On-line diagnosis method for transformer insulation fault based on multi-source information fusion

Abstract

The present application belongs to the technical field of intelligent sensor, and relates to a transformer insulation fault online diagnosis method based on multi-source information fusion. The ultra-high frequency, ultrasonic wave and pulse current sensor are provided with a synchronous clock and an edge computing chip, a synchronous trigger signal is generated when any sensor captures a partial discharge pulse, the data of each sensor is forced to be truncated under the same time reference through hard wiring, and a multi-dimensional feature vector is generated on the edge side. A master station constructs a physical conduction delay compensation model according to the propagation speed difference of electromagnetic waves and ultrasonic waves in oil-paper insulation medium and the physical coordinates of the sensor, performs time-space alignment and feature-level fusion on the multi-dimensional feature vector, inputs the classifier and outputs a diagnosis result. The present application moves the time alignment to the physical acquisition end, eliminates the time deviation introduced by network transmission jitter, reduces the communication bandwidth occupation, and improves the time synchronization accuracy of multi-source signals.

Description

Technical Field

[0001] This invention belongs to the field of intelligent sensor technology and relates to an online diagnosis method for transformer insulation faults based on multi-source information fusion. Background Technology

[0002] Online diagnosis of transformer insulation faults typically employs multi-source monitoring using ultra-high frequency sensors, ultrasonic sensors, and pulse current sensors. In existing conventional solutions, each sensor independently acquires raw waveform data of partial discharge, and this raw data is aggregated and transmitted to a backend server via an industrial network. Upon receiving the multiple raw data streams, the backend server performs data-level fusion, i.e., performs cross-correlation calculations or direct stitching at the raw waveform level, and combines this with the physical coordinates of each sensor for fault location and diagnostic analysis. During this process, the sensors only act as data acquisition channels and do not perform time-referenced processing; time alignment relies entirely on the backend server for post-calibration based on the network reception timestamp of the data packets or the system log time.

[0003] The core technical problem with the existing solutions is that the centralized time alignment method in the background cannot eliminate the random time deviation introduced by network transmission jitter. Partial discharge signals are nanosecond-level transient pulses. When multiple raw waveform data are transmitted to the background server via an industrial network, the arrival time of each data stream is subject to uncontrollable jitter due to network bandwidth fluctuations and protocol scheduling mechanisms. The background server's time reference calibration based on the network reception timestamp causes misalignment of the multi-source heterogeneous signals that should be strictly aligned on the time axis. This time axis misalignment directly leads to the loss of physical synchronization for subsequent feature extraction and fusion algorithms, resulting in misjudgments in insulation fault diagnosis. Summary of the Invention

[0004] The purpose of this invention is to provide an online diagnostic method for transformer insulation faults based on multi-source information fusion, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for online diagnosis of transformer insulation faults based on multi-source information fusion is applied to ultra-high frequency sensors, ultrasonic sensors, and pulse current sensors deployed on the transformer body. The ultra-high frequency sensors, ultrasonic sensors, and pulse current sensors all have built-in synchronous clocks and edge computing chips. The method includes: when any one of the ultra-high frequency sensors, ultrasonic sensors, and pulse current sensors captures a partial discharge pulse exceeding a preset threshold, a synchronous trigger signal is generated, and the synchronous trigger signal is transmitted to the other sensors through hardwiring, forcing the ultra-high frequency sensors, ultrasonic sensors, and pulse current sensors to truncate data under the same time reference. The ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor perform fast Fourier transform and phase-resolved extraction on the truncated original signal at the edge side to generate a multi-dimensional feature vector containing discharge amplitude, phase, and spectral energy. The main station receives the multi-dimensional feature vectors uploaded by the UHF sensor, the ultrasonic sensor, and the pulse current sensor. Based on the difference between the propagation speed of UHF electromagnetic waves in transformer oil-paper insulation medium and the propagation speed of ultrasonic waves in insulating oil, and combined with the physical installation coordinates of the UHF sensor, the ultrasonic sensor, and the pulse current sensor, a physical conduction delay compensation model is constructed. The main station performs spatiotemporal alignment and feature-level fusion on the multidimensional feature vector based on the physical conduction delay compensation model, generates a fused diagnostic feature sequence, inputs it into the classifier, and outputs insulation fault diagnosis results.

[0006] Preferably, the step of generating a synchronization trigger signal and transmitting the synchronization trigger signal to the other sensors via hardwire to force the ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor to truncate data under the same time reference includes: when the first sensor captures the partial discharge pulse, the first sensor calls the built-in field-programmable gate array to generate the synchronization trigger signal with a fixed pulse width. The first sensor connects the synchronization trigger signal to the hardware interrupt pins of the second and third sensors via a shielded twisted pair cable; When the second sensor and the third sensor detect a level change in the hardware interrupt pin, they lock the current count value of the synchronization clock, and use the time corresponding to the count value as the zero point of the time base, and extract the original signal before and after the zero point for a preset time length and store it in a first-in-first-out buffer queue.

[0007] Preferably, the ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor perform fast Fourier transform and phase-resolved extraction on the truncated original signal at the edge side to generate a multi-dimensional feature vector containing discharge amplitude, phase, and spectral energy. This includes: the ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor performing windowing processing on the original signal, and performing fast Fourier transform on the windowed original signal to obtain a frequency domain sequence. The energy values ​​of multiple frequency band intervals in the frequency domain sequence are extracted as the spectral energy; Simultaneously, the ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor acquire the zero-crossing point of the transformer's power frequency voltage, calculate the phase angle of each pulse waveform in the original signal relative to the zero-crossing point as the phase, extract the peak value of each pulse waveform as the discharge amplitude, and combine the discharge amplitude, the phase, and the spectral energy into the multidimensional feature vector.

[0008] Preferably, the physical conduction delay compensation model is constructed based on the difference between the propagation speed of ultra-high frequency electromagnetic waves in the transformer oil-paper insulation medium and the propagation speed of ultrasonic waves in the insulating oil, combined with the physical installation coordinates of the ultra-high frequency sensor, the ultrasonic sensor and the pulse current sensor. This includes: pre-storing a three-dimensional spatial gridded distribution map of the transformer's internal insulating oil, insulating paperboard and metal components in the main station. The main station calculates the first path length of the ultra-high frequency electromagnetic wave from the preset power source location to the ultra-high frequency sensor and the second path length of the ultrasonic wave from the preset power source location to the ultrasonic sensor based on the ray tracing algorithm in the three-dimensional spatial gridded distribution map. The first theoretical delay time is obtained by dividing the first path length by the propagation speed of the ultra-high frequency electromagnetic wave, and the second theoretical delay time is obtained by dividing the second path length by the propagation speed of the ultrasonic wave. The physical conduction delay compensation model is constructed using the difference between the first theoretical delay time and the second theoretical delay time.

[0009] Preferably, the main station performs spatiotemporal alignment and feature-level fusion on the multidimensional feature vector based on the physical conduction delay compensation model to generate a fused diagnostic feature sequence, including: the main station extracts the first multidimensional feature vector uploaded by the ultra-high frequency sensor, the second multidimensional feature vector uploaded by the ultrasonic sensor, and the third multidimensional feature vector uploaded by the pulse current sensor; The main station shifts the first multidimensional feature vector on the time axis by the first theoretical delay time in the physical conduction delay compensation model, and shifts the second multidimensional feature vector on the time axis by the second theoretical delay time in the physical conduction delay compensation model, so that the first multidimensional feature vector, the second multidimensional feature vector and the third multidimensional feature vector are aligned on the time axis. The aligned first multidimensional feature vector, second multidimensional feature vector, and third multidimensional feature vector are concatenated along the feature dimension to generate the fused diagnostic feature sequence.

