A fault diagnosis method for oil-immersed transformer

Abstract

The present application relates to the technical field of power equipment fault diagnosis, and discloses a kind of oil-immersed transformer fault diagnosis method, including synchronous acquisition transformer's operating phase voltage discrete sequence, load current discrete sequence, top oil temperature discrete sequence and box wall acceleration sequence;In light load steady state condition, identify transfer function parameter matrix, reconstruct core vibration reference component in dynamic condition and difference extraction winding vibration component;Relative viscosity coefficient is calculated using top oil temperature to dampen winding vibration component;In load transient time window, extract excitation force characteristics, macroscopic displacement characteristics and high-frequency distortion rate characteristics, map to three-dimensional state space, calculate space distortion volume and compare with reference threshold, output diagnosis result.The present application reduces the damping influence of core background vibration interference and oil temperature fluctuation, quantifies multi-source nonlinear degradation characteristics into volume index, and improves the recognition ability of early micro loose state of transformer winding.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault diagnosis technology, specifically a fault diagnosis method for oil-immersed transformers. Background Technology

[0002] Oil-immersed power transformers are crucial equipment in power grid systems. Under long-term electromagnetic stress, thermal aging, and short-circuit current impacts, the internal winding clamping structure of the transformer is prone to loosening, leading to winding mechanical instability or insulation damage. Currently, non-invasive monitoring technology based on the analysis of vibration signals from the transformer tank wall is a standard method for assessing the mechanical condition of transformer windings.

[0003] However, existing transformer vibration diagnosis methods have certain limitations in practical applications. The vibration signal measured from the transformer tank wall is a mixture of core magnetostrictive vibration and winding electromagnetic vibration. Because their characteristic frequencies overlap, conventional signal processing methods struggle to effectively separate the winding vibration component, making the extracted winding characteristics susceptible to interference from the core background vibration. Simultaneously, the insulating oil inside oil-immersed transformers provides physical damping for vibration propagation, and its viscosity changes with transformer operating load and ambient temperature. Existing technologies typically do not consider fluid damping changes caused by oil temperature fluctuations and lack dynamic temperature compensation mechanisms for vibration signals, leading to deviations in vibration measurement results under different temperature conditions. Furthermore, when early minor loosening occurs in the windings, the macroscopic structural deformation is not significant, and conventional low-frequency vibration analysis methods are insensitive to high-frequency signals generated by minor frictional slippage between insulating pads. Most existing methods rely on steady-state characteristics for single-dimensional evaluation, lacking the ability to capture and comprehensively calculate multi-dimensional response characteristics during transient load changes, making it difficult to accurately reflect the early degradation trend of the winding's mechanical state, thus limiting the accuracy of transformer fault early warning. Summary of the Invention

[0004] The technical problem solved by this invention is that existing transformer vibration diagnosis methods are difficult to effectively separate the vibration signals of the core and windings under complex operating conditions, and do not consider the interference of fluid damping fluctuations caused by oil temperature changes on vibration characteristics, resulting in low diagnostic accuracy. At the same time, conventional low-frequency vibration analysis is not sensitive to the early micro-friction slippage of the windings and lacks the ability to capture dynamic characteristics of the load transient process, making it difficult to achieve effective early warning of the mechanical state of the transformer windings.

[0005] To address the above problems, the present invention provides the following technical solution: The first aspect of this invention provides a fault diagnosis method for an oil-immersed transformer, comprising the following steps: The discrete sequences of the operating phase voltage, load current, top oil temperature, and tank wall acceleration of the transformer are collected synchronously. Under light load steady-state conditions, adaptive filtering is performed based on the discrete sequence of operating phase voltages and the sequence of box wall accelerations to obtain the locked transfer function parameter matrix; Under dynamic operating conditions, the core vibration reference component is reconstructed using the locked transfer function parameter matrix, and the winding vibration component is extracted from the differential decoupling from the box wall acceleration sequence. The relative viscosity coefficient is calculated based on the discrete sequence of the top oil temperature and a damping compensation factor is generated. The damping compensation factor is then applied to the extracted winding vibration components to obtain a standardized winding vibration signal. When the rate of change of the discrete sequence of the load current exceeds a set threshold, a time window is extracted. Within the time window, the excitation force characteristics are extracted based on the discrete sequence of the load current, and the macroscopic displacement characteristics and high-frequency distortion rate characteristics are extracted based on the standardized winding vibration signal, respectively. The excitation force characteristics, macroscopic displacement characteristics, and high-frequency distortion rate characteristics are mapped to a three-dimensional state space. The spatial distortion volume is calculated and compared with a preset benchmark threshold to output the diagnostic results of the transformer winding mechanical state.

[0006] Furthermore, the transfer function parameter matrix for obtaining the lock includes: The square of the discrete sequence of the running phase voltage is calculated to construct the real-time input sequence, and the box wall acceleration sequence is used as the desired output sequence. The filter tap weight column vector is updated using the normalized minimum mean square error algorithm; When the moving average energy of the transient estimation error is continuously lower than the steady-state error threshold, the currently converged weight coefficients are locked as the locked transfer function parameter matrix.

[0007] Furthermore, under dynamic operating conditions, the steps of reconstructing the core vibration reference components using the locked transfer function parameter matrix and extracting the winding vibration components from the box wall acceleration sequence through differential decoupling include: The square of the discrete sequence of the operating phase voltage under dynamic conditions is convolved with the locked transfer function parameter matrix in discrete time to generate the core vibration reference component. In the time domain, the box wall acceleration sequence is subtracted from the core vibration reference component point by point to extract the winding vibration component by difference. The winding vibration component is subjected to bidirectional bandpass filtering to obtain the filtered winding vibration component.

[0008] Further, the steps of calculating the relative viscosity coefficient and generating a damping compensation factor based on the discrete sequence of the top oil temperature, and applying the damping compensation factor to the extracted winding vibration components to obtain a standardized winding vibration signal include: The top-level oil temperature discrete sequence is combined with the equivalent average oil temperature discrete sequence to convert it into an absolute temperature sequence; the relative viscosity coefficient of the absolute temperature sequence relative to the reference absolute temperature is calculated based on the Arrhenius equation and used as the damping compensation factor. The vibration components of the winding are multiplied by the damping compensation factor to obtain the temperature damping compensated vibration sequence. Calculate the effective value of the discrete sequence of load current within the current analysis time window, and use the square of the effective value as a normalization factor to perform a division scaling operation on the temperature damping compensated vibration sequence to generate an amplitude-normalized vibration sequence as the normalized winding vibration signal.

[0009] Furthermore, the process of extracting a time window when the rate of change of the discrete sequence of the load current exceeds a set threshold includes: The continuously generated standardized winding vibration signal and the discrete sequence of load current are written into a ring-shaped data buffer of fixed depth in real time. The effective value sequence of sliding current is continuously calculated according to the set sliding time window, and the first-order forward differential sequence is calculated to monitor the instantaneous slope of grid load fluctuations. When the first-order forward difference sequence exceeds the load transient threshold for multiple consecutive sampling periods, a hardware interrupt is generated and the discrete time point of the first over-limit is locked as the transient zero-point index. Historical data with a length equal to the number of preceding sampling points is read back from the circular data buffer, and the latest data with a length equal to the number of subsequent sampling points is continuously collected. The data is then spliced ​​together to obtain the complete time window. After the data is extracted, the dead zone shielding time is activated to pause the over-limit detection.

[0010] Furthermore, the process of extracting excitation force features based on the discrete sequence of the load current and extracting high-frequency distortion rate features based on the standardized winding vibration signal includes: An equivalent electromagnetic excitation force sequence is constructed by performing point-by-point squaring operations on the discrete sequence of load current within the time window, and the peak value of the equivalent electromagnetic excitation force sequence is extracted as the excitation force feature. The standardized winding vibration signal is subjected to Fourier transform to the frequency domain. Based on the set low-frequency truncation threshold, a second-order frequency domain integral operation is performed, and then the signal is restored by inverse Fourier transform to generate a macroscopic displacement time discrete sequence. The peak-to-peak value of the macroscopic displacement time discrete sequence is extracted as the macroscopic displacement feature.

[0011] Furthermore, the process of extracting high-frequency distortion rate features based on the standardized winding vibration signal includes: The box wall acceleration sequence within the time window is subjected to high-pass filtering to separate the high-frequency sequence of frictional acoustic emission; The high-frequency sequence of triboacoustic emission is processed using a discrete energy operator to generate a nonlinear instantaneous energy sequence that reflects the dual abrupt changes in amplitude and frequency. Calculate the average high-frequency background energy within the transient zero-point pre-steady-state interval, calculate the relative error based on the nonlinear instantaneous energy sequence and the average high-frequency background energy, and generate a dynamic high-frequency energy distortion rate sequence. The global maximum value of the high-frequency energy distortion rate sequence is extracted as the high-frequency distortion rate feature.

[0012] Furthermore, the process of mapping the excitation force characteristics, the macroscopic displacement characteristics, and the high-frequency distortion rate characteristics to the three-dimensional state space includes: Using the mean and standard deviation of historical health condition feature data, linear translation and scaling standardization are performed on the currently extracted excitation force feature, macroscopic displacement feature and high frequency distortion rate feature to generate a dimensionless three-dimensional coordinate vector. An exponentially weighted moving average algorithm is used to calculate the weight allocation between the current three-dimensional coordinate vector and the historical evolution trajectory, and to update and generate smooth evolution trajectory coordinates, thereby constructing a manifold evolution curve in the three-dimensional state space.

[0013] Furthermore, the process of calculating the spatial distortion volume includes: calculating the phase transition critical index using the outer product operation of the spatial difference vectors of the continuous smooth evolution trajectory coordinates; when the phase transition critical index exceeds the set critical threshold, marking the corresponding transient event index as the phase transition critical point. The reference centroid is calculated using the coordinate set of the historical healthy smooth evolution trajectory prior to the phase transition critical point; Starting from the phase transition critical point, the coordinates of three adjacent smooth evolution trajectories and the reference centroid are used to form a spatial infinitesimal tetrahedron. Discrete cumulative integration is performed on the scalar volume of multiple continuous spatial infinitesimal tetrahedrons within the evaluation event span to obtain the spatial distortion volume.

