Transformer operation state evaluation system and method

Abstract

The invention provides a transformer operation state evaluation system and method, an online monitoring module of the system continuously collects n kinds of operation data of a transformer in real time, and n is greater than or equal to 1; the state evaluation module evaluates the n kinds of operation data based on a first state evaluation model to obtain n single-parameter evaluation results; performing multi-source information fusion on the n single-parameter evaluation results to obtain fusion information; and finally, inputting the fusion information into a second state evaluation model for quantitative evaluation to obtain a transformer operation state evaluation result. The first state evaluation model has a parameter theta; the incremental learning module inputs the operation data continuously received from the online monitoring module into an incremental learning algorithm formula to carry out iterative optimization on the parameter theta until a preset iteration stop condition is met, and an updated value theta'of the parameter theta is obtained; and the first state evaluation model obtains a single-parameter evaluation result based on the updated value. According to the method, the state evaluation model is dynamically optimized through the incremental learning algorithm, so that the state evaluation is more accurate.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and particularly to a transformer operation state evaluation system and method. Background Art

[0002] As a core device for power transmission and distribution, the operation state of a transformer is directly related to the stability and safety of the power system. Therefore, it is of great significance to evaluate the operation state of a transformer in real time and accurately for maintaining the continuous and healthy operation of the equipment, preventing equipment failures, extending the service life of the equipment, and ensuring the continuous and reliable supply of power.

[0003] Traditional methods for evaluating the operation state of a transformer mainly rely on regular manual inspections and off-line tests. Although they can reflect the operation state of the transformer to a certain extent, there are problems such as long evaluation cycles, poor real-time performance, and high labor costs. With the development of sensor technology, signal processing technology, and machine learning technology, the power industry has begun to explore data-based on-line monitoring and evaluation methods, which process and analyze the data obtained by real-time monitoring of the operation of the transformer to achieve real-time evaluation of the transformer state.

[0004] Although data-based on-line monitoring and evaluation methods have been applied in the power industry, there are still some technical defects: 1. Based on the state evaluation model trained by using a static learning method, the training data set comes from the historical operation data of the transformer. After the training is completed, the model parameters are closed, and then the state evaluation model is used to evaluate and predict the operation state of the transformer to obtain the evaluation result. Therefore, the evaluation result is only related to the historical operation data and has nothing to do with the newly generated operation data that continues to be produced subsequently. If the newly generated operation data cannot find the same or similar data in the historical data, it will inevitably lead to a significant reduction in the accuracy of the evaluation result.

[0005] 2. Only focus on the electrical and environmental parameters of the transformer, such as voltage, current, and temperature, while ignoring the sound characteristics during the operation of the transformer, resulting in incomplete and inaccurate evaluation results.

[0006] 3. The evaluation result is displayed in a very professional data format, lacking visualization display and alarm prompt functions. It is difficult for users to simply, intuitively, and accurately know the current operation state of the transformer. What they see is not what they get, and it is easy to miss the opportunity to take timely remedial measures. Summary of the Invention

[0007] Aiming at the defects of incomplete transformer status evaluation, poor real-time performance, and insufficient model adaptability in the prior art, the purpose of the present invention is to provide a transformer operation status evaluation system and method, which dynamically optimize the status evaluation model with continuously received new data through an incremental learning algorithm, realize accurate and real-time evaluation of the transformer operation status, and provide a strong guarantee for the safe and stable operation of the power system.

[0008] To achieve the above invention purpose, in the first aspect, the present invention provides a transformer operation status evaluation system, including: an online monitoring module for continuously and real-time collecting n kinds of operation data of the transformer, where n≥1; a status evaluation module for first evaluating the n kinds of operation data respectively based on a first status evaluation model to obtain n single-parameter evaluation results corresponding one-to-one to the n kinds of operation data; then performing multi-source information fusion on the n single-parameter evaluation results to obtain fusion information; finally inputting the fusion information into a second status evaluation model for quantitative evaluation to obtain a transformer operation status evaluation result; the first status evaluation model has parameters θ; an incremental learning module for inputting the n kinds of operation data continuously received from the online monitoring module into an incremental learning algorithm formula to iteratively optimize the parameters θ of the first status evaluation model until a preset iteration stop condition is met, and obtaining an updated value θ' of the parameters θ; the first status evaluation model obtains each single-parameter evaluation result based on the updated value θ'.

[0009] Preferably, the incremental learning algorithm formula is ; in the formula, is the parameter of the first status evaluation model at time t, η is the learning rate, is the gradient of the loss function L with respect to the parameter θ, is the operation data input into the incremental learning module at time t, is the true label at time t.

[0010] Preferably, the iteration stop condition is: after T iterations, the parameter θ is updated from to ; or, the iteration stop condition is: the value of the loss function L is less than a preset threshold.

[0011] Preferably, it further includes a sound field analysis module for analyzing the sound characteristics during the operation of the transformer and inputting the sound characteristics into the first status evaluation model to obtain a sound characteristic evaluation result.

[0012] Preferably, the sound field analysis module includes a sound feature extraction unit and a sound feature matching unit; the sound feature extraction unit is used to extract feature information from the sound generated during the operation of the transformer; a sound feature library is preset in the sound feature matching unit, and the sound feature library includes a plurality of feature pairs (c, d), where c is a sound element and d is the transformer state corresponding to c. The extracted feature information is compared with each of the sound elements respectively until they match, and then the transformer state corresponding to the feature information is the transformer state corresponding to the matched feature element.

[0013] Preferably, it further includes a visualization display module for graphically or tabularly displaying the evaluation result of the transformer operation state.

