Mine wind power supply system monitoring method

Abstract

The invention belongs to the technical field of power supply system monitoring, and particularly relates to a mine wind power supply system monitoring method. In order to solve the problem that the fault category of a training set is different from the fault category of a test set, the invention provides a zero sample learning-based fan fault diagnosis method. Comprising the following steps: S1, building a fan fault diagnosis framework considering unknown faults; s2, embedding a fan fault attribute space in the fault sample space and the fault category space, and establishing a mapping classification of fault samples, fault attributes and fault categories; s3, describing fault attributes to form a fault attribute feature set; s4, performing data preprocessing on the fault attribute feature set by adopting a supervised principal component analysis method, and extracting key fault features; s5, based on the preprocessed fault features, fault classification is carried out; and S6, dynamically updating the diagnosis capability by adopting an incremental diagnosis method.

Description

Technical Field

[0001] This invention belongs to the field of power supply system monitoring technology, and particularly relates to a monitoring method for a wind power supply system in a mine. Background Technology

[0002] Wind turbines are prone to failure in harsh operating environments such as storms, rain, and snow, causing significant economic losses. Prolonged shutdowns can even lead to power shortages, impacting people's livelihoods and the industrial economy. Therefore, it is necessary to conduct research on wind turbine fault diagnosis methods to promptly understand the operating status of wind turbines and ensure the safe and stable operation of the new power system.

[0003] In the field of wind turbine condition monitoring and fault diagnosis, there are two main categories of methods: mechanism-driven modeling and data-driven modeling. Mechanism-driven modeling requires simplifying the working principle of wind turbines to describe the relationship between model inputs and outputs in a deterministic and quantitative way. Data-driven modeling obtains diagnostic results by reasoning and analyzing a large amount of wind turbine data, and is currently the mainstream research method in this field.

[0004] Traditional data-driven wind turbine fault diagnosis methods generally fall into two categories. One involves setting thresholds for fault diagnosis; a deviation exceeding the threshold indicates an abnormal turbine condition. However, this approach heavily relies on human experience. The other approach uses supervised learning of the mapping relationship between fault samples and fault states via artificial intelligence algorithms. Both methods are based on historical fault datasets from the same turbine unit. However, in real-world operating environments, wind turbines may experience faults outside the training set. This means the training and test sets may contain different categories, and faults outside the training set could be types that have never occurred or have occurred but were not labeled. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a monitoring method for wind power supply systems in mines. To resolve the issue of discrepancies between fault categories in the training and test sets, it proposes a wind turbine fault diagnosis method based on zero-shot learning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a monitoring method for a mine wind power supply system, comprising the following steps:

[0007] S1. Establish a wind turbine fault diagnosis framework that considers unknown faults;

[0008] S2. Embed the wind turbine fault attribute space into the fault sample space and fault category space, and establish a mapping classification of fault sample → fault attribute → fault category.

[0009] S3. Describe the fault attributes and form a fault attribute feature set;

[0010] S4. Supervised principal component analysis is used to preprocess the fault attribute feature set and extract key fault features.

[0011] S5. Based on the preprocessed fault characteristics, classify the faults.

[0012] S6. Use incremental diagnostic methods to dynamically update diagnostic capabilities.

[0013] Furthermore, S1 specifically includes:

[0014] S1.1 Construct the equivalent circuit model of the grid-side converter of the direct-drive wind turbine, and establish the active power P g and reactive power Q g The equation is:

[0015] P g =P0+P1cos(2ω) g t)+P2sin(2ω g t)

[0016] Q g =Q0+Q1cos(2ω g t)+Q2sin(2ω g t)

[0017] in:

[0018]

[0019] In the formula, These are the positive and negative sequence components of the grid voltage. ω represents the positive and negative sequence components of the three-phase current on the grid side. g The angular frequency of the power grid;

[0020] S1.2, By real-time monitoring P g and Q g The amplitude deviations of the harmonic components (P1, P2, Q1, Q2) are used to establish a mapping relationship between power fluctuation parameters and fault characteristics.

[0021] Furthermore, S2 specifically includes:

[0022] S2.1 Based on the power fluctuation parameters in S1.2, generate fault attribute descriptions and construct a fault attribute matrix. Each row of the matrix corresponds to a fault sample, and each column corresponds to an attribute feature, including harmonic amplitude deviation and timing phase offset.

[0023] S2.2. A fault attribute learner is constructed using a first convolutional neural network (CNN), with the power fluctuation parameters from S1.2 as input and the fault attribute vector as output.

