UWB channel environment state identification method under limited mark data

Abstract

The invention provides a UWB channel environment state identification method under limited mark data. The method comprises the following steps of: selecting UWB (Ultra Wide Band, Ultra Wide Band) CIR (Channel Impulse Response) data fragments, and converting UWB channel impulse response data into a two-dimensional image through continuous wavelet transform and Gramer angle field transform; the UWB channel state identification method comprises the following steps: constructing a hybrid network fusing a Vision Transform, a Wide Residual Network, a Wide ResNet and a channel attention mechanism compression and excitation (SE) module, inputting a two-dimensional image into the hybrid network, and combining the hybrid network with a semi-supervised learning algorithm SoftMatch to obtain a UWB channel state classification identification result under limited tag data. According to the method, the semi-supervised learning algorithm is applied to the UWB CIR signal-based channel state recognition task, so that the accuracy of channel state recognition can be remarkably improved under the condition of scarcity of marked data, and meanwhile, the time and economic cost required by manual marking are greatly reduced.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision and wireless positioning technology, and in particular to a method for identifying a UWB channel environment state under limited marking data. Background Art

[0002] With the rapid development of wireless positioning technology for products in the information age, location-based services have become an important function in many end-user applications, including military, civil, disaster relief and peacekeeping missions, in recent years.

[0003] In exposed space, i.e., areas without obstructions from buildings or other objects, global satellite navigation and positioning systems, such as the Beidou satellite navigation system and the global positioning system, can achieve high-precision positioning in an open and unobstructed environment through four or more satellites with known positions. Due to environmental structures, obstructions, electromagnetic interference and other reasons, in non-exposed space, i.e., indoor and semi-indoor environments, accurate location information cannot be obtained through traditional satellite navigation systems or other conventional location services. Currently, technologies such as Bluetooth, wireless fidelity, ultra-wideband (UWB), ZigBee, wireless radio frequency identification and pseudo-satellite have been widely used in non-exposed space. Among them, UWB has been widely used in non-exposed spaces such as urban rail transit systems, tunnels, coal mines, hospitals, shopping malls and factories to achieve accurate positioning based on its advantages of strong multipath resolution and high distance resolution.

[0004] Although the theoretical positioning accuracy of UWB can reach the centimeter level, the actual positioning accuracy varies greatly under different complex indoor environmental conditions. In order to achieve accurate positioning based on UWB technology, it is usually necessary to first consider the recognition of UWB signals. Currently, common UWB signal recognition methods are mainly divided into three categories: distance-based methods, location-based methods, and channel-based methods. The distance-based method is achieved through the statistical difference of distance information and has a strong dependence on the environment; the location-based method is achieved through the difference between location information and has poor universality; the channel-based method mainly uses the channel impulse response, which can reflect the environment of the propagation path and is a key parameter to characterize the impact of the channel on the signal. In recent years, based on the rapid development of high-performance computing hardware, deep learning has become one of the most influential methodologies in the current field of artificial intelligence.

[0005] Since there are large differences between CIR waveforms in different environments, once the environment is modified, sample collection and labeling need to be performed again. At present, the disadvantages of the channel state identification method in the prior art include: the available label data is relatively scarce. Summary of the invention

[0006] The embodiment of the present invention provides a method for identifying the UWB channel environment state under limited label data, so as to effectively identify the UWB channel environment state under limited label data.

[0007] In order to achieve the above object, the present invention adopts the following technical scheme.

[0008] A method for identifying an ultra-wideband (UWB) channel environment state under limited labeled data, comprising:

[0009] Selecting an ultra-wideband UWB channel impulse response CIR data segment, and converting the UWB channel impulse response data into a two-dimensional image by continuous wavelet transform and Gram angular field transform;

[0010] Construct a hybrid network that integrates a visual converter, a wide residual network, and a channel attention mechanism, and train the hybrid network using labeled data and unlabeled data through a joint supervised learning and unsupervised learning strategy, that is, introduce a semi-supervised learning algorithm SoftMatch to optimize the training of the hybrid network;

[0011] The two-dimensional image is input into a hybrid network, and the hybrid network is combined with a SoftMatch algorithm to obtain a UWB channel state classification and recognition result under limited labeled data.

[0012] Preferably, the selecting of UWB CIR data segments and converting the UWB channel impulse response data into a two-dimensional image by continuous wavelet transform and Gram angular field transform comprises:

[0013] Obtaining a log file of CIR data generated by a UWB device, preprocessing the log file, and extracting a specific segment of the one-dimensional time series CIR;

[0014] A two-dimensional time-frequency domain image is generated by using a specific segment of the one-dimensional time series CIR through continuous wavelet transform to obtain a synchronous representation of the CIR signal in the time and frequency dimensions. The generalized Morse wavelet is selected as the basis function, and the Fourier transform of the generalized Morse wavelet is:

[0015]

[0016] Among them, U(ω) is a unit step function, β and γ are parameters that control the attenuation of the wavelet in the time domain and frequency domain, and determine the wavelet waveform;

[0017] The formula for continuous wavelet transform is defined as follows:

[0018]

[0019] Where x(t) is the original one-dimensional CIR signal, τ is the translation factor, a is the scale factor, Ψ* (t) is the complex conjugate of the basis function, WT(a,τ) is the wavelet transform value, and RGB color mapping is performed on the wavelet transform value WT(a,τ) to generate a two-dimensional time-frequency domain image;

[0020] The specific segment of the one-dimensional time series CIR is converted and encoded into a two-dimensional image by Gram angular field transform. n ] is normalized and scaled to between 0 and 1, as follows:

[0021]

[0022] Among them, x i is a data point in the original CIR time series.

