A method for fault diagnosis of pelletizing fan

Abstract

A pellet fan fault diagnosis method of the present invention is intended to solve the problems of low efficiency and insufficient accuracy caused by the reliance on manual inspections and empirical judgments in traditional fan fault diagnosis. The method first collects the fan operating parameters to form a data set. After data cleaning and normalization, random sampling and Gaussian noise addition are used for data enhancement. The ConvNeXt network is used for feature extraction, the data set is divided and the fault category feature prototype vector is calculated, and the results are predicted through distance measurement and classification decision making. An L2 regularization term is added to the prototype network model loss function, the model parameters are adjusted through back propagation, and the model is deployed after the performance is stable after evaluation on the test set. The preprocessing and enhancement steps are repeated on the real-time data. If the prediction result is inconsistent or there is new data, the model is retrained. This method improves the diagnostic accuracy and efficiency through the collaboration of multiple technologies, effectively prevents model overfitting, improves the ability to capture fault features, and demonstrates the advantages of complete functions and technological foresight.

Description

Technical Field

[0001] The present invention relates to the technical field of blower fault diagnosis, and more particularly to a pelletizing blower fault diagnosis method. Background Art

[0002] Pellet fans play a vital role in steel production, and their performance and reliability are directly related to the stable operation of the entire steel plant. In the pelletizing line, pellet fans are responsible for providing the necessary oxygen and fuel for pellet roasting. Their efficient and stable operation plays a decisive role in the quality and yield of pellets. However, due to the long-term operation of pellet fans in harsh working environments (such as high temperature, high humidity, and high dust), key components such as bearings and blades within them may experience performance degradation and even failure due to wear and aging. If these failures are not discovered and addressed promptly, they will not only lead to reduced production efficiency but may also cause safety accidents, seriously impacting the company's economic benefits and production safety.

[0003] Faced with this challenge, traditional pellet blower fault diagnosis methods rely primarily on manual inspections and empirical judgment. This approach suffers from inefficiency and insufficient accuracy when dealing with small fault samples. Therefore, the use of a small-sample training method is particularly important. This method effectively trains a fault diagnosis model to address small fault sample problems, even with a small number of fault parameters. To further improve the accuracy and efficiency of fault diagnosis, data augmentation techniques can be used to increase the diversity of training data. This method simulates different fault scenarios to generate more training samples, thereby improving the model's ability to identify various potential faults. Furthermore, L2 regularization effectively prevents model overfitting, which can occur during small-sample training, ensuring model stability and reliability.

[0004] These technological innovations and process improvements not only enhance the stability and reliability of pellet blowers, but also drive technological progress and industrial upgrading across the entire pellet material preparation industry. By employing methods such as data augmentation, L2 regularization, and prototype network models for fault diagnosis, pellet blower maintenance and fault diagnosis have become more efficient and accurate, providing strong technical support for intelligent manufacturing and automation in the steel industry and laying a solid technical foundation for its sustainable development. Summary of the Invention

[0005] The present invention provides a pelletizing fan fault diagnosis method to solve the problems of low efficiency and insufficient accuracy caused by reliance on manual inspection and experience-based judgment in the existing traditional pelletizing fan fault diagnosis.

[0006] According to one aspect of the present invention, a pellet blower fault diagnosis method is provided, comprising the following steps:

[0007] Step S1, collecting historical operating parameters of the pelletizing blower to form a data set, removing outliers and noise interference in the data set through data cleaning, and then normalizing the data;

[0008] Step S2, data enhancement is performed on the dataset using a combination of random sampling and Gaussian noise addition;

[0009] Step S3: After feature extraction through the convolutional neural network, the dataset is divided into a training set, a validation set, and a test set. The training set is divided into a support set and a query set. The feature prototype vector of each fault category in the support set is calculated. The distance between the query sample and the prototype is then calculated using a distance metric, and the prediction result is made through classification decision.

[0010] Step S4: Add an L2 regularization term to the loss function of the prototype network model, determine the L2 regularization coefficient to balance the model's fitting ability and generalization ability, adjust the model parameters through backpropagation, regularly calculate the validation set loss, and determine whether the training termination condition is met;

[0011] Step S5: Input the test set data into the trained model for prediction. Based on the prediction results and the true labels, calculate the accuracy, recall rate, and F1 value and evaluate them. If the performance is stable, deploy the model.

[0012] In step S6, repeat steps S1 and S2 to process and enhance the real-time data. The real-time feature vector is predicted through the prototype network model constructed in S3, and the category with the highest probability is output as the predicted fault type. If it is found that the prediction result of the model is obviously inconsistent with the actual situation and new data is collected, return to step S3 to re-divide the data set and retrain the model.

[0013] Preferably, based on the above solution, in step S1, the specific steps are:

[0014] Step S11, collecting parameters including: vibration signal data provided by the vibration sensor, temperature data collected by the temperature sensor, and current parameters recorded by the current sensor under known abnormal conditions in a normal state; processing the collected data set using a box plot method, and treating data points outside the upper and lower limits of the box plot as outliers;

[0015] Step S12, selecting low-pass filtering, high-pass filtering, and median filtering for noise reduction according to the characteristics of the abnormal value;

[0016] Step S13, replacing the marked outliers by Lagrange interpolation method;

[0017] Step S14: Convert the one-dimensional vibration signal into a two-dimensional time-frequency diagram through short-time Fourier transform. Use Z-score normalization to unify the temperature and current data to the same scale range as additional channels for splicing with the vibration time-frequency diagram. Z-score normalization converts the data into a distribution with a mean of 0 and a standard deviation of 1. The calculation formula is:

[0018]

[0019] Where x is the original temperature and current data value, μ is the mean of the data, σ is the standard deviation, and x normalized It is the temperature and current data value after normalization and then unified scale;

[0020] Among them, in step S14, a time-frequency diagram of size N×M can be obtained through short-time Fourier transform, where N is the resolution of the frequency axis and M is the resolution of the time axis; after normalization, the temperature and current data can be expanded into a matrix of the same size as the vibration time-frequency diagram, each as a channel, and the shape of the input data is N×M×C, where C is the number of channels.

[0021] Preferably, based on the above solution, in step S2, the specific steps are:

[0022] S21, set the minimum and maximum sampling duration limits for the time series dataset, randomly generate a sampling duration, and randomly select a starting point within the total duration range, extract the data segment of the corresponding duration starting from this point, and obtain a new sequence with random length and similar characteristics;

[0023] S22, sampling the feature dimension data to determine the total number of feature dimensions of the data, randomly selecting a portion of the feature dimensions, and randomly selecting feature dimensions from the portion to form a new feature subset;

[0024] S23, determine the mean and standard deviation of the Gaussian noise, independently generate a random noise value that obeys the Gaussian distribution of the mean and standard deviation for each data point in the original data, and add it to the original data point, and obtain new enhanced data after point-by-point operation.

