An ultra-wideband non-line-of-sight signal identification method, device and computer equipment

Abstract

The application relates to an ultra-wideband non-line-of-sight signal identification method, device and computer equipment, which comprises the following steps: obtaining an original ultra-wideband signal sample set, constructing a one-dimensional array discrete sequence according to original ultra-wideband signal samples in the original ultra-wideband signal sample set; determining a data deletion discrete interval of each original ultra-wideband signal sample according to the one-dimensional array discrete sequence, obtaining a plurality of intermediate ultra-wideband signal samples according to the data deletion discrete interval; filling the data deletion discrete interval with a piece of data randomly selected from other original ultra-wideband signal samples for each data deletion discrete interval in the intermediate ultra-wideband signal samples, obtaining a plurality of mixed ultra-wideband signal samples; training a pre-constructed identification model by using the mixed ultra-wideband signal samples to obtain a trained identification model, and identifying the ultra-wideband non-line-of-sight signal by using the trained identification model. The application considers the precision and speed of ultra-wideband signal identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, in particular to an ultra-wideband non-line-of-sight signal identification method and device and computer equipment. BACKGROUND

[0002] With the increasing demand for positioning of personnel, goods and other targets in indoor scenarios such as large shopping malls, underground parking lots and storage warehouses, ultra-wideband (UWB) signals have become a reliable choice for short-range wireless positioning due to their low power consumption, high precision and strong penetration. UWB signal propagation has two cases: line of sight (LOS) and non-line of sight (NLOS). In the LOS case, the distance between the target and each base station is determined by the transmission time of the UWB signal, and further positioning is performed, which can currently achieve centimeter-level positioning accuracy. However, various obstacles often exist in actual environments, blocking UWB signals to form NLOS propagation, which seriously damages the time of arrival parameters required for positioning, resulting in increased positioning error. Therefore, in the positioning process based on UWB signals, NLOS and LOS are classified and identified, which can ensure accurate positioning of LOS propagation signals, or after identifying NLOS signals, the positioning result is obtained after error compensation of NLOS signals.

[0003] Existing UWB signal NLOS identification methods include: 1. Constructing a double-flow network, one path uses a residual network to extract features of CIR signals at multiple times, and the other path uses a convolutional neural network to extract features of multiple additional parameters. The disadvantage of this method is that the residual network is generally used to process time series signals; 2. Combining UWB and inertial measurement units to establish a combined positioning model, or pre-constructing a pedestrian model, etc., to jointly determine whether NLOS exists. This method requires additional sensor support; 3. Using artificially selected features to identify NLOS, which is not sufficient for information utilization and cannot achieve accurate classification; 4. Based on the statistical information of UWB ranging values, the method uses hypothesis testing to compare the statistical values of multiple distance measurements with a decision threshold to make identification, which requires multiple measurements at each positioning location to obtain sufficient sample size and calculate the statistical parameters of the measurement data for comparison with known parameters, resulting in reduced real-time positioning. SUMMARY

[0004] Therefore, the purpose of the present application is to provide an ultra-wideband non-line-of-sight signal identification method, device and computer equipment that takes into account the accuracy and speed of ultra-wideband non-line-of-sight signal identification.

[0005] An ultra-wideband non-line-of-sight signal identification method, the method comprising:

[0006] obtaining a set of original ultra-wideband signal samples, constructing a one-dimensional array discrete sequence according to original ultra-wideband signal samples in the set of original ultra-wideband signal samples; the one-dimensional array discrete sequence is connected in series by one-dimensional arrays;

[0007] determining a data deletion discrete interval of each original ultra-wideband signal sample according to the one-dimensional array discrete sequence, and obtaining a plurality of intermediate ultra-wideband signal samples according to the data deletion discrete interval;

[0008] filling each data deletion discrete interval in the intermediate ultra-wideband signal sample with data randomly selected from other original ultra-wideband signal samples to obtain a plurality of mixed ultra-wideband signal samples; the other original ultra-wideband signal samples are any original ultra-wideband signal samples other than the original ultra-wideband signal sample corresponding to the current intermediate ultra-wideband signal sample; the length of the randomly selected data is equal to the length of the corresponding data deletion discrete interval;

[0009] training a pre-constructed identification model by using the mixed ultra-wideband signal samples to obtain a trained identification model;

[0010] identifying an ultra-wideband non-line-of-sight signal by using the trained identification model.

[0011] In one embodiment, the determining a data deletion discrete interval of each original ultra-wideband signal sample according to the one-dimensional array discrete sequence, and obtaining a plurality of intermediate ultra-wideband signal samples according to the data deletion discrete interval comprises:

[0012] aligning the one-dimensional array discrete sequence with each original ultra-wideband signal sample respectively;

[0013] determining a data deletion discrete interval in the original ultra-wideband signal sample according to the position of each one-dimensional array in the one-dimensional array discrete sequence;

[0014] locally deleting data in the corresponding original ultra-wideband signal sample according to the data deletion interval to obtain a plurality of intermediate ultra-wideband signal samples.

[0015] In one embodiment, the constructing a one-dimensional array discrete sequence according to original ultra-wideband signal samples in the set of original ultra-wideband signal samples comprises:

[0016] designing the number of one-dimensional arrays in the one-dimensional array discrete sequence and the interval change interval of adjacent one-dimensional arrays;

[0017] randomly selecting a value in the interval change interval as the interval of each pair of adjacent one-dimensional arrays;

[0018] constructing a data deletion rate according to a sum of a data length of the original ultra-wideband signal sample in the original ultra-wideband signal sample set and the interval of each one-dimensional array discrete sequence;

[0019] designing a data length of each one-dimensional array according to the preset data deletion rate, the data length of the original ultra-wideband signal sample and the number of one-dimensional arrays in the one-dimensional array discrete sequence;

[0020] constructing a one-dimensional array discrete sequence according to the number of one-dimensional arrays, the interval of each pair of adjacent one-dimensional arrays and the data length of each one-dimensional array.

[0021] In one embodiment, the pre-constructed recognition model is trained by using the mixed ultra-wideband signal sample to obtain a trained recognition model, including:

[0022] obtaining a pre-constructed recognition model; the recognition model includes a time convolution network layer and a self-attention layer;

[0023] using the mixed ultra-wideband signal sample as a training sample of the recognition model;

[0024] inputting the training sample into the recognition model, constructing a fitness function according to a recognition accuracy and a convergence time of the recognition model, and optimizing network parameters of the recognition model by using a particle swarm optimization algorithm to obtain a trained recognition model, wherein the network parameters include a convolution kernel size and a number of the time convolution network layer, a number of time convolution network layers, a learning rate and a batch size.