[0023] Convert the normalized signal to polar coordinates:

[0024]

[0025] Among them, Φ i is the polar angle after the arc cosine function transformation, r i is the polar radius, i is the timestamp, and n is a coefficient used to adjust the coverage of the polar coordinate system.

[0026] The Gram and angle field is selected and calculated using the phase angle and phase angle of the time series points after polar coordinate transformation. The specific formula is as follows:

[0027]

[0028] Use Python programming to splice the two-dimensional time-frequency domain image obtained by continuous wavelet transform and the two-dimensional image obtained by Gram angular field transform up and down to obtain a two-dimensional spliced ​​image.

[0029] Preferably, the construction of a hybrid network integrating a visual converter, a wide residual network and a channel attention mechanism comprises:

[0030] Determine the parameters of the ViT network. The ViT network consists of an image block layer, a position embedding layer, and a Transformer encoder layer.

[0031] Determine the WideResNet network type and parameters, determine the SE module parameters, and add the SE module to the three residual blocks of the WideResNet network structure;

[0032] The ViT network and the WideResNet network mixed with the SE module are fused to obtain a hybrid network.

[0033] Preferably, the hybrid network is trained using labeled data and unlabeled data by combining supervised learning and unsupervised learning strategies, that is, a semi-supervised learning algorithm SoftMatch is introduced to optimize the training of the hybrid network, including:

[0034] (1) Initialization: The transformation for labeled data includes weak data augmentation and normalization, and the transformation for unlabeled data includes strong and weak data augmentation and normalization; weak data augmentation uses padding random cropping and horizontal flipping with a probability of 50%, and strong data augmentation uses padding random cropping, horizontal flipping with a probability of 50%, and random augmentation. The total number of labeled data is set to 180, the number of data categories is set to 3, and the unlabeled data prediction results use hard labels; set distribution alignment to adjust the predicted probability distribution of unlabeled data to make the predicted probability distribution closer to uniform distribution; the unlabeled data prediction process uses EMA, and the weight value is set to 0.999 to capture the model's prediction trend for unlabeled data in the long term; the parameter μ t and σ t The weight value when using EMA is set to 0.999, and the batch size of unlabeled data during training is twice that of labeled data;

[0035] (2) Training: mean μ of weighted function λ(p) t The variance σ increases t Reduce, dynamically adjust the sample weight distribution, and learn the current underfitting category features through UA operation.

[0036] Preferably, the two-dimensional image is input into the hybrid network, and the hybrid network is combined with the semi-supervised learning SoftMatch algorithm to obtain the UWB channel state classification and recognition result under limited labeled data, including:

[0037] The loss function and optimization strategy of the SoftMatch algorithm are set as follows:

[0038] The loss function of SoftMatch is composed of supervised loss With unsupervised loss Composition, the labeled dataset is The unlabeled dataset is The specific calculation formula is as follows:

[0039]

[0040] Where p is p(y|ω(x u )) abbreviation, is the one-hot encoding value argmax(p) of the pseudo label, λ(p) is the sample weighting function, B L is the batch size of labeled data, B Uis the batch size of unlabeled data, Ω(x u ) is strong data enhancement, ω(x u ) is weak data enhancement;

[0041] The sample weighting function λ(p i ) is calculated as follows:

[0042] The truncated Gaussian distribution is selected as the weighting function λ(p), and the specific calculation formula is as follows:

[0043]

[0044] The parameter μ is calculated based on the model's past prediction output for unlabeled data t and σ t , the calculation formula for the tth iteration is as follows:

[0045]

[0046] Based on the exponential moving average of historical batch data, a more reliable parameter μ is obtained t and σ t :

[0047]

[0048] in, The value is 1 / C, The value is 1.0 and m is the weight.

[0049] In order to solve the problem of unbalanced distribution of pseudo-label categories, uniform alignment is used to force the model to cover each category more balanced when generating pseudo-labels. The specific formula is as follows:

[0050]

[0051] Where u(C) is a uniform distribution and Normalize() is a normalization operation.

[0052] The two-dimensional image is input into the hybrid network, the WideResNet network learns local area features, the SE module learns the importance weights between channels, the WideResNet network of the hybrid SE module outputs two-dimensional CIR hybrid image features, the ViT network learns the relationship between global features and features, and the ViT network outputs two-dimensional CIR hybrid image features;

[0053] The two-dimensional CIR mixed image features output by the WideResNet network of the ViT network and the hybrid SE module are spliced ​​to generate a feature vector of the two-dimensional image. The UWB channel state classification and recognition result is obtained using a fully connected layer according to the feature vector of the two-dimensional image. For the labeled data: is the UWB channel state classification and recognition result obtained by hybrid network prediction after weak data enhancement of the spliced ​​image; for unlabeled data: It is the UWB channel status classification and recognition result obtained by hybrid network prediction after strong data enhancement of the spliced ​​image; It is the pseudo label obtained by predicting the hybrid network after weak data enhancement of the spliced ​​image.