[0025] Preferably, based on the above solution, in step S3, the specific steps are:

[0026] S311, using a ConvNeXt network with a standard convolutional module to perform architecture design and extract features of the time-frequency dataset; using the N×M×C time-frequency graph after data enhancement as the input layer;

[0027] S312: The first convolution layer selects a 3×3 convolution kernel and sets 32 convolution kernels to extract different basic features. It adopts the "same" padding method and a step size of 1. It uses the ReLU activation function. The parameters of the second and third convolution layers are consistent with those of the first convolution layer. The output feature map has a shape of N×M×32.

[0028] S313, add pooling windows after the first and second convolutional layers respectively, use a 2×2 maximum pooling window with a step size of 2; perform a pooling operation on the feature map output by the convolutional layer, reduce the size of the feature map by half, and the shape of the output feature map becomes

[0029] The second convolution layer also uses a 3×3 convolution kernel, sets 64 convolution kernels, the padding method and step size remain unchanged, the activation function is still ReLU, and the input is the output of the previous layer. The output feature map shape is,

[0030] A maximum pooling window of 2×2 is used with a step size of 2; a pooling operation is performed on the feature map output by the convolutional layer, reducing the size of the feature map by half, and the shape of the output feature map becomes

[0031] The third convolution layer still uses a 3×3 convolution kernel, sets 128 convolution kernels, "same" padding and step size to 1, the activation function is ReLU, and the input is the output of the previous layer. The output feature map shape is

[0032] S314, in the two-dimensional convolution, the calculation formula for each element (i, j, k) in the output feature map (where i and j are spatial position indices and k is a channel index) is:

[0033]

[0034] Among them, I i+m，j+n，l Represents the element value at the corresponding position in the input feature map, K m，n，l，k Represents the weight value of the corresponding position in the convolution kernel;

[0035] S315, adds BatchNormalization after each convolution layer to speed up training and improve model stability;

[0036] S316, after a series of convolutional layers and pooling layers, obtain the feature map of shape N×M×C, where N1 and M1 are the spatial dimensions of the feature map and C is the number of channels; flatten the feature map into a one-dimensional vector with a length of N1×M1×C;

[0037] S317, the first fully connected layer has L neurons, and the output of the fully connected layer is obtained by linear transformation through a weight matrix W1 (size is N1×M1×C×L) and a bias vector b1 (length is L); the formula is:

[0038] z=f(W1x+b1);

[0039] Where x is the flattened feature vector and f is the ReLU function;

[0040] S318, the second fully connected layer has K neurons (K is the number of fault classes), and a weight matrix W2 (size is N1×M1×C×K) and a bias vector b2 (length is K) are linearly transformed to obtain the output of the fully connected layer; the formula is:

[0041] y=f(W2z+b2);

[0042] Where z is the output of the first fully connected layer, and f is the Softmax function;

[0043] S32 divides the data set: the feature vectors obtained in S31 are divided into 70% training set, 20% validation set, and 10% test set. A random partitioning method is used to ensure that each category has a certain distribution in different data sets. The training set is divided into 75% support set and 25% query set.

[0044] The S321 training set is used for the model to learn data patterns and regularities, adjust parameters and establish mapping relationships; the validation set is used to evaluate performance during training, help adjust hyperparameters, and determine overfitting or underfitting; the test set is used to evaluate the final performance after training and measure generalization ability; the support set (part of the training set) is used to calculate the category prototype; the query set is used to test the model's ability to classify new samples, and make classification decisions based on the distance from the category prototype obtained based on the support set.

[0045] S33 computing prototype:

[0046] Computation prototype initialization phase: Divide the samples in the support set according to the fault category; mark the K fault categories as C1, C2, ..., C K ; For each fault category C k (k=1,2,…,K), represents the kth fault category, K is the total number of fault categories, and a small number of initial samples are randomly selected to form the initial sample set:

[0047]

[0048] Where n is the number of initial samples; calculate the initial prototype for each fault category That is, the average value of the initial sample feature vector of this category; the calculation formula is:

[0049]

[0050] n: The number of initial samples selected for each fault category during the initialization phase

[0051] x k，i : The feature vector of the i-th sample in the k-th fault category in the initial sample set

[0052] Calculation prototype batch update phase: As new samples are continuously added, the prototype is gradually updated in batch mode; the current number of iterations is t, and the newly added batch of samples is

[0053]

[0054] A new batch of samples added to the k-th fault category at the th iteration

[0055] y k，m : is the feature vector of the mth sample newly added to the sample set of the kth fault category at the tth iteration;

[0056] m: The number of samples newly added to the kth fault category at the tth iteration, that is, the sample set The number of samples in ;

[0057] The existing prototype is Calculate the total number of processed samples;

[0058]

[0059] Its i is the number of new samples added in the i-th iteration; the prototype is updated after iteration The calculation formula is:

[0060]

[0061] S34 distance metric:

[0062] The faults are measured and classified by the nearest neighbor classification (K-NN); the sample feature vector y of the query set is calculated using the Euclidean distance method k，n (y k，n represents the feature vector of the nth sample in the kth fault category in the query set) and the prototype set The distance between two feature vectors y=(y1,y2,…,y n ) and x=(x1,x2,…,x n), (x and y are any two feature vectors in the query set and prototype set respectively, and n is the dimension of the feature vector)

[0063] The formula for calculating Euclidean distance is:

[0064]

[0065] n is the dimension of the feature vector;

[0066] S35 decision classification:

[0067] According to the calculated distance, find the k training samples closest to each other, which are called y k，n The k nearest neighbors of ; count the frequency of occurrence of each category in these k nearest neighbor samples;

[0068] Take the negative exponential of the distance and then convert it into a probability through the softmax function. For sample x, the probability of belonging to the i-th fault category is calculated as follows:

[0069]

[0070] Where N is the total number of fault categories;

[0071] d(x,c i ): represents the relationship between sample x and the i-th fault category prototype c i the distance between them;

[0072] c i : represents the prototype vector of the i-th fault category;

[0073] y k，n The class with the highest probability is classified as this class, which is considered to best represent the characteristics of the new sample.

[0074] Based on the above solution, preferably, in step S4, the specific steps are:

[0075] S41, use functions such as cross entropy loss as the basic loss function.

