A method for generating multimodal datasets of satellite search and rescue jamming signals
By generating a multimodal dataset of satellite search and rescue interference signals through three-dimensional hierarchical segmentation, the problem of lack of high fidelity and standardization in existing datasets is solved. This enables effective training of multimodal deep learning algorithms and improves their anti-interference performance, making it suitable for interference identification and electromagnetic compatibility testing in satellite search and rescue systems.
Patent Information
- Application Number
- CN202611110372.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-24
- Publication Date
- 2026-11-10
AI Technical Summary
The existing interference signal datasets of satellite search and rescue systems lack high-fidelity, standardized, multimodal, and reproducible datasets, which cannot effectively support interference identification and anti-interference algorithm verification. Furthermore, the existing datasets do not strictly follow the COSPAS-SARSAT international standard, making it difficult to simulate the superposition of multiple types of interference in actual electromagnetic environments.
A three-dimensional hierarchical segmentation method was used to generate a multi-modal dataset of satellite search and rescue interference signals. The dataset strictly followed the C/S T.018, ITU-R M.1478 and ITU-R SM.2258 standards, generating 11 typical COSPAS-SARSAT interference signals. Through three-dimensional hierarchical segmentation (class, SNR, JSR), the dataset output three modes: time domain, frequency domain single-sided spectrum and time-frequency diagram, and stored in HDF5 format.
It achieves high fidelity and standardization of the dataset, supports training of multimodal deep learning algorithms, improves the accuracy and anti-interference performance of signal-to-noise ratio estimation, ensures the reproducibility and flexibility of the dataset, and is suitable for a variety of downstream tasks.
Smart Images

Figure CN123074558A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of satellite communication signal simulation, artificial intelligence dataset construction, and electromagnetic compatibility testing technology. In particular, it relates to a method for generating a multimodal dataset of satellite search and rescue interference signals, which is a method for generating a multimodal dataset of satellite search and rescue interference signals based on three-dimensional hierarchical segmentation, conforming to the COSPAS-SARSAT international standard C / S T.018 and ITU-R M.1478 and ITU-R SM.2258. Background Technology
[0002] COSPAS-SARSAT is an international satellite search and rescue system operating in the 406 MHz band. Since its official operation in 1985, it has assisted in the rescue of more than 50,000 people in distress and is one of the world's most important emergency communication infrastructures. The system's second-generation beacon, SGB, operates in the 406.0 MHz to 406.1 MHz band, using OQPSK-DSSS modulation, conforming to the international standard C / S T.018, with a spreading chip rate of 38.4 kchips / s, a data rate of 300 bps, and a burst duration of 1000 ms. With the increasingly complex electromagnetic environment in this band, SGB signals face various interferences such as radar pulses, frequency-hopping communication, single-tone / multi-tone interference, broadband noise, and phase modulation. Therefore, downstream tasks such as SARSAT interference identification, anti-interference algorithm verification, and satellite payload electromagnetic compatibility testing urgently require high-fidelity, standardized, multimodal, and reproducible satellite interference signal datasets. However, due to three factors—the privacy and military sensitivity of real signals in search and rescue, the limited types of interference actually recorded, and the high cost of manual annotation—this field still lacks publicly available high-quality interference signal datasets.
[0003] Among existing publicly available datasets, the RadioML series (RadioML 2016.10A and RadioML 2018.01A), released by DeepSig, primarily targets general modulation recognition tasks (such as QPSK and 16QAM), and does not model the OQPSK-DSSS signal structure specific to SARSAT. Its signal generation uses a simplified channel model and does not adhere to key parameters of the C / S T.018 specification; for example, it does not implement a 1000 ms long burst, and does not correctly configure the 24 dB spreading gain and corresponding spreading code sequence. Furthermore, this dataset only provides IQ complex time-domain representations and lacks multimodal data such as frequency-domain single-sided spectra and time-frequency plots, making it difficult to support the training and evaluation of multimodal deep learning algorithms. The HisarMod 2019.1 dataset contains 26 types of modulated signals, but its design goal is modulation identification, completely ignoring interference signal scenarios and failing to introduce the crucial signal-to-interference ratio (JSR). Therefore, it cannot be used to evaluate the receiver's anti-interference performance under interference conditions. Furthermore, simplified simulation datasets built by some research institutions generally suffer from the following problems: incomplete physical parameterization, failing to strictly adhere to relevant ITU-R recommendations; random data partitioning, leading to statistical bias in the signal-to-noise ratio (SNR) and JSR dimensions, resulting in unreliable performance evaluation results; and most datasets only model a single type of interference, failing to consider the combination of multiple types of interference commonly encountered in actual search and rescue environments (such as frequency sweeping interference, partial-band noise interference, and multipath interference combinations), resulting in a significant discrepancy with real-world application scenarios. Summary of the Invention
[0004] To address the aforementioned problems, the present invention aims to provide a method for generating a multimodal dataset of satellite search and rescue jamming signals, which is based on three-dimensional hierarchical segmentation. This invention strictly adheres to the international standards C / S T.018, ITU-R M.1478, and ITU-R SM.2258, systematically generating 11 typical COSPAS-SARSAT jamming signals. Three-dimensional hierarchical segmentation (class, SNR, JSR) ensures a balanced distribution of the test set, and simultaneously outputs time-domain, frequency-domain single-sided spectra, and time-frequency... Figure 3 Types of modes.