[0054] It can be seen from the technical solution provided by the above-mentioned embodiments of the present invention that the method of the present invention applies the semi-supervised learning algorithm to the channel state recognition task based on UWB CIR signals, which can significantly improve the accuracy of channel state recognition when labeled data is scarce, while greatly reducing the time and economic cost required for manual labeling.

[0055] Additional aspects and advantages of the present invention will be given in part in the following description, which will become obvious from the following description, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0057] Figure 1 A processing flow chart of a method for identifying a UWB channel environment state under limited label data provided by an embodiment of the present invention;

[0058] Figure 2 The CIR original signal waveform diagram of three propagation conditions LOS, DP_NLOS, and NDP_NLOS provided in the embodiment of the present invention;

[0059] Figure 3 Continuous wavelet transform images of CIR signals of three propagation conditions LOS, DP_NLOS, and NDP_NLOS provided by the embodiment of the present invention;

[0060] Figure 4 Gram angular field transformation images of CIR signals under three propagation conditions, LOS, DP_NLOS and NDP_NLOS, provided in an embodiment of the present invention;

[0061] Figure 5 A diagram of the CIR feature extraction and classification network architecture based on a fusion visual converter, a wide residual network, and a channel attention mechanism provided in an embodiment of the present invention;

[0062] Figure 6 A simplified diagram of the processing flow of the semi-supervised learning algorithm SoftMatch for unlabeled data provided by an embodiment of the present invention;

[0063] Figure 7 The present invention provides an overall detailed flow chart of a method for identifying UWB channel environment status based on limited-mark CIR data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0065] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or coupling. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items.

[0066] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0067] To facilitate understanding of the embodiments of the present invention, several specific embodiments will be further explained below with reference to the accompanying drawings, and each embodiment does not constitute a limitation on the embodiments of the present invention.

[0068] The present invention selects CIR signals as the basis for judging the channel environment status in complex environments, and uses a network that integrates visual converters, wide residual networks and channel attention mechanisms to achieve feature extraction and classification.

[0069] Example 1

[0070] The processing flow of a method for identifying UWB channel environment status under limited label data provided by an embodiment of the present invention is as follows: Figure 1 As shown, the following processing steps are included:

[0071] Step S1, selecting a UWB channel impulse response data segment of appropriate length, preprocessing the UWB channel impulse response data segment, and converting the UWB channel impulse response data segment into a two-dimensional image through continuous wavelet transform and Gram angular field transform;

[0072] Step S2, constructing a hybrid network integrating a visual converter, a wide residual network and a channel attention mechanism, and training the hybrid network using labeled data and unlabeled data by combining supervised learning and unsupervised learning strategies, that is, introducing a semi-supervised learning algorithm SoftMatch to optimize the training of the hybrid network;

[0073] Step S3, inputting the two-dimensional image into a hybrid network, and combining the hybrid network with a SoftMatch algorithm to obtain a UWB channel state classification and recognition result under limited labeled data.

[0074] Wherein, step S1 specifically includes the following steps:

[0075] S1.1, obtain channel impulse response data and perform preprocessing;

[0076] The specific process includes the following steps:

[0077] S1.1.1, read the log file: based on the log file generated by the UWB device, read 1×1016 CIR complex values ​​in each data and the first path peak index HLP obtained by using the leading edge detection algorithm;

[0078] S1.1.2, according to whether there is a direct line of sight path between the two ends of the signal transmission and reception, the propagation conditions of the CIR signal are divided into three categories: Line of sight (LOS), Direct Path-Non line of sight (DP-NLOS), Non Direct Path-Non line of sight (NDP-NLOS). The direct path component of the LOS condition is clear and can be used directly; the DP_NLOS condition refers to a slight non-line of sight situation, in which the moderate deviation value generated can be subsequently alleviated and used for positioning; the direct path component of the NDP_NLOS condition cannot be correctly received and analyzed. At this time, the direct path component identified by the receiver is actually one of the reflected multipath components, and the introduction of a large delay of several nanoseconds will result in an erroneous distance of meters, which is a serious NLOS situation;

[0079] S1.1.3, select CIR segment: extract a segment of length 1×128 from the original signal of length 1×1016. The segment is based on the HLP index as the reference point, and selects the 5 values ​​before the index and the 122 consecutive values ​​after the index as the initial input segment. The original signal waveforms of the CIR segment under three propagation conditions are as follows: Figure 2 As shown;

[0080] S1.2, for the extracted one-dimensional time series CIR specific segment, continuous wavelet transform and Gram angular field transform are used to transform it into a more informative two-dimensional image representation, which specifically includes the following steps:

[0081] S1.2.1, continuous wavelet transform:

[0082] Continuous wavelet transform uses one-dimensional time series CIR to generate two-dimensional time-frequency domain images to obtain synchronous representation of CIR signals in time and frequency dimensions. The continuous wavelet transform images of CIR signals under three propagation conditions are shown in the figure below. Figure 3 As shown. The fully analytical generalized Morse wavelet with high time-frequency resolution is selected as the basis function, and its Fourier transform is:

[0083]

[0084] Among them, U(ω) is the unit step function, β and γ are parameters that control the attenuation of the wavelet in the time domain and frequency domain, and determine the wavelet waveform.