[0024] S2.3 Calculate the Euclidean distance between the attribute vector of the unknown fault sample and the true attribute vector. If the distance is less than the dynamic threshold, it is classified into the corresponding known fault category; otherwise, incremental learning is triggered.

[0025] Furthermore, the fault attribute feature set in S3 includes pitch system fault attributes, yaw system fault attributes, and general fault attributes;

[0026] Among them, the fault attributes of the pitch system include the abnormal angle detected by the pitch control cabinet or the safety chain alarm triggered by the excessive angle difference of the three blades; the periodic angle change and temperature rise characteristics caused by the damage of the pitch motor encoder, driver failure or bearing damage; and the temperature alarm signal caused by the frequent operation of the motor under wind speed fluctuations.

[0027] Yaw system fault attributes include excessive yaw acceleration caused by mechanical vibration or brake lock-up, and abnormal wind vane signal directly related to wind direction measurement deviation; general fault attributes include amplitude and timing phase offset of power harmonic components (P1, P2, Q1, Q2).

[0028] Furthermore, the method for generating general fault attributes is as follows:

[0029] Based on the power equation of the equivalent circuit of the grid-side converter of the direct-drive wind turbine, the harmonic components of P1, P2, Q1, and Q2 are extracted.

[0030] The time-domain power fluctuations are converted into frequency-domain features by normalization and Fast Fourier Transform (FFT) and used as fault attribute inputs.

[0031] Furthermore, the data preprocessing for the fault attribute feature set in S4 specifically includes:

[0032] S4.1 Based on the pitch system, yaw system and general fault attributes defined in S3, generate 12-dimensional attribute labels for each fault sample;

[0033] S4.2 Use three datasets, A / B / C, each including a training set and a test set, and normalize the original data;

[0034] The training set consists of 4 known faults × 800 samples × 25-dimensional original parameters (including S3 parameters such as pitch angle, motor temperature, and yaw acceleration).

[0035] The test set consists of: 2 new fault types × 400 samples × 25-dimensional original parameters;

[0036] S4.3. Supervised principal component analysis (SPCA) is used to reduce the dimensionality of the 25-dimensional original data to a 12-dimensional feature vector, where: the 12-dimensional features after dimensionality reduction correspond to the 12 types of fault attributes in S3 (including pitch, yaw and power harmonic features); during the dimensionality reduction process, feature correlation is calculated based on the fault attribute labels, and the 12 principal components with the strongest correlation with the labels are retained.

[0037] S4.4 After SPCA processing, each sample generates 144 intermediate features, and each group of 12 features corresponds to one fault attribute.

[0038] Furthermore, the 25-dimensional original data is reduced to a 12-dimensional feature vector, where: the first 3 dimensions correspond to propeller system attributes (angle, temperature, voltage); the fourth and fifth dimensions correspond to yaw system attributes (acceleration, wind vane); and the sixth to twelfth dimensions correspond to general electrical attributes (harmonic amplitude and phase).

[0039] Furthermore, the SPCA process in S4.3 includes:

[0040] a) Mapping the original fault features to a high-dimensional space based on radial basis functions (RBF);

[0041] b) Calculate the vector distance between the high-dimensional features and the 12-dimensional attribute labels in S4.1. The closer the distance, the stronger the correlation.

[0042] c) Select the 12 principal components whose cumulative variance contribution rate reaches 99% or more for the first time as the final features.

[0043] Furthermore, the preprocessed 144-dimensional features are input into a second convolutional neural network (CNN-2), and fault classification is achieved by calculating the Euclidean distance between the output attribute vector and the true vector; specifically including:

[0044] S5.1. Use the second convolutional neural network (CNN-2) as a fault diagnosis learner, with the input being the 144-dimensional intermediate features (12 groups × 12 dimensions) generated in S4.4;

[0045] S5.2. Learn the mapping relationship between fault samples and fault attributes through CNN, and output a 12-dimensional fault attribute prediction vector;

[0046] S5.3 Calculate the Euclidean distance between the predicted attribute vector (i.e., the true fault attribute vector) and the true attribute vector:

[0047]

[0048] In the formula, A i To predict the attribute vector, A' iLet i represent the fault attribute vector output by the fault diagnosis learner, where i represents the i-th sampling point, i = 0, 1, ... 999; j represents the j-th attribute in the fault attribute vector, j = 0, 1, ... 11;

[0049] S5.4 The fault classifier categorizes samples into the fault category with the smallest Euclidean distance.