[0014] Preferably, the visualization display module includes an alarm prompt unit, which triggers an alarm prompt when the evaluation result of the transformer operation state reaches a preset alarm threshold.

[0015] In a second aspect, the present invention provides a method for evaluating the operation state of a transformer, including: data acquisition: continuously and real-time online collecting n kinds of operation data of the transformer, n≥1; state evaluation: first evaluating each of the n kinds of operation data based on a first state evaluation model to obtain n single-parameter evaluation results; then performing multi-source information fusion on the n single-parameter evaluation results to obtain fusion information; finally inputting the fusion information into a second state evaluation model for quantitative evaluation to obtain an evaluation result of the transformer operation state; the first state evaluation model has a parameter θ; parameter update: inputting at least one of the operation data continuously received from the online monitoring module into an incremental learning algorithm formula to iteratively optimize the parameter θ until a preset iteration stop condition is met, and obtaining an updated value θ' of the parameter θ; the first state evaluation module obtains the single-parameter evaluation result based on the updated value θ'.

[0016] Preferably, the incremental learning algorithm formula is ; in the formula, is the parameter of the model at time t, η is the learning rate, is the gradient of the loss function L with respect to the parameter θ, is the data input at time t, is the true label at time t.

[0017] Preferably, the iteration stop condition is: after T iterations, the parameter θ is updated from to ; or, the iteration stop condition is: the value of the loss function L is less than a preset threshold.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention dynamically optimizes the transformer operation state evaluation model through an incremental learning algorithm. The continuously newly received transformer operation data is input into the incremental learning algorithm formula, and the parameters of the state evaluation model are iteratively updated, so that the state evaluation result is not only related to the historical operation data, but also closely related to the latest operation data, realizing accurate and real-time evaluation of the transformer operation state, thereby timely discovering potential safety hazards, providing a strong reference basis for eliminating operation faults in the first time, and providing a strong guarantee for the safe and stable operation of the power system. Other technical effects brought by the additional features will be further elaborated in the corresponding embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Other features, objects, and advantages of the present invention will become more apparent by reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 is a schematic block diagram of an embodiment of the transformer operation state evaluation system of the present invention; Figure 2 is a schematic block diagram of another embodiment of the transformer operation state evaluation system of the present invention; Figure 3 is a schematic block diagram of still another embodiment of the transformer operation state evaluation system of the present invention; Figure 4 is a schematic block diagram of yet another embodiment of the transformer operation state evaluation system of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that those of ordinary skill in the art can make several modifications and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0021] As Figure 1 shown, an embodiment of the transformer operation state evaluation system of the present invention includes: an on-line monitoring module, a state evaluation module, and an incremental learning module, wherein: The on-line monitoring module continuously and real-time collects n kinds of operation data of the transformer, where n ≥ 1; The state evaluation module first evaluates the n kinds of operation data respectively based on the first state evaluation model to obtain n single-parameter evaluation results corresponding one-to-one to the n kinds of operation data, then performs multi-source information fusion on the n single-parameter evaluation results to obtain fusion information, and finally inputs the fusion information into the second state evaluation model for quantitative evaluation to obtain the transformer operation state evaluation result, wherein the first state evaluation model has parameters θ; The incremental learning module inputs the n types of operation data continuously received from the online monitoring module into the incremental learning algorithm formula respectively to iteratively optimize the parameters θ of the first state evaluation model until a preset iteration stop condition is met, and obtains an updated value θ' of the parameter θ; the first state evaluation model obtains each single-parameter evaluation result based on the updated value θ'.

[0022] In this embodiment, the online monitoring module includes various sensors installed on the transformer, such as voltage sensors, current sensors, temperature sensors, and oil chromatograph monitoring devices, etc. These sensors collect the operation data of the transformer in real time, including electrical parameters such as voltage, current, and power factor, winding temperature, and characteristic gas parameters representing insulation characteristics, etc. It can be seen that the type n of the operation data is at least 5. These operation data can reflect the operation state and load condition of the transformer and provide basic data for state evaluation. Then at least one of these operation data is transmitted to the incremental learning module through the data interface. The characteristic gases include hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2), carbon monoxide (CO), carbon dioxide (CO2), etc.

[0023] In this embodiment, the incremental learning module adopts an incremental learning algorithm based on machine learning, which can continuously receive new data and dynamically optimize the state evaluation model. The specific implementation method is as follows: (1) Select the incremental learning algorithm based on Stochastic Gradient Descent (SGD) as the optimization algorithm, which can quickly update the model parameters each time new data is received, so as to realize the dynamic optimization of the model.

[0024] SGD algorithm formula: , where represents the parameter of the first state evaluation model at time t, η represents the learning rate, represents the gradient of the loss function L with respect to the parameter θ, and represent the operation data input into the incremental learning module at time t and the corresponding true label respectively.

[0025] In the above algorithm, for the learning rate η, the common range of the learning rate is generally between (0,1). Although in some cases, the value can be set larger, such as 1; or smaller, such as . But in most cases, the commonly selected range is [0.001,0.1]. There are many influencing factors for the specific selection of the learning rate, including but not limited to model complexity, dataset scale and features, gradient magnitude, optimization algorithm, training stage, etc.

[0026] For example, the more complex the model and the more parameters it has, the smaller the learning rate usually needs to be set. Complex models are prone to oscillations and instability during training, and a smaller learning rate helps with stable convergence. This is because complex models optimize in a high-dimensional parameter space, increasing their sensitivity to the learning rate. An overly large learning rate may cause the model to fail to converge.

[0027] Again, a larger dataset may require a higher learning rate because the model can be effectively updated and learned through more data samples. In this way, when the dataset is large, the direction and magnitude of the update are more significantly affected by the data, and an effective learning rate can quickly utilize this information.