[0076] The true category of the sample is y, and the predicted category is The cross entropy loss function is

[0077]

[0078] If the sample belongs to the jth fault category, then the true category vector y = (0, 0, ..., 1, ..., 0), where the jth position is 1 and the rest of the positions are 0

[0079] S42, calculate the L2 regularization term based on the model parameters. The formula of the L2 regularization term is:

[0080] λ∑ w ||w|| 2 , where w is the parameter of the model and λ is the regularization coefficient;

[0081] S45, add the L2 regularization term to the basic loss function to obtain a new loss function:

[0082]

[0083] S43 back propagation;

[0084] S431 calculates the gradient of the loss function with respect to the model parameters:

[0085] Calculate the predicted probability of a sample belonging to each fault category, the distance function, and the gradient of the L2 regularization term to the model parameters, so that the optimizer can be used to update the model parameters and improve the performance of the model and the accuracy of fault diagnosis;

[0086] Use the chain rule to gradually calculate the gradient of the loss function with respect to the model parameters. For each parameter, calculate

[0087] First calculate the gradient of the cross entropy loss function with respect to the predicted probability

[0088] Among them, L is the cross entropy loss function, y i is the true category label of the sample, is the predicted probability that the sample belongs to the i-th fault category;

[0089] According to the calculation formula of predicted probability, where d(x,c i ) is the distance between the sample and the prototype of the i-th category, and N is the total number of fault categories;

[0090] calculate where d(x,c i ) is the distance between the sample and the prototype of the i-th category, and N is the total number of fault categories:

[0091] Combined with the distance function d(x,c i ) With respect to the derivative of the parameters, we can obtain the gradient of the predicted probability of each fault category with respect to the model parameters:

[0092]

[0093] The formula for the L2 regularization term is λ∑ m＝1 ∥w∥ 2 , where w is the parameter of the model and λ is the regularization coefficient: the gradient of the L2 regularization term with respect to the parameter is calculated as 2λw:

[0094] Finally, the total gradient is

[0095] S432, use the optimizer to update the model parameters:

[0096] S4321, initialize Adam optimizer parameters:

[0097] Set the learning rate α, usually between 0.001 and 0.01, the exponential decay rate β1 of the first-order moment estimate is 0.9; the exponential decay rate β2 of the second-order moment estimate is 0.999; set a very small constant ∈ to 1e-8 to prevent the denominator from being zero;

[0098] For each model parameter w;

[0099] Initialize the first-order moment estimate m0=0 and the second-order moment estimate v0=0;

[0100] At iteration t:

[0101] Compute the first moment estimate:

[0102] m t First-order moment estimate, representing the exponentially weighted average of the gradient;

[0103] Compute the second moment estimate:

[0104] v t The second-order moment estimate represents the exponentially weighted average of the square of the gradient, which is used to estimate the variance of the gradient and help adjust the step size of the parameter update;

[0105] Compute the corrected first moment estimate:

[0106] Compute the corrected second moment estimates:

[0107] Update parameters:

[0108] By repeating the above steps, the Adam optimizer is used to update the model parameters in each iteration, so that the model is gradually optimized during the training process.

[0109] S4322, regularly calculate the validation set loss

[0110] During model training, the loss value is calculated on the validation set in each training cycle. The loss function is the cross entropy loss plus the L2 regularization term, that is, L total =L+λ∑ w ∥w∥ 2 , where L is the cross entropy loss function, λ is the regularization coefficient, and w is the model parameter:

[0111] For each sample in the validation set, input it into the model for forward propagation, calculate the prediction result, then calculate the loss value based on the true label, and finally average the loss values ​​of all samples to get the average loss of the validation set;

[0112] Stop training when the change in validation set loss is less than 5% over several consecutive training cycles. If the loss is unstable, continue training and adjust the model parameters.

[0113] Preferably, based on the above solution, in step S5, the specific steps are:

[0114] S5, model evaluation and testing further includes: S51 using the test set to evaluate the trained model; S52 analyzing the diagnosis results of the model to check whether there is any misclassification or missed diagnosis; S53 deploying the model.

[0115] Preferably, based on the above solution, in step S5,

[0116] S51, evaluating the trained model using the test set further includes: calculating the harmonic mean of F1 and recall, where the F1 value is the harmonic mean of precision and recall, which comprehensively considers the precision and recall ability of the model;

[0117] Input the samples of the test set into the trained model and calculate the performance indicators of the model on the test set, such as accuracy, recall rate, F1 value and other parameters;

[0118] Accuracy calculation formula:

[0119]

[0120] Recall calculation formula:

[0121]

[0122] The accuracy formula is:

[0123]

[0124] F1 formula:

[0125]

[0126] Based on the evaluation results on the validation set, adjust the model's hyperparameters, such as learning rate, regularization coefficient, and number of neighbors, to find the best hyperparameter combination;

[0127] S52, analyzing the diagnosis results of the model to check whether there are any misclassifications or missed diagnoses further includes: traversing each sample in the test set and comparing the model's predicted category with the actual category; if the predicted category is inconsistent with the actual category, it is a misclassification; recording the features of the misclassified sample and the actual category and predicted category to facilitate cause analysis; for missed diagnoses, checking whether there is a situation where a fault actually exists but is not detected by the model;

[0128] If a high number of misclassifications or missed diagnoses are detected, the following three adjustment methods can be used: Adjust model hyperparameters, such as increasing regularization strength, adjusting the learning rate, changing the number of neighbors, retraining the model, and evaluating it; perform data augmentation, such as adding more noise, rotating, flipping, and other operations to expand the dataset, and retrain the model to improve its generalization ability; Analyze the feature extraction process, which may require improving the feature extraction method or adding new features to better distinguish different fault categories;

[0129] If the performance is stable, the model is deployed and loaded into the local server.

[0130] Based on the above solution, preferably, in step S52, the specific steps are:

[0131] S6 fault diagnosis and analysis:

[0132] Repeat steps S1 and S2 to process and enhance the real-time data. Use the prototype network model built in S3 to predict the real-time feature vector and output the category with the highest probability as the predicted fault type. If it is found that the model's prediction result is obviously inconsistent with the actual situation or new data is collected, return to step S3 to re-divide the data set and retrain the model to improve its accuracy and adaptability.

[0133] The pellet blower fault diagnosis method of the present invention has the following advantages compared with the prior art:

[0134] 1. Diagnostic Accuracy and Efficiency: The fault diagnosis method of the present invention not only improves the accuracy and efficiency of pellet blower fault diagnosis through the synergistic effects of data augmentation, L2 regularization, and a prototype network model, but also optimizes the fault detection process through an intelligent diagnostic process. The fault diagnosis method of the present invention analyzes the operating status of the pellet blower from different angles. Data augmentation enriches the training data, L2 regularization ensures model stability, and the prototype network model accurately captures fault characteristics. Intelligent diagnosis ensures the consistency and reliability of the results.

