Starlink terminal signal recognition method and device based on multi-dimensional feature fusion and CNN-LSTM

Abstract

The present invention discloses a Starlink terminal signal recognition method and device based on multi-dimensional feature fusion and CNN-LSTM, including: extracting multi-dimensional features from the filtered Starlink terminal signal data, obtaining a multi-dimensional feature matrix according to high-order cumulative quantities and OFMD signal parameter features; sending the multi-dimensional feature matrix to a convolutional neural network, which performs feature fusion and perception on the original data through a convolution kernel, extracts the deep features of the input waveform data using local connections and weight sharing, and performs pooling operations on the convolved data by a pooling layer, and performs regularization after two convolutional pooling layers to obtain high-level abstract features; sending the regularized data to a long short-term memory network to obtain time series data features; inputting the obtained time series data features into a fully connected layer through two layers of Dense, and classifying the output of the fully connected layer using a Softmax function. The present invention realizes the classification and recognition of beacon signals and data transmission signals within the Starlink frequency band.

Description

Technical Field

[0001] The present invention relates to the field of signal processing, and specifically to a Starlink terminal signal recognition method and device based on multi-dimensional feature fusion and CNN-LSTM. Background Art

[0002] In recent years, with the development of satellite communication technology and changes in the Internet application environment, in response to the current limited coverage of the ground network as a whole, insufficient fiber penetration, low network speed, and especially the shortage of broadband in rural areas, a wave of satellite Internet development has emerged in the world, which is positioned to provide Internet services to people who are difficult to cover by fiber and mobile phone base stations. Starlink is the most representative of all current satellite Internet constellations. It has the largest planned network scale, the largest number of satellites launched, and independently owns the ownership and control of the entire industry chain. Therefore, it is necessary to focus on the signal characteristics of the Starlink terminal, which has become the key point of current signal identification. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a Starlink terminal signal recognition method based on multi-dimensional feature fusion and CNN-LSTM, comprising the following steps:

[0004] Step 1: extract multi-dimensional features from the filtered Starlink terminal signal data, where the multi-dimensional features include high-order cumulants and OFMD signal parameter features, and obtain a multi-dimensional feature matrix based on the high-order cumulants and OFMD signal parameter features;

[0005] Step 2: The multi-dimensional feature matrix is ​​fed into the convolutional neural network. The convolutional neural network uses the convolution kernel to fuse and perceive the original data, and uses local connections and weight sharing to extract the deep features of the input waveform data. The pooling layer performs a pooling operation on the convolved data, and regularizes it after two convolutional pooling layers to obtain high-level abstract features.

[0006] Step 3: Send the regularized data into the long short-term memory network to obtain the time series data features;

[0007] Step 4: Input the obtained time series data features into the fully connected layer through two layers of Dense, and use the Softmax function to classify the output of the fully connected layer.

[0008] Furthermore, the high-order cumulant extraction includes:

[0009] By collecting the original signal as an IQ data set, the received data is:

[0010] S(i)=S I (i)+S Q (i)

[0011] Where S(i) represents the i-th signal sample; S I (i) and S Q (i) respectively represent the in-phase component and the orthogonal component of the i-th signal; through the cumulant-moment formula and the moment-cumulant formula, the P-order moment and q-order cumulant of the zero-mean complex random signal S[i] are expressed as:

[0012]

[0013] In the formula, * represents complex conjugation; p is the order (p>q); q is the number of sequences to take complex conjugation; E represents the mean;

[0014] The high-order cumulant expression of S(i) is obtained as:

[0015] C 21 =M 21

[0016]

[0017] C 41 =M 41 -3*M 20 *M 21

[0018]

[0019] Furthermore, the OFMD signal parameter feature extraction includes:

[0020] Let the OFDM signal in baseband be:

[0021]

[0022] In the formula, m a,b represents the ath OFDM signal on the bth subcarrier, N represents the number of subcarriers, P is the signal power factor, h(t) is the pulse function, T s Indicates the symbol period length, T s =Tu+Tg, where Tu represents the useful symbol length, Tg is the cyclic prefix, Δf represents the interval between subcarriers, and Δf=1 / Tu;

[0023] The Tu value is obtained according to Tu=Fs / Nu, where Fs is the signal sampling rate;

[0024] After the signal is delayed by Tu, it is autocorrelated. The amplitude spectrum of the autocorrelation will have equally spaced peaks, and the equally spaced difference is the number of symbol points Ns.