[0010] Preferably, the step of generating the fusion diagnostic feature sequence input to the classifier and outputting the insulation fault diagnosis result includes: the classifier includes a dimensionality reduction layer, a fully connected layer and an output layer connected in sequence; The dimensionality reduction layer receives the fused diagnostic feature sequence and maps the fused diagnostic feature sequence to a low-dimensional latent space through a preset projection matrix to obtain dimensionality reduction features; The fully connected layer performs a nonlinear transformation on the dimensionality-reduced features. The output layer contains multiple output nodes corresponding to various transformer insulation fault types. The output layer performs Softmax processing on the nonlinearly transformed features and outputs the probability values ​​corresponding to the multiple output nodes respectively. The fault type corresponding to the maximum probability value is taken as the insulation fault diagnosis result.

[0011] Preferably, the first sensor calls the built-in field-programmable gate array to generate the synchronization trigger signal with a fixed pulse width, including: when the field-programmable gate array detects the partial discharge pulse, it starts an internal counter to count, and when the count value of the internal counter reaches a preset pulse width threshold, it controls the level of the synchronization trigger signal to return to the initial state. During the transmission of the synchronization trigger signal, when the second sensor and the third sensor do not detect the level change of the hardware interrupt pin within multiple consecutive power frequency cycles, the second sensor and the third sensor send a frame loss flag to the master station. When the master station receives the frame loss flag, it discards the multidimensional feature vector uploaded by the first sensor in the current power frequency cycle.

[0012] Preferably, the ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor perform windowing processing on the original signal, including: the ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor acquiring the signal power spectral density of the original signal, and identifying the discrete white noise frequency band in the signal power spectral density; The ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor dynamically generate a windowing function with corresponding notch characteristics based on the center frequency of the discrete white noise frequency band. The ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor perform time-domain convolution operations between the windowing function and the original signal to filter out the interference waveforms corresponding to the discrete white noise frequency band, and output the filtered original signal to perform the fast Fourier transform.

[0013] Preferably, the main station pre-stores a three-dimensional spatial gridded distribution map of the insulating oil, insulating paperboard, and metal components inside the transformer, including: the main station obtains the real-time temperature and real-time moisture content of the insulating oil inside the transformer through a fiber optic grating sensor; The main station queries a preset oil medium characteristic mapping table based on the real-time temperature and the real-time moisture content to obtain the corrected propagation speed of the ultra-high frequency electromagnetic wave and the corrected propagation speed of the ultrasonic wave corresponding to the real-time temperature and the real-time moisture content. The main station replaces the propagation speed of the UHF electromagnetic wave with the corrected propagation speed of the UHF electromagnetic wave, and replaces the propagation speed of the ultrasonic wave with the corrected propagation speed of the ultrasonic wave, and performs the path length calculation based on the ray tracing algorithm.

[0014] Preferably, the extraction of the first multidimensional feature vector uploaded by the UHF sensor, the second multidimensional feature vector uploaded by the ultrasonic sensor, and the third multidimensional feature vector uploaded by the pulse current sensor by the master station includes: when the master station only receives the first multidimensional feature vector and the second multidimensional feature vector in the current alignment period, but does not receive the third multidimensional feature vector, the master station extracts the historical third multidimensional feature vector that is adjacent to the timestamp of the first multidimensional feature vector in the previous alignment period; The main station calculates the covariance matrix between the historical third multidimensional feature vector and the first multidimensional feature vector, uses the covariance matrix to perform a linear transformation on the historical third multidimensional feature vector, and uses the linearly transformed historical third multidimensional feature vector as the third multidimensional feature vector of the current alignment period to participate in the translation operation on the time axis.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, by configuring a synchronous clock and edge computing chip within the intelligent sensor, generates a synchronous trigger signal when any sensor captures a partial discharge pulse and transmits it to the remaining sensors via hardwired connection, forcing all sensors to truncate data under the same time reference. This mechanism moves the time alignment process forward to the physical acquisition end of the sensor, utilizing the deterministic transmission of hardwired connection to avoid random time deviations introduced by transmission jitter in industrial networks. Each sensor performs Fast Fourier Transform and phase-resolved extraction on the truncated data at the edge side, generating multi-dimensional feature vectors before uploading them. The main station constructs a physical conduction delay compensation model based on the propagation speed difference between UHF electromagnetic waves and ultrasonic waves in the oil-paper insulating medium and the physical coordinates of the sensors, performing spatiotemporal alignment and feature-level fusion of the multi-dimensional feature vectors. This invention eliminates the time reference offset of multi-source signals at the physical level, ensuring that the fused diagnostic feature sequence is established on the basis of physical time synchronization.

[0016] 2. This invention eliminates erroneous data interference caused by synchronization link anomalies by introducing an internal counter into a field-programmable gate array (FPGA) to generate a fixed-width synchronization trigger signal and sending a frame loss flag to trigger the master station to discard corresponding data when no interruption transition is detected. By identifying discrete white noise frequency bands based on the signal power spectral density of the original signal and dynamically generating a windowing function with notch characteristics, specific frequency band interference is filtered out before the Fast Fourier Transform at the edge. By introducing a fiber optic grating sensor to obtain the real-time temperature and moisture content of the insulating oil and querying the oil medium characteristic mapping table, the wave velocity in the physical conduction delay compensation model is corrected, reducing the impact of environmental parameter changes on wave velocity calculation. When eigenvectors are missing, the covariance matrix is ​​calculated using historical eigenvectors and a linear transformation is performed to replace the missing data, maintaining the continuity of the fusion diagnostic process. Attached Figure Description

[0017] Figure 1 This is a process for multi-sensor synchronous triggering and data truncation based on hardwiring; Figure 2 The process for generating fixed pulse width synchronization trigger signals and handling frame loss marking is as follows; Figure 3 This is a windowing process and multi-dimensional feature vector generation process with notch characteristics; Figure 4 The process for constructing a physical conduction delay compensation model based on ray tracing algorithm; Figure 5 This describes the spatiotemporal alignment and feature-level fusion process for multi-source feature vectors. Figure 6 This describes a process for missing data completion and insulation fault classification based on the covariance matrix. Detailed Implementation

[0018] The specific implementation methods of this technical solution are described in detail below. It should be understood that the specific implementation methods described herein are for illustration and explanation only and are not intended to limit this technical solution.

[0019] Please refer to the attached document. Figure 1In one embodiment, the online transformer insulation fault diagnosis method based on multi-source information fusion is applied to ultra-high frequency sensors, ultrasonic sensors, and pulse current sensors deployed on the transformer body. All three sensors have built-in synchronous clocks and edge computing chips. The synchronous clock provides a unified, continuously increasing time counting reference for all sensors. The counting source of the synchronous clock is a common clock signal, ensuring that the counting steps of all sensors are completely consistent and there is no cumulative deviation. The edge computing chip is directly connected to the signal acquisition channel of the sensor and is used to perform local signal processing operations at the sensor end. It does not rely on the computing resources of the main station and can execute the corresponding processing operations immediately after signal acquisition.