[0135] 2. Prevent model overfitting: The present invention effectively avoids the model overfitting phenomenon caused by small sample data through L2 regularization, ensures that the fault diagnosis model can still accurately identify faults under limited sample data, improves the generalization ability of the model, ensures the reliability of pellet blower fault diagnosis, and prevents misjudgment and missed judgment problems caused by overfitting.

[0136] 3. Improved Fault Characteristic Capture: This invention utilizes a prototype network model to accurately capture the characteristic patterns of pellet blower faults. Each diagnostic step promptly identifies potential fault hazards, preventing them from further deteriorating and reducing downtime and repair costs. Furthermore, the application of the prototype network model makes fault diagnosis more intelligent, providing strong support for rapid and accurate troubleshooting.

[0137] 4. Functional Improvement and Technological Foresight: While the fault diagnosis method of this invention is still under development and improvement, it has already demonstrated significant potential. Data augmentation provides a path for acquiring more diverse data in the future. L2 regularization and continuous optimization of the prototype network model also leave room for development in pelletizing blower fault diagnosis under diverse operating conditions. This design approach reflects a comprehensive consideration of the foresight and sustainability of fault diagnosis technology, ensuring that the method can adapt to the higher requirements and development trends for blower fault diagnosis in future pelletizing production. BRIEF DESCRIPTION OF THE DRAWINGS

[0138] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. 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 work.

[0139] In the picture:

[0140] Figure 1 is a fault diagnosis flow chart of the present invention; DETAILED DESCRIPTION

[0141] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0142] It will be understood that when used in this specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections.

[0143] To simplify the drawings, only the parts relevant to the present invention are schematically shown in each figure. They do not represent the actual structure of the product. Furthermore, to simplify the drawings and facilitate understanding, in some figures, only one of the components with the same structure or function is schematically depicted or labeled. As used herein, "one" not only means "only one" but also "more than one."

[0144] It should be further understood that the term "and / or" used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0145] In the embodiments shown in the accompanying drawings, directional indications (such as up, down, left, right, front, and rear) used to explain the structure and movement of various components of the present invention are not absolute but relative. These descriptions are applicable when the components are in the positions shown in the accompanying drawings. If the descriptions of the positions of these components are changed, the directional indications will also change accordingly.

[0146] In addition, in the description of the present application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0147] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the specific embodiments of the present invention will be described below with reference to the accompanying drawings. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings and other embodiments can be obtained based on these drawings without inventive work.

[0148] See also Figure 1 The present invention provides a method for diagnosing fan faults, comprising the following steps:

[0149] S1, data collection and preprocessing: collect data sets of various operating parameters of the wind turbine, remove outliers and noise interference through data cleaning, and then normalize the data;

[0150] The data sets of various operating parameters of the wind turbine collected in step S1 include: collecting vibration signal data provided by the vibration sensor under normal and known abnormal conditions, collecting temperature data from the temperature sensor, and recording current data from the current sensor.

[0151] Step S11, the detailed steps of data cleaning are: using the box plot method to process the collected data set, treating the data points outside the upper and lower limits of the box plot as outliers, and determining the center position and dispersion of the data by calculating the lower quartile (Q1), median (Q2) and upper quartile (Q3) of the data set. According to the 1.5 times interquartile range (IQR) rule, points below Q1-1.5×IQR or above Q3+1.5×IQR are marked as outliers.

[0152] Step S12: According to the characteristics of the abnormal point, a low-pass filter is used to remove high-frequency components, a high-pass filter is used to remove low-frequency drift or trend items, and a median filter is used to remove the interference of impulse noise.

[0153] Step S13: replace the marked outlier by Lagrange interpolation; select n normal data points around the outlier as interpolation nodes;

[0154] Construct the Lagrange interpolation polynomial. For the given n interpolation nodes (x0, y0), (x1, y1), ..., (x n ,y n ), the Lagrange interpolation polynomial L(x) is defined as:

[0155]

[0156] Among them, L i (x) is the Lagrangian basis function, defined as:

[0157]

[0158] Calculate the estimated value of the outlier: For the point x corresponding to the outlier, substitute it into the Lagrange interpolation polynomial L(x) for calculation, and calculate each Lagrange basis function L in turn i (x) value, and then with the corresponding y i Multiply them together and add up all the products to get an estimate of the outlier.

[0159] The calculated estimated values ​​of the outliers are used to replace the outliers in the original data, thus making the data more reasonable for subsequent fault diagnosis and analysis.

[0160] Step S14: convert the one-dimensional vibration signal into a two-dimensional time-frequency diagram through short-time Fourier transform. The size of the time-frequency diagram is N×M, where N is the resolution of the frequency axis and M is the resolution of the time axis.

[0161] Z-score normalization is used to unify the temperature and current data to the same scale range as additional channels and splice them with the vibration time-frequency graph. Z-score normalization converts the data into a distribution with a mean of 0 and a standard deviation of 1. The calculation formula is:

[0162]

[0163] Where x is the original temperature and current data value, μ is the mean of the data, σ is the standard deviation, and x normalized It is the temperature and current data value after normalization and then unified scale.

[0164] Among them, in step S14, a time-frequency diagram of size N×M can be obtained through short-time Fourier transform, where N is the resolution of the frequency axis and M is the resolution of the time axis; after normalization, the temperature and current data can be expanded into a matrix of the same size as the vibration time-frequency diagram, each as a channel, and the shape of the input data is N×M×C, where C is the number of channels.

[0165] After normalization, the temperature and current data can be expanded into a matrix of the same size as the vibration time-frequency graph, each serving as a channel. The shape of the input data is N × M × C, where C is the number of channels (the vibration time-frequency graph plus the number of temperature and current channels).

[0166] Through the above data collection and preprocessing steps, high-quality and reliable data can be provided for the subsequent fan fault diagnosis model, thereby improving the accuracy and generalization ability of the model.

[0167] S2 data enhancement: A combination of random sampling and Gaussian noise addition is used to enhance the data set, increasing data diversity and improving the generalization ability of the model;

[0168] The specific S2 steps include:

[0169] S21, set the minimum and maximum sampling duration limits for the time series dataset, randomly generate a sampling duration, and randomly select a starting point within the total duration range, extract the data segment of the corresponding duration starting from this point, and obtain a new sequence with random length and similar characteristics;

[0170] S22, sampling the feature dimension data to determine the total number of feature dimensions of the data, randomly selecting a portion of the feature dimensions and randomly selecting feature dimensions therefrom to form a new feature subset;

[0171] The detailed steps of step S21 are as follows: the total duration of the original time series data is T, and a minimum sampling duration t-min and a maximum sampling duration t-max are set. A sampling duration t is randomly generated, satisfying t-min ≤ t ≤ t-max. A starting time point s is randomly selected such that s + t does not exceed the total duration T. Data for a continuous time period of length t starting from time point s is extracted as the sampling result. In this way, a new time series segment is obtained, whose length is random and has similar characteristics to the original time series, but not exactly the same.