[0028] Also, the magnitude of the gradient affects the choice of learning rate. If the gradient value is large, the learning rate may need to be decreased to prevent the update from being too large and causing the model to diverge; if the gradient is small, increasing the learning rate can be considered. This is because the gradient reflects the change of the loss function, and an overly large gradient will lead to an overly large update amplitude, thus missing the optimal solution.

[0029] In addition, different optimization algorithms have different sensitivities to the learning rate. For example, adaptive learning rate optimization algorithms such as Adam and RMSprop usually dynamically adjust the learning rate based on historical gradients and allow the use of a larger initial learning rate. Adaptive algorithms adjust the learning rate to adapt to gradient changes, helping the model converge more effectively at different stages.

[0030] Moreover, it is also different at different training stages. Because a large learning rate in the early stage can effectively explore the parameter space, while a small learning rate in the later stage can find the optimal solution more precisely. Therefore, at the beginning of training, a larger learning rate may be required to quickly converge to a better solution, while in the later stage, the learning rate usually needs to be decreased to finely adjust the parameters.

[0031] Based on the above various influencing factors, in practice, an optimal learning rate generally needs to be determined. However, this optimal value is not a fixed optimal learning rate value, and it usually depends on the specific problem, model architecture, and dataset. In practical applications, the best learning rate is usually obtained through experiments and tuning, and it can also be obtained by using search methods, namely, Learning Rate Finder: One method is to gradually increase the learning rate and observe the change of the loss function to find the learning rate value within the range where the loss decreases; another can be grid search or random search, which trains on a series of preset learning rates and selects the one with the best performance. Of course, in practical applications, there can also be other methods.

[0032] (2) Each time new data is received from the online monitoring module, use the SGD algorithm to update the parameters of the state evaluation model.

[0033] Calculate the gradient of the loss function \(L\) with respect to the parameter \(\theta\), and update the model parameters according to the SGD formula. Gradient calculation formula: ; where, represents the predicted output of the model under the parameter \(\theta\) and the input \(x\) t and \(y\) t represents the true label. represents the predicted output gradient with respect to the parameter \(\theta\).

[0034] (3) Through continuous iterative optimization, the model can adapt to the changes in data in real time, improving the accuracy and timeliness of the evaluation results.

[0035] Set an iteration number or a stopping condition (such as the loss function value is less than a certain threshold). In each iteration, receive new data from the online monitoring module, update the model parameters using the SGD algorithm, and calculate the loss function value to evaluate the model performance. The value range of the loss function value threshold usually depends on the specific task and the loss function itself. Generally speaking, the threshold can be a positive number or zero. Common value ranges include: for regression problems, the threshold can be between \((0, \infty)\), and the specific value depends on the data distribution and task requirements. For classification problems, especially binary classification problems, the threshold can be between \([0, 1]\), usually a positive number less than 1, such as 0.01, 0.1, etc.

[0036] In practice, the selection of an appropriate value for the loss function threshold is influenced by multiple factors, mainly including: task type, characteristics of the dataset, model complexity, fault tolerance of the application scenario, training stage, etc. Among them, regarding the task type, different tasks (regression, classification, clustering, etc.) usually correspond to different loss functions, so the selection of the threshold will also vary. For regression problems, loss functions such as mean squared error (MSE) are generally used. In this case, the threshold needs to consider a reasonable error range between the predicted value and the true value. For classification problems, cross-entropy loss may be used. At this time, the threshold is usually related to the classification accuracy. Regarding the characteristics of the dataset, the distribution, number of samples, and characteristics of the data will affect the magnitude of the loss and the performance of the model. Therefore, an appropriate threshold needs to be set to measure the effectiveness of the model. For example, sample imbalance in the dataset may lead to a relatively high value of the loss function, so the threshold may need to be adjusted accordingly to ensure that the model can still effectively classify minority class samples. Regarding model complexity, a model with higher complexity (such as a deep neural network) may require a more stringent threshold to avoid overfitting. A complex model may perform well on the training set but may not perform well on the validation set or test set. Therefore, setting a lower threshold can ensure the generalization ability of the model on unseen data. Regarding the fault tolerance of the application scenario, some application scenarios (such as medical diagnosis) have a low tolerance for errors and may require a more stringent threshold, while in other applications (such as recommendation systems), the threshold can be relatively loose. In critical tasks, it is particularly important to ensure the high accuracy and low error rate of the model. Therefore, the threshold should be set lower to ensure the reliability of the model. At different stages of training, the selection of the threshold may vary. In the initial stage, a relatively high threshold can be set to allow the model to explore, while in the later stage, the threshold can be lowered to pursue higher accuracy. As training progresses, the model gradually learns the patterns in the data, and lowering the threshold helps to improve the performance and accuracy of the model.

[0037] The optimal threshold usually depends on the specific task, model architecture, and dataset. In practical applications, the optimal threshold is usually determined through experiments and tuning.

[0038] The iterative process can be expressed as: for t = 1 to T (or until the stopping condition is met): Receive new data (x t ,y t ) Calculate the gradient Update the parameters Evaluate the model performance (calculate the value of the loss function, etc.). Here, T represents the preset number of iterations or the number of iterations when the stopping condition is met.

[0039] In this embodiment, the first state evaluation model can also be called a single-parameter evaluation model, which is used to evaluate a single transformer operation data, such as electrical data, sound data, etc., to obtain single-parameter evaluation results, such as electrical evaluation results, sound evaluation results, etc.

[0040] Taking two input features, including sound feature (Ssound S sound) and electrical parameter (Selec S elec), an independent evaluation model is established for each input feature, that is, a single-parameter evaluation model. Taking the sound feature as an example, assuming that a linear regression model is used to evaluate the sound feature, its expression can be written as: S sound =f sound (x) = θ T sound (x) + b sound S sound is the evaluation result of the sound feature.