[0050] Furthermore, the incremental diagnostic method for S6 specifically includes:

[0051] S6.1 Adaptive Sample Library Construction:

[0052] In each incremental phase b, from the training set D b Select Z representative samples to construct a sample library E b ;

[0053] By calculating sample features and class centers The Euclidean distance is used to select the Z-axis with the smallest distance in each class. b Each sample is stored in the sample library;

[0054] Dynamically eliminate samples from the old sample library that are farthest from the class center, keeping the library size Z constant;

[0055] S6.2, Dynamic Weight Correction:

[0056] The training weights are dynamically adjusted based on the model's prediction error for each fault category, with categories having higher training weights for categories with larger errors.

[0057] The model parameters are updated using a weighted cross-entropy loss function;

[0058] S6.3, Incremental Phase Execution:

[0059] In the initial stage, the model was trained using five types of basic fault data and an initial sample library was built.

[0060] In subsequent incremental phases, new fault data are added sequentially, and the model is fine-tuned by combining it with historical samples from the sample library.

[0061] The test set for each incremental phase includes all learned fault types.

[0062] Among them, CNN-1 is used to learn the mapping from power fluctuation parameters to fault attributes; CNN-2 is used to learn the mapping from dimensionality reduction features to fault classification; the parameters of the two networks are independent of each other.

[0063] Compared with the prior art, the present invention has the following advantages.

[0064] This invention inserts a fault attribute space into the fault sample space and fault category space based on a fault attribute description matrix, providing a common knowledge base for fault diagnosis that can characterize the commonalities and differences of different faults. Fault diagnosis is divided into two steps: establishing a mapping relationship from fault samples to fault attribute vectors, and then establishing a mapping relationship from fault attribute vectors to fault categories, thus avoiding the requirement in traditional machine learning to match fault training samples with fault test categories. Furthermore, the fault attribute learner based on CNN can effectively extract key features from fault samples. Choosing appropriate batch processing methods and network structures can improve the learning ability of the fault attribute learner and better fit nonlinear relationships. Attached Figure Description

[0065] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The scope of protection of the present invention is not limited to the following description.

[0066] Figure 1 This is the equivalent circuit diagram of the grid-side converter for a direct-drive wind turbine.

[0067] Figure 2 This is a framework diagram for wind turbine fault diagnosis that takes into account unknown faults.

[0068] Figure 3 This is a flowchart for diagnosing wind turbine faults.

[0069] Figure 4 This is a fault attribute matrix diagram of a wind turbine.

[0070] Figure 5 This is a diagram showing the structure and parameters of a CNN fault attribute learner.

[0071] Figure 6 This is an incremental diagnostic framework diagram. Detailed Implementation

[0072] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0073] The terminology used in the embodiments of this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of this disclosure. The singular forms “a,” “the,” and “the” as used in the embodiments of this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0074] Depending on the context, words such as “if” or “suppose” used here can be interpreted as “when”, “in response to determination”, or “in response to detection”.

[0075] For ease of understanding, the embodiments of this disclosure will be described in detail first.

[0076] like Figure 1-6 As shown, a monitoring method for a mine wind power supply system includes the following steps:

[0077] S1. Establish a wind turbine fault diagnosis framework that considers unknown faults;

[0078] S2. Embed the wind turbine fault attribute space into the fault sample space and fault category space, and establish a mapping classification of fault sample → fault attribute → fault category.

[0079] S3. Describe the fault attributes and form a fault attribute feature set;

[0080] S4. Supervised principal component analysis is used to preprocess the fault attribute feature set and extract key fault features.

[0081] S5. Based on the preprocessed fault characteristics, classify the faults.

[0082] S6. Use incremental diagnostic methods to dynamically update diagnostic capabilities.

[0083] In one embodiment, in S1.1, the equivalent circuit diagram of the direct-drive wind turbine grid-side converter is as follows: Figure 1 As shown in the figure, the positive direction of the current is marked, E a E b E c Indicates the three-phase voltage of the power grid; i ga i gb i gc U represents the three-phase current on the grid side; a U b U c This represents the three-phase output voltage of the inverter; L and R are the line inductance and resistance, respectively; U dc This is the DC side voltage.