[0025] According to T s =Fs / Ns to get T s Value, where Fs is the signal sampling rate;

[0026] According to Tg = T s -Tu obtains the cyclic prefix length;

[0027] The interval between subcarriers is obtained according to Δf=1 / Tu;

[0028] The number of subcarriers is obtained according to the formula SubCarryNum = Fs / Δf;

[0029] Extract the OFDM signal symbol period length T s , cyclic prefix Tg, interval Δf between subcarriers, and number of subcarriers SubCarryNum.

[0030] Furthermore, the multi-dimensional feature matrix is ​​obtained according to the high-order cumulants and OFDM signal parameter features, including: fusing the 8 parameter features of the high-order cumulants and the OFDM signal parameter features to construct a 12-dimensional feature matrix V = [C 21 ,C 40 ,C 41 ,C 42 ,C 60 ,C 61 ,C 63 ,C 80 ,T s , Tg, Δf, SubCarryNum].

[0031] Furthermore, the regularized data is fed into the long short-term memory network to obtain time series data features, including:

[0032] The main component of the long short-term memory network recurrent unit structure is the forget gate f t , input gate i t And the output gate o t , where the forget gate is:

[0033] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0034] where f t is the output of the forget gate, W f is the linear coefficient, b i is the bias coefficient, σ represents the sigmoid function, h t-1 is the output at the previous moment, and the input signal x t ;

[0035] The input gate is:

[0036] i t=σ(W i ·[h t-1 ,x t ]+b i )

[0037]

[0038] Among them, i t and are the outputs of the Sigmoid activation function and the tanh activation function respectively; W i , W C is the linear coefficient; b i 、b C is the bias coefficient, and the LSTM unit state is:

[0039]

[0040] Among them, C t is the current LSTM unit state, ⊙ represents the Hadamard product;

[0041] The output gate is:

[0042] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0043] h t =o t ⊙tanh(C t )

[0044] Among them, t is the output gate output, h t is the hidden layer output, Wo is the weight, and bo is the initial bias coefficient.

[0045] A Starlink terminal signal recognition system based on multi-dimensional feature fusion and CNN-LSTM, applying any of the Starlink terminal signal recognition methods based on multi-dimensional feature fusion and CNN-LSTM, comprising: a nine-element uniform phased circular array, a switch matrix, a wide-band and large-bandwidth receiver, a computer, and a radio frequency switch driver;

[0046] The nine-element uniform phased circular array uses a nine-element phased array antenna, and the directional pattern of the circular array is 360° omnidirectional without array ambiguity;

[0047] The switch matrix connects different antenna combinations through program control to achieve automatic synthesis of antenna patterns;

[0048] The radio frequency switch driver is used to control the working state of the matrix switch through the radio frequency switch driver according to the instructions given by the computer;

[0049] The wide-band and wide-bandwidth receiver provides the measured signal data received by the two antenna array elements to the computer for processing;

[0050] The computer: The computer realizes human-computer dialogue, controls the start and end of data collection, processes the collected data through software, and displays the results to the user through the software interface after the collected signals.

[0051] The beneficial effect of the present invention is that the beacon signals and data transmission signals within the Starlink frequency band can be classified and identified through the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flowchart of the Starlink terminal signal recognition method based on multi-dimensional feature fusion and CNN-LSTM;

[0053] Figure 2 This is the structural diagram of the 16-element phased array Starlink terminal signal acquisition system;

[0054] Figure 3 Provides the signal anti-interference algorithm workflow;

[0055] Figure 4 The workflow of the Starlink terminal signal classification and recognition algorithm. DETAILED DESCRIPTION

[0056] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0057] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0058] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative work are within the scope of protection of the present invention. It should be noted that relational terms such as the terms "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0059] Moreover, the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes a list of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or apparatus. In the absence of more restrictions, an element defined by the phrase "comprises a ..." does not exclude the presence of additional identical elements in the process, method, article or apparatus that includes the element.

