Transformer internal partial discharge detection method, device and equipment and storage medium

Abstract

The invention provides a transformer internal partial discharge detection method, device and equipment and a storage medium, and belongs to the field of transformer monitoring, and the method comprises the steps: obtaining an ultrahigh frequency electric signal, an optical signal and an acoustic signal generated by transformer internal partial discharge through an ultrahigh frequency sensor, an optical sensor and an optical fiber Fabry-Perot ultrasonic sensor; a multi-state quantity integrated sensing SoC chip on the circuit board is used for conditioning and digitizing the signals to generate a digitized signal flow; inputting the signal flow into a state sensing special edge computing NPU chip, processing the signal flow by using a graph attention network algorithm, calculating an attention coefficient between sensor nodes, and fusing the multi-modal signals to generate a fusion signal; and determining the partial discharge type according to the fusion signal. According to the invention, three sensors are internally integrated, signal conditioning digitalization is carried out by using the SoC chip, multi-mode signals are fused through the NPU chip, high-precision and high-reliability identification is realized, and the difficulty of an external sensor is overcome.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of transformer monitoring, and particularly relates to a transformer internal partial discharge detection method, device, equipment and storage medium. BACKGROUND

[0002] The detection of transformer partial discharge mainly includes ultrasonic wave method and ultra-high frequency (UHF) method, which is used for monitoring the partial discharge caused by internal insulation degradation of the transformer, so as to prevent insulation fault accidents and ensure the safe and stable operation of the power grid.

[0003] The prior art specifically adopts an external sensor, such as an ultrasonic sensor or an ultra-high frequency sensor, which is installed outside the transformer oil tank. In this way, the sensor detects the acoustic or electrical signals generated by the partial discharge, but due to the need for the signals to penetrate the oil-paper insulation and the thick metal tank shell, serious signal attenuation and waveform distortion will occur; at the same time, the external sensor is easily affected by the strong external electromagnetic interference and background noise in the substation site, resulting in low signal-to-noise ratio and insufficient sensitivity of the detection system, and it is difficult to accurately capture early weak discharge signals.

[0004] In the process of working independently through the external sensor, since the installation position of the sensor depends on experience, there is a lack of unified standard, the data is difficult to effectively fuse, and it is impossible to form multi-dimensional cross verification, which exists a monitoring blind area. This together causes frequent false positives and false negatives, and it is impossible to realize high-reliability partial discharge identification. SUMMARY

[0005] The purpose of the present application is to overcome the defects in the prior art, and to provide a transformer internal partial discharge detection method, device, equipment and storage medium.

[0006] The present application provides a transformer internal partial discharge detection method, comprising:

[0007] Obtaining multi-modal sensing signals from an ultra-high frequency sensor, an optical sensor and a fiber-optic Fabry-Perot ultrasonic sensor, the ultra-high frequency sensor being used for detecting ultra-high frequency electrical signals generated by the transformer internal partial discharge, the optical sensor being used for detecting optical signals generated by the transformer internal partial discharge, and the fiber-optic Fabry-Perot ultrasonic sensor being used for detecting acoustic signals generated by the transformer internal partial discharge;

[0008] The multi-modal sensing signals are conditioned and digitized by a multi-state quantity integrated perception SoC chip on a circuit board to generate a digitized signal stream;

[0009] The digitized signal stream is input into a state perception special-purpose edge computing NPU chip, the NPU chip uses a graph attention network algorithm to process the digitized signal stream, calculates attention coefficients between sensor nodes, and fuses multi-modal signals based on the attention coefficients to generate a fused signal;

[0010] According to the fusion signal, a type identification result of internal partial discharge of the transformer is determined.

[0011] Optionally, the multi-modal sensing signal obtained from the ultra-high frequency sensor, the optical sensor and the fiber Fabry-Perot ultrasonic sensor comprises:

[0012] The fiber Fabry-Perot ultrasonic sensor converts the acoustic signal into an optical signal through a Fabry-Perot cavity, the Fabry-Perot cavity is composed of a vibrating diaphragm and a single-mode optical fiber end face, and the vibrating diaphragm modulates the optical signal in response to acoustic waves;

[0013] The optical signal is transmitted to the circuit board through an optical fiber.

[0014] Optionally, the multi-modal sensing signal is conditioned and digitized by a multi-state quantity integrated perception SoC chip on the circuit board to generate a digitized signal stream, comprising:

[0015] The SoC chip receives the multi-modal sensing signal through a reconfigurable front-end interface, and the reconfigurable front-end interface is adapted to voltage type, current type and capacitance type signals;

[0016] The SoC chip performs special conditioning on the multi-modal sensing signal, including broadband amplification, transimpedance amplification and filtering;

[0017] The digitized signal stream in a unified format is generated.

[0018] Optionally, the multi-modal sensing signal is conditioned and digitized by a multi-state quantity integrated perception SoC chip on the circuit board to generate a digitized signal stream, comprising:

[0019] The SoC chip adopts an event-driven adaptive working mechanism and is in a dormant state when there is no signal, and is awakened by a signal pulse;

[0020] After being awakened, the multi-modal sensing signal is conditioned and digitized in real time.

