Sensor-based transformer adaptive fault diagnosis method, system and equipment

Abstract

The invention relates to the technical field of fault diagnosis, and discloses a transformer adaptive fault diagnosis method, system and equipment based on a sensor, and the method comprises the steps: collecting an operation signal of a transformer through a sensing device, carrying out the feature extraction of the operation signal, and obtaining a plurality of signal feature data; each signal feature data is input into each corresponding signal fault diagnosis model to obtain a plurality of fault diagnosis sub-results, and each signal fault diagnosis model is constructed by adopting a probabilistic neural network model; according to a preset weight value, weighted summation is carried out on the fault diagnosis sub-results, a fault diagnosis result of the transformer is obtained, and the weight value is obtained through calculation based on the diagnosis accuracy of each signal fault diagnosis model. According to the invention, comprehensive data processing and diagnostic analysis are carried out on various sensor signals, so that efficient and accurate diagnosis of the transformer fault can be realized, and thus a powerful guarantee is provided for stable operation of a power system.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and particularly to a transformer adaptive fault diagnosis method, system and device based on sensors. Background Art

[0002] In modern power systems, transformers, as the core hubs for electric energy conversion and transmission, their safe and stable operation is of crucial significance for ensuring the reliability and continuity of the entire power network. Once a transformer fails, it will have extremely serious impacts on both industrial production and residential life.

[0003] To ensure the safety and stability of power system operation, transformer fault diagnosis technology has always been a key research direction in the field of power engineering. By detecting and resolving potential faults in a timely manner, the reliable operation of transformers is ensured. Currently, the existing transformer fault diagnosis technologies mainly include dissolved gas analysis in oil technology, electrical test analysis technology, vibration analysis technology, etc. However, these existing diagnosis technologies all have certain limitations. On the one hand, although the dissolved gas analysis in oil technology can judge faults based on the components and contents of dissolved gases in oil, the gas dissolution and diffusion processes are slow, the detection results are lagging, and it is impossible to respond to sudden faults in a timely manner. The processes of oil sample collection, transportation and analysis are cumbersome, requiring professional equipment and personnel, with a long detection cycle, and it is difficult to achieve real-time monitoring. Moreover, external factors such as oil temperature, oil flow velocity, and tank sealing conditions are likely to affect the gas solubility and diffusion rate, resulting in deviations in diagnosis results. On the other hand, although the electrical test analysis technology covers various items such as insulation resistance test, direct current leakage current test, dielectric loss tangent test, turns ratio test, etc., and can measure electrical parameters to evaluate the transformer status under a power-off state. However, these tests are complex and time-consuming. The power-off operation not only affects normal power supply but also impacts the stability of the power system. The test results are easily affected by factors such as environmental humidity, temperature, and electric field interference. There may be mutual interference between different tests, and professional personnel are required to comprehensively analyze multiple test results, which is highly subjective and prone to misjudgment. In addition, the vibration analysis technology can collect signals using vibration sensors and judge faults based on vibration characteristics. However, vibration sensors are easily affected by external vibration sources such as nearby motors, mechanical vibrations, and electromagnetic interference, resulting in large noise and low signal-to-noise ratio in the collected signals, affecting the accuracy of fault feature extraction. Moreover, the internal structure of the transformer is complex, and the vibration signals of different fault types may be similar, increasing the difficulty of fault discrimination. It is difficult to comprehensively reflect the changes in the electrical and thermal performance of the transformer through a single vibration signal, and its fault diagnosis capabilities for insulation aging, partial discharge, etc. are limited. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a sensor-based adaptive fault diagnosis method, system, and device for transformers. By comprehensively analyzing various sensor signals, accurate and efficient transformer fault diagnosis is achieved, and the technical effects of improving the safety and stability of power system operation are realized.

[0005] In a first aspect, the present invention provides a sensor-based adaptive fault diagnosis method for transformers, the method comprising:

[0006] Collecting the operation signals of the transformer through a sensing device, extracting features from the operation signals to obtain a plurality of signal feature data, the signal feature data including vibration signal features, oil temperature signal features, load current signal features, and partial discharge signal features;

[0007] Respectively inputting each of the signal feature data into corresponding signal fault diagnosis models to obtain a plurality of fault diagnosis sub-results. Each signal fault diagnosis model is constructed using a probabilistic neural network model and trained using a sample data set of the corresponding signal type;

[0008] According to a preset weight value, performing weighted summation on the fault diagnosis sub-results to obtain a fault diagnosis result of the transformer, the weight value being calculated based on the diagnosis accuracy of each signal fault diagnosis model.

[0009] Further, the step of collecting the operation signals of the transformer through a sensing device, extracting features from the operation signals to obtain a plurality of signal feature data includes:

[0010] Collecting the operation signals of the transformer through a sensing device, the operation signals including vibration signals, oil temperature signals, load current signals, and partial discharge signals;

[0011] Performing time-domain analysis and frequency-domain analysis on the vibration signals to obtain time-domain features and frequency-domain features, and using the time-domain features and the frequency-domain features as vibration signal features. The time-domain features include peak value, mean value, and variance, and the frequency-domain features include energy distribution;

[0012] Calculating the oil temperature change rate and the oil temperature gradient according to the oil temperature signals at adjacent sampling moments, and using the oil temperature change rate and the oil temperature gradient as oil temperature signal features;

[0013] Calculating harmonic content features according to the harmonic amplitude and fundamental wave amplitude of the load current signal, calculating current unbalance degree features according to the three-phase currents of the load current signal, and using the harmonic content features and the current unbalance degree features as load current signal features;

[0014] Based on the discharge amount and the number of discharge pulses of the partial discharge signal per unit time, the discharge amount feature and the discharge number feature are obtained. Based on the number of discharge pulses of the partial discharge signal within the power frequency cycle, the discharge phase feature is obtained, and the discharge amount feature, the discharge number feature, and the discharge phase feature are used as partial discharge signal features.

[0015] Further, before the step of performing time-domain analysis and frequency-domain analysis on the vibration signal, the following steps are also included:

[0016] The wavelet transform algorithm is used to remove noise from the vibration signal to obtain a denoised vibration signal;

[0017] The moving average filtering algorithm is used to smooth the oil temperature signal to obtain a smoothed oil temperature signal;

[0018] The fast Fourier transform is used to perform time-frequency conversion and interference harmonic removal on the load current signal to obtain a pure load current signal;

[0019] The noise threshold is used to remove noise from the partial discharge signal to obtain a denoised partial discharge signal.

[0020] Further, the signal fault diagnosis model is constructed by using an improved probabilistic neural network model, and includes an input layer, a sample layer, a summation layer, and an output layer;

[0021] The input layer is used to calculate the sample distance between the input sample and the training sample by using an improved distance formula, and input the sample distance into the sample layer;

[0022] The sample layer is used to calculate the similarity between the input sample and the training sample by using a probability density formula based on a smoothing factor according to the sample distance, and input the similarity into the summation layer;

[0023] The summation layer is used to calculate the conditional probability density according to the similarity, and input the conditional probability density into the output layer;

[0024] The output layer is used to select the corresponding fault category as the output data according to the conditional probability density;

[0025] Among them, the sample distance is represented by the following formula:

[0026]

[0027] In the formula, X represents the input sample, X i represents the i-th training sample, m represents the number of features of the training sample, w j represents the weight of the j-th feature, k represents the distance exponent, x jrepresents the j-th eigenvalue of the input sample X, x ij represents the i-th training sample X i of the j-th eigenvalue;

[0028] The similarity is expressed by the following formula:

[0029]

[0030] In the formula, Φ i (X) represents the similarity between the input sample X and the training sample X i between, and σ represents the smoothing factor.