[0020] When any of the UHF sensor, ultrasonic sensor, and pulse current sensor detects a partial discharge pulse exceeding a preset threshold, a synchronization trigger signal is generated and transmitted to the other sensors via hardwiring. This forces the UHF sensor, ultrasonic sensor, and pulse current sensor to truncate data at the same time reference. The preset threshold is a judgment threshold pre-set based on the background noise amplitude under normal transformer operation. The amplitude of the preset threshold is higher than the maximum amplitude of the background noise under normal transformer operation, used to distinguish between normal background noise and partial discharge events, avoiding false triggering caused by fluctuations in background noise. The synchronization trigger signal is transmitted via hardwiring using a shielded twisted-pair cable structure with grounded shielding. This suppresses interference from the strong electromagnetic environment at the transformer site on the trigger signal transmission process, ensuring a fixed and negligible transmission delay and eliminating random time jitter caused by bandwidth fluctuations and protocol scheduling during industrial network transmission. The two ends of the hardwiring are connected to the trigger signal output port and trigger signal input port of the corresponding sensor, forming a point-to-point direct transmission link without any intermediate routing or forwarding devices, ensuring the uniqueness and determinism of the transmission path.

[0021] The data truncation operation is based on a unified time reference for all sensors. All sensors use the moment of synchronous trigger signal locking as the zero point of the time reference, truncating the raw signals for a preset time length before and after the zero point. This ensures that the raw signals truncated by all sensors correspond to the complete waveform of the same partial discharge event, without any misalignment on the time axis. The truncated raw signals are discrete-time digital signals obtained by analog-to-digital conversion from the sensor's analog acquisition channel, containing complete time-domain waveform information of the partial discharge pulse and background noise information.

[0022] Ultra-high frequency (UHF) sensors, ultrasonic sensors, and pulse current sensors perform Fast Fourier Transform (FFT) and phase-resolved extraction on the truncated raw signal at the edge, generating a multi-dimensional feature vector containing discharge amplitude, phase, and spectral energy. The edge computing chip completes all signal processing and feature extraction locally on the sensor, eliminating the need to upload the raw time-domain signal to the main station. Only the final generated multi-dimensional feature vector needs to be uploaded, significantly reducing the amount of data to be transmitted and lowering the bandwidth requirements of the industrial Ethernet. This also avoids distortion, packet loss, and timestamp offset issues that occur during long-distance transmission of the raw time-domain signal. The FFT converts the raw time-domain signal into a frequency-domain sequence, extracting the signal's frequency distribution characteristics. Phase-resolved extraction obtains the phase distribution information of the partial discharge pulse relative to the zero-crossing point of the transformer's power frequency voltage. Combined with discharge amplitude and spectral energy characteristics, a multi-dimensional feature vector is constructed, enabling a multi-dimensional representation of the partial discharge signal.

[0023] For the truncated original time-domain signal x(n), where n is the index of the discrete-time sequence and N is the total length of the discrete sequence, performing a Fast Fourier Transform on it yields the corresponding frequency-domain sequence X(k). The specific calculation process is defined by the following formula: Where k is the frequency index of the frequency domain sequence, corresponding to different frequency components, j is the imaginary unit, and X(k) is the complex value of the frequency domain corresponding to the k-th frequency point, which contains the amplitude and phase information of the frequency component.

[0024] The rated frequency of the transformer's power frequency voltage is f, and the zero-crossing time of the power frequency voltage is... The peak value of the partial discharge pulse corresponds to the time when The phase angle φ corresponding to this pulse is defined by the following formula: in, The value of φ is the time difference between the peak moment of the partial discharge pulse and the nearest zero-crossing point of the power frequency voltage. The value of φ ranges from [0, 2π), corresponding to the phase interval from 0° to 360° within the power frequency cycle, which can completely characterize the phase distribution position of the partial discharge pulse within the power frequency cycle.

[0025] For a frequency domain sequence X(k), it is pre-divided into m consecutive frequency bands. The frequency index range corresponding to the m-th frequency band is: Spectral energy within this frequency band Defined by the following formula: Where |X(k)| is the amplitude of the frequency domain sequence at the k-th frequency point, obtained by taking the modulus of the complex value X(k) in the frequency domain. The sum of squares of the amplitudes at all frequency points within this frequency band can characterize the energy distribution of the original signal within this frequency band.

[0026] The main station receives multi-dimensional feature vectors uploaded by UHF sensors, ultrasonic sensors, and pulse current sensors. Based on the difference between the propagation speed of UHF electromagnetic waves in the transformer's oil-paper insulation medium and the propagation speed of ultrasound in the insulating oil, and combined with the physical installation coordinates of the UHF, ultrasonic, and pulse current sensors, a physical conduction delay compensation model is constructed. The main station is an industrial control device deployed in the transformer's local monitoring room. It establishes a bidirectional communication connection with all sensors via industrial Ethernet, stably receiving multi-dimensional feature vectors uploaded by the sensors, and simultaneously sending configuration parameters and control commands to the sensors. The propagation speed of UHF electromagnetic waves in the transformer's oil-paper insulation medium differs by a fixed order of magnitude from the propagation speed of ultrasound in the insulating oil. This difference results in a fixed delay difference in the arrival time of the UHF signal and the ultrasonic signal generated by the same partial discharge power source at the corresponding sensor. Combining the physical installation coordinates of the sensors on the transformer body, this delay difference can be quantitatively modeled to obtain a physical conduction delay compensation model, used to eliminate the inherent time offset caused by signal propagation in the transformer's internal medium.

[0027] The first path length of the ultra-high frequency electromagnetic wave from the discharge source to the ultra-high frequency sensor is: The propagation speed of ultra-high frequency electromagnetic waves in oil-paper insulating medium is The corresponding first theoretical delay time Defined by the following formula: The second path length of the ultrasonic wave from the discharge source to the ultrasonic sensor is The speed of ultrasonic waves in insulating oil is The corresponding second theoretical delay time Defined by the following formula: in, This is the theoretical time required for the UHF signal to propagate from the discharge source location to the UHF sensor. The theoretical time required for the ultrasonic signal to propagate from the discharge source to the ultrasonic sensor is given. The difference between the two is the propagation delay difference between different types of signals, which is the core parameter of the physical conduction delay compensation model.

[0028] The main station performs spatiotemporal alignment and feature-level fusion on multidimensional feature vectors based on a physical conduction delay compensation model, generating a fused diagnostic feature sequence that is input into a classifier and outputs insulation fault diagnosis results. The spatiotemporal alignment operation adjusts multidimensional feature vectors from different sensors corresponding to the same partial discharge event to the same reference point on the time axis, eliminating the inherent time delay caused by signal propagation in the medium and ensuring that all feature vectors correspond to the same partial discharge event at the same moment. Feature-level fusion concatenates the aligned multi-source feature vectors along the feature dimensions to obtain a fused diagnostic feature sequence containing multi-source information. Compared to feature vectors from a single sensor, this sequence contains more comprehensive partial discharge signal features, effectively improving the discriminative power of insulation fault diagnosis.

[0029] The first multidimensional feature vector corresponding to the aligned UHF sensor is The feature dimension is The second multidimensional feature vector corresponding to the ultrasonic sensor is The feature dimension is The third multidimensional feature vector corresponding to the pulse current sensor is The feature dimension is The spliced ​​fusion diagnostic feature sequence F is defined by the following formula: Where [・] represents the concatenation operation of feature dimensions, and the total dimension of the fused diagnostic feature sequence F is... It fully contains all the characteristic information of three types of sensors: ultra-high frequency, ultrasonic, and pulse current.