[0021] Optionally, the digitized signal stream is input into a state perception special edge computing NPU chip, the NPU chip uses a graph attention network algorithm to process the digitized signal stream, calculates attention coefficients between sensor nodes, and fuses multi-modal signals based on the attention coefficients to generate a fusion signal, comprising:

[0022] The NPU chip executes the graph attention network algorithm through a neural network hardware accelerator, and the accelerator is based on a RISC-V processor core and an NPU unit.

[0023] The accelerator calculates the attention coefficients and signal fusion in parallel.

[0024] Optionally, the digital signal stream is input into a state-aware special-purpose edge computing NPU chip, the NPU chip uses a graph attention network algorithm to process the digital signal stream, calculates attention coefficients between sensor nodes, and fuses multi-modal signals based on the attention coefficients to generate a fused signal, including:

[0025] The NPU chip performs model pruning and INT8 quantization processing on the graph attention network algorithm to compress the model size;

[0026] The compressed algorithm is used to process the digital signal stream.

[0027] Optionally, the digital signal stream is input into a state-aware special-purpose edge computing NPU chip, the NPU chip uses a graph attention network algorithm to process the digital signal stream, calculates attention coefficients between sensor nodes, and fuses multi-modal signals based on the attention coefficients to generate a fused signal, including:

[0028] The NPU chip evaluates the signal-to-noise ratio of each sensor signal and dynamically calculates the weight according to the signal-to-noise ratio;

[0029] The attention coefficients are adjusted based on the weight, and the multi-modal signals are fused to generate the fused signal.

[0030] The application also provides a transformer internal partial discharge detection device, comprising:

[0031] An acquisition module acquires multi-modal sensing signals from a UHF sensor, an optical sensor, and a fiber-optic Fabry-Perot ultrasonic sensor, the UHF sensor is used to detect UHF electrical signals generated by internal partial discharge of a transformer, the optical sensor is used to detect optical signals generated by internal partial discharge of a transformer, and the fiber-optic Fabry-Perot ultrasonic sensor is used to detect acoustic signals generated by internal partial discharge of a transformer;

[0032] A modulation module modulates and digitizes the multi-modal sensing signals by a multi-state quantity integrated perception SoC chip on a circuit board to generate a digital signal stream;

[0033] A fusion module inputs the digital signal stream into a state-aware special-purpose edge computing NPU chip, the NPU chip uses a graph attention network algorithm to process the digital signal stream, calculates attention coefficients between sensor nodes, and fuses multi-modal signals based on the attention coefficients to generate a fused signal;

[0034] An identification module determines a type identification result of internal partial discharge of a transformer according to the fused signal.

[0035] Optionally, the acquisition module acquires multimodal sensing signals from a UHF sensor, an optical sensor, and a fiber optic Fabry-Perot ultrasonic sensor, including:

[0036] The fiber optic Fabry-Perot ultrasonic sensor converts acoustic signals into optical signals through a Fabry-Perot cavity, which is composed of a vibrating diaphragm and a single-mode fiber end face. The vibrating diaphragm modulates the optical signal in response to the vibration of the acoustic wave.

[0037] The optical signal is transmitted to the circuit board via optical fiber.

[0038] Optionally, the modulation module uses a multi-state integrated sensing SoC chip on a circuit board to condition and digitize the multimodal sensing signals to generate a digital signal stream, including:

[0039] The SoC chip receives the multimodal sensing signals through a reconfigurable front-end interface, which is compatible with voltage-type, current-type, and capacitive signals.

[0040] The SoC chip performs dedicated conditioning on the multimodal sensing signal, including broadband amplification, transimpedance amplification, and filtering.

[0041] Generate the digital signal stream in a uniform format.

[0042] Optionally, the modulation module uses a multi-state integrated sensing SoC chip on a circuit board to condition and digitize the multimodal sensing signals to generate a digital signal stream, including:

[0043] The SoC chip adopts an event-driven adaptive working mechanism, which is in a sleep state when there is no signal and is woken up by a signal pulse.

[0044] After being woken up, the multimodal sensing signals are conditioned and digitized in real time.

[0045] Optionally, the fusion module inputs the digitized signal stream into a state-aware dedicated edge computing NPU chip. The NPU chip processes the digitized signal stream using a graph attention network algorithm, calculates the attention coefficients between sensor nodes, and fuses multimodal signals based on the attention coefficients to generate a fused signal, including:

[0046] The NPU chip executes the graph attention network algorithm through a neural network hardware accelerator, which is based on the collaboration between the RISC-V processor core and the NPU unit.

[0047] The accelerator performs parallel calculations of the attention coefficients and signal fusion.

[0048] Optionally, the fusion module inputs the digitized signal stream into a state-aware dedicated edge computing NPU chip. The NPU chip processes the digitized signal stream using a graph attention network algorithm, calculates the attention coefficients between sensor nodes, and fuses multimodal signals based on the attention coefficients to generate a fused signal, including:

[0049] The NPU chip performs model pruning and INT8 quantization on the graph attention network algorithm to compress the model size.

[0050] The digitized signal stream is processed using a compressed algorithm.

[0051] Optionally, the fusion module inputs the digitized signal stream into a state-aware dedicated edge computing NPU chip. The NPU chip processes the digitized signal stream using a graph attention network algorithm, calculates the attention coefficients between sensor nodes, and fuses multimodal signals based on the attention coefficients to generate a fused signal, including:

[0052] The NPU chip evaluates the signal-to-noise ratio of each sensor signal and dynamically calculates the weights based on the signal-to-noise ratio.

