Partial discharge signal fingerprint identification method based on Rydberg atoms and deep learning

Abstract

The invention provides a partial discharge signal fingerprint identification method based on Rydberg atoms and deep learning, and belongs to the technical field of state monitoring of power systems. The method comprises the following steps: (1) broadband non-intrusive partial discharge detection: based on a Rydberg atom sensor, directly mapping a partial discharge electromagnetic signal into a spectrum fingerprint by using EIT and Stark effects, and realizing MHz-GHz broadband coverage; (2) spectrum fingerprint construction: converting the electric field characteristics of the partial discharge signals into optical transmission spectrum characteristics to form spectrum fingerprints capable of distinguishing different discharge types; the spectrum fingerprints are divided into a training set and a test set, the training set is used for training the deep learning network, and the test set is used for testing the recognition performance of the model; and (3) performing deep learning automatic identification: performing feature extraction and classification on the time domain data by using a deep learning model. The method has the characteristics of broadband coverage, non-intrusive type, automatic feature extraction, high robustness under a low signal-to-noise ratio, and strong system integration and expansibility.

Description

Technical Field

[0001] This invention provides a partial discharge signal fingerprinting method based on Rydberg atoms and deep learning, belonging to the field of power system state monitoring technology. Background Technology

[0002] Partial discharge (PD) is a common early breakdown phenomenon in the insulation materials of high-voltage electrical equipment, caused by microscopic defects or impurities. It is characterized by nanosecond-level pulses, electromagnetic radiation, and a wide frequency spectrum, covering MHz to GHz. The presence of partial discharge is often a precursor to insulation degradation and potential equipment failure; therefore, its detection and identification are of great significance in power system condition monitoring and fault early warning.

[0003] Electromagnetically induced transparency (EIT) is a quantum interference effect where, when two energy levels of an atomic system are simultaneously excited by probe light and coupling light, a coherent superposition occurs between the two transition paths, making the medium that originally absorbed the probe light transparent under certain conditions. In the three-level ladder structure of a Rydberg atom, the probe light excites the ground state to an intermediate state, and the coupling light excites the intermediate state to a highly excited state (Rydberg state). When the two-photon resonance condition is satisfied, a narrow-linewidth transparent window appears in the transmission spectrum.

[0004] The Ac Stark effect: Under the influence of an alternating electric field, the energy levels of atoms shift due to their interaction with the external field. The amount of displacement is related to the strength and frequency of the applied electric field. This effect is particularly significant for highly excited Rydberg states due to their extremely high polarizability. The AC Stark effect alters the center frequency and linewidth of the EIT spectral lines, thus mapping the amplitude information of the external electromagnetic field into the optical transmission spectrum. This effect can be used to achieve sensitive detection and characterization of broadband microwave or radio frequency electric fields.

[0005] One existing technology is a partial discharge detection method based on current / voltage measurement. This method acquires the current / voltage signal of the partial discharge pulse by connecting a coupling capacitor or sensing impedance to the port of high-voltage equipment, and then observes its amplitude, phase, and charge transfer characteristics using an oscilloscope or spectrum analyzer. Since the signal often experiences high-frequency attenuation and waveform distortion after transmission through LC networks such as transformer windings, analog or digital filters are typically used to extract pulse characteristics in the 40–200 kHz frequency band for analyzing the initiation voltage, extinction voltage, and defect type of the partial discharge.

[0006] This method has the following technical drawbacks:

[0007] 1. Limited bandwidth: Due to limitations of sensors and sampling systems, it is difficult to cover the wide bandwidth (MHz–GHz) of PD signals.

[0008] 2. Highly invasive: It requires the installation of coupling devices on the equipment, which affects operational safety;

[0009] 3. Features rely on human experience: Feature parameters are usually selected manually, making it difficult to guarantee robustness;

[0010] 4. High noise sensitivity: The signal is severely attenuated at long distances or in complex environments, resulting in low recognition accuracy.