[0172] S23, Gaussian noise addition, involves determining the mean and standard deviation of the Gaussian noise, independently generating random noise values ​​that follow a Gaussian distribution with the mean and standard deviation dui for each data point in the original data, and adding these noise values ​​to the original data points. This step, performed point by point, yields new enhanced data. This combination effectively increases data diversity and robustness, improving the model's generalization across diverse data scenarios.

[0173] S3 builds a prototype network model: After performing feature extraction using a convolutional neural network, the dataset is divided into training, validation, and test sets. The training set is divided into a support set and a query set. The feature prototype vector of each fault category in the support set is calculated. The distance between the query sample and the prototype is then calculated using a distance metric. The prediction result is made through classification decision making.

[0174] Using the convolutional neural network feature extraction method, the enhanced dataset is input into the feature extractor to extract high-dimensional feature vectors. For each fault category, the average value of its feature vector is calculated as the prototype of the fault category. For the query sample, the distance between its feature vector and the prototype of each fault category is calculated. The calculated distance is converted into a probability distribution using the softmax function, and the category with the highest probability is selected as the predicted category.

[0175] The detailed steps of the S3 step include:

[0176] S311, using a ConvNeXt network with a standard convolutional module to perform architecture design and extract features of the time-frequency dataset; using the N×M×C time-frequency graph after data enhancement as the input layer;

[0177] S312: The first convolution layer selects a 3×3 convolution kernel and sets 32 convolution kernels to extract different basic features. It adopts the "same" padding method and a step size of 1. It uses the ReLU activation function. The parameters of the second and third convolution layers are consistent with those of the first convolution layer. The output feature map has a shape of N×M×32.

[0178] S313, add pooling windows after the first and second convolutional layers respectively, using a 2×2 maximum pooling window with a step size of 2. Pooling operation is performed on the feature map output by the convolutional layer, reducing the size of the feature map by half, and the shape of the output feature map becomes

[0179] The second convolution layer also uses a 3×3 convolution kernel, sets 64 convolution kernels, the padding method and step size remain unchanged, the activation function is still ReLU, and the input is the output of the previous layer. The output feature map shape is,

[0180] A maximum pooling window of 2×2 is used with a step size of 2. Pooling operation is performed on the feature map output by the convolution layer, reducing the size of the feature map by half, and the shape of the output feature map becomes

[0181] The third convolution layer still uses a 3×3 convolution kernel, sets 128 convolution kernels, "same" padding and step size to 1, the activation function is ReLU, and the input is the output of the previous layer. The output feature map shape is

[0182] S314, in the two-dimensional convolution, the calculation formula for each element (i, j, k) in the output feature map (where i and j are spatial position indices and k is a channel index) is:

[0183]

[0184] Among them, I i+m，j+n，l Represents the element value at the corresponding position in the input feature map, K m，n，l，k Represents the weight value of the corresponding position in the convolution kernel.

[0185] S315, adds BatchNormalization after each convolution layer to speed up training and improve model stability;

[0186] S316, after a series of convolutional and pooling layers, a feature map of shape N×M×C is obtained, where N1 and M1 are the spatial dimensions of the feature map, and C is the number of channels. This feature map is flattened into a one-dimensional vector of length N1×M1×C;

[0187] S317, the first fully connected layer has L neurons, and the output of the fully connected layer is obtained by linear transformation through a weight matrix W1 (size is N1×M1×C×L) and a bias vector b1 (length is L); the formula is:

[0188] z=f(W1x+b1);

[0189] Where x is the flattened feature vector and f is the ReLU function;

[0190] S318, the second fully connected layer has K neurons (K is the number of fault classes), and a weight matrix W2 (size is N1×M1×C×K) and a bias vector b2 (length is K) are linearly transformed to obtain the output of the fully connected layer; the formula is:

[0191] y=f(W2z+b2);

[0192] Where z is the output of the first fully connected layer and f is the Softmax function.

[0193] S32 divides the data set: the feature vectors obtained in S31 are divided into 70% training set, 20% validation set, and 10% test set ratios; a random partitioning method is used to ensure that each category has a certain proportional distribution in different data sets; the training set is divided into 75% support set and 25% query set.

[0194] The S321 training set is used for the model to learn data patterns and regularities, adjust parameters and establish mapping relationships; the validation set is used to evaluate performance during training, help adjust hyperparameters, and determine overfitting or underfitting; the test set is used to evaluate the final performance after training and measure generalization ability; the support set (part of the training set) is used to calculate the category prototype; the query set is used to test the model's ability to classify new samples, and make classification decisions based on the distance from the category prototype obtained based on the support set.

[0195] S33 computing prototype:

[0196] Computation prototype initialization phase: Divide the samples in the support set according to the fault category. Label the K fault categories as C1, C2, ..., C K For each fault category C k , randomly select a small number of initial samples to form an initial sample set:

[0197]

[0198] Where n is the number of initial samples. Calculate the initial prototype of each fault category That is, the average value of the initial sample feature vector of this category; the calculation formula is:

[0199]

[0200] Calculation prototype batch update phase: As new samples are continuously added, the prototype is gradually updated using batch processing. The current number of iterations is t, and the newly added batch of samples is

[0201]

[0202] The existing prototype is Calculate the total number of processed samples

[0203]

[0204] Its i The number of samples newly added in the i-th iteration. Update the prototype after iteration The calculation formula is:

[0205]

[0206] S34 distance metric

[0207] The faults are measured and classified by the nearest neighbor classification (K-NN); the sample feature vector y of the query set is calculated using the Euclidean distance method k，n With prototype set The distance between two feature vectors x = (x1, x2, ..., x n ) and y=(y1,y2,…,y n ), the Euclidean distance calculation formula is:

[0208]

[0209] n is the dimension of the feature vector.

[0210] S35 Decision Classification

[0211] According to the calculated distance, find the k training samples closest to each other, which are called y k，n Count the frequency of each category in these k nearest neighbor samples.

[0212] Take the negative exponential of the distance and then convert it into a probability through the softmax function. For sample x, the probability of belonging to the i-th fault category is calculated as

[0213]

[0214] Where N is the total number of fault categories

[0215] y k，n The class with the highest probability is classified as this class, which is considered to best represent the characteristics of the new sample.