[0030] The classifier is a deep learning model pre-trained using labeled transformer insulation fault sample data. Deployed in the main station's computing unit, it processes the input fused diagnostic feature sequence and outputs the corresponding insulation fault diagnosis results. The classifier's output layer contains multiple output nodes corresponding to various transformer insulation fault types. Softmax processing is applied to the nonlinearly transformed features, outputting the probability value corresponding to each output node. For the input vector z of the output layer, the vector contains C elements, corresponding to C transformer insulation fault types, and the probability value corresponding to the i-th output node. Defined by the following formula: in, Let i be the input value of the i-th node in the output layer. To determine the probability of occurrence of a corresponding insulation fault type, the sum of the probability values ​​of all output nodes is 1. The insulation fault type corresponding to the highest probability value is taken as the final insulation fault diagnosis result.

[0031] In this embodiment, the dimensional definition and physical meaning of the multidimensional feature vector are clearly defined in Table 1, ensuring that the master station can accurately identify the physical information corresponding to each dimension feature after receiving the feature vector, providing a unified feature benchmark for subsequent spatiotemporal alignment and feature fusion.

[0032] Table 1. Definition and Physical Meaning of Dimensions of Multidimensional Feature Vectors In Table 1, the multidimensional feature vectors corresponding to the three types of sensors all adopt a unified dimension division rule. The multidimensional feature vector generated by each sensor has a dimension of 30, and the total dimension of the fused diagnostic feature sequence after splicing is 90, which can completely characterize the core features of the partial discharge signals collected by the three types of sensors.

[0033] This embodiment achieves data truncation from multiple sensors on the same time base by setting a synchronous clock and an edge computing chip at the sensor end. When a partial discharge pulse is captured, a synchronous trigger signal is transmitted via hardwired transmission. This moves the time synchronization action forward to the signal acquisition end, eliminating random time deviations caused by network transmission jitter. By generating multi-dimensional feature vectors through signal processing and feature extraction at the edge, the amount of data transmitted is reduced, thus lowering network bandwidth requirements. The main station achieves spatiotemporal alignment and feature-level fusion of multi-source features by constructing a physical conduction delay compensation model. This ensures that the fused diagnostic feature sequence is based on strict physical time synchronization, providing a reliable feature foundation for online diagnosis of transformer insulation faults.

[0034] refer to Figure 2 In a preferred embodiment, when the first sensor detects a partial discharge pulse, it invokes its built-in field-programmable gate array (FPGA) to generate a synchronous trigger signal with a fixed pulse width. The first sensor is the one that first detects a partial discharge pulse exceeding a preset threshold among ultra-high frequency sensors, ultrasonic sensors, and pulse current sensors. The second and third sensors are the remaining two sensors. The FPGA is directly connected to the sensor's signal acquisition channel, allowing parallel acquisition of real-time signal data. When the signal amplitude exceeds the preset threshold, the FPGA can generate the trigger signal within a nanosecond response time, avoiding response delays caused by serial software processing. The fixed-pulse-width synchronous trigger signal is a level-transition signal, transitioning from an initial low level to a high level. The duration of the high level is the fixed pulse width, set by a preset pulse width threshold. This ensures that the hardware interrupt pins of the second and third sensors can stably detect the level transition, avoiding missed interrupt detection due to an excessively narrow pulse width or subsequent trigger signal conflicts due to an excessively wide pulse width.

[0035] When a partial discharge pulse is detected, the field-programmable gate array (FPGA) starts an internal counter. When the counter reaches a preset pulse width threshold, the level of the synchronization trigger signal is restored to its initial state. This internal hardware counter is synchronized with the sensor's built-in clock, ensuring identical counting frequency and step-by-step accuracy. When the amplitude of the detected partial discharge pulse exceeds the preset threshold, the counter increments from 0, and the synchronization trigger signal is set to active. Upon reaching the preset pulse width threshold, the signal is immediately restored to its initial inactive state, completing the generation of the fixed-pulse-width synchronization trigger signal. This hardware counting method ensures that the pulse width accuracy of the synchronization trigger signal is perfectly consistent with the accuracy of the synchronization clock, avoiding pulse width errors and jitter introduced by software counting.

[0036] The preset pulse width threshold is The counting period of the synchronous clock is The effective pulse width of the corresponding synchronization trigger signal Defined by the following formula: in, The preset pulse width threshold for the internal counter of the field-programmable gate array (FPGA) is a positive integer. The counting period of the synchronization clock is equal to the reciprocal of the synchronization clock frequency. The effective level duration of the synchronous trigger signal, i.e., the fixed pulse width.

[0037] The first sensor connects its synchronization trigger signal to the hardware interrupt pins of the second and third sensors via a shielded twisted-pair cable. These hardware interrupt pins are external interrupt input pins of the sensor's built-in edge computing chip or field-programmable gate array (FPGA). These pins are pre-configured for edge-triggered mode, triggering the interrupt service routine only when a level transition is detected, thus avoiding false triggering during the duration of the level transition and improving the speed and accuracy of the interrupt response. Impedance matching resistors are provided at both ends of the shielded twisted-pair cable to prevent signal reflection during transmission and ensure the waveform integrity of the synchronization trigger signal.

[0038] When the second and third sensors detect a level transition on the hardware interrupt pin, they lock the current synchronization clock count value and use the time corresponding to the count value as the zero point of the time base. They then extract the raw signal before and after the zero point for a preset time length and store it in a first-in-first-out buffer queue. When the hardware interrupt pin detects a level transition edge of the synchronization trigger signal, it immediately triggers the highest priority interrupt service routine. The first operation of the interrupt service routine is to lock the real-time count value of the current synchronization clock. The synchronization clock count value is a continuously increasing value, corresponding to absolute time. The time corresponding to the locked count value is the unified time base zero point for all sensors, ensuring that the time base of all sensors is completely consistent and there is no time deviation caused by software processing.

[0039] The captured raw signal includes a pre-trigger time length before the time reference zero point and a post-trigger time length after the zero point. The pre-trigger time length is used to preserve the background noise data before the partial discharge pulse occurs, providing a noise reference for subsequent signal processing; the post-trigger time length is used to preserve the complete partial discharge pulse waveform data, ensuring that the complete waveform of the partial discharge pulse is captured. The pre-trigger time length is... The subsequent trigger time length is The total length of the original signal captured Defined by the following formula: in, The length of the signal truncation before the zero point of the time base. The length of the signal truncation after the zero point of the time base. The total time length of the original signal captured after a single trigger corresponds to the length of the discrete signal sequence. The product of the sensor sampling frequency.

[0040] The captured raw signal is stored in a first-in-first-out (FIFO) buffer queue. The storage space of the buffer queue is pre-allocated in the sensor's memory to temporarily store the truncated raw signal, awaiting subsequent signal processing and feature extraction by the edge computing chip. The FIFO storage mechanism ensures that the signal data is processed in the order of acquisition time, avoiding data disorder. It can also buffer raw signals triggered multiple times consecutively, preventing data loss when partial discharge events occur consecutively.