[0014] Furthermore, the process of calculating the spatial distortion volume and comparing it with a preset benchmark threshold to output the diagnostic results of the transformer winding mechanical condition includes: The ratio of the macroscopic displacement feature to the excitation force feature is calculated to generate a dynamic equivalent compliance coefficient. The cumulative damage index is generated by calculating the ratio between the spatial distortion volume and the reference envelope volume. The dynamic equivalent compliance coefficient, the high-frequency distortion rate feature, and the cumulative damage index are compared with the set independent physical thresholds in a binary form to generate a three-dimensional state determination Boolean vector. The three-dimensional state determination Boolean vector is input into a preset collaborative diagnosis rule matrix for bit-by-bit matching, and the winding mechanical state diagnosis results are output, indicating normal operation, overall loosening of clamping force, local loosening of pads, or risk of pad falling off.

[0015] A second aspect of the present invention provides a fault diagnosis system for an oil-immersed transformer, comprising: The multi-source data synchronous acquisition module is used to synchronously acquire the discrete sequence of the operating phase voltage, the discrete sequence of the load current, the discrete sequence of the top oil temperature, and the tank wall acceleration sequence of the transformer. The background vibration decoupling module is used to identify and obtain the locked transfer function parameter matrix based on the phase voltage and acceleration sequence under light load steady-state conditions, and to reconstruct the core vibration reference component under dynamic conditions, thereby differentially extracting the winding vibration component from the box wall acceleration sequence. The temperature damping compensation module is used to calculate the relative viscosity coefficient based on the discrete sequence of top oil temperature using the Arrhenius equation, and to apply a damping compensation factor to the winding vibration components to generate a standardized winding vibration signal. The transient feature extraction module is used to extract excitation force features, macroscopic displacement features based on normalized vibration, and high-frequency distortion rate features based on high-frequency frictional acoustic emission signals within the time window triggered by the load current change rate. The spatial manifold diagnostic module is used to map the extracted multidimensional features to a three-dimensional state space, calculate the spatial distortion volume formed by the degradation trajectory, and comprehensively determine the diagnostic results of the mechanical state of the output transformer winding.

[0016] This invention provides a fault diagnosis method for oil-immersed transformers. It has the following beneficial effects: 1. This invention obtains the transfer function parameter matrix by filtering and identifying the phase voltage and box wall acceleration sequence under light load steady-state conditions. Then, under dynamic load conditions, it reconstructs the core vibration reference component and performs differential decoupling, effectively separating the mixed vibration signal of the core and winding, removing the background interference caused by the magnetostriction of the core, thereby obtaining the objective and true winding vibration component and improving the data accuracy of subsequent fault feature extraction.

[0017] 2. This invention calculates the relative viscosity coefficient by combining the discrete sequence of top oil temperature with the Arrhenius equation, and generates a damping compensation factor to scale and compensate the winding vibration component. This overcomes the problem of fluid damping changes caused by oil temperature rise and fall in actual transformer operation, eliminates the physical interference of ambient temperature fluctuations on vibration signal propagation attenuation, and ensures the consistency of diagnostic results under complex temperature variation conditions in substations.

[0018] 3. This invention utilizes the transient window of load current to simultaneously extract macroscopic displacement features in the low-frequency dimension and distortion rate features in the high-frequency dimension, and maps them to a three-dimensional state space to calculate the spatial distortion volume. At the same time, it takes into account the macroscopic structural flexible deformation of the winding and the early microscopic pad friction slip, transforming the multi-source nonlinear degradation process into a single volume index for quantitative comparison, thereby improving the ability to identify and judge early minor loosening faults in transformers. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the system hardware architecture according to an embodiment of the present invention; Figure 2 This is a flowchart of the method steps in an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the comprehensive performance indicators of the dynamic model experimental samples of the present invention. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See attached document Figure 1 The present invention provides a fault diagnosis system for oil-immersed transformers, comprising: a data acquisition module, a data processing module, and a fault early warning module.

[0022] The data acquisition module is configured in the transformer body and the secondary circuit of the substation. The data acquisition module includes voltage transformers, current transformers, oil temperature sensors, and piezoelectric vibration sensors.

[0023] A voltage transformer is connected to the transformer's operating phase voltage circuit to obtain the operating phase voltage signal. A current transformer is connected to the transformer's load current circuit to obtain the load current signal. An oil temperature sensor is installed at the top oil tank of the transformer to obtain the top oil temperature signal. A piezoelectric vibration sensor is fixed to the surface of the transformer tank to obtain the total vibration acceleration signal of the tank surface.

[0024] The piezoelectric vibration sensor's response bandwidth covers both the low-frequency structural mechanical vibration band and the high-frequency friction stress wave band. The data acquisition module performs analog-to-digital conversion on the physical sensing signal at a set synchronous sampling rate, generating a discrete-time series. The synchronous sampling rate satisfies the Nyquist sampling theorem's sampling requirements for the high-frequency friction stress wave band.

[0025] See attached document Figure 2This invention provides a fault diagnosis method for oil-immersed transformers. Based on the aforementioned oil-immersed transformer fault diagnosis system, it achieves the diagnostic result of determining the mechanical condition of the output transformer windings. The method includes the following steps: S10, Synchronous Acquisition and Monitoring of Multi-parameter Physical Signals: Monitoring the operating condition of the synchronously acquired discrete sequences of operating phase voltage, load current, top oil temperature, and tank wall acceleration. S20, under light load steady-state conditions, uses the phase voltage square sequence as the system input and the total vibration sequence of the box wall as the system output. Iteratively solves the finite-length unit impulse response filter coefficients through an adaptive filtering algorithm to lock the mechanical transmission link parameters from the iron core to the box wall. Under dynamic conditions, the locked filter coefficients are called and discrete-time convolution is performed with the real-time phase voltage square sequence to reconstruct the iron core vibration reference component, and the winding vibration component is differentially extracted from the total vibration acceleration sequence of the box surface. S30 establishes a physical mapping relationship between insulating oil temperature and kinematic viscosity based on the Arrhenius equation, calculates the relative viscosity coefficient based on the real-time top oil temperature and generates a damping compensation factor, applies the damping compensation factor to the differentially extracted winding vibration component, completes the amplitude and phase standardization processing of the vibration signal in the fluid medium, and outputs a standardized winding vibration signal. S40, calculate the real-time load current change rate, trigger diagnostic logic and record continuous time window when the current change rate exceeds the set threshold; within the time window, extract the low-frequency envelope of the load current square sequence as the excitation force feature, extract the low-frequency response envelope of the standardized winding vibration signal as the macroscopic displacement feature, extract the basic low-frequency band energy and the high-frequency friction stress wave band energy in the standardized winding vibration signal, and calculate the ratio of the two as the high-frequency distortion rate feature; The S50 uses excitation force characteristics, macroscopic displacement characteristics, and high-frequency distortion rate characteristics as three-dimensional orthogonal coordinates in Euclidean space to map the data sequence within the time window into a three-dimensional spatiotemporal evolution manifold; it monitors the first-order discrete derivative of the high-frequency distortion rate characteristics to determine the phase transition critical point, calculates the spatial distortion volume integral of the envelope of the three-dimensional spatiotemporal evolution manifold; compares the spatial distortion volume with the preset health benchmark threshold, and determines the mechanical connection status of the transformer winding in combination with the phase transition critical point trigger state, and outputs a normal operation signal or a winding loosening fault warning signal.

[0026] The data acquisition module is configured in the transformer body and the secondary circuit of the substation to acquire multi-dimensional physical operating status parameters of the transformer in real time. The physical basis for acquiring these multi-dimensional parameters is that the mechanical vibration of the transformer tank wall is a comprehensive response jointly excited by the magnetostriction of the iron core and the electrodynamic force of the winding, while the insulating oil temperature directly changes the dynamic damping of the transmission medium.

[0027] Based on the above principles, the data processing module controls the data acquisition module to perform synchronous sampling operations, converting continuous analog signals into discrete time series. For the synchronous acquisition process of multi-parameter physical signals, this embodiment specifically includes the following acquisition and processing logic: S101, the data acquisition module acquires the operating phase voltage signal and load current signal of the transformer through voltage transformers and current transformers, respectively. As a preferred method, the voltage transformers are connected in parallel to the transformer bus, and the current transformers are connected in series to the phase sequence circuit. While the operating phase voltage signal reflects the change in alternating magnetic flux density inside the iron core, the load current signal characterizes the transient fluctuation state of the Lorentz force on the winding conductors. The data acquisition module is internally equipped with an analog-to-digital conversion circuit with a fixed synchronous sampling rate. The analog signal is discretized to generate a discrete sequence of operating phase voltages. and load current discrete sequence .in, This represents a discrete-time series index. For the specific wiring method and range configuration of the current transformer, those skilled in the art can make conventional selections based on the transformer's rated capacity. The hardware connection method is well-known technology in this field and will not be elaborated upon here.

[0028] S102, the data acquisition module acquires the transformer top-level oil temperature signal through an oil temperature sensor to capture reference data on the variation of fluid medium viscosity with thermodynamic state. The oil temperature sensor is installed in a reserved temperature measurement hole in the transformer top-level oil tank to sense the temperature of the insulating oil fluid medium. The analog-to-digital conversion circuit simultaneously converts this analog temperature signal into a discrete sequence of top-level oil temperature. .

[0029] S103, the data acquisition module acquires the total vibration acceleration signal of the transformer tank surface through a piezoelectric vibration sensor. The piezoelectric vibration sensor is rigidly connected to the side wall of the transformer tank. The response frequency band of the piezoelectric vibration sensor is set to simultaneously cover the low-frequency structural mechanical vibration band and the high-frequency frictional stress wave band. Specifically, the low-frequency structural mechanical vibration band is set to a frequency range of 0 to 500 Hz, covering the fundamental frequency and low harmonics of the core magnetostriction and winding electrodynamic vibration. In contrast, the high-frequency frictional stress wave band is set to a frequency range of 1 kHz to 5 kHz, used to capture the mechanical friction acoustic emission signal generated when the winding pads undergo microscopic sliding. The analog-to-digital conversion circuit converts the acquired analog acceleration signal into a tank wall acceleration sequence. .

[0030] S104, Before performing the aforementioned analog-to-digital conversion, to prevent spectral folding caused by high-frequency electromagnetic interference or broadband white noise above the target analysis frequency band in the environment during sampling, the data acquisition module is equipped with a hardware anti-aliasing low-pass filter with a cutoff frequency matched to the target upper limit frequency. After filtering, the data acquisition module operates at the set synchronous sampling rate. Perform synchronous analog-to-digital conversion on the signals from the four channels. Synchronous sampling rate. The numerical settings satisfy the Nyquist sampling theorem. To ensure that the piezoelectric vibration sensor can completely acquire the high-frequency friction stress wave band with an upper limit frequency of 5kHz, the synchronous sampling rate is... The lower threshold is set to twice the highest analysis frequency of the high-frequency friction stress wave band. In specific implementation, the synchronous sampling rate of the data acquisition module is... Configured to be greater than or equal to 20kHz.

