UWB positioning precision optimization method based on IWPT-BILSTM

Abstract

The invention discloses a UWB positioning precision optimization method based on IWPT-BILSTM, and relates to the field of UWB positioning. A lightweight attention module is introduced to improve a traditional WPT, the improved WPT (IWPT) has a self-adaptive sub-band screening and weighted reconstruction mechanism, accurate time-frequency decomposition and effective feature extraction are performed on original UWB signals by using the IWPT, noise in data is weakened to reserve more effective time-frequency signals, and the UWB signals are subjected to time-frequency decomposition and effective feature extraction. And then low-noise and high-validity data are sent to a BI-LSTM network for training, the BI-LSTM network is used for carrying out time sequence correlation and dynamic error correction on effective characteristic signals after IWPT processing, and finally nonlinear time sequence characteristics generated in the time sequence correlation process are accurately mapped into a UWB positioning result. By adopting the method, the positioning error can be obviously reduced, the positioning robustness and generalization ability in a complex environment can be improved, and the method has a relatively good engineering application value.

Description

TECHNICAL FIELD

[0001] The application relates to the field of ultra-wideband (UWB) positioning technology, and particularly relates to a UWB positioning precision optimization method based on an IWPT-BILSTM. BACKGROUND

[0002] With the rapid development and deep penetration of UWB technology in the field of high-precision positioning, the UWB-based positioning system has shown significant technical potential and application advantages in multiple scenarios. UWB technology, with its ultra-large bandwidth characteristics (usually defined as a relative bandwidth of more than 20% or an absolute bandwidth of more than 500MHz), ultra-low power design and excellent anti-multipath interference capability, has attracted widespread attention from academia and industry in the fields of indoor complex environment positioning, industrial Internet of Things (IIoT) device cooperation, intelligent networked vehicle environment perception, unmanned aerial vehicle autonomous navigation and obstacle avoidance, and smart warehouse dynamic tracking. UWB positioning technology has gradually become the focus of interdisciplinary research. Although UWB technology has the technical endowment of centimeter-level positioning in the theoretical level, the positioning accuracy and system robustness in its actual deployment are still affected by multiple complex factors: firstly, the signal attenuation and time delay caused by non-line-of-sight (NLOS) propagation significantly reduces the estimation accuracy of core ranging parameters such as time difference of arrival (TDoA) or time of arrival (ToA); secondly, the signal reflection, diffraction and scattering caused by multipath effect form the superposition of redundant observation information and noise interference; thirdly, the random obstruction of obstacles and the spectral superposition of electromagnetic interference sources (such as industrial equipment and wireless communication signals) in the dynamic environment further aggravate the uncertainty of the positioning system; in addition, device-level errors such as hardware clock drift and antenna phase deviation also have a non-negligible impact on the positioning performance. Therefore, how to break through the positioning precision bottleneck in complex scenarios through signal processing algorithm optimization, multi-source information fusion, environment adaptive modeling and other technical paths, and improve the reliability and stability of the UWB system in the dynamic interference environment, has become a key scientific problem and engineering challenge to be solved in the current positioning field. SUMMARY

[0003] Based on this, this invention conducts an in-depth study of the current UWB positioning system in the ADS-TWR (Asymmetric Double Sided Two-Way Ranging) scenario. Combining the improved wavelet packet transform (IWPT) and bidirectional long short-term memory network (BI-LSTM), it proposes a UWB positioning accuracy optimization method based on IWPT-BILSTM to improve the positioning accuracy and robustness of the current UWB positioning system in complex environments.

[0004] The technical solution of this invention is:

[0005] A method for optimizing UWB positioning accuracy based on IWPT-BILSTM, comprising the following steps:

[0006] Step 1: The original UWB signal is decomposed and denoised using the improved WPT, i.e., IWPT, to extract effective feature signals that are strongly correlated with the positioning, and the optimized positioning is solved at the same time.

[0007] Step 2: Use the BI-LSTM network to perform temporal correlation and dynamic error correction on the effective feature signals extracted in Step 1 to obtain the predicted localization;

[0008] Step 3: Subtract the optimized positioning from the predicted positioning to obtain the prediction residual. Correct the predicted positioning using the prediction residual to obtain the optimal predicted positioning.

[0009] Optionally, based on the UWB positioning accuracy optimization method, a lightweight attention module is introduced into the subband selection and denoising stage of traditional WPT to obtain IWPT.

[0010] Optionally, according to the UWB positioning accuracy optimization method, the lightweight attention module is a fully connected neural network architecture containing only a 2D input layer, a single hidden layer, and an output layer, with its number of parameters strictly controlled to within 65.

[0011] Optionally, according to the UWB positioning accuracy optimization method, step 1 further includes the following steps:

[0012] Step 1.1: Calculate the initial distance between the UWB tag and each UWB base station;

[0013] Step 1.2: Use IWPT to perform multi-scale decomposition and denoising on the original UWB signal, extract effective feature signals that are strongly correlated with the positioning, and reconstruct the effective feature signals to obtain an optimized feature set;

[0014] Step 1.3: Based on the initial distance between the UWB tag and each UWB base station, extract the propagation time between the UWB tag and each UWB base station from the optimized feature set, calculate the optimized distance through the propagation time, and then solve for the optimized positioning through the optimized distance.