[0011] The second existing technology is a broadband electromagnetic detection method based on spectrum analysis. This method uses UHF, VHF, or broadband antennas to receive electromagnetic radiation signals generated by partial discharge, and then performs frequency domain analysis on the signals using a spectrum analyzer or high-speed acquisition equipment. Subsequently, it extracts the energy distribution or spectral characteristics of different frequency bands and combines this with pattern recognition or machine learning algorithms to identify the type of partial discharge. This method has the following technical drawbacks:

[0012] 1. Reliance on large-scale instruments: Spectrum analyzers and oscilloscopes are expensive and bulky, making them unsuitable for online monitoring;

[0013] 2. Poor reliability in noisy environments: When detecting at long distances, partial discharge signals are difficult to distinguish from background noise;

[0014] 3. Feature extraction still relies on manual design, the algorithm lacks versatility, and the recognition rate drops significantly when transferred to complex working conditions. Summary of the Invention

[0015] To address the aforementioned technical problems, the present invention aims to propose a partial discharge signal fingerprinting method based on Rydberg atoms and deep learning, which can achieve the following functions:

[0016] (1) Capture broadband discharge signals in a non-contact, non-invasive manner;

[0017] (2) Convert the signal into a unique “spectral fingerprint”;

[0018] (3) High-accuracy automatic recognition can still be achieved in low signal-to-noise ratio environments through deep learning models. The complete technical solution provided by this invention is as follows:

[0019] A partial discharge signal fingerprinting method based on Rydberg atoms and deep learning includes the following steps:

[0020] (1) Wideband non-invasive partial discharge detection: Based on the Rydberg atom sensor, the partial discharge electromagnetic signal is directly mapped into a spectral fingerprint using the EIT and Stark effects, achieving wideband coverage of MHz–GHz. No metal probes or electrical coupling devices are required, ensuring non-invasiveness and safe operation of the equipment;

[0021] The energy level principle and system construction are as follows:

[0022] Using the two-photon EIT scheme, rubidium atoms are moved from their ground state 5 Excited to the Ridgberg state 58 Two-photon Rydberg EIT excitation employs a 780nm probe beam and a 480nm coupling beam, which are input into the rubidium atom gas cell via a dichroic mirror. Partial discharge signals are generated by a partial discharge source, received by an antenna, and radiated to the atom gas cell location via a waveguide. The time-domain signal received by the atomic system is transmitted to a deep learning model, which uses a ResNet residual neural network architecture.

[0023] (2) Construction of spectral fingerprint: The electric field characteristics of the partial discharge signal are converted into optical transmission spectrum characteristics to form a spectral fingerprint that can distinguish different discharge types. The spectral fingerprint is then divided into a training set and a test set. The training set is used to train the deep learning network, and the test set is used to test the model's recognition performance. This mapping process is unique and stable.

[0024] (3) Deep learning automatic recognition: 1D ResNet or equivalent deep learning model is used to directly extract features and classify time-domain data. This avoids the shortcomings of manual feature extraction, and can still maintain a high recognition accuracy (>90%), especially under low signal-to-noise ratio conditions.

[0025] The beneficial effects of the technical solution of this invention are as follows:

[0026] 1. Wideband coverage, non-invasive: Utilizing the direct interaction between Rydberg atoms and electromagnetic fields, it eliminates the need for metal probes or coupling capacitors, avoiding interference with equipment operation, and can cover a wide bandwidth (MHz–GHz) of discharge signals.

[0027] 2. Automatic feature extraction: By combining deep learning models, “spectral fingerprint” features can be directly extracted from time-domain signals, avoiding the limitations of manually designed features.

[0028] 3. High robustness under low signal-to-noise ratio: Experiments show that even when the signal-to-noise ratio drops to 2 dB, the model can still achieve a recognition accuracy of about 94%, which is significantly better than traditional methods.

[0029] 4. Strong system integration and scalability: The solution of this invention can be connected with existing power monitoring platforms to realize online monitoring; at the same time, it has the self-calibration and high sensitivity advantages unique to quantum sensing, and is suitable for various complex electromagnetic environments. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the energy level principle of the present invention;

[0031] Figure 2 This is a schematic diagram of the measurement system of the present invention;

[0032] Figure 3 This is a wideband air-gap partial discharge signal diagram of the present invention;

[0033] Figure 4 The time-domain waveform of the air gap partial discharge signal obtained by the Rydberg atom sensing system of the present invention is shown.

[0034] Figure 5 This invention relates to the deep learning model's focus on different time regions;

[0035] Figure 6 This is the spectral fingerprint recognition process of the present invention;

[0036] Figure 7 These are the recognition results at different training stages of the present invention.

