by Sparse Regularization[J].IEEE internationaljointconference on neural networks,2017,2684-2691.

[0141]

[17] Yadav Hemant, Thakkar Amit, et al. NOA-LSTM: An efficient LSTM cellarchitecture for time series forecasting[J]. Expert systems with applications, 2023, 122333.

[0142]

[18] Luo Long, et al. Research and analysis of manipulator control method based on deep learning [J]. Journal of engineering research, 2024, 12(3), 412-420.

[0143]

[19]

[0144]

[20] Changhui Jiang,Jichun Shen,Shuai Chen,et al.UWB NLOS / LOSClassification Using Deep Learning Method[J].IEEE Transactions on signalprocessing,2020,24(10),2226-2230..

[0145] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. An indoor integrated positioning system based on a distributed extended Kalman filter deep learning optimization algorithm, characterized in that: include: Ultra-wideband UWB module, collects ultra-wideband UWB ranging data, including ultra-wideband UWB channel impulse response CIR data; Inertial measurement unit IMU module, collects IMU ranging data; Based on the distributed extended Kalman filter deep learning optimization algorithm module, an optimization algorithm CNN-LSTM-DEKF that integrates the convolutional long short-term memory network CNN-LSTM and the distributed extended Kalman filter DEKF is proposed. The CNN-LSTM is used to extract features and classify the CIR data of the ultra-wideband UWB channel impulse response. The classified prediction data is then fused with the inertial measurement unit IMU ranging data through DEKF to suppress the cumulative error caused by the drift of the IMU over time, thereby ensuring the accuracy of the ranging data while also ensuring that the fusion of each sensor data will not have cumulative errors and time synchronization problems.

2. The indoor integrated positioning system based on the distributed extended Kalman filter deep learning optimization algorithm according to claim 1 is characterized in that: The network framework of CNN-LSTM includes two basic components: convolutional neural network CNN and long short-term memory LSTM. In CNN-LSTM, CIR data is used as the input vector of CNN. CNN uses the non-temporal structure of the input signal for feature extraction by learning the frequency characteristics of CIR data. The output of CNN is then input into LSTM to learn the dependency of time series data for classification and prediction of non-line-of-sight NLOS signals and line-of-sight LOS signals.

3. The indoor integrated positioning system based on the distributed extended Kalman filter deep learning optimization algorithm according to claim 1 or 2, characterized in that: CNN-LSTM consists of a convolutional layer, a maximum pooling layer, a fully connected layer, an LSTM layer and a fully connected layer, and the generalization ability is improved through Dropout regularization.

4. The indoor integrated positioning system based on the distributed extended Kalman filter deep learning optimization algorithm according to claim 2 is characterized in that: CNN-LSTM performs feature extraction and classification prediction on the channel impulse response CIR data of ultra-wideband UWB, which specifically includes the following stages: CNN components: CNN processes the raw CIR data and extracts spatial features that capture the signal pattern changes caused by NLOS conditions. Multiple convolutional layers are used, and each convolutional layer uses the same activation function to detect the input data for features. To ensure that the activation function is nonlinear, ReLU is used, and its equation is as follows: x represents the parameters of the ReLU function; After the convolutional layer outputs to the pooling layer, the pooling layer reduces the dimension and prevents overfitting by downsampling; then, the pooling layer output is input to the fully connected layer, and the fully connected layer selects the classifier as the output through the SoftMax function; LSTM components: The output of the CNN is passed to the LSTM to capture the temporal dependencies in the CIR data sequence. The gating architecture of the LSTM enables it to model long-term relationships in the data. After the data extracted by CNN features is input into LSTM, LSTM is trained using a labeled dataset with a cross entropy loss function to ensure that the UWB signal is accurately classified into LOS or NLOS.

5. The indoor integrated positioning system based on the distributed extended Kalman filter deep learning optimization algorithm according to claim 4 is characterized in that: The gating architecture of LSTM includes input, forget, and output gates.