[0084] Active power P g and reactive power Q g They are respectively:

[0085] P g =P0+P1cos(2ω) g t)+P2sin(2ω g t)

[0086] Q g =Q0+Q1cos(2ωg t)+Q2sin(2ω g t)

[0087] in:

[0088]

[0089] Due to the presence of DC capacitors, the generator side is minimally affected by grid voltage imbalances. Therefore, without any intervention, the generator side maintains its rated output power even under fault conditions. Figure 1 It is easy to see from the DC-side circuit shown that the operating characteristics of the DC side under unbalanced conditions can be described as follows:

[0090]

[0091] In the formula: C is the DC side capacitance value; P m This refers to the power input from the machine-side rectifier to the DC side. Therefore, even with a constant output power at the machine side, the DC-side capacitor voltage will oscillate at twice the power frequency due to unbalanced operating conditions. Traditional fault ride-through control strategies typically only consider three-phase symmetrical operating conditions and cannot suppress the aforementioned current and power imbalances.

[0092] In S1.2, a wind turbine fault attribute space is embedded in the fault sample space and fault category space, transforming the mapping relationship from fault sample → fault category to fault sample → fault attribute → fault category. This avoids directly classifying unknown faults. The specific framework is as follows: Figure 2 As shown, the fault diagnosis framework first analyzes the possible fault types of wind turbines, providing a description for each fault and forming a wind turbine fault attribute matrix containing all fault attributes. The wind turbine fault attribute description is key information to assist in fault classification. Second, a fault attribute learner is established based on a convolutional neural network (CNN). During training, the fault attribute learner learns the mapping relationship between known faults and fault attributes, and infers the fault attributes of unknown faults during the testing phase. Finally, the Euclidean distance between the predicted attributes and the true attributes is used as the classification criterion to achieve the classification of unknown faults.

[0093] Section S2 discusses the wind turbine fault diagnosis process in three subsections: fault attribute description based on fault principles, fault data preprocessing based on supervised principal component analysis, and fault diagnosis. The wind turbine fault diagnosis process is as follows: Figure 3 As shown, the fault data is first preprocessed, dividing it into training and test sets. Then, fault attribute learning is performed based on CNN, and finally, a fault classifier is used for diagnosis.

[0094] S2.1 Based on the fault principle, analyze and describe the fault attributes.

[0095] The pitch and yaw systems are crucial control systems for ensuring wind turbine power generation efficiency. The pitch system adjusts the blade angle based on wind speed, thus changing the wind-facing area; the yaw system changes the blade position based on wind direction to achieve wind alignment. Because the pitch and yaw systems operate frequently, and some components are exposed to the external environment, they are susceptible to dust, rain, and snow corrosion, making them high-frequency failure points in wind turbines. Furthermore, both systems input wind power information and output control commands to the blades, resulting in highly similar fault characteristics. Common causes of internal safety chain failure alarms in the pitch control cabinet include abnormal pitch angle detection or excessive angle difference between the three blades; in cases of malfunction of the pitch generator position sensor, the sensor emits abnormal voltage signals, increasing the difference between the pitch angle and the theoretical value. Wind turbines typically operate in harsh natural environments, and external climatic conditions such as wind speed are strongly correlated with damage to the turbine's mechanical components. Wear and aging of mechanical components used to detect the pitch angle, damage to the pitch motor encoder, driver, or the pitch motor itself can all lead to excessively large pitch angle differences. When wind speed fluctuates rapidly, the motor needs to operate frequently to maintain stable speed, resulting in periodic changes in the pitch angle and increased motor temperature, triggering a high-temperature alarm. Other causes include motor stall, cooling fan failure, and pitch bearing damage. Excessive yaw acceleration is generally caused by mechanical vibration and yaw brake lock-up; abnormal yaw vane readings directly affect wind direction measurement. Based on the above knowledge of wind turbine faults, this invention proposes a fault attribute matrix as follows: Figure 4 As shown.

[0096] In one embodiment, the fault samples obtained by this invention come from three wind farms, A, B, and C, and contain a total of 6 faults. Four of these are pitch system faults: internal safety chain fault (#1), pitch generator position sensor malfunction (#2), large pitch position comparison deviation (#3), and high pitch motor temperature (#4). The other two are yaw system faults: acceleration exceeding limits (#5) and wind vane malfunction (#6). Specifically, the historical dataset (one year ago) for wind farm A contains faults #1, #2, #3, and #4, while the recent dataset (within one year) contains faults #5 and #6. The historical dataset for wind farm B contains faults #3, #4, #5, and #6, while the recent dataset contains faults #1 and #2. The historical dataset for wind farm C contains faults #1, #2, #5, and #6, while the recent dataset contains faults #3 and #4.