[0005] To achieve the above objectives, the solution of the present invention is as follows:
[0006] A method for generating a multimodal dataset of satellite search and rescue interference signals is provided, which is based on three-dimensional hierarchical segmentation. The method includes the following steps:
[0007] S1. Generate an SGB clean signal according to the COSPAS-SARSAT international standard C / S T.018. The SGB clean signal has a center frequency of 406.05 MHz, a modulation method of OQPSK-DSSS, a burst length of 1000 ms, a data rate of 300bps, a spreading chip rate of 38.4 kchips / s, a spreading gain of 24 dB, and an effective isotropic radiated power of 5W.
[0008] S2. Construct various types of interference signals;
[0009] S3. Combine the SGB clean signal and various types of interference signals, and add additive white Gaussian noise to form a triplet time-domain sample.
[0010] The triple is class, SNR, JSR:
[0011] class: Multiple categories of interference;
[0012] SNR: Signal-to-noise power ratio (dB);
[0013] JSR: Interference / Signal Power Ratio (dB)
[0014] The values of both SNR and JSR are discrete, equally spaced sets:
[0015]
[0016] ;
[0017] S4. Perform a three-dimensional hierarchical segmentation of the triplet time-domain samples into training, validation, and test sets, with segmentation ratios of 70%, 20%, and 10%, respectively.
[0018] S5. Simultaneously generate time-domain and frequency-domain one-sided spectra and time-frequency spectra for each triplet time-domain sample. Figure 3 The data are processed in multiple modes and stored as HDF5 format files to form a multimodal dataset of satellite search and rescue interference signals.
[0019] The solution further includes: the mathematical model for the SGB clean signal is:
[0020] ;
[0021] in: The data symbols are OQPSK modulated. A spreading sequence of 128 chips per symbol; Roll-off factor Raised cosine pulse shaping filter; The chip cycle; t is the average power of the transmitted signal (dB); n is the continuous time variable; N is the mathematical symbol sequence; N is the total number of data symbols in a frame; fc is the carrier center frequency.
[0022] The solution further includes: In step S3, the synthesis formula for the time-domain samples is:
[0023] ;
[0024] in: For time-domain samples; For SGB clean signal; It is additive white Gaussian noise; Different types of interference signals; among them, the signal-to-noise ratio (SNR) of additive white Gaussian noise is defined as:
[0025] .
[0026] The solution further specifies that the various types of interference signals fall into 11 categories, namely: clean, single_tone, multi_tone, chirp, pulse, wideband_noise, narrowband_noise, bpsk_modulated, comb_spectrum, frequency_hopping, and composite.
[0027] The "clean" class refers to signals without interference, containing only additive white Gaussian noise.
[0028] The single_tone class expression is: ;
[0029] The multi_tone class expression is: ;
[0030] The chirp class expression is: ;
[0031] The pulse class expression is: ;
[0032] The wideband_noise class expression is: ;
[0033] The narrowband_noise class expression is: ;
[0034] The bpsk_modulated class expression is: ;
[0035] The comb_spectrum class expression is: ;
[0036] The expression for the frequency_hopping class is: ;
[0037] The expression for the composite class is: ;
[0038] in:
[0039] To interfere with the prototype signal; , , All are amplitudes, among which , They are respectively polyphonic. Each component, comb spectrum The amplitude of the spectral lines; For simulating intermediate frequency; , This is the frequency offset relative to the intermediate frequency; , , , As the initial phase, in Uniformly distributed values within the range; The number of notes in a polyphonic system; This is the start frequency for the frequency sweep. This is the frequency modulation slope, which is the change in instantaneous frequency per unit time. It is a periodic rectangular envelope used to characterize the on / off state of pulse interference; It is white Gaussian noise; , Bandwidth broadband, center frequency bandwidth Narrowband bandpass filtering; For taking values The binary non-return-to-zero baseband waveform; This represents the number of spectral lines in the comb pattern. For the first Spectral line frequencies (interval between adjacent lines) ); For frequency hopping instantaneous carrier frequencies, there are period hopping intervals between several discrete carrier frequencies. Switch; The number of components of the composite interference (take 2 to 4). For the first One basic interference component Use its power weighting coefficient.