[0031] To reduce time skew between heterogeneous signals from multiple channels, the data acquisition module uses a shared external crystal oscillator as a unified global clock pulse to drive multi-channel synchronous sample-and-hold circuits to simultaneously lock transient physical quantities. This hardware architecture ensures the operation of discrete phase voltage sequences. Discrete sequence of load current Discrete sequence of top oil temperature and the box wall acceleration sequence Maintaining the required synchronization alignment on the physical time scale ensures that the same discrete-time series index exists in each channel sequence. The corresponding physical sampling times are basically consistent, thus meeting the engineering requirements for phase consistency in subsequent signal differential and transfer function decoupling operations.

[0032] Transformer tank wall vibration is dominated by different physical sources at different operating stages. During light load or no-load periods, the electrodynamic force on the transformer winding conductors is approximately zero, and the mechanical vibration of the tank wall mainly originates from the magnetostrictive effect of the iron core. This physical characteristic provides a quantifiable benchmark for separating iron core vibration from winding vibration. Based on this principle, the data processing module executes a set of transformer operating condition state machine judgment logic, dividing the continuous equipment operation cycle into specific operating conditions to guide the parameter updates and diagnostic triggering of subsequent adaptive filtering algorithms.

[0033] In this embodiment, the transformer operating condition state machine determination logic specifically includes the following determination and state switching sub-steps: S105, The data processing module receives the discrete sequence of load current generated by the data acquisition module in real time. Feature extraction is performed to quantify the current load level and fluctuations. To reduce the interference of the instantaneous zero-crossing of the power frequency AC current on amplitude determination, the data processing module calculates the short-time effective value sequence of the load current by setting a time sliding window. The specific calculation formula is as follows: ,in, Indicates the number of sampling points contained within the time sliding window; This indicates that the number of sampling points included within the time window is reduced by one; It is a discrete sequence of load currents; These are historical sampling points within the time window; It is the short-time effective value of the previous discrete moment; This represents the index for backtracking and summing historical samples within the sliding window. To ensure calculation accuracy, this value is set to the number of samples covering at least one complete power frequency cycle. As a preferred approach, if the power grid frequency of the system is 50Hz and the data acquisition synchronous sampling rate is... If set to 20kHz, then The specific value is set to 400. Based on the calculated short-time effective value sequence, the data processing module further calculates the first-order discrete forward difference to generate a current change rate sequence characterizing load fluctuations. : For the memory data caching of the time sliding window in the effective value calculation process and the register shifting operation in the differential calculation, those skilled in the art can use the low-level data read and write instructions of conventional digital signal processors to implement them. The specific addressing and storage mechanism is a well-known technology in the field and will not be described in detail here.

[0034] S106, based on the extracted feature sequence, the data processing module executes the state machine switching logic. In this step, the system determines whether the current operating environment meets the light-load steady-state condition. If the determination condition is met... and and continues to exceed the preset time threshold. At this time, the data processing module switches the system state machine identifier to the light-load steady-state condition. Where, This represents a sequence of short-time effective values ​​of the load current. This indicates the set light load current threshold. This represents a sequence of current change rates. Based on the transformer nameplate parameters, this threshold is typically set to 1% to 5% of the transformer's rated load current. This represents the steady-state current change rate threshold, used to shield the power grid base load ripple interference. Its value is determined based on the static current fluctuation variance of the background power grid when there is no switching operation. The steady-state time threshold is set based on the transformer mechanical vibration damping decay time constant, which is usually taken as 5 to 15 minutes.

[0035] S107, when the load current characteristics break the above-mentioned light-load steady-state conditions, the system exits the parameter identification mechanism. As a preferred approach, the data processing module continuously monitors the current change rate sequence after exiting the light-load steady-state condition. When the condition is detected to be met... At that time, the state machine determines that the system enters a dynamic transient condition. In the formula, The set load transient threshold characterizes the load step that the transformer is experiencing due to the start-up and shutdown of large equipment.

[0036] Apart from the two extreme operating conditions mentioned above, the transformer operates under normal load fluctuations for most of its operating time. If the load current... At the light load current threshold If the load is between the rated full load and the dynamic transient condition is not met, the state machine determines that the transformer is in a normal load condition. In this state, the data processing module calls the existing filter parameters to perform routine real-time source decoupling calculations, but does not update the filter parameters or trigger fault diagnosis logic.

[0037] The magnetostrictive deformation force of the transformer core exhibits a high correlation with the square of the main magnetic flux density within the core, which numerically corresponds to the integral of the operating phase voltage. In discrete-time signal processing, considering that the power frequency AC system is in steady state, the integration operation on the sinusoidal signal is essentially equivalent to a fixed phase shift and amplitude scaling, and the multi-tap structure of the finite-length unit impulse response filter itself possesses full-band amplitude and phase compensation capability. Based on this principle, the data processing module can directly use the square sequence of the operating phase voltage to equivalently characterize the excitation force source of the core. When the state machine is in a light-load steady-state condition, the current flowing through the transformer windings is usually at a low level, and the winding vibration component excited by the Lorentz force is negligible. At this time, the total vibration acceleration signal collected from the tank surface mainly represents the mechanical response of the core vibration transmitted to the tank wall through the insulating oil and mechanical structure. The data processing module utilizes this specific physical window to execute an adaptive filtering algorithm, calculate and lock the spatial mechanical transmission link parameters from the core to the tank wall.

[0038] In this embodiment, the steady-state transfer function identification process based on the NLMS algorithm specifically includes the following data processing sub-steps: S201, After confirming that the system has entered a light-load steady-state operating condition, the data processing module extracts the discrete sequence of the operating phase voltages within that time period. and the box wall acceleration sequence To reduce the interference of DC bias components in the power grid on parameter identification, the data processing module pre-processes the phase voltage sequence with zero mean. After processing, the system calculates the square of the discrete phase voltage sequence to construct the real-time input sequence. ,Right now Meanwhile, the acceleration sequence of the box wall under this operating condition is directly used as the desired output sequence of the system. ,Right now .

[0039] S202, after acquiring the input and desired output, the data processing module constructs a finite-length unit impulse response filter model. This filter is used to simulate the dynamic response process of a space mechanical transmission link, and its actual output sequence... Indicates the real-time input sequence The estimated core vibration response generated by the excitation. This is the actual output sequence. The calculation formula is: In the formula, Let the input delay column vector consist of current and historical input data, with dimension 1. Specifically, it is expressed as: ; Let be the column vector of filter tap weights at the current iteration, with the same dimension . , represented as: ,in, Represents the order of the filter; This represents the matrix transpose. As a preferred method, the order... The setting is based on the product of the physical delay time span of the sound wave inside the transformer from the iron core to the tank wall and the synchronous sampling rate, which is usually between 200 and 500 to ensure that the main energy area of ​​the complete structural impulse response can be covered.

[0040] S203, As the signal sequence is input step by step, the data processing module calculates the system's desired output sequence. With the actual output sequence Transient estimation error between : Based on this error, the system uses the normalized minimum mean square error algorithm to update the filter tap weight column vector. The iterative update formula for the filter tap weight column vector is: In the formula, This is the column vector of filter tap weights at the current iteration time; This is the weight column vector for the next time step; Represents the input delay column vector; It represents the square of the Euclidean norm of the input delay column vector, that is, the energy of the input signal within the current time window; This is the iteration step size factor of the algorithm, with a value ranging from 0 to 2. To achieve a balance between steady-state convergence speed and system offset error, The standard value is set to 0.01 to 0.1; Ɛ represents a regularization constant to prevent numerical overflow caused by minimizing the denominator, and is usually taken as a small positive real number, such as 10. -6 .

[0041] S204, During the sustained light-load steady-state operation, the data processing module executes the aforementioned error calculation and weight update operations point by point, guiding the filter tap weight column vector to approximate the actual physical space transfer function. To avoid blindly locking parameters before the algorithm fully converges, the data processing module simultaneously calculates the moving average energy of the transient estimation error. When the moving average energy is lower than the preset steady-state error threshold for several consecutive power frequency cycles, the system determines that the adaptive filter has reached convergence. Under this premise, if the state machine detects load current fluctuations and determines to exit the light-load steady-state operation, the data processing module stops the iterative weight update process, fixes the converged weight column vector at this time, and saves it as a locked transfer function parameter matrix. The successfully locked parameter matrix fully maps the inherent mechanical characteristics of the transmission of the core excitation force to the box wall under specific oil temperature and structural tightening conditions.

[0042] After a transformer leaves a light-load steady state and enters a dynamic transient or normal load condition, the load current flowing through the winding conductors increases. The Lorentz force generated by the alternating current acts on the winding structure, inducing mechanical vibration. To assess the mechanical connection status of the windings, the data processing module uses previously locked mechanical transmission link parameters to separate the core vibration component from the mixed vibration signal, thereby achieving physical decoupling of the vibration source.

[0043] In this embodiment, the discrete-time convolution decoupling process between the core and the winding vibration source specifically includes the following calculation sub-steps: S205, under dynamic operating conditions, the data processing module receives in real time the discrete sequence of operating phase voltages synchronously converted by the data acquisition module. and the box wall acceleration sequence To ensure the consistency of the calculation benchmark, the data processing module performs DC bias removal processing on the discrete sequence of running phase voltages. As a preferred method, this DC bias removal process is achieved by calculating the average value of the voltage sequence within the current observation time window and subtracting it point-by-point from the original sequence. After processing, the squared values ​​are calculated to form the real-time input sequence. Due to the change in operating conditions, the filter tap weights are no longer adaptively updated. The data processing module calls the transfer function parameter matrix that was previously locked under light-load steady-state conditions. Compare it with the real-time input sequence Perform discrete-time convolution operations.

[0044] S206, the physical significance of discrete-time convolution lies in reconstructing the vibration response of the core on the tank surface at the current moment using known excitation force sources and system spatial transmission characteristics. This operation is based on the theoretical assumption that the transformer's mechanical transmission path is approximately a linear time-invariant system within a finite observation window. The reconstructed core vibration reference component... The calculation formula is: In the formula, Typically represents a discrete time index; The transfer function parameter matrix representing the lock The first in Each weighting coefficient; Indicates the process The input sequence value after a delay of one sampling period; The filter order was set in advance; Decrease the filter order set in the initial stage by one. This applies to the initial phase of the operation. In cases where historical data is missing, the system maintains the convolution length by padding with zeros. For the underlying hardware implementation of this finite-length discrete convolution, those skilled in the art can use a hardware multiplier-accumulator within a digital signal processor in conjunction with a circular addressing register. The allocation of computational resources and the parallel processing methods are well-known technologies in this field and will not be elaborated upon here.