[0025] In one embodiment, the inputting the training sample into the recognition model, constructing a fitness function according to a recognition accuracy and a convergence time of the recognition model, and optimizing network parameters of the recognition model by using a particle swarm optimization algorithm to obtain a trained recognition model, includes:

[0026] constructing the fitness function according to the recognition accuracy and the convergence time of the recognition model as:

[0027] f(X) = a(X) - a · t(X)

[0028] wherein a(X) represents a recognition accuracy of the recognition model corresponding to the particle, t(X) represents a convergence time of the recognition model, and a represents a preset weight coefficient;

[0029] setting initial parameters of the particle swarm; the initial parameters include a maximum iteration number of the particle swarm optimization algorithm, a particle number, a search dimension, and two learning factors;

[0030] setting an inertia weight according to the maximum iteration number:

[0031]

[0032] wherein w represents an inertia weight, w max represents an upper bound of the value of the inertia weight, w min represents a lower bound of the value of the inertia weight, Iter represents a maximum number of iterations, and iter represents a current number of iterations;

[0033] Randomly initializing a position of a particle in each search dimension in a particle swarm:

[0034] X i = (m, n, l, lr, b)

[0035] wherein X i represents a position of the i-th particle, m represents a convolution kernel size of a time convolution network layer, n represents a number of convolution kernels of the time convolution network layer, l represents a number of time convolution network layers, lr represents a learning rate, and b represents a batch size;

[0036] Randomly initializing a velocity of a particle in each search dimension in the particle swarm;

[0037] In an iteration process, each particle constantly updates a velocity and a position according to a swarm optimal position and an individual optimal position, and the updating formulae of the particle velocity and the particle position are as follows:

[0038]

[0039]

[0040] wherein represents a velocity of the i-th particle in the dimension d at the k-th iteration, represents a velocity of the i-th particle in the dimension d at the (k-1)-th iteration, c1 and c2 are preset learning factors, r1 and r2 are two random numbers with a value range of [0, 1], gbest d represents a value of the particle with the optimal fitness value in all iterations of all particles in the dimension d, and pbest id represents a value of the i-th particle in the dimension d when the i-th particle obtains the optimal fitness value in all iterations;

[0041] When the maximum number of iterations is reached, a corresponding optimal network parameter and an optimized recognition model are obtained;

[0042] The training sample is input into the optimized recognition model, and a trained recognition model is obtained.

[0043] In an embodiment, the trained recognition model is used for recognition of an ultra-wideband non-line-of-sight signal, including:

[0044] obtaining an ultra-wideband signal to be recognized, mapping the ultra-wideband signal into a channel impulse response sequence;

[0045] inputting the channel impulse response sequence into the time convolution network layer to obtain a corresponding time sequence; a time step of the time sequence is same as a time step of the channel impulse response sequence; the time convolution network layer comprises a causal convolution layer, a dilated convolution layer and a residual connection layer;

[0046] inputting the time sequence into the self-attention layer, performing flattening processing on a corresponding output result, and inputting the flattened result into a fully connected layer of the recognition model, and using a Sigmoid activation function to recognize the ultra-wideband non-line-of-sight signal.

[0047] In one embodiment, according to the data deletion interval, data in a corresponding original ultra-wideband signal sample is locally deleted to obtain a plurality of intermediate ultra-wideband signal samples, comprising:

[0048] According to the data deletion interval, data in a corresponding original ultra-wideband signal sample is locally deleted to obtain a plurality of intermediate ultra-wideband signal samples, comprising:

[0049]

[0050] wherein, indicates an intermediate ultra-wideband signal sample, X indicates an original ultra-wideband signal sample, M e {0, 1} W indicates a data processing condition at a corresponding position, M = 0 indicates deletion, M = 1 indicates retention, W indicates a data length of the original ultra-wideband signal sample, {W i , i.e. the data deletion interval.

[0051] In one embodiment, according to a pre-set data deletion rate, a data length of the original ultra-wideband signal sample and a number of one-dimensional arrays in the one-dimensional array discrete sequence, a data length of each one-dimensional array is designed, comprising:

[0052] setting each one-dimensional array in the one-dimensional array discrete sequence to be of the same length;

[0053] According to a pre-set data deletion rate, a data length of the original ultra-wideband signal sample and a number of one-dimensional arrays in the one-dimensional array discrete sequence, a data length of each one-dimensional array is designed, comprising:

[0054]

[0055] wherein, l r indicates a data length of the rth one-dimensional array, c indicates a data deletion rate, and R indicates a number of one-dimensional arrays in the one-dimensional array discrete sequence.

[0056] An ultra-wideband non-line-of-sight signal recognition device, characterized in that the device comprises:

[0057] A discrete sequence construction module is configured to acquire an original ultra-wideband signal sample set and construct a one-dimensional array discrete sequence according to original ultra-wideband signal samples in the original ultra-wideband signal sample set; the one-dimensional array discrete sequence is connected in series.

[0058] An intermediate sample acquisition module is configured to determine a data deletion discrete interval of each original ultra-wideband signal sample according to the one-dimensional array discrete sequence and obtain a plurality of intermediate ultra-wideband signal samples according to the data deletion discrete interval.

[0059] A data filling module is configured to fill each data deletion discrete interval in the intermediate ultra-wideband signal sample with data randomly selected from other original ultra-wideband signal samples to obtain a plurality of mixed ultra-wideband signal samples; the other original ultra-wideband signal samples are any original ultra-wideband signal sample other than the original ultra-wideband signal sample corresponding to the current intermediate ultra-wideband signal sample; the length of the randomly selected data is equal to the length of the corresponding data deletion discrete interval.

[0060] A model training module is configured to train a pre-constructed recognition model using the mixed ultra-wideband signal samples to obtain a trained recognition model.

[0061] A signal recognition module is configured to recognize an ultra-wideband non-line-of-sight signal using the trained recognition model.

[0062] A computer device comprises a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0063] An original ultra-wideband signal sample set is acquired, and a one-dimensional array discrete sequence is constructed according to original ultra-wideband signal samples in the original ultra-wideband signal sample set; the one-dimensional array discrete sequence is connected in series.

[0064] A data deletion discrete interval of each original ultra-wideband signal sample is determined according to the one-dimensional array discrete sequence, and a plurality of intermediate ultra-wideband signal samples are obtained according to the data deletion discrete interval.