[0015] Optionally, according to the UWB positioning accuracy optimization method, step 1.2 further includes the following steps:

[0016] Step 1.2.1: Select WPT basis functions that are adapted to the characteristics of the original UWB signal;

[0017] Step 1.2.2: Perform multi-scale frequency band decomposition on the original UWB signal based on the selected WPT basis functions to obtain multiple sub-bands;

[0018] Step 1.2.3: Extract the features of the decomposed subbands, introduce a lightweight attention module to select the subbands that contribute more to the localization from the decomposed subbands and perform weighted processing on them;

[0019] Step 1.2.4: Reconstruct the weighted subbands to obtain the optimized feature set.

[0020] Optionally, according to the UWB positioning accuracy optimization method, step 1.2.3 further includes the following steps:

[0021] Step 1.2.3.1: Extract the features of the decomposed subbands, including the energy of each subband. Pearson correlation coefficients between each sub-band and the actual location tag ;

[0022] Step 1.2.3.2: Based on the features of the decomposed sub-bands, a lightweight attention module is introduced to calculate the attention weights of each sub-band;

[0023] Step 1.2.3.3: Based on the attention weight of each subband and the set weight threshold, adaptively filter out the subbands with high attention weights and weight them.

[0024] Optionally, according to the UWB positioning accuracy optimization method, in step 1.2.3.3, the selected subbands with high attention weights are weighted according to equations (16) and (17):

[0025] (16)

[0026] (17)

[0027] in Represented as the first Layer The coefficients of each screening subband The coefficients of invalid subbands are set to zero to avoid interference with the reconstructed signal; This indicates the set weight threshold; Represented as the soft thresholding denoising coefficient of the sub-band; Indicates the first Layer Attention weights for each selected subband; Represented as the first Layer The weighting coefficients of each sub-band.

[0028] Optionally, according to the UWB positioning accuracy optimization method, step 2 further includes the following steps:

[0029] Step 2.1: Preprocess the optimized feature sequences in the optimized feature set, including normalization and construction of time series data samples;

[0030] Step 2.2: Using the preprocessed data as input data for the BI-LSTM model, the BI-LSTM network model is used to extract core temporal correlation features and fuse the extracted core temporal correlation features. Finally, the fused features are mapped to 3D positioning coordinates.

[0031] Step 2.3: Perform dynamic error correction on the 3D positioning coordinates using the BI-LSTM model to obtain the predicted positioning.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] (1) In view of the shortcomings of the traditional WPT algorithm in selecting effective subbands by setting a manual threshold, this invention introduces a lightweight attention module to improve the traditional WPT algorithm. The improved WPT algorithm (IWPT) has an adaptive subband selection and weighted reconstruction mechanism, which solves the shortcomings of the traditional WPT algorithm in selecting effective subbands by setting a manual threshold.

[0034] (2) For the original positioning signal with large errors generated in UWB positioning, the method of this invention improves positioning performance through signal preprocessing and deep modeling. First, IWPT is used to decompose the original UWB signal into multiple scales, filter out electromagnetic and device acquisition noise, extract effective frequency bands strongly correlated with positioning, and output a smooth and feature-clear preprocessed signal; then, it is input into the BI-LSTM network, which uses bidirectional units to capture the temporal correlation features of the signal, and combines a gating mechanism to strengthen key positioning information, and finally outputs accurate positioning coordinates. It can be seen that this positioning accuracy optimization method consists of only two algorithms, and the fusion algorithm requires low computing power when performing UWB positioning optimization.

[0035] (3) Using IWPT to perform multi-scale decomposition and soft threshold denoising on the original UWB signal can accurately separate the effective signal from multipath interference and electromagnetic noise, thereby improving the signal quality by denoising from the source of the original UWB signal and greatly reducing the noise error of subsequent positioning calculation.

[0036] (4) The subbands selected by the IWPT algorithm in this invention do not participate in the reconstruction equally, but are weighted according to attention weight. The subbands with higher weights contribute more to the reconstructed signal, which is equivalent to amplifying the effective features and suppressing residual noise, thereby further improving the upper limit of the accuracy of the reconstructed signal.

[0037] (5) The lightweight attention module introduced in this invention adopts a minimalist fully connected structure—a three-layer network structure including a 2D input layer, a single hidden layer and an output layer. The number of model parameters is strictly controlled to within 65, which is significantly reduced compared to traditional attention-based models, ensuring the lightweight nature and universality of the model.

[0038] (6) This invention uses BI-LSTM for time series modeling, which can simultaneously capture the motion inertia of UWB positioning signals from "past to present" and the trend correlation from "future to present", adapt to the motion characteristics of dynamic targets, and effectively improve the positioning accuracy in dynamic scenes.

[0039] (7) Introducing dynamic threshold adjustment and anomaly handling strategies, adaptively optimizing residual correction thresholds for different motion speeds, and enabling weighted fusion logic when residuals exceed limits can effectively improve the robustness of the system in complex interference environments and avoid a sharp drop in accuracy caused by a single correction logic.

[0040] (8) The dual-benchmark residual correction mechanism of IWPT optimized positioning and BI-LSTM prediction constructed in this invention balances the reliability of the two types of positioning results through adaptive correction coefficients, effectively suppresses model generalization error and hardware-level deviation, and enables static positioning accuracy to break through the centimeter level, meeting the stringent requirements of industrial scenarios for positioning stability.

[0041] (9) The lightweight design enables IWPT denoising and BI-LSTM prediction to be performed in real time on the edge processor. The data processing time per frame is ≤8ms, which meets the real-time positioning requirements in industrial and engineering scenarios and avoids positioning delays caused by excessive algorithm complexity.