[0045] S207, based on the reconstructed core vibration reference components, the data processing module performs time-domain difference extraction. Due to the box wall acceleration sequence... It is composed of the linear superposition of core vibration, winding vibration, and background noise. The discrete-time convolution operation already contains the group delay information of the signal propagation in physical space. The reconstructed component is basically consistent with the measured total vibration in terms of physical propagation delay. The data processing module subtracts the box wall acceleration sequence from the reconstructed core vibration reference component point by point in the time domain, and extracts the winding vibration component by difference.

[0046] S208, the specific differential calculation formula for the winding vibration component is as follows: In the formula, This represents the vibration components of the decoupled winding; This represents the reconstructed core vibration reference component; Represents the sequence of accelerations of the box wall; This typically represents a discrete time index. Considering the high-frequency broadband background noise interference present in actual measurements, the data processing module applies a digital bandpass filter with a passband range matched to the transformer's mechanical resonant frequency before outputting the differential results. In this embodiment, the lower cutoff frequency of the bandpass filter is set to 10Hz to filter out slowly varying trend terms, and the upper cutoff frequency is set to 500Hz to retain the fundamental frequency of the electrodynamics and the main higher harmonic energy. To avoid the nonlinear phase distortion caused by traditional recursive filters affecting the alignment accuracy of subsequent feature extraction, the data processing module performs forward and backward bidirectional filtering on the data within the captured time window, thereby achieving zero-phase-shift filtering in offline or near-real-time analysis frameworks. The winding vibration component after this filtering process... It effectively eliminates low-frequency background interference dominated by iron core magnetostriction, and its amplitude and phase characteristics can reflect the mechanical response state of the winding under the current Lorentz force excitation, providing an independent data source for subsequent dynamic compensation and feature space construction.

[0047] The insulating oil inside a transformer not only serves as an insulating and heat dissipation medium, but also plays a role in transmitting mechanical waves and providing structural damping at the mechanodynamic level. The viscosity of the insulating oil changes non-linearly with thermodynamic temperature. When the transformer load increases, leading to a rise in oil temperature, the viscosity of the insulating oil decreases, and the fluid damping decreases accordingly. This causes a change in the amplitude of the winding vibration response measured under the same electrodynamic excitation.

[0048] In this embodiment, the modeling and calculation process for the relative viscosity coefficient of insulating oil specifically includes the following sub-steps: S301, the data processing module synchronously extracts the discrete sequence of top oil temperature within the current operating time window. Due to the internal thermal convection circulation of fluids within the transformer, the top oil temperature is typically higher than the average oil temperature around the windings, while mechanical vibrations mainly propagate within the viscosity environment corresponding to the average oil temperature. To correct this thermodynamic discrepancy, the data processing module calculates the equivalent average oil temperature discrete sequence based on the heat conduction model of the transformer radiator. As a preferred method, the empirical formula for calculating the equivalent average oil temperature is: ,in This is the average correction value for the top and bottom oil temperature difference, which is usually set between 10 and 15 degrees Celsius based on transformer temperature rise test data. This represents the discrete sequence of top-layer oil temperature. After correction, because the Arrhenius model is based on the theory of molecular thermal motion, the input parameters need to use an absolute thermodynamic temperature scale. The data processing module performs a temperature scale conversion operation on the equivalent average oil temperature discrete sequence to generate an absolute temperature sequence. The specific conversion formula is as follows: In the formula, The unit is Celsius. The unit is Kelvin. For the basic scalar addition and floating-point conversion logic in digital signal processing systems, those skilled in the art can implement it using underlying basic arithmetic logic units. The instruction set calls are well-known techniques in this field and will not be elaborated upon here.

[0049] S302, based on the converted absolute temperature sequence, the data processing module uses the Arrhenius equation to calculate the estimated dynamic viscosity of the insulating oil under the current thermodynamic state. This physical model describes the energy relationship required for fluid molecules to overcome intermolecular forces and undergo relative displacement. The calculation formula is as follows: In the formula, Represents an absolute temperature sequence; is the base of the natural logarithm; Indicates the first Estimated dynamic absolute viscosity of insulating oil at any given time; This is a pre-exponential factor related to the basic physical properties of insulating oil, and its value depends on the molecular weight and initial structure of the oil. The activation energy of insulating oil is the energy barrier required for microscopic molecular transitions during macroscopic fluid flow, and its unit is usually joules per mole. The ideal gas constant is 8.314. Pre-exponential factor. With flow activation energy The specific values ​​can be obtained by referring to the petrochemical industry standard oil property parameter table based on the specific grade of insulating oil added to the transformer, or by conducting a constant temperature gradient calibration test using a standard rotational viscometer before the equipment is put into operation.

[0050] S303: Directly using absolute viscosity values ​​in subsequent system decoupling calculations introduces complex dimensional conversions. To simplify the damping compensation logic, the data processing module calculates the current absolute temperature sequence. Relative viscosity coefficient relative to reference temperature By setting a reference absolute temperature The data processing module calculates the ratio of the current viscosity to the reference viscosity. The simplified relative viscosity coefficient is then calculated. The calculation formula is: In the formula, The selected reference absolute temperature; is the base of the natural logarithm; The activation energy for the flow of insulating oil; This is the ideal gas constant. In practical engineering applications, the absolute temperature corresponding to the rated ambient temperature of 20 degrees Celsius for a transformer is usually set to 293.15 K. The relative viscosity coefficient It is a dimensionless scaling factor that reflects the ratio of the system damping deviation from the reference state caused by the current oil temperature.

[0051] S304. Considering that the oil temperature sensor may experience transient abnormal fluctuations due to electromagnetic interference, causing the calculated relative viscosity coefficient to diverge exponentially, the system sets numerical boundary clamping logic before outputting this coefficient. The data processing module judges the calculated absolute temperature sequence. Whether it is within the reasonable physical operating range of the transformer. Specifically, the lower limit of this range is usually set at 233.15K, corresponding to -40 degrees Celsius, and the upper limit is set at the safe critical value of the transformer insulating oil flash point, usually 403.15K, corresponding to 130 degrees Celsius. When When the value exceeds this specific range, the data processing module stops refreshing the relative viscosity coefficient and uses the calculation results from the previous valid sampling period.

[0052] After separating the winding vibration components and obtaining the relative viscosity coefficient of the insulating oil, directly comparing vibration data at different times often leads to inconsistencies in data benchmarks due to varying operating conditions. From a physical perspective, the Lorentz electrodynamic force on the transformer winding is proportional to the square of the load current, resulting in differences in vibration amplitude under different load levels. Simultaneously, variations in insulating oil damping also affect the decay process of mechanical vibration. To construct a feature space that accurately reflects the inherent state of the winding's mechanical structure, the data processing module performs amplitude and phase standardization on the winding vibration components, mapping the dynamic physical response to uniform excitation force and damping boundary conditions.

[0053] In this embodiment, the amplitude and phase normalization process for the winding vibration signal specifically includes the following calculation and conversion sub-steps: S305, the data processing module extracts the decoupled winding vibration components within the current analysis time window. With relative viscosity coefficient Based on the extracted data, the system uses the relative viscosity coefficient to compensate for the vibration amplitude with fluid damping. Considering that under harmonic excitation, the acceleration response amplitude of the transformer winding structure is approximately inversely proportional to the viscous damping coefficient of the system within a certain range, the data processing module multiplies the decoupled winding vibration component sequence point-by-point with the relative viscosity coefficient as a scalar to generate a temperature-damped compensated vibration sequence. The calculation formula is as follows: The physical mechanism of this compensation operation is to use the relative viscosity coefficient as a proportional gain to calculate the vibration amplitude attenuation or amplification effect caused by the change of fluid damping under different oil temperature backgrounds, so that the processed vibration sequence is equivalent to the physical response measured at a set reference thermodynamic temperature.

[0054] S306, after completing damping compensation, in order to further reduce the impact of load fluctuations on the amplitude of the excitation force source, the data processing module simultaneously processes the discrete sequence of the load current. This is used to extract the steady-state electrodynamic amplitude benchmark within the corresponding time window. The data processing module calculates the effective value of the discrete sequence of load current within the current analysis time window. According to the principles of electromagnetism, the magnitude of the Lorentz force on the transformer winding conductors is proportional to the square of the current flowing through them. Utilizing this physical relationship, the data processing module uses the square of the effective value of the load current as a normalization factor to perform a division scalar scaling operation on the temperature-damped compensated vibration sequence, generating an amplitude-normalized vibration sequence. The specific calculation formula is as follows: In the formula, A small regularization constant is used to prevent the denominator from approaching zero at the moment of transformer light load or no load cut-off, which would lead to numerical overflow. This represents a temperature-damped compensated vibration sequence. As a preferred method, The value is determined by the data processing system, where the floating-point precision is typically set to 10. -6 The amplitude-normalized vibration sequence output after this step essentially characterizes the pure mechanical amplitude response of the winding under the action of a unit current square excitation force.

[0055] In S307, when assessing mechanical degradation such as decreased winding clamping force or loose pads, changes in structural stiffness often lead to a shift in the mechanical resonant frequency. This change in the system's inherent properties, under electrodynamic excitation at a fixed frequency, manifests as a lag or lead in the measured response phase. To accurately extract this degradation characteristic, the system needs to establish a stable phase reference. The data processing module uses a synchronously sampled discrete sequence of load currents. As a phase reference for the excitation source, since the main frequency of the grid current is concentrated at the power frequency, while the main frequency of the electrodynamic force is located at twice the power frequency, the data processing module first performs point-to-point squaring on the discrete sequence of the load current. Subsequently, the system uses a digital bandpass filter to extract the twice-power frequency component of this squared sequence, generating the excitation force reference sequence. As a preferred approach, for a 50Hz power frequency system, the lower cutoff frequency of the digital bandpass filter is set to 95Hz and the upper cutoff frequency is set to 105Hz to ensure full extraction of the main frequency energy of the electric drive and suppression of harmonic interference.

[0056] S308, after acquiring the reference sequence, the data processing module processes the amplitude-normalized vibration sequence. and excitation force reference sequence Phase extraction and synchronization alignment are performed. The system uses Hilbert transform to analytically reconstruct the two real signals and calculates the amplitude of the analytic signals in the complex plane, thereby obtaining the instantaneous phase sequences of the vibration. With the instantaneous phase sequence of the excitation force Next, by performing point-by-point subtraction in the time domain, an instantaneous phase difference sequence reflecting the dynamic delay of the mechanical structure is extracted. The specific formula is as follows: In the formula, This represents the phase unwinding operation. By detecting whether the phase transition between adjacent sampling points exceeds pi, it adaptively adds or subtracts integer multiples of pi to reduce the wrapping discontinuities caused by the principal value range limitation of the arctangent function, thereby restoring the continuous physical phase evolution trajectory.