[0031] Furthermore, the smoothing factor is determined by an adaptive parameter optimization method, and the specific steps include:

[0032] Randomly extract a first sample set and a second sample set from the sample dataset, and traverse the preset search interval according to the preset step size to obtain a candidate factor sequence containing multiple candidate smoothing factors;

[0033] Iteratively calculate each candidate smoothing factor in the candidate factor sequence according to the model classification accuracy to obtain a candidate factor set, where the steps of one iteration include:

[0034] Take the signal fault diagnosis model using the candidate smoothing factor in the candidate factor sequence as the first diagnosis model, train the first diagnosis model with the first sample set, and calculate the first classification accuracy according to the output result of the trained first diagnosis model;

[0035] If the first classification accuracy is greater than the preset optimal accuracy, take the first classification accuracy as the optimal accuracy and add the candidate smoothing factor to the candidate factor set;

[0036] Take the signal fault diagnosis model using the candidate smoothing factor in the candidate factor set as the second diagnosis model, train each second diagnosis model with the second sample set respectively, and obtain the corresponding second classification accuracy according to the output result of each trained second diagnosis model;

[0037] Select the candidate smoothing factor corresponding to the maximum value from each of the second classification accuracies as the smoothing factor of the signal fault diagnosis model.

[0038] Furthermore, the step of selecting the corresponding fault category as the output data according to the conditional probability density includes:

[0039] Compare each of the conditional probability densities, and select the fault category corresponding to the maximum conditional probability density as the output data according to the comparison result.

[0040] Further, the step of selecting a corresponding fault category as output data according to the conditional probability density includes:

[0041] Calculating the posterior probability using Bayes' formula based on the conditional probability density and a preset prior probability, and calculating the conditional risk value based on the posterior probability and a preset risk assessment matrix;

[0042] Comparing the conditional risk values and selecting the fault category corresponding to the minimum conditional risk value as the output data.

[0043] Further, the calculation step of the weight value includes:

[0044] Inputting the sample data sets of each signal type into the corresponding signal fault diagnosis model for fault diagnosis respectively, and calculating the diagnosis accuracy of each signal fault diagnosis model according to the output result and the sample classification label;

[0045] Calculating the weight value of the fault diagnosis sub-result according to the diagnosis accuracy;

[0046] Secondly, the weight value is represented by the following formula:

[0047]

[0048] where ω i represents the weight value of the i-th fault diagnosis sub-result, Acc i represents the diagnosis accuracy of the i-th signal fault diagnosis model, and n represents the total number of signal fault diagnosis models.

[0049] In a second aspect, the present invention provides a transformer adaptive fault diagnosis system based on sensors, and the system includes:

[0050] A feature extraction module, configured to collect the operation signals of the transformer through a sensing device, preprocess and extract features from the operation signals to obtain a plurality of signal feature data, and the signal feature data includes vibration signal features, oil temperature signal features, load current signal features, and partial discharge signal features;

[0051] A classification diagnosis module, configured to input each of the signal feature data into the corresponding signal fault diagnosis models respectively to obtain a plurality of fault diagnosis sub-results, and each signal fault diagnosis model is constructed by using a probabilistic neural network model and trained by using the sample data sets of the corresponding signal types;

[0052] A comprehensive evaluation module is configured to perform weighted summation on the fault diagnosis sub-results according to preset weight values to obtain a fault diagnosis result of the transformer, where the weight values are calculated based on the diagnosis accuracy rates of each signal fault diagnosis model.

[0053] In a third aspect, an embodiment of the present invention further provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented.

[0054] The present invention provides a sensor-based adaptive fault diagnosis method, system, and device for transformers. By performing multi-dimensional real-time perception of the operating state of the transformer and efficient processing and in-depth feature mining of various signals, the present invention can obtain representative and discriminative fault features, thereby more accurately locating the root cause of the fault; by constructing a signal fault diagnosis model based on an improved probabilistic neural network and combining a risk assessment strategy, the accuracy of model diagnosis can be improved, and the risk of misjudgment can be effectively reduced; and by calculating the diagnosis accuracy rates of each signal to dynamically allocate weights, the fusion process can flexibly adjust the influence of each signal according to different fault scenarios, avoiding diagnostic biases caused by fixed weights, making the diagnostic result more consistent with the actual fault situation, and further improving the accuracy of the diagnostic result. Description of the Drawings

[0055] Figure 1 is a schematic flowchart of the sensor-based adaptive fault diagnosis method for transformers in an embodiment of the present invention;

[0056] Figure 2 is a schematic structural diagram of the sensor-based adaptive fault diagnosis system for transformers in an embodiment of the present invention;

[0057] Figure 3 is an internal structural diagram of the computer device in an embodiment of the present invention. Detailed Embodiments

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0059] Please refer to Figure 1 , a sensor-based adaptive fault diagnosis method proposed in the first embodiment of the present invention, which includes steps S10 to S30:

[0060] Step S10, collect the operation signals of the transformer through a sensing device, preprocess and extract features from the operation signals to obtain multiple signal feature data, where the signal feature data includes vibration signal features, oil temperature signal features, load current signal features, and partial discharge signal features;

[0061] Step S20, input each of the signal feature data into the corresponding signal fault diagnosis models respectively to obtain multiple fault diagnosis sub-results. Each signal fault diagnosis model is constructed using a probabilistic neural network model and trained using a sample data set of the corresponding signal type;

[0062] Step S30, perform weighted summation on the fault diagnosis sub-results according to the preset weight values to obtain the fault diagnosis result of the transformer. The weight values are calculated based on the diagnostic accuracy of each signal fault diagnosis model.

[0063] In the present invention, a pre-constructed sensing device accurately and comprehensively collects various key information during the operation of the transformer. The sensing device integrates various types of sensors such as vibration sensors, oil temperature sensors, load current sensors, and partial discharge sensors. Among them, the vibration sensor, relying on the piezoelectric effect, precisely converts the internal mechanical vibration of the transformer into an electrical signal for capturing the operation state of the mechanical structure; the oil temperature sensor, relying on the sensitive characteristic of platinum resistance to temperature, monitors the temperature change of the oil in real time, and its data is the key basis for judging thermal faults; the load current sensor uses the Hall effect to accurately measure the current situation entering and leaving the transformer, reflecting the electrical performance relationship between the load and the winding; the partial discharge sensor uses the capacitance principle to keenly capture the discharge phenomenon in the insulation structure, facilitating the early detection of insulation hidden dangers.

[0064] Specifically, a high-precision piezoelectric acceleration sensor is selected for the vibration sensor. Its working principle is based on the piezoelectric effect. When the sensor is vibrated to generate acceleration, the internal piezoelectric crystal will be subjected to mechanical stress and generate charges. According to the piezoelectric effect formula Q = dF (where Q is the charge quantity, d is the piezoelectric constant, and F is the acting force), and Newton's second law F = ma (where m is the mass and a is the acceleration), the relationship between the charge quantity Q and the acceleration a can be deduced as Q = dma. By measuring the charge quantity Q and knowing the piezoelectric constant d and mass m of the sensor, the vibration acceleration a can be calculated, and thus the vibration signal V(t) can be obtained.

[0065] The vibration sensor is precisely installed at key positions on the surface of the transformer box body, such as the core clamping parts, winding ends, etc., and in the vicinity of core components such as the core and winding, to ensure that the vibration details of the internal mechanical components of the transformer can be captured to the greatest extent. Whether it is a slight looseness of the winding or an abnormal vibration of the core, it can be keenly sensed, providing a key basis for diagnosing faults related to the mechanical structure.