[0042] (10) The technical architecture of the present invention has good scalability, supports dynamic expansion of the number of UWB tags and UWB base stations, supports two-dimensional and three-dimensional scenes, and is compatible with multi-sensor data fusion such as inertial measurement unit and lidar, breaking through the limitations of single UWB technology.

[0043] In summary, the method of this invention can significantly reduce positioning errors, improve positioning robustness and generalization ability in complex environments, and has good engineering application value. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the lightweight attention module in this embodiment;

[0045] Figure 2 This is a flowchart of the UWB positioning accuracy optimization method based on WPT-BILSTM in this embodiment;

[0046] Figure 3 A three-dimensional UWB system positioning diagram provided for this embodiment;

[0047] Figure 4 This is a tree-like structure diagram of WPT;

[0048] Figure 5 This is a structural diagram of BI-LSTM. Detailed Implementation

[0049] To facilitate understanding of this application, a more comprehensive description of this application will be provided below with reference to the accompanying drawings.

[0050] The basic idea of ​​this invention is as follows: First, the improved Wavelet Packet Transform (IWPT) is used to perform accurate time-frequency decomposition and effective feature extraction on the original UWB signal, reducing noise in the data to retain more effective time-frequency signals. Then, the low-noise, high-effectiveness data is fed into the BI-LSTM network for training. The BI-LSTM network is used to perform temporal correlation and dynamic error correction on the effective feature signals after IWPT processing. Finally, the nonlinear temporal features generated by the temporal correlation process are accurately mapped to the UWB positioning results.

[0051] Traditional WPT (Wideband Perspective Testing) primarily uses a manually set threshold, typically a percentage of the total energy of all subbands. This manually set percentage is usually between 5% and 10%, and the selection logic is based on comparing subband energies. The threshold value is used to remove noisy subbands. Traditional WPT filters effective subbands by manually setting a threshold. If the threshold is set too high, more noisy subbands can be removed, but more effective subbands will be mistakenly deleted. If the threshold is set too low, more effective subbands can be retained, but a large number of noisy subbands will be retained, resulting in incomplete denoising.

[0052] To address this deficiency, this implementation introduces a lightweight attention module into the subband selection and denoising stages of traditional WPT, and further optimizes the subsequent signal reconstruction stage. The purpose of introducing a lightweight attention module into traditional WPT is to accurately reflect the value of subbands to localization during subband selection, clearly distinguishing the magnitude of each subband's contribution to localization. This implementation selects only two localization-related features, including subband energy. Pearson correlation coefficient between sub-band and real location tag .

[0053] This lightweight attention module is a lightweight fully connected attention module, such as... Figure 1 As shown, a minimalist fully connected neural network architecture is adopted, containing only a three-layer network structure: a 2D input layer, a single hidden layer, and an output layer. The number of model parameters is strictly controlled to within 65, which is significantly reduced compared to traditional attention-based models, effectively achieving the goal of lightweight design. The input layer uses a two-dimensional standardized feature vector. The input data is 16 neurons in the hidden layer, which uses the ReLU activation function. The output layer is a single neuron structure that outputs attention weights using the Sigmoid activation function.

[0054] This lightweight attention module calculates the localization contribution weight. Based on location contribution weight The subbands with high contribution weights are selected, and the retained subbands with high weights are weighted and then the signal is reconstructed. The original UWB signal is denoised by the improved WPT (IWPT), which makes the denoised signal cleaner and lower in noise, providing more accurate and stable input data for the subsequent input to the BI-LSTM network.

[0055] Figure 2 This is a flowchart of the UWB positioning accuracy optimization method based on IWPT-BILSTM in this embodiment. Figure 2 As shown, the UWB positioning accuracy optimization method based on IWPT-BILSTM includes the following steps:

[0056] Step 1: The original UWB signal is decomposed and denoised using the improved WPT, i.e., IWPT, to extract effective feature signals that are strongly correlated with the positioning, and the optimized positioning is solved at the same time.

[0057] Step 1.1: Calculate the initial distance between the UWB tag and each UWB base station.

[0058] In a UWB positioning system within an ADS-TWR scenario, the distance between the UWB tag and each UWB base station needs to be calculated by measuring the time it takes for the signal emitted by the UWB tag to reach each UWB base station. Then, the location of the UWB tag is determined using these calculated distances. Firstly, as...Figure 3 As shown, multiple, for example, are set within the positioning area. The first UWB base station transmits a ranging request frame from the UWB tag, while a local timer records the local timestamp at that moment, recording the distance to the first UWB base station. The time it takes for a UWB base station to send a ranging request frame ;No. Upon receiving a ranging request frame, the UWB base station immediately records the timestamp of the reception time and records the first... The time it takes for each UWB base station to successfully receive the request frame is ;No. Each UWB base station waits for a preset fixed response delay. Then, a response frame is sent to the UWB tag and the timestamp of the received moment is immediately recorded, recording the first... The time it takes for a UWB base station to send a response frame to a UWB tag is The UWB tag receives the response frame sent by the base station and immediately records the timestamp of the reception time, recording the number of times the UWB tag receives the first response frame. The time of each base station is .

[0059] Assuming the signal is in the UWB tag and the first The total time for one round trip between UWB base stations is The one-way propagation time of the signal can then be calculated using the four timestamps mentioned above according to equation (1). .