[0040] The scheme further includes: generating interference for each of the 11 types of interference signals. K samples are evenly distributed across 13×13=169 (SNR, JSR) case grids, such that each triplet contains K samples. There are 10 independent samples, and the total number of samples in the entire dataset is:
[0041] .
[0042] The solution further includes the following: For the 11 types of interference signals, the three-dimensional hierarchical segmentation method in step S4 specifically satisfies the following: For each of the 11×13×13=1859 triples, it is independently segmented into a training set, a validation set, and a test set at ratios of 70%, 20%, and 10%, respectively, so that the training set, validation set, and test set contain 38,500, 11,000, and 5,500 samples, respectively, and the training set, validation set, and test set completely cover all 1859 triples.
[0043] A further step in the scheme is: in step S5, the frequency domain single-sided spectrum is obtained by taking the modulus through Fast Fourier Transform (FFT).
[0044] Length is 8192, frequency resolution:
[0045] .
[0046] The scheme further includes: In step S5, the time-frequency graph is obtained by calculating using a short-time Fourier transform (STFT), with parameters of window length nperseg = 512, overlap noverlap = 256, and an output tensor size of 224×224. The time resolution is:
[0047] ;
[0048] Frequency resolution:
[0049] .
[0050] The solution further includes: In step S5, the HDF5 format file contains three data groups: training set (train), validation set (val), and test set (test). Each data group stores three modal data types: time_domain, freq_domain, and tf_image, as well as three types of label data: class_labels, snr_labels, and jsr_labels.
[0051] The solution further includes: In step S5, the data representation of the three modalities specifically includes:
[0052] I. Time Domain Representation: Composed of the real and imaginary parts of the directly sampled time domain samples, with a data shape of (2, 16384);
[0053] II. Frequency Domain Single-Sided Spectrum Representation: The data is obtained by taking the first 8192 points after taking the modulus of the directly sampled time-domain sample using an FFT of length 16384 and normalizing it. The data shape is (8192,).
[0054] III. The time-frequency diagram is obtained by directly sampling the time-domain samples, interpolating to 224×224 after STFT mode taking with nperseg=512 and noverlap=256, and normalizing. The data shape is (224, 224).
[0055] The beneficial effects of this invention are:
[0056] 1. Compliance of the dataset: This invention strictly follows the C / S T.018 specification to model SARSAT signals, covers 11 typical interference types, employs three-dimensional hierarchical segmentation, and simultaneously provides time domain / frequency domain / time-frequency data. Figure 3 It features modalities and includes composite interference, providing higher physical fidelity in SARSAT scenarios compared to RadioML 2018 and HisarMod 2019.1.
[0057] Second, signal-to-noise ratio monotonicity: The SNR-accuracy curve of the deep learning model trained using the dataset of this invention shows an overall monotonically increasing trend on the test set. In the low signal-to-noise ratio range (−10 dB ~ 0 dB), it strictly monotonically increases by about +24.8 pp (from 64.66% to 89.46%), and in the high signal-to-noise ratio range (≥0 dB), the jitter amplitude does not exceed 1.7 pp, which is much smaller than the abnormal jitter (amplitude greater than 3 pp) of "−4 dB accuracy > +2 dB" that appeared in the era of traditional random segmentation.
[0058] Third, downstream task support: The dataset of this invention can support downstream tasks such as 11 types of interference identification, signal-to-noise ratio estimation, anti-interference algorithm verification, satellite payload electromagnetic compatibility testing, and multi-modal fusion method research.
[0059] Fourth, reproducibility: This invention makes the dataset generation process completely reproducible by explicitly recording the (class, SNR, JSR) triplet label and random seed for each sample, and allows for flexible adjustment of the number of samples K in each class, the set of discrete SNR / JSR points, or the segmentation ratio in different experimental scenarios as needed.
