Power data imbalance fault prediction method based on improved SMOTE algorithm

Abstract

The invention discloses a power data imbalance fault prediction method based on an improved SMOTE algorithm, and the method comprises the steps: carrying out the intelligent selection of a neighbor through fusing the Euclidean distance, the local density of a sample, and the composite measurement of physical state information defined by an autoencoder reconstruction error, and carrying out the interpolation, and generating a candidate synthesis sample; according to a preset power system physical constraint model, for candidate samples which do not meet physical authenticity, correcting the candidate samples into physically effective final synthetic samples by solving an optimization problem taking a minimum correction constant as a target; and finally, training a fault prediction model by using a data set containing the high-fidelity final synthetic sample. According to the method, the physical authenticity of the synthetic sample is ensured, data pollution is avoided, and the accuracy, generalization ability and reliability of a data-driven fault prediction model in actual power system application are improved.

Description

Technical Field

[0001] This invention relates to the field of fault prediction technology, and in particular to a method for predicting power data imbalance faults based on an improved SMOTE algorithm. Background Technology

[0002] With the deepening development of smart grids, the operation and management model of power systems is rapidly evolving from traditional planned maintenance to condition-based predictive maintenance (PdM). The core driving force behind this evolution lies in utilizing massive amounts of sensor data and data-driven analysis and prediction models to achieve refined management of the entire lifecycle of power equipment, optimize maintenance resource allocation, and reduce the risk of unplanned outages. Against this backdrop, fault prediction methods based on machine learning and deep learning have become a research hotspot. These methods can build intelligent models that provide early warnings of potential faults by learning the complex mapping relationship between historical operating data and fault events. However, the performance of these methods is highly dependent on the quality of the training data, especially when dealing with the inherent data imbalance problem in power systems, where the number of normal operating samples far exceeds the number of fault samples, leading to model training bias towards the majority class and a severe deficiency in the ability to identify minority class fault samples. To alleviate this problem, data augmentation algorithms, represented by Synthetic Minority Oversampling Technique (SMOTE), have emerged. This algorithm artificially synthesizes new samples by performing linear interpolation between minority class samples, thereby balancing the dataset.

[0003] However, existing SMOTE-based data augmentation methods often have limitations when applied to engineering systems with strong physical constraints (such as power systems). The standard SMOTE algorithm performs "blind" linear interpolation in the feature space, completely ignoring the underlying physical laws that variables must follow, such as Kirchhoff's laws and power balance relationships. This approach easily generates "pseudo-samples" that are mathematically valid but physically impossible. Injecting these physically unconstrained pseudo-samples into the training set not only fails to effectively improve the model's generalization ability but also pollutes the data, misleading the model to learn incorrect or distorted feature relationships. This causes the decision boundary to deviate from the true physical manifold, ultimately reducing the model's prediction accuracy and reliability under real-world conditions. Therefore, ensuring the physical authenticity of synthetic samples while maintaining statistical similarity has become a technical bottleneck restricting the effectiveness of data-driven power fault prediction methods in critical management decisions. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a power data imbalance fault prediction method based on an improved SMOTE algorithm to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a power data imbalance fault prediction method based on an improved SMOTE algorithm, comprising:

[0007] S1: Obtain power equipment feature data containing minority fault samples, and generate a candidate synthetic sample by interpolating the minority fault samples;

[0008] S2: Based on the preset power system physical constraint model, determine whether the candidate synthetic sample meets the physical authenticity. When the candidate synthetic sample does not meet the physical authenticity, correct it to meet the physical constraint model and obtain a final synthetic sample.

[0009] S3: Using the dataset containing the final synthetic samples, train a fault prediction model.

[0010] As a preferred embodiment of the power data imbalance fault prediction method based on the improved SMOTE algorithm described in this invention, wherein: the interpolation in S1 includes:

[0011] Linear interpolation is performed between a selected minority class sample and its nearest neighbor samples.

[0012] As a preferred embodiment of the power data imbalance fault prediction method based on the improved SMOTE algorithm described in this invention, the nearest neighbor sample is determined by a composite metric that integrates Euclidean distance, sample local density information, and sample physical state information.