[0060] The features and performance of the present invention are further described in detail below in conjunction with the embodiments.

[0061] like Figure 1 As shown in the figure, the Starlink terminal signal recognition method based on multi-dimensional feature fusion and CNN-LSTM includes the following steps:

[0062] Step 1: extract multi-dimensional features from the filtered Starlink terminal signal data, where the multi-dimensional features include high-order cumulants and OFMD signal parameter features, and obtain a multi-dimensional feature matrix based on the high-order cumulants and OFMD signal parameter features;

[0063] Step 2: The multi-dimensional feature matrix is ​​fed into the convolutional neural network. The convolutional neural network uses the convolution kernel to fuse and perceive the original data, and uses local connections and weight sharing to extract the deep features of the input waveform data. The pooling layer performs a pooling operation on the convolved data, and regularizes it after two convolutional pooling layers to obtain high-level abstract features.

[0064] Step 3: Send the regularized data into the long short-term memory network to obtain the time series data features;

[0065] Step 4: Input the obtained time series data features into the fully connected layer through two layers of Dense, and use the Softmax function to classify the output of the fully connected layer.

[0066] The high-order cumulant extraction includes:

[0067] By collecting the original signal as an IQ data set, the received data is:

[0068] S(i)=S I (i)+S Q (i)

[0069] Where S(i) represents the i-th signal sample; S I (i) and S Q(i) respectively represent the in-phase component and the orthogonal component of the i-th signal; through the cumulant-moment formula and the moment-cumulant formula, the P-order moment and q-order cumulant of the zero-mean complex random signal S[i] are expressed as:

[0070]

[0071] In the formula, * represents complex conjugation; p is the order (p>q); q is the number of sequences to take complex conjugation; E represents the mean;

[0072] The high-order cumulant expression of S(i) is obtained as:

[0073] C 21 =M 21

[0074]

[0075] C 41 =M 41 -3*M 20 *M 21

[0076]

[0077] The OFMD signal parameter feature extraction includes:

[0078] Let the OFDM signal in baseband be:

[0079]

[0080] In the formula, m a,b represents the ath OFDM signal on the bth subcarrier, N represents the number of subcarriers, P is the signal power factor, h(t) is the pulse function, T s Indicates the symbol period length, T s =Tu+Tg, where Tu represents the useful symbol length, Tg is the cyclic prefix, Δf represents the interval between subcarriers, and Δf=1 / Tu;

[0081] The Tu value is obtained according to Tu=Fs / Nu, where Fs is the signal sampling rate;

[0082] After the signal is delayed by Tu, it is autocorrelated. The amplitude spectrum of the autocorrelation will have equally spaced peaks, and the equally spaced difference is the number of symbol points Ns.

[0083] According to T s =Fs / Ns to get T s Value, where Fs is the signal sampling rate;

[0084] According to Tg = T s -Tu obtains the cyclic prefix length;

[0085] The interval between subcarriers is obtained according to Δf=1 / Tu;

[0086] The number of subcarriers is obtained according to the formula SubCarryNum = Fs / Δf;

[0087] Extract the OFDM signal symbol period length T s , cyclic prefix Tg, interval Δf between subcarriers, and number of subcarriers SubCarryNum.

[0088] The multi-dimensional feature matrix is ​​obtained according to the high-order cumulants and OFDM signal parameter features, including: fusing the 8 parameter features of the high-order cumulants and the OFDM signal parameter features to construct a 12-dimensional feature matrix V = [C 21 ,C 40 ,C 41 ,C 42 ,C 60 ,C 61 ,C 63 ,C 80 ,T s , Tg, Δf, SubCarryNum].

[0089] The regularized data is fed into the long short-term memory network to obtain time series data features, including:

[0090] The main component of the long short-term memory network recurrent unit structure is the forget gate f t , input gate i t And the output gate o t , where the forget gate is:

[0091] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0092] where f t is the output of the forget gate, W f is the linear coefficient, b i is the bias coefficient, σ represents the sigmoid function, h t-1 is the output at the previous moment, and the input signal x t ;

[0093] The input gate is:

[0094] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0095]

[0096] Among them, i t and are the outputs of the Sigmoid activation function and the tanh activation function respectively; W i , W C is the linear coefficient; b i , b C is the bias coefficient, and the LSTM unit state is:

[0097]

[0098] Among them, C t is the current LSTM unit state, ⊙ represents the Hadamard product;

[0099] The output gate is:

[0100] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0101] h t =o t ⊙tanh(C t )

[0102] Among them, t is the output gate output, h t is the hidden layer output, Wo is the weight, and bo is the initial bias coefficient.