[0060] Fifth, compliance with standards: The SGB signal modeling described in this invention strictly follows the OQPSK-DSSS modulation specifications, burst structure, and spread spectrum parameters of the C / S T.018 standard; the noise and interference modeling follows the relevant recommendations of ITU-R M.1478 (on the electromagnetic environment of the 406MHz band) and ITU-R SM.2258 (on radio monitoring and spectrum management), and is a satellite search and rescue interference signal simulation method that simultaneously complies with the above three international standards.
[0061] The invention will be further explained in detail below with reference to the accompanying drawings and specific embodiments. Attached Figure Description
[0062] Figure 1 A five-stage overall flowchart generated for the dataset of this invention;
[0063] Figure 2 This is a summary image of representative samples of 11 types of signals in the dataset of this invention. It is a summary of the time domain, frequency domain and video images of 11 types of COSPAS-STRSAT interference signals (SNR=15dB, JSR=5dB, duration=20ms, frequency: single-sided spectrum, IF50kHz=RF406.05MHz).
[0064] Figure 3 This is a schematic diagram of the three-dimensional hierarchical segmentation strategy (class, SNR, JSR) of the present invention; Figure 4 This is an SNR-accuracy curve of the dataset in this invention under three-dimensional hierarchical segmentation. Detailed Implementation
[0065] This embodiment aims to solve the following technical problems: (1) How to strictly simulate SGB (second-generation beacon) signals according to the C / S T.018 specification; (2) How to systematically model multiple typical interferences to cover the interference modes that may be encountered in actual search and rescue scenarios; (3) How to scientifically segment the dataset to avoid statistical bias in the signal-to-noise ratio and interference-to-signal ratio dimensions; (4) How to simultaneously generate time-domain, frequency-domain single-sided spectra and time-frequency... Figure 3 Various modal data are used to support multimodal deep learning research; to this end:
[0066] A method for generating a multimodal dataset of satellite search and rescue interference signals is provided, which is based on three-dimensional hierarchical segmentation. The dataset is used for training, verification, and testing of neural networks to discriminate satellite search and rescue interference signals. The method includes the following steps:
[0067] S1. Generate a clean SGB (second-generation beacon) signal according to the COSPAS-SARSAT international standard C / S T.018. The SGB clean signal has a center frequency of 406.05 MHz, a modulation method of OQPSK-DSSS, a burst length of 1000 ms, a data rate of 300 bps, a spreading chip rate of 38.4 kchips / s, a spreading gain of 24 dB, and an effective omnidirectional radiated power of 5W.
[0068] S2. Construct various types of interference signals;
[0069] S3. Receive the transmitted SGB clean signal, synthesize the received SGB clean signal and various constructed interference signals of different types, and add additive white Gaussian noise to form a triplet time-domain sample.
[0070] The triple is class, SNR, JSR:
[0071] class: Multiple categories of interference;
[0072] SNR: Signal-to-noise power ratio (dB);
[0073] JSR: Interference / Signal Power Ratio (dB)
[0074] The values of both SNR and JSR are discrete, equally spaced sets:
[0075]
[0076] ;
[0077] S4. Perform a three-dimensional hierarchical segmentation of the triplet time-domain samples into training, validation, and test sets, with segmentation ratios of 70%, 20%, and 10%, respectively.
[0078] S5. Simultaneously generate time-domain and frequency-domain one-sided spectra and time-frequency spectra for each triplet time-domain sample. Figure 3 The data are processed in multiple modes and stored as HDF5 format files to form a multimodal dataset of satellite search and rescue interference signals.
[0079] The calculation methods and data shapes for the three modes are shown in Table 1.
[0080] Table 1 Multimodal Data Specifications
[0081]
[0082] The HDF5 dataset contains three data groups: train, val, and test. Each data group stores three modalities of data: time_domain, freq_domain, and tf_image, as well as three types of label data: class_labels, snr_labels, and jsr_labels.
[0083] Wherein: the mathematical model of the SGB clean signal is:
[0084] ;
[0085] Among them: Among them: The data symbol is modulated using OQPSK (Offset Quad Phase Shift Keying). A spreading sequence of 128 chips per symbol; Roll-off factor Raised cosine pulse shaping filter; The chip cycle; t is the average power of the transmitted signal (dB); n is the continuous time variable; N is the mathematical symbol sequence; N is the total number of data symbols in a frame; fc is the carrier center frequency.
[0086] In step S3, the synthesis formula for the time-domain samples is:
[0087] ;
[0088] in: For time-domain samples; For SGB clean signal; It is additive white Gaussian noise; Different types of interference signals; among them, the signal-to-noise ratio (SNR) of additive white Gaussian noise is defined as:
[0089] .