[0013] As a preferred embodiment of the power data imbalance fault prediction method based on the improved SMOTE algorithm described in this invention, wherein the correction in S2 includes:

[0014] The correction is to solve an optimization problem;

[0015] The objective of the optimization problem is to minimize the distance between the candidate synthetic sample and the corrected sample, with the constraint that the corrected sample satisfies the physical constraint model.

[0016] As a preferred embodiment of the power data imbalance fault prediction method based on the improved SMOTE algorithm described in this invention, the physical constraint model includes at least one expression of the following constraints:

[0017] Three-phase system component constraints, power balance constraints, or chemical or physical relationship constraints between equipment characteristic parameters.

[0018] As a preferred embodiment of the power data imbalance fault prediction method based on the improved SMOTE algorithm described in this invention, the method further includes the following steps before the execution of step S3:

[0019] Feature selection is performed on the dataset to filter out the feature subset that contributes the most to fault prediction.

[0020] As a preferred embodiment of the power data imbalance fault prediction method based on the improved SMOTE algorithm described in this invention, the fault prediction model is a neural network model based on a self-attention mechanism.

[0021] As a preferred embodiment of the power data imbalance fault prediction method based on the improved SMOTE algorithm described in this invention, the method further includes, before training a fault prediction model:

[0022] Using an autoencoder model trained on health status data, a health index is defined for each sample by calculating the reconstruction error of the feature data.

[0023] As a preferred embodiment of the power data imbalance fault prediction method based on the improved SMOTE algorithm described in this invention, wherein: the S2 is applied to generate equipment degradation trajectory samples, the process of which includes:

[0024] The interpolation and correction operations are performed between subsets of samples representing different health levels based on health indices to generate synthetic samples simulating continuous degradation of device performance.

[0025] As a preferred embodiment of the power data imbalance fault prediction method based on the improved SMOTE algorithm described in this invention, a health management model is trained using the degradation trajectory samples, with the historical health index sequence of the equipment as input and the remaining service life of the equipment as output.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] 1. By introducing a two-step process of "interpolation followed by correction," and especially by projecting candidate samples that violate physical constraints into the physically feasible region through solving a constrained optimization problem, this approach fundamentally ensures that every final synthesized sample strictly adheres to the underlying physical laws of the power system (such as power balance, three-phase component constraints, and chemical relationships). Compared to the existing SMOTE method, which "blindly" interpolates and generates a large number of physical pseudo-samples, this method effectively avoids contaminating the training set and ensures the high fidelity of the augmented dataset.

[0028] 2. Because the training data is physically real and reliable, the feature relationships and decision boundaries learned by the model can more accurately reflect the real fault mechanisms of power equipment, rather than being misled by spurious samples. This gives the model stronger generalization ability and higher prediction accuracy when facing real-world, unseen data, while also effectively reducing the false alarm or missed alarm rate caused by model misjudgment, thus enhancing the reliability of prediction results.

[0029] 3. In selecting interpolation nearest neighbors, this invention employs a composite metric that integrates Euclidean distance, local sample density information (safety), and physical state information (health index) defined by the autoencoder. This makes interpolation no longer limited to geometric proximity, but rather tends to be performed between "safe" and "physically similar" samples, avoiding the generation of meaningless data around noise points or between samples with vastly different health states, thus improving the quality of candidate synthetic samples from the source. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0031] Figure 1 This is a flowchart illustrating the overall process of a power data imbalance fault prediction method based on an improved SMOTE algorithm, as described in one embodiment of the present invention. Detailed Implementation

[0032] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0033] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0034] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0035] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0036] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0037] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0038] Example 1

[0039] Reference Figure 1 This is the first embodiment of the present invention, which provides a method for predicting power data imbalance faults based on an improved SMOTE algorithm, including:

[0040] S1: Obtain power equipment feature data containing minority fault samples, and generate a candidate synthetic sample by interpolating the minority fault samples.