[0041] During the transmission of the synchronization trigger signal, if the second and third sensors fail to detect a level transition on the hardware interrupt pin within multiple consecutive power frequency cycles, they send a frame loss flag to the master station. Upon receiving the frame loss flag, the master station discards the multi-dimensional feature vector uploaded by the first sensor within the current power frequency cycle. The transmission link of the synchronization trigger signal is hardwired. If an open circuit or short circuit occurs in the transmission link, the field-programmable gate array of the first sensor malfunctions, or the amplitude of the partial discharge pulse does not reach the detection threshold of the other sensors, the hardware interrupt pins of the second and third sensors cannot detect the corresponding level transition, and the time base locking and data truncation cannot be completed. In this case, if the second and third sensors fail to detect a level transition within multiple consecutive power frequency cycles, they generate a frame loss flag. The frame loss flag contains the sensor's unique number, the corresponding power frequency cycle number, and timestamp information, and is uploaded to the master station via the industrial Ethernet. After receiving the frame loss flag, the master station determines that the synchronization of multi-source data in the current power frequency cycle has failed. It discards the multi-dimensional feature vector uploaded by the first sensor in the current power frequency cycle to prevent asynchronous feature data from entering the subsequent fusion and diagnosis process, which could lead to misjudgment of the diagnosis results.

[0042] In this embodiment, the matching relationship between the pulse width parameter of the synchronous trigger signal and the interrupt response is clearly defined in Table 2. The corresponding preset pulse width threshold can be selected according to the electromagnetic interference environment at the transformer site and the hardware interrupt characteristics of the sensor to ensure stable transmission and reliable detection of the synchronous trigger signal.

[0043] Table 2 Matching Table of Synchronous Trigger Signal Pulse Width Parameters and Interrupt Response In Table 2, as the preset pulse width threshold increases, the high-level duration of the synchronous trigger signal increases, and the corresponding minimum detection level amplitude decreases. This can adapt to field environments with stronger electromagnetic interference, avoid false triggering caused by interference signals, and at the same time ensure that the interrupt pin can stably detect the level transition of the synchronous trigger signal.

[0044] This embodiment generates a fixed-pulse-width synchronous trigger signal using a hardware counter in a field-programmable gate array (FPGA) to ensure the accuracy and stability of the trigger signal; it locks the synchronous clock count value using the edge-triggered mode of a hardware interrupt pin to achieve strict uniformity of the time base for multiple sensors; it temporarily stores truncated original signals using a first-in-first-out (FIFO) buffer queue to ensure the sequentiality and continuity of signal processing; and it eliminates interference from erroneous data caused by synchronization link anomalies in the diagnostic process by generating frame loss markers and discarding corresponding data, further improving the reliability of multi-source data synchronization.

[0045] refer to Figure 3In another preferred embodiment, the ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor perform windowing processing on the original signal, and then perform a Fast Fourier Transform (FFT) on the windowed original signal to obtain a frequency domain sequence. Windowing is used to suppress spectral leakage caused by signal truncation during the FFT process. Simultaneously, the windowing function in this embodiment possesses notch characteristics, which can filter out discrete white noise frequency band interference in the original signal during time-domain processing, eliminating the need for additional digital filtering algorithms, reducing the computational load on the edge computing chip, and improving signal processing efficiency.

[0046] Ultra-high frequency (UHF) sensors, ultrasonic sensors, and pulse current sensors acquire the signal power spectral density of the original signal and identify the discrete white noise frequency band within the signal power spectral density. The signal power spectral density describes the distribution of signal power at different frequency points and is obtained by performing a discrete Fourier transform on the autocorrelation function of the original signal. The original time-domain signal sequence is x(n), with a sequence length of N, and the corresponding autocorrelation function... Defined by the following formula: Where m is the delay amount. The autocorrelation function of the original signal is used to describe the correlation of the signal under different delays and can reflect the periodicity and frequency distribution characteristics of the signal.

[0047] Based on autocorrelation function The corresponding signal power spectral density Defined by the following formula: in, The signal power spectral density at the k-th frequency point can accurately characterize the distribution of the original signal power at different frequency points.

[0048] The discrete white noise band is a continuous frequency range where the power spectral density amplitude exceeds a preset noise threshold. Signals within this range consist of fixed-frequency narrowband interference at the transformer site, including carrier communication signals, inverter switching frequency interference, and radiated interference from surrounding power equipment. This type of interference is discrete narrowband noise and can interfere with the feature extraction of partial discharge signals. By identifying the center frequency and bandwidth of the discrete white noise band, a corresponding windowing function with notch characteristics can be generated to filter out narrowband interference.

[0049] Ultra-high frequency (UHF) sensors, ultrasonic sensors, and pulsed current sensors dynamically generate windowing functions with corresponding notch characteristics based on the center frequency of the discrete white noise frequency band. The base window function is a Hanning window, Hamming window, or Blackman window, which possesses good spectral leakage suppression capabilities. A notch point is set at the center frequency position of the discrete white noise frequency band in the frequency domain response of the base window function. The attenuation depth and bandwidth of the notch point are adaptively adjusted according to the power and bandwidth of the discrete white noise frequency band, resulting in a windowing function with notch characteristics. This windowing function can suppress spectral leakage while filtering out specific narrowband interference without requiring additional filtering operations.

[0050] Ultra-high frequency (UHF) sensors, ultrasonic sensors, and pulse current sensors perform time-domain convolution operations between a windowed function and the original signal to filter out interference waveforms corresponding to discrete white noise frequency bands. The filtered original signal is then output for Fast Fourier Transform (FFT). Time-domain convolution is equivalent to frequency-domain multiplication, allowing for the filtering of interference in specific frequency bands within the time domain. The operation is simple and suitable for implementation in edge computing chips. The original signal is x(n), the time-domain sequence of the windowed function with notch characteristics is w(n), and the sequence length is M. The filtered signal y(n) after convolution is defined by the following formula: Where w(k) is the time-domain sequence of a windowed function with notch characteristics, and y(n) is the filtered time-domain signal sequence used for subsequent fast Fourier transform and feature extraction.

[0051] After performing a Fast Fourier Transform on the filtered original signal to obtain a frequency domain sequence, the energy values ​​of multiple frequency bands in the frequency domain sequence are extracted as spectral energy. For the signal frequency characteristics of different types of sensors, corresponding characteristic frequency bands are pre-defined. Different frequency bands correspond to the characteristic frequency bands of different types of partial discharge signals. Each characteristic frequency band is divided into multiple continuous sub-frequency bands, and the spectral energy within each sub-frequency band is extracted as the spectral energy feature in the multi-dimensional feature vector.

[0052] Simultaneously, ultra-high frequency sensors, ultrasonic sensors, and pulse current sensors acquire the zero-crossing points of the transformer's power frequency voltage. The phase angle of each pulse waveform in the original signal relative to the zero-crossing point is calculated as the phase, and the peak value of each pulse waveform is extracted as the discharge amplitude. The sensors acquire the real-time waveform of the power frequency voltage through the transformer's voltage transformer, and identify the zero-crossing points of the power frequency voltage in real time. The zero-crossing points include the rising edge zero-crossing point of the power frequency voltage transitioning from negative to positive and the falling edge zero-crossing point of the power frequency voltage transitioning from positive to negative. The time interval between two adjacent rising edge zero-crossing points constitutes a complete power frequency cycle. For each partial discharge pulse waveform in the original signal, the peak value is identified using a peak detection algorithm, and the time difference between this moment and the nearest power frequency voltage zero-crossing point is calculated. The corresponding phase angle is obtained using the phase angle calculation formula and used as the phase feature. The peak amplitude of the pulse waveform is extracted as the discharge amplitude feature.