[0041] x is the input feature vector (which may include frequency, amplitude, etc.).

[0042] θ sound is the weight parameter of the sound feature model.

[0043] b sound is the bias term.

[0044] Similarly, an evaluation model is established for the electrical parameter, and its expression is: S elec =f elec (y) = θ T elec (y) + b elec S elec is the evaluation result of the electrical parameter.

[0045] y is the electrical parameter feature vector (such as voltage, current, etc.).

[0046] θ elec is the weight parameter of the electrical parameter model.

[0047] belec is the bias term.

[0048] In this embodiment, the final state score or state level is evaluated by the second state evaluation model, specifically as follows: After obtaining the single-parameter evaluation results, these results need to be fused to obtain the final state score or state level of the transformer. A comprehensive evaluation model, such as a weighted sum model, can be used to achieve this. The expression of this weighted sum model is: S = f(S sound ,S elec ,…)=w 1 S sound +w 2 S elec +…+w n S n S is the final state score or state level.

[0049] w 1, w 2, …, w n are the weight coefficients of each evaluation result, reflecting their importance in the comprehensive evaluation.

[0050] S n represents other possible evaluation results.

[0051] The above-mentioned evaluation models are further explained and assigned calculations below.

[0052] The description of the bias term b is as follows: In many machine learning models, the bias term is an important parameter that can help the model better fit the data. For a linear model, the bias term is usually added to the input features, making the model more flexible.

[0053] Meaning: The bias term b is a constant term in the model used to adjust the baseline of the prediction result. It allows the model to still give a non-zero prediction when there are no input features.

[0054] Logical relationship: After the bias term is added to the model, it allows the model to output a fixed value when the feature value is zero (or when the influence of all features is canceled).

[0055] The description of the parameter θ is as follows: Meaning: θ is the weight parameter of the model, representing the linear relationship between the input features and the prediction result. Each feature corresponds to a weight, reflecting the degree of influence of that feature on the final prediction result.

[0056] Logical relationship: In the model, the linear combination of θ and the input feature x jointly affects the output result.

[0057] The prediction process of the first - state evaluation model or single - parameter evaluation model is as follows: Suppose there are two single - parameter evaluation models, one for evaluating sound features, S sound , and the other for evaluating electrical features, S elec .

[0058] The sound feature evaluation result, S sound is usually obtained through a series of steps, including feature extraction, model construction, training, and evaluation. First, a suitable model can be selected from traditional machine learning methods (such as support vector machines, decision trees, random forests, etc.) or deep - learning methods (such as convolutional neural networks, recurrent neural networks, etc.). The selection depends on specific task requirements and data characteristics.

[0059] In the model, sound features can include frequency features (such as Mel - Frequency Cepstral Coefficients, MFCCs), time - domain features (such as Short - Time Energy, STE, Zero - Crossing Rate, ZCR), and spectral features (such as spectrogram and spectral centroid). The next step is data preparation. Collect an audio dataset containing the target sound, such as environmental sounds, music, and speech, etc., and perform pre - processing, including denoising, framing, normalization, etc., to prepare the audio signal for feature extraction.

[0060] In the feature extraction step, signal - processing techniques are used to extract features from the audio data. For example, calculate MFCCs, short - time energy, and zero - crossing rate from the audio signal, and generate spectrograms for subsequent analysis. Subsequently, a labeled audio - feature dataset is used for model training. This process usually includes dividing the dataset into training set, validation set, and test set, selecting a loss function, optimizing the model parameters using the training - set data, and adjusting the hyperparameters of the model through the validation set.

[0061] In the model evaluation stage, evaluate the model performance on the test set to obtain the evaluation result of S sound . Evaluation metrics can include accuracy, precision, recall, and F1 - score, etc., to comprehensively measure the performance of the model. The finally generated evaluation result, S sound can be the recognition result of specific categories (such as "normal", "abnormal") or the numerical score of sound features (such as loudness, frequency range, etc.).

[0062] For example, assume we want to evaluate the state of environmental noise. First, a set of environmental noise audio is collected and labeled into categories such as "quiet", "noisy", "medium", etc. When extracting features, MFCC features and short-time energy features are extracted from the audio files. Then, a random forest classifier is selected, and the model is trained on the training set using the extracted features. When evaluating the model on the test set, metrics such as accuracy and F1 score are calculated. Finally, S sound can represent the classification result of the model (such as "the environmental noise is noisy") or a numerical evaluation of the sound features (such as "the sound intensity is 85 dB"). These evaluation results can be used for subsequent decision support, alarm systems, or other applications.

[0063] The evaluation result S of electrical parameters elec is usually obtained through a series of well-defined steps, including data acquisition, feature extraction, model selection, training, and evaluation. First, the evaluation process starts with data acquisition, usually by installing sensors or measuring instruments to obtain relevant parameters of electrical equipment, such as voltage, current, power, frequency, and phase. This data can come from the actual operating environment or be obtained through laboratory tests.

[0064] In the data preparation stage, the collected electrical data needs to be preprocessed to ensure data quality and consistency. Preprocessing may include steps such as denoising, smoothing, and normalization to eliminate the interference of outliers and noise and ensure the accuracy of subsequent feature extraction. The feature extraction stage is to extract meaningful features from the original electrical signals, such as calculating the effective values of voltage and current, power factor, harmonic analysis, etc., to capture the operating state and performance characteristics of the equipment.