[0041] Furthermore, historical operational datasets of the power equipment are obtained. These datasets can originate from the power grid's SCADA system, PMU devices, or online monitoring systems for specific equipment (such as online monitoring of dissolved gas in transformer oil using DGA). These datasets are then combined into a multi-dimensional time series. Features characterizing the equipment's operating status are extracted from this multi-dimensional time series to obtain a feature vector. In this embodiment, taking a transformer fault prediction scenario as an example, the feature vector may include: voltage, current, active / reactive power, winding temperature of each phase, oil level, and the concentrations of various gas components monitored by DGA (such as hydrogen). methane Acetylene ethylene (etc.). Combining these feature vectors with their corresponding labels (e.g., "normal", "winding short circuit fault", "partial discharge fault", etc.) yields the original feature dataset. The original feature dataset The dataset exhibits typical data imbalance characteristics, meaning that samples labeled "normal" constitute the majority class, while samples labeled with various fault categories constitute the minority class. To eliminate the influence of different feature units, this feature dataset needs to be... Perform traditional normalization, such as using max-min normalization, to scale all feature values ​​to the [0,1] interval.

[0042] Furthermore, in order to perceive the differences in physical state between samples during the subsequent interpolation process, the present invention introduces a health index (HI) as a quantitative representation of the physical state of the samples.

[0043] Specifically, a deep autoencoder model is constructed. This model consists of an encoder and a decoder, both of which are multi-layer fully connected neural networks. The encoder is composed of three fully connected layers, and the network dimension changes to the normalized feature vector. The dimensions → 64 → 32. That is, through the first layer... The input dimension is compressed to 64 dimensions, and then further compressed to a 32-dimensional latent space representation through a second layer. Except for the output layer, all layers of the encoder use ReLU as the activation function. The decoder structure is symmetrical to the encoder, consisting of three fully connected layers, with the network dimension changing from 32 to 64 to the normalized feature vector. The goal is to reconstruct the original feature vector from the 32-dimensional latent space representation. The dimension is [0, 1]. Similarly, in the decoder, all layers except the output layer use the ReLU activation function, and the output layer uses the Sigmoid activation function to ensure that the reconstructed values ​​are in the range [0, 1], corresponding to the normalized input. Based on the above structure, the encoder receives a normalized feature vector. As input, it is compressed into a low-dimensional latent space representation through several fully connected layers with non-linear activation functions (such as ReLU). The decoder then receives... By using a fully connected layer whose structure is symmetrical to the encoder, an attempt is made to reconstruct the original input vector, resulting in a reconstructed vector. It is important to emphasize that this autoencoder is trained only on samples from the dataset that belong to the majority class (i.e., the healthy state). The goal of training is to minimize the reconstruction error between the input and the reconstructed output, for example, using a mean squared error loss function. Through this training, the model learns the "normal" manifold structure of healthy samples in the feature space. Once training is complete, for any sample in the dataset... (Whether it's a healthy sample or a faulty sample), its health index All of these are defined as the reconstruction error after passing through the autoencoder, which is mathematically expressed as:

[0044]

[0045] in, It is a sample Normalized feature vectors This represents the trained autoencoder model. That is, the reconstructed vector , This represents the square of the L2 norm (i.e., the square of the Euclidean distance).

[0046] It should be noted that in this mathematical expression, the smaller the health index value of a sample, the closer its characteristic pattern is to a healthy state, and the healthier the equipment; the larger the health index value, the more it deviates from the normal pattern, indicating that the equipment has an abnormality or malfunction, and the value can reflect the degree of deviation. Therefore, this health index value can also be used as information about the physical state of the sample.

[0047] Furthermore, by iterating through each minority class sample, one or more candidate synthetic samples are generated for it. For a selected minority class sample... The execution process is as follows:

[0048] S101. The present invention does not use the traditional SMOTE algorithm, which only uses Euclidean distance to find nearest neighbors, but instead uses a composite metric. To evaluate Similar neighboring candidates The "comprehensive distance" between them. This composite metric aims to simultaneously satisfy geometric proximity, data distribution security, and physical similarity, and is expressed as:

[0049]