[0060] (1)

[0061] In ADS-TWR scenarios, clock synchronization errors can be offset, i.e., the one-way propagation time of the signal. The calculation results are unaffected by clock drift; the signal propagates in the air as an electromagnetic wave at a speed approximately equal to the speed of light. ( ), combined with one-way transmission time The distance between the UWB tag and the first tag can be calculated according to formula (2). The distance between UWB base stations.

[0062] (2)

[0063] in, Indicates the distance of the UWB tag from the first... The distance between each UWB base station is [number]. The number of UWB base stations is mainly based on the actual number of UWB base stations deployed.

[0064] In practical applications, due to the existence of measurement errors, further data processing and algorithm optimization are needed to improve the accuracy and reliability of positioning. Through these steps, the initialization work for positioning in the ADS-TWR scenario can be completed, and the necessary data and models can be prepared for subsequent data processing and location calculation.

[0065] Step 1.2: Use IWPT to perform multi-scale decomposition and denoising on the original UWB signal, extract effective feature signals that are strongly correlated with localization, and reconstruct the effective feature signals to obtain an optimized feature set.

[0066] Step 1.2.1: Select WPT basis functions that are adapted to the characteristics of the original UWB signal;

[0067] WPT decomposition relies on scaling functions for low-frequency analysis and wavelet functions for high-frequency analysis, such as Figure 4 As shown, this enables the analysis of the entire frequency band of the original UWB signal. In engineering practice, the db4 wavelet packet is usually selected as the WPT basis function for analysis. Specifically, the scaling function shown in equation (3) is used for low-frequency analysis to extract the signal trend.

[0068] (3)

[0069] in, This represents the db4 scaling function, and the output is a low-frequency filter kernel. Indicates a scaling filter; express Scale by 2 times and translate This operation ensures that the filter core is adapted to different frequency domain scales.

[0070] The wavelet function shown in equation (4) is used for high-frequency analysis to extract signal details and noise.

[0071] (4)

[0072] in, This represents the db4 wavelet function, and its output is a high-frequency filter kernel; This represents a db4 wavelet filter that satisfies... It has orthogonality, which can ensure that the decomposed signal can be reconstructed without distortion.

[0073] Step 1.2.2: Perform multi-scale frequency band decomposition on the original UWB signal based on the selected WPT basis functions to obtain multiple sub-bands;

[0074] The original signal time series dataset output in the ADS-TWR scenario As input, where , For the number of base stations, Given the observation time length, the sampling rate of the original signal in the ADS-TWR scenario is [missing information]. (Typically 5Hz is chosen to satisfy the Nyquist criterion for UWB signals) and effective frequency band (3~10HZ), let the number of decomposition layers be... Furthermore, it conforms to the core principles of decomposition: the sub-band bandwidth after decomposition is ≤ the effective frequency band width, and the center frequency of the highest sub-band is ≥ the upper limit of the effective frequency band. Based on the frequency domain decomposition characteristics of wavelet packet transform, the first... After layer decomposition, the bandwidth of each subband is For the original signal time series dataset Perform Mallat algorithm decomposition, such as Figure 4 As shown, the recursive formula includes:

[0075] No. Layer approximation coefficients (low frequency, effective signal trend):

[0076] (5)

[0077] No. Layer detail factor (high frequency, detail / noise):

[0078] (6)

[0079] Wavelet packet extension (further decomposition of high-frequency subbands):

[0080] (7)

[0081] (8)

[0082] in, It is the first Layer approximation coefficient, The coefficient index is the length of the original signal sequence. ; It is the first Layer detail coefficients mainly contain the high-frequency components of the signal (multipath interference, electromagnetic noise); It is the first Layer Approximate coefficients for each high-frequency subband; No. Layer Detail coefficients of each high-frequency subband; It is a scaling filter. It is a wavelet filter that satisfies The above decomposition can achieve full-frequency domain decomposition of the signal without omission, and can extract the effective signal more efficiently.

[0083] original Sequence decomposition, the original sequence length is The sampling frequency is The observation time is T, and it satisfies ,through After layer decomposition, we obtain Sub-band coefficient Each sub-band length is 1 / 3 of the original sequence length. .

[0084] in, correspond The mapped frequency band includes The direct component of the signal belongs to the effective signal frequency band, and its energy accounts for ≥10%; correspond The mapped frequency band mainly contains signals with multipath interference and noise, with an energy percentage of ≤5%.

[0085] Step 1.2.3: Extract the features of the decomposed subbands, introduce a lightweight attention module to select the subbands with high localization contribution from the decomposed subbands and perform weighted processing on them;

[0086] Step 1.2.3.1: Extract the subband features after decomposition, mainly including subband energy. Pearson correlation coefficient between sub-band and real location tag ;

[0087] Energy is a fundamental indicator of the effective information contained in a subband. The energy of each subband can be calculated according to equations (9) and (10). With energy percentage ;

[0088] Sub-band energy:

[0089] (9)

[0090] Energy percentage:

[0091] (10)

[0092] in, Indicates the first In the layer decomposition The energy carried by a person; This represents the index of a sample point within a sub-band, used to iterate through each sample point within the sub-band. Indicates the first Layer In the sub-band, the first The wavelet packet coefficients corresponding to each sample point reflect the signal characteristics of that sample point in the corresponding sub-band; The subband length varies with the number of decomposition layers. The increase in the value decreases as the value increases, meaning that the more detailed the decomposition, the fewer sample points each sub-band contains; Indicates the first All layers The total energy carried by each individual.