[0065] delete a discrete interval of data from each of the intermediate ultra-wideband signal samples, fill the data-deleted discrete interval with a piece of data randomly selected from other original ultra-wideband signal samples to obtain a plurality of mixed ultra-wideband signal samples corresponding to the data-deleted discrete interval; the other original ultra-wideband signal samples are any original ultra-wideband signal samples other than the original ultra-wideband signal sample corresponding to the current intermediate ultra-wideband signal sample; the length of the randomly selected data is equal to the length of the corresponding data-deleted discrete interval;

[0066] train the pre-constructed identification model by using the mixed ultra-wideband signal samples to obtain a trained identification model;

[0067] identify the ultra-wideband non-line-of-sight signal by using the trained identification model.

[0068] The above-mentioned ultra-wideband non-line-of-sight signal identification method, device and computer equipment first acquire an original ultra-wideband signal sample set, construct a one-dimensional array discrete sequence according to the original ultra-wideband signal samples in the original ultra-wideband signal sample set, and serially connect one-dimensional array strings in the one-dimensional array discrete sequence; then determine data-deleted discrete intervals of the original ultra-wideband signal samples according to the one-dimensional array discrete sequence, and obtain a plurality of intermediate ultra-wideband signal samples corresponding to the data-deleted discrete intervals; then fill the data-deleted discrete intervals with a piece of data randomly selected from other original ultra-wideband signal samples to obtain a plurality of mixed ultra-wideband signal samples corresponding to the data-deleted discrete intervals, the other original ultra-wideband signal samples are any original ultra-wideband signal samples other than the original ultra-wideband signal sample corresponding to the current intermediate ultra-wideband signal sample, and the length of the randomly selected data is equal to the length of the corresponding data-deleted discrete interval; then train the pre-constructed identification model by using the mixed ultra-wideband signal samples to obtain a trained identification model, and finally identify the ultra-wideband non-line-of-sight signal by using the trained identification model. The present application enhances the diversity of samples by introducing a one-dimensional array discrete sequence to locally delete the original ultra-wideband signal samples and filling the deletion area according to other original ultra-wideband signal samples, so that the generalization ability of the network parameters of the subsequent trained identification model is stronger, and the precision and speed of ultra-wideband signal identification are greatly improved. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1 A flowchart of an embodiment of an ultra-wideband non-line-of-sight signal identification method;

[0070] Figure 2 A schematic diagram of a one-dimensional array discrete sequence in an embodiment;

[0071] Figure 3 A schematic diagram of the positional relationship between a one-dimensional array discrete sequence and an original ultra-wideband signal sample in an embodiment;

[0072] Figure 4 An architecture diagram for identifying a model in another embodiment;

[0073] Figure 5 A flowchart of a particle swarm algorithm in one embodiment;

[0074] Figure 6 A causal dilated convolution diagram in one embodiment;

[0075] Figure 7 A structure block diagram of an ultra-wideband non-line-of-sight signal identification device in one embodiment;

[0076] Figure 8 An internal structure diagram of a computer device in one embodiment. DETAILED DESCRIPTION

[0077] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0078] In one embodiment, as shown in Figure 1 , an ultra-wideband non-line-of-sight signal identification method is provided, comprising the following steps:

[0079] Step 102, obtaining an original ultra-wideband signal sample set, and constructing a one-dimensional array discrete sequence according to the original ultra-wideband signal samples in the original ultra-wideband signal sample set.

[0080] As shown in Figure 2 , a schematic diagram of the one-dimensional array discrete sequence is provided, and it can be seen that the one-dimensional arrays in the one-dimensional array discrete sequence are connected in series, and there is a gap between adjacent one-dimensional arrays, where d r represents the spacing between the rth one-dimensional array and the (r-1)th one-dimensional array, and l r represents the data length of the rth one-dimensional array. The spacing of each pair of adjacent one-dimensional arrays can be the same or different, and the data length of different one-dimensional arrays can be the same or different, that is, the size and spacing of the one-dimensional arrays in the one-dimensional array discrete sequence can be adjusted, Figure 2 for display only, and actual design can be made according to specific conditions.

[0081] Constructing a one-dimensional array discrete sequence according to the original ultra-wideband signal samples in the original ultra-wideband signal sample set means that the total length of the data in the constructed one-dimensional array sequence and the total value of the spacing are at least equal to the data length of the original ultra-wideband sample.

[0082] Step 104, determining the data deletion discrete interval of each original ultra-wideband signal sample according to the one-dimensional array discrete sequence, and obtaining a plurality of intermediate ultra-wideband signal samples according to the data deletion discrete interval.

[0083] As shown in the position relation diagram between the one-dimensional array discrete sequence and the original ultra-wideband signal sample, it can be seen that the position of each one-dimensional array in the one-dimensional array discrete sequence is the data deletion discrete interval of the original ultra-wideband signal sample, and the dashed box in the figure is the data deletion discrete interval. Figure 3

[0084] The data in the data deletion discrete interval is deleted, that is, the data value in the data deletion discrete interval is changed to 0, so that the excessive deletion is effectively avoided when the original sample data information is discarded, and the continuity of the target information can be better preserved.

[0085] Step 106, filling the data deletion discrete interval in each intermediate ultra-wideband signal sample with a piece of data randomly selected from other original ultra-wideband signal samples to obtain a plurality of mixed ultra-wideband signal samples.

[0086] The other original ultra-wideband signal sample is any original ultra-wideband signal sample other than the original ultra-wideband signal sample corresponding to the current intermediate ultra-wideband signal sample, and the length of the randomly selected data is equal to the length of the corresponding data deletion discrete interval.

[0087] Directly changing the value of part of the data in the input data to 0 will lose a large amount of content information, resulting in the destruction of features. In order to ensure that subsequent processing can extract as rich features as possible, the present application fuses the deleted region with other samples, so that no information sample appears in the training process, and the training efficiency is improved. The data obtained from the corresponding position of other original ultra-wideband signal samples is interpolated into the corresponding position of the current input sample data, that is, the data of the other sample original ultra-wideband signal sample is interpolated into the data deletion interval, so that no information input sample appears in the training process. After this processing, the diversity of the sample is enhanced, and the quality of the network model obtained by training can be improved.

[0088] Step 108, training the pre-constructed recognition model using the mixed ultra-wideband signal sample to obtain a trained recognition model.

[0089] Step 110, using the trained recognition model to recognize the ultra-wideband non-line-of-sight signal.