[0053] The attention coefficients are adjusted based on the weights, and the multimodal signals are fused to generate the fused signal.

[0054] This application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0055] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the above-described method.

[0056] The beneficial effects of this application are:

[0057] This application provides a method for detecting partial discharge inside a transformer, comprising: acquiring multimodal sensing signals from a UHF sensor, an optical sensor, and a fiber optic Fabry-Perot ultrasonic sensor; wherein the UHF sensor is used to detect UHF electrical signals generated by partial discharge inside the transformer, the optical sensor is used to detect optical signals generated by partial discharge inside the transformer, and the fiber optic Fabry-Perot ultrasonic sensor is used to detect acoustic signals generated by partial discharge inside the transformer; conditioning and digitizing the multimodal sensing signals by a multi-state integrated sensing SoC chip on a circuit board to generate a digital signal stream; inputting the digital signal stream into a state-aware dedicated edge computing NPU chip; the NPU chip using a graph attention network algorithm to process the digital signal stream, calculating attention coefficients between sensor nodes, and fusing the multimodal signals based on the attention coefficients to generate a fused signal; and determining the type identification result of partial discharge inside the transformer based on the fused signal. This application integrates a high-frequency sensor, an optical sensor, and a fiber optic Fabry-Perot ultrasonic sensor, and uses a multi-state integrated sensing SoC chip for signal conditioning and digitization. Then, it uses a state-aware dedicated edge computing NPU chip to fuse multi-modal signals using a graph attention network algorithm, achieving high-precision and high-reliability identification of partial discharge types inside transformers. This overcomes the challenges of signal attenuation, interference, and data fusion associated with external sensors. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the partial discharge detection process inside the transformer in this application;

[0059] Figure 2 This is a schematic diagram of the built-in acoustic-optical-electric integrated coupling sensor structure in this application;

[0060] Figure 3 This is a schematic diagram of the built-in ultra-high frequency sensor structure in this application;

[0061] Figure 4 This is a schematic diagram of the optical sensor structure in this application;

[0062] Figure 5 This is a schematic diagram of the fiber optic Fabry-Perot sensor structure in this application;

[0063] Figure 6 This is a schematic diagram of the transformer mounting cover structure in this application;

[0064] Figure 7 This is a schematic diagram of the rain cover structure in this application. Detailed Implementation

[0065] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is to be understood that various forms of implementation of the present disclosure are intended and should not be limited to the embodiments set forth herein. Rather, the embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0066] Please refer to Figure 1 This application provides a method for detecting partial discharge inside a transformer, applied in the field of transformer monitoring, to solve the problems of severe signal attenuation, susceptibility to interference, and low sensitivity of traditional external detection methods. The method includes:

[0067] S101. Obtain multimodal sensing signals from an ultra-high frequency sensor, an optical sensor, and a fiber optic Fabry-Perot ultrasonic sensor. The ultra-high frequency sensor is used to detect ultra-high frequency electrical signals generated by partial discharge inside the transformer. The optical sensor is used to detect optical signals generated by partial discharge inside the transformer. The fiber optic Fabry-Perot ultrasonic sensor is used to detect acoustic signals generated by partial discharge inside the transformer.

[0068] like Figure 2 and Figure 3 As shown, the UHF sensor 2 consists of three parts: antenna 2.1, housing 2.2, and N-type connector 2.3. The antenna 2.1 is located on the top of the sensor and is dish-shaped. It is used to receive UHF signals generated by partial discharge. The housing 2.2 has an external thread at the tail, which cooperates with the threaded hole 1.2 in the mounting cover plate 1 to fix the sensor. The inside of the housing is filled with epoxy resin for insulation, which insulates the antenna 2.1 and the housing 2.2. The N-type connector 2.3 is installed at the tail of the housing 2.2. The connector extends into the housing and connects to the antenna 2.1 to transmit the UHF signals received by the antenna 2.1.

[0069] like Figure 2 and Figure 4As shown, the optical sensor 3 consists of four parts: a light guide post 3.1, a fixing flange 3.2, a cover 3.4, and a cap. The light guide post 3.1 is made of a material with good light transmission and is cylindrical in shape. The fixing flange 3.2 has eight holes around its perimeter, including four threaded holes and four through holes. The fixing flange 3.3 has eight identical countersunk holes around its perimeter. Four of these countersunk holes mate with the threaded holes on the fixing flange 3.2 to fix the two together and secure the light guide post 3.1. The other four holes mate with the through holes on the fixing flange 3.2. Four screws pass through the four holes and then fix the optical sensor 3 to the cap 1. The light guide post 3.1 extends into the transformer tank through the hole 1.1 in the cap 1 to receive light signals. The fixing flange 3.2 has a sealing groove for sealing after installing a sealing ring. The cap 1 is connected to the fixing flange 3.3 and has a fiber optic interface at the end for connecting optical fibers for signal transmission.

[0070] like Figure 2 and Figure 5 As shown, the fiber optic Fabry-Perot ultrasonic sensor 4 converts acoustic signals into optical signals through a Fabry-Perot cavity. The Fabry-Perot cavity is composed of a vibrating diaphragm and a single-mode optical fiber end face. The vibrating diaphragm vibrates in response to the acoustic wave to modulate the optical signal, and the optical signal is transmitted to the circuit board through the optical fiber.