[0097] S2.2. Based on supervised principal component analysis (PCA), fault data preprocessing is performed on the dataset in S2.1. Three datasets, A, B, and C, are used. Historical data in each dataset contains four types of faults, with 800 samples per fault type as the training set (some fault samples were insufficient, so the sample set was augmented). Recent data contains two types of faults, with 400 samples per fault type as the test set. The fault categories in the training and test sets are mutually exclusive, and each fault type has 25 physical parameters. Since high-dimensional wind turbine spatiotemporal data would lead to excessively long training times, supervised principal component analysis (SPCA) is used to reduce the dimensionality of the wind turbine fault data. Before dimensionality reduction, fault samples are normalized to unify the data value range and achieve a uniform data distribution. Furthermore, the labels of fault samples are derived from the fault attribute descriptions of the corresponding categories; that is, each fault sample has a 12-dimensional label.

[0098] SPCA can reduce data dimensionality while retaining features strongly correlated with fault attribute labels. Since the cumulative variance contribution rate initially reaches over 99% and all singular values ​​are greater than 1 when the principal component dimension is 12, the 12-dimensional data can be considered to provide sufficient explanatory information. Supervised processing is reflected in SPCA's measurement of the correlation between wind turbine fault features and each fault attribute based on the fault attribute label. Calculating the correlation requires first mapping the fault features to a higher-dimensional space based on the radial basis function (RBF), and then calculating the distance between vectors; the closer the distance, the stronger the correlation. The fault sample after SPCA processing has 144 features, with each set of 12 features corresponding to one fault attribute.

[0099] S2.3. Perform fault diagnosis. S2.3.1. Use CNN as the fault diagnosis learner. The fault diagnosis learner established in this invention has 12 layers, and its structure is as follows: Figure 5 As shown, the basic structure of a CNN-based fault attribute learner includes convolutional layers (Conv), pooling layers (Pooling), flattening layers (Flatten), and fully connected layers (FC). Figure 5 In this context, ReLU represents the activation function, filter_num represents the number of convolution kernels, and stride represents the stride.

[0100] In one embodiment, 1) the first layer is a convolutional layer. The function of the convolutional layer is to extract features from the two-dimensional data by performing convolution operations on the input data using convolution kernels. Eight 3×3 convolution kernels perform convolution operations on the input data matrix from left to right, obtaining eight types of features for the wind turbine fault data. Multiple convolution kernels achieve multi-dimensional extraction of fault data. Multi-dimensionality means that the extracted features are at the same level, but reflect different fault information. Using ReLU as the activation function shapes the non-linear relationship between network layers, which can weaken the dependence of parameters and prevent the network from overfitting during training.

[0101] 2) The second layer is the pooling layer, which typically follows the convolutional layer. It reduces network parameters, retaining only important features useful for the final attribute classification result. A 2×2 pooling kernel maximizes the input data matrix from left to right, effectively selecting important features highly correlated with wind turbine faults and reducing training computation. All pooled important features are concatenated before being input into the next convolutional layer. Since convolution reduces feature dimensions, to ensure consistent data shape before and after convolution, this invention pads the edges of the convolutional feature matrix with zeros. Multiple convolutions achieve depth-based feature extraction; depth refers to repeated convolution within the same feature to extract deeper levels of fault information.

[0102] 3) The extracted features are finally input into the flattening layer to transform them into one-dimensional data, and then enter the fully connected layer. The fully connected layer is connected to all nodes of the flattening layer, integrates the features extracted from the previous layer, and feeds them into the classifier, finally outputting the fault diagnosis result at the output layer. The fault attribute learning process is essentially training a CNN classifier for each fault attribute. The input is the training fault sample and its corresponding fault attribute, and the output is a 0 / 1 value (representing whether the sample has that fault attribute). Since there are 12 fault attributes, the training process needs to be repeated 12 times.

[0103] S2.3.2, Construction of a fault classifier based on Euclidean distance.

[0104] The fault attribute learner establishes a mapping relationship between fault samples and fault attributes, while the mapping between fault attributes and fault categories is established by the fault classifier. The Euclidean distance between the fault attribute output by the fault attribute learner and the true fault attribute vector is calculated according to the following formula. The fault classifier considers the sample to belong to the same fault category as the nearest true fault attribute vector.

[0105]

[0106] In the formula, A i Let A' be the actual fault attribute vector. iThis is the fault attribute vector output by the fault attribute learner. `i` represents the th sampling point, `i = 0, 1, ..., 999`; `j` represents the j-th attribute in the fault attribute vector, `j = 0, 1, ..., 11`.