[0050] in, for and The normalized Euclidean distance between them is a fundamental term that ensures geometric closeness. This is a measure of the difference in physical state between two points, which can be directly represented by the absolute value of the difference in health index calculated in the previous step: This feature enables the SMOTE algorithm to tend to select samples with similar health status (or degree of deterioration) for interpolation, avoiding the generation of meaningless intermediate states between "early failure" and "severe failure". It is a sample The local density safety factor, which is used to evaluate Whether a point is a "safe" intra-class point or potentially a noise point, in this embodiment, we use the Local Outlier Factor (LOF) to calculate this coefficient, and the process is as follows: First, calculate... The LOF value is used to determine the likelihood of an outlier (noise). A higher LOF value indicates a greater likelihood of an outlier (noise). A safety factor is then set. Inversely proportional to the LOF value, in this embodiment, it can be calculated using the following function: The higher the safety factor, the better. The higher the safety factor, the smaller the overall distance. In the composite metric formula, subtracting the weighted value of the safety factor means that a point with a higher safety factor has a smaller overall distance, and thus a greater probability of being selected as a nearest neighbor. These are three non-negative weighted hyperparameters, summing to 1. These three weights control the relative importance of Euclidean distance, physical state difference, and sample safety in nearest neighbor selection, respectively. Technicians can adjust these weights according to specific application scenarios. For example, in a scenario with extremely high requirements for the continuity of physical state evolution (such as predicting bearing wear), the weights can be appropriately increased. Weights, such as setting In a scenario with a lot of noisy data, it can improve... The weight is used to prioritize the selection of safe points, such as setting... In a typical case of this embodiment, the equalization setting can be configured as follows: .

[0051] S102, Based on the calculated composite metric ,from Among all similar samples, find the one with the smallest "comprehensive distance". Each sample constitutes of -Nearest neighbor set.

[0052] S103, from this - Randomly select a nearest neighbor sample from the nearest neighbor set. and in and Linear interpolation is performed between them to generate a candidate synthetic sample. :

[0053]

[0054] in, It is a random number uniformly sampled from the interval (0,1).

[0055] It should be noted that the present invention can complete a candidate synthetic sample through the above steps. The generation of candidate synthetic samples is guided by both data distribution security and physical state similarity, ensuring that the sample lies between two real minority class samples in the feature space. This provides a basis for subsequent physical constraint testing. By repeating the above process on all minority class samples, the required number of candidate synthetic samples can be generated.

[0056] S2: Based on the preset power system physical constraint model, determine whether the candidate synthetic sample meets the physical authenticity. When the candidate synthetic sample does not meet the physical authenticity, correct it to meet the physical constraint model and obtain a final synthetic sample.

[0057] Furthermore, the generally accepted physical laws, equipment operating mechanisms, or expert experience in the power system field are transformed into mathematical constraint expressions that can be used to calculate and judge the feature vectors. These constraints constitute a "feasible region" in the feature space, within which all physically possible real data points should fall. Based on this, this embodiment uses the following three typical power system physical constraint models as examples, but is not limited to them:

[0058] The first constraint model: based on power balance constraints. For the same monitoring point, active power... reactive power and apparent power The Pythagorean theorem must be satisfied between them. If the eigenvector contains these three quantities, then the physical constraint can be expressed as an equality constraint or an inequality constraint that takes into account measurement errors:

[0059]

[0060] in, It is a positive number close to 0, used to tolerate measurement and calculation errors.

[0061] It should be noted that, under this inequality constraint, any sample point that does not satisfy this relationship is physically invalid.

[0062] The second constraint: three-phase system component constraint. In a three-phase balanced system, the zero-sequence component (voltage or current) should theoretically be zero. Therefore, for a system containing three-phase currents... The eigenvectors of can have their zero-order components expressed as . However, considering the slight imbalances that may exist in actual operation, this constraint can be expressed as:

[0063]

[0064] in, It is a threshold set according to the unbalance requirements of a three-phase balanced system.

[0065] The third constraint: Chemical relationship constraint of dissolved gases (DGA) in transformer oil. Based on the Duval triangle or pentagon method, characteristic gases (such as methane) produced by different fault types (e.g., overheating, arc discharge) are considered. Acetylene ethylene The relative proportions of gases (such as methane, etc.) lie within a specific geometric region. For example, the Duval triangle divides the fault region into 7 parts, requiring the percentage of gas components (methane, etc.) at a sample point. %, acetylene %, ethylene The point (%) must lie within the plane of the triangle. Therefore, this can be transformed into a set of linear inequality constraints, for example, the point (%). The conditions inside the triangle are:

[0066]

[0067] in, The coordinates are derived from the percentage conversion of the three gases.