[0037] Figure 8 This is one of the training dynamics of the present invention;

[0038] Figure 9 This is the second training dynamic of the present invention;

[0039] Figure 10 This is a visualization result of t-SNE in this invention. Detailed Implementation

[0040] The partial discharge signal fingerprinting method based on Rydberg atoms and deep learning provided by this invention includes the following steps:

[0041] Step 1: Energy Level Principle and System Construction

[0042] Energy level principle, such as Figure 1 As shown, rubidium atoms are moved from their ground state 5 by a two-photon EIT scheme. Excited to the Ridgberg state 58 Measurement systems such as Figure 2As shown, the system consists of a laser module, an atomic gas cell, a partial discharge (PD) signal source, and a deep learning model. Two-photon Rydberg EIT excitation utilizes a 780nm probe beam and a 480nm coupling beam, which are input into the rubidium atomic gas cell (Rb Vapor Cell) via a dichroic mirror. The PD signal is generated by a PD signal source, received by an antenna, and radiated to the atomic gas cell location via a waveguide. The time-domain signal received by the atomic system is transmitted to... Figure 2 The deep learning model shown uses the ResNet residual neural network architecture.

[0043] Step 2: Construction and Recognition of Spectral Fingerprints

[0044] The process of constructing spectral fingerprints is as follows: Figures 3 to 5 As shown, Figure 3 This demonstrates a wideband air-gap partial discharge (PGD) signal, with a frequency range of 0-6 GHz. The time-domain waveform of the PDD signal obtained by the Rydberg atom sensing system is shown below. Figure 4 As shown, the duration is 100ms. Figure 5 This demonstrates the focus of deep learning models on different time periods, showing that the models pay more attention to points where the signal amplitude changes abruptly.

[0045] The process of spectral fingerprint recognition is as follows: Figure 6 As shown, the partial discharge signal is encoded into a spectral fingerprint by the Rydberg atom sensing system, and then divided into two parts: a training set and a test set. The training set is used to train the deep learning network, and the test set is used to test the model's recognition performance.

[0046] Step 3: Spectrum Fingerprint Recognition Results

[0047] As the distance between the antenna and the partial discharge source increases, the signal attenuates rapidly, while noise increases, leading to a significant decrease in the signal-to-noise ratio (SNR). When the distance increases from 1 cm to 30 cm, the SNR drops from 16 dB to 2 dB. In this situation, the salient features of the signal become very weak, making it necessary to use deep learning models to extract the latent representations of the signal for reliable identification.

[0048] Figure 7 The recognition results at different training stages were summarized. As shown in the matrix, in the early stages of training, the network incorrectly classified the signal as a corona, resulting in a low overall accuracy of only 46.5%. By the 8th training iteration, the predictions were still incorrect, but the overall accuracy improved to 71.5%. After sufficient training, the model correctly identified the signal as an air gap signal, and the classification accuracy improved to 93.5%.

[0049] Training dynamics such as Figure 8 , Figure 9As shown, due to the extremely low signal-to-noise ratio, the model requires nearly 50 training iterations to converge, and the learning curve exhibits significant oscillations before stabilizing. Ultimately, the training accuracy approaches 100%, while the validation accuracy remains stable at 94%, indicating that although the generalization ability is somewhat limited, the model's learning performance is still good. Figure 10 The t-SNE visualization further reflects this trend; the four signal categories remain separable, but the clusters appear less compact and partially overlapping, suggesting that noise contamination reduces class separability in the latent space.

Claims

1. A partial discharge signal fingerprinting method based on Rydberg atoms and deep learning, characterized in that, Includes the following steps: (1) Wideband non-invasive partial discharge detection: Based on the Rydberg atom sensor, the partial discharge electromagnetic signal is directly mapped into a spectral fingerprint using the EIT and Stark effects to achieve wideband coverage of MHz–GHz; (2) Construction of spectral fingerprint: The electric field characteristics of the partial discharge signal are transformed into optical transmission spectrum characteristics to form a spectral fingerprint that can distinguish different discharge types; the spectral fingerprint is divided into two parts: training set and test set. The training set is used to train the deep learning network, and the test set is used to test the recognition performance of the model. (3) Deep learning automatic recognition: use deep learning models to extract features and classify time-domain data.

2. The partial discharge signal fingerprinting method based on Rydberg atoms and deep learning according to claim 1, characterized in that, Step (1) Using the two-photon EIT scheme, rubidium atoms are moved from the ground state 5 Excited to the Ridgberg state 58 The two-photon Rydberg EIT excitation uses a 780nm probe beam and a 480nm coupling beam, which are input into the rubidium atom gas cell in opposite directions through a dichroic mirror. The partial discharge signal is generated by a partial discharge signal source, received by an antenna, and radiated to the atom gas cell location through a waveguide. The time-domain signal received by the atom system is transmitted to the deep learning model.

3. The partial discharge signal fingerprinting method based on Rydberg atoms and deep learning according to claim 1, characterized in that, The deep learning model uses 1D ResNet or an equivalent deep learning model.

Images