[0103] A Starlink terminal signal recognition system based on multi-dimensional feature fusion and CNN-LSTM, and a Starlink terminal signal recognition method based on multi-dimensional feature fusion and CNN-LSTM as described in the application claims, comprising: a nine-element uniform phased circular array, a switch matrix, a wide-band and large-bandwidth receiver, a computer, and a radio frequency switch driver;

[0104] The nine-element uniform phased circular array uses a nine-element phased array antenna, and the directional pattern of the circular array is 360° omnidirectional without array ambiguity;

[0105] The switch matrix connects different antenna combinations through program control to achieve automatic synthesis of antenna patterns;

[0106] The radio frequency switch driver is used to control the working state of the matrix switch through the radio frequency switch driver according to the instructions given by the computer;

[0107] The wide-band and wide-bandwidth receiver provides the measured signal data received by the two antenna array elements to the computer for processing;

[0108] The computer: The computer realizes human-computer dialogue, controls the start and end of data collection, processes the collected data through software, and displays the results to the user through the software interface after the collected signals.

[0109] Specifically, Figure 2 As shown, the nine-element phased array Starlink terminal signal acquisition system. Figure 2 It can be seen that the equipment includes: 16-element uniform phased circular array, switch matrix, wide-band and large-bandwidth receiver, computer, RF switch driver and other hardware components.

[0110] Circular array: Using a 16-element phased array antenna, the circular array's directional pattern is 360° all-round without array ambiguity. Through the phased array synthesis method, the antenna obtains high gain and improves the system monitoring sensitivity. This design not only solves the problem that the traditional linear array cannot omnidirectionally monitor the Starlink terminal signal, but also solves the problem of insufficient system monitoring sensitivity. It provides a powerful means for the subsequent optimization of Starlink terminal signal data.

[0111] Switch matrix: Through program control, different antenna combinations are connected to achieve automatic synthesis of antenna patterns, thereby achieving the purpose of high system gain.

[0112] RF switch driver: The instructions given by the computer control the working state of the matrix switch through the RF switch driver.

[0113] Wide-band and large-bandwidth receiver: This system uses a wide-band and large-bandwidth receiver to provide the measured signal data received by two antenna array elements in a time-sharing manner to the computer for processing.

[0114] Computer: The computer realizes human-computer dialogue, controls the start and end of data collection, and processes the collected data through software. After the signal is collected, the result is intuitively displayed to the user through the software interface.

[0115] Figure 3 This is the workflow of the signal anti-interference algorithm. After the digital intermediate frequency signal is converted into the data baseband IQ through digital down-conversion, the handshake protocol signal is first sent to communicate with the antenna control board. Then, the antennas that need to be connected are encoded and the dwell time parameters are set. Secondly, the anti-interference processing is performed on the data collected by each group of antennas. Finally, it is determined whether all the encoded antennas are connected. The specific steps of anti-interference processing are as follows:

[0116] Step 1. Send a handshake protocol signal to communicate with the antenna control board. If it fails, start the delayed retransmission mechanism. If it fails to retransmit for 10 consecutive times, send an error type code to the host computer to inform the personnel to check whether the circuit connection and system board are working properly.

[0117] Step 2: Encode the antennas that need to be connected (0 to N) and set the dwell time parameters.

[0118] Step 3. Start the antenna control trigger signal.

[0119] Step 4. Check whether the antenna i of the current group is connected successfully. If it is not connected, resend the trigger signal. If the antenna cannot be connected after 5 consecutive resends of the trigger signal, send an error flag code to the upper computer to inform the operator to check whether the high and low terminals of this group of antennas are damaged.

[0120] Step 5. If the antenna i of this group is successfully connected, delay de-jittering is performed to ensure that the antenna is in a normal connection state and the collected data has no jitter.