[0090] ​The method comprises the following steps: first, obtaining a set of original ultra-wideband signal samples, and constructing a one-dimensional array discrete sequence according to the original ultra-wideband signal samples in the set of original ultra-wideband signal samples, wherein the one-dimensional array discrete sequence is formed by serially connecting one-dimensional arrays; then, determining a data deletion discrete interval of each original ultra-wideband signal sample according to the one-dimensional array discrete sequence, and obtaining a plurality of intermediate ultra-wideband signal samples according to the data deletion discrete interval; then, filling the data deletion discrete interval in each intermediate ultra-wideband signal sample with data randomly selected from other original ultra-wideband signal samples to obtain a plurality of mixed ultra-wideband signal samples, wherein the other original ultra-wideband signal samples are any original ultra-wideband signal samples other than the original ultra-wideband signal sample corresponding to the current intermediate ultra-wideband signal sample, and the length of the randomly selected data is equal to the length of the corresponding data deletion discrete interval; then, training a pre-constructed identification model by using the mixed ultra-wideband signal samples to obtain a trained identification model; and finally, identifying the ultra-wideband non-line-of-sight signal by using the trained identification model. The original ultra-wideband signal samples are partially deleted by introducing the one-dimensional array discrete sequence, and the deletion area is filled according to other original ultra-wideband signal samples, so as to enhance the diversity of the samples, make the generalization ability of the network parameters of the subsequently trained identification model stronger, and greatly improve the precision and speed of the ultra-wideband signal identification.

[0091] In one embodiment, the one-dimensional array discrete sequence is constructed according to the original ultra-wideband signal samples in the set of original ultra-wideband signal samples, comprising:

[0092] The number of one-dimensional arrays in the one-dimensional array discrete sequence and the interval change range of adjacent one-dimensional arrays are designed, and the interval of each pair of adjacent one-dimensional arrays is randomly selected as the interval in the interval change range. Assuming that there are r deletion arrays in the sequence, the parameter d r The density of the one-dimensional arrays in the sequence is set, and the method introduces randomness to enhance the diversity of the samples in a big data environment. Specifically, d min and d max represent the minimum and maximum values of the intervals between adjacent arrays in the sequence, and d r is randomly selected in the range of [d min , d max ]:

[0093] d r = random[d min , d max ].

[0094] The data deletion rate is constructed according to the sum of the data length of the original ultra-wideband signal sample in the original ultra-wideband signal sample set and the interval of each one-dimensional array discrete sequence. The greater the data deletion rate is, the more elements are deleted. If the data deletion rate is too small, the deleted elements are too few, and it is difficult to achieve the purpose of enhancing sample diversity. For example, the data deletion rate can be set to 0.1.

[0095] The data length of each one-dimensional array is designed according to the pre-set data deletion rate, the data length of the original ultra-wideband signal sample, and the number of one-dimensional arrays in the one-dimensional array discrete sequence. For simplicity, the data length of each one-dimensional array can be set to r The value is equal, and the value is calculated as:

[0096]

[0097] Wherein, l r The data length of the rth one-dimensional array is represented by r, c represents the data deletion rate, and R represents the number of one-dimensional arrays in the one-dimensional array discrete sequence.

[0098] The one-dimensional array discrete sequence is constructed according to the number of one-dimensional arrays, the interval of each pair of adjacent one-dimensional arrays, and the data length of each one-dimensional array.

[0099] In one embodiment, the data deletion discrete interval of each original ultra-wideband signal sample is determined according to the one-dimensional array discrete sequence, and the corresponding plurality of intermediate ultra-wideband signal samples are obtained according to the data deletion discrete interval, including:

[0100] The one-dimensional array discrete sequence is aligned with each original ultra-wideband signal sample respectively, the data deletion discrete interval in the original ultra-wideband signal sample is determined according to the position of each one-dimensional array in the one-dimensional array discrete sequence, and the data in the corresponding original ultra-wideband signal sample is partially deleted according to the data deletion interval to obtain the corresponding plurality of intermediate ultra-wideband signal samples.

[0101] In specific implementation, the data in the corresponding original ultra-wideband signal sample is partially deleted according to the data deletion interval to obtain the corresponding plurality of intermediate ultra-wideband signal samples, which is:

[0102]

[0103] Wherein, The intermediate ultra-wideband signal sample is represented by X, the original ultra-wideband signal sample is represented by X, M is an element of {0,1} W The data processing condition at the corresponding position is represented by M, M=0 represents deletion, M=1 represents retention, W represents the data length of the original ultra-wideband signal sample, and {W i M=0} is the data deletion interval.

[0104] In one embodiment, a piece of data is randomly selected from other original ultra-wideband signal samples to fill the data deletion discrete interval, and a plurality of mixed ultra-wideband signal samples corresponding to the data deletion discrete interval are obtained, and the specific steps are as follows:

[0105] It is known , which represents the intermediate ultra-wideband signal sample after the current local data deletion is completed, The position of the deletion operation in the array has r arrays, and the mixing operation is performed as follows:

[0106] Step 1: The first one-dimensional array i = 1, and the length of the array is l r ;

[0107] Step 2: If i > r, exit the mixing operation; otherwise, generate a random number p = random[0, 1], and if p ≤ 0.5, i = i + 1, and step 2 is executed again, otherwise, step 3 is executed;

[0108] Step 3: Randomly select another original ultra-wideband signal sample X' i ∈ R W from the original ultra-wideband signal sample set as the interpolation data source, and randomly select a piece of data with a length of l r from X' i , that is, the starting position of the piece of data is random[1, W-l r ], and the piece of data is copied to the corresponding elements of the i arrays deleted in ;

[0109] Step 4: i = i + 1, and jump to step 2.

[0110] In one embodiment, the pre-constructed recognition model is trained using the mixed ultra-wideband signal sample to obtain a trained recognition model, including:

[0111] The pre-constructed recognition model is obtained, and the recognition model includes a time convolution network layer and a self-attention layer. As Figure 4 shown, an architecture diagram of the recognition model is provided.