[0216] S4 model training:

[0217] Add an L2 regularization term to the loss function of the prototype network model and determine the L2 regularization coefficient to balance the model's fitting ability and generalization ability;

[0218] Step S4 further includes: S41 calculating a basic loss function; S42 adding an L2 regularization term;

[0219] S41 calculates the basic loss function further including:

[0220] Use cross entropy loss and other functions as the basic loss function. The true category of the sample is y, and the predicted category is The cross entropy loss function is

[0221]

[0222] If the sample belongs to the jth fault category, then the true category vector y = 0, 0, ..., 1, ..., 0), where the jth position is 1 and the rest of the positions are 0

[0223] S42 adding the L2 regularization term further includes: adding to the new loss function through the L2 regularization term formula:

[0224] The L2 regularization term is calculated based on the model parameters. The formula for the L2 regularization term is λ∑ w ∥w∥ 2 , where w is the parameter of the model and λ is the regularization coefficient; the L2 regularization term is added to the basic loss function to obtain a new loss function:

[0225]

[0226] S43 Backpropagation

[0227] S431 calculates the gradient of the loss function with respect to the model parameters

[0228] Calculate the predicted probability of the sample belonging to each fault category, the distance function, and the gradient of the L2 regularization term to the model parameters, so as to use the optimizer to update the model parameters and improve the performance of the model and the accuracy of fault diagnosis. Use the chain rule to gradually calculate the gradient of the loss function to the model parameters. For each parameter, calculate

[0229] First calculate the gradient of the cross entropy loss function with respect to the predicted probability

[0230] Where L is the cross entropy loss function, y i is the true category label of the sample, is the predicted probability that the sample belongs to the i-th fault category.

[0231] Then according to the calculation formula of predicted probability, where d(x,c i ) is the distance between the sample and the prototype of the i-th category, and N is the total number of fault categories.

[0232] Further calculation where d(x,c i ) is the distance between the sample and the prototype of the i-th category, and N is the total number of fault categories.

[0233] Combined with the distance function d(x,c i ) With respect to the derivative of the parameters, we can get the gradient of the predicted probability of each fault category with respect to the model parameters.

[0234]

[0235] The formula for the L2 regularization term is λ∑m ＝1 ∥w∥ 2 , where w is the model parameter and λ is the regularization coefficient. The gradient of the L2 regularization term with respect to the parameter is calculated as 2λw.

[0236] Finally, the total gradient is

[0237] S432 uses the optimizer to update model parameters

[0238] S4321 Initialize Adam optimizer parameters

[0239] Set the learning rate α, which is usually between 0.001 and 0.01. The exponential decay rate β1 of the first-order moment estimate is 0.9; the exponential decay rate β2 of the second-order moment estimate is 0.999; set a very small constant ∈ to 1e-8 to prevent the denominator from being zero.

[0240] For each model parameter w:

[0241] Initialize the first-order moment estimate m0=0 and the second-order moment estimate v0=0.

[0242] At iteration t:

[0243] Compute the first moment estimate:

[0244] m t The first-order moment estimate represents the exponentially weighted average of the gradient. It reflects the overall trend of past gradients and helps stabilize the direction of parameter updates during training.

[0245] Compute the second moment estimate:

[0246] v t The second-order moment estimate represents the exponentially weighted average of the square of the gradient. It is used to estimate the variance of the gradient and help adjust the step size of the parameter update.

[0247] Compute the corrected first moment estimate:

[0248] Compute the corrected second moment estimates:

[0249] Update parameters:

[0250] By repeating the above steps, the Adam optimizer is used to update the model parameters in each iteration, so that the model is gradually optimized during the training process.

[0251] S4322 Regularly calculate validation set loss

[0252] During model training, the loss value is calculated on the validation set in each training cycle. The loss function is the cross entropy loss plus the L2 regularization term, that is, L total =L+λ∑ w ||w|| 2 , where L is the cross entropy loss function, λ is the regularization coefficient, and w is the model parameter.

[0253] For each sample in the validation set, it is fed into the model for forward propagation, the prediction result is calculated, and then the loss value is calculated based on the true label. Finally, the loss values ​​of all samples are averaged to obtain the average loss of the validation set.

[0254] Stop training when the change in validation set loss is less than 5% over several consecutive training cycles. If the loss is unstable, continue training and adjust the model parameters.

[0255] S5 model testing and evaluation

[0256] Feed the trained model with the test set data for prediction. Based on the prediction results and the true labels, calculate precision, recall, F1 score, and other metrics for evaluation. These metrics measure the model's performance on the test set. If performance is stable, deploy the model.

[0257] S5 model evaluation and testing further includes: S51 using the test set to evaluate the trained model, S52 analyzing the model's diagnostic results to check whether there are any misclassifications or missed diagnoses; and S53 model deployment.

[0258] S51 uses the test set to evaluate the trained model, further including: calculating the harmonic mean of F1 and recall. The F1 value is the harmonic mean of precision and recall, which comprehensively considers the precision and recall ability of the model;

[0259] Input the samples of the test set into the trained model and calculate the performance indicators of the model on the test set, such as accuracy, recall rate, F1 value and other parameters.

[0260] Accuracy calculation formula:

[0261]

[0262] Recall calculation formula:

[0263]

[0264] The accuracy formula is:

[0265]

[0266] F1 formula:

[0267]

[0268] Based on the evaluation results on the validation set, adjust the model's hyperparameters, such as learning rate, regularization coefficient, and number of neighbors, to find the best hyperparameter combination;

[0269] S52 analyzes the diagnosis results of the model to check whether there are any misclassifications or missed diagnoses, further comprising: traversing each sample in the test set and comparing the predicted category of the model with the actual category; if the predicted category is inconsistent with the actual category, it is a misclassification; recording the features of the misclassified sample and the actual category and predicted category to facilitate cause analysis; for missed diagnoses, checking whether there is a situation where there is actually a fault but the model has not detected the fault;

[0270] If a large number of misclassifications or missed diagnoses are found, the following three adjustment methods are used: adjust the model hyperparameters, such as increasing the regularization strength, adjusting the learning rate, changing the number of neighbors, etc., retrain the model and evaluate it; perform data enhancement, such as adding more noise, rotating, flipping and other operations to expand the data set, and train the model again to improve the model's generalization ability; analyze the feature extraction process, which may require improving the feature extraction method or adding new features to better distinguish different fault categories.

[0271] If the performance is stable, the model is deployed and loaded into the local server.

[0272] S6 fault diagnosis and analysis

[0273] Repeat steps S1 and S2 to process and enhance the real-time data. Use the prototype network model built in S3 to predict the real-time feature vector and output the category with the highest probability as the predicted fault type. If the model's predictions are significantly inconsistent with the actual situation or new data is collected, return to step S3 to repartition the dataset and retrain the model to improve its accuracy and adaptability.