[0066] For the oil temperature sensor, a high-precision platinum resistance temperature sensor is adopted. It utilizes the characteristic that the resistance value of the platinum resistance changes with temperature to accurately measure the oil temperature. The resistance value of the platinum resistance and temperature satisfy a specific temperature coefficient formula R t =R 0 (1 + αt) (where R t is the resistance value at temperature t, R 0 is the resistance value at the reference temperature, and α is the temperature coefficient). By measuring the resistance value R t of the platinum resistance, and knowing the resistance value R 0 at the reference temperature and the temperature coefficient α, the current oil temperature t can be calculated, thus obtaining the oil temperature signal T(t).

[0067] The oil temperature sensor is deeply installed inside the transformer oil tank and can reflect the temperature change of the transformer oil in real time and accurately. As an important indicator of the thermal state of the transformer, abnormal changes in the oil temperature often indicate potential faults such as overheating inside. Therefore, accurate oil temperature monitoring is crucial for timely detecting potential problems.

[0068] The load current sensor based on the Hall effect principle is installed at the input and output ends of the transformer to accurately measure the magnitude of the load current. When current I passes through the Hall element placed in magnetic field B, a Hall voltage U H will be generated in the direction perpendicular to the current and the magnetic field. Its magnitude is proportional to the current I, the magnetic field strength B, and the sensitivity coefficient K H of the Hall element, that is, U H =K H IB. By measuring the Hall voltage U H , and knowing the magnetic field strength B and the sensitivity coefficient K H of the Hall element, the load current I can be calculated, thus obtaining the load current signal I(t). The change of the load current is closely related to the load condition of the transformer and the electrical performance of the internal windings, and is of great significance for judging whether there are electrical faults such as short circuits and overloads in the windings.

[0069] For the partial discharge sensor, a high-sensitivity capacitive partial discharge sensor is selected. It utilizes the change of the electric field between the capacitor plates to capture the instantaneous charge change generated by partial discharge. According to the definition formula of capacitance C = Q / V (where C is the capacitance value, Q is the charge quantity, and V is the voltage), when partial discharge occurs, the instantaneous change of charge will cause a change in the voltage between the capacitor plates. By detecting the voltage change, the change of the charge quantity is indirectly measured, thereby obtaining the partial discharge signal PD(t).

[0070] The partial discharge sensor is placed near the internal insulation structure of the transformer and can promptly detect the possible partial discharge phenomena inside the transformer. Partial discharge is an important early sign of the deterioration of the transformer insulation. Early detection and monitoring of partial discharge play a crucial role in preventing the occurrence and development of insulation faults.

[0071] Through the above-mentioned sensing device, key operating signals such as vibration signals, oil temperature signals, load current signals, and partial discharge signals during the operation of the transformer can be collected. Then, by extracting features and analyzing these data, it can be determined whether there are potential faults in the transformer.

[0072] In a preferred embodiment, in order to make the extracted features more accurate and reliable, before extracting features from the operating signals, the operating data is first preprocessed to achieve in-depth optimization and purification of the original signals. The specific preprocessing steps include:

[0073] The wavelet transform algorithm is used to remove noise from the vibration signal to obtain a denoised vibration signal;

[0074] The moving average filtering algorithm is used to smooth the oil temperature signal to obtain a smoothed oil temperature signal;

[0075] The fast Fourier transform is used to perform time-frequency conversion and remove interfering harmonics from the load current signal to obtain a pure load current signal;

[0076] The noise threshold is used to remove noise from the partial discharge signal to obtain a denoised partial discharge signal.

[0077] In this embodiment, for the above four different types of sensor data, different preprocessing methods are adopted. For the vibration signal, the wavelet transform method is used for denoising. Preferably, the Daubechies wavelet basis function db4 with good time-frequency localization characteristics is used to perform wavelet decomposition on the vibration signal V(t). Through wavelet decomposition, the signal is decomposed into a low-frequency approximation part (represented by the scaling coefficient) and a high-frequency detail part (represented by the wavelet coefficient). Since the energy of the noise signal is mainly concentrated in the high-frequency part, the wavelet coefficients are refined by setting a threshold λ to remove the wavelet coefficients corresponding to the noise. Finally, the denoised vibration signal V d (t) is obtained through wavelet reconstruction. After the wavelet transform denoising process, the signal-to-noise ratio of the vibration signal is significantly improved, laying a solid foundation for subsequent accurate feature extraction and fault diagnosis.

[0078] For the oil temperature signal, a moving average filtering method is used for smoothing. Assuming that the sampling sequence of the oil temperature signal is T(n) (n = 1, 2, …, N), and the window length of the moving average is M (accurately calculated according to the change frequency and sampling frequency of the oil temperature signal, generally taking 5 to 10 sampling points), then the smoothed oil temperature signal T s (n) is calculated by the formula:

[0079]

[0080] By calculating the average value of the oil temperature values within the window as the smoothed value of the current sampling point, random fluctuations and short-term interferences in the oil temperature signal can be effectively eliminated, and a smoother and more stable oil temperature signal T s (t) can be obtained, which helps to more accurately monitor the change trend of the oil temperature and timely detect potential fault signs such as abnormal increase or fluctuation of the oil temperature.

[0081] For the load current signal I(t), a fast Fourier transform (FFT) is performed to convert the time-domain signal into a frequency-domain signal for analyzing its spectral characteristics. By analyzing the spectral characteristics of the frequency-domain signal through FFT, the fundamental wave and harmonic components in the load current can be clearly identified. Then, a strict harmonic threshold is set according to the harmonic content range during the normal operation of the transformer, and the harmonic components with amplitudes exceeding the harmonic threshold in the frequency domain are filtered out to obtain a pure load current signal I p (t). The load current signal after removing harmonic interference can more accurately reflect the load characteristics of the transformer and the electrical state of the winding, providing a more reliable basis for judging whether there are faults such as short circuits, overloads, and core saturation in the winding.

[0082] For the partial discharge signal, a reasonable threshold is set according to its pulse characteristics for noise removal. Assuming that the sampling sequence of the partial discharge signal is PD(n) (n = 0, 1, …, N - 1), through in-depth statistical analysis of the signal amplitude, the threshold PD th is determined. Preferably, the mean plus three times the standard deviation is used as the threshold, PD th = μ PD + 3σ PD where μ PD is the signal mean and σ PD is the signal standard deviation. Signals below the threshold are regarded as background noise and removed to obtain the processed partial discharge signal PD p (t). After threshold processing, the noise in the partial discharge signal is effectively suppressed, highlighting the real partial discharge pulse signal, which helps to more accurately detect and analyze the characteristics of partial discharge and provides an important basis for judging the type, severity, and development trend of partial discharge inside the transformer insulation.

[0083] In this embodiment, by preprocessing the sensor signals, the quality of each signal is significantly improved, thus providing a more accurate and reliable data basis for subsequent feature extraction and fault diagnosis.

[0084] In a preferred embodiment, for the collected sensor data, the present invention uses a variety of algorithms to mine the fault features in the signals from different angles. The specific steps include:

[0085] Collect the operation signals of the transformer through the sensing device, and the operation signals include vibration signals, oil temperature signals, load current signals, and partial discharge signals;

[0086] Perform time-domain analysis and frequency-domain analysis on the vibration signal to obtain time-domain features and frequency-domain features, and use the time-domain features and the frequency-domain features as vibration signal features. The time-domain features include peak value, mean value, and variance, and the frequency-domain features include energy distribution;

[0087] According to the oil temperature signals at adjacent sampling times, calculate the oil temperature change rate and the oil temperature gradient, and use the oil temperature change rate and the oil temperature gradient as oil temperature signal features;

[0088] According to the harmonic amplitude and fundamental wave amplitude of the load current signal, calculate the harmonic content feature, and according to the three-phase currents of the load current signal, calculate the current unbalance degree feature, and use the harmonic content feature and the current unbalance degree feature as load current signal features;

[0089] According to the discharge amount and the number of discharge pulses of the partial discharge signal within a unit time, obtain the discharge amount feature and the discharge number feature, and according to the number of discharge pulses of the partial discharge signal within the power frequency period, obtain the discharge phase feature, and use the discharge amount feature, the discharge number feature, and the discharge phase feature as partial discharge signal features.