[0057] Considering that the instantaneous phase is susceptible to severe fluctuations due to broadband background noise when the signal envelope approaches zero, leading to phase estimation distortion, the data processing module sets an envelope threshold mask logic when calculating the instantaneous phase difference sequence. When the instantaneous envelope amplitude of the excitation force reference signal or vibration signal is less than twice the preset root mean square (RMS) value of the background noise, the system determines that the phase calculation for that discrete moment is invalid. To ensure the reliability of this threshold, the aforementioned RMS value of the background noise is usually pre-recorded and calculated by the data acquisition module in a static state with the transformer de-energized. For discrete moments determined to be invalid, the system uses the phase values ​​of valid moments before and after that point for linear interpolation to fill in the gap. If the proportion of invalid points within the current time window exceeds the set data quality threshold, the system will discard the phase calculation results for that time window to avoid outputting false phase characteristics.

[0058] The mechanical response of a transformer winding under steady-state operation exhibits an approximately periodic characteristic, containing relatively limited information about its structural dynamic evolution. When the transformer load undergoes a step change, the surge in Lorentz electrodynamic force exerts a broadband mechanical shock on the winding. This transient shock can excite potential mechanical loosening or deformation responses within the winding structure. To extract dynamic characteristics reflecting the winding's health status, the data processing module executes a data window capture logic to extract standardized multidimensional sequences containing the complete transient evolution process.

[0059] In this embodiment, the logic for capturing load transient data windows specifically includes the following sub-steps: S401, the data processing module maintains a fixed-depth annular data buffer during normal transformer operation. This buffer employs a first-in, first-out (FIFO) memory read / write mechanism to store continuously generated amplitude-normalized vibration sequences, instantaneous phase difference sequences, and discrete load current sequences in real time. The depth of the annular data buffer affects the fault lead time that the system can trace. As a preferred approach, the buffer depth is set to cover at least fifty complete cycles of sampling data under the power frequency system, thereby preserving the steady-state reference mechanical response of the winding before load abrupt changes. Simultaneously, the data processing module continuously calculates the effective value sequence of the sliding current according to a set sliding time window for the discrete load current sequence. This serves as a benchmark for monitoring load fluctuations. For the circular shift control of the memory address pointer in the circular buffer, those skilled in the art can implement it using the direct memory access controller of the underlying microprocessor. The hardware address mapping rules are well-known in the field and will not be elaborated upon here.

[0060] S402, during the continuous writing of data to the buffer, the data processing module bases its data on the effective value sequence of the sliding load current. Real-time computation of its first-order forward difference sequence The calculation formula is: This is to monitor the instantaneous slope of power grid load fluctuations. When a condition is detected... At that time, the system determines that the transformer has been subjected to a sudden load impact. In the formula, This indicates the set load transient threshold. The setting of this threshold must avoid false triggering caused by slow load changes in the conventional power grid. In actual engineering deployments, this threshold is determined in conjunction with the transformer nameplate capacity parameters. Typically, this is set to 5% to 15% of the transformer's rated full-load current RMS value. To prevent false triggering caused by high-frequency noise spikes, as a preferred approach, the system requires this over-limit condition to occur continuously. Each sampling period (usually) A valid trigger can only be confirmed if the value (ranging from 3 to 5) is continuously true. Once all the above conditions are met, the data processing module generates an internal hardware interrupt, locking the current first time point exceeding the limit discrete time point as the transient zero-point index. .

[0061] S403, relying on the locked transient zero-point index The data processing module initiates a time window interception mechanism. The system uses... Based on this, the length read back from the historical cache is... The pre-sampling point, and simultaneously continuously sampled along the positive time axis for a length of... The post-sampling points. The total length of the complete transient data window after stitching. Determined by the following formula: In the formula, Determined by the system's need to trace the steady-state baseline before the mutation, its value is equal to the preceding physical time span. With synchronous sampling rate The product of It is usually set to 1 to 2 seconds. Determined by the mechanical damping decay time constant of the transformer windings, its value is equal to the subsequent physical time span. With synchronous sampling rate The product of The time interval is typically set to 3 to 5 seconds to ensure that the mechanically damped aftershocks caused by the Lorentz force impact are included. The constructed multidimensional data matrix synchronously includes amplitude-normalized vibration sequences, instantaneous phase difference sequences, and discrete load current sequences, which fully record the three continuous physical stages: steady state before the abrupt change, impact transition state, and steady state after the abrupt change.

[0062] S404. Considering that the transformer may experience multiple consecutive load surges within a short period, frequent triggering of the capture logic could easily lead to overlapping data windows, thereby overloading the computational resources of subsequent feature extraction algorithms. Therefore, after completing a full data window capture, the data processing module initiates a dead-zone masking period. During this masking period, the system suspends processing of the normal load transient threshold. For over-limit detection, newly incoming monitoring data is simply cached normally without triggering a new round of window capture. As a preferred approach, the dead-zone shielding time is set to 1.5 times the length of the aforementioned subsequent data segment. However, to avoid missing more severe short-circuit impact faults during the shielding period, the data processing module sets a limit short-circuit threshold in parallel. This short-circuit threshold is typically set to 2 to 3 times the rated current. During dead-zone shielding, if a short circuit is detected... The system will forcibly break the shielding state, abandon or seal the currently captured data window, and use the new fault trigger point as... Restart the S403 interception process. By introducing this anti-jitter and multi-level threshold interrupt mechanism, the algorithm scheduling continuity of the diagnostic system in complex power grid environments is ensured, and the logic dead zone caused by continuous transients is effectively eliminated.

[0063] After capturing the transient load data window, the system acquired a multi-dimensional data matrix encompassing the entire process of the sudden change. The vibration acceleration signal measured on the transformer tank surface is highly sensitive to high-frequency local vibrations. Low-frequency macroscopic displacements have a more direct physical characterization significance when reflecting overall winding structural loosening or a decrease in clamping force. When structural stiffness degrades, the physical deformation of the windings under the same excitation force usually increases. To construct an evaluation index that can quantify this change in mechanical state, the data processing module simultaneously extracts transient electromagnetic excitation force characteristics and winding macroscopic displacement characteristics within the captured data window, establishing a dynamic mapping relationship between force and displacement.

[0064] In this embodiment, the extraction process of excitation force and macroscopic displacement characteristics specifically includes the following calculation and transformation sub-steps: S405, the data processing module extracts excitation force characteristics based on the discrete sequence of load current within the capture window. The Lorentz electrodynamic force experienced by the transformer winding under load conditions is positively correlated with the square of the current flowing through the winding. The data processing module processes the discrete sequence of load current within the data window... Perform point-by-point squaring to construct an equivalent electromagnetic excitation force sequence. The calculation formula is: Based on this equivalent sequence, the data processing module extracts the excitation force characteristic quantity reflecting the impact intensity. As a preferred method, the system extracts the peak value of the equivalent excitation force within the window. and the step change in excitation force .in, for The maximum value in the sequence represents the maximum mechanical load the winding experiences during a transient process. To accurately calculate... The system combines the locked transient zero-point index ,Will Previous Each sampling point is defined as the steady-state interval before the mutation, and the length of the end of the data window is equal to... The sampling points are defined as the steady-state interval after the mutation. The difference between the mean of the steady-state interval after the mutation and the mean of the steady-state interval before the mutation is calculated and used to characterize the migration range of the working reference point of the mechanical structure.

[0065] S406, In order to obtain the physical displacement state of the winding, the data processing module processes the amplitude-normalized vibration sequence. Discrete-time integration is performed. In kinematics, a single integration of the acceleration signal generates a velocity signal, and a double integration generates a displacement signal. However, due to low-frequency thermal noise and quantization truncation errors in practical sensors, direct numerical integration in the time domain leads to the continuous accumulation of constant bias, causing baseline drift in the calculation results. To suppress this drift, the data processing module uses an algorithm combining frequency domain integration and time domain denoising to reconstruct the macroscopic displacement sequence. Before frequency domain transformation, the data processing module... A smoothing window function, such as the Hanning window, is applied to reduce edge discontinuities and spectral leakage caused by data truncation. Subsequently, a Fast Fourier Transform is performed on the windowed sequence to transform it into the frequency domain, yielding the acceleration spectrum. .

[0066] S407, in the frequency domain, the data processing module calculates the displacement spectrum using the frequency domain integration theorem. The mathematical mechanism of the second integral in the frequency domain is equivalent to dividing the original frequency domain signal by the square of the angular frequency. The specific calculation formula is as follows: In the formula, For the first The displacement frequency domain complex value corresponding to each frequency node; For the first The physical frequency of each frequency node; This is the acceleration spectrum obtained from frequency domain transformation; the negative sign in the constant term originates from the phase flip caused by two consecutive derivatives of the complex exponent. This is to avoid [further issues] at lower frequencies. To address the numerical divergence dead zone caused by values ​​approaching zero, the data processing module sets a low-frequency cutoff threshold during frequency domain integration. .when At that time, the system forced an order Based on the actual modal distribution of transformer mechanical vibration, this low-frequency cutoff threshold... The frequency is typically set between 5Hz and 10Hz. For the underlying instruction scheduling of Fast Fourier Transform and frequency domain complex division, those skilled in the art can rely on the built-in mathematical acceleration hardware of conventional digital signal processing chips. The memory mapping rules are well-known in the field and will not be elaborated upon here.

[0067] S408, after completing frequency domain division and low-frequency truncation, the data processing module processes the displacement spectrum. Perform an inverse fast Fourier transform to restore the time domain, generating a macroscopic displacement-time discrete sequence. Based on this displacement sequence, the data processing module extracts displacement characteristic quantities that reflect the mechanical state. The system calculates the peak-to-peak value of the displacement sequence within the transient window. This value equals The difference between the global maximum and global minimum values ​​in the sequence represents the maximum allowable movement of the winding under transient impact. The data processing module also extracts the displacement envelope decay time constant. As a specific implementation, the system utilizes the Hilbert transform to extract... An absolute envelope is constructed from the analytical signal amplitudes of the sequence, and the highest peak point of the envelope after the impact is identified. Starting from this peak point, a natural exponential decay function is used to fit the subsequent envelope curve, extracting the time span required for the envelope amplitude to decrease to 36.8% of the initial peak. This time constant... The system damping ratio, which is related to the winding structure and the mechanical fastening state, decreases when the winding clamping force decreases, and the decay time constant usually becomes longer.