[0274] Finally, the method of this application is only a preferred embodiment and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A pellet blower fault diagnosis method, characterized in that: The following steps are involved: Step S1, collecting historical operating parameters of the pelletizing blower to form a data set, removing outliers and noise interference in the data set through data cleaning, and then normalizing the data; Step S2, data enhancement is performed on the dataset using a combination of random sampling and Gaussian noise addition; Step S3: After feature extraction using a convolutional neural network, the dataset is divided into a training set, a validation set, and a test set. The training set is divided into a support set and a query set. The feature prototype vector of each fault category in the support set is calculated. The distance between the query sample and the prototype is then calculated using a distance metric, and the prediction result is made through classification decision. Step S4: Add an L2 regularization term to the loss function of the prototype network model, determine the L2 regularization coefficient to balance the fitting ability and generalization ability of the prototype network model, adjust the model parameters through backpropagation, regularly calculate the validation set loss, and determine whether the training termination condition is met; In step S5, the test set data is input into the trained balanced prototype network model for prediction. Based on the prediction results and the true labels, the precision, recall, and F1 value are calculated and evaluated. If the performance is stable, the model is deployed. The F1 value is the harmonic mean of the precision and recall rates. Step S6: Repeat steps S1 and S2 to process and enhance the real-time data, predict the real-time feature vector using the prototype network model constructed in S3, and output the category with the highest probability as the predicted fault type; If the model's prediction results are found to be significantly inconsistent with the actual situation or new data is collected, return to step S3 to re-partition the dataset and retrain the model; In step S1, the specific steps are: Step S11, collecting parameters including: vibration signal data provided by the vibration sensor, temperature data collected by the temperature sensor, and current parameters recorded by the current sensor under known abnormal conditions in a normal state; processing the collected data set using a box plot method, and treating data points outside the upper and lower limits of the box plot as outliers; Step S12, selecting low-pass filtering, high-pass filtering, and median filtering for noise reduction according to the characteristics of the abnormal value; Step S13, replacing the marked outliers by Lagrange interpolation method; Step S14: Convert the one-dimensional vibration signal into a two-dimensional time-frequency diagram through short-time Fourier transform. Use Z-score normalization to unify the temperature and current data to the same scale range as additional channels for splicing with the vibration time-frequency diagram. Z-score normalization converts the data into a distribution with a mean of 0 and a standard deviation of 1. The calculation formula is: ； Where x is the original temperature and current data value, is the mean of the data, is the standard deviation, are the temperature and current data values ​​after normalization on a unified scale; Among them, in step S14, a value of The time-frequency diagram is as follows: N is the resolution of the frequency axis and M is the resolution of the time axis. After normalization, the temperature and current data can be expanded into a matrix of the same size as the vibration time-frequency diagram, each as a channel. The shape of the input data is , C is the number of channels; In step S2, the specific steps are: S21, set the minimum and maximum sampling duration limits for the time series dataset, randomly generate a sampling duration, and randomly select a starting point within the total duration range, extract the data segment of the corresponding duration starting from this point, and obtain a new sequence with random length and similar characteristics; S22, sampling the feature dimension data to determine the total number of feature dimensions of the data, randomly selecting a portion of the feature dimensions, and randomly selecting feature dimensions from the portion to form a new feature subset; S23, determine the mean and standard deviation of the Gaussian noise, independently generate a random noise value that obeys the Gaussian distribution of the mean and standard deviation for each data point in the original data, and add it to the original data point, and obtain new enhanced data after point-by-point operation.

2. A pelletizing blower fault diagnosis method according to claim 1, characterized in that: In step S3, the specific steps are: S311, using the ConvNeXt network of the standard convolution module for architecture design, extracting the features of the time-frequency dataset; using the data enhanced The time-frequency map is used as the input layer; S312, first convolution layer selection The convolution kernel is set to 32 to extract different basic features. The "same" filling method and the step size are 1. The ReLU activation function is used. The parameters of the second and third convolution layers are consistent with those of the first convolution layer. The output feature map is ; S313, add pooling windows after the first and second convolutional layers respectively, using The maximum pooling window is 2, and the step size is 2; the feature map output by the convolutional layer is pooled to reduce the size of the feature map by half, and the shape of the output feature map becomes ; The second convolutional layer also uses The convolution kernel is set to 64, the filling method and step size remain unchanged, the activation function is still ReLU, and the input is the output of the previous layer. , the output feature map shape is ; use The maximum pooling window is 2, and the step size is 2; the feature map output by the convolutional layer is pooled to reduce the size of the feature map by half, and the shape of the output feature map becomes ; The third convolutional layer still chooses The convolution kernel is set to 128, with "same" padding and stride of 1, the activation function is ReLU, and the input is the output of the previous layer. , the output feature map shape is ; S314, in the two-dimensional convolution, each element in the output feature map The calculation formula is: ； in, Represents the element value at the corresponding position in the input feature map, Represents the weight value of the corresponding position in the convolution kernel, and is the spatial position index, is the channel index; S315, adds BatchNormalization after each convolution layer to speed up training and improve model stability; S316, after a series of convolutional layers and pooling layers, the shape is obtained as The feature map of and is the spatial size of the feature map, Is the number of channels; flatten this feature map into a one-dimensional vector with a length of ; S317, the first fully connected layer has neurons, through a weight matrix and the bias vector Perform linear transformation to obtain the output of the fully connected layer; the formula is: ； in, is the flattened eigenvector, is the ReLU function, the weight matrix Size , bias vector Length is ; S318, the second fully connected layer has K neurons, K is the number of fault classes, through a weight matrix and the bias vector Perform linear transformation to obtain the output of the fully connected layer; the formula is: ； in is the output of the first fully connected layer, is the softmax function, the weight matrix Size , bias vector Length is ; S32, divide the feature vectors obtained in S31 into 70% training set, 20% validation set, and 10% test set; use random partitioning to ensure that each category has a certain proportion distribution in different data sets; divide the training set into 75% support set and 25% query set; S321, the training set is used by the model to learn data patterns and regularities, adjust parameters and establish mapping relationships; the validation set is used to evaluate performance during training, helps adjust hyperparameters, and determine overfitting or underfitting; the test set is used to evaluate the final performance after training and measure generalization ability; the support set is used to calculate category prototypes; the query set is used to test the model's ability to classify new samples, and make classification decisions based on the distance from the category prototype obtained based on the support set; S33, calculation prototype: Computation prototype initialization phase: Divide the samples in the support set according to the fault category; mark the K fault categories as ;For each fault category , represents the kth fault category, K is the total number of fault categories, and a small number of initial samples are randomly selected to form the initial sample set: ； in, is the number of initial samples; calculate the initial prototype for each fault category , that is, the average value of the feature vector of the initial samples of this category; the calculation formula is: ； n represents the number of initial samples selected for each fault category in the initialization phase; Represents the feature vector of the i-th sample in the k-th fault category in the initial sample set; Calculation prototype batch update phase: As new samples are continuously added, the prototype is gradually updated using batch processing; the current number of iterations is , the newly added batch of samples are: ; represents a set of samples newly added to the k-th fault category at the t-th iteration; represents the feature vector of the mth sample newly added to the sample set of the kth fault category at the tth iteration; m represents the number of samples newly added to the kth fault category at the tth iteration, that is, the sample set The number of samples in ; The existing prototype is ;Calculate the total number of processed samples; ； For the The number of new samples added in the iteration; the prototype is updated after the iteration The calculation formula is: ； S34, distance metric: The fault categories are measured and classified by the nearest neighbor classification K-NN; the sample feature vector of the query set is calculated using the Euclidean distance method With prototype set The distance between two eigenvectors and , where x and y are any two feature vectors in the query set and prototype set respectively, and n is the dimension of the feature vector. Represents the feature vector of the nth sample in the kth fault category in the query set; The formula for calculating Euclidean distance is: ； n represents the dimension of the feature vector; S35, Decision Classification: According to the calculated distance, find the k nearest training samples, which are called The k nearest neighbors of ; count the frequency of occurrence of each category in these k nearest neighbor samples; Take the negative exponent of the distance and convert it into probability through the softmax function; for the sample , belongs to The probability calculation formula for each fault category is: ； in, is the total number of fault categories; Represents the sample x and the prototype of the i-th fault category the distance between them; represents the prototype vector of the i-th fault category; Will The class with the highest probability is classified as this class, which is considered to best represent the characteristics of the new sample.