[0085] The basic idea of ​​wavelet transform is to project the signal to be analyzed onto a set of wavelet basis functions to obtain the frequency components of the signal at different scales. The formula of continuous wavelet transform is defined as follows:

[0086]

[0087] Where x(t) is the original one-dimensional CIR signal, τ is the translation factor, a is the scale factor, Ψ * (t) is the complex conjugate of the basis function. Formula (2) is the formula for continuous wavelet transform. By calculating different translation factors and scale factors, the wavelet transform value WT(a,τ) corresponding to each pixel in the image is obtained. The wavelet transform value WT(a,τ) is subjected to RGB color mapping to generate a two-dimensional time-frequency domain image.

[0088] S1.2.2, Gram angular field transform:

[0089] The Gram angular field transform converts and encodes a one-dimensional time series CIR fragment into a feature-rich two-dimensional image, where the two-dimensional image is an image of the correlation between CIR sequence points generated by the Gram angular field transform of the one-dimensional time series CIR fragment; the two-dimensional time-frequency domain image is an image with the horizontal axis being time and the vertical axis being frequency generated by the continuous wavelet transform of the one-dimensional time series CIR fragment. The present invention believes that the Gram angular field transform image data can further supplement the correlation of the one-dimensional time series CIR fragment itself on the basis of using the continuous wavelet transform image data, and the combined use enables the model to learn a more comprehensive feature representation.

[0090] The Gram angular field transform can alleviate the dependency of CIR data on time series. The Gram angular field transform images of CIR signals under three propagation conditions are shown in Figure 2. Figure 4 shown.

[0091] First, we need to analyze the CIR data X=[x0,x1,x2,…x n ] is normalized and scaled to between 0 and 1, as follows:

[0092]

[0093] Among them, x i is a data point in the original CIR time series.

[0094] Convert the normalized signal to polar coordinates:

[0095]

[0096] Among them, Φ i is the polar angle after the arc cosine function transformation, r i is the polar radius, i is the timestamp, and n is a coefficient used to adjust the coverage of the polar coordinate system.

[0097] The Gram and angle field is selected and calculated using the phase angle and angle of the time series points after polar coordinate transformation. The specific formula is as follows:

[0098]

[0099] Python programming is used to splice the two-dimensional time-frequency domain image obtained by continuous wavelet transform and the two-dimensional image obtained by Gram angular field transform up and down to obtain a two-dimensional spliced ​​image corresponding to the specific segment of the above one-dimensional time series CIR.

[0100] like Figure 5 As shown in the figure, the CIR feature extraction and classification network architecture diagram based on ViT, WideResNet and channel attention mechanism, that is, step S2 and step S3 specifically include the following steps:

[0101] S2.1, 2D CIR mixed image feature extraction based on ViT network:

[0102] Determine the parameters of the ViT network; the ViT network is a Transformer network dedicated to computer vision tasks. The ViT network mainly consists of an image block layer, a position embedding layer, and a Transformer encoder layer.

[0103] The image segmentation layer divides the input image into fixed-size image blocks and projects them into a high-dimensional space, adds classification tags, and then flattens them into a one-dimensional vector, which not only reduces the computational complexity but also enhances the expressiveness of the model. The position embedding layer adds position embedding to each image block embedding to preserve the spatial structure information of the image. The Transformer encoder layer uses a multi-head self-attention mechanism to capture the long-range dependencies between the various parts of the image, and uses a feedforward neural network to deeply process and transform the features of each position.

[0104] The first part of the ViT network uses a 2×2 two-dimensional convolution kernel Conv2d, a step size of 2, 192 filters, no padding, and a splicing size of 1×192 classification tags. The second part uses a learnable position embedding matrix of size 257×192. The third part consists of 12 identical Transformer encoders; in the Transformer encoder, the input features are first layer normalized and then passed into the MHA module. MHA uses 3 parallel attention heads, and the feature dimension of each head is 64; LN is added to the front end of the FFN of each Transformer encoder. FFN uses a two-layer fully connected network: the first layer has 192 input units, 768 output units, and uses a Gaussian error linear activation function; the second layer has 768 input units and 192 output units. The output features of the Transformer encoder layer pass through LN to extract the classification tag features of size 1×192.

[0105] S2.2, 2D CIR mixed image feature extraction based on WideResNet network:

[0106] Determine the WideResNet network type and parameters; WideResNet increases the network width based on the residual network and is a special architecture of the CNN network. WideResNet is mainly composed of two-dimensional convolutional layers and pooling layers.

[0107] Among them, the two-dimensional convolution layer extracts the local time-frequency features of the CWT image and the local correlation features of the GAF image. The two-dimensional convolution formula is as follows:

[0108]

[0109] O(i,j) represents the convolution result at position (i,j), K is the convolution kernel used for convolution with the input matrix, I is the feature map that needs to be convolved, and * is the convolution operation.