[0071] Specifically, the fiber optic Fabry-Perot ultrasonic sensor 4 consists of three parts: the fiber optic Fabry-Perot ultrasonic sensor 4.1, the encapsulation shell 4.2, and the acoustically transparent membrane 4.3. The fiber optic Fabry-Perot ultrasonic sensor 4.1 converts acoustic signals into optical signals using the Fabry-Perot interference principle. Its core structure comprises a highly sensitive vibrating diaphragm and the end face of a single-mode optical fiber, forming a Fabry-Perot resonant cavity. When the ultrasonic signal generated by external partial discharge causes the sensitive diaphragm to vibrate, it causes a slight change in the cavity length of the Fabry-Perot cavity, thereby modulating the intensity of the interference light of the reflected light. By detecting this change in light intensity, the original ultrasonic signal can be demodulated. The encapsulation shell 4.2 is composed of three cylinders of different diameters connected together. Inside is a countersunk hole that is open at both ends. The fiber optic Fabry-Perot ultrasonic sensor 4.1 is placed at the bottom of the countersunk hole, and an optical fiber extends from the tail end. The top of the countersunk hole is sealed with a dustproof and waterproof acoustically transparent membrane 4.3. The ultrasonic signal generated by partial discharge inside the transformer can pass through the acoustically transparent membrane 4.3 and be received by the fiber optic Fabry-Perot ultrasonic sensor 4.1 and transmitted through the optical fiber.

[0072] like Figure 2 and Figure 6As shown, all the sensors are mounted on the transformer mounting cover 1. The cover 1 is circular or square in shape, with a ring of through holes around the edge for fixing to the outer wall of the transformer tank. A through hole 1.1 is located slightly above the center line, and around this through hole 1.1 are four threaded blind holes for fixing the optical sensor 3. The light guide extends into the transformer tank through the through hole 1.1. A larger threaded through hole 1.2 is located slightly below the center line of the cover 1 for mounting the ultra-high frequency sensor 2. The threaded hole 1.2 is slightly recessed on the inner side of the transformer tank, with a sealing groove at the bottom for placing a sealing ring. During installation, the ultra-high frequency sensor 2 is embedded in the recessed part and the sealing ring is pressed tightly for sealing. Slightly to the left between the through hole 1.1 and the threaded through hole 1.2 is a countersunk hole 1.3. The diameter of the hole is larger on the inner side of the transformer tank and smaller on the outer side. The fiber optic Fabry-Perot sensor 4 is installed by passing through the countersunk hole 1.3 from the inner side of the transformer tank.

[0073] like Figure 2 and Figure 7 As shown, after installation, the inside of the countersunk hole 1.3 is sealed with epoxy resin. On the outer side of the transformer tank in the cover plate 1, there is a ring of eight threaded blind holes for installing the rain cover 5. The rain cover 5 consists of two parts: an outer shell 5.1 and a cover 5.2, used to protect the part of the sensor exposed outside the transformer tank. Due to its outdoor environment, it is made of stainless steel. The bottom of the outer shell 5.1 has a ring of eight holes that fit with the threaded holes in the cover plate for fixation. The top of the outer shell 5.1 has a ring of eight threaded holes for fixing the cover 5.2. The side of the outer shell 5.1 has a through hole for cable routing; the cables of the three sensors pass through this hole and connect to the terminal. The cover 5.2, located inside the rain cover 5, has four posts for mounting the circuit board 6.

[0074] This application achieves high sensitivity and high anti-interference through an internal layout. All sensing cores are located inside the transformer tank, directly capturing signals at the signal source, avoiding the signal attenuation problem of traditional external sensors. At the same time, the transformer's metal casing serves as a natural electromagnetic shielding layer, effectively resisting external interference.

[0075] S102. The multi-state integrated sensing SoC chip on the circuit board conditions and digitizes the multimodal sensing signal to generate a digital signal stream;

[0076] The SoC chip receives multimodal sensing signals through a reconfigurable front-end interface, which is compatible with voltage-type, current-type, and capacitive signals. The SoC chip performs dedicated conditioning on the multimodal sensing signals, including broadband amplification, transimpedance amplification, and filtering.

[0077] The ultra-high frequency channel integrates a broadband amplifier and peak hold circuitry. The optical channel integrates a high-sensitivity transimpedance amplifier using a reconfigurable current interface to convert the weak current pulses output from the photomultiplier tube of the optical sensor into voltage signals. The ultrasonic channel integrates a low-noise amplifier and a bandpass filter.

[0078] In addition, the SoC chip employs reconfigurable interface technology to be compatible with different types of sensor outputs.

[0079] In the digitization process, the SoC chip is equipped with a successive approximation ADC for high-speed signals such as ultra-high frequency signals and a ΣΔ ADC for low-frequency high-precision signals such as ultrasound signals, generating a digital signal stream in a unified format.

[0080] Meanwhile, the SoC chip employs an event-driven adaptive operating mechanism, entering a sleep state when there is no signal and waking up when triggered by a signal pulse. After waking up, it performs real-time conditioning and digitization of multimodal sensing signals. Specifically, the chip features an event-driven adaptive operating mechanism, entering a sleep state with static power consumption of less than 1μW when there is no signal, and waking up when triggered by a randomly input signal pulse, achieving extremely low operating power consumption, with a power consumption of 40-60mW per channel.