[0090] In this embodiment, for different sensor signals, different feature extraction methods are adopted. Specifically, for the vibration signal, it is analyzed in the time domain and the frequency domain respectively to extract the time-domain features and frequency-domain features of the signal. The time-domain features here include the peak value P of the vibration signal V , mean value μ V and variance σ V . These time-domain features can reflect the amplitude change characteristics of the vibration signal and are closely related to the mechanical state of the internal components of the transformer. For example, the change of the peak value may indicate faults such as winding looseness or increased core vibration.

[0091] The frequency-domain features refer to obtaining the spectrum through the FFT transformation of the vibration signal. According to the transformer fault characteristic frequency range, the spectrum can be divided into multiple frequency bands, such as the low-frequency band (0 - 100 Hz), the medium-frequency band (100 - 500 Hz), and the high-frequency band (above 500 Hz). By calculating the energy in each frequency band, the energy distribution E of the vibration signal can be obtained. V , thus effectively extracting the frequency component information related to different fault types in the vibration signal. It should be noted that the time-domain analysis and frequency-domain analysis of the signal in this embodiment can refer to the conventional signal analysis steps, which will not be elaborated here one by one.

[0092] For the oil temperature signal, the oil temperature change rate R and the oil temperature gradient G are calculated through the oil temperature T(n) and T(n - 1) at adjacent sampling times and the sampling time interval Δt. s (n) and T s (n - 1), as well as the sampling time interval Δt, to calculate the oil temperature change rate R T and the oil temperature gradient G T , where the oil temperature change rate can be expressed as:

[0093]

[0094] The oil temperature gradient can be expressed as:

[0095] G T =T s (n)-T s (n - 1)

[0096] The oil temperature change rate and the oil temperature gradient can intuitively reflect the speed and trend of the oil temperature change, which is of great significance for judging the development process of internal thermal faults in the transformer. For example, a sudden increase in the oil temperature change rate or an abnormal change in the oil temperature gradient may indicate problems such as overheating faults or poor heat dissipation inside the transformer.

[0097] For the load current signal, through FFT analysis, the harmonic amplitudes of each order are calculated, and the harmonic amplitudes are compared with the fundamental wave amplitude to obtain the harmonic content feature H I , and its formula is expressed as:

[0098]

[0099] In the formula, A m represents the mth harmonic amplitude, and A 1 represents the fundamental wave amplitude.

[0100] By analyzing the harmonic content feature, it can be judged whether there are faults such as non-linear loads, short circuits, and core saturation in the transformer winding, because these faults often lead to changes in the harmonic content in the load current.

[0101] In addition, according to the three-phase currents I in the load current signala 、I b 、I c , calculate the current unbalance degree feature:

[0102]

[0103] The unbalance degree can accurately reflect whether there are faults such as inter-turn short circuits and three-phase load unbalances in the transformer winding. When the current unbalance degree feature exceeds the normal range, it may indicate that there is a fault in the transformer winding or the three-phase load distribution is uneven, and further inspection and processing are required.

[0104] Regarding the partial discharge signal feature, the discharge quantity feature Q is obtained by statistically summing the discharge quantities within a unit time PD , let the sampling sequence of the partial discharge signal be PD p (n) (n = 1, 2,..., N)), the discharge quantity quantization step is Δq, and the calculation formula for the discharge quantity feature is:

[0105]

[0106] By statistically counting the number of discharge pulses within a unit time, the discharge times feature N is obtained PD , specifically, it can be accurately counted by detecting the number of times the signal amplitude exceeds the set threshold.

[0107] In addition, by analyzing the distribution of discharge pulses in terms of phase, the discharge phase feature φ is obtained PD . Specifically, a power frequency cycle (2π radians) is divided into K intervals (e.g., K = 36), and the number of discharge pulses in each interval is counted. The discharge phase feature φ PD is the normalized distribution vector of the number of discharge pulses in each interval, that is:

[0108]

[0109] where n i is the number of discharge pulses in the i-th interval.

[0110] These partial discharge features can provide important bases for accurately judging the type, severity, and development trend of partial discharges inside the transformer insulation. For example, different types of partial discharges (such as internal discharges, surface discharges, etc.) may exhibit different patterns in terms of discharge quantity, discharge times, and discharge phase features. By analyzing these features, the nature of partial discharge faults can be identified more accurately.

[0111] Through the above feature extraction steps, rich fault features are extracted from different types of signals. These features will be used as the input of the subsequent signal fault diagnosis model, providing strong support for accurately diagnosing transformer faults.

[0112] In this embodiment, separate signal fault diagnosis models are used to analyze and predict the characteristic data of different types of sensor signals. First, a large number of sample data are collected for the models of each type of signal, including samples in normal operating states and various fault states, and these data are divided into a training sample set and a test sample set. According to the feature extraction method described above, the training sample set is transformed into the form of feature vectors. For example, the vibration signal feature vector contains the energy distribution in different frequency bands and time-domain characteristics (peak value, mean value, and variance), that is, X = [E V , P V , μ V , σ V T ; the oil temperature signal feature vector contains the change rate and gradient, that is, X = [R T , G T T ; the load current feature vector contains the harmonic content and the three-phase unbalance degree, that is, X = [H I , U I T ; the partial discharge feature vector contains the discharge amount, the number of discharges, and the phase distribution, that is, X = [Q PD , N PD , φ PD T . Datasets corresponding to different signal types are established to train the signal fault diagnosis model.

[0113] In this embodiment, the architectures of the signal fault diagnosis models corresponding to each signal type are the same, and the difference lies in the different datasets for training. Therefore, the models for different signal types are described uniformly. In this embodiment, the probability neural network model PNN is used for the signal fault diagnosis model. It is mainly used for pattern classification and is a feedforward neural network based on the Bayesian strategy. The PNN model is mainly divided into four layers, namely the input layer, the sample layer, the summation layer, and the output layer. In the input layer, the number of neurons is the dimension of the feature vector. This layer calculates the distance between the input vector and all training sample vectors. The sample layer calculates the initial probability density matrix through the activation function, and this matrix is equivalent to the matching degree between the training samples and the samples to be recognized. The number of neurons in the summation layer is the number of categories. The summation layer adds the outputs of the sample layer by category. The number of neurons in the output layer is 1, that is, the output decision result, and the category with the largest probability value is output. The specific model architecture and the model construction steps can refer to the conventional architecture and model construction steps of the PNN model, and will not be introduced in detail here.

[0114] ​​​​In a preferred embodiment, the present invention improves the conventional PNN model to make the prediction results of the signal fault diagnosis model more accurate and reliable. The main difference between the improved probabilistic neural network model and the conventional probabilistic neural network model in this embodiment is as follows:

[0115] When calculating the sample distance in the input layer, this model uses an improved distance calculation formula to calculate the distance d(X, X i (i = 1, 2,..., n)) between the input sample X and each training sample X i . Most conventional sample distances are calculated using the Euclidean distance. However, this calculation method ignores the different influence degrees of different features on different fault types. Therefore, in this embodiment, first, through expert knowledge and combined with data analysis, the weights of each feature are set according to the influence degree of each feature on different fault types, and the sample distance is calculated according to the feature weights. The formula can be expressed as:

[0116]

[0117] In the formula, m represents the number of features of the training sample, w j represents the weight of the jth feature, k represents the distance exponent, x j represents the jth feature value of the input sample X, and x ij represents the jth feature value of the ith training sample X i .