[0068] When transformer windings become loose or insulation pads wear, microscopic relative slippage can easily occur between conductors and insulators during macroscopic displacement caused by load transients. Dry friction during relative motion typically excites high-frequency transient elastic waves, i.e., acoustic emission signals. The frequency range of these signals is much higher than the inherent mechanical resonant frequency of the winding structure.

[0069] In this embodiment, the extraction process for the high-frequency distortion rate characteristics of triboacoustic emission specifically includes the following calculation and transformation sub-steps: S409, the data processing module extracts the original broadband vibration acceleration discrete sequence from the captured load transient data window. Considering that the acoustic emission signals generated by dry friction are concentrated in the higher frequency band, the system is pre-configured with a system sampling frequency of at least 50kHz for the broadband vibration sensor to meet the Nyquist sampling theorem requirement for high-frequency signal acquisition. The system uses a digital high-pass filter to truncate the original broadband vibration acceleration discrete sequence, separating the high-frequency sequence of friction acoustic emission. As a preferred approach, the cutoff frequency of this digital high-pass filter is typically set to 10kHz to filter out the electrodynamic main frequency response, mechanical resonance response, and low-frequency vibrations from the environmental background. The configuration of the coefficients of the multi-order difference equations and the discrete-time convolution operation for the digital high-pass filter can be achieved by those skilled in the art using standard filter design toolkits; the system transfer function design rules are well-known in the field and will not be elaborated upon here.

[0070] After acquiring the high-frequency components, S410 directly analyzes the instantaneous amplitude in the time domain, which is insufficient to stably reflect the energy burst of transient friction. The data processing module calculates the instantaneous energy envelope of the high-frequency sequence of triboacoustic emission. To improve the sensitivity to high-frequency transient impacts and suppress smooth background high-frequency noise, the system introduces a discrete energy operator to process the sequence. This operator, in its physical mechanism, can simultaneously respond to the square of the signal amplitude and the change in instantaneous frequency, making it suitable for capturing transient friction impact signals accompanied by amplitude modulation and frequency modulation characteristics. The data processing module calculates the nonlinear instantaneous energy sequence. The specific calculation formula is as follows: In the formula, This represents the high-frequency vibration amplitude at the current sampling point. These represent the amplitudes of adjacent sampling points before and after.

[0071] S411. To quantify anomalous energy mutations, a reference baseline under steady-state frictionless conditions needs to be established. The data processing module, relying on previously locked transient zero-point indices, calculates the average high-frequency background energy within the steady-state range before the mutation. The system extracts steady-state window data with a length equal to the preset number of preceding sampling points before the transient zero-point index, and calculates the arithmetic mean sum of the absolute values ​​of its instantaneous energy sequences. The calculation formula is as follows: In the formula, For transient zero-point indexing; Let be the instantaneous energy of the i-th sampling point; This represents the absolute value of the instantaneous energy at the i-th sampling point; This represents the set number of pre-sampling points. This average value characterizes the inherent high-frequency background noise energy level of the transformer under stable electromagnetic load, generated by high-frequency harmonics from the magnetostriction of the iron core and the fluid flow in the cooling system. To prevent sensor channel failures from causing benchmark calculation failures, the system sets a sensor noise floor threshold. If... If the noise level is below the noise floor threshold, the system determines that the current vibration acquisition channel is abnormal and stops the subsequent high-frequency feature extraction calculation for that channel.

[0072] S412, based on the extracted instantaneous energy and background reference, the data processing module calculates the dynamic high-frequency energy distortion rate sequence. This sequence reflects the relative proportion of deviations from the steady-state background noise at each moment during the shock transient process. The system calculates this by taking the relative error: In the formula, To prevent the denominator from approaching zero in low-noise environments and causing a dead zone in the numerical divergence, a regularization constant is preferably set to 10. -6 ; This represents the average high-frequency background energy. This is the current instantaneous energy sequence; This represents the absolute value of the instantaneous energy at time n. The data processing module extracts the global maximum value of the high-frequency energy distortion rate sequence within the data window and defines it as the high-frequency distortion rate feature. When the internal insulation pads of the transformer are loose, the dry friction accompanying macroscopic displacement usually causes this characteristic value to increase. To avoid the algorithm extracting random noise spikes when there are no abnormal fluctuations, the data processing module sets an effective threshold for distortion rate. This effective threshold is determined based on the maximum noise fluctuation multiple in the transformer's normal operating history data, and its value range is generally between 3.0 and 5.0. If the calculated global maximum value is lower than this effective threshold, the system determines that the current impact process has not triggered effective frictional acoustic emission and forcibly reduces the high-frequency distortion rate characteristic value. Setting it to zero ensures the purity of the feature space for subsequent state diagnosis.

[0073] After acquiring the characteristics of a single load transient process, the system needs to assess the state degradation trend of the transformer windings during long-term operation. The decline in the mechanical fastening condition of the windings is usually not a linear evolution process, but rather exhibits staged abrupt changes. When the internal pads wear down or the clamping force loosens to a certain boundary, the system's mechanical response mapping logic will be restructured. To capture the critical moment of this nonlinear degradation, the state diagnostic module maps the historically extracted multidimensional physical features to a high-dimensional space, identifying state abrupt changes through changes in geometric topological properties.

[0074] In this embodiment, the process of capturing the three-dimensional spatiotemporal evolution manifold mapping and phase transition critical points specifically includes the following calculation and transformation sub-steps: S501, the condition diagnosis module extracts feature data corresponding to each load transient event and constructs a three-dimensional physical state space. Let... This is a discrete index for the time series of recorded transient events. For the first... In this event, the system selects the equivalent excitation force peak value, displacement peak value, and high-frequency distortion rate characteristic quantities, and combines them into an original state vector. Different physical characteristics have inherent differences in dimensions and numerical ranges; if geometric distances are directly calculated in multidimensional space, smaller feature dimensions are easily masked. The state diagnosis module performs standardization processing on the original state vector to generate a dimensionless three-dimensional coordinate vector. In its implementation, the system uses the mean and standard deviation of each feature dimension under the transformer's historical health status to linearly shift and scale the current original feature components, mapping them to a unified data distribution range. As a preferred method, the coordinate axes... Corresponding to the peak value of the equivalent excitation force after standardization, The corresponding peak-to-peak displacement value after standardization. This corresponds to the standardized high-frequency distortion rate characteristic. This explicit dimensional mapping establishes the physical benchmark for state-space analysis.

[0075] S502, grid disturbances in the transformer's operating environment can easily cause outlier fluctuations in the feature vector of a single event. Directly connecting discrete spatial points into an evolution trajectory can lead to the geometric features being corrupted by high-frequency background noise. The state diagnosis module uses an exponentially weighted moving average algorithm to smooth the manifold trajectory in the three-dimensional state space. The system calculates the... Smooth evolution trajectory coordinates of the secondary event The updated formula is: In the formula, This indicates the generation of a dimensionless three-dimensional coordinate vector; This is a smoothing coefficient used to adjust the weighting ratio between the current observation and the historical evolution trajectory; The smoothed coordinates correspond to the previous event. Considering the timescale attribute of the slow degradation of the transformer's mechanical state, the smoothing coefficient is chosen as a preferred method. The value is set between 0.1 and 0.2 to suppress measurement abrupt noise while preserving the system's long-term true evolution trend. For the initial state... The system defaults to setting it to the average characteristic coordinates under known healthy operating conditions.

[0076] S503, smoothed evolution trajectory coordinates A continuous manifold evolution curve is plotted in three-dimensional space. When the winding is in a stable, tight state, this curve exhibits a random walk within a small range in space. Once the tight state crosses the critical point and enters the loose phase, the trajectory shows a significant deflection in the direction of displacement response and high-frequency distortion mapping. The state diagnosis module quantifies the degree of this deflection by calculating the order of magnitude of the local spatial curvature of the discrete manifold curve. The system extracts the smoothed coordinates of three consecutive transient events and calculates the phase transition critical exponent. : In the formula, This represents the spatial difference vector between the smoothed coordinates of the current event and the previous event; The spatial difference vector representing the smoothed coordinates of the previous event is specifically calculated as follows: The operator × represents the outer product operation of vectors in three-dimensional space. This represents the Euclidean norm 2 of a vector. A regularization constant is used to prevent denominator overflow. When the winding state is in a stable period, the high overlap of trajectory points causes the difference vector norm to approach zero. The system adjusts the regularization constant based on the precision of floating-point operations. Set to 10 -6 This eliminates the algorithm divergence dead zone in the curvature calculation process.

[0077] S504, after obtaining the phase transition critical exponent sequence, the state diagnosis module initiates the critical point capture logic. The system will then calculate the critical point... With the preset critical threshold Perform a comparison. This critical threshold. The benchmark values ​​are derived from the baseline characteristic data of the transformer during its initial commissioning. The condition diagnosis module calculates the corresponding time curvature sequence from the historical characteristic data of the six months prior to transformer commissioning, extracts the mean of this sequence, and adds three times the standard deviation as the benchmark. When an inequality is detected Upon establishment, the system determines that the transformer winding state has entered the nonlinear relaxation stage and marks the current event index. This represents the critical point of phase transition leading to state degradation. Considering the local oscillations in the trajectory near the critical point, the state diagnosis module activates an event masking dead zone after triggering a valid capture. Within the continuous events covered by this masking dead zone, the system only performs coordinate update calculations and suspends threshold comparison and judgment. To prevent the algorithm from permanently getting stuck in the masked state, the length of this dead zone is set to a preset fixed number of events. As a preferred approach, the number of masking events is set to 5 to 10 historical transient events. This anti-jitter and limited masking mechanism ensures the stability of the state assessment conclusion output and the long-term monitoring capability of the system.

[0078] After capturing the critical phase transition point of transformer winding degradation, the diagnostic system needs to further quantify the cumulative damage caused by structural relaxation. In the three-dimensional physical state space, the multidimensional feature mapping points under healthy operating conditions typically cluster within a finite geometric envelope. When the mechanical fastening force crosses the critical point and enters the loosening phase state, the feature mapping points triggered by subsequent load transient events gradually deviate from this reference range, forming an outward-expanding evolutionary manifold trajectory in the state space. By calculating the geometric volume swept and enclosed by this distorted trajectory relative to the healthy reference envelope, the system can synthesize the degradation magnitude of the three-dimensional physical features, providing a quantitative assessment index for the transformer's health status.