[0081] The analog front-end design of the SoC chip ensures high-fidelity signal conditioning. The broadband amplifier in the ultra-high frequency channel processes high-frequency signals, the transimpedance amplifier in the optical channel converts weak photocurrents, and the filter in the ultrasonic channel eliminates noise. These conditioning steps, combined with an event-driven mechanism, significantly reduce system power consumption and improve real-time performance.

[0082] This application integrates edge computing to achieve intelligent front-end diagnosis and rapid response, and replaces the traditional massive data remote transmission mode by processing signals locally at the sensor front end.

[0083] S103. The digital signal stream is input into a state-aware edge computing NPU chip. The NPU chip uses a graph attention network algorithm to process the digital signal stream, calculates the attention coefficients between sensor nodes, and fuses multimodal signals based on the attention coefficients to generate a fused signal.

[0084] The NPU chip executes graph attention network algorithms through a neural network hardware accelerator. This accelerator is based on the collaboration between a RISC-V processor core and the NPU unit, and it computes attention coefficients and signal fusion in parallel. Specifically, the NPU chip adopts a heterogeneous architecture combining a RISC-V processor core and a neural network hardware accelerator, achieving an AI computing power of up to 635 GOPS with extremely low power consumption.

[0085] Meanwhile, the NPU chip performs model pruning and INT8 quantization on the graph attention network algorithm to compress the model size, and uses the compressed algorithm to process the digital signal stream. Specifically, to efficiently run diagnostic algorithms on this NPU, the neural network model is first compressed using techniques such as model pruning and INT8 quantization, and then the algorithm is deployed into the chip through a custom simplified instruction set and assembler.

[0086] In addition, the NPU chip evaluates the signal-to-noise ratio (SNR) of each sensor signal, dynamically calculates weights based on the SNR, adjusts the attention coefficients based on the weights, and fuses multimodal signals to generate a fused signal, which specifically includes two stages:

[0087] The first stage is the determination of the existence of internal discharge, which is used for real-time triggering and preliminary assessment.

[0088] The chip first standardizes the signals from each sensing channel and extracts multi-dimensional features. The extracted feature vectors include time-domain features and frequency-domain features. The time-domain features include skewness, kurtosis, approximate entropy, and sample entropy, while the frequency-domain features include spectral centroid and spectral entropy, providing rich input information for subsequent accurate judgment.

[0089] The formula for calculating skewness is:

[0090]

[0091] Where g1 is the skewness, x is the signal value, μ is the mean, and σ is the standard deviation.

[0092] The formula for calculating kurtosis is:

[0093]

[0094] Wherein, g2 is the kurtosis.

[0095] The formula for calculating the centroid of the spectrum is:

[0096]

[0097] Where SC is the spectral centroid, f i p(f) is the frequency value. i () represents the power spectrum.

[0098] The formula for calculating spectral entropy is:

[0099]

[0100] Where SE is the spectral entropy, p i It represents a probability distribution.

[0101] Subsequently, the chip evaluates the signal-to-noise ratio of the signals from each sensing channel and performs adaptive weighted fusion according to the following formula:

[0102]

[0103] Among them, w i It is the dynamic weight of the i-th sensor, SBR i is the signal-to-noise ratio of the i-th sensor, and N is the number of sensors.

[0104] The fused multidimensional features are fed into a hardware-accelerated deep belief network, which is trained offline using a contrastive divergence algorithm. Its energy function is defined as:

[0105]

[0106] Where E(v,h) is the energy function, v is the visible layer state, h is the hidden layer state, a and b are bias vectors, and W is the weight matrix.

[0107] The conditional probability distribution of each layer of the network is as follows:

[0108]

[0109]

[0110] Among them, p(h j =1|v) is the conditional probability of the hidden layer, p(v) j =1|h) is the conditional probability of the visible layer, and σ is the sigmoid function.

[0111] Using the above model, the chip can accurately determine whether a real partial discharge exists. This determination result forms a state sequence, which is then modeled by the forward algorithm of the Hidden Markov Model.

[0112]

[0113] Among them, a t+1 (j) is the state probability at time t+1, a t (i) is the state probability at time t, a ij It is the state transition probability, b j (O t+1 ) is the observation probability.

[0114] This hidden Markov model is trained using the Baum-Welch algorithm, enabling it to determine the fault development trend by analyzing the state transition probability.

[0115] The second stage is partial discharge type identification. After determining that a discharge exists, the NPU immediately calls the second stage algorithm for fine classification. This algorithm adopts a model that integrates a multi-branch convolutional neural network and a graph attention network.

[0116] First, the signal is fed into a multi-branch convolutional neural network, which uses convolutional kernels of different scales to extract multi-scale temporal features of each signal in parallel. The convolution operation is as follows:

[0117]

[0118] Among them, y i For output features, x i For the input signal, k ij For the convolution kernel, b j f is the bias, and f is the activation function.

[0119] A batch normalization layer is added after each convolutional layer to accelerate convergence and stabilize training. The batch normalization is calculated as follows:

[0120]

[0121] in, Here, E[x] is the normalized value, E[x] is the mean, and Var[x] is the variance. It is a small constant. The output is γ, and β are learnable parameters.

[0122] Feature maps extracted from different branches are fused through an attention-weighted mechanism, which includes channel attention and spatial attention modules. Channel attention is achieved through compression and activation operations.