[0068] It should be noted that, under this constraint, the proportion of gas components at points outside this region is chemically unreasonable.

[0069] Furthermore, by combining all the physical constraints applicable to the current scenario, a set of physical constraint models can be formed. Each of them They are all in the shape of or The mathematical expression for .

[0070] Furthermore, for each generated candidate synthetic sample Substitute it into the preset set of physical constraint models Then, each item is checked individually. If... Satisfy all constraints, that is, for all They all If the condition is met, the sample is determined to be a physically valid sample and is directly used as the final synthesized sample. Then proceed to the next step. If at least one constraint exists... If this condition is not met, then the candidate synthetic sample is deemed invalid. This violates physical authenticity and a correction process needs to be initiated.

[0071] Furthermore, the purpose of the correction is not simply to discard invalid samples, nor to arbitrarily pull them back into the valid region, but to find a point within the physically feasible region that is closest to the original candidate synthesized sample as the final corrected sample. Utilizing this principle of minimum perturbation, the original data distribution information from the interpolation in step S1 can be preserved to the greatest extent possible. Based on this, we model this correction procedure as a quadratic programming (QP) problem or a general constrained optimization problem.

[0072] Specifically, for a candidate synthetic sample that does not meet the physical constraints The final synthesized sample can be obtained by solving the following optimization problem. :

[0073]

[0074] in, It is the objective function, which minimizes the final sample With candidate synthetic samples The square of the Euclidean distance between the points is geometrically equivalent to the square of the distance between the points. Orthogonally projected onto the boundary or interior of the "feasible region" defined by all physical constraints. These are the corrected sample points that we need to solve for. The constraint condition requires the solution to be... All the physical constraints constructed above must be satisfied.

[0075] It should be noted that, through the above-described modification method, compared to traditional heuristic rules, it is possible to find a theoretically unique, distance-based... The nearest efficient point maximizes the preservation of the original interpolation direction, avoiding the introduction of new biases due to over-correction and ensuring the fidelity of the synthesized samples. Furthermore, because the framework is universal, regardless of whether the physical constraints are linear or nonlinear, equal or inequalities, they can all be uniformly incorporated into the constraints of this optimization problem. When new physical knowledge needs to be added, only new mathematical expressions need to be added to the constraint set; the entire method framework remains unchanged, demonstrating scalability. In addition, when the physical constraints are linear and quadratic, the above problem becomes a standard quadratic programming problem, which can be efficiently solved using mature optimization solvers (such as CVXPY, Gurobi, etc.). The optimal solution obtained is... This results in the final synthetic sample that satisfies both physical realism and retains the original interpolation information, which is what we need. Repeating this process for all generated candidate synthetic samples yields the final augmented dataset.

[0076] S3: Train a fault prediction model using a dataset containing the final synthetic samples.

[0077] It should be noted that although the original feature set already contains rich information, not all features contribute equally to the prediction of a specific fault. Redundant or irrelevant features may increase model complexity, prolong training time, or even introduce noise, thus degrading model performance. Therefore, a feature selection step will be performed before formally training the model.

[0078] Specifically, in this embodiment, we employ a feature selection method based on mutual information. This method selects each feature from the augmented dataset... Calculate its relationship with the target fault label Mutual information values ​​between It should be explained that mutual information can measure the amount of information that one feature brings to another variable, thus effectively capturing linear and nonlinear relationships.

[0079] Furthermore, after calculating the mutual information values ​​of all features, a threshold is set, or the top mutual information values ​​are directly selected. The features constitute the final feature subset used for model training. For example, it can be set We selected 20 key features that are most discriminative for fault classification.

[0080] Furthermore, the operational feature vectors of the equipment are mapped to specific fault categories. Since power equipment operational data is typically time-series, we chose to construct a classification model based on a Transformer encoder.

[0081] Specifically, the input to this classification model is a sequence of device operation data within a time window. More specifically, this device operation data sequence is a continuous sequence of data... Feature vectors at each time point (after feature selection) Stacked together, forming a shape of size A matrix. For example, using sampled data from each minute over the past hour, then... .