[0118] In this embodiment, by introducing feature weights into the distance calculation formula, the features that have an important impact on fault diagnosis can be more prominent, making the distance calculation more in line with the actual fault diagnosis requirements.

[0119] In the pattern layer, the output value is calculated through a specific formula according to the distance between the input sample and the training sample in the input pattern layer. Let the ith neuron in the pattern layer correspond to the training sample X i , then the output of the pattern layer is:

[0120]

[0121] This value reflects the similarity degree between the input sample X and the ith training sample X i , providing a basis for the subsequent summation layer and output layer calculations. Its principle is based on the idea of the probability density function. This formula is similar to the Gaussian function form. σ controls the width of the function and determines the influence range of the sample points in the feature space. d(X, X i ) 2 measures the square of the distance between samples. The closer the distance, the smaller the value of the exponential part. Φ iThe larger the value of (X), the more similar the input sample is to the training sample, and vice versa.

[0122] In this embodiment, the difference between the output formula of the pattern layer and the conventional formula lies not only in the different distance calculation formulas, but also in the optimization of the smoothing factor σ. The smoothing factor affects the performance of the model because the value of σ directly affects the shape and distribution of the output Φ i (X), which in turn affects the calculation results of the subsequent summation layer and output layer. A suitable value of σ can enable the model to better fit the data, avoid overfitting or underfitting problems, and thus improve the classification accuracy. Therefore, to improve the performance of the model, this embodiment provides an adaptive parameter optimization method to optimize the smoothing factor. The specific steps include:

[0123] Randomly extract a first sample set and a second sample set from the sample dataset, and traverse the preset search interval according to the preset step size to obtain a candidate factor sequence containing multiple candidate smoothing factors;

[0124] Iteratively calculate each candidate smoothing factor in the candidate factor sequence according to the model classification accuracy to obtain a candidate factor set. Among them, the steps of one iteration include:

[0125] Use the signal fault diagnosis model with the candidate smoothing factor in the candidate factor sequence as the first diagnosis model, train the first diagnosis model with the first sample set, and calculate the first classification accuracy according to the output result of the trained first diagnosis model;

[0126] If the first classification accuracy is greater than the preset optimal accuracy, then use the first classification accuracy as the optimal accuracy and add the candidate smoothing factor to the candidate factor set;

[0127] Use the signal fault diagnosis model with the candidate smoothing factor in the candidate factor set as the second diagnosis model, train each second diagnosis model with the second sample set respectively, and obtain the corresponding second classification accuracy according to the output result of each trained second diagnosis model;

[0128] Select the candidate smoothing factor corresponding to the maximum value from each of the second classification accuracies as the smoothing factor of the signal fault diagnosis model.

[0129] In this embodiment, two small sample sets S 1 and S 2 are randomly extracted from the training sample set. The number of samples in sample set S 1 and S 2 are n 1 and n 2 respectively. The search interval of the smoothing factor is [σmin , σ max , the search step size is Δσ. Assume Δσ = 0.1, σ min = 0.1, σ max = 10, and n 1 = n 2 = 0.2n, where n represents the number of samples in the training sample set.

[0130] According to the above step size, traverse and search for the smoothing factor σ within the search interval, and form a candidate factor sequence [σ 1 , …… σ m of the searched smoothing factors. Then, perform iterative processing on each candidate smoothing factor in the sequence. Taking σ i as an example, the steps of one iteration are described as follows:

[0131] Set the smoothing factor in the output formula of the pattern layer in the signal fault diagnosis model to σ i , and use the sample set S 1 to train the model. For the trained model, input the samples in the sample set S 1 into the model one by one, obtain the predicted classification results, and compare the predicted results with the actual classification labels of the samples to calculate the proportion of the number of correctly classified samples to the total number of samples, which is the classification accuracy Acc 1 . Compare Acc 1 with the preset optimal accuracy Acc best . If Acc 1 is greater than the current optimal accuracy Acc best , then update Acc best = Acc 1 , and add σ i to the candidate factor set. Then, select the next candidate smoothing factor σ i+1 and perform iterative processing according to the above steps until all σ in the sequence are traversed to obtain the candidate factor set.

[0132] The candidate factor set contains some relatively good smoothing factors. Then, use the sample set S 2 to further screen out the optimal smoothing factor. Specifically, substitute each candidate smoothing factor in the candidate factor set into the model respectively, and then use the sample set S 2 to train the model, and calculate its classification accuracy Acc 2 for the trained model. Then, compare the classification accuracies Acc 2 corresponding to each smoothing factor with each other, select the model with the highest accuracy from multiple classification accuracies Acc 2 , and use the smoothing factor in this model as the optimal smoothing factor of the signal fault diagnosis model.

[0133] Through the adaptive parameter optimization method provided in this embodiment, the smoothing factor most suitable for the model can be found, thereby improving the performance and accuracy of the model.

[0134] In the summation layer, its processing steps are the same as those of the conventional PNN model, that is, the conditional probability density function is calculated in the summation layer. For each type of pattern i, the outputs Φ ij (j = 1, 2, …, n i ) of the training samples belonging to this type of pattern are summed, and then divided by the number of training samples n i of this type of pattern to obtain the conditional probability density function v i .

[0135] The specific calculation process is as follows: First, sum all Φ ij belonging to the i-th type of pattern to obtain Then calculate Its significance lies in that through the summarization and normalization of the outputs of the pattern layer, the probability estimate of each category under the current input sample is obtained, reflecting the likelihood of the input sample belonging to each category.

[0136] In the output layer, by judging the conditional probability density input from the summation layer, the corresponding fault category is selected for output. In the present invention, two methods are given to implement the judgment of the output layer. One of the methods is to compare the conditional probability density functions v i of various patterns, and select the pattern category corresponding to the maximum value as the output classification result. Assuming that the number of fault categories is c, it is specifically divided according to the common fault types of transformers, such as overheating faults, insulation faults, winding faults, and other faults, and set as c = 5. The calculation process is as follows: Compare v 1 , v 2 , …, v c to find the maximum value v max , record its corresponding index i max , then the output classification result y = i max . This step is based on the maximum likelihood principle, that is, it is considered that the input sample is most likely to belong to the category with the largest probability density function value.

[0137] In the actual operation process of the power system, the risks brought by misjudgments of different types of transformer faults are also different. In order to effectively reduce the risk of misjudgment, especially the misjudgment of serious faults, and enable the model to make more accurate and robust judgments in complex fault diagnosis scenarios, the present invention also provides another output judgment method based on the risk loss matrix. The specific steps include:

[0138] Calculate the posterior probability using Bayes' formula according to the conditional probability density and the preset prior probability, and calculate the conditional risk value according to the posterior probability and the preset risk assessment matrix;

[0139] Compare each of the conditional risk values, and select the fault category corresponding to the minimum conditional risk value as the output data.

[0140] In this embodiment, first calculate the posterior probability according to the preset prior probability and the conditional probability density output by the summation layer. Denote the conditional probability density v output by the summation layer i by p(X|ω j ), which is the probability density function of the sample X under the fault mode ω j . Assume that P(ω j ) is the prior probability of the fault mode ω j . This value can be obtained by statistically analyzing a large amount of historical data to obtain the occurrence probabilities of different fault types. For example, assume that the prior probability of overheating fault P(ω 1 ) = 0.2, the prior probability of insulation fault P(ω 2 ) = 0.3, the prior probability of winding fault P(ω 3 ) = 0.3, the prior probability of other faults P(ω 4 ) = 0.1, and the prior probability of normal state P(ω 5 ) = 0.1. Then calculate the posterior probability according to the following formula:

[0141]

[0142] The calculation of the posterior probability combines the sample characteristics (reflected by p(X|ω j )) and the prior knowledge P(ω j ), making the classification decision more scientific and reasonable.