[0090] In this embodiment: the various types of interference signals are 11 categories, namely: clean, single tone, multi tone, chirp, pulse, wideband noise, narrowband noise, bpsk modulated, comb spectrum, frequency hopping, and composite.
[0091] The "clean" class refers to signals without interference, containing only additive white Gaussian noise.
[0092] The single_tone class expression is: ;
[0093] The multi_tone class expression is: ;
[0094] The chirp class expression is: ;
[0095] The pulse class expression is: ;
[0096] The wideband_noise class expression is: ;
[0097] The narrowband_noise class expression is: ;
[0098] The bpsk_modulated class expression is: ;
[0099] The comb_spectrum class expression is: ;
[0100] The expression for the frequency_hopping class is: ;
[0101] The expression for the composite class is: ;
[0102] in:
[0103] To interfere with the prototype signal; , , All are amplitudes, among which , They are respectively polyphonic. Each component, comb spectrum The amplitude of the spectral lines; For simulating intermediate frequency; , This is the frequency offset relative to the intermediate frequency; , , , As the initial phase, in Uniformly distributed values within the range; The number of notes in a polyphonic system; This is the start frequency for the frequency sweep. The frequency modulation slope is the change in instantaneous frequency per unit time. It is a periodic rectangular envelope used to characterize the on / off state of pulse interference; It is white Gaussian noise; , Bandwidth broadband, center frequency bandwidth Narrowband bandpass filtering; For taking values The binary non-return-to-zero baseband waveform; This represents the number of spectral lines in the comb pattern. For the first Spectral line frequencies (interval between adjacent lines) ); For frequency hopping instantaneous carrier frequencies, there are period hopping intervals between several discrete carrier frequencies. Switch; The number of components of the composite interference (take 2 to 4). For the first One basic interference component Use its power weighting coefficient.
[0104] The power of each type of interference is controlled by the interference / signal power ratio (JSR), defined as:
[0105] ;
[0106] Its value range is {−10, −8, −6, −4, −2, 0, 2, 4, 6, 8, 10, 12, 14} dB, with a total of 13 equally spaced points.
[0107] Wherein: each of the 11 types of interference signals generates interference. K samples are evenly distributed across 13×13=169 (SNR, JSR) case grids, such that each triplet contains K samples. There are 10 independent samples, and the total number of samples in the entire dataset is:
[0108] .
[0109] For the 11 types of interference signals, the three-dimensional hierarchical segmentation method in step S4 specifically satisfies the following: for each of the 11×13×13=1859 triples, it is independently segmented into a training set, a validation set, and a test set at ratios of 70%, 20%, and 10%, respectively, so that the training set, validation set, and test set have 38,500, 11,000, and 5,500 samples, respectively, and the training set, validation set, and test set completely cover all 1859 triples.
[0110] In step S5, the frequency domain single-sided spectrum is obtained by taking the modulus of the Fast Fourier Transform (FFT), with a length of 8192 and a frequency resolution of:
[0111] .
[0112] In step S5, the time-frequency graph is calculated using a short-time Fourier transform (STFT) with parameters of window length nperseg = 512, overlap noverlap = 256, and an output tensor size of 224×224. The time resolution is:
[0113] ;
[0114] Frequency resolution:
[0115] .
[0116] In step S5, the HDF5 format file contains three data groups: training set (train), validation set (val), and test set (test). Each data group stores three modal data types: time_domain, freq_domain, and tf_image, as well as three types of label data: class_labels, snr_labels, and jsr_labels.
[0117] In step S5, the data representation of the three modalities specifically includes:
[0118] I. Time Domain Representation: Composed of the real and imaginary parts of the directly sampled time domain samples, with a data shape of (2, 16384);
[0119] II. Frequency Domain Single-Sided Spectrum Representation: The data is obtained by taking the first 8192 points after taking the modulus of the directly sampled time-domain sample using an FFT of length 16384 and normalizing it. The data shape is (8192,).
[0120] III. The time-frequency diagram is obtained by directly sampling the time-domain samples, interpolating to 224×224 after STFT mode taking with nperseg=512 and noverlap=256, and normalizing. The data shape is (224, 224).