[0082] Specifically, the output of this classification model is a representation within a time window. At the end of the output is a probability distribution vector of the device belonging to each fault category. For example, if there are three categories: "normal", "fault A", and "fault B", the output is a 3-dimensional vector. Their sum is 1.

[0083] Specifically, in this embodiment, the classification model is mainly composed of the following stacked modules:

[0084] Input embedding layer: Since the Transformer itself does not process sequence position information, a position encoding layer is first needed to add a unique position vector to each time point of the input sequence. Then, a fully connected layer (or a 1D convolutional layer) is used to encode each time point... 1D feature vectors are mapped to a higher-dimensional model internal representation space (e.g., dimension).

[0085] Transformer encoder layer: Each Transformer encoder layer consists of two main sub-modules:

[0086] Multi-head self-attention mechanism module: This module allows the classification model to simultaneously "attention" to information from all other time points in the sequence while processing a specific time point. It achieves this by calculating a query, key, and value matrix. For the input sequence... The calculation process can be simplified as follows:

[0087]

[0088] in, It is a matrix obtained by applying different linear transformations to the input of the classification model. It is the dimension of the key vector. The multi-head mechanism executes this process multiple times in parallel (e.g., ...). (size), and concatenate the results so that the classification model can learn different dependencies from different representation subspaces.

[0089] Feedforward Neural Network Module: This module is a simple network consisting of two fully connected layers and a ReLU activation function in the middle. It is mainly used to perform further nonlinear transformations on the output of the multi-head self-attention mechanism module.

[0090] Specifically, the dimensionality changes of this feedforward neural network module need to follow the standard Transformer settings, i.e. In this embodiment, The dimensions of the middle layer It is usually set to 4× That is, 512. Therefore, the dimension changes from 128 to 512 to 128.

[0091] It is important to emphasize that the output of each submodule of the Transformer encoder layer described above needs to undergo residual connections and layer normalization to stabilize the training process. In this embodiment, four such Transformer encoder layers can be stacked.

[0092] Output layer: The sequence representation output by the last Transformer encoder layer (usually the output of the first time step, or global average pooling of the entire sequence) is fed into a linear layer whose output dimension is equal to the number of fault categories. A Softmax activation function is then applied to obtain the final probability distribution.

[0093] Specifically, for the model training dataset, a training set obtained by merging the original dataset with the final synthetic samples generated through steps S1 and S2 is used. Meanwhile, for the model loss, a traditional cross-entropy loss function is employed, equipped with the Adam optimizer, and the learning rate is set to... The training set is divided into multiple batches, and iterative training is performed over multiple epochs until the performance on the validation set no longer improves (early stopping).

[0094] It should be noted that the Transformer model based on self-attention was chosen because it is particularly good at capturing complex temporal dependencies in long-sequence data, which in turn correspond to the typical characteristics of power equipment fault evolution (for example, a winding fault may not be determined by instantaneous values, but rather by the coordinated change patterns of multiple parameters such as temperature and current over the past few minutes or hours).

[0095] Furthermore, the present invention utilizes the Health Index (HI) calculated for each sample in step S1 to sort and group all samples (including original and synthetic samples) according to their HI values. For example, a HI of [0, 0.2] can be defined as "healthy," [0.2, 0.5] as "sub-healthy," [0.5, 1.0] as "minor malfunction," and a HI > 1.0 as "significant malfunction." Secondly, to generate trajectory samples simulating continuous degradation of the device, steps S1 and S2 of the present invention are applied between subsets of samples with different health levels. For example, a "sub-healthy" sample is randomly selected as... And find its nearest neighbors in the set of "minor fault" samples. (Based on composite measurement) Then, candidate samples are generated through linear interpolation. The final synthetic sample was obtained by correcting the physical constraint model. Repeating this process generates a large number of physically realistic, continuous degradation trajectory samples that simulate the smooth transition of equipment from a sub-healthy state to a minor failure state. This effectively solves the problem of obtaining complete degradation process data in the real world. Finally, using these generated degradation trajectory samples, a health management model specifically for predicting the remaining useful life (RUL) of equipment is trained.