[0121] Step 6. After delay and debouncing, collect a section of data according to the host computer instructions.

[0122] Step 7. Sort the collected data and remove abnormal data values. This will greatly improve the data accuracy and achieve the purpose of anti-interference of Starlink terminal signals.

[0123] Step 8. Determine whether all antenna groups are connected according to the dwell time interval. If not, start the next antenna connection process. If all antenna groups are connected, end the polling and wait for the next acquisition command.

[0124] Step 9. Complete the above processing and give the optimized Starlink terminal signal data.

[0125] Figure 4 This is the workflow of the Starlink terminal signal classification and recognition algorithm. The computer processes the signal data received by the receiver. The specific steps of Starlink terminal signal classification and recognition are as follows:

[0126] Step 1. For the selected Starlink terminal signal data, multi-dimensional feature extraction is required first, including high-order cumulants and OFMD signal parameter features.

[0127] High-order cumulant feature extraction

[0128] The original signal is collected as an IQ data set, so the received data can be expressed as:

[0129] S(i)=S I (i)+S Q (i) (1)

[0130] Where S(i) represents the i-th signal sample; S I (i) and S Q (i) represent the in-phase component and quadrature component of the ith signal, respectively. Since the statistics of Gaussian noise above the second order are zero, high-order cumulants are robust to Gaussian noise and are therefore often used in signal recognition.

[0131] Through the cumulant-moment formula (CM formula) and the moment-cumulant formula (MC formula), the P-order moment and q-order cumulant of the zero-mean complex random signal S[i] can be expressed as:

[0132]

[0133] In the formula, * represents complex conjugation; p is the order (p>q); q is the number of sequences to take the complex conjugation; E represents the mean, the same below. After derivation, the commonly used high-order cumulant expression of S(i) can be obtained as:

[0134] C 21 =M 21 , (4)

[0135]

[0136] C 41 =M 41 -3*M 20 *M 21 , (6)

[0137]

[0138] OFMD signal parameter extraction

[0139] It is known that OFDM signals are composed of multiple independent and orthogonal subcarriers. These subcarriers are not correlated with each other, but the same subcarrier has the characteristic of autocorrelation. Therefore, the premise of this method also needs to prove the periodic characteristics of the same subcarrier of the OFDM signal, as shown below.

[0140] Let the OFDM signal in baseband be:

[0141]

[0142] In the formula, m a,b It represents the ath OFDM signal on the bth subcarrier, N represents the number of subcarriers, P is the signal power factor, h(t) is the pulse function, T s Represents the symbol period length, which is defined as T s =Tu+Tg, where Tu represents the useful symbol length, Tg is the cyclic prefix, and Δf represents the interval between subcarriers, which is defined as Δf=1 / Tu.

[0143] After the signal autocorrelation function is transformed by FFT, the signal autocorrelation function is obtained. The amplitude spectrum of the autocorrelation will show a very obvious symmetrical peak, and the position of the peak is the number of valid symbol points Nu. The Tu value is calculated using the formula Tu = Fs / Nu, where Fs is the signal sampling rate.

[0144] After the signal is delayed by Tu, the autocorrelation is performed. The amplitude spectrum of the autocorrelation will show obvious equally spaced peaks, and the equally spaced difference is the number of symbol points Ns. s =Fs / Ns formula to calculate T s Value, where Fs is the signal sampling rate.

[0145] Using the formula Tg = T s -Tu calculates the cyclic prefix length.

[0146] The interval between subcarriers is calculated using the formula Δf=1 / Tu.

[0147] The number of subcarriers is calculated using the formula SubCarryNum=Fs / Δf.

[0148] Thus, the OFDM signal symbol period length T is extracted s , cyclic prefix Tg, interval Δf between subcarriers, and number of subcarriers SubCarryNum.

[0149] The 12-dimensional feature matrix V = [C 21 ,C 40 ,C 41 ,C 42 ,C 60 ,C 61 ,C 63 ,C 80 ,T s ,Tg,Δf,SubCarryNum].