[0110] The two-dimensional pooling layer uses average pooling to extract global features while improving the translation invariance of the model. The residual network alleviates the gradient vanishing or exploding problem through jump connections, while extracting more complex feature representations; increasing the network width can reduce the dependence on network depth.

[0111] The WideResNet network is mainly composed of three residual blocks. Before entering the residual block, a 3×3 two-dimensional convolution kernel Conv2d is used, with a step size of 1, a filter number of 16, and a padding of 1. The first residual block is batch normalized, and then the rectified linear activation function is input, and then a 3×3 two-dimensional convolution kernel Conv2d is used, with a step size of 1, a filter number of 32, and a padding of 1. BN and RELU are added later, and the convolution kernel with the same configuration as above is used again; the residual connection is to match the output dimension of the main path, and a 1×1 two-dimensional convolution kernel Conv2d is used, with a step size of 1, a filter number of 32, and no padding is used. The second residual block uses a 3×3 2D convolution kernel Conv2d with a stride of 2, 64 filters, and 1 padding. BN and RELU are added later, and the same convolution kernel is used again. The residual connection uses a 1×1 2D convolution kernel Conv2d with a stride of 2, 64 filters, and no padding to match the output dimension of the main path. The overall structure of the third residual block is the same as the second residual block, but the third residual block uses a 3×3 2D convolution kernel Conv2d with a stride of 2, 128 filters, and 1 padding. The residual connection uses a 1×1 2D convolution kernel Conv2d with a stride of 2, 128 filters, and no padding. BN and RELU are applied to the output features of the three residual blocks, and then a 2D average pooling layer is used to average pool each channel of the 8×8 input feature map into 1×1 units.

[0112] S2.3, 2D CIR mixed image feature extraction based on channel attention mechanism compression and excitation module:

[0113] Determine the parameters of the SE module; the SE module is mainly composed of a two-dimensional average pooling layer and a fully connected layer, and enhances the expressiveness of features by learning the importance weights between channels; in order to achieve integration with the WideResNet network, the SE module is integrated into the residual block of the WideResNet network.

[0114] The SE module is mainly composed of four parts. The first part is a two-dimensional average pooling layer, which pools each channel of the input feature map of size 32×32 into a 1×1 unit; the second part is a fully connected layer with 32 input units and 2 output units, followed by a RELU; the third part is a fully connected layer with 2 input units and 32 output units, followed by an S-type activation function; the fourth part is channel weighting, which reweights the input feature map according to the channel weight. The SE module is added to the three residual blocks of the WideResNet network structure (such as Figure 5 As shown in Figure 2, the WideResNet network mainly learns local area features, and the SE module learns the importance weights between channels. The spliced ​​image generated in step S1 is input into the network structure that mixes the SE module and WideResNet, and the output is the corresponding two-dimensional CIR mixed image feature.

[0115] The two-dimensional CIR mixed image features output by the WideResNet network of the ViT network and the hybrid SE module are concatenated to generate a feature vector of a two-dimensional image of size 1×320. The final classification result of the two-dimensional image is obtained using a fully connected layer with 320 input units and 3 output units.

[0116] S2.4, initialize and train the hybrid network SEWideResViT of hybrid ViT, WideResNet and channel attention mechanism, specifically including the following steps:

[0117] (1) Initialization: The convolutional layers and fully connected layers in the ViT network use He Initialization to initialize weights w, and use uniform distribution to initialize biases; the position embedding and classification tags are initialized to 0. The convolutional layers of the WideResNet network use He Initialization to initialize weights w; the fully connected layers use Glorot Initialization to initialize weights w, and the bias is initialized to 0; the BN layer weights are initialized to 1, and the bias is initialized to 0. The SE module fully connected layer uses He Initialization to initialize weights w, and uses uniform distribution to initialize biases.

[0118] (2) Training: The SEWideResViT network uses the AdamW optimization algorithm that adaptively estimates the parameter update step size and independently performs weight decay. The learning rate is set to 0.0005, the weight decay coefficient is set to 0.0005, and the learning rate scheduler with preheating plus cosine annealing is used to dynamically adjust the basic learning rate during training. The exponential moving average is used to smooth the network parameters during training. The SEWideResViT network parameters are optimized to minimize the loss function. Since the present invention uses a semi-supervised learning algorithm, the total loss function consists of supervised loss and unsupervised consistency regularization loss. Cross entropy is used as the loss metric calculation. The formula is as follows:

[0119]

[0120] total_loss=sup_loss+λ*unsup_loss (8)

[0121] Where p is the predicted probability vector, y is the label / pseudo-label vector, C is the total number of categories, and λ is the loss weight for unlabeled data.

[0122] The hyperparameter settings during the training of the hybrid network SEWideResViT are as follows:

[0123]

[0124] like Figure 6 The figure shows a simplified process flow of the semi-supervised learning algorithm SoftMatch for unlabeled data. The weakly data-enhanced image is input into the network of the hybrid ViT, WideResNet, and channel attention mechanism to obtain the pseudo-label, and the weighted function value under the corresponding probability is calculated. At the same time, the prediction of the hybrid ViT, WideResNet, and channel attention mechanism network for the strong data enhancement of the same image is calculated. The model is trained so that its prediction of the strong data enhancement version matches the pseudo-label through the cross entropy loss.