[0107] S3, Incremental diagnostic methods.

[0108] S3.1 Construction of the adaptive sample library, assuming the training set for incremental phase b is... The sample library that needs to be constructed after training can be expressed as:

[0109]

[0110] In the formula, Z represents the number of samples stored in the sample library. It should be noted that Z is a fixed value and will not increase with the increment stage, avoiding gradual redundancy in the sample library during the increment process. The selection of samples must be effectively filtered before being added to subsequent increment stages. This requires that the statistical characteristics of the sample library be similar to those of the training set to ensure the representativeness of the sample library. Therefore, an adaptive sample library module based on the herding effect is constructed. This module adaptively extracts representative samples for each category, then updates and stores the sample library. It includes four S's: class center calculation, class sample extraction, sample self-updating, and sample library storage, which can be summarized as follows.

[0111] 1) Class center calculation: Model f in the incremental stage b b After training, the mean center of each class is calculated, which can be expressed as follows:

[0112]

[0113] In the formula, N represents the mean center of category c in incremental stage b; c C represents the number of training samples for category c; b This is the number of new categories added in incremental phase b (if it is the initial phase, then C). 0 (This refers to the total number of categories in the initial phase).

[0114] 2) Sample Refinement: The selection criteria for samples are closely related to the quality of the sample library. The number of samples Z for each category in incremental stage b is allocated based on the Z-value of the sample library. b .

[0115]

[0116] Then, a metric function is used to calculate the center of gravity of each sample in each category at the current stage relative to the class mean. distance

[0117]

[0118] The smaller the value, the more representative the sample is. Finally, the system automatically selects samples from each class whose distance from the class mean center is the most significant. Z b The smallest The value serves as a sample that needs to be saved at the current stage.

[0119] 3) Sample Self-Update: Adds new samples extracted in the current stage to the sample library E. b And the old sample library E stored in the previous incremental phase b-1 The number of samples to be eliminated in the old category is Z. b -Z b-1 That is, removing Z b -Z b-1 The mean center of the distance class The furthest example. Self-updating of examples is key to ensuring class balance in the example library and avoiding redundancy.

[0120] 4) Sample Library Storage: The reorganized sample library E b Saved to memory and compared with the new training set D in the next incremental phase. b+1 Mix to update the model.

[0121] S3.2, Dynamic Weight Correction Algorithm.

[0122] The loss function has a direct impact on model performance, making the design of a suitable loss function crucial. In the incremental stage b, the classic cross-entropy loss function... This can be expressed as follows:

[0123]

[0124] In the formula, p ic This indicates that the training sample belongs to category x. ic The predicted probability. Cross-entropy loss function. Assigning fixed weights to each fault type limits the model's robustness in learning features across categories. In the multi-incremental diagnosis of wind turbines, the obtained fault data is often limited, and the quality of data collected at different stages varies, resulting in different levels of difficulty for the model to extract features from each category. If the proportion of new knowledge in the model is too high, it will inhibit its ability to retain old knowledge; conversely, it may easily fall into a predicament of low plasticity. Therefore, a dynamic weight adjustment algorithm was designed to help the model adjust the importance of nodes in each category in real time, so as to flexibly learn diagnostic knowledge from limited fault data.

[0125] First, dynamic weight coefficients are introduced. To measure the model's fit to the features of each data category after t iterations in the incremental phase b, it can be expressed as follows:

[0126]

[0127] In the formula, Indicates training sample x ic The predicted probability of belonging to class c after t iterations; This represents the number of training samples for category c in incremental phase b. Let be the loss value of category c after t iterations in the incremental phase b. If The smaller the value, the faster the model learns the data features of category c in t iterations, and the smaller the weight should be assigned to the data features of that category. Conversely, if A larger value indicates a worse fit of the model to the data features of that category, requiring a larger weight to improve the model's learning speed for those features. Therefore, the weight correction loss function in the (t+1)th iteration can be expressed as follows:

[0128]

[0129] When t=1, the dynamic weight coefficients of each category Initialize to 1. The loss value l in the above equation. b(t+1) During the iteration process, the model's adaptability to new failure modes can be dynamically balanced with the stability of existing failure knowledge. Finally, the loss value is calculated along the model's learned parameters θ. b(t+1) Backpropagation is performed, and l is minimized using an adaptive moment estimator optimizer. b(t+1) The learning parameters are used to train the model.

[0130] S3.3 Incremental Diagnostic Framework and Process.