[0079] In this embodiment, the integral calculation process for spatial distortion volume specifically includes the following calculation and transformation sub-steps: S505, the state diagnosis module constructs a baseline space under historical healthy operating conditions, using the marked phase transition critical point index as the boundary. Let the discrete index of the phase transition critical point locked by the system be... The status diagnosis module extracts indexes ranging from 0 to... Coordinates of all historical smooth evolutionary trajectories between These points form a set of health status feature points. For this set of points, the system calculates its geometric centroid coordinates in three-dimensional space. Specifically, this is calculated as the arithmetic mean of all historical health trajectory points on each coordinate axis. Subsequently, the system calculates the minimum convex polyhedron volume enclosing the set of health points, defining it as the baseline envelope volume. For calculating the volume of the convex hull of a discrete point set in three-dimensional space, those skilled in the art can rely on the fast convex hull algorithm in computational geometry. The topological patch construction rules and volume calculation steps are well-known technologies in this field and will not be elaborated here.

[0080] S506, after determining the healthy baseline centroid and envelope volume, the state diagnosis module initiates a micro-element volume partitioning mechanism for the distorted trajectory. The system starts from the phase transition critical point. Begin by extracting continuously occurring events along the forward timeline. The coordinates of the smoothed evolution trajectory of a transient event. As a preferred method, this assesses the event span. The number of iterations is typically set to 10 to 20 to ensure sufficient evolutionary nodes are collected to support spatial geometric operations. In the three-dimensional state space, the system utilizes the coordinates of three adjacent evolutionary trajectories and the centroid of the health baseline. Together they form a spatial infinitesimal tetrahedron. Let the local sliding index of the extracted evolution trajectory be... Its value range is from to .

[0081] S507, the state diagnosis module calculates the volume of the infinitesimal tetrahedron element one by one according to the rules of vector algebra and performs discrete cumulative integration. According to the principles of spatial analytic geometry, the volume of a tetrahedron formed by four vertices is equal to one-sixth of the absolute value of the triple product of the three spatial vectors sharing the same vertex. The system uses the reference centroid... As a reference vertex, calculate the... The volume of a tiny tetrahedron The calculation formula is: In the formula, , and These are three consecutive spatial coordinate points on the evolutionary trajectory; Represents the geometric centroid coordinates of a three-dimensional point set under a healthy baseline state; symbol The symbol × represents the inner product operation of vectors; the symbol × represents the outer product operation of vectors. This indicates the operation of taking the absolute value. Considering the geometric degradation where adjacent evolutionary trajectory points are approximately collinear, the volume of a single infinitesimal element may be smaller due to floating-point errors. The system calculates each... Perform a validity check; if it is lower than the preset volumetric noise floor threshold (usually set to 10), it will be considered valid. -5 If the volume of the infinitesimal element is zero, then the volume of that infinitesimal element is forcibly set to zero. After completing a single volume infinitesimal element calculation, the state diagnosis module performs discrete accumulation within the set evaluation event span to obtain the absolute spatial distortion volume of the distortion trajectory relative to the reference centroid in three-dimensional space. The formula for summation of integrals is: In the formula, Represents the absolute spatial distortion volume obtained from the cumulative calculation; For the local sliding index of the evolution trajectory used as the summation variable; This represents the discrete index of the phase transition critical point, corresponding to the lower limit of the summation interval; To assess the duration of the event; Then it means the first The scalar volume of a spatial infinitesimal tetrahedron.

[0082] In step S508, to transform the calculated absolute spatial volume into an evaluation metric with general engineering guidance significance, the condition diagnosis module performs a normalization mapping to generate a cumulative damage index. The system converts the absolute spatial distortion volume... Compared with the aforementioned baseline envelope volume Perform ratio calculations to determine the dimensionless cumulative damage. : In the formula, To prevent computational overflow caused by the reference convex hull volume approaching zero due to high overlap of historical feature points, a constant is typically set to 10, taking into account the floating-point precision of 3D feature vectors. -4 Cumulative damage This reflects the relative severity of the current winding condition deviating from the healthy baseline. The condition diagnosis module has built-in multi-level early warning logic. When a warning is detected... When the minor damage threshold is exceeded, the system generates an early loosening warning log. When the severe damage threshold is exceeded, the system determines that the winding clamping force has been severely lost, triggering a high-priority hardware interrupt to prompt the upper-level main control unit to arrange transformer shutdown and maintenance. By introducing a multi-level threshold mapping mechanism, the executability of diagnostic conclusions in the entire life cycle operation and maintenance management of the transformer is ensured.

[0083] After quantifying damage accumulation in high-dimensional space, the diagnostic system needs to transform multi-dimensional feature parameters into equipment physical state conclusions that are easy for engineers to understand. The mechanical degradation mechanism of transformer windings involves multiple modes, such as overall clamping force loss and local pad wear, and relying on a single feature is prone to false alarms. The condition diagnosis module introduces a multi-dimensional feature collaborative judgment mechanism, integrating macroscopic dynamic flexibility, microscopic high-frequency friction distortion, and long-term cumulative damage volume to establish a definite diagnostic mapping rule.

[0084] In this embodiment, the output process of the collaborative diagnostic criteria for the winding mechanical condition specifically includes the following calculation and conversion sub-steps: S509, the state diagnosis module extracts the peak value of the equivalent excitation force synchronously captured by the current transient event. With macroscopic displacement peak value To reduce the absolute numerical impact of load transient amplitude fluctuations on displacement response, the system calculates the dynamic equivalent compliance coefficient of the winding structure under the current state. The specific calculation formula is as follows: In the formula, This represents the peak-to-peak value of the winding vibration displacement after standardization. The peak value of the equivalent excitation force extracted synchronously; To prevent calculation divergence when the excitation force is small, a regularization constant is used. As a preferred approach, considering the mechanical dimensions of the transformer, this constant is typically set to 10% of the excitation force measurement range.-5 The coefficient mentioned above reflects the deformation amplitude of the winding under a unit electromagnetic excitation force. When the overall preload of the winding decreases, the mechanical stiffness of the structure decreases, which usually leads to a corresponding increase in this dynamic equivalent compliance coefficient.

[0085] S510, after obtaining the dynamic equivalent compliance coefficient, the state diagnosis module combines the previously extracted high-frequency distortion rate feature. With cumulative damage Construct a three-dimensional state determination Boolean vector The system sets independent state thresholds for each feature dimension for binary evaluation. Specifically, when the dynamic equivalent compliance coefficient... Greater than the flexibility safety threshold Time, endow ,otherwise When the high-frequency distortion rate characteristic quantity Greater than the effective friction threshold Time, endow ,otherwise When the cumulative damage Greater than the mild injury threshold Time, endow ,otherwise 0. The above-mentioned threshold values ​​are calibrated based on type test data of transformers of the same model at the time of manufacture and baseline data of similar equipment in operation. For example, the flexibility safety threshold. It is typically set between 1.5 and 2.0 times the average baseline flexibility of the transformer in the first month of operation to accommodate the manufacturing dispersion of different batches of equipment.

[0086] S511, the state diagnosis module establishes a collaborative diagnosis rule matrix based on Boolean vectors and performs physical mapping of discrete states. The system has built-in feature encoding tables for multiple typical mechanical states, and uses the currently generated Boolean vectors... Perform a bit-by-bit matching with the encoding table. When the vector is... When this occurs, the system outputs a diagnostic conclusion that the mechanical fastening condition of the winding is healthy. If the vector matching is... or This indicates that the macroscopic flexibility increased but did not induce high-frequency dry friction, and the system outputs an overall pressure relaxation warning. When the vector is presented as or When the system determines that there is wear and microscopic relative slippage in a local insulating pad, it outputs a warning of local pad loosening accompanied by friction. For vector presentation as... or In this scenario, the current macroscopic stiffness has not yet decreased, but the spatial evolution trajectory reflects that historical accumulated damage has exceeded the limit or is accompanied by occasional friction. Based on this, the system outputs an early gradual degradation warning to cover all state combinations in a closed loop. Once the vector is in a fully set state... This means that the transformer windings have become macroscopically loose, accompanied by internal dry friction and long-term spatial trajectory deterioration. The system then outputs a high-risk warning criterion for the detachment of insulation pads.

[0087] In step S512, after generating specific collaborative diagnostic criteria, the status diagnostic module serializes and encapsulates the criterion conclusion, trigger timestamp, and corresponding original characteristic parameters. The system sends the encapsulated diagnostic report message to the substation's upper-level main control unit or remote operation and maintenance platform via a standard industrial communication bus to trigger corresponding production scheduling and maintenance processes. For the data packet encapsulation and network transmission protocol parsing of the industrial communication bus, those skilled in the art can implement it using the standard Modbus or IEC61850 protocol stack. The data packet encapsulation and communication handshake interaction mechanism are well-known technologies in the field and will not be elaborated upon here.

[0088] Specific application examples: In a practical application scenario, taking a 110kV oil-immersed power transformer that has been in operation for 5 years in a substation as an example, this transformer has long been responsible for supplying power to a downstream industrial park, with frequent load fluctuations. The system adopts the latest deployed data acquisition hardware, with a synchronous sampling rate of... It is configured to 50kHz to meet the acquisition requirements of high-frequency frictional acoustic emission signals (>10kHz).

[0089] During implementation and operation, the following operating conditions may occur: Operating Condition 1: Transfer Function Identification under Light Load Steady State From 2:00 AM to 3:00 AM, the industrial park shuts down, and the transformer load current drops to 4% of its rated value (meeting the requirements). The system acquires phase voltage and tank wall vibration acceleration during this period. An NLMS adaptive filtering algorithm is used (filter order set to 400, step size factor...). After approximately 8 minutes, steady-state error convergence was achieved, and the spatial mechanical transfer parameter matrix from the iron core to the oil and then to the tank wall was successfully extracted and locked. .

[0090] Operating Condition 2: Temperature Damping Compensation and Load Transient Triggering At 8:00 AM, a large rolling mill in the industrial park starts up, and the effective value of the load current jumps from 150A to 420A within 2 seconds (triggering transient threshold). At this point, the top oil temperature sensor read 65℃ (equivalent to an equivalent average absolute temperature of approximately 328.15K). The system then calculates the relative viscosity coefficient using the Arrhenius equation. (Compared to the 20℃ baseline), the amplitude of the decoupled winding vibration component was amplified and compensated, reducing the vibration measurement error caused by the decrease in fluid damping due to high temperature.

[0091] Working Condition 3: Multidimensional Feature Extraction and Collaborative Diagnosis Within the extracted data window containing 2 seconds before and after the impact: Excitation force and displacement: Extract the peak value of the equivalent excitation force and perform a second-order frequency domain integral to obtain the peak-to-peak value of the macroscopic displacement. The dynamic equivalent compliance coefficient was calculated and found to be 1.6 times higher than the benchmark value.