[0125] S2.5, the loss function and optimization strategy of the SoftMatch algorithm are as follows:

[0126] The SoftMatch algorithm combines pseudo labels with consistency regularization, which not only improves the efficiency of pseudo labels but also ensures the overall quality of the pseudo labels used. The loss function of SoftMatch is composed of supervised loss With unsupervised loss Composition, the labeled dataset is The unlabeled dataset is The specific calculation formula is as follows:

[0127]

[0128] Where p is p(y|ω(x u )) abbreviation, is the one-hot encoding value argmax(p) of the pseudo label, λ(p) is the sample weighting function, B L is the batch size of labeled data, B U is the batch size of unlabeled data, Ω(x u ) is strong data enhancement, ω(x u ) is weak data enhancement.

[0129] Weak data augmentation uses mild image transformations, which retains key basic features, effectively generates high-quality pseudo labels, and provides slight regularization. Strong data augmentation uses more aggressive transformations, which enhances the model's recognition ability under diverse conditions and improves its robustness to input perturbations.

[0130] For labeled data: is the UWB channel state classification and recognition result obtained by performing weak data enhancement on the spliced ​​image obtained in step S1 and predicting it through the hybrid network. The y in formula (10) i is the true category. For unlabeled data: is the UWB channel state classification and recognition result obtained by performing strong data enhancement on the spliced ​​image obtained in step S1 and predicting through a hybrid network; It is a pseudo label obtained by performing weak data enhancement on the spliced ​​image obtained in step S1 and predicting it through a hybrid network.

[0131] The calculation process of the following formulas (12)-(17) is used to calculate the sample weighting function λ(p) in formula (11).

[0132] The truncated Gaussian distribution is selected as the weighting function λ(p), and the specific calculation formula is as follows:

[0133]

[0134] The parameter μ is calculated based on the model's past prediction output for unlabeled data t and σ t , the calculation formula for the tth iteration is as follows:

[0135]

[0136] Based on the exponential moving average of historical batch data, a more reliable parameter μ is obtained t and σ t :

[0137]

[0138] in, The value is 1 / C, The value is 1.0 and m is the weight.

[0139] In order to solve the problem of unbalanced distribution of pseudo-label categories, uniform alignment is used to force the model to cover each category more balanced when generating pseudo-labels. The specific formula is as follows:

[0140]

[0141] Where u(C) is a uniform distribution and Normalize() is a normalization operation.

[0142] The SoftMatch algorithm assigns non-zero weights to low-confidence samples by truncating the Gaussian function, thereby improving the utilization of pseudo-labels. Through dynamic weight allocation and UA mechanism, it maintains a high pseudo-label quality while improving the utilization.

[0143] S3.2, initialize and train the hybrid network based on the SoftMatch algorithm, specifically including the following steps:

[0144] (1) Initialization: The transformation for labeled data includes weak data augmentation and normalization, and the transformation for unlabeled data includes strong and weak data augmentation and normalization; weak data augmentation uses padding random cropping and 50% probability of horizontal flipping, and strong data augmentation uses padding random cropping, 50% probability of horizontal flipping and random augmentation (operations include rotation, translation, shearing, brightness adjustment, contrast adjustment, color adjustment, adaptive sharpening and equalization, etc., 3 operations are randomly selected from them, and the enhancement strength of each operation is 6). The total number of labeled data is set to 180, and the number of data categories is set to 3. The prediction results of unlabeled data use hard labels; set distribution alignment to adjust the predicted probability distribution of unlabeled data to make it closer to uniform distribution; the unlabeled data prediction process uses EMA, and the weight value is set to 0.999 to better capture the model's prediction trend for unlabeled data in the long term; the parameter μ t and σ t The weight value when using EMA is set to 0.999. The batch size of unlabeled data during training is twice that of labeled data. (2) Training: As the model performance improves, the prediction tends to be stable, and the mean μ of the weighting function λ(p) t The variance σ increases t Reduce and dynamically adjust the sample weight distribution. At the same time, through UA operation, the algorithm focuses on learning the current underfitting category features, thereby enhancing the overall generalization ability of the model.

[0145] The SoftMatch algorithm hyperparameter settings are as follows:

[0146]

[0147]

[0148] Summarizing the above technical features, the overall detailed flow chart of the method for identifying the UWB channel environment state based on limited-mark CIR data is as follows: Figure 7 shown.

[0149] Example 2

[0150] Taking the collected CIR signals in a complex environment (divided into three propagation conditions: LOS, DP_NLOS, and NDP_NLOS, with a total of 9000 data, of which 60 labeled data are of each category, and the training set, validation set, and test set are extracted in a ratio of 7:1:2) as an example, the method of the present invention is further illustrated.