[0053] Discharge amplitude, phase, and spectral energy are combined into a multi-dimensional feature vector. All discharge amplitude features, phase features, and spectral energy features within the same power frequency cycle are combined in a preset fixed order to generate a multi-dimensional feature vector with fixed dimensions. This ensures that the main station can parse and process the feature vector in a unified format, avoiding parsing errors caused by inconsistent feature dimensions.

[0054] In this embodiment, the windowing function parameters and feature frequency band division rules corresponding to different types of sensors are specified in Table 3. For the signal frequency characteristics of different types of sensors, the corresponding windowing function parameters and frequency band division methods are set to ensure the accuracy and relevance of feature extraction.

[0055] Table 3. Windowing function parameters and characteristic frequency band division for different sensors In Table 3, for the high-frequency signal characteristics of UHF sensors, the Blackman window with greater sidelobe attenuation is used as the basic window function, which can better suppress spectral leakage. For the mid-to-low frequency signal characteristics of ultrasonic sensors and pulse current sensors, the Hamming window and Hanning window are used as the basic window functions, respectively. While suppressing spectral leakage, they have a narrower main lobe width and improve frequency resolution.

[0056] This embodiment identifies discrete white noise frequency bands by calculating the power spectral density of the original signal, dynamically generates a windowing function with corresponding notch characteristics, filters out narrowband interference through time-domain convolution operations, and suppresses spectral leakage during the fast Fourier transform process. By dividing multiple characteristic frequency bands according to different sensor characteristics to extract spectral energy features, and combining discharge amplitude and phase features, a multi-dimensional feature vector is generated. Signal preprocessing and feature extraction are completed at the edge side, reducing the computational burden on the main station, while improving the anti-interference capability and fault differentiation of the features.

[0057] refer to Figure 4 In another preferred embodiment, the main station pre-stores a three-dimensional spatial gridded distribution map of the transformer's internal insulating oil, insulating paperboard, and metal components. This three-dimensional spatial gridded distribution map is constructed based on the transformer's design drawings and actual installation parameters. It divides the transformer's internal enclosed space into multiple uniform three-dimensional cubic grid units, each labeled with its corresponding medium type. Medium types include insulating oil, insulating paperboard, and metal components, etc. Different medium types correspond to different electromagnetic wave and ultrasonic wave propagation characteristics. Metal components correspond to total internal reflection propagation characteristics, while insulating oil and insulating paperboard correspond to different propagation velocities and refractive indices. The three-dimensional spatial gridded distribution map also labels the physical installation coordinates of the UHF sensor, ultrasonic sensor, and pulse current sensor on the transformer body, providing a spatial reference for subsequent propagation path calculations.

[0058] The main station uses a ray tracing algorithm to calculate the first path length of UHF electromagnetic waves from a preset discharge source location to the UHF sensor, and the second path length of ultrasonic waves from a preset discharge source location to the ultrasonic sensor, based on a 3D spatial gridded distribution map. The ray tracing algorithm simulates the propagation paths of electromagnetic waves and ultrasonic waves in different media inside the transformer. Starting from the preset discharge source location, multiple discrete rays are emitted in various directions in 3D space. When a ray reaches the interface between different media, the refraction and reflection directions of the ray are calculated according to the laws of refraction and reflection until the ray reaches the sensor installation location. The total propagation path length from the discharge source location to the sensor installation location is recorded, which is the corresponding path length. For UHF electromagnetic waves, total internal reflection occurs on the surface of metal components, and refraction and reflection occur at the interface between insulating oil and insulating paperboard, allowing propagation within the insulating oil and insulating paperboard medium. For ultrasonic waves, propagation is limited to the insulating oil medium, and total internal reflection occurs on the surface of metal components and insulating paperboard. Therefore, the propagation paths of the two differ, and their corresponding path lengths also differ.

[0059] The first theoretical delay time is obtained by dividing the first path length by the propagation speed of ultra-high frequency electromagnetic waves, and the second theoretical delay time is obtained by dividing the second path length by the propagation speed of ultrasonic waves. A physical conduction delay compensation model is constructed using the difference between the first and second theoretical delay times. This model quantifies the time delay difference between different types of signals generated by the same partial discharge source reaching their corresponding sensors, providing a compensation benchmark for the spatiotemporal alignment of multi-source feature vectors.

[0060] The main station acquires the real-time temperature and moisture content of the insulating oil inside the transformer using fiber Bragg grating (FBG) sensors. These FBG sensors are deployed at multiple locations inside the transformer tank, enabling real-time data collection of the insulating oil's temperature and moisture content. The measurement process is unaffected by the strong electromagnetic environment at the transformer site, ensuring accurate acquisition of the insulating oil's real-time status parameters. The propagation speed of the insulating oil inside the transformer is significantly affected by temperature and moisture content. Changes in temperature significantly alter the propagation speed of ultrasonic waves and ultra-high frequency (UHF) electromagnetic waves. Conversely, increased moisture content alters the dielectric constant of the insulating oil, leading to changes in the UHF electromagnetic wave propagation speed. Therefore, corrections to the propagation speed based on the real-time status parameters of the insulating oil are necessary.

[0061] The main station queries a pre-defined oil dielectric property mapping table based on real-time temperature and moisture content to obtain the corrected propagation speeds of UHF electromagnetic waves and ultrasound waves corresponding to the real-time temperature and moisture content. The pre-defined oil dielectric property mapping table is a table that, obtained through standard testing, shows the correspondence between the propagation speeds of UHF electromagnetic waves and ultrasound waves in insulating oil under different combinations of temperature and moisture content. The main station uses the obtained real-time temperature and moisture content as query indexes to find the corresponding corrected propagation speeds in the oil dielectric property mapping table.

[0062] The main station replaces the propagation speed of UHF electromagnetic waves with a corrected propagation speed, and replaces the propagation speed of ultrasonic waves with a corrected propagation speed, performing path length calculations based on a ray tracing algorithm. By calculating the theoretical delay time using the corrected propagation speeds, the impact of changes in insulating oil state parameters on propagation delay calculations can be effectively reduced, improving the calculation accuracy of the physical conduction delay compensation model.

[0063] refer to Figure 5 The main station extracts the first multi-dimensional feature vector uploaded by the UHF sensor, the second multi-dimensional feature vector uploaded by the ultrasonic sensor, and the third multi-dimensional feature vector uploaded by the pulse current sensor. The main station receives all multi-dimensional feature vectors uploaded by the sensors. Each feature vector carries a corresponding timestamp, which is the absolute time corresponding to the synchronization clock count value locked by the sensor. The main station divides the feature vectors corresponding to the same partial discharge event into the same alignment period according to the timestamps, ensuring that the feature vectors participating in the alignment correspond to the same partial discharge event.

[0064] refer to Figure 6When the master station receives only the first and second multidimensional feature vectors within the current alignment period, but not the third multidimensional feature vector, it extracts the historical third multidimensional feature vector from the previous alignment period that is adjacent to the timestamp of the first multidimensional feature vector. When a single sensor experiences a communication failure, data packet loss, or abnormal acquisition, the master station cannot receive the corresponding multidimensional feature vector within the current alignment period. In this case, it extracts the historical feature vector from the previous complete alignment period as the basis for data completion, avoiding interruption of the entire diagnostic process due to missing data from a single sensor.