[0092] High-frequency distortion rate: For high-frequency signals above 10kHz, the discrete energy operator is calculated to identify energy spikes; this is the characteristic quantity of high-frequency distortion rate. (far exceeding the threshold of 3.0), indicating that the macroscopic displacement process was accompanied by the frictional slippage of the microscopic insulating pad.

[0093] Spatial volume integral: Mapping the current coordinates into a three-dimensional feature space, calculations revealed a sharp increase in the curvature of the evolution trajectory space (triggering a critical phase transition point), and a distortion envelope volume. It reaches 4.2 times the healthy baseline volume.

[0094] Diagnostic result output: Comprehensive Boolean vector The system output a high-risk warning of loose insulation pads accompanied by micro-friction, prompting a shutdown for spring maintenance and troubleshooting. Subsequent unpacking and lifting inspection confirmed a 2mm gap and signs of friction carbonization in the clamping pads at the upper end of the C-phase high-voltage winding, verifying the system's accuracy.

[0095] Experimental verification and effect comparison: See attached document Figure 3 To verify the advanced nature of the present invention, a comparative test was conducted on the 10kV transformer dynamic model platform of the State Grid Corporation's High Voltage Key Laboratory.

[0096] The experiment was conducted using a hydraulic jack and pad wear simulation device to artificially create three transformer winding states: State 1 (healthy pre-tightening), State 2 (early slight loosening), and State 3 (severe loosening with friction). Different gradients of transient load impact were applied to the model transformer within a variable temperature cycling chamber ranging from -10℃ to 80℃. A total of 600 test samples were collected.

[0097] Comparison Algorithms: Traditional Method A: Vibration energy method based on pure frequency domain analysis (threshold alarm, without separating the core and winding).

[0098] Traditional Method B: Steady-state decoupling method based on current square-vibration transfer function (without temperature and viscosity compensation and high-frequency friction analysis).

[0099] The method of this invention is: multi-source decoupling + Arrhenius temperature compensation + three-dimensional manifold volume diagnosis.

[0100] Comparison results data table: Conclusion: Experimental data fully demonstrate that the system and method provided by this invention have engineering advantages in terms of resistance to temperature-induced interference, sensitivity to early detection of minor loosening, and forward-looking fault warning.

Claims

1. A fault diagnosis method for an oil-immersed transformer, characterized in that, Includes the following steps: The discrete sequences of the operating phase voltage, load current, top oil temperature, and tank wall acceleration of the transformer are collected synchronously. Under light load steady-state conditions, adaptive filtering is performed based on the discrete sequence of operating phase voltages and the sequence of box wall accelerations to obtain the locked transfer function parameter matrix; Under dynamic operating conditions, the core vibration reference component is reconstructed using the locked transfer function parameter matrix, and the winding vibration component is extracted from the differential decoupling from the box wall acceleration sequence. The relative viscosity coefficient is calculated based on the discrete sequence of the top oil temperature and a damping compensation factor is generated. The damping compensation factor is then applied to the extracted winding vibration components to obtain a standardized winding vibration signal. When the rate of change of the discrete sequence of the load current exceeds a set threshold, a time window is extracted. Within the time window, the excitation force characteristics are extracted based on the discrete sequence of the load current, and the macroscopic displacement characteristics and high-frequency distortion rate characteristics are extracted based on the standardized winding vibration signal, respectively. The excitation force characteristics, macroscopic displacement characteristics, and high-frequency distortion rate characteristics are mapped to a three-dimensional state space. The spatial distortion volume is calculated and compared with a preset benchmark threshold to output the diagnostic results of the transformer winding mechanical state.

2. The method for fault diagnosis of an oil-immersed transformer according to claim 1, characterized in that, The transfer function parameter matrix for acquiring the lock includes: The square of the discrete sequence of the running phase voltage is calculated to construct the real-time input sequence, and the box wall acceleration sequence is used as the desired output sequence. The filter tap weight column vector is updated using the normalized minimum mean square error algorithm; When the moving average energy of the transient estimation error is continuously lower than the steady-state error threshold, the currently converged weight coefficients are locked as the locked transfer function parameter matrix.

3. The method for fault diagnosis of an oil-immersed transformer according to claim 1, characterized in that, Under dynamic operating conditions, the steps of reconstructing the core vibration reference components using the locked transfer function parameter matrix and extracting the winding vibration components from the box wall acceleration sequence through differential decoupling include: The square of the discrete sequence of the operating phase voltage under dynamic conditions is convolved with the locked transfer function parameter matrix in discrete time to generate the core vibration reference component. In the time domain, the box wall acceleration sequence is subtracted from the core vibration reference component point by point to extract the winding vibration component by difference. The winding vibration component is subjected to bidirectional bandpass filtering to obtain the filtered winding vibration component.

4. The method for fault diagnosis of an oil-immersed transformer according to claim 1, characterized in that, The steps of calculating the relative viscosity coefficient and generating a damping compensation factor based on the discrete sequence of the top oil temperature, and applying the damping compensation factor to the extracted winding vibration components to obtain a standardized winding vibration signal include: The top-level oil temperature discrete sequence is combined with the equivalent average oil temperature discrete sequence to convert it into an absolute temperature sequence. The relative viscosity coefficient of the absolute temperature sequence relative to the reference absolute temperature is calculated based on the Arrhenius equation and used as the damping compensation factor. The vibration components of the winding are multiplied by the damping compensation factor to obtain the temperature damping compensated vibration sequence. Calculate the effective value of the discrete sequence of load current within the current analysis time window, and use the square of the effective value as a normalization factor to perform a division scaling operation on the temperature damping compensated vibration sequence to generate an amplitude-normalized vibration sequence as the normalized winding vibration signal.

5. The method for fault diagnosis of an oil-immersed transformer according to claim 1, characterized in that, The process of extracting a time window when the rate of change of the discrete sequence of load current exceeds a set threshold includes: The continuously generated standardized winding vibration signal and the discrete sequence of load current are written into a ring-shaped data buffer of fixed depth in real time. The effective value sequence of sliding current is continuously calculated according to the set sliding time window, and the first-order forward differential sequence is calculated to monitor the instantaneous slope of grid load fluctuations. When the first-order forward difference sequence exceeds the load transient threshold for multiple consecutive sampling periods, a hardware interrupt is generated and the discrete time point of the first over-limit is locked as the transient zero-point index. Historical data with a length equal to the number of preceding sampling points is read back from the circular data buffer, and the latest data with a length equal to the number of subsequent sampling points is continuously collected. The data is then spliced ​​together to obtain the complete time window. After the data is extracted, the dead zone shielding time is activated to pause the over-limit detection.

6. The method for fault diagnosis of an oil-immersed transformer according to claim 1, characterized in that, The process of extracting excitation force features based on the discrete sequence of load current and extracting high-frequency distortion rate features based on the standardized winding vibration signal includes: An equivalent electromagnetic excitation force sequence is constructed by performing point-by-point squaring operations on the discrete sequence of load current within the time window, and the peak value of the equivalent electromagnetic excitation force sequence is extracted as the excitation force feature. The standardized winding vibration signal is subjected to Fourier transform to the frequency domain. Based on the set low-frequency truncation threshold, a second-order frequency domain integral operation is performed, and then the signal is restored by inverse Fourier transform to generate a macroscopic displacement time discrete sequence. The peak-to-peak value of the macroscopic displacement time discrete sequence is extracted as the macroscopic displacement feature.

7. The method for fault diagnosis of an oil-immersed transformer according to claim 1, characterized in that, The process of extracting high-frequency distortion rate features based on the standardized winding vibration signal includes: The box wall acceleration sequence within the time window is subjected to high-pass filtering to separate the high-frequency sequence of frictional acoustic emission; The high-frequency sequence of triboacoustic emission is processed using a discrete energy operator to generate a nonlinear instantaneous energy sequence that reflects the dual abrupt changes in amplitude and frequency. Calculate the average high-frequency background energy within the transient zero-point pre-steady-state interval, calculate the relative error based on the nonlinear instantaneous energy sequence and the average high-frequency background energy, and generate a dynamic high-frequency energy distortion rate sequence. The global maximum value of the high-frequency energy distortion rate sequence is extracted as the high-frequency distortion rate feature.

8. The method for fault diagnosis of an oil-immersed transformer according to claim 1, characterized in that, The process of mapping the excitation force characteristics, the macroscopic displacement characteristics, and the high-frequency distortion rate characteristics to a three-dimensional state space includes: Using the mean and standard deviation of historical health condition feature data, linear translation and scaling standardization are performed on the currently extracted excitation force feature, macroscopic displacement feature and high frequency distortion rate feature to generate a dimensionless three-dimensional coordinate vector. An exponentially weighted moving average algorithm is used to calculate the weight allocation between the current three-dimensional coordinate vector and the historical evolution trajectory, and to update and generate smooth evolution trajectory coordinates, thereby constructing a manifold evolution curve in the three-dimensional state space.

9. The method for fault diagnosis of an oil-immersed transformer according to claim 1, characterized in that, The process of calculating the spatial distortion volume includes: The phase transition critical index is calculated using the outer product operation of the spatial difference vectors of the continuous smooth evolution trajectory coordinates. When the phase transition critical index exceeds the set critical threshold, the corresponding transient event index is marked as the phase transition critical point. The reference centroid is calculated using the coordinate set of the historical healthy smooth evolution trajectory prior to the phase transition critical point; Starting from the phase transition critical point, the coordinates of three adjacent smooth evolution trajectories and the reference centroid are used to form a spatial infinitesimal tetrahedron. Discrete cumulative integration is performed on the scalar volume of multiple continuous spatial infinitesimal tetrahedrons within the evaluation event span to obtain the spatial distortion volume.

10. The method for fault diagnosis of an oil-immersed transformer according to claim 1, characterized in that, The process of calculating the spatial distortion volume and comparing it with a preset benchmark threshold to output the diagnostic results of the transformer winding mechanical condition includes: The ratio of the macroscopic displacement feature to the excitation force feature is calculated to generate a dynamic equivalent compliance coefficient. The cumulative damage index is generated by calculating the ratio between the spatial distortion volume and the reference envelope volume. The dynamic equivalent compliance coefficient, the high-frequency distortion rate feature, and the cumulative damage index are compared with the set independent physical thresholds in a binary form to generate a three-dimensional state determination Boolean vector. The three-dimensional state determination Boolean vector is input into a preset collaborative diagnosis rule matrix for bit-by-bit matching, and the winding mechanical state diagnosis results are output, indicating normal operation, overall loosening of clamping force, local loosening of pads, or risk of pad falling off.

Images