Dryer operation fault monitoring system and method based on machine learning

Abstract

The invention discloses a dryer operation fault monitoring system and method based on machine learning, and relates to the technical field of machine learning, and the method comprises the following steps: determining a dryer state category set, and defining the judgment standard of each category; collecting operation monitoring data; segmenting the data, and performing category labeling on samples to form a sample set; extracting features to form feature vectors, integrating to obtain a new sample set, and dividing into a training set, a verification set and a test set; constructing a multi-classification model to obtain a reference model; defining a core index to judge whether the effect of the reference model meets a set requirement; if not, judging whether the root cause is sample category imbalance or not; when the category is judged to be unbalanced, calculating a category weight adjustment weight loss function; retraining to obtain an optimization model; and verifying the effect, and performing prediction and feedback based on the optimization model when a set requirement is met. According to the method, the condition that the dryer fault classification prediction sample category imbalance influences identification in the prior art can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, specifically a machine learning-based dryer operation fault monitoring system and method. Background Technology

[0002] Dryers are key equipment in industrial production and agricultural processing, and their operational stability directly affects production efficiency and product quality. Traditional fault monitoring relies on manual inspection or single sensor threshold judgment, which has limitations such as response lag and high false negative rates. With the integration of IoT and machine learning technologies, intelligent fault diagnosis has become the mainstream development direction in this field by collecting data such as sensor signals and equipment operating parameters, and then performing feature extraction and model training, significantly improving the automation and intelligence level of monitoring.

[0003] In machine learning-driven dryer fault diagnosis, class imbalance is a prominent technical challenge. In real-world scenarios, the number of normally operating samples (majority class) far exceeds the number of faulty samples (minority class), causing the loss function to be dominated by the majority class during model training. This leads to overfitting of the model to majority class features, resulting in extremely low accuracy and frequent false negatives in minority class fault identification. Existing general imbalance handling methods lack scenario-specific adaptability and struggle to address the core issue of degraded fault identification performance in this field, failing to meet actual monitoring needs. Summary of the Invention

[0004] The purpose of this invention is to provide a machine learning-based dryer operation fault monitoring system and method to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based method for monitoring dryer operation faults, the method comprising the following steps: Step 1: Define the set of dryer status categories and define the criteria for each category; collect dryer operation monitoring data; divide the data into fixed time windows, and label the samples according to the criteria to form a sample set; Step 2: Extract features from the sample set to form feature vectors, and integrate the sample set and feature vectors to obtain a new sample set; based on the new sample set, divide it into a training set, a validation set, and a test set; Step 3: Construct a multi-class classification model based on the training set to obtain the benchmark model; define core metrics to quantify the performance of the benchmark model; determine whether the benchmark model's performance meets the set requirements based on the core metrics; when the benchmark model is determined to have failed to meet the set requirements, determine whether the root cause is class imbalance. Step 4: When the root cause is determined to be class imbalance, calculate the class weights based on the number of samples in each category of the new sample set; adjust the weighted loss function based on the class weights. Step 5: Based on the training set and validation set, retrain the multi-classification model using the adjusted weighted loss function to obtain the optimized model; verify the effect of the optimized model based on the test set; when the optimized model meets the set requirements, make predictions based on the optimized model and provide feedback.