[0151] The specific implementation process of the method of the present invention comprises the following steps:

[0152] Step S1: Select channel impulse response data of appropriate length, and convert the one-dimensional time series into a two-dimensional image through continuous wavelet transform and Gram angular field transform; the specific steps are as follows:

[0153] S1.1: 9000 CIR data in a complex environment are collected through a UWB device based on the DW1000 chip. The format of each original CIR data is 1×1016. The starting point of the CIR data segment to be extracted is determined by the first path peak index HLP provided by the DW1000 chip, that is: take 5 values ​​forward from the index point and take 122 consecutive values ​​backward. That is to say, the present invention uses a CIR segment with a length of 128 as the initial input.

[0154] S1.2: Apply continuous wavelet transform and Gram angular field transform to the 1×128 CIR segment; set the signal length of CWT to 128, the sampling frequency to 1e9, the number of divisions per octave to 18, and select the generalized Morse wavelet with parameters γ=3 and β=20 as the mother wavelet, so the output CWT image pixels are 128×74, and the CWT image size is adjusted to 32×16; set the image size of GAF transform to 128, so the output GAF image pixels are 128×128, and the GAF image size is adjusted to 32×16. The adjusted CWT image and GAF ​​image are spliced ​​vertically to obtain an RGB image with pixels of 32×32.

[0155] Step S2: Use the neural network architecture SEWideResViT that integrates the visual converter, wide residual network and channel attention mechanism for feature extraction and classification; the specific steps are as follows:

[0156] S2.1: For the ViT network part, the input image size is 3×32×32, and it goes through the image block layer, position embedding layer, Transformer encoder layer and other structures, and finally only 1×192 classification mark features are extracted;

[0157] S2.2: For the WideResNet part, the input image size is 3×32×32 and passes through the main three wide residual blocks and other structures, and finally obtains the output feature of 1×128;

[0158] S2.3: For the SE module, since the SE module is integrated inside WideResNet, the input feature dimension of the first SE module is 32×32×32, the input feature dimension of the second SE module is 64×16×16, and the input feature dimension of the third SE module is 128×8×8. The output feature dimension is equal to the input feature dimension.

[0159] The 1×192 dimensional features of the ViT network and the 1×128 dimensional features of WideResNet are concatenated to obtain 1×320 dimensional features, and a classification result with a shape of 1×3 is generated through a fully connected layer with 320 input units and 3 output units.

[0160] Step S3: Select the semi-supervised learning algorithm SoftMatch to achieve the final channel state classification and recognition under the premise of less labeled data.

[0161] 9000 pieces of data collected by UWB devices were partially annotated, with 60 samples annotated for each category. Weak data enhancement was used for labeled data, and the model was used to predict the data after weak data enhancement, and the cross entropy loss was calculated between the prediction result and the label; strong and weak data enhancement were used for unlabeled data, and the image after weak data enhancement was input into the model to obtain the pseudo label and calculate the weighted function value under the corresponding probability. At the same time, the model's prediction result of strong data enhancement for the same image was calculated, and matched with the pseudo label through cross entropy loss.

[0162] The classification model parameter configuration and final accuracy evaluation results are shown in the following table:

[0163]

[0164] In summary, compared with other UWB channel state identification methods, the embodiments of the present invention have the following advantages: it can effectively parse and extract channel pulse response data features, and can still achieve accurate classification in actual application scenarios where labeled data is scarce, while maintaining strong environmental adaptability and anti-interference capabilities, which helps to improve the positioning effect of UWB technology in complex environments.

[0165] Those skilled in the art can understand that the accompanying drawings are only schematic diagrams of an embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present invention.

[0166] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present invention can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present invention or certain parts of the embodiments.

[0167] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0168] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for identifying the state of an ultra-wideband (UWB) channel environment under limited labeled data, characterized in that: include: Selecting an ultra-wideband UWB channel impulse response CIR data segment, and converting the UWB channel impulse response data into a two-dimensional image by continuous wavelet transform and Gram angular field transform; Construct a hybrid network that integrates a visual converter, a wide residual network, and a channel attention mechanism, and train the hybrid network using labeled data and unlabeled data through a joint supervised learning and unsupervised learning strategy, that is, introduce a semi-supervised learning algorithm SoftMatch to optimize the training of the hybrid network; The two-dimensional image is input into a hybrid network, and the hybrid network is combined with a SoftMatch algorithm to obtain a UWB channel state classification and recognition result under limited labeled data.