[0065] Next, selecting a suitable evaluation model is a crucial step. Common models include linear regression, support vector machines, decision trees, or more complex deep learning models. When selecting a model, specific application requirements and data characteristics need to be considered. In the model training stage, the extracted features are matched with known electrical parameter labels, and appropriate algorithms are used to optimize the model parameters. This process usually involves dividing the dataset into a training set and a test set, using the training set to learn the model, and adjusting the hyperparameters through a validation set to improve the generalization ability of the model.

[0066] In the model evaluation stage, the trained model is evaluated using the test set to obtain the final evaluation result S of electrical parameters elec . Evaluation metrics can include mean squared error (MSE), coefficient of determination (R²), etc., to quantify the performance of the model in predicting electrical parameters. The final evaluation result can be a classification of the operating state of the electrical equipment (such as "normal", "abnormal") or a numerical prediction of specific electrical parameters (such as "the voltage is 230 V").

[0067] For example, when evaluating the electrical parameters of a transformer, first collect the voltage and current data during its operation through sensors and record the data. After preprocessing, extract features such as the effective value, power factor, and harmonics, and then select a linear regression model for training. After optimizing the model parameters with the training set and evaluating on the test set, the evaluation results of the transformer's electrical parameters can be obtained, such as whether its operating state is normal or the specific power output value.

[0068] The assignment calculation process for sound features and electrical features is as follows: Model expression: Sound feature model: S sound = θ sound T x + b sound Electrical feature model: S elec = θ elec T y + b elec Parameter assignment: Assume the following parameters: For the sound feature model: (Assume there are two features) b sound = 1.0 For the electrical feature model: θ elec = [0.4, 0.6] (Assume there are two features) b elec = 0.5 Input features: Sound feature x = [2.0, 3.0] Electrical feature y = [1.5, 2.5] Prediction process: Calculate the evaluation result of the sound feature: 。

[0069] Calculate the evaluation result of the electrical feature: 。

[0070] The prediction process of the final state scoring model or the second state evaluation model is as follows: After obtaining the single-parameter evaluation results, fuse these results to obtain the final state score or state level of the transformer.

[0071] Model expression: S = w1S sound + w2Selec Parameter assignment: Assume the following weights: w1 = 0.7 (weight of sound features) w2 = 0.3 (weight of electrical features) Prediction process: Single-parameter evaluation results obtained from the above calculations: S sound = 1.4 S elec = 2.6 Calculate the final state score: S = 0.7×1.4 + 0.3×2.6 = 0.98 + 0.78 = 1.76.

[0072] Through the above process, starting from the single-parameter evaluation model, the evaluation results of sound and electrical features are calculated using the set parameters and input features. Then, the weighted sum model is used to combine the two evaluation results to obtain the final state score S = 1.76.

[0073] This prediction process demonstrates how the working principle of the model and the parameters jointly affect the final result.

[0074] In this embodiment, through the collaborative work of the online monitoring module and the incremental learning module, new data can be collected and processed in real time, and the state evaluation model can be dynamically optimized. This enables the system to reflect the operating state of the transformer in real time and continuously update the evaluation results as the data changes, enhancing the real-time performance and dynamic adaptability of the system. The application of the incremental learning algorithm enables the system to update the model without retraining the entire model, thereby reducing the computational complexity and improving the evaluation efficiency. At the same time, since the transformer state can be evaluated in real time, the frequency of manual inspections and offline tests is reduced, and the operation and maintenance costs are lowered.

[0075] In this embodiment, as Figure 2 shown, it may further include a sound field analysis module for analyzing the sound features during the operation of the transformer and inputting the sound features into the first state evaluation model to obtain the sound feature evaluation results. Preferably, the sound field analysis module includes a sound feature extraction unit and a sound feature matching unit; the sound feature extraction unit is used to extract feature information from the sound generated during the operation of the transformer; a sound feature library is preset in the sound feature matching unit, and the sound feature library includes a plurality of feature pairs (a, b), where a is a sound element and b is the transformer state corresponding to a. The extracted feature information is compared with each of the sound elements until they match, and then the transformer state corresponding to the feature information is the transformer state corresponding to the matched feature element. Specific examples of the above feature pairs are illustrated as follows: Feature Pair 1: c: Frequency d: Sound Pressure Level, SPL Description: Frequency represents the pitch of a sound, while the sound pressure level represents the intensity of a sound. The combination of the two can analyze the nature and impact of a sound.

[0076] Feature Pair 2: c: Time-domain features (such as Zero Crossing Rate, ZCR) d: Short-Time Energy Description: The zero crossing rate can reflect the complexity of a signal, while the short-time energy is used to represent the intensity of a signal. This is very useful for the analysis of speech or music.

[0077] Feature Pair 3: c: Mel-frequency cepstral coefficients, MFCCs d: Fundamental Frequency, F0 Description: MFCCs are widely used in speech recognition and can effectively capture the features of speech, while the fundamental frequency is an important parameter of pitch.

[0078] Feature Pair 4: c: Spectrogram features d: Timbre Description: The spectrogram provides the frequency components of a sound over time, while the timbre can help identify the texture and characteristics of a sound, applicable to music and environmental sound recognition.

[0079] Feature Pair 5: c: High-Frequency Energy d: Low-Frequency Energy Description: The ratio of high-frequency and low-frequency energy can be used for the classification of audio signals, such as in the separation of speech and noise.

[0080] Feature Pair 6: c: Loudness d: Audio Duration Description: Loudness reflects the subjective intensity of a sound, while the duration can be used to analyze the characteristics of an event, such as the length of a music segment.