[0006] In step 1, a high-quality, structured labeled sample set is constructed to provide a data foundation for subsequent model training and avoid the model performance being affected by ambiguous category definitions or data quality issues. Based on the dryer's operating mechanism, historical maintenance records, and industrial application scenarios, the overall category set C = {F0, F1, ..., F...} is defined. k Define the criteria for each category; for example: the criteria for F1 (motor failure) is "motor current fluctuation exceeds ±15% and lasts for ≥30s", the criteria for F2 (heating system failure) is "the difference between actual temperature and set temperature is ≥10℃ and lasts for ≥1min", and the criteria for F0 (normal state) is "all operating parameters are within the preset normal range". Where F0 represents the normal operating state, F1~F k The set of fault categories F is formed (e.g., F1 is motor fault, F2 is heating system fault, F3 is sensor fault, F4 is transmission system fault, F5 is control system fault, i.e., k=5), where k is the total number of fault categories; the total number of categories m=k+1; Collect operational monitoring data of the dryer under all operating conditions, including sensor output signals (such as real-time output data of temperature sensor, humidity sensor, current sensor, and vibration sensor) and equipment operating parameter records (such as start-up time, shutdown time, load, set drying temperature, and wind speed). The operational monitoring data is processed, including missing value imputation (using conventional data processing methods such as linear interpolation and mean imputation, for example: for missing values ​​caused by temporary sensor interruption, the mean of the five valid data points before and after is used to imput them) and outlier removal (using the 3σ criterion, box plot method, etc., for example: when a sensor data exceeds the range of μ±3σ (μ is the mean of the sensor data, σ is the standard deviation), it is judged as an outlier and removed). The continuous data after processing is divided into fixed time windows (the window duration can be set according to the operating characteristics of the dryer, such as 5 minutes / window, to ensure that each window can fully reflect the characteristics of a period of operation). Each dataset corresponding to a time window is defined as 1 sample. Each sample's Y is labeled by professionals according to the criteria of category set C. i All labeled samples constitute the sample set S; Where i represents any sample index; Y i ∈C represents the sample label and its corresponding category.

[0007] In step 2, the raw data is transformed into low-dimensional, high-discrimination feature vectors that the model can recognize, and the dataset is reasonably divided to avoid data leakage and ensure the effectiveness of model training and validation. For each sample, basic features in the time domain, frequency domain, or time series are extracted based on statistical analysis (such as mean, variance, peak value, kurtosis, skewness, etc., used to extract time-domain features; for example, the peak value of the motor current signal can reflect abnormal motor load) and signal transformation (such as Fast Fourier Transform (FFT), wavelet transform, etc., used to extract frequency-domain features; for example, by converting the vibration signal into a frequency-domain signal through FFT, peak features at a specific frequency can be extracted to reflect the bearing wear state). Features are then screened based on variance analysis and mutual information (the F-statistic of each feature is calculated through variance analysis, and features with F-statistics greater than a preset threshold are screened to remove redundant features with low discriminative power). The screened features are then standardized (using Z-score standardization). Feature dimensionality reduction techniques (such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), etc.) are used to remove redundant information between features and reduce the computational complexity of the model, resulting in the feature vector X for each sample. i ; Integrate the sample set S and the feature vector X of each sample i The new sample set S'={(X1,Y1),(X2,Y2),...,(X N ,Y N )}; Where N represents the total number of samples; i∈{1,2,…,N}; Based on the new sample set S', a hierarchical partitioning strategy is adopted to divide S' into a training set S'. tr Validation set S val With test set S te Maintain the proportion of each category of samples in the training set, validation set, and test set consistent with S', and avoid the absence of samples of a certain category in a certain dataset; In step 3, the problems of the original dataset are verified through the benchmark model, and the core root cause of sample class imbalance is accurately located, providing a basis for subsequent adjustment of the loss function; Based on a multi-class classification model, using the training set S tr Perform initial training to obtain the baseline model M0; Choose mature models in the field of fault diagnosis, such as tree models (XGBoost, LightGBM, suitable for small sample and high-dimensional feature scenarios) or deep learning models (CNN, LSTM, suitable for large data volume and time-series feature scenarios), and select the model according to the actual number of samples and data types. The training process does not introduce any class imbalance handling strategies, and is used to simulate the performance of regular machine learning models under imbalanced samples. Define core metrics to quantitatively evaluate M0 performance, specifically including minority class recall, weighted F1 score, and confusion matrix: Confusion matrix: Visually displays the prediction accuracy for each class of samples; for example: rows represent the true class, columns represent the predicted class, and matrix elements CM c1,c2 This indicates the number of samples whose true class is c1 but are predicted as c2, clearly showing whether minority class samples are being misclassified as majority class samples. Minority recall rate R min :R min =TP / (TP+FN); R min It directly reflects the model's ability to identify minority class faults; the higher the value, the lower the false negative rate of minority class faults. Among them, the minority class refers to the class whose sample size is less than a% of the total sample size, where a is set by professionals; TP represents the number of true samples in this class that are correctly predicted, and FN represents the number of true samples in this class that are incorrectly predicted; Weighted F1 score F w :F w =Σ c=0 m-1 n c / N·F1 c ; F w Taking into account the proportion of samples in each category, we avoid the problem of inflated accuracy caused by a high proportion of samples in the majority class; Among them, F1 c Let represent the F1 score of class c; the number of samples in class c is denoted as n. c ;c∈{0,1,...,m-1},Σ c=0 m-1 n c =N; When R min <θ1 and F w When the value is less than θ2, the baseline model is deemed to have failed to meet the set requirements. When the baseline model is determined to fail to meet the set requirements, the root cause is determined to be class imbalance: when the sample set S' contains both a majority class and a minority class, and the confusion matrix shows that the majority class recall is ≥α1 and the minority class recall is ≥α1. min When α < 2, the root cause is determined by class imbalance in the sample. Among them, the majority class refers to the class whose sample size accounts for more than b% of the total sample size, where b is set by professionals; θ1, θ2, α1, and α2 represent the thresholds preset based on the actual needs of industrial fault diagnosis.