[0131] Incremental diagnostic framework such as Figure 6 As shown, it mainly consists of a feature extractor. Linear classifier Sample Library E b It consists of a weight-adjusted loss function and a feature extractor. Linear classifiers are used to mine high-dimensional feature representations of inputs. This involves projecting high-dimensional features into a category space, and the proposed method employs a popular residual neural network. Example library E b A small storage unit, independent of the diagnostic model, is used to store representative old class samples for incremental model updates. A weight-adjusted loss function helps the model adjust the importance of each class of data during iteration, allowing for flexible learning of diagnostic knowledge from limited fault data.

[0132] In one embodiment, the multi-incremental diagnostic process for the wind turbine is as follows:

[0133] 1) Initial Stage: Initial data including health status and 5 common wind turbine fault types were collected, and preprocessing operations such as sample partitioning, normalization, and fast Fourier transform were performed to obtain the training set. training set D 0 Input model f 0 Used for iterative training. After training, a sample library is constructed from the initial dataset, and the initial test set T is used. 0 Validation model f 0 Diagnostic performance.

[0134] 2) Incremental Phase 1: Collect an incremental training set D containing two new spindle faults. 1 And combined with the sample library E 0 Used for fine-tuning and updating model f 1 It is important to note that f 1 Inherited f 0 The trained learning parameters are used to preserve the initial diagnostic bias; this is also known as network propagation. Furthermore, linear classifiers... The number of output nodes also needs to be updated to match newly emerging fault categories, and the learning parameters of the new nodes are randomly initialized. During the update process, a dynamic weight correction algorithm is used to adjust the importance of each category of nodes in real time. After training, the overall D... 1 and E 0 To adaptively update the sample library E 1 and using the incremental test set T 1 Validate the new model f 1 The incremental diagnostic performance. The incremental test set can be expressed as follows:

[0135]

[0136] In the formula, and These represent the test samples and corresponding labels in incremental phase 1, respectively; M 1 It is the test set T 1 The number of samples in the sample. It should be noted that... This indicates that the distribution space of the test set labels includes all categories from the initial stage and incremental stage 1, in order to verify the stability and plasticity of the model.

[0137] ...

[0138] 3) Incremental Phase b: Collect an incremental training set D containing faults in fresh air unit components. b And combined with the sample library E b-1 Used for fine-tuning and updating model f b Subsequent operations are similar to S in the incremental phase b-1.

[0139] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "preferred embodiment," "detailed description," or "preferred embodiment," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0140] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Therefore, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A monitoring method for a wind power supply system in a mine, characterized in that: The method comprises the following steps: S1, building a wind turbine fault diagnosis framework considering unknown faults; S2, embedding a wind turbine fault attribute space in a fault sample space and a fault category space, and establishing a mapping classification of fault samples→fault attributes→fault categories; S3, describing fault attributes to form a fault attribute feature set; S4, using a supervised principal component analysis method to preprocess the fault attribute feature set and extract key fault features; S5, based on the preprocessed fault features, performing fault classification; S6, using an incremental diagnosis method to dynamically update the diagnosis capability.

2. The method of claim 1, wherein: The S1 specifically comprises: S1.1, Construct the equivalent circuit model of the direct-drive wind turbine grid-side converter, and establish the active power P g and the reactive power Q g The equation is: P g = P0+ P1cos(2ω g t)+ P2sin(2ω g t) Q g = Q0+ Q1cos(2ω g t)+ Q2sin(2ω g t) Wherein: wherein are positive and negative sequence components of the grid voltage, are positive and negative sequence components of the grid-side three-phase current, ω g is the grid angular frequency; S1.2, the power fluctuation parameter is established by monitoring the amplitude deviation of the harmonic component of P g and Q g in real time. the mapping relationship between the power fluctuation parameter and the fault characteristics.

3. The method of claim 2, wherein: The S2 specifically comprises: S2.1, based on the power fluctuation parameters of S1.2, generating fault attribute descriptions and constructing a fault attribute matrix, wherein each row of the matrix corresponds to a fault sample, and each column corresponds to an attribute feature, including harmonic amplitude deviation and timing phase shift; S2.2, using a first convolutional neural network to construct a fault attribute learner, the input being the power fluctuation parameters of S1.2, and the output being a fault attribute vector; S2.3, calculating the Euclidean distance between the attribute vector of the unknown fault sample and the true attribute vector, if the distance is less than a dynamic threshold, then it is classified into the corresponding known fault category, otherwise the incremental learning is triggered.