[0143] Furthermore, considering the risk of misjudging faults, in this embodiment, according to the misjudgment risks of different faults, a risk loss matrix is designed. The elements in this matrix are λ ij , which represents the risk loss of misjudging a sample actually belonging to the j-th type of fault as the i-th type of fault. This value is set according to the actual risk situation of transformer faults. For example, set the risk loss of misjudging a serious fault as the normal state to a relatively high value, such as λ ij = 5 (when i = 1 (normal state) and j = 3, 4, 5) (severe fault state)), and set the risk loss of misjudging the normal state as a minor fault to a relatively low value, such as λ ij= 1 (when i = 2, 3 (minor fault state) and j = 1 (normal state)). The setting of the risk loss matrix is based on the assessment of the severity of the consequences of transformer faults, aiming to guide the model to be more cautious in dealing with high-risk misjudgment situations during classification decisions.

[0144] After obtaining the posterior probability and the risk loss matrix through the above steps, the conditional risk value of the model can be calculated by the following formula:

[0145]

[0146] where R(α i |X) represents the risk value of the decision-making behavior α i under the condition of the sample X, and the decision-making behavior α i corresponds to the output classification result.

[0147] After obtaining the conditional risk values of all decision-making behaviors, find the minimum risk value R min , record the corresponding decision-making behavior index i min , then the final output of the output layer is f(X) = i min .

[0148] In this embodiment, by introducing risk assessment, the risk cost of misjudgment is fully considered during classification decisions, enabling the model to make decisions more cautiously, effectively reducing the risk of misjudgment, especially significantly reducing the risk of misjudgment for serious faults. Through the above process of constructing and optimizing the reclassification model, the diagnostic accuracy and reliability of the model for transformer faults can be improved.

[0149] Since the present invention makes separate determinations for different types of sensor data, the final determination of transformer faults still requires a comprehensive determination of the prediction results output by each model. Specifically, by performing weighted summation on each output result, the final fault diagnosis result is obtained. Among them, the weight value of each model output result can be set with a fixed weight value in advance. To improve the accuracy of the diagnosis result, in a preferred embodiment, the present invention also provides a method for calculating the weight value based on the diagnosis accuracy rate, and the specific steps include:

[0150] Input the sample data sets of each signal type into the corresponding signal fault diagnosis model for fault diagnosis, and calculate the diagnosis accuracy rates of each signal fault diagnosis model according to the output results and the sample classification labels;

[0151] Calculate the weight value of the fault diagnosis sub-result according to the diagnosis accuracy rate.

[0152] In this embodiment, for each trained signal fault diagnosis model, its dataset is used as input data and input into the corresponding model. Then, by comparing the output result with the actual classification label and calculating the ratio of the number of correctly classified samples to the total number of samples, the diagnostic accuracy of each model corresponding to the corresponding signal is obtained. The diagnostic accuracy can evaluate the effectiveness and reliability of each signal in fault diagnosis.

[0153] According to the diagnostic accuracy of each signal, the corresponding weight value is set. Specifically, first calculate the sum of the diagnostic accuracies of each signal:

[0154] Acc total = Acc V + Acc T + Acc I + Acc PD

[0155] In the formula, Acc V represents the diagnostic accuracy of the model corresponding to the vibration signal, Acc T represents the diagnostic accuracy of the model corresponding to the oil temperature signal, Acc I represents the diagnostic accuracy of the model corresponding to the load current signal, Acc PD represents the diagnostic accuracy of the model corresponding to the partial discharge signal.

[0156] Then, divide the diagnostic accuracy of each signal by Acc total , to obtain the weight of each signal:

[0157]

[0158] In the formula, w V represents the weight value of the output result of the model corresponding to the vibration signal, w T represents the weight value of the output result of the model corresponding to the oil temperature signal, w I represents the weight value of the output result of the model corresponding to the load current signal, w PD represents the weight value of the output result of the model corresponding to the partial discharge signal.

[0159] Then, perform weighted summation on the diagnostic results of each signal and the corresponding weights to obtain the final fault diagnosis result:

[0160] F = w V r V + w T r T + w I r I + w PD r PD

[0161] In the formula, rV represents the model output result corresponding to the vibration signal, r T represents the model output result corresponding to the oil temperature signal, r I represents the model output result corresponding to the load current signal, r PD represents the model output result corresponding to the partial discharge signal.

[0162] Finally, according to the correspondence between the pre-set fault categories and numerical ranges, the fault category to which the final fault diagnosis result F belongs is obtained.

[0163] The adaptive weighted fusion mechanism provided in this embodiment can give full play to the advantages of each signal, improve the credibility and stability of the diagnosis result, and make the diagnosis result more accurate and reliable.

[0164] Furthermore, for the fault diagnosis results output by the model, corresponding handling measures and maintenance plans can also be formulated according to the fault type and severity, in combination with the operation regulations and maintenance experience of the transformer. In this embodiment, a corresponding severity is set for each fault type, and different handling methods are required for faults of different severities. Specifically, when F falls within the numerical range corresponding to severe faults (such as winding short circuit, severe core overheating, etc.), the system immediately triggers an emergency alarm, notifies the operation and maintenance personnel to arrange power outage maintenance as soon as possible, and provides detailed fault location and possible fault cause information, so that the maintenance personnel can quickly locate the problem and repair it. At the same time, according to historical fault data and maintenance experience, components and maintenance tools that may need to be replaced are recommended to the maintenance personnel. When F is within the range corresponding to moderate faults (such as relatively severe partial discharge, mild insulation aging, etc.), the system issues a warning signal to remind the operation and maintenance personnel to closely monitor the operation status of the transformer and arrange power outage inspection and maintenance at an appropriate time. The operation and maintenance personnel can formulate a detailed inspection plan based on the fault type and relevant data provided by the system, including further detecting the transformer using professional detection equipment, such as conducting partial discharge location tests, detailed insulation resistance measurements, etc., to determine the specific situation and scope of the fault. At the same time, in combination with the operation history and current load conditions of the transformer, evaluate the potential impact of the fault on the operation of the transformer, and adjust the operation parameters or take measures such as temporarily restricting the load when necessary to ensure the safe operation of the transformer before the fault is completely repaired. For the situation where F shows a minor fault (such as a slightly elevated oil temperature, slightly abnormal vibration, etc.), the system records the fault information and continuously monitors the change of the operation parameters of the transformer. The operation and maintenance personnel, according to the system's suggestion, first check the external factors that may cause the fault, such as checking whether the load is balanced, whether the cooling system is operating normally (including whether the cooling fan is rotating normally, whether the pressure of the cooling oil pump is normal, whether the radiator is clean, etc.), whether the oil temperature sensor and other monitoring equipment are working properly, etc. If the external factor check is normal, but the fault signal still exists or shows an increasing trend, further inspection and maintenance work, such as conducting more detailed electrical performance tests and oil quality analysis on the transformer, can be considered to determine whether there are potential internal fault hazards. It should be noted that the specific handling method can be flexibly set according to the actual situation, and this is only a preference rather than a specific limitation here.

[0165] Furthermore, the present invention also provides a human-machine interaction processing and data storage function. Specifically, the diagnosis results, treatment measures, and maintenance plans are displayed to the operator through the human-machine interface of the monitoring system, enabling the operator to intuitively understand the operating conditions of the transformer and the measures to be taken. The human-machine interface can adopt a graphical interface design, clearly presenting various operating parameters, fault diagnosis results, and corresponding treatment suggestions of the transformer in the form of charts, indicator lights, alarm information, etc. For example, different colored indicator lights are used to represent the normal operation, minor faults, medium faults, and severe fault states of the transformer; the changing trends of parameters such as oil temperature and load current are shown in the form of charts; and information such as fault types, occurrence times, and urgency levels are listed in detail in the alarm information column.