[0065] The main station calculates the covariance matrix between the historical third-dimensional eigenvector and the first-dimensional eigenvector. Using this covariance matrix, a linear transformation is performed on the historical third-dimensional eigenvector. This transformed historical third-dimensional eigenvector is then used as the third-dimensional eigenvector for the current alignment period in the time-axis translation operation. The covariance matrix describes the linear correlation between two eigenvectors. Based on the covariance matrix, a linear transformation matrix adapted to the current alignment period can be constructed. This matrix is ​​then used to linearly transform the historical eigenvector, obtaining a complete eigenvector that matches the eigenvector of the current alignment period, thus maintaining the continuity of the diagnostic process.

[0066] The third historical multidimensional feature vector is The feature dimension is d, the number of samples is n, and the current first multi-dimensional feature vector is Given a feature dimension of d and a sample size of n, the covariance matrix Cov of the two is defined by the following formula: in, For the i-th sample of the third historical multidimensional feature vector, The mean vector of the third historical multidimensional eigenvector. For the i-th sample of the current first multidimensional feature vector, This is the mean vector of the current first multidimensional feature vector. This is the transpose operation of the matrix, where Cov is the covariance matrix between two eigenvectors, with dimensions d×d.

[0067] The linear transformation matrix W is constructed based on the covariance matrix, the bias vector is b, and the transformed third multidimensional eigenvector of the current alignment period is... Defined by the following formula: Where W is the linear transformation matrix obtained based on the eigenvalue decomposition of the covariance matrix, and b is the bias vector. The third multidimensional feature vector of the current alignment period after completion can be directly involved in subsequent spatiotemporal alignment and feature fusion operations.

[0068] The main station shifts the first multidimensional feature vector along the time axis by the first theoretical delay time in the physical conduction delay compensation model, and shifts the second multidimensional feature vector along the time axis by the second theoretical delay time in the physical conduction delay compensation model, so that the first, second, and third multidimensional feature vectors are aligned along the time axis. The pulse current sensor is installed on the transformer's grounding wire. The pulse current generated by the partial discharge pulse propagates directly to the pulse current sensor through the grounding wire, and the propagation delay is negligible. Therefore, using the time reference of the third multidimensional feature vector as a reference, the first and second multidimensional feature vectors are shifted along the time axis, so that the positions of the partial discharge pulses corresponding to the three feature vectors are completely aligned along the time axis, eliminating the inherent time delay caused by signal propagation in the medium.

[0069] The aligned first, second, and third multidimensional feature vectors are concatenated along their feature dimensions to generate a fused diagnostic feature sequence. This concatenated fused diagnostic feature sequence fully encompasses the feature information from the three types of sensors, comprehensively characterizing the properties of partial discharge events and providing a comprehensive feature foundation for subsequent fault diagnosis.

[0070] The generated fused diagnostic feature sequence is input into the classifier, which outputs the insulation fault diagnosis result. The classifier consists of a dimensionality reduction layer, a fully connected layer, and an output layer connected in sequence. The dimensionality reduction layer receives the fused diagnostic feature sequence and maps it to a low-dimensional latent space using a pre-defined projection matrix to obtain dimensionality-reduced features. The projection matrix is ​​pre-obtained using principal component analysis, which reduces the dimensionality of the features while preserving the main information of the fused diagnostic feature sequence, thereby reducing the computational load of subsequent layers and improving diagnostic efficiency. The fully connected layer performs a nonlinear transformation on the dimensionality-reduced features. The fully connected layer can be one or more layers, each containing multiple neurons. After a linear transformation of the dimensionality-reduced features using a weight matrix and a bias vector, a nonlinear transformation is performed using a nonlinear activation function to enhance the model's feature representation capability. The output layer contains multiple output nodes corresponding to various transformer insulation fault types. The output layer performs Softmax processing on the nonlinearly transformed features and outputs the probability values ​​corresponding to the multiple output nodes. The fault type corresponding to the highest probability value is taken as the insulation fault diagnosis result. Transformer insulation fault types include surface discharge of insulating paperboard, oil bubble discharge, metal tip discharge, floating potential discharge, and insulation oil deterioration. Each fault type corresponds to an output node of the output layer.

[0071] In this embodiment, the propagation velocity correction parameters of the insulating oil medium under different combinations of temperature and moisture content are specified in Table 4. This table is an example of a preset oil medium characteristic mapping table. The main station obtains the wave velocity parameters of the insulating oil under real-time conditions by querying this table, thereby improving the calculation accuracy of the physical conduction delay compensation model.

[0072] Table 4 Correction table for the propagation velocity of insulating oil medium under different temperatures and moisture contents. Table 4 shows that as the temperature and moisture content of the insulating oil increase, the propagation speed of both ultra-high frequency electromagnetic waves and ultrasonic waves in the insulating oil decreases. This table allows for accurate lookup of the corrected propagation speed under different conditions, eliminating the impact of changes in the state of the insulating oil on the delay calculation.

[0073] This embodiment accurately calculates the signal propagation path length and theoretical delay time using a three-dimensional spatial gridded distribution map inside the transformer and a ray tracing algorithm, constructing a physical conduction delay compensation model. Real-time temperature and moisture content of the insulating oil are obtained using fiber optic grating sensors, and the propagation speed is corrected in real time, improving the accuracy of delay compensation. A linear transformation of historical feature vectors is performed using the covariance matrix to complete missing feature vectors, maintaining the continuity of the diagnostic process. A classifier containing a dimensionality reduction layer, a fully connected layer, and an output layer processes the fused diagnostic feature sequence and outputs insulation fault diagnosis results, achieving accurate online diagnosis of transformer insulation faults.

Claims

1. A method for online diagnosis of transformer insulation faults based on multi-source information fusion, characterized in that, The method applies to ultra-high frequency sensors, ultrasonic sensors, and pulse current sensors deployed on the transformer body. Each of the ultra-high frequency sensor, ultrasonic sensor, and pulse current sensor has a built-in synchronous clock and edge computing chip. The method includes: when any one of the ultra-high frequency sensor, ultrasonic sensor, and pulse current sensor captures a partial discharge pulse exceeding a preset threshold, generating a synchronous trigger signal, and transmitting the synchronous trigger signal to the other sensors via hardwired connection, forcing the ultra-high frequency sensor, ultrasonic sensor, and pulse current sensor to truncate data under the same time reference; The ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor perform fast Fourier transform and phase-resolved extraction on the truncated original signal at the edge side to generate a multi-dimensional feature vector containing discharge amplitude, phase, and spectral energy. The main station receives the multi-dimensional feature vectors uploaded by the UHF sensor, the ultrasonic sensor, and the pulse current sensor. Based on the difference between the propagation speed of UHF electromagnetic waves in transformer oil-paper insulation medium and the propagation speed of ultrasonic waves in insulating oil, and combined with the physical installation coordinates of the UHF sensor, the ultrasonic sensor, and the pulse current sensor, a physical conduction delay compensation model is constructed. The main station performs spatiotemporal alignment and feature-level fusion on the multidimensional feature vector based on the physical conduction delay compensation model, generates a fused diagnostic feature sequence, inputs it into the classifier, and outputs insulation fault diagnosis results.

2. The online diagnosis method for transformer insulation faults based on multi-source information fusion according to claim 1, characterized in that, The step of generating a synchronization trigger signal and transmitting the synchronization trigger signal to the other sensors via hardwired connection, forcing the ultra-high frequency sensor, the ultrasonic sensor and the pulse current sensor to truncate data under the same time reference, includes: when the first sensor captures the partial discharge pulse, the first sensor calls the built-in field programmable gate array to generate the synchronization trigger signal with a fixed pulse width; The first sensor connects the synchronization trigger signal to the hardware interrupt pins of the second and third sensors via a shielded twisted pair cable; When the second sensor and the third sensor detect a level change in the hardware interrupt pin, they lock the current count value of the synchronization clock, and use the time corresponding to the count value as the zero point of the time base, and extract the original signal before and after the zero point for a preset time length and store it in a first-in-first-out buffer queue.