[0112] The mixed ultra-wideband signal sample is used as a training sample of the recognition model, the training sample is input into the recognition model, and an adaptability function is constructed according to the recognition accuracy and convergence time of the recognition model:

[0113] f(X) = a(X) - a·t(X)

[0114] Wherein, a(X) represents the recognition accuracy of the recognition model corresponding to the particle, t(X) represents the convergence time of the recognition model, and a represents a preset weight coefficient, which can be 2×10 -4The time convolution network has many parameters, and traditional manual parameter adjustment is not only complex to operate but also easy to make the network fall into a local optimal solution. In addition, the ultra-wideband signal transceiver device is often a small and integrated device, and usually has small specifications, limited storage capacity and computing capacity. Therefore, when selecting network parameters, not only the network identification accuracy should be considered, but also the processing speed problem should be fully considered. The particle swarm optimization algorithm is used to select the parameters of the time convolution network. In each iteration, each particle determines a neural network according to its own parameters, and trains the data. The convergence time of the network is often proportional to the complexity of the network structure, therefore, the fitness of the particle is determined by the accuracy and complexity of the network. When the network accuracy is high and the network structure is relatively simple, the fitness function can obtain a larger value.

[0115] The particle swarm optimization algorithm is used to optimize the network parameters of the identification model to obtain a trained identification model, the network parameters include the size and number of the convolution kernel of the time convolution network layer, the number of time convolution network layers, the learning rate and the batch size, and the five network parameters correspond to five search dimensions of the particle swarm optimization algorithm. A set of parameters of each particle corresponds to the parameters of a time convolution network. After initializing the particle swarm parameters, the velocity and position of the particle swarm are calculated according to the velocity formula and displacement formula, and the global optimal position and individual optimal position of the particle swarm are constantly updated until the maximum iteration number of the particle swarm is reached. After the particle swarm optimization algorithm ends, an optimal global position can be obtained, which corresponds to the best network parameters. The specific algorithm flow chart is shown in Figure 5 .

[0116] The initial parameters of the particle swarm are set, including the maximum iteration number of the particle swarm optimization algorithm, the number of particles, the search dimension, and two learning factors. For example, the maximum iteration number Iter of the particle swarm optimization algorithm can be set to 25, the number of particles is 10, the search dimension is 5, and the two learning factors c1 and c2 are set to 2. The iteration number of each particle corresponding to the network is set to 20 times.

[0117] The inertia weight is set according to the maximum iteration number:

[0118]

[0119] wherein w represents the inertia weight, w max represents the upper limit of the value of the inertia weight, which can be 0.9, and w min represents the lower limit of the value of the inertia weight, which can be 0.4, Iter represents the maximum iteration number, and iter represents the current iteration number.

[0120] The positions of the particles in the particle swarm in each search dimension are randomly initialized:

[0121] Xi = (m, n, l, lr, b)

[0122] wherein X i represents the position of the i-th particle, m represents the convolution kernel size of the time convolution network layer, n represents the number of convolution kernels of the time convolution network layer, l represents the number of time convolution network layers, lr represents the learning rate, and b represents the batch size. The position of the particle swarm is initialized, that is, a group of particles in the value range are randomly generated, and the value range of each search dimension is: m∈[2, 8], n∈[10, 200], l∈[2, 6], lr∈

[0123] [0.001, 0.01], and b∈[64, 256].

[0124] The speed of the particle in the particle swarm is randomly initialized in each search dimension, and the speed of the particle in each dimension is randomly generated, and the speed of the particle in a certain dimension is limited to be always less than 20% of the maximum value range of the dimension.

[0125] In the iteration process, the speed and position of each particle are constantly updated according to the population optimal position and the individual optimal position, and the update formula of the particle speed and position is as follows:

[0126]

[0127]

[0128] wherein, represents the speed of the i-th particle in dimension d at the k-th iteration, represents the speed of the i-th particle in dimension d at the k-1-th iteration, c1 and c2 are preset learning factors, r1 and r2 are two random numbers with a value range of [0, 1], and gbest d represents the value of the particle with the optimal fitness value in dimension d in all iterations in the history, and pbest id represents the value of the i-th particle in dimension d when the optimal fitness is obtained in all iterations, and the individual optimal position in dimension d;

[0129] The entire particle swarm will concentrate on the position with high fitness, and finally find the optimal particle position and determine the optimal network parameters. The particle swarm optimization algorithm is used to determine the convolution kernel size, the number of filters, the number of network layers, the learning rate, and the batch size and other parameters, which overcomes the low efficiency of traditional manual parameter adjustment, and is more likely to quickly find the global optimal parameters, and further improves the NLOS recognition accuracy of UWB. When the maximum number of iterations is reached, the corresponding optimal network parameters and the optimized recognition model are obtained, the training samples are input into the optimized recognition model, and the trained recognition model is obtained.

[0130] In one embodiment, the trained recognition model is used for recognition of ultra-wideband non-line-of-sight signals, comprising:

[0131] An ultra-wideband signal to be recognized is obtained, and the ultra-wideband signal is mapped into a channel impulse response sequence.

[0132] The channel impulse response sequence is input into a time convolution network layer to obtain a corresponding time sequence; the time step of the time sequence is the same as the time step of the channel impulse response sequence; the time convolution network layer includes a causal convolution layer, a dilated convolution layer, and a residual connection layer. Using the time convolution network is more suitable for processing time series signals, and the NLOS recognition accuracy of UWB is higher.

[0133] The input of the recognition model is a channel impulse response (CIR) sequence with a time step of n. When the CIR sequence length is low, the accuracy of the network gradually increases with the CIR length. Experiments show that when the CIR length increases to more than 250, the accuracy of the network remains basically unchanged. However, the average convergence time of the network increases linearly with the increase of the CIR length. Therefore, in practical applications, the network accuracy and network size can be balanced according to the needs. In application scenarios where the NLOS recognition accuracy is not high, the length of the CIR sequence used can be appropriately reduced to obtain smaller storage space occupation and faster processing speed. In application scenarios where accurate NLOS recognition is required, the length of the CIR sequence used can be increased, but attention should be paid to the upper limit of the network accuracy improvement. When the network size is increased to reach the upper limit, the calculation amount will be increased without significant improvement in network accuracy, for example, n=200 is selected in the present application.

[0134] The one-dimensional channel impulse response CIR input is first input into the time convolution network layer. The network operates on the CIR through causal convolution, dilated convolution, and residual connection and outputs a time sequence with the same time step. The causal convolution module is one of the important modules of the TCN network model. The causal convolution can be represented as Figure 6 Intuitively, the value at time t of the previous layer only depends on the values at time t and before of the next layer.