[0121] Below are specific examples:
[0122] Example 1: In this embodiment, the SGB clean signal is generated according to the C / S T.018 specification. The core pseudocode for generating the SGB signal is as follows:
[0123] def generate_sgb_signal(fs=1e6, duration=16.384e-3,
[0124] bit_rate=300, chip_rate=38400, sf=128):
[0125] "Generate a clean SGB signal according to C / S T.018 specification."
[0126] n_samples = int(fs * duration)
[0127] n_bits = int(bit_rate * duration)
[0128] # Step 1: Generate random bits and split the I / Q channels
[0129] bits = np.random.randint(0, 2, n_bits)
[0130] i_bits, q_bits = bits[::2], bits[1::2]
[0131] i_symbols, q_symbols = 2 * i_bits - 1, 2 * q_bits - 1
[0132] # Step 2: DSSS spreading, spreading factor sf=128
[0133] spread_code = np.random.choice([-1, 1], sf)
[0134] i_spread = np.repeat(i_symbols, sf) * np.tile(spread_code, len(i_symbols))
[0135] q_spread = np.repeat(q_symbols, sf) * np.tile(spread_code, len(q_symbols))
[0136] # Step 3: RRC shaping filter with β=0.5 (design_rrc_filter is a known filter design function)
[0137] sps = int(fs / chip_rate)
[0138] rrc = design_rrc_filter(beta=0.5, span=8, sps=sps)
[0139] i_signal = upfirdn(rrc, i_spread, sps)
[0140] q_signal = upfirdn(rrc, q_spread, sps)
[0141] # Step 4: Delay the Q channel by half a chip to form OQPSK modulation.
[0142] q_signal = np.roll(q_signal, sps / / 2)
[0143] s = (i_signal + 1j * q_signal)[:n_samples]
[0144] return s.
[0145] Example 2: Generation of 11 types of interference signals. Taking chirp, comb_spectrum, and frequency_hopping as examples, the core pseudocode is as follows:
[0146] def generate_chirp(t, fc=406.05e6, bw=30e3, duration=1e-3):
[0147] "Linear frequency modulation (chirp) interference: sweep bandwidth bw, duration."
[0148] mu = bw / duration # Frequency modulation slope
[0149] return np.exp(1j * 2 * np.pi * (fc * t + 0.5 * mu * t**2))
[0150] def generate_comb_spectrum(t, n_combs=8, f_spacing=4e3):
[0151] "Comb-like interference: n_combs of equally spaced carrier frequencies superimposed, with a spacing of f_spacing."
[0152] s = np.zeros(len(t), dtype=complex)
[0153] for k in range(n_combs):
[0154] f_k = (k - n_combs / / 2) * f_spacing
[0155] s += np.exp(1j * 2 * np.pi * f_k * t) / np.sqrt(n_combs)
[0156] return s
[0157] def generate_freq_hopping(t, hop_rate=500, n_hops=8):
[0158] Frequency hopping interference: Switching between n_hops discrete carrier frequencies every 1 / hop_rate second.
[0159] hop_duration = 1.0 / hop_rate
[0160] f_set = np.linspace(-30e3, 30e3, n_hops)
[0161] s = np.zeros(len(t), dtype=complex)
[0162] for i, ti in enumerate(t):
[0163] idx = int(ti / hop_duration) % n_hops
[0164] s[i] = np.exp(1j * 2 * np.pi * f_set[idx] * ti)
[0165] return s.
[0166] Example 3: Three-dimensional layered segmentation. Its core pseudocode is as follows:
[0167] def stratified_3d_split(classes, snrs, jsrs,
[0168] train_ratio=0.7, val_ratio=0.2):
[0169] "For each (class, snr, jsr) triple, it is independently split into layers in a 7:2:1 ratio."
[0170] train_idx, val_idx, test_idx = [], [], []
[0171] # Step 1: Enumerate all 1859 (class, snr, jsr) triples
[0172] unique_groups = set(zip(classes, snrs, jsrs))
[0173] for cls, snr, jsr in unique_groups:
[0174] # Step 2: Locate all samples corresponding to the current triplet
[0175] mask = (classes == cls) & (snrs == snr) & (jsrs == jsr)
[0176] idx = np.where(mask)[0]
[0177] np.random.shuffle(idx)
[0178] # Step 3: Independently split the triplet according to 70% / 20% / 10%.
[0179] n = len(idx)
[0180] n_train = int(n * train_ratio)
[0181] n_val = int(n * val_ratio)
[0182] train_idx.extend(idx[:n_train])
[0183] val_idx.extend(idx[n_train:n_train + n_val])
[0184] test_idx.extend(idx[n_train + n_val:])
[0185] return (np.array(train_idx),
[0186] np.array(val_idx),
[0187] np.array(test_idx)).