[0166] Meanwhile, the relevant data is stored in the database for subsequent query, analysis, and statistics. The data stored in the database includes the real-time operating data of the transformer (such as signal data collected by various sensors, diagnosis results, treatment measures, etc.), historical operating data (including operating parameters, fault records, maintenance records, etc. over a certain period of time), and equipment information related to the transformer (such as model, production date, last maintenance time, etc.). By regularly mining and analyzing this data, the laws and trends of transformer faults can be summarized. For example, the occurrence frequencies and type distributions of transformer faults under different seasons and different load conditions can be found. It is found that certain fault types are more likely to occur during high-temperature or high-load operation in summer, so a summer maintenance plan can be formulated accordingly, such as increasing the inspection frequency, strengthening the maintenance of the cooling system (cleaning the radiator fins in advance, checking the coolant level and quality, etc.), and optimizing the load distribution.

[0167] In addition, based on the historical fault data and maintenance records, the parameters and feature weights in the fault reclassification can also be optimized. For example, if it is found that a certain feature plays a key role in the diagnosis of a specific fault type but has a low weight in the model, its weight can be appropriately increased; conversely, if a certain feature makes little contribution to the diagnosis result or causes interference, its weight can be reduced. By continuously optimizing the model parameters and feature weights, the accuracy and efficiency of the diagnosis are improved, enabling the fault diagnosis system to better adapt to the operating changes of the transformer and provide more reliable diagnostic services.

[0168] Through the above diagnosis result output and decision-making process, the effective management and maintenance of transformer faults are realized, the reliability and safety of transformer operation are improved, and a strong guarantee is provided for the stable operation of the power system. At the same time, through human-machine interaction and data storage analysis, the diagnosis and maintenance strategies are continuously optimized, realizing the intelligent and scientific management of transformer operation and maintenance.

[0169] A transformer adaptive fault diagnosis method based on sensors provided in this embodiment. Through the integration of multiple sensors by the sensing device, the present invention realizes multi-dimensional real-time perception of the operating state of the transformer, providing rich and accurate original data for subsequent precise diagnosis; through efficient processing and in-depth feature mining of different types of signals, representative and discriminative fault features can be mined from a variety of preprocessed signals, enabling more accurate positioning of the root cause of the fault, effectively avoiding misdiagnosis and missed diagnosis, and greatly improving the reliability of fault diagnosis; the present invention constructs a signal fault diagnosis model based on an improved probabilistic neural network and combines a risk assessment strategy. Through training and optimization with a large amount of sample data, it can automatically learn the complex relationship between different fault features and types. By introducing signal feature weights in distance calculation, the model pays more attention to key features and improves diagnostic accuracy. At the same time, by setting a risk loss matrix, the misjudgment risk is quantitatively evaluated, and the serious consequences of different misjudgment situations are fully considered in the decision-making process, effectively reducing the misjudgment risk. Especially in the face of complex fault situations, more scientific and reasonable diagnostic conclusions can be made, providing a strong guarantee for the safe operation of the transformer; at the same time, the present invention also dynamically allocates weights by calculating the diagnostic accuracy of each signal, enabling the fusion process to flexibly adjust the influence of each signal according to different fault scenarios, giving full play to the advantages of each signal, avoiding diagnostic deviations caused by fixed weights, making the diagnostic results more in line with the actual fault situation, and significantly improving the accuracy of the diagnostic results.

[0170] Please refer to Figure 2 , based on the same inventive concept, a transformer adaptive fault diagnosis system proposed in the second embodiment of the present invention includes:

[0171] A feature extraction module 10, configured to collect the operating signals of the transformer through a sensing device, preprocess and extract features from the operating signals to obtain a plurality of signal feature data, where the signal feature data includes vibration signal features, oil temperature signal features, load current signal features, and partial discharge signal features;

[0172] A classification diagnosis module 20, configured to input each of the signal feature data into corresponding signal fault diagnosis models respectively to obtain a plurality of fault diagnosis sub-results. Each signal fault diagnosis model is constructed using a probabilistic neural network model and trained with a sample data set corresponding to the signal type;

[0173] A comprehensive evaluation module 30, configured to perform weighted summation on the fault diagnosis sub-results according to a preset weight value to obtain a fault diagnosis result of the transformer, where the weight value is calculated based on the diagnostic accuracy of each signal fault diagnosis model.

[0174] The technical features and effects of the sensor-based transformer adaptive fault diagnosis system proposed in the embodiments of the present invention are the same as those of the method proposed in the embodiments of the present invention, and will not be elaborated herein. Each module in the above-mentioned sensor-based transformer adaptive fault diagnosis system can be implemented in whole or in part by software, hardware, and their combinations. The above-mentioned modules can be embedded in the processor of the computer device in the form of hardware or independent of it, or stored in the memory of the computer device in the form of software, so as to facilitate the processor to call and execute the operations corresponding to the above-mentioned modules.

[0175] In addition, an embodiment of the present invention also proposes a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented.

[0176] Please refer to Figure 3 , the internal structure diagram of the computer device in one embodiment. The computer device can specifically be a terminal or a server. The computer device includes a processor, a memory, a network interface, a display, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it realizes the sensor-based transformer adaptive fault diagnosis method. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the housing of the computer device, or an external keyboard, a touchpad, or a mouse, etc.

[0177] Those of ordinary skill in the art can understand that Figure 3 the structure shown in

[0178] In summary, a sensor-based transformer adaptive fault diagnosis method, system, and device proposed in an embodiment of the present invention. The method collects the operating signals of the transformer through a sensing device, extracts features from the operating signals to obtain multiple signal feature data, and the signal feature data includes vibration signal features, oil temperature signal features, load current signal features, and partial discharge signal features; each of the signal feature data is respectively input into corresponding signal fault diagnosis models to obtain multiple fault diagnosis sub-results. Each signal fault diagnosis model is constructed using a probabilistic neural network model and trained using a sample data set of the corresponding signal type; according to a preset weight value, the fault diagnosis sub-results are weighted and summed to obtain the fault diagnosis result of the transformer, and the weight value is calculated based on the diagnostic accuracy of each signal fault diagnosis model. The present invention integrates multiple sensors through a sensing device, realizing multi-dimensional real-time perception of the operating state of the transformer, providing rich and accurate raw data for subsequent precise diagnosis; through efficient processing and in-depth feature mining of different types of signals, representative and discriminative fault features can be mined from various preprocessed signals, thereby being able to more accurately locate the root cause of the fault, effectively avoiding misdiagnosis and missed diagnosis, and greatly improving the reliability of fault diagnosis; the present invention constructs a signal fault diagnosis model based on an improved probabilistic neural network and combines a risk assessment strategy. Through training and optimization with a large amount of sample data, it can automatically learn the complex relationship between different fault features and types. The signal feature weight is introduced in the distance calculation to make the model pay more attention to key features and improve the diagnostic accuracy. At the same time, by setting a risk loss matrix, the misjudgment risk is quantitatively evaluated, and the serious consequences of different misjudgment situations are fully considered in the decision-making process, effectively reducing the misjudgment risk. Especially in the face of complex fault situations, a more scientific and reasonable diagnostic conclusion can be made, providing a strong guarantee for the safe operation of the transformer; at the same time, the present invention also dynamically allocates weights by calculating the diagnostic accuracy of each signal, enabling the fusion process to flexibly adjust the influence of each signal according to different fault scenarios, giving full play to the advantages of each signal, avoiding diagnostic deviations caused by fixed weights, making the diagnostic result more in line with the actual fault situation, and significantly improving the accuracy of the diagnostic result.

[0179] The various embodiments in this specification are described in a progressive manner. For the parts that are the same or similar in each embodiment, reference can be made to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and for the relevant parts, reference can be made to the description of the method embodiment. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as within the scope described in this specification.