[0081] In this embodiment, a high-sensitivity microphone array can be arranged around the transformer to collect the sound signals generated during the operation of the transformer. The collected sound signals are preprocessed, including filtering, denoising, etc., to improve the signal quality. Signal processing techniques are used to extract sound features, such as frequency, amplitude, harmonic components, etc. Specific extraction methods can use mathematical tools such as Fourier transform and wavelet transform.

[0082] Sound feature extraction formula: f = n / T, where f is the sound frequency, n is the harmonic order, and T is the period.

[0083] , where A is the sound amplitude, and a and b are the real and imaginary parts of the sound signal, respectively.

[0084] These sound features can reflect the state information such as mechanical vibration and electromagnetic noise inside the transformer, providing an important reference for the assessment of the transformer operation state.

[0085] In this embodiment, the state assessment module quantitatively assesses the operation state of the transformer, and the specific steps are as follows: (1) Adopt multi-source information fusion technology to comprehensively analyze information such as sound features and electrical parameters.

[0086] Multi-source information fusion technology formula: , where S is the comprehensive assessment result, S sound is the assessment result of sound features, S elec is the assessment result of electrical parameters, and w1, w2,..., w n are the weight coefficients of the assessment results of each single parameter. S n It also includes the types and contents of oil chromatographic characteristic gases, oil temperature, oil pressure, etc.

[0087] (2) Use machine learning algorithms (such as support vector machines, neural networks, etc.) to establish a state assessment model.

[0088] The steps for model establishment are as follows: 1. Data preparation 1.1 Data collection Collect data related to the assessment target, which can be sensor data, laboratory test results, or historical records. The quality of the data directly affects the performance of the model, so the accuracy and integrity of the data must be ensured.

[0089] 1.2 Data preprocessing Preprocess the collected data, including: Denoising: Removing noise or outliers from the data.

[0090] Standardization / Normalization: Scaling the data to a unified range to improve the convergence speed and performance of the model.

[0091] Missing Value Handling: Filling or discarding missing data to ensure the integrity of the dataset.

[0092] 2. Feature Extraction Extract meaningful features from the preprocessed data, which will be used to train the model. The steps of feature extraction include: Feature Selection: Select important features related to state assessment through domain knowledge.

[0093] Feature Generation: Use algorithms to extract new features, such as extracting frequency-domain features through Fourier transform, or calculating statistical features (such as mean, variance, etc.).

[0094] 3. Model Selection Select a suitable machine learning algorithm according to the data characteristics and evaluation tasks. Neural Network: Suitable for modeling complex non-linear relationships, especially performing well when the data volume is large.

[0095] 4. Model Training Train the selected model through the following steps: Data Partitioning: Divide the dataset into training set, validation set and test set, generally in the ratio of 70%-15%-15%.

[0096] Model Training: Use the training set data to train the model, and gradient descent method or other optimization algorithms can be used for parameter optimization.

[0097] Hyperparameter Tuning: Adjust the hyperparameters of the model, such as regularization parameter, learning rate, number of layers, etc., through the validation set to improve the model performance.

[0098] 5. Model Evaluation Evaluate the model performance on the test set, and the evaluation metrics can include: Accuracy: The proportion of correct classifications of the classification model.

[0099] Precision and Recall: Used to analyze the recognition ability of the classification model for positive classes.

[0100] F1 Score: An indicator that comprehensively considers precision and recall.

[0101] Mean Squared Error (MSE): The average squared difference between the predicted values and the actual values of the regression model.

[0102] 6. Model Application Once the model is trained and validated, it can be applied to actual state assessment tasks. The actual application can be carried out through the following steps: Real-time data input: Input real-time or new data into the model for evaluation.

[0103] State prediction: Output the evaluation results, such as "normal", "abnormal", or specific numerical predictions.

[0104] In a neural network model, the output of a neuron is usually achieved through a series of linear and non-linear transformations. The basic expression of a neural network can be described as: , Here, l represents the number of layers in the network. The following will explain each parameter in the expression in detail: a(l): This is the activation value or output of the l-th layer. For the output layer, a(l) is the final output of the neural network.

[0105] W(l): This is the weight matrix of the l-th layer. The weight matrix is one of the trainable parameters in the neural network and determines how the input features are linearly combined into the input of the next layer. The size of W(l) is usually "the number of neurons in the l-th layer × the number of neurons in the l-th layer".

[0106] b(l): This is the bias vector of the l-th layer. The bias is another trainable parameter that is added to the linear combination to allow the model to be adjusted without relying on the input being zero.

[0107] The dimension of b(l) is usually equal to the number of neurons in the l-th layer.

[0108] : This is the activation value or output of the l-th layer, that is, the input of the l-th layer. For the input layer, a(0) is the input feature vector.

[0109] : This is the activation function, which is applied to the result of the linear combination of each neuron. The activation function introduces non-linearity, enabling the neural network to learn complex patterns. Commonly used activation functions include ReLU (Rectified Linear Unit), Sigmoid (Logistic function), and Tanh (Hyperbolic Tangent function), etc.

[0110] Specifically, the calculation of a neural network can be divided into the following steps: Linear transformation: First, perform matrix multiplication on the input vector and the weight matrix, and then add the bias, that is .

[0111] Nonlinear transformation: Pass the result of the linear transformation through an activation function, i.e., a(l) = f(z(l)), to introduce non-linearity.

[0112] Multi-layer propagation: Repeatedly apply the above steps to each layer of the neural network until the output layer is reached.

[0113] Throughout the neural network, these steps allow the model to extract features from the input and map them to the output. During training, through the backpropagation algorithm, the weights and biases of the network are adjusted according to the loss function features to minimize the output error.

[0114] (3) Input the fused information into the state evaluation model to obtain the state score or state level of the transformer.