[0188] Example 4: Multimodal Data Generation and HDF5 Storage. From a single time-domain complex signal... The core pseudocode for generating three modalities and writing them to an HDF5 file is as follows:
[0189] def generate_multimodal_data(x_t, fs=1e6):
[0190] """Time-domain / frequency-domain / time-frequency generation is achieved synchronously from a single complex time-domain signal." Figure 3 Modality.
[0191] # Mode 1: Time-domain complex signal split into real / imaginary dual channels
[0192] time_domain = np.stack([x_t.real, x_t.imag], axis=0)
[0193] # Modality 2: FFT modulus yields 8192-point one-sided spectrum, which is then normalized.
[0194] freq_domain = np.abs(np.fft.rfft(x_t, n=16384))[:8192]
[0195] freq_domain = freq_domain / np.max(freq_domain)
[0196] # Modality 3: STFT post-interpolation of 224×224 time-frequency plot and normalization
[0197] f, t, Zxx = stft(x_t, fs=fs, nperseg=512,
[0198] noverlap=256, return_onesided=False)
[0199] tf_image = resize(np.abs(Zxx), (224, 224))
[0200] tf_image = tf_image / np.max(tf_image)
[0201] return time_domain, freq_domain, tf_image
[0202] # Create an HDF5 dataset (train, val, and test datasets) with h5py.File('cospas_sarsat_dataset.h5', 'w') as f:
[0203] for split, N in [('train', 38500), ('val', 11000), ('test',5500)]:
[0204] grp = f.create_group(split)
[0205] grp.create_dataset('time_domain', shape=(N, 2, 16384), dtype='float32')
[0206] grp.create_dataset('freq_domain', shape=(N, 8192), dtype='float32')
[0207] grp.create_dataset('tf_image', shape=(N, 224, 224), dtype='float32')
[0208] grp.create_dataset('class_labels', shape=(N,), dtype='int32')
[0209] grp.create_dataset('snr_labels', shape=(N,), dtype='int32')
[0210] grp.create_dataset('jsr_labels', shape=(N,), dtype='int32').
[0211] Following the steps described in the four examples above, a total of 55,000 HDF5 format datasets are generated, which can be used for downstream tasks such as satellite search and rescue interference identification, anti-interference algorithm verification, and satellite payload electromagnetic compatibility testing.
Claims
1. A method for generating a multimodal dataset of satellite search and rescue interference signals, which is based on three-dimensional hierarchical segmentation, characterized in that... The steps of the method include: S1. Generate an SGB clean signal according to the COSPAS-SARSAT international standard C / S T.
018. The SGB clean signal has a center frequency of 406.05 MHz, a modulation method of OQPSK-DSSS, a burst length of 1000 ms, a data rate of 300 bps, a spreading chip rate of 38.4 kchips / s, a spreading gain of 24 dB, and an effective omnidirectional radiated power of 5W. S2. Construct various types of interference signals; S3. Combine the SGB clean signal and various types of interference signals, and add additive white Gaussian noise to form a triplet time-domain sample. The triple is class, SNR, JSR: class: Multiple categories of interference; SNR: Signal-to-noise power ratio (dB); JSR: Interference / Signal Power Ratio (dB) The values of both SNR and JSR are discrete, equally spaced sets: ; ; S4. Perform a three-dimensional hierarchical segmentation of the triplet time-domain samples into training, validation, and test sets, with segmentation ratios of 70%, 20%, and 10%, respectively. S5. For each triplet time-domain sample, generate three modes simultaneously: time-domain, frequency-domain single-sided spectrum, and time-frequency graph, and store them as HDF5 format files to form a multi-modal dataset of satellite search and rescue interference signals.
2. The method for generating a multimodal dataset of satellite search and rescue interference signals according to claim 1, characterized in that, The mathematical model for the SGB clean signal is as follows: ; in: The data symbols are OQPSK modulated. A spread spectrum sequence of 128 chips per symbol; Roll-off factor Raised cosine pulse shaping filter; The chip cycle; t is the average power of the transmitted signal (dB); n is the continuous time variable; N is the mathematical symbol sequence; N is the total number of data symbols in a frame; fc is the carrier center frequency.