[0096] Furthermore, these generated degradation trajectory samples can be used to train a health management model specifically for predicting the remaining useful life of equipment.

[0097] It is important to emphasize that for this health management model, either a Long Short-Term Memory Network (LSTM) or a Gated Recurrent Unit (GRU) can be selected, as these variants of Recurrent Neural Networks (RNNs) are well-suited for processing and predicting sequence data.

[0098] Specifically, in this embodiment, a Long Short-Term Memory (LSTM) network is chosen as an example. This LSTM model consists of two stacked LSTM layers, with each LSTM layer having 64 hidden units. After the output of the second LSTM layer, a Dropout layer (with a dropout rate of 0.2) is connected to prevent overfitting. Finally, a fully connected layer outputs a scalar value, the predicted RUL. The input to this model is: the device's data over a past period (e.g., the past...). The model outputs a daily Health Index (HI) sequence for each day. The model's output is: the device's... The remaining useful life at any given time is the time remaining before the device reaches a preset "failure threshold" (e.g., health index > 1.5). When training this health management model, the generated degradation trajectory samples are used to label each time point with its actual remaining useful life, thus constituting a supervised learning problem. Using mean squared error (MSE) as the loss function and training the model with the Adam optimizer, a health management model capable of accurately predicting the remaining useful life of equipment can be obtained.

[0099] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0100] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0101] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0102] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0103] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0104] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A power data imbalance fault prediction method based on an improved SMOTE algorithm, characterized in that, include: S1: Obtain power equipment feature data containing minority fault samples, and generate a candidate synthetic sample by interpolating the minority fault samples; S2: Based on the preset power system physical constraint model, determine whether the candidate synthetic sample meets the physical authenticity. When the candidate synthetic sample does not meet the physical authenticity, correct it to meet the physical constraint model and obtain a final synthetic sample. S3: Using the dataset containing the final synthetic samples, train a fault prediction model.

2. The power data imbalance fault prediction method based on the improved SMOTE algorithm as described in claim 1, characterized in that, The interpolation in S1 includes: Linear interpolation is performed between a selected minority class sample and its nearest neighbor samples.

3. The power data imbalance fault prediction method based on the improved SMOTE algorithm as described in claim 2, characterized in that, The nearest neighbor samples are determined by a composite metric that integrates Euclidean distance, local density information of the samples, and physical state information of the samples.

4. The power data imbalance fault prediction method based on the improved SMOTE algorithm as described in claim 1, characterized in that, The correction in S2 includes: The correction is to solve an optimization problem; The objective of the optimization problem is to minimize the distance between the candidate synthetic sample and the corrected sample, with the constraint that the corrected sample satisfies the physical constraint model.

5. The power data imbalance fault prediction method based on the improved SMOTE algorithm as described in claim 1, characterized in that, The physical constraint model includes at least one of the following forms of expression for constraints: Three-phase system component constraints, power balance constraints, and chemical or physical relationship constraints between equipment characteristic parameters.

6. The power data imbalance fault prediction method based on the improved SMOTE algorithm as described in claim 1, characterized in that, The process also includes the following steps before S3 is executed: Feature selection is performed on the dataset to filter out the feature subset that contributes the most to fault prediction.

7. The power data imbalance fault prediction method based on the improved SMOTE algorithm as described in claim 1, characterized in that, The fault prediction model is a neural network model based on a self-attention mechanism.

8. The power data imbalance fault prediction method based on the improved SMOTE algorithm as described in claim 1, characterized in that, Before training a fault prediction model, the following is also included: Using an autoencoder model trained on health status data, a health index is defined for each sample by calculating the reconstruction error of the feature data.

9. The power data imbalance fault prediction method based on the improved SMOTE algorithm as described in claim 7, characterized in that, The process of applying S2 to generate device degradation trajectory samples includes: The interpolation and correction operations are performed between subsets of samples representing different health levels based on health indices to generate synthetic samples simulating continuous degradation of device performance.

10. The power data imbalance fault prediction method based on the improved SMOTE algorithm as described in claim 8, characterized in that, Using the degradation trajectory samples, a health management model is trained that takes the device's historical health index sequence as input and the device's remaining service life as output.

Images