[0115] The state scoring system is set as follows: Assume that the range of scores from the state evaluation model is from 0 to 100, and the higher the score, the better the state of the transformer. According to the score, the state of the transformer can be divided into the following levels: Excellent (80 - 100): The transformer is in good operating condition and no additional treatment is required.

[0116] Good (60 - 79): The transformer is in relatively good operating condition, and regular monitoring is recommended.

[0117] Fair (40 - 59): The transformer shows slight abnormalities, and maintenance is recommended.

[0118] Poor (20 - 39): The transformer is in poor condition, and repair is recommended as soon as possible.

[0119] Dangerous (0 - 19): The transformer may have serious faults, and operation should be stopped immediately for repair.

[0120] Case analysis and score calculation Assume that a set of key data of the transformer, including temperature, voltage, current, and vibration frequency, etc., is collected through sensors. After preprocessing and feature extraction, these data are input into the state evaluation model, and the state score output by the model is 45.

[0121] According to the processing result of the score The score is 45, belonging to the "Fair" state (40 - 59): Processing result: The transformer shows slight abnormalities. This may be due to high temperature or abnormal vibration frequency. It is recommended to arrange technicians for on-site inspection and maintenance to prevent potential problems. This can include cleaning the radiator, checking the quality of the insulating oil, ensuring the stability of the wiring, etc.

[0122] Processing measures for other scores Excellent (80 - 100): Processing result: No special treatment is required. Continue to run according to the plan. Regular routine inspections are sufficient.

[0123] Good (60 - 79): Processing result: The transformer is in good condition and can continue to operate. However, it is recommended to increase the monitoring frequency to respond promptly when the condition changes.

[0124] Poor (20 - 39): Processing result: The condition is not good. There may be problems such as component wear and insulation aging. Shut - down for maintenance should be arranged as soon as possible, and special attention should be paid to the condition of key components, such as windings and iron cores.

[0125] Dangerous (0 - 19): Processing result: The transformer may have serious faults, such as overheating and short - circuit risks. It should be stopped immediately for a comprehensive inspection and repair to avoid equipment damage or safety accidents.

[0126] The status level system is set as follows: Suppose there are the following key transformer status indicators and their weights for status scoring: Temperature: maximum 100 points, accounting for 30% of the total score Voltage: maximum 100 points, accounting for 25% of the total score Current: maximum 100 points, accounting for 25% of the total score Vibration frequency: maximum 100 points, accounting for 20% of the total score The actual scores obtained through sensors and data processing are as follows: Temperature score: 70 Voltage score: 80 Current score: 75 Vibration frequency score: 60 Calculate the total score According to the weights, calculate the total status score of the transformer: Total score = 0.3×70 + 0.25×80 + 0.25×75 + 0.2×60 Total score = 21 + 20 + 18.75 + 12 = 71.75 Determine the status level According to the total score, we can classify the transformer status level as follows: Excellent (80 - 100) Good (60 - 79) Fair (40 - 59) Poor (20 - 39) Dangerous (0 - 19) In this case, the total score is 71.75, belonging to the "good" level.

[0127] Treatment results for this level Good (60 - 79): Treatment results: The transformer is in good condition and can continue to operate. It is recommended to increase the monitoring frequency, for example, check the key indicators once a week, to ensure timely response when the condition changes. Quarterly maintenance can be considered to ensure long-term safe operation.

[0128] Treatment measures under different levels Excellent (80 - 100): Treatment results: No special treatment is required, and continue to operate according to the plan. Regular routine inspections can be carried out, for example, a comprehensive inspection once a quarter.

[0129] Good (60 - 79): Treatment results: The transformer is in good condition and can continue to operate, but it is recommended to increase the monitoring frequency to ensure timely capture of changes in the condition.

[0130] Average (40 - 59): Treatment results: The transformer shows slight abnormalities, which may be caused by environmental factors or equipment aging. It is recommended to arrange maintenance, such as cleaning the radiator, checking the wiring and insulation conditions.

[0131] Poor (20 - 39): Treatment results: The condition is not good. It is recommended to arrange shutdown for maintenance as soon as possible, and focus on checking key components, such as transformer windings and iron cores, to prevent the expansion of faults.

[0132] Dangerous (0 - 19): Treatment results: The transformer may have a serious fault risk. It should be immediately stopped and a comprehensive inspection and repair should be carried out to avoid equipment damage or safety accidents.

[0133] In this embodiment, the transformer condition is comprehensively evaluated from multiple dimensions such as sound characteristics and electrical parameters. Sound characteristics can reflect the state information such as mechanical vibration and electromagnetic noise inside the transformer, providing richer references for evaluation, thereby improving the comprehensiveness and accuracy of the evaluation.

[0134] In this embodiment, as Figures 3 - 4 shown, it also includes a visualization display module for graphically or tabularly displaying the evaluation results of the transformer operating state. Preferably, the visualization display module includes an alarm prompt unit, which triggers an alarm prompt when the evaluation results of the transformer operating state reach a preset alarm threshold.

[0135] In this embodiment, the visualization display module is responsible for converting complex evaluation data and results into an intuitive and easy-to-understand graphical or tabular form, specifically: (1) Design a dynamically updated chart to display the real-time status of the transformer, including the changing trends of electrical parameters, temperature distribution, and sound characteristics.

[0136] (2) Present the evaluation results output by the status evaluation model in a graphical way, such as status scores, status levels, etc.

[0137] (3) Provide a comparison function for historical data to help users analyze the performance changes of the transformer.

[0138] The following are some specific examples: Load change analysis: Users can view the comparison of load characteristics in different time periods (such as daily, weekly, monthly). Assuming that the load of the transformer has gradually increased in the past few months, users can analyze whether the transformer can operate safely under the current configuration or whether capacity expansion is needed by comparing the load curves.