[0093] The Pearson correlation coefficient between the sub-band and the actual location tag is calculated according to equation (11). This coefficient is used to measure the linear correlation between the sub-band coefficient and the actual distance from the UWB tag to the UWB base station. The stronger the correlation, the greater the contribution of the sub-band to the location.

[0094] (11)

[0095] in, For the first Layer Pearson correlation coefficient of individual bands ; For the first The actual distance between the UWB tag and the UWB base station at any given time corresponds to... The truth value, in meters; for The mean, ; for The mean, .

[0096] To effectively eliminate the significant difference in numerical magnitude between energy and correlation coefficient, this invention introduces the Z-score standardization method, the standardization formula of which is as follows:

[0097] (12)

[0098] in, For the first Layer The standardized feature vectors of each sub-band are the input to the subsequent attention module; Indicates the first In the layer decomposition The energy of each sub-band is calculated according to formula (5); yes The average energy carried by each individual ; yes The standard deviation of energy carried by size ; yes The mean of the correlation coefficients of each subband. ; yes The standard deviation of the correlation coefficient of each sub-band .

[0099] Step 1.2.3.2: Based on the features of the decomposed sub-bands, a lightweight attention module is introduced to calculate the attention weights of each sub-band;

[0100] To achieve more accurate subband selection, this invention constructs a lightweight fully connected attention module. It employs a simplified fully connected neural network architecture, building a three-layer network structure consisting of a 2D input layer, a single hidden layer, and an output layer. The number of model parameters is strictly controlled to within 65, significantly reducing the number compared to traditional attention-based models, effectively achieving the lightweight design goal. The localization contribution weight of each subband is calculated through an automatic learning mechanism. The specific implementation process is as follows:

[0101] Standardized feature vector This is the input layer data for the attention module, and its network structure is as follows:

[0102] Input layer: Receives two-dimensional standardized feature vectors ;

[0103] Hidden layer: Contains 16 neurons, using the ReLU activation function to avoid gradient vanishing, and can be represented as... ,in, Represented as the first Layer The 16-dimensional output vector is obtained after the input feature vectors are processed by the hidden layer. Each dimension corresponds to the output of a neuron in the hidden layer.

[0104] Output layer: Single neuron structure, outputting attention weights through a sigmoid activation function. ,in, Represented as the first Layer The attention weights obtained from the input feature vectors have values ​​ranging from 0 to 1 and are used to measure the importance of the feature vector in subsequent processing.

[0105] Let the weight matrix from the input layer to the hidden layer be... The bias vector is The weight matrix from the hidden layer to the output layer is: The bias scalar is ,but;

[0106] Hidden layer output:

[0107] (13)

[0108] in, Element-wise application to vectors ;

[0109] Sub-band positioning contribution weight output:

[0110] (14)

[0111] in, Represented as the first Layer Hidden layer output of each sub-band; Represented as the first Layer Attention weights are obtained from each input feature vector. The higher the weight, the more significant the improvement in positioning accuracy of the corresponding sub-band. .

[0112] Meanwhile, a loss function is designed to guide the lightweight attention module to learn the weights of the actual localization contribution: the subband that reduces localization error is given high weight, and the noisy subband is given low weight. In order to achieve adaptive weight allocation of subband features, a supervision mechanism for localization error optimization is constructed, and a new generalization loss function is designed as follows.

[0113] (15)

[0114] in, Indicates the total network loss; Indicates the number of samples in the training set; Represented as the first The first sample Positioning error of each sub-band ,in, Indicated as using only the first The distance of the reconstructed sub-band; Represented as the first The true distance of each sample; Represented as the first The first sample Layer The weight of each sub-band.

[0115] Step 1.2.3.3: Attention weights based on each subband With a set weight threshold, the system adaptively filters out subbands with high weights and then weights the remaining subbands with high weights.

[0116] To address the core issues of artificial energy thresholds being highly subjective and prone to accidental deletion of effective high-frequency subbands, a weighted threshold optimized through a validation set is set. ,default ,then At this time, 30% to 50% of invalid subbands can be removed, reducing the computing power requirements for subsequent reconstruction and realizing intelligent subband selection. The selection formula is as follows:

[0117] (16)

[0118] in, Represented as the first Layer The coefficients of each screening subband The coefficients of invalid subbands are set to zero to avoid interference with the reconstructed signal; It is represented as the soft threshold denoising coefficient of the subband.

[0119] Traditional WPT signal reconstruction only performs an inverse transform on the selected subband coefficients, without distinguishing the subband localization contribution. This invention introduces a subband localization contribution weight. This allows effective subbands with higher weights to contribute more to the reconstruction, further improving the signal-to-noise ratio of the output features. The process is as follows:

[0120] Weighted subband coefficient calculation; the filtered subband coefficients With the corresponding attention weights Multiplying these values ​​yields the weighting coefficients for the subbands, highlighting the characteristic contributions of the effective subbands. The formula is as follows:

[0121] (17)

[0122] in, Represented as the first Layer The weighting coefficients of each sub-band;

[0123] Step 1.2.4: Reconstruct the weighted subbands to obtain the optimized feature set;

[0124] Based on Mallat inverse transform logic, from the highest decomposition level... The signal is reconstructed layer by layer down to the original scale, layer 0, with the same length as the original signal. Weighted subband coefficients are incorporated during the reconstruction process. The formula is as follows:

[0125] (18)

[0126] in, Represented as the first The reconstruction approximation coefficients of the layer This reflects the low-frequency effective characteristics of this layer, ultimately This is the reconstructed signal at the original scale; The scaling filter coefficients, represented as db4 wavelet packets, are used for the reconstruction of effective low-frequency features; The wavelet filter coefficients, represented as db4 wavelet packets, are used for the reconstruction of effective high-frequency features; Represented as the first Layer Each translation term with weighted coefficients, This is represented as a translation index entry; because the db4 basis functions satisfy orthogonality, there is no need to invert the filter coefficients; direct substitution into the calculation ensures that the reconstructed signal is distortion-free and its length is consistent with the original signal; the length is... .