[0150] Step 2. Send the V feature matrix to the convolutional neural network (CNN). CNN uses the convolution kernel to fuse and perceive the original data, and uses local connections and weight sharing to extract the deep features of the input waveform data. The pooling layer performs pooling operations on the convolved data to reduce the number of parameters and model complexity. Generally, the CNN layer and the pooling layer appear in pairs, and the convolution pooling layer is passed through twice. Then, regularization discards some parameters to prevent overfitting (Dropout). The advantage of CNN in the field of signal processing is that it can extract the hidden features of the data and obtain high-level abstract features through step-by-step fusion. However, it lacks memory function and does not extract enough time series correlation features for the entire time domain segment.

[0151] Step 3. Since there is a strong temporal correlation between the symbols of the Starlink terminal signal, the data after Dropout is sent to the Long Short-Term Memory Network (LSTM). Recursive Neural Network (RNN) is one of the early machine learning models used to calculate time series. LSTM is an upgraded version of RNN. It solves the gradient vanishing problem of RNN through the forget gate structure. This special gate structure is also the reason why it can effectively preserve the characteristics of time series data.

[0152] The main component of the LSTM recurrent unit structure is the forget gate f t , input gate i t And the output gate o t The calculation process is:

[0153] (1) Forget Gate

[0154] f t =σ(W f ·[h t-1 ,x t ]+b f ) (13)

[0155] where f t is the output of the forget gate, W f is the linear coefficient, b i is the bias coefficient, σ represents the sigmoid function (between 0 and 1), h t-1 is the output at the previous moment, and the input signal x t .

[0156] (2) Input Gate

[0157] i t =σ(W i ·[h t-1 ,x t ]+b i ) (14)

[0158]

[0159] Among them, i t and are the outputs of the Sigmoid activation function and the tanh activation function respectively; W i , W C is the linear coefficient; b i 、b C is the bias coefficient. At this time, the LSTM unit state can be expressed as:

[0160]

[0161] Among them, Ct is the current LSTM unit state, ⊙ represents the Hadamard product.

[0162] (3) Output gate

[0163] o t =σ(W o ·[h t-1 ,x t ]+b o ) (17)

[0164] h t =o t ⊙tanh(C t ) (18)

[0165] Among them, t is the output gate output, h t is the hidden layer output, Wo is the weight, and bo is the initial bias coefficient.

[0166] The special gating structure of LSTM can selectively control the retention and forgetting of features in time series data. For Starlink terminal data containing pilot sequence, start flag and end flag fields, similar time series structures will interfere with feature screening. The output of the CNN structure can be used as the input of the LSTM layer to improve the performance of the model.

[0167] Step 4. Pass the two layers of Dense and input the obtained feature vector into the fully connected layer, and use the Softmax function for classification.

[0168] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall not deviate from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. The Starlink terminal signal recognition method based on multi-dimensional feature fusion and CNN-LSTM is characterized by: The steps include: Step 1: extract multi-dimensional features from the filtered Starlink terminal signal data, where the multi-dimensional features include high-order cumulants and OFMD signal parameter features, and obtain a multi-dimensional feature matrix based on the high-order cumulants and OFMD signal parameter features; Step 2: The multi-dimensional feature matrix is ​​fed into the convolutional neural network. The convolutional neural network uses the convolution kernel to fuse and perceive the original data, and uses local connections and weight sharing to extract the deep features of the input waveform data. The pooling layer performs a pooling operation on the convolved data, and regularizes it after two convolutional pooling layers to obtain high-level abstract features. Step 3: Send the regularized data into the long short-term memory network to obtain the time series data features; Step 4: Input the obtained time series data features into the fully connected layer through two layers of Dense, and use the Softmax function to classify the output of the fully connected layer.