3. A pelletizing blower fault diagnosis method according to claim 2, characterized in that: In step S4, the specific steps are: S41, using functions such as cross entropy loss as the basic loss function; The true category of the sample is , the predicted category is , the cross entropy loss function is: ； If the sample belongs to fault categories, then the true category vector , among which The first position is 1, and the rest are 0; S42, calculate the L2 regularization term based on the model parameters. The formula of the L2 regularization term is: ,in, are the parameters of the model, is the regularization coefficient; S45, add the L2 regularization term to the basic loss function to obtain a new loss function: ； S43, back propagation; S431, calculate the gradient of the loss function with respect to the model parameters: Calculate the predicted probability of a sample belonging to each fault category, the distance function, and the gradient of the L2 regularization term to the model parameters, so that the optimizer can be used to update the model parameters to improve the performance of the model and the accuracy of fault diagnosis; Use the chain rule to gradually calculate the gradient of the loss function with respect to the model parameters. For each parameter, calculate ; First, calculate the gradient of the cross entropy loss function with respect to the predicted probability ; in, is the cross entropy loss function, is the true category label of the sample, The sample belongs to The predicted probability of each fault category; According to the calculation formula of predicted probability, ,in Is the sample and The distance between the prototypes of the categories, is the total number of fault categories; calculate ,in Is the sample and The distance between the prototypes of the categories, is the total number of fault categories: Combined distance function The derivative of the parameters gives the gradient of the predicted probability of each fault category with respect to the model parameters: ； The formula for the L2 regularization term is ,in are the parameters of the model, is the regularization coefficient: the gradient of the L2 regularization term with respect to the parameter is calculated as : Finally, the total gradient is : S432, use the optimizer to update the model parameters: S4321, initialize Adam optimizer parameters: Setting the learning rate , usually between 0.001 and 0.01, the exponential decay rate of the first-order moment estimate Take 0.9; exponential decay rate of the second-order moment estimate Take 0.999; set a very small constant for , used to prevent the denominator from being zero; For each model parameter ; Initialize the first-order moment estimate and second-order moment estimates ; In the At iteration: Compute the first moment estimate: ; First-order moment estimate, which represents the exponentially weighted average of the gradient; Compute the second moment estimate: ; The second-order moment estimate represents the exponentially weighted average of the square of the gradient, which is used to estimate the variance of the gradient and help adjust the step size of the parameter update; Compute the corrected first moment estimate: ; Compute the corrected second moment estimates: ; Update parameters: ; By repeating the above steps, the Adam optimizer is used to update the model parameters in each iteration, so that the model is gradually optimized during the training process. S4322, regularly calculate the validation set loss: During model training, the loss value is calculated on the validation set in each training cycle. The loss function is the cross entropy loss plus the L2 regularization term, that is, ,in is the cross entropy loss function, is the regularization coefficient, are the model parameters: For each sample in the validation set, input it into the model for forward propagation, calculate the prediction result, then calculate the loss value based on the true label, and finally average the loss values ​​of all samples to get the average loss of the validation set; Stop training when the change in validation set loss is less than 5% over several consecutive training cycles. If the loss is unstable, continue training and adjust the model parameters.

4. A pelletizing blower fault diagnosis method according to claim 1, characterized in that: In step S5, the specific steps are: S5, model evaluation and testing further includes: S51 using the test set to evaluate the trained model; S52 analyzing the diagnosis results of the model to check whether there is any misclassification or missed diagnosis; S53 deploying the model.

5. A pelletizing blower fault diagnosis method according to claim 4, characterized in that: In the step S5, S51, using the test set to evaluate the trained model further includes: calculating the harmonic mean of F1 and recall, where the F1 value is the harmonic mean of precision and recall, which comprehensively considers the precision and recall capabilities of the model; Input the samples of the test set into the trained model and calculate the performance indicators of the model on the test set, such as accuracy, recall rate, F1 value and other parameters; Accuracy calculation formula: ； Recall calculation formula: ； The accuracy formula is: ； F1 formula: ； Based on the evaluation results on the validation set, adjust the model's hyperparameters, such as learning rate, regularization coefficient, and number of neighbors, to find the best hyperparameter combination; S52, analyzing the diagnosis results of the model to check whether there are any misclassifications or missed diagnoses further includes: traversing each sample in the test set and comparing the model's predicted category with the actual category; if the predicted category is inconsistent with the actual category, it is a misclassification; recording the features of the misclassified sample and the actual category and predicted category to facilitate cause analysis; for missed diagnoses, checking whether there is a situation where a fault actually exists but is not detected by the model; If the performance is stable, the model is deployed and loaded into the local server.

6. A pelletizing blower fault diagnosis method according to claim 5, characterized in that: S6 fault diagnosis and analysis, the specific steps are as follows: Repeat steps S1 and S2 to process and enhance the real-time data. Use the prototype network model built in S3 to predict the real-time feature vector and output the category with the highest probability as the predicted fault type. If it is found that the model's prediction results are obviously inconsistent with the actual situation or new data is collected, return to step S3 to re-divide the data set and retrain the model to improve its accuracy and adaptability.

Images