[0180] The above-described embodiments only represent several preferred embodiments of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be pointed out that for those of ordinary skill in the art in this technical field, without departing from the technical principle of the present invention, several improvements and substitutions can be made, and these improvements and substitutions should also be regarded as within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the protection scope of the claims.

Claims

1. A transformer adaptive fault diagnosis method based on sensors, characterized in that: include: Collecting the operation signal of the transformer through the sensor device, extracting the characteristics of the operation signal, and obtaining a plurality of signal characteristic data, wherein the signal characteristic data includes vibration signal characteristics, oil temperature signal characteristics, load current signal characteristics and partial discharge signal characteristics; Inputting each of the signal feature data into the corresponding signal fault diagnosis model to obtain multiple fault diagnosis sub-results, each signal fault diagnosis model is constructed using a probabilistic neural network model and trained using a sample data set of the corresponding signal type; The fault diagnosis sub-results are weighted and summed according to a preset weight value to obtain a fault diagnosis result of the transformer, wherein the weight value is calculated based on the diagnosis accuracy of each signal fault diagnosis model.

2. The sensor-based transformer adaptive fault diagnosis method according to claim 1, characterized in that: The step of collecting the transformer operation signal through the sensor device, extracting the characteristics of the operation signal, and obtaining a plurality of signal characteristic data comprises: Collecting transformer operation signals through a sensor device, wherein the operation signals include vibration signals, oil temperature signals, load current signals and partial discharge signals; Performing time domain analysis and frequency domain analysis on the vibration signal to obtain time domain features and frequency domain features, and using the time domain features and the frequency domain features as vibration signal features, the time domain features including peak value, mean value and variance, and the frequency domain features including energy distribution; Calculating the oil temperature change rate and the oil temperature gradient according to the oil temperature signals at adjacent sampling moments, and using the oil temperature change rate and the oil temperature gradient as oil temperature signal features; Calculate the harmonic content feature according to the harmonic amplitude and fundamental amplitude of the load current signal, calculate the current imbalance feature according to the three-phase current of the load current signal, and use the harmonic content feature and the current imbalance feature as the load current signal feature; According to the discharge amount and the number of discharge pulses of the local discharge signal in unit time, a discharge amount characteristic and a discharge number characteristic are obtained; according to the number of discharge pulses of the local discharge signal in an industrial frequency cycle, a discharge phase characteristic is obtained, and the discharge amount characteristic, the discharge number characteristic and the discharge phase characteristic are used as local discharge signal characteristics.

3. The sensor-based transformer adaptive fault diagnosis method according to claim 2 is characterized in that: Before the step of performing time domain analysis and frequency domain analysis on the vibration signal, the method further comprises: Using a wavelet transform algorithm to remove noise from the vibration signal to obtain a denoised vibration signal; The oil temperature signal is smoothed by using a moving average filtering algorithm to obtain a smoothed oil temperature signal; Using fast Fourier transform to perform time-frequency conversion and remove interference harmonics on the load current signal to obtain a pure load current signal; The noise threshold is used to remove noise from the partial discharge signal to obtain a denoised partial discharge signal.

4. The sensor-based transformer adaptive fault diagnosis method according to claim 1, characterized in that: The signal fault diagnosis model is constructed by adopting an improved probabilistic neural network model, and comprises an input layer, a sample layer, a summation layer and an output layer; The input layer is used to calculate the sample distance between the input sample and the training sample using an improved distance formula, and input the sample distance into the sample layer; The sample layer is used to calculate the similarity between the input sample and the training sample according to the sample distance using a probability density formula based on a smoothing factor, and input the similarity into the summation layer; The summation layer is used to calculate the conditional probability density according to the similarity, and input the conditional probability density into the output layer; The output layer is used to select the corresponding fault category as output data according to the conditional probability density; The sample distance is expressed by the following formula: In the formula, X represents the input sample, X i represents the i-th training sample, m represents the number of features of the training sample, and w j represents the weight of the jth feature, k represents the distance index, x j represents the jth eigenvalue of the input sample X, x ij represents the i-th training sample X i The j-th eigenvalue of The similarity is expressed by the following formula: In the formula, Φ i (X) represents the input sample X and the training sample X i The similarity between them, σ represents the smoothing factor.

5. The sensor-based transformer adaptive fault diagnosis method according to claim 4 is characterized in that: The smoothing factor is determined by an adaptive parameter optimization method, and the specific steps include: Randomly extracting a first sample set and a second sample set from the sample data set, and traversing a preset search interval according to a preset step length to obtain a candidate factor sequence including a plurality of candidate smoothing factors; Iteratively calculate each candidate smoothing factor in the candidate factor sequence in turn according to the model classification accuracy to obtain a candidate factor set, wherein the steps of one iteration include: Using a signal fault diagnosis model that uses a candidate smoothing factor in the candidate factor sequence as a first diagnosis model, using the first sample set to train the first diagnosis model, and calculating a first classification accuracy rate based on an output result of the trained first diagnosis model; If the first classification accuracy is greater than a preset optimal accuracy, the first classification accuracy is used as the optimal accuracy, and the candidate smoothing factor is added to the candidate factor set; Using the signal fault diagnosis model of the candidate smoothing factor in the candidate factor set as the second diagnosis model, using the second sample set to train each second diagnosis model respectively, and obtaining the corresponding second classification accuracy according to the output results of each trained second diagnosis model; The candidate smoothing factor corresponding to the maximum value is selected from each of the second classification accuracy rates as the smoothing factor of the signal fault diagnosis model.

6. The sensor-based transformer adaptive fault diagnosis method according to claim 4, characterized in that: The step of selecting a corresponding fault category as output data according to the conditional probability density comprises: The conditional probability densities are compared, and according to the comparison result, the fault category corresponding to the maximum conditional probability density is selected as the output data.

7. The sensor-based transformer adaptive fault diagnosis method according to claim 4, characterized in that: The step of selecting a corresponding fault category as output data according to the conditional probability density comprises: According to the conditional probability density and the preset prior probability, the posterior probability is calculated using the Bayesian formula, and the conditional risk value is calculated according to the posterior probability and the preset risk assessment matrix; The conditional risk values ​​are compared, and the fault category corresponding to the minimum conditional risk value is selected as output data.

8. The sensor-based transformer adaptive fault diagnosis method according to claim 1, characterized in that: The step of calculating the weight value comprises: The sample data sets of each signal type are input into the corresponding signal fault diagnosis model for fault diagnosis, and the diagnostic accuracy of each signal fault diagnosis model is calculated based on the output results and sample classification labels; Calculating the weight value of the fault diagnosis sub-result according to the diagnosis accuracy; Secondly, the weight value is expressed by the following formula: In the formula, ω i represents the weight value of the i-th fault diagnosis sub-result, Acc i represents the diagnostic accuracy of the i-th signal fault diagnosis model, and n represents the total number of signal fault diagnosis models.

9. A transformer adaptive fault diagnosis system based on sensors, characterized in that: include: A feature extraction module is used to collect the operation signal of the transformer through a sensor device, preprocess and extract the feature of the operation signal, and obtain a plurality of signal feature data, wherein the signal feature data includes vibration signal features, oil temperature signal features, load current signal features and partial discharge signal features; A classification diagnosis module is used to input each of the signal feature data into the corresponding signal fault diagnosis model to obtain multiple fault diagnosis sub-results, each signal fault diagnosis model is constructed using a probabilistic neural network model and trained using a sample data set of the corresponding signal type; The comprehensive evaluation module is used to perform weighted summation on the fault diagnosis sub-results according to a preset weight value to obtain a fault diagnosis result of the transformer, wherein the weight value is calculated based on the diagnosis accuracy of each signal fault diagnosis model.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images