2. The Starlink terminal signal recognition method based on multi-dimensional feature fusion and CNN-LSTM according to claim 1 is characterized in that: The high-order cumulant extraction includes: By collecting the original signal as an IQ data set, the received data is: S(i)=S I (i)+S Q (i) Where S(i) represents the i-th signal sample; S I (i) and S Q (i) respectively represent the in-phase component and the orthogonal component of the i-th signal; through the cumulant-moment formula and the moment-cumulant formula, the P-order moment and q-order cumulant of the zero-mean complex random signal S[i] are expressed as: In the formula, * represents complex conjugation; p is the order (p>q); q is the number of sequences to take complex conjugation; E represents the mean; The high-order cumulant expression of S(i) is obtained as: C 21 =M 21 C 41 =M 41 -3*M 20 *M 21 3. The Starlink terminal signal recognition method based on multi-dimensional feature fusion and CNN-LSTM according to claim 2 is characterized in that: The OFMD signal parameter feature extraction includes: Let the OFDM signal in baseband be: In the formula, m a,b represents the ath OFDM signal on the bth subcarrier, N represents the number of subcarriers, P is the signal power factor, h(t) is the pulse function, T s Indicates the symbol period length, T s =Tu+Tg, where Tu represents the useful symbol length, Tg is the cyclic prefix, Δf represents the interval between subcarriers, and Δf=1 / Tu; The Tu value is obtained according to Tu=Fs / Nu, where Fs is the signal sampling rate; After the signal is delayed by Tu, it is autocorrelated. The amplitude spectrum of the autocorrelation will have equally spaced peaks, and the equally spaced difference is the number of symbol points Ns. According to T s =Fs / Ns to get T s Value, where Fs is the signal sampling rate; According to Tg = T s -Tu obtains the cyclic prefix length; The interval between subcarriers is obtained according to Δf=1 / Tu; The number of subcarriers is obtained according to the formula SubCarryNum = Fs / Δf; Extract the OFDM signal symbol period length T s , cyclic prefix Tg, interval Δf between subcarriers, and number of subcarriers SubCarryNum.

4. The Starlink terminal signal recognition method based on multi-dimensional feature fusion and CNN-LSTM according to claim 3 is characterized in that: The multi-dimensional feature matrix is ​​obtained according to the high-order cumulants and OFDM signal parameter features, including: fusing the 8 parameter features of the high-order cumulants and the OFDM signal parameter features to construct a 12-dimensional feature matrix V = [C 21 ,C 40 ,C 41 ,C 42 ,C 60 ,C 61 ,C 63 ,C 80 ,T s , Tg, Δf, SubCarryNum].

5. The Starlink terminal signal recognition method based on multi-dimensional feature fusion and CNN-LSTM according to claim 4 is characterized in that: The regularized data is fed into the long short-term memory network to obtain time series data features, including: The main component of the long short-term memory network recurrent unit structure is the forget gate f t , input gate i t And the output gate o t , where the forget gate is: f t =σ(W f ·[h t-1 ,x t ]+b f ) where f t is the output of the forget gate, W f is the linear coefficient, b i is the bias coefficient, σ represents the sigmoid function, h t-1 is the output at the previous moment, and the input signal x t ; The input gate is: i t =σ(W i ·[h t-1 ,x t ]+b i ) Among them, i t and are the outputs of the Sigmoid activation function and the tanh activation function respectively; W i , W C is the linear coefficient; b i 、b C is the bias coefficient, and the LSTM unit state is: Among them, C t is the current LSTM unit state, ⊙ represents the Hadamard product; The output gate is: the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t ⊙tanh(C t ) Among them, t is the output gate output, h t is the hidden layer output, Wo is the weight, and bo is the initial bias coefficient.

6. The Starlink terminal signal recognition system based on multi-dimensional feature fusion and CNN-LSTM is characterized by: The Starlink terminal signal recognition method based on multi-dimensional feature fusion and CNN-LSTM described in any one of claims 1 to 5 is applied, comprising: a nine-element uniform phased circular array, a switch matrix, a wide-band and large-bandwidth receiver, a computer, and a radio frequency switch driver; The nine-element uniform phased circular array uses a nine-element phased array antenna, and the directional pattern of the circular array is 360° omnidirectional without array ambiguity; The switch matrix connects different antenna combinations through program control to achieve automatic synthesis of antenna patterns; The radio frequency switch driver is used to control the working state of the matrix switch through the radio frequency switch driver according to the instructions given by the computer; The wide-band and wide-bandwidth receiver provides the measured signal data received by the two antenna array elements to the computer for processing; The computer: The computer realizes human-computer dialogue, controls the start and end of data collection, processes the collected data through software, and displays the results to the user through the software interface after the collected signals.

Images