[0135] The receptive field of the simple causal convolution is limited by the size of the convolution kernel. If a longer time sequence, such as the CIR sequence in this paper, is to be processed, more layers need to be stacked. In Figure 6 Dilated convolution is used to make the convolution kernel applicable to a region beyond its own length by skipping part of the input, thereby effectively expanding the receptive field. For a one-dimensional input CIR sequence x∈R n and a filter f:{0,…,k-1}→R, the dilated convolution operation F on the sequence element s is defined as:

[0136]

[0137] Since the receptive field of a TCN depends on the network depth *l*, filter size *m*, and expansion factor *d*, deeper networks generally offer greater expressive power and better performance. However, this depth can also lead to problems such as vanishing or exploding gradients. Residual connections have been shown to effectively address these issues, making it easier to train deeper networks. Essentially, a residual network adds skip connections to a simple forward propagation layer, allowing the input of one layer to skip one or more layers and be passed directly to deeper layers of the neural network.

[0138] o = Activation(x + F(x))

[0139] The time series data is input into the attention layer, and the corresponding output is flattened before being input into the fully connected layer of the recognition model. The Sigmoid activation function is used to recognize ultra-wideband non-line-of-sight signals.

[0140] Self-attention mechanisms can strengthen or ignore information by adjusting weights and can dynamically adjust weights, thus enhancing the importance of selecting information in different situations. This invention uses a self-attention layer to analyze the vector output by the TCN layer, calculates the correlation between each input vector and all input vectors, and assigns different weights to different time steps.

[0141] The computation process of the self-attention mechanism can be divided into three steps: calculating the query vector, key vector, and value vector for each input vector; calculating the correlation α (i.e., self-attention score) between different input vectors; and extracting important information based on the self-attention score.

[0142] First, for a set of TCN layer output vectors Y = [y1, y2, ... y] of length n n ], each vector y i (i∈[1,n]) represents an input information. Each input vector y in Y... i Multiply by matrix W respectively q W k and W v Get the query vector q i key vector k i and the value vector v i .

[0143] Then, calculate a vector y i q i With another vector y j k j The dot product of the vectors and divided by a constant, which is typically the square root of the vector dimension, is denoted as: You can then get yi and y j between the relevance a ij At the same time, Softmax normalization needs to be done for each self-attention score:

[0144]

[0145]

[0146] Finally, for each input vector y i in Y, multiply it with the self-attention score a' ij of other input vectors y j and the corresponding value vector v j , and then accumulate to obtain the output vector b i :

[0147]

[0148] In the time convolution network, the self-attention layer is added, which can dynamically adjust the weight according to the importance of the influence of the input data on the classification, enhance the NLOS recognition accuracy of UWB, and the output of the self-attention layer is flattened and fully connected with a unit, so as to reduce the dimension of the extracted high-dimensional features. In combination with the Sigmoid activation function, the LOS and NLOS conditions are classified.

[0149] It should be understood that, although each step in the flowchart of Figure 1 is shown in sequence according to the arrow, these steps are not necessarily executed in sequence according to the arrow. Unless otherwise specified in this article, the execution of these steps has no strict order restriction, and these steps can be executed in other order. Moreover, Figure 1 At least part of the steps in may include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or sub-steps or stages of other steps.

[0150] Figure 7 In one embodiment, as shown in , an ultra-wideband non-line-of-sight signal recognition device is provided, comprising a discrete sequence construction module, an intermediate sample acquisition module, a data padding module, a model training module and a signal recognition module, wherein:

[0151] The discrete sequence construction module is used to acquire an original ultra-wideband signal sample set, and construct a one-dimensional array discrete sequence according to the original ultra-wideband signal samples in the original ultra-wideband signal sample set; the one-dimensional array discrete sequence is connected in series;

[0152] an intermediate sample obtaining module, configured to determine a data deletion discrete interval of each original ultra-wideband signal sample according to the one-dimensional array discrete sequence, and obtain a plurality of intermediate ultra-wideband signal samples according to the data deletion discrete interval;

[0153] a data filling module, configured to fill each data deletion discrete interval in the intermediate ultra-wideband signal sample with data randomly selected from other original ultra-wideband signal samples to obtain a plurality of hybrid ultra-wideband signal samples; the other original ultra-wideband signal samples are any original ultra-wideband signal sample other than the original ultra-wideband signal sample corresponding to the current intermediate ultra-wideband signal sample; the length of the randomly selected data is equal to the length of the corresponding data deletion discrete interval;

[0154] a model training module, configured to train a pre-constructed identification model by using the hybrid ultra-wideband signal samples to obtain a trained identification model;

[0155] a signal identification module, configured to identify the ultra-wideband non-line-of-sight signal by using the trained identification model.

[0156] In an embodiment, the discrete sequence constructing module is further configured to design the number of one-dimensional arrays in the one-dimensional array discrete sequence and the interval variation interval of adjacent one-dimensional arrays;

[0157] randomly taking a value in the interval variation interval as the interval of each pair of adjacent one-dimensional arrays;

[0158] constructing a data deletion rate according to the sum of the data length of the original ultra-wideband signal sample and the interval of each one-dimensional array discrete sequence;

[0159] designing the data length of each one-dimensional array according to the pre-set data deletion rate, the data length of the original ultra-wideband signal sample and the number of one-dimensional arrays in the one-dimensional array discrete sequence;

[0160] constructing the one-dimensional array discrete sequence according to the number of one-dimensional arrays, the interval of each pair of adjacent one-dimensional arrays and the data length of each one-dimensional array.

[0161] In an embodiment, the discrete sequence constructing module is further configured to design the data length of each one-dimensional array according to the pre-set data deletion rate, the data length of the original ultra-wideband signal sample and the number of one-dimensional arrays in the one-dimensional array discrete sequence as follows:

[0162]

[0163] wherein, l rrepresents the data length of the rth one-dimensional array, c represents the data deletion rate, and R represents the number of one-dimensional arrays in the one-dimensional array discrete sequence.

[0164] In one embodiment, the intermediate sample acquisition module is further configured to align the one-dimensional array discrete sequence with each original ultra-wideband signal sample respectively;

[0165] According to the position of each one-dimensional array in the one-dimensional array discrete sequence, a data deletion discrete interval in the original ultra-wideband signal sample is determined.

[0166] According to the data deletion interval, the data in the corresponding original ultra-wideband signal sample is partially deleted to obtain a plurality of corresponding intermediate ultra-wideband signal samples.

[0167] In one embodiment, the intermediate sample acquisition module is further configured to obtain a plurality of corresponding intermediate ultra-wideband signal samples by partially deleting the data in the corresponding original ultra-wideband signal sample according to the data deletion interval.