[0127] Step 1.3: Based on the initial distance between the UWB tag and each UWB base station, extract the propagation time between the UWB tag and each UWB base station from the optimized feature set, calculate the optimized distance through the propagation time, and then solve for the optimized positioning through the optimized distance;

[0128] Step 2.2 is performed iteratively on all dimensional features of the original UWB signal to obtain the attention-optimized feature set. :

[0129] (19)

[0130] Among them, the Feature vector at each observation time ; Represents the total feature dimension ( For the number of base stations, each of the three feature dimensions indivual); This represents the number of sampling times. The optimized feature vector at each time step. Composed of denoised multi-base station feature values, only the propagation time between UWB tags and each UWB base station is extracted. , .

[0131] After IWPT denoising, the corresponding optimized distance is obtained:

[0132] (20)

[0133] Based on the optimized distance, a system of equations is established using a trilateration algorithm to solve for the optimized positioning coordinates. For each UWB base station, the equations for trilateration are:

[0134] (twenty one)

[0135] The optimized label location can be obtained through linear elimination:

[0136] (twenty two)

[0137] in, , , They are the first The coordinates of each UWB base station.

[0138] IWPT achieves denoising and effective feature signal frequency band extraction of UWB positioning signals through multi-scale decomposition and threshold filtering, significantly improving signal quality. However, due to the inherent nature of static signal processing, it suffers from inherent limitations such as missing temporal correlation capture, mapping discontinuities between effective features and label positioning coordinates, and insufficient adaptability to dynamic environments. In contrast, BI-LSTM can reconstruct the spatiotemporal dependence of signals through forward and backward dual units, construct nonlinear mappings from features to positioning coordinates based on gating mechanisms, and enhance dynamic anti-interference capabilities through data-driven capabilities. It specifically compensates for the limitations of IWPT in processing UWB signals. The two work together to form a preprocessing-deep modeling closed loop, which is the key to improving the accuracy and robustness of UWB positioning. The optimized positioning coordinates, which initially filter noise, provide high-quality input data for the subsequent BI-LSTM model.

[0139] Step 2: Use the BI-LSTM network to perform temporal correlation and dynamic error correction on the effective feature signals extracted in Step 1 to obtain the predicted localization;

[0140] Step 2.1: Preprocess the optimized feature sequences in the optimized feature set; the preprocessing includes normalization and construction of time series data samples;

[0141] Optimized feature set output by IWPT It requires two steps: normalization and time series sample construction, to convert it into an input format that BI-LSTM can accept, thus avoiding differences in feature magnitude and loss of time series information.

[0142] The significant differences in the magnitudes of the features necessitate Min-Max normalization to map them to the [0,1] interval. The formula is as follows:

[0143] (twenty three)

[0144] in, No. Time of the first The normalized values ​​of each feature (output, range [0,1]) eliminate the bias of magnitude difference on model training; No. Time of the first IWPT optimization values ​​for each feature; and They are the first One feature in all The minimum and maximum values ​​at each time point are used to ensure that all feature magnitudes are uniform after normalization.

[0145] The normalized feature set is: , as input for constructing time series samples.

[0146] Using the sliding window method This is converted into time-series data samples that serve as input to the BI-LSTM. Each time series data sample is defined as:

[0147] (twenty four)

[0148] in, Represented as the first A time series data sample, containing A series of consecutive moments Dimensional features, For data sample index, the range of values ​​is... , co-generated One data sample; Represented as continuous Normalized features at each moment; This is represented as the dimension of the data sample.

[0149] Simultaneously build a tag set , No. The true label of a moment is defined as: ;in, It is the first The actual planar coordinates of the target at any given time; They are the first The time-point labels are located at their actual positions on the x, y, and z axes, serving as supervision signals for model training.

[0150] Step 2.2: Using the preprocessed data as input data for the BI-LSTM model, the BI-LSTM network model is used to extract core temporal correlation features and fuse the extracted core temporal correlation features. Finally, the fused features are mapped to 3D positioning coordinates.

[0151] Bi-LSTM is derived from forward LSTM (from Time (capturing "past → present" temporal dependencies) and backward LSTM (from The time frame is composed of two parts (capturing the "current → past" time sequence dependency). The outputs of the two parts are concatenated and mapped to positioning coordinates through a fully connected layer. At the same time, all variable dimensions are consistent with the previous description.

[0152] LSTM controls the flow of information through a "forget gate - input gate - cell state - output gate". time steps ( Corresponding time series samples The The gating formula for (at each time point) is:

[0153] Forget Gate (Filtering the percentage of historical information retained):

[0154] (25)

[0155] Represented as the first The output of the time step forget gate has a value of , Indicates complete forgetting. This indicates that past information is completely preserved; This represents the sigmoid activation function ( ), map the output to Simulates the gating switch effect; This represents the forget gate weight matrix, with dimensions... Among them, the former The column corresponds to the weight of the historical hidden state, and then... The column corresponds to the weight of the current input feature; Indicates the first Hidden state (dimension) of a time-step forward LSTM Before storage The temporal characteristics at each moment; Indicates the first Normalization characteristics of time steps ( ), from time series samples The OK; This represents the concatenation operation between the hidden state and the input features (dimension). This allows for the integration of historical information with current characteristics. It is represented as the forget gate bias vector, which adjusts the baseline value of the gated output.