3. The method for generating a multimodal dataset of satellite search and rescue interference signals according to claim 1, characterized in that, In step S3, the synthesis formula for the time-domain samples is: ; in: For time-domain samples; For SGB clean signal; It is additive white Gaussian noise; Different types of interference signals; among them, the signal-to-noise ratio (SNR) of additive white Gaussian noise is defined as: 。 4. The method for generating a multimodal dataset of satellite search and rescue interference signals according to claim 1, characterized in that, The various types of interference signals are 11 categories, namely: clean, single_tone, multi_tone, chirp, pulse, wideband_noise, narrowband_noise, bpsk_modulated, comb_spectrum, frequency_hopping, and composite. The "clean" class refers to signals without interference, containing only additive white Gaussian noise. The single_tone class expression is: ; The multi_tone class expression is: ; The chirp class expression is: ; The pulse class expression is: ; The wideband_noise class expression is: ; The narrowband_noise class expression is: ; The bpsk_modulated class expression is: ; The comb_spectrum class expression is: ; The expression for the frequency_hopping class is: ; The composite class expression is: ; in: To interfere with the prototype signal; , , All are amplitudes, among which , They are respectively polyphonic. Each component, comb spectrum The amplitude of the spectral lines; For simulating intermediate frequency; , This is the frequency offset relative to the intermediate frequency; , , , As the initial phase, in Uniformly distributed values within the range; The number of notes in a polyphonic system; This is the start frequency for the frequency sweep. This is the frequency modulation slope, which is the change in instantaneous frequency per unit time. It is a periodic rectangular envelope used to characterize the on / off state of pulse interference; It is white Gaussian noise; , Bandwidth broadband, center frequency bandwidth Narrowband bandpass filtering; For taking values The binary non-return-to-zero baseband waveform; This represents the number of spectral lines in the comb pattern. For the first Spectral line frequencies (interval between adjacent lines) ); For frequency hopping instantaneous carrier frequencies, there are period hopping intervals between several discrete carrier frequencies. Switch; The number of components of the composite interference (take 2 to 4). For the first One basic interference component Use its power weighting coefficient.
5. The method for generating a multimodal dataset of satellite search and rescue interference signals according to claim 4, wherein its features are as follows: The characteristic is that each of the 11 types of interference signals generates... K samples are evenly distributed across 13×13=169 (SNR, JSR) case grids, such that each triplet contains K samples. There are 10 independent samples, and the total number of samples in the entire dataset is: 。 6. The method for generating a multimodal dataset of satellite search and rescue interference signals according to claim 5, wherein its features are as follows: The key feature is that, for the 11 types of interference signals, the three-dimensional hierarchical segmentation method in step S4 specifically satisfies the following: for each of the 11×13×13=1859 triples, it is independently segmented into a training set, a validation set, and a test set at ratios of 70%, 20%, and 10%, respectively, so that the training set, validation set, and test set have 38,500, 11,000, and 5,500 samples, respectively, and the training set, validation set, and test set all completely cover all 1859 triples.
7. The method for generating a multimodal dataset of satellite search and rescue interference signals according to claim 1, wherein... The characteristic is that, in step S5, the frequency domain single-sided spectrum is obtained by taking the modulus through Fast Fourier Transform (FFT). Length is 8192, frequency resolution: 。 8. The method for generating a multimodal dataset of satellite search and rescue interference signals according to claim 1, wherein... The characteristic is that, in step S5, the time-frequency diagram is calculated using a short-time Fourier transform (STFT), with parameters including window length nperseg = 512, overlap nooverlap = 256, and an output tensor size of 224×224. Time resolution: ; Frequency resolution: 。 9. The method for generating a multimodal dataset of satellite search and rescue interference signals according to claim 1, wherein... The feature is that, in step S5, the HDF5 format file contains three data groups: training set (train), validation set (val), and test set (test). Each data group stores three modal data types: time_domain, freq_domain, and tf_image, as well as three types of label data: class_labels, snr_labels, and jsr_labels.
10. The method for generating a multimodal dataset of satellite search and rescue interference signals according to claim 1, wherein... The characteristic is that, in step S5, the data representation of the three modalities specifically includes: I. Time Domain Representation: Composed of the real and imaginary parts of the directly sampled time domain samples, with a data shape of (2, 16384); II. Frequency Domain Single-Sided Spectrum Representation: The data is obtained by taking the first 8192 points after taking the modulus of the directly sampled time-domain sample using an FFT of length 16384 and normalizing it. The data shape is (8192,). III. The time-frequency diagram is obtained by interpolating the directly sampled time-domain samples to 224×224 after STFT mode taking with nperseg=512 and noverlap=256 and normalizing. The data shape is (224, 224).