4. The method of claim 1, wherein: The fault attribute feature set in S3 includes variable pitch system fault attributes, yaw system fault attributes, and general fault attributes; Wherein, the variable pitch system fault attributes include angle abnormalities detected by the variable pitch cabinet or three-pitch angle difference exceeding the limit triggering safety chain alarm; periodic angle changes and temperature rise characteristics caused by variable pitch motor encoder damage, driver failure or bearing damage; temperature alarm signals caused by frequent motor actions under wind speed fluctuations; The yaw system fault attributes include yaw acceleration exceeding the limit caused by mechanical vibration or brake seizure, and wind vane signal abnormalities directly related to wind direction measurement deviation; the general fault attributes include amplitude and timing phase shift of power harmonic components.

5. The method of claim 4, wherein: The general fault attribute generation method is: Based on the power equation of the direct-drive wind turbine grid-side converter equivalent circuit, extract P1, P2, Q1, Q2 harmonic components; Through normalization and fast Fourier transform, convert the time domain power fluctuation into frequency domain features as fault attribute input.

6. The method of claim 4, wherein: The data preprocessing of the fault attribute feature set in S4 specifically comprises: S4.1, based on the variable pitch system, yaw system and general fault attributes defined in S3, generate 12-dimensional attribute labels for each fault sample; S4.2, use A / B / C three data sets, each data set includes training set and test set, and normalize the original data; Wherein, the training set is: 4 kinds of known faults × 800 samples × 25-dimensional original parameters; The test set is: 2 kinds of new faults × 400 samples × 25-dimensional original parameters; S4.3, use a supervised principal component analysis method to reduce the 25-dimensional original data to 12-dimensional feature vectors, wherein: the 12-dimensional features after dimension reduction correspond to the 12 fault attributes in S3; during the dimension reduction process, calculate the feature correlation according to the fault attribute label, and retain the 12 principal components with the strongest correlation with the label; S4.4, 144 intermediate features are generated for each sample after SPCA processing, and 12 features correspond to one fault attribute.

7. The method of claim 6, wherein: The 25-dimensional original data is reduced to a 12-dimensional feature vector, wherein: the first to third dimensions correspond to the variable pitch system attribute; the fourth to fifth dimensions correspond to the yaw system attribute; and the sixth to twelfth dimensions correspond to the general electrical attribute.

8. The method of claim 6, wherein: The SPCA process of S4.3 includes: a) mapping the original fault features to a high-dimensional space based on a radial basis function; b) calculating the vector distance between the high-dimensional features and the 12-dimensional attribute labels in S4.1, and the closer the distance, the stronger the correlation; c) selecting 12 principal components whose cumulative variance contribution rate first reaches more than 99% as the final features.

9. The method of claim 6, wherein: The preprocessed 144-dimensional features are input into the second convolutional neural network, and the Euclidean distance between the output attribute vector and the real vector is calculated to realize fault classification; specifically including: S5.1, the second convolutional neural network is used as a fault diagnosis learner, and the input is the 144-dimensional intermediate features generated in S4.4; S5.2, the mapping relationship between the fault samples and the fault attributes is learned through CNN, and a 12-dimensional fault attribute prediction vector is output; S5.3, the Euclidean distance between the predicted attribute vector and the real attribute vector is calculated: where A i is the predicted attribute vector, A i is the fault attribute vector output by the fault diagnosis learner, i represents the th sampling point, i = 0, 1,... 999; j represents the jth attribute in the fault attribute vector, j = 0, 1,... 11; S5.4, the fault classifier classifies the samples into the fault category with the smallest Euclidean distance.

10. The method of claim 1, wherein: The incremental diagnosis method of S6 specifically includes: S6.1, adaptive example library construction: At each incremental stage b, Z representative samples are selected from the training set D b to construct an example library E b ; By calculating sample features and class centers The Euclidean distance is used to select the Z-axis with the smallest distance in each class. b Each sample is stored in the sample library; old samples in the example library are dynamically eliminated, and the distance from the class center is maintained constant; S6.2, dynamic weight correction: the training weight is dynamically adjusted according to the prediction error of the model for each fault category, and the category with greater error obtains higher training weight; a weighted cross-entropy loss function is used to update the model parameters; S6.3, incremental stage execution: in the initial stage, the model is trained using 5 categories of basic fault data and an initial example library is constructed; in the subsequent incremental stage, new fault data is added in turn, and the model is fine-tuned in combination with the historical samples of the example library; the test set of each incremental stage includes all learned fault types.

Images