[0123]

[0124] To recalibrate the characteristic response of the channel.

[0125] Where s is the attention weight, z is the input feature, W1 and W2 are weight matrices, δ is the ReLU activation function, and σ is the sigmoid function.

[0126] Spatial attention is used to capture the importance of a location:

[0127]

[0128] Among them, M s F represents the spatial attention map, F represents the feature map, AugPool represents average pooling, and MaxPool represents max pooling. 7×7 It is a 7×7 convolution.

[0129] Subsequently, the fused features from multiple sensors are fed into a graph attention network for higher-dimensional information fusion. This network treats each sensor as a graph node and aggregates multimodal information by weighting the attention coefficients between nodes. The formula for calculating the attention coefficients is as follows:

[0130]

[0131] Where, α ij Here, a is the attention coefficient between nodes i and j, a is the attention vector, W is the weight matrix, and h is the weight coefficient. i and h j For node features, N i The set of neighbors of node i, where LeakyReLU is the activation function.

[0132] This effectively integrates the spatial and functional relationships between sound, light, and electrical signals.

[0133] Finally, the model training uses a composite loss function:

[0134]

[0135] Among them, the classification loss L class Employing the focus loss function:

[0136]

[0137] Among them, FL(p t ) represents the focal loss, p t To predict the probability, α t γ is the equilibrium parameter, and γ is the adjustment factor.

[0138] Regression loss L reg Use Huber loss to improve robustness to outliers.

[0139] Furthermore, the NPU chip's algorithm implementation is based on dedicated hardware acceleration. The deep belief network is trained through a contrastive divergence algorithm to learn the deep features of the signal, the hidden Markov model analyzes temporal changes, the multi-branch convolutional neural network extracts multi-scale features, and the graph attention network fuses multi-modal information. The combination of these algorithms ensures the accuracy of discharge presence judgment and type identification.

[0140] This application utilizes multi-physical quantity collaborative detection to provide richer diagnostic information. Acoustic, optical, and electrical sensors simultaneously capture signals, enabling cross-verification and avoiding misjudgment or missed diagnosis.

[0141] S104. Based on the fused signal, determine the type identification result of the partial discharge inside the transformer.

[0142] Based on the fused signal processed by the NPU chip, the system can identify the type of partial discharge, such as tip discharge, floating discharge or surface discharge, and transmit the diagnostic results to the back-end main system through the cable outlet of the rainproof cover, so as to achieve real-time response and intelligent diagnosis.

[0143] Furthermore, the internal circuit board of the rain cover integrates edge computing functionality.

[0144] likeFigure 7 As shown, circuit board 6 integrates a sensing SoC and an NPU chip to achieve real-time processing of sound, light, and electrical signals.

[0145] The advantages of this application include the high anti-interference capability of the built-in layout, the fast response of edge computing, and the reliability of multimodal fusion. These advantages ensure the high efficiency and accuracy of transformer partial discharge detection.

[0146] This application also provides a transformer internal partial discharge detection device, comprising:

[0147] The acquisition module acquires multimodal sensing signals from an ultra-high frequency sensor, an optical sensor, and a fiber optic Fabry-Perot ultrasonic sensor. The ultra-high frequency sensor is used to detect ultra-high frequency electrical signals generated by partial discharge inside the transformer, the optical sensor is used to detect optical signals generated by partial discharge inside the transformer, and the fiber optic Fabry-Perot ultrasonic sensor is used to detect acoustic signals generated by partial discharge inside the transformer.

[0148] The modulation module, consisting of a multi-state integrated sensing SoC chip on a circuit board, conditions and digitizes the multimodal sensing signals to generate a digital signal stream;

[0149] The fusion module inputs the digital signal stream into a state-aware edge computing NPU chip. The NPU chip uses a graph attention network algorithm to process the digital signal stream, calculates the attention coefficients between sensor nodes, and fuses multimodal signals based on the attention coefficients to generate a fused signal.

[0150] The identification module determines the type of partial discharge inside the transformer based on the fused signal.

[0151] Optionally, the acquisition module acquires multimodal sensing signals from a UHF sensor, an optical sensor, and a fiber optic Fabry-Perot ultrasonic sensor, including:

[0152] The fiber optic Fabry-Perot ultrasonic sensor converts acoustic signals into optical signals through a Fabry-Perot cavity, which is composed of a vibrating diaphragm and a single-mode fiber end face. The vibrating diaphragm modulates the optical signal in response to the vibration of the acoustic wave.

[0153] The optical signal is transmitted to the circuit board via optical fiber.

[0154] Optionally, the modulation module uses a multi-state integrated sensing SoC chip on a circuit board to condition and digitize the multimodal sensing signals to generate a digital signal stream, including:

[0155] The SoC chip receives the multimodal sensing signals through a reconfigurable front-end interface, which is compatible with voltage-type, current-type, and capacitive signals.

[0156] The SoC chip performs dedicated conditioning on the multimodal sensing signal, including broadband amplification, transimpedance amplification, and filtering.

[0157] Generate the digital signal stream in a uniform format.

[0158] Optionally, the modulation module uses a multi-state integrated sensing SoC chip on a circuit board to condition and digitize the multimodal sensing signals to generate a digital signal stream, including:

[0159] The SoC chip adopts an event-driven adaptive working mechanism, which is in a sleep state when there is no signal and is woken up by a signal pulse.

