| CPC G01S 7/40 (2013.01) [G01S 13/02 (2013.01); G06N 3/0442 (2023.01); G06N 3/0464 (2023.01); G06N 3/09 (2023.01); H01Q 3/267 (2013.01); G01S 2013/0245 (2013.01); H01Q 1/288 (2013.01)] | 8 Claims |
|
1. A phased array automatic calibration device based on convolutional neural network-long short-term memory (CNN-LSTM) using power measurement, comprising: at least one processor; and a memory communicatively coupled to the at least one processor, wherein the memory stores program instructions executable by the at least one processor, and the processor is configured to:
S1: model a phased array based on an array signal theory, set a amplitude and phase for each channel randomly by program within an amplitude-phase error range, calculate and store corresponding array radiated field data, repeat above operations to generate phased array signal data with amplitude-phase error information that meets a preset quantity condition;
S2: preprocess the phased array signal data to obtain a training dataset;
S3: obtain a calibration model by building a CNN-LSTM neural network model and input training data containing a label into a network for training until the network is in a state of convergence; or
obtain an initial calibration model from a cloud server, the initial calibration model being obtained by training a CNN-LSTM neural network model constructed based on neural network structure parameters by the cloud server; the neural network structure parameters being obtained base on feature groupings of a phased array to be calibrated;
S4: control an array elements of the phased array to be calibrated to shift phase;
measure the phased array to be calibrated, preprocess acquired measured data, and input the acquired measured data into the calibration model to obtain a calibration result;
wherein an input of the calibration model further includes a sequence of interference features in an environment in which the phased array to be calibrated is, wherein the interference features are obtained based on a terminal processing device being captured by an environmental monitoring device; and the processor is further configured to:
determine an acquisition frequency of the interference features based on a number of array elements in a phased array feature, and an average frequency of measured data after correction processing;
control the environmental monitoring device to obtain the interference features based on the acquisition frequency;
wherein the S3 is accomplished by following sub-steps:
S3.1: building the CNN-LSTM neural network model;
S3.2: designing a loss function for training the CNN-LSTM neural network model;
S3.3: obtaining a phased array calibration model by inputting the training dataset with the label into the network for training until the network is in the state of convergence;
the S3.1 is accomplished by following sub-steps:
S3.1.1: the CNN-LSTM neural network including a Convolutional Neural Network (CNN), a Long Short-Term Memory (LSTM), and a complex fully connected network, wherein the CNN is configured to adaptively extract spatial feature information from input sequence data, and the LSTM is configured to utilize spatial features extracted by the CNN and combine temporal information of the sequence data for automatic modeling, and output a final prediction result by using the complex fully connected network immediately after the LSTM;
the S3.2 is accomplished by following sub-steps:
S3.2.1: calculating mean square errors for both the real and imaginary parts, respectively, and adding the two mean square errors as the loss function, the loss function is denoted as:
 wherein; and ŷi denote yi predicted value and a true value, respectively,
 (·) denotes obtaining a complex real part, and  (·) denotes obtaining a complex imaginary part;the S1 is accomplished by following sub-steps:
S1.1: establishing a phased array signal model; setting antenna to be measured as a two-dimensional planar phased array of size Nx×Ny, arranged in a rectangle, describing a spatial orientation in terms of a pitch angle and an azimuth angle, and when phased array beam to be measured is directed to (θs,ϕs), a radiated electric field of the phased array in a measurement orientation (θ, ϕ) being expressed as:
 wherein k=2π/λ,λ is wavelength dx and dy denote a row spacing and column spacing of array element, respectively, m and n denote a row and column number of the array element, respectively, Gn,m(θ,ϕ) and In,m denote a normalized independent directivity coefficient and a complex excitation including an amplitude-phase error of array element (m, n), respectively, and In,m is denoted as:
 wherein αn,m is a relative amplitude of initial complex excitation of the array element (m, n), a theoretical range of αn,m is 0 to 1, and δn,m denotes a phase of the initial complex excitation of the array element (m, n), a theoretical range of δn,m is −180° to +180°, data for training the model is generated based on signal models described in formula (2) and formula (3);
S1.2: taking (θs,ϕs)=(0,0) for beam pointing and (0,ϕ)=(0,0) for observation orientation, making measured array beam pointing in a normal direction, and placing a probe antenna in the normal direction for measuring, wherein measured value is E(0,0), making Gn,m(θ, ϕ)=1, substituting (θs, ϕ)=(0,0) and (θ,ϕ)=(0,0) according to the formula (2) and (3) to obtain:
 wherein the formula (4) is Nx×Ny cumulative form, there is no sequential relationship among the items, and the items of the formula (4) are rearranged and rewritten to obtain:
 wherein N=Nx×Ny denotes a total number of array elements, and the formula (5) is written in matrix form as:
 wherein Q denotes an array initial complex excitation matrix containing error components, and S denotes a phase setting matrix of measurement, a count of columns of the phase setting matrix are equal to a count of array elements N and the count of rows is determined by a count of measurements, each row describes a phase shift value of each array element in a single measurement, elements of S are continuously changed during a calibration process, simultaneously measuring and recording corresponding E(0,0) to obtain a data sequence, and estimating the array initial complex excitation matrix Q based on the data sequence;
S1.3: solving the array initial complex excitation matrix Q by performing 2N+1 measurements, wherein all array elements are kept in an initial state without phase shifting for a first measurement, and after which each array element is sequentially phase shifted by 90° and 180° for measurement, wherein the 2N+1 measurements corresponding to the phase setting matrix S is expressed as:
 wherein when generating the training data by simulating measurements with the program, each element of the array initial complex excitation matrix Q in the formula (6) is randomly generated in accordance with a preset error range, a amplitude of each element ranges from 0.1 to 1, and a phase of each element ranges from −180° to +180°, and after generating a set of initial excitations randomly each time to obtain the array initial complex excitation matrix Q substituting the phase setting matrix S and the array initial complex excitation matrix Q into the formula (6) to calculate an electric field vector sequence with a length of 2N+1, and representing the sequence by a matrix E of size (2N+1)×1:
 wherein each element Ei=Ei=(0,0)i, i=0,1 . . . ,2N+1 represents a radiation electric field vector measured by phase-shifting the phased array according to i-th row of the phase alignment matrix S;
S1.4: randomly generating the array initial complex excitation matrix Q calculating the electric field vector sequence E by substituting the array initial complex excitation matrix Q and the phase setting matrix S in the formula (7) into the formula (6), and saving the array initial complex excitation matrix Q and the electric field vector sequence E one by one corresponding to each other, and repeating the S1.4 until generated data meets a training requirement, and completing acquisition of raw data.
|