3. The online diagnosis method for transformer insulation faults based on multi-source information fusion according to claim 1, characterized in that, The ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor perform fast Fourier transform and phase-resolved extraction on the truncated original signal at the edge side to generate a multi-dimensional feature vector containing discharge amplitude, phase, and spectral energy. This includes: the ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor performing windowing processing on the original signal, and performing fast Fourier transform on the windowed original signal to obtain a frequency domain sequence. The energy values ​​of multiple frequency band intervals in the frequency domain sequence are extracted as the spectral energy; Simultaneously, the ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor acquire the zero-crossing point of the transformer's power frequency voltage, calculate the phase angle of each pulse waveform in the original signal relative to the zero-crossing point as the phase, extract the peak value of each pulse waveform as the discharge amplitude, and combine the discharge amplitude, the phase, and the spectral energy into the multidimensional feature vector.

4. The online diagnosis method for transformer insulation faults based on multi-source information fusion according to claim 1, characterized in that, Based on the difference between the propagation speed of ultra-high frequency electromagnetic waves in the transformer oil-paper insulation medium and the propagation speed of ultrasonic waves in the insulating oil, and combined with the physical installation coordinates of the ultra-high frequency sensor, the ultrasonic sensor and the pulse current sensor, a physical conduction delay compensation model is constructed, including: a three-dimensional spatial gridded distribution map of the transformer's internal insulating oil, insulating paperboard and metal components is pre-stored in the main station. The main station calculates the first path length of the ultra-high frequency electromagnetic wave from the preset power source location to the ultra-high frequency sensor and the second path length of the ultrasonic wave from the preset power source location to the ultrasonic sensor based on the ray tracing algorithm in the three-dimensional spatial gridded distribution map. The first theoretical delay time is obtained by dividing the first path length by the propagation speed of the ultra-high frequency electromagnetic wave, and the second theoretical delay time is obtained by dividing the second path length by the propagation speed of the ultrasonic wave. The physical conduction delay compensation model is constructed using the difference between the first theoretical delay time and the second theoretical delay time.

5. The online diagnosis method for transformer insulation faults based on multi-source information fusion according to claim 1, characterized in that, The main station performs spatiotemporal alignment and feature-level fusion on the multidimensional feature vector based on the physical conduction delay compensation model to generate a fused diagnostic feature sequence, including: the main station extracts the first multidimensional feature vector uploaded by the ultra-high frequency sensor, the second multidimensional feature vector uploaded by the ultrasonic sensor, and the third multidimensional feature vector uploaded by the pulse current sensor. The main station shifts the first multidimensional feature vector on the time axis by the first theoretical delay time in the physical conduction delay compensation model, and shifts the second multidimensional feature vector on the time axis by the second theoretical delay time in the physical conduction delay compensation model, so that the first multidimensional feature vector, the second multidimensional feature vector and the third multidimensional feature vector are aligned on the time axis. The aligned first multidimensional feature vector, second multidimensional feature vector, and third multidimensional feature vector are concatenated along the feature dimension to generate the fused diagnostic feature sequence.

6. The online diagnosis method for transformer insulation faults based on multi-source information fusion according to claim 1, characterized in that, The process of generating fusion diagnostic feature sequences and inputting them into a classifier to output insulation fault diagnostic results includes: the classifier comprising a dimension reduction layer, a fully connected layer, and an output layer connected in sequence. The dimensionality reduction layer receives the fused diagnostic feature sequence and maps the fused diagnostic feature sequence to a low-dimensional latent space through a preset projection matrix to obtain dimensionality reduction features; The fully connected layer performs a nonlinear transformation on the dimensionality-reduced features. The output layer contains multiple output nodes corresponding to various transformer insulation fault types. The output layer performs Softmax processing on the nonlinearly transformed features and outputs the probability values ​​corresponding to the multiple output nodes respectively. The fault type corresponding to the maximum probability value is taken as the insulation fault diagnosis result.

7. The online diagnosis method for transformer insulation faults based on multi-source information fusion according to claim 2, characterized in that, The first sensor calls the built-in field-programmable gate array to generate the synchronization trigger signal with a fixed pulse width, including: when the field-programmable gate array detects the partial discharge pulse, it starts an internal counter to count; when the count value of the internal counter reaches a preset pulse width threshold, it controls the level of the synchronization trigger signal to return to the initial state. During the transmission of the synchronization trigger signal, when the second sensor and the third sensor do not detect the level change of the hardware interrupt pin within multiple consecutive power frequency cycles, the second sensor and the third sensor send a frame loss flag to the master station. When the master station receives the frame loss flag, it discards the multidimensional feature vector uploaded by the first sensor in the current power frequency cycle.

8. The online diagnosis method for transformer insulation faults based on multi-source information fusion according to claim 3, characterized in that, The ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor perform windowing processing on the original signal, including: the ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor acquiring the signal power spectral density of the original signal and identifying the discrete white noise frequency band in the signal power spectral density; The ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor dynamically generate a windowing function with corresponding notch characteristics based on the center frequency of the discrete white noise frequency band. The ultra-high frequency sensor, the ultrasonic sensor, and the pulse current sensor perform time-domain convolution operations between the windowing function and the original signal to filter out the interference waveforms corresponding to the discrete white noise frequency band, and output the filtered original signal to perform the fast Fourier transform.

9. The online diagnosis method for transformer insulation faults based on multi-source information fusion according to claim 4, characterized in that, The master station pre-stores a three-dimensional spatial gridded distribution map of the transformer's internal insulating oil, insulating paperboard, and metal components, including: the master station acquires the real-time temperature and real-time moisture content of the transformer's internal insulating oil through a fiber optic grating sensor; The main station queries a preset oil medium characteristic mapping table based on the real-time temperature and the real-time moisture content to obtain the corrected propagation speed of the ultra-high frequency electromagnetic wave and the corrected propagation speed of the ultrasonic wave corresponding to the real-time temperature and the real-time moisture content. The main station replaces the propagation speed of the UHF electromagnetic wave with the corrected propagation speed of the UHF electromagnetic wave, and replaces the propagation speed of the ultrasonic wave with the corrected propagation speed of the ultrasonic wave, and performs the path length calculation based on the ray tracing algorithm.

10. The online diagnosis method for transformer insulation faults based on multi-source information fusion according to claim 5, characterized in that, The main station extracts the first multi-dimensional feature vector uploaded by the UHF sensor, the second multi-dimensional feature vector uploaded by the ultrasonic sensor, and the third multi-dimensional feature vector uploaded by the pulse current sensor, including: when the main station only receives the first multi-dimensional feature vector and the second multi-dimensional feature vector in the current alignment period, but does not receive the third multi-dimensional feature vector, the main station extracts the historical third multi-dimensional feature vector that is adjacent to the timestamp of the first multi-dimensional feature vector in the previous alignment period; The main station calculates the covariance matrix between the historical third multidimensional feature vector and the first multidimensional feature vector, uses the covariance matrix to perform a linear transformation on the historical third multidimensional feature vector, and uses the linearly transformed historical third multidimensional feature vector as the third multidimensional feature vector of the current alignment period to participate in the translation operation on the time axis.

Images