[0160] After being woken up, the multimodal sensing signals are conditioned and digitized in real time.

[0161] Optionally, the fusion module inputs the digitized signal stream into a state-aware dedicated edge computing NPU chip. The NPU chip processes the digitized signal stream using a graph attention network algorithm, calculates the attention coefficients between sensor nodes, and fuses multimodal signals based on the attention coefficients to generate a fused signal, including:

[0162] The NPU chip executes the graph attention network algorithm through a neural network hardware accelerator, which is based on the collaboration between the RISC-V processor core and the NPU unit.

[0163] The accelerator performs parallel calculations of the attention coefficients and signal fusion.

[0164] Optionally, the fusion module inputs the digitized signal stream into a state-aware dedicated edge computing NPU chip. The NPU chip processes the digitized signal stream using a graph attention network algorithm, calculates the attention coefficients between sensor nodes, and fuses multimodal signals based on the attention coefficients to generate a fused signal, including:

[0165] The NPU chip performs model pruning and INT8 quantization on the graph attention network algorithm to compress the model size.

[0166] The digitized signal stream is processed using a compressed algorithm.

[0167] Optionally, the fusion module inputs the digitized signal stream into a state-aware dedicated edge computing NPU chip. The NPU chip processes the digitized signal stream using a graph attention network algorithm, calculates the attention coefficients between sensor nodes, and fuses multimodal signals based on the attention coefficients to generate a fused signal, including:

[0168] The NPU chip evaluates the signal-to-noise ratio of each sensor signal and dynamically calculates the weights based on the signal-to-noise ratio.

[0169] The attention coefficients are adjusted based on the weights, and the multimodal signals are fused to generate the fused signal.

[0170] This application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0171] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the above-described method.

[0172] The above description of the embodiments is provided to enable those skilled in the art to understand and apply this application. Those skilled in the art will readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without inventive effort. Therefore, this application is not limited to the above embodiments, and any improvements and modifications made to this application based on the disclosure thereof should be within the scope of protection of this application.

Claims

1. A method of detecting partial discharge inside a transformer, characterized by, The method comprises the following steps: obtaining multi-modal sensing signals from a very high frequency sensor, an optical sensor and a fiber-optic Fabry-Perot ultrasonic sensor, wherein the very high frequency sensor is used to detect very high frequency electrical signals generated by internal partial discharge of a transformer, the optical sensor is used to detect optical signals generated by internal partial discharge of the transformer, and the fiber-optic Fabry-Perot ultrasonic sensor is used to detect acoustic signals generated by internal partial discharge of the transformer; processing and digitizing the multi-modal sensing signals by a multi-state quantity integrated perception SoC chip on a circuit board to generate a digitized signal stream; inputting the digitized signal stream into a state perception special-purpose edge computing NPU chip, processing the digitized signal stream by the NPU chip using a graph attention network algorithm, calculating attention coefficients between sensor nodes, fusing multi-modal signals based on the attention coefficients, and generating a fused signal; determining a type recognition result of internal partial discharge of the transformer according to the fused signal.

2. The method of claim 1, wherein, The method of obtaining multi-modal sensing signals from a very high frequency sensor, an optical sensor and a fiber-optic Fabry-Perot ultrasonic sensor comprises the following steps: the fiber-optic Fabry-Perot ultrasonic sensor converts acoustic signals into optical signals through a Fabry-Perot cavity, the Fabry-Perot cavity is composed of a vibrating diaphragm and a single-mode optical fiber end face, and the vibrating diaphragm modulates optical signals in response to acoustic wave vibration; the optical signals are transmitted to the circuit board through an optical fiber.

3. The method of claim 1, wherein, The method of processing and digitizing the multi-modal sensing signals by a multi-state quantity integrated perception SoC chip on a circuit board to generate a digitized signal stream comprises the following steps: the SoC chip receives the multi-modal sensing signals through a reconfigurable front-end interface, and the reconfigurable front-end interface is adapted to voltage-type, current-type and capacitance-type signals; the SoC chip performs special processing on the multi-modal sensing signals, including wideband amplification, transimpedance amplification and filtering; a unified format of the digitized signal stream is generated.

4. The method of claim 1, wherein, The method of processing and digitizing the multi-modal sensing signals by a multi-state quantity integrated perception SoC chip on a circuit board to generate a digitized signal stream comprises the following steps: the SoC chip adopts an event-driven adaptive working mechanism, and is in a dormant state when there is no signal, and is triggered to wake up by a signal pulse; after being woken up, the multi-modal sensing signals are processed and digitized in real time.

5. The method of claim 1, wherein, The method of inputting the digitized signal stream into a state perception special-purpose edge computing NPU chip, processing the digitized signal stream by the NPU chip using a graph attention network algorithm, calculating attention coefficients between sensor nodes, fusing multi-modal signals based on the attention coefficients, and generating a fused signal comprises the following steps: the NPU chip executes the graph attention network algorithm through a neural network hardware accelerator, and the accelerator is based on a RISC-V processor core and an NPU unit; the accelerator calculates the attention coefficients and signal fusion in parallel.