2. The method according to claim 1, characterized in that: The selecting of UWB CIR data segments and converting the UWB channel impulse response data into a two-dimensional image by continuous wavelet transform and Gram angular field transform comprises: Obtaining a log file of CIR data generated by a UWB device, preprocessing the log file, and extracting a specific segment of the one-dimensional time series CIR; A two-dimensional time-frequency domain image is generated by using a specific segment of the one-dimensional time series CIR through continuous wavelet transform to obtain a synchronous representation of the CIR signal in the time and frequency dimensions. The generalized Morse wavelet is selected as the basis function, and the Fourier transform of the generalized Morse wavelet is: Among them, U(ω) is a unit step function, β and γ are parameters that control the attenuation of the wavelet in the time domain and frequency domain, and determine the wavelet waveform; The formula for continuous wavelet transform is defined as follows: Where x(t) is the original one-dimensional CIR signal, τ is the translation factor, a is the scale factor, Ψ * (t) is the complex conjugate of the basis function, WT(a,τ) is the wavelet transform value, and RGB color mapping is performed on the wavelet transform value WT(a,τ) to generate a two-dimensional time-frequency domain image; The specific segment of the one-dimensional time series CIR is converted and encoded into a two-dimensional image by Gram angular field transform. n ] is normalized and scaled to between 0 and 1, as follows: Among them, x i is the data point in the original CIR time series; Convert the normalized signal to polar coordinates: Among them, Φ i is the polar angle after the arc cosine function transformation, r i is the polar diameter, i is the timestamp, and n is a coefficient used to adjust the coverage of the polar coordinate system; The Gram and angle field is selected and calculated using the phase angle and phase angle of the time series points after polar coordinate transformation. The specific formula is as follows: Use Python programming to splice the two-dimensional time-frequency domain image obtained by continuous wavelet transform and the two-dimensional image obtained by Gram angular field transform up and down to obtain a two-dimensional spliced ​​image.

3. The method according to claim 1, characterized in that The hybrid network of the fusion visual converter, wide residual network and channel attention mechanism includes: Determine the parameters of the ViT network. The ViT network consists of an image block layer, a position embedding layer, and a Transformer encoder layer. Determine the WideResNet network type and parameters, determine the SE module parameters, and add the SE module to the three residual blocks of the WideResNet network structure; The ViT network and the WideResNet network mixed with the SE module are fused to obtain a hybrid network.

4. The method according to claim 1, characterized in that: The hybrid network is trained using labeled data and unlabeled data by combining supervised learning and unsupervised learning strategies, that is, the semi-supervised learning algorithm SoftMatch is introduced to optimize the training of the hybrid network, including: (1) Initialization: The transformation for labeled data includes weak data augmentation and normalization, and the transformation for unlabeled data includes strong and weak data augmentation and normalization; weak data augmentation uses padding random cropping and horizontal flipping with a probability of 50%, and strong data augmentation uses padding random cropping, horizontal flipping with a probability of 50%, and random augmentation. The total number of labeled data is set to 180, the number of data categories is set to 3, and the unlabeled data prediction results use hard labels; set distribution alignment to adjust the predicted probability distribution of unlabeled data to make the predicted probability distribution closer to uniform distribution; the unlabeled data prediction process uses EMA, and the weight value is set to 0.999 to capture the model's prediction trend for unlabeled data in the long term; the parameter μ t and σ t The weight value when using EMA is set to 0.999, and the batch size of unlabeled data during training is twice that of labeled data; (2) Training: mean μ of weighted function λ(p) t The variance σ increases t Reduce, dynamically adjust the sample weight distribution, and learn the current underfitting category features through UA operation.

5. The method according to claim 4, characterized in that The two-dimensional image is input into the hybrid network, and the hybrid network is combined with the semi-supervised learning SoftMatch algorithm to obtain the UWB channel state classification and recognition result under limited labeled data, including: The loss function and optimization strategy of the SoftMatch algorithm are set as follows: The loss function of SoftMatch is composed of supervised loss With unsupervised loss Composition, the labeled dataset is The unlabeled dataset is The specific calculation formula is as follows: Where p is p(y|ω(x u )) abbreviation, is the one-hot encoding value argmax(p) of the pseudo label, λ(p) is the sample weighting function, B L is the batch size of labeled data, B U is the batch size of unlabeled data, Ω(x u ) is strong data enhancement, ω(x u ) is weak data enhancement; The sample weighting function λ(p i ) is calculated as follows: The truncated Gaussian distribution is selected as the weighting function λ(p), and the specific calculation formula is as follows: The parameter μ is calculated based on the model's past prediction output for unlabeled data t and σ t , the calculation formula for the tth iteration is as follows: Based on the exponential moving average of historical batch data, a more reliable parameter μ is obtained t and σ t : in, The value is 1 / C, The value is 1.0, and m is the weight; In order to solve the problem of unbalanced distribution of pseudo-label categories, uniform alignment is used to force the model to cover each category more balanced when generating pseudo-labels. The specific formula is as follows: Among them, u(C) is uniform distribution, Normalize() is normalization operation; The two-dimensional image is input into the hybrid network, the WideResNet network learns local area features, the SE module learns the importance weights between channels, the WideResNet network of the hybrid SE module outputs two-dimensional CIR hybrid image features, the ViT network learns the relationship between global features and features, and the ViT network outputs two-dimensional CIR hybrid image features; The two-dimensional CIR mixed image features output by the WideResNet network of the ViT network and the hybrid SE module are spliced ​​to generate a feature vector of the two-dimensional image. The UWB channel state classification and recognition result is obtained using a fully connected layer according to the feature vector of the two-dimensional image. For the labeled data: is the UWB channel state classification and recognition result obtained by hybrid network prediction after weak data enhancement of the spliced ​​image; for unlabeled data: It is the UWB channel status classification and recognition result obtained by hybrid network prediction after strong data enhancement of the spliced ​​image; It is the pseudo label obtained by predicting the hybrid network after weak data enhancement of the spliced ​​image.

Images