[0156] The input gate consists of two steps: "information filtering" and "candidate information generation," with the following formulas:

[0157] (26)

[0158] (27)

[0159] Indicates the first Time step input gate output (dimension) ), with a value of , Indicates no update. This indicates a complete update of the current information; Represented as the input gate weight matrix (dimension) ); Represented as the input gate bias vector; Indicates the first The candidate cell states at each time step store new feature information for the current time step; Represents the hyperbolic tangent activation function ( Output range This avoids gradient explosion caused by excessively large eigenvalues; Represents the candidate cell state weight matrix; This represents the candidate cell state bias vector.

[0160] Cellular state (the long-term memory of LSTM) is updated collaboratively with the input gate through the forget gate, as shown in the formula:

[0161] (28)

[0162] Indicates the first Cell state at time step ( ), storing from 1 to Key timing information for each time step; Indicates the first Cellular states at time steps, i.e., historical long-term memory; first through Preserve some historical memories, and then through Add some new information to achieve iterative updates of memory.

[0163] The output gate controls which information from the cell state is passed to the hidden state (LSTM's "short-term memory"), and the formula is:

[0164] (29)

[0165] (30)

[0166] Represented as the first Time step output gate output (dimension) ), range of values Control the output ratio of cell state; Represented as the input gate weight matrix (dimension) ); This represents the input gate bias vector; Indicates the first The hidden state of the forward LSTM at time step, i.e., the hidden state of the forward LSTM at the time step. The output of the time step includes historical time series features; This indicates mapping cell state values ​​to... To avoid excessively large output values, a forward LSTM is used to traverse the time series samples. of After each time step, output the final forward hidden state. This represents the temporal sequence from history to the present.

[0167] Backward LSTM and forward LSTM exhibit mirror-symmetric characteristics in structure; their parameter configurations are completely identical, differing only in the direction of information transmission. For example... Figure 5 As shown, the forward LSTM propagates information in forward chronological order, while the backward LSTM operates in reverse chronological order, starting from the first... Time Steps Information is transmitted through time steps. This backward propagation mechanism aims to uncover the temporal dependencies between future and current moments in a data sequence, effectively capturing dynamic features such as the inertial trend of target motion. Define the backward time step index. This index is related to the forward time step. A one-to-one correspondence is established, and the gating calculation formula is as follows:

[0168] Forgotten Gate:

[0169] (31)

[0170] Input Gate:

[0171] (32)

[0172] (33)

[0173] Cell status update:

[0174] (34)

[0175] Output gate:

[0176] (35)

[0177] (36)

[0178] Indicates the parameters of the backward LSTM; Indicates the first The hidden state of the backward LSTM at the next time step; Indicates backward The input features at each time step correspond to the forward feed. time step This ensures that the temporal direction of the input features is opposite; , , The forward LSTM is defined the same as the forward LSTM, only the timing direction is reversed; the backward LSTM traverses... After one time step, output the final backward hidden state. , representing the temporal characteristics of the future → present.

[0179] The core advantage of BI-LSTM is that it integrates forward and backward temporal features, achieved through hidden state concatenation:

[0180] (37)

[0181] Indicates the first Time series samples BI-LSTM fusion features; This represents the concatenation operation between the forward and backward hidden states, while simultaneously doubling the feature dimension (dimension). ).

[0182] To map the fused features to 3D localization coordinates, a fully connected layer is added:

[0183] (38)

[0184] Indicates the first Predicted location coordinates at any given time; Indicates the first Moment tags Predicted position of the axis; Indicates the first Moment tags Predicted position of the axis; Indicates the first Moment tags Predicted position of the axis; This represents the weight matrix of the fully connected layer; This represents the bias vector of the fully connected layer.

[0185] Step 2.3: Through the training process of the BI-LSTM model, dynamic error correction of 3D positioning coordinates is achieved to obtain predicted positioning;

[0186] The goal of model training is to optimize network parameters. Minimize the predicted value Compared with the true value By minimizing the deviations, the model with the best generalization ability is finally obtained.

[0187] Since the positioning coordinates are continuous values, the mean squared error (MSE) is used to measure the prediction deviation, and the loss function is:

[0188] (39)

[0189] This represents the overall training loss of the model; a smaller value indicates higher prediction accuracy. express The square error in the axial direction reflects Axis prediction deviation; Show The square error in the axial direction reflects Axis prediction deviation; Show The square error in the axial direction reflects Axis prediction deviation; for all The deviations of each sample are summed and then averaged to avoid the sample size affecting the loss value.