[0139] Temperature change trend: By comparing the temperature data of each area of the transformer, users can identify the trend of temperature increase. For example, if the winding temperature has been continuously higher than the normal range in the past few months, it may indicate that the transformer has an overload or insulation problem, and users can perform maintenance accordingly.

[0140] Electrical parameter fluctuation: Users can analyze the fluctuations of electrical parameters such as voltage, current, and power factor. For example, if it is found that the current fluctuates frequently and is higher than the historical average value during a certain period, it may reflect signs of unstable transformer load or equipment failure.

[0141] Sound characteristic comparison: By making a historical comparison of the sound characteristics generated during the operation of the transformer, users can identify changes in abnormal sounds. For example, if the humming sound emitted by the transformer becomes more obvious during a certain period, it may indicate mechanical wear or failure, and further inspection is required.

[0142] Failure rate analysis: Users can analyze the frequency and type of faults through historical feature comparison. For example, if the failure rate of a certain type of transformer has increased significantly in the past year, users can consider evaluating and improving this type to avoid future failures.

[0143] (4) When the evaluation result shows an abnormality, automatically trigger an alarm prompt to remind the user to take corresponding measures.

[0144] In the above embodiments of the present invention, a visualization display function is provided, which can display the state evaluation results of the transformer in a graphical or tabular form, enabling users to intuitively understand the operating state and potential risks of the transformer. In addition, the system can automatically trigger alarm prompts according to the evaluation results, realizing intelligent operation and maintenance management.

[0145] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various deformations or modifications within the scope of the claims, which does not affect the essence of the present invention. The above preferred features can be used in any combination without conflict.

Claims

1. A transformer operating state evaluation system, characterized in that Including: An on-line monitoring module, which continuously and real-timely collects n kinds of operation data of the transformer, where n≥1; A state evaluation module, which first evaluates the n kinds of operation data respectively based on a first state evaluation model to obtain n single-parameter evaluation results corresponding one by one to the n kinds of operation data, then performs multi-source information fusion on the n single-parameter evaluation results to obtain fusion information, and finally inputs the fusion information into a second state evaluation model for quantitative evaluation to obtain a transformer operation state evaluation result, where the first state evaluation model has a parameter θ; An incremental learning module, which is used to input the n kinds of operation data continuously received from the on-line monitoring module into an incremental learning algorithm formula respectively to iteratively optimize the parameter θ of the first state evaluation model until a preset iteration stop condition is met, and obtain an updated value θ' of the parameter θ; The first state evaluation model obtains each of the single-parameter evaluation results based on the updated value θ'.

2. The transformer operation state evaluation system according to claim 1, wherein The formula of the incremental learning algorithm is ; In the formula, is the parameter of the first state evaluation model at time t, η is the learning rate, is the gradient of the loss function L with respect to the parameter θ, is the operation data input to the incremental learning module at time t, is the true label at time t.

3. The transformer operation state evaluation system according to claim 2, characterized in that, The iteration stop condition is: After T iterations, the parameter θ is updated from to ; Alternatively, the iteration stop condition is that the value of the loss function L is less than a preset threshold.

4. The transformer operation state evaluation system according to claim 1, characterized in that It further includes a sound field analysis module, which is used to analyze the sound characteristics during the operation of the transformer and input the sound characteristics into the first state evaluation model to obtain a sound characteristic evaluation result.

5. The transformer operation state evaluation system according to claim 4, wherein, The sound field analysis module includes a sound characteristic extraction unit and a sound characteristic matching unit; The sound characteristic extraction unit is used to extract characteristic information from the sound generated during the operation of the transformer; a sound characteristic library is preset in the sound characteristic matching unit, and the sound characteristic library includes a plurality of feature pairs (c, d), where c is a sound element and d is the transformer state corresponding to c. The extracted characteristic information is compared with each of the sound elements respectively until they match, and then the transformer state corresponding to the characteristic information is the transformer state corresponding to the matched characteristic element.

6. The transformer operation state evaluation system according to any one of claims 1-5, characterized in that, It further includes a visualization display module, which is used to graphically or tabularly display the transformer operation state evaluation result.

7. The transformer operation status evaluation system according to claim 6, wherein The visualization display module includes an alarm prompt unit, which triggers an alarm prompt when the transformer operation state evaluation result reaches a preset alarm threshold.

8. A method for evaluating the operating state of a transformer, characterized in that, Including: Data collection: Continuously and real-timely collect n kinds of operation data of the transformer online, where n≥1; State evaluation: First evaluate the n kinds of operation data respectively based on a first state evaluation model to obtain n single-parameter evaluation results; then perform multi-source information fusion on the n single-parameter evaluation results to obtain fusion information; finally input the fusion information into a second state evaluation model for quantitative evaluation to obtain a transformer operation state evaluation result; the first state evaluation model has a parameter θ; Parameter update: Input at least one of the operation data continuously received from the on-line monitoring module into an incremental learning algorithm formula to iteratively optimize the parameter θ until a preset iteration stop condition is met, and obtain an updated value θ' of the parameter θ; The first state evaluation module obtains the single-parameter evaluation result based on the updated value θ'.

9. The transformer operation state evaluation method according to claim 8, wherein The formula of the incremental learning algorithm is ; In the formula, is the parameter of the model at time t, η is the learning rate, is the gradient of the loss function L with respect to the parameter θ, is the data input at time t, is the true label at time t.

10. The transformer operation state evaluation method according to claim 9, characterized in that, The iteration stop condition is: After T iterations, the parameter θ is updated from to ; Alternatively, the iteration stop condition is that the value of the loss function L is less than a preset threshold.

Images