6. The method of claim 1, wherein, The method of inputting the digitized signal stream into a state perception special-purpose edge computing NPU chip, processing the digitized signal stream by the NPU chip using a graph attention network algorithm, calculating attention coefficients between sensor nodes, fusing multi-modal signals based on the attention coefficients, and generating a fused signal comprises the following steps: The NPU chip performs model pruning and INT8 quantization processing on the graph attention network algorithm, compressing the model size; The compressed algorithm is used to process the digitized signal stream.

7. The method of claim 1, wherein, The digitized signal stream is input into a state-aware special edge computing NPU chip, which uses a graph attention network algorithm to process the digitized signal stream, calculates attention coefficients between sensor nodes, and fuses multi-modal signals based on the attention coefficients to generate a fused signal, including: The NPU chip evaluates the signal-to-noise ratio of each sensor signal and dynamically calculates the weight according to the signal-to-noise ratio; Based on the weight, the attention coefficients are adjusted to fuse multi-modal signals to generate the fused signal.

8. A transformer internal partial discharge detection apparatus, characterized by, Including: The acquisition module acquires multi-modal sensing signals from a very high frequency sensor, an optical sensor, and a fiber-optic Fabry-Perot ultrasonic sensor, the very high frequency sensor is used to detect very high frequency electrical signals generated by internal partial discharge of a transformer, the optical sensor is used to detect optical signals generated by internal partial discharge of a transformer, and the fiber-optic Fabry-Perot ultrasonic sensor is used to detect acoustic signals generated by internal partial discharge of a transformer; The modulation module modulates and digitizes the multi-modal sensing signals by a multi-state quantity integrated perception SoC chip on a circuit board to generate a digitized signal stream; The fusion module inputs the digitized signal stream into a state-aware special edge computing NPU chip, which uses a graph attention network algorithm to process the digitized signal stream, calculates attention coefficients between sensor nodes, and fuses multi-modal signals based on the attention coefficients to generate a fused signal; The identification module determines the type identification result of the internal partial discharge of the transformer according to the fused signal.

9. The apparatus of claim 8, wherein, The acquisition module acquires multi-modal sensing signals from a very high frequency sensor, an optical sensor, and a fiber-optic Fabry-Perot ultrasonic sensor, including: The fiber-optic Fabry-Perot ultrasonic sensor converts acoustic signals into optical signals through a Fabry-Perot cavity, the Fabry-Perot cavity is composed of a vibrating diaphragm and a single-mode optical fiber end face, and the vibrating diaphragm modulates optical signals in response to acoustic wave vibration; The optical signals are transmitted to the circuit board through an optical fiber.

10. The apparatus of claim 8, wherein, The modulation module modulates and digitizes the multi-modal sensing signals by a multi-state quantity integrated perception SoC chip on a circuit board to generate a digitized signal stream, including: The SoC chip receives the multi-modal sensing signals through a reconfigurable front-end interface, and the reconfigurable front-end interface is adapted to voltage-type, current-type, and capacitance-type signals; The SoC chip performs special conditioning on the multi-modal sensing signals, including wideband amplification, transimpedance amplification, and filtering; A unified format of the digitized signal stream is generated.

11. The apparatus of claim 8, wherein, The modulation module modulates and digitizes the multi-modal sensing signals by a multi-state quantity integrated perception SoC chip on a circuit board to generate a digitized signal stream, including: The SoC chip adopts an event-driven adaptive working mechanism and is in a dormant state when there is no signal, and is triggered to wake up by a signal pulse; After waking up, the multi-modal sensing signals are real-time conditioned and digitized.

12. The apparatus of claim 8, wherein, The fusion module inputs the digitized signal stream into a state-aware special-purpose edge computing NPU chip, the NPU chip uses a graph attention network algorithm to process the digitized signal stream, calculates attention coefficients between sensor nodes, and fuses multi-modal signals based on the attention coefficients to generate a fused signal, including: The NPU chip executes the graph attention network algorithm through a neural network hardware accelerator, and the accelerator is based on a RISC-V processor core and an NPU unit in cooperation; The accelerator calculates the attention coefficients and signal fusion in parallel.

13. The apparatus of claim 8, wherein, The fusion module inputs the digitized signal stream into a state-aware special-purpose edge computing NPU chip, the NPU chip uses a graph attention network algorithm to process the digitized signal stream, calculates attention coefficients between sensor nodes, and fuses multi-modal signals based on the attention coefficients to generate a fused signal, including: The NPU chip performs model pruning and INT8 quantization processing on the graph attention network algorithm to compress the model size; The compressed algorithm is used to process the digitized signal stream.

14. The apparatus of claim 8, wherein, The fusion module inputs the digitized signal stream into a state-aware special-purpose edge computing NPU chip, the NPU chip uses a graph attention network algorithm to process the digitized signal stream, calculates attention coefficients between sensor nodes, and fuses multi-modal signals based on the attention coefficients to generate a fused signal, including: The NPU chip evaluates the signal-to-noise ratio of each sensor signal and dynamically calculates the weight according to the signal-to-noise ratio; Based on the weight, the attention coefficients are adjusted, and the multi-modal signals are fused to generate the fused signal.

15. An electronic device, comprising: A memory and a processor, the memory stores a computer program, and the processor executes the computer program to realize the method of claims 1-7.

16. A computer-readable storage medium, characterized in that, A computer program is stored thereon, and when the computer program is executed in a computer, the computer executes the method of claims 1-7.

Images