[0190] The Adam optimizer is used for adaptive moment estimation, and the output minimizes the loss function. Its advantage lies in avoiding the slow convergence and local optima problems of traditional SGD (Stochastic Gradient Descent) through adaptive learning rates. Taking the forget gate weight matrix as an example, the parameter update formula is:

[0191] First-order moment momentum estimation:

[0192] (40)

[0193] Second-order moment variance estimation:

[0194] (41)

[0195] Bias correction to eliminate initial moment estimation bias:

[0196] (42)

[0197] Parameter update:

[0198] (43)

[0199] This represents the training epoch index, and the initial value. ; Indicates the first Rounds The first moment estimate, This represents the exponential decay rate of the first moment; Indicates the first Rounds The second moment estimate, This represents the exponential decay rate of the second moment; Represented as loss function right The gradient reflects The extent of the impact on the loss; , These represent the deviation correction terms for the first and second moments, respectively, both of which are 0 at the initial time. This represents the initial learning rate. Representing minute values ​​can avoid the need for fractions in the denominator. This leads to data anomalies; Indicates the first After the round update Other network parameters The update formula is exactly the same; only the parameter symbols need to be replaced.

[0200] Step 3: Subtract the optimized positioning obtained from the initial noise filtering from the predicted positioning to obtain the prediction residual. Correct the predicted positioning using the prediction residual to obtain the optimal predicted positioning coordinates.

[0201] After obtaining the optimal model's prediction results for the test set samples, the following measures are taken to further improve the positioning accuracy:

[0202] Define the BI-LSTM prediction residual as:

[0203] (44)

[0204] The final positioning result is obtained by correcting the predicted residuals.

[0205] (45)

[0206] (46)

[0207] in, It is the residual correction factor, set The weighting ratio between the predicted residual and the predicted location is balanced.

[0208] It should be understood that, inspired by the technical concept of this invention, those skilled in the art can make various improvements or modifications based on the above content without departing from the scope of this invention, and these modifications still fall within the protection scope of this invention.

Claims

1. A method for optimizing UWB positioning accuracy based on IWPT-BILSTM, characterized in that, The method includes the following steps: Step 1: The original UWB signal is decomposed and denoised using the improved WPT, i.e., IWPT, to extract effective feature signals that are strongly correlated with the positioning, and the optimized positioning is solved at the same time. Step 2: Use the BI-LSTM network to perform temporal correlation and dynamic error correction on the effective feature signals extracted in Step 1 to obtain the predicted localization; Step 3: Subtract the optimized positioning from the predicted positioning to obtain the prediction residual. Correct the predicted positioning using the prediction residual to obtain the optimal predicted positioning.

2. The UWB positioning accuracy optimization method according to claim 1, characterized in that, IWPT is obtained by introducing a lightweight attention module into the subband selection and denoising stage of traditional WPT.

3. The UWB positioning accuracy optimization method according to claim 2, characterized in that, The lightweight attention module is a fully connected neural network architecture containing only a 2D input layer, a single hidden layer, and an output layer, with its number of parameters strictly controlled to within 65.

4. The UWB positioning accuracy optimization method according to claim 3, characterized in that, Step 1 further includes the following steps: Step 1.1: Calculate the initial distance between the UWB tag and each UWB base station; Step 1.2: Use IWPT to perform multi-scale decomposition and denoising on the original UWB signal, extract effective feature signals that are strongly correlated with the positioning, and reconstruct the effective feature signals to obtain an optimized feature set; Step 1.3: Based on the initial distance between the UWB tag and each UWB base station, extract the propagation time between the UWB tag and each UWB base station from the optimized feature set, calculate the optimized distance through the propagation time, and then solve for the optimized positioning through the optimized distance.

5. The UWB positioning accuracy optimization method according to claim 3, characterized in that, Step 1.2 further includes the following steps: Step 1.2.1: Select WPT basis functions that are adapted to the characteristics of the original UWB signal; Step 1.2.2: Perform multi-scale frequency band decomposition on the original UWB signal based on the selected WPT basis functions to obtain multiple sub-bands; Step 1.2.3: Extract the features of the decomposed subbands, introduce a lightweight attention module to select the subbands that contribute more to the localization from the decomposed subbands and perform weighted processing on them; Step 1.2.4: Reconstruct the weighted subbands to obtain the optimized feature set.

6. The UWB positioning accuracy optimization method according to claim 5, characterized in that, Step 1.2.3 further includes the following steps: Step 1.2.3.1: Extract the features of the decomposed subbands, including the energy of each subband. Pearson correlation coefficients between each sub-band and the actual location tag ; Step 1.2.3.2: Based on the features of the decomposed sub-bands, a lightweight attention module is introduced to calculate the attention weights of each sub-band; Step 1.2.3.3: Based on the attention weight of each subband and the set weight threshold, adaptively filter out the subbands with high attention weights and weight them.

7. The UWB positioning accuracy optimization method according to claim 6, characterized in that, In step 1.2.3.3, the selected subbands with high attention weights are weighted according to equations (16) and (17): (16) (17) in Represented as the first Layer The coefficients of each screening subband The coefficients of invalid subbands are set to zero to avoid interference with the reconstructed signal; This indicates the set weight threshold; Represented as the soft thresholding denoising coefficient of the sub-band; Indicates the first Layer Attention weights for each selected subband; Represented as the first Layer The weighting coefficients of each sub-band.

8. The UWB positioning accuracy optimization method according to claim 1, characterized in that, Step 2 further includes the following steps: Step 2.1: Preprocess the optimized feature sequences in the optimized feature set, including normalization and construction of time series data samples; Step 2.2: Using the preprocessed data as input data for the BI-LSTM model, the BI-LSTM network model is used to extract core temporal correlation features and fuse the extracted core temporal correlation features. Finally, the fused features are mapped to 3D positioning coordinates. Step 2.3: Perform dynamic error correction on the 3D positioning coordinates using the BI-LSTM model to obtain the predicted positioning.

Images