6. The indoor integrated positioning system based on the distributed extended Kalman filter deep learning optimization algorithm according to claim 4 or 5, characterized in that: The input gate and output gate functions in the LSTM gating architecture are described as follows: i t =σ(W i ·[h t-1 ,x t ]+b i ) the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t ·tanh(C t ) Where i t is the input gate, which determines whether the spatial features extracted by CNN, i.e., the classification-related information including NLOS and LOS, are used to update the time series state; h t-1 represents the hidden state of the previous time step, which contains the dynamic information of the feature sequence of the past time step; x t Represents the feature vector of the CNN output at the current time step, which contains the UWB CIR spatial features of the current time step; W i 、b i represents the weight and bias parameters of the input gate, W o 、b o Represents the weight and bias parameters of the output gate, which are optimized by gradient descent during training to adapt to the influence of different input features; is a candidate state, which is used to generate a new candidate state for the current time step. It captures the potential features of the CNN features of the current time step and the historical hidden state, and retains important information related to LOS / NLOS. Tanh represents the hyperbolic tangent activation function, which compresses the candidate state to the range of (-1,1) to represent possible state updates. W C 、b C The weights and bias parameters representing the candidate states can effectively characterize the nonlinear relationship of the time series after optimization; o t is the output gate, which extracts the most relevant information for LOS / NLOS classification from the candidate states and outputs it to h t ;h t Represents the hidden state of the current time step, which is the output of LSTM and one of the inputs of the next time step; tanh(C t ) means to perform a nonlinear transformation on the current time series state so that its range remains in (-1,1).

7. The indoor integrated positioning system based on the distributed extended Kalman filter deep learning optimization algorithm according to claim 1 is characterized in that: The DEKF calculation process is as follows: Step 1: Local prediction stage Each node i runs the prediction step independently, updating its state estimate based on the local model: is the predicted state estimate of node i at the current time k compared to the previous time k-1; represents the predicted value of the state of node i at time k-1; u k-1,i represents the control input of node i at time k-1; P k∣k-1,i is the prediction error covariance matrix of node i compared to the previous moment k-1 and the current moment k, F k,i represents the state transition function of node i at time k; P k-1∣k-1,i represents the prediction error covariance matrix of node i at time k-1; represents the state transfer matrix obtained by derivation of the state transfer function; Q k,i Represents the system noise covariance matrix, describing the uncertainty of state transition; Step 2: Local update phase Each node updates its local state based on the observed value: P k∣k,i =(I-K k,i H k,i )P k∣k-1,i K k,i is the Kalman gain of node i, which is used to weigh the influence of prediction and observation; H k,i The Jacobian matrix of the observation function; z k,i represents the measurement value of node i; h i Represents a nonlinear observation function, describing the mapping relationship from state to observation value; R k,i represents the observation noise covariance, describing the uncertainty of sensor measurement; Step 3: Information sharing stage: each node shares local state estimation and covariance information through communication to form a global information set. The shared content includes P k∣k,i , K k,i ; Step 4: Global fusion stage: obtain the optimal state estimate of the system as a whole through global information fusion, and perform weighted summation of the estimates of each node according to the covariance matrix weights through the weighted average method: represents the optimal state estimation after global fusion, W k,i represents the weight of the i-th node, represents the local state estimate of the i-th node; Step 5: In the global update feedback phase, the fusion results are fed back to each node and the local state is adjusted to maintain global consistency.

8. The indoor integrated positioning system based on the distributed extended Kalman filter deep learning optimization algorithm according to claim 7 is characterized in that: DEKF is validated by MAE and RMSE.

9. The indoor integrated positioning system based on the distributed extended Kalman filter deep learning optimization algorithm according to claim 1 is characterized in that: The ultra-wideband (UWB) module collects the ranging data of the anchor point through the ultra-wideband (UWB) tag.

10. The indoor integrated positioning system based on the distributed extended Kalman filter deep learning optimization algorithm according to claim 1, characterized in that: The inertial measurement unit (IMU) module collects ranging data through a three-axis acceleration sensor.

Images