[0168]

[0169] wherein, represents an intermediate ultra-wideband signal sample, X represents an original ultra-wideband signal sample, and M is an element of {0, 1} W represents the data processing condition at the corresponding position, M = 0 represents deletion, M = 1 represents retention, W represents the data length of the original ultra-wideband signal sample, and {W i | M = 0} is the data deletion interval.

[0170] In one embodiment, the model training module is further configured to obtain a pre-constructed identification model; the identification model includes a time convolution network layer and a self-attention layer.

[0171] The mixed ultra-wideband signal sample is used as a training sample of the identification model.

[0172] The training sample is input into the identification model, an adaptability function is constructed according to the identification accuracy and convergence time of the identification model, a particle swarm optimization algorithm is adopted to optimize the network parameters of the identification model to obtain a trained identification model, and the network parameters include the convolution kernel size and number of the time convolution network layer, the number of time convolution network layers, the learning rate, and the batch size.

[0173] In one embodiment, the model training module is further configured to construct the adaptability function according to the identification accuracy and convergence time of the identification model as:

[0174] f(X) = a(X) - a-t(X)

[0175] wherein a(X) represents the recognition accuracy of the recognition model corresponding to the particle, t(X) represents the convergence time of the recognition model, and a represents a preset weight coefficient;

[0176] Set initial parameters of the particle swarm; the initial parameters include the maximum number of iterations of the particle swarm optimization algorithm, the number of particles, the search dimension, and two learning factors;

[0177] Set the inertia weight according to the maximum number of iterations:

[0178]

[0179] wherein w represents the inertia weight, w max represents the upper limit of the value of the inertia weight, w min represents the lower limit of the value of the inertia weight, Iter represents the maximum number of iterations, and iter represents the current number of iterations;

[0180] Randomly initialize the position of a particle in each search dimension in the particle swarm:

[0181] X i = (m, n, l, lr, b)

[0182] wherein X i represents the position of the i-th particle, m represents the convolution kernel size of the time convolution network layer, n represents the number of convolution kernels of the time convolution network layer, l represents the number of time convolution network layers, lr represents the learning rate, and b represents the batch size;

[0183] Randomly initialize the speed of a particle in each search dimension in the particle swarm;

[0184] In the iteration process, the speed and position of each particle are constantly updated according to the population optimal position and the individual optimal position, and the update formula of the particle speed and position is as follows:

[0185]

[0186]

[0187] wherein, represents the speed of the i-th particle in dimension d at the k-th iteration, represents the speed of the i-th particle in dimension d at the k-1-th iteration, c1 and c2 are preset learning factors, r1 and r2 are two random numbers with a value range of [0, 1], gbest d represents the value of the particle with the optimal fitness value in dimension d among all particles in all iterations in history, and pbest id represents the value of the i-th particle in dimension d when the optimal fitness is obtained in all iterations;

[0188] When the maximum number of iterations is reached, the corresponding optimal network parameters and the optimized identification model are obtained;

[0189] The training sample is input into the optimized identification model to obtain a trained identification model.

[0190] In one embodiment, the signal identification module is further configured to obtain an ultra-wideband signal to be identified, and map the ultra-wideband signal into a channel impulse response sequence;

[0191] The channel impulse response sequence is input into a time convolution network layer to obtain a corresponding time sequence; the time step of the time sequence is the same as the time step of the channel impulse response sequence; the time convolution network layer includes a causal convolution layer, a dilated convolution layer, and a residual connection layer;

[0192] The time sequence is input into a self-attention layer and the corresponding output result is flattened and input into a fully connected layer of the identification model, and a Sigmoid activation function is used to identify the ultra-wideband non-line-of-sight signal.

[0193] The specific limitations of the ultra-wideband non-line-of-sight signal identification device can be referred to the limitations of the ultra-wideband non-line-of-sight signal identification method described above, which will not be repeated here. Each module in the above ultra-wideband non-line-of-sight signal identification device can be realized by software, hardware, and a combination thereof, in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0194] In one embodiment, a computer device is provided, which can be a server, and its internal structure diagram can be as shown in Figure 8 The computer device includes a processor, a memory, a network interface, and a database connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The database of the computer device is configured to store ultra-wideband signal data. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement an ultra-wideband non-line-of-sight signal identification method.

[0195] Those skilled in the art can understand that, Figure 8The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0196] In one embodiment, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the method in the above embodiments when executing the computer program.

[0197] A person of ordinary skill in the art can understand that all or part of the processes in the above embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium and can include the processes of the above embodiments when executed. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synch link) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0198] Any combination of the technical features of the above embodiments can be made. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0199] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the patent protection scope of the present application should be subject to the appended claims.

Claims

1. A method for identifying ultra-wideband non-line-of-sight signals, characterized in that, The method comprises: acquiring an original ultra-wideband signal sample set, and constructing a one-dimensional array discrete sequence according to original ultra-wideband signal samples in the original ultra-wideband signal sample set; the one-dimensional array discrete sequence is connected in series by one-dimensional arrays; designing the number of one-dimensional arrays in the one-dimensional array discrete sequence and a spacing variation interval of adjacent one-dimensional arrays; randomly selecting values in the spacing variation interval as the spacing of each pair of adjacent one-dimensional arrays; constructing a data deletion rate according to the data length of the original ultra-wideband signal samples in the original ultra-wideband signal sample set and the sum of the spacings of the one-dimensional array discrete sequences; designing the data length of each one-dimensional array according to the pre-set data deletion rate, the data length of the original ultra-wideband signal samples and the number of one-dimensional arrays in the one-dimensional array discrete sequence; constructing a one-dimensional array discrete sequence according to the number of one-dimensional arrays, the spacing of each pair of adjacent one-dimensional arrays and the data length of each one-dimensional array; determining a data deletion discrete interval of each original ultra-wideband signal sample according to the one-dimensional array discrete sequence, and obtaining a plurality of intermediate ultra-wideband signal samples corresponding to the data deletion discrete interval; aligning the one-dimensional array discrete sequence with each original ultra-wideband signal sample respectively; determining the data deletion discrete interval in the original ultra-wideband signal sample according to the position of each one-dimensional array in the one-dimensional array discrete sequence; locally deleting data in the corresponding original ultra-wideband signal sample according to the data deletion discrete interval to obtain a plurality of intermediate ultra-wideband signal samples corresponding to the data deletion discrete interval; for each data deletion discrete interval in the intermediate ultra-wideband signal sample, randomly selecting a piece of data from other original ultra-wideband signal samples to fill the data deletion discrete interval to obtain a plurality of mixed ultra-wideband signal samples corresponding to the data deletion discrete interval; the other original ultra-wideband signal samples are any original ultra-wideband signal sample other than the original ultra-wideband signal sample corresponding to the current intermediate ultra-wideband signal sample; the length of the randomly selected data is equal to the length of the corresponding data deletion discrete interval; training a pre-constructed identification model using the mixed ultra-wideband signal samples to obtain a trained identification model; using the trained identification model to identify an ultra-wideband non-line-of-sight signal.

2. The method of claim 1, wherein, Training a pre-constructed identification model using the mixed ultra-wideband signal samples to obtain a trained identification model, comprising: acquiring a pre-constructed identification model; the identification model comprises a time convolution network layer and a self-attention layer; using the mixed ultra-wideband signal samples as training samples of the identification model; inputting the training samples into the identification model, constructing a fitness function according to the identification accuracy and convergence time of the identification model, and optimizing the network parameters of the identification model to obtain a trained identification model using a particle swarm optimization algorithm with the maximum fitness function as the target; the network parameters include the size and number of convolution kernels of the time convolution network layer, the number of time convolution network layers, the learning rate and the batch size.

3. The method of claim 2, wherein, The process of inputting the training samples into the recognition model, constructing a fitness function based on the recognition model's recognition accuracy and convergence time, and optimizing the network parameters of the recognition model using a particle swarm optimization algorithm with the objective of maximizing the fitness function to obtain a trained recognition model includes: Based on the recognition accuracy and convergence time of the recognition model, the fitness function is constructed as follows: wherein, represents the recognition accuracy of the recognition model corresponding to the particle, represents the convergence time of the recognition model, represents a preset weight coefficient; Set the initial parameters for the particle swarm optimization algorithm; the initial parameters include the maximum number of iterations of the particle swarm optimization algorithm, the number of particles, the search dimension, and two learning factors; Set the inertia weight based on the maximum number of iterations: wherein, denotes an inertia weight, denotes an upper bound of the inertia weight, denotes a lower bound of the inertia weight, denotes a maximum number of iterations, denotes a current number of iterations; Randomly initialize the positions of particles in the particle swarm across each search dimension: wherein, represents the position of the th particle, represents the convolution kernel size of the time convolution network layer, represents the number of convolution kernels of the time convolution network layer, represents the number of time convolution network layers, represents the learning rate, represents the batch size; Randomly initialize the velocities of particles in the particle swarm across each search dimension; During the iteration process, each particle continuously updates its velocity and position based on the population-optimal position and the individual-optimal position. The formulas for updating particle velocity and position are as follows: wherein, denotes the velocity of the i-th particle in dimension d at the j-th iteration, denotes the velocity of the i-th particle in dimension d at the j-th iteration, , is a preset learning factor, , are two random numbers with values in the range [0, 1], represents the value of the particle with the best fitness value in all iterations of all particles in history in dimension d, represents the value of the particle in dimension d when it obtains the best fitness in all iterations;​​​​​​ When the maximum number of iterations is reached, obtain the corresponding optimal network parameters and optimize the recognition model; The training samples are input into the optimized recognition model to obtain the trained recognition model.

4. The method of claim 2, wherein, The process of using the trained recognition model to identify ultra-wideband non-line-of-sight signals includes: Acquire the ultra-wideband signal to be identified, and map the ultra-wideband signal into a channel impulse response sequence; The channel impulse response sequence is input into the temporal convolutional network layer to obtain the corresponding time series; the time step of the time series is the same as the time step of the channel impulse response sequence; the temporal convolutional network layer includes a causal convolutional layer, a dilated convolutional layer, and a residual connection layer. The time series is input into the self-attention layer, and the corresponding output is flattened before being input into the fully connected layer of the recognition model. The Sigmoid activation function is used to recognize ultra-wideband non-line-of-sight signals.

5. The method of claim 2, wherein, Based on the data deletion discrete interval, the data in the corresponding original ultra-wideband signal sample is locally deleted to obtain multiple corresponding intermediate ultra-wideband signal samples, including: Based on the data deletion discrete interval, the data in the corresponding original ultra-wideband signal sample is locally deleted to obtain multiple corresponding intermediate ultra-wideband signal samples: wherein, represents an intermediate ultra-wideband signal sample, represents an original ultra-wideband signal sample, represents a data processing condition at a corresponding position, represents deletion, represents retention, represents a data length of the original ultra-wideband signal sample, i.e., a data deletion discrete interval.

6. The method of claim 2, wherein, The data length of each one-dimensional array is designed based on a pre-set data deletion rate, the data length of the original ultra-wideband signal sample, and the number of one-dimensional arrays in the one-dimensional array discrete sequence, including: Set all one-dimensional arrays in the discrete sequence of the one-dimensional array to have the same length; Based on the preset data deletion rate, the data length of the original ultra-wideband signal sample, and the number of one-dimensional arrays in the one-dimensional array discrete sequence, the data length of each one-dimensional array is designed as follows: wherein, represents the data length of the first one-dimensional array, represents the data deletion rate, represents the number of one-dimensional arrays in the discrete sequence of one-dimensional arrays.

7. An ultra-wideband non-line-of-sight signal identification apparatus, comprising: The ultra-wideband non-line-of-sight signal identification device is used to implement the method according to any one of claims 1 to 6, the device comprising: A discrete sequence construction module is used to obtain an original ultra-wideband signal sample set and construct a one-dimensional array discrete sequence based on the original ultra-wideband signal samples in the original ultra-wideband signal sample set; the one-dimensional arrays in the one-dimensional array discrete sequence are connected serially. An intermediate sample obtaining module is configured to determine a data deletion discrete interval of each of the original UWB signal samples according to the one-dimensional array discrete sequence, and obtain a plurality of intermediate UWB signal samples according to the data deletion discrete interval; A data filling module is configured to fill each of the data deletion discrete intervals in the intermediate UWB signal samples with data randomly selected from other original UWB signal samples to obtain a plurality of mixed UWB signal samples; the other original UWB signal samples are any original UWB signal samples other than the original UWB signal sample corresponding to the current intermediate UWB signal sample; the length of the randomly selected data is equal to the length of the corresponding data deletion discrete interval; A model training module is configured to train a pre-constructed recognition model using the mixed UWB signal samples to obtain a trained recognition model; A signal recognition module is configured to recognize UWB NLOS signals using the trained recognition model.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the steps of the method in any one of claims 1 to 6.

Images