[0008] In step 4, the loss function is adjusted by class weights so that the model training process focuses on minority class samples, thus solving the problem of majority class dominating the loss. When the root cause is determined to be class imbalance, the class weight w for each class is calculated based on the number of samples nc in each class of the sample set S'. c :w c =N / (m·n c ); The weight calculation logic is as follows: the smaller the number of minority class samples, the larger the weight, ensuring that the loss from minority class prediction errors is amplified; Based on w c Choose one of the following two suitable weighted loss functions to correct the bias: Weighted cross-entropy loss L WCE :L WCE =-Σ c=0 m-1 w c ·y c log(p c ); Among them, y c This represents the one-hot encoding of the sample's true label; y c =1 indicates that the sample belongs to class c; otherwise, it is 0. c This indicates the probability that the model predicts the sample belongs to class c; Weighted focus loss L WFL :L WFL =-Σ c=0 m-1 w c ·(1-p c ) γ ·y c log(p c ); Where γ≥0 represents the focusing parameter, with a default value of 2, which is used to reduce the loss proportion of easily classified samples, further strengthen the model's attention to difficult-to-classify samples such as minority classes, and avoid the model being overly biased towards the majority class; In step 5, the optimized model is trained and validated to ensure that it can meet the actual fault monitoring needs, and ultimately achieve real-time and accurate fault identification and alarm. Based on training set S tr and validation set S val The multi-class classification model was retrained using the adjusted weighted loss function, and during training, S... val Real-time monitoring R min With F w Optimize model hyperparameters (using conventional hyperparameter optimization methods such as grid search and random search; for example, XGBoost hyperparameters include learning rate, tree depth, regularization coefficient, etc., through S...) valThe optimal hyperparameter combination is selected based on the feedback of the indicators, resulting in the optimized model M1. If M1 does not meet the above criteria, repeat the training and validation process in steps 4-5 until the model meets the criteria or reaches the preset termination condition.

[0009] Based on test set S te Verify the effect of M1 by comparing the R values ​​of M1 and M0. min F w and confusion matrix, when R min <θ1 and F w When θ < 2, deploy M1 to the dryer real-time monitoring system: collect real-time operation monitoring data and extract feature vector X. real (Ensure that the real-time feature extraction method is consistent with the training set to avoid data distribution drift), predict the output class Y using M1. pred And the confidence level P (P represents the model's confidence in the prediction results); When Y pred When ∈F and P≥ preset threshold θ, the system immediately triggers an alarm and records the fault type, occurrence time and corresponding feature vector to achieve real-time fault monitoring.

[0010] A machine learning-based dryer operation fault monitoring system includes a data preprocessing module, a feature engineering module, a baseline model module, an imbalance processing module, and an optimization model module. The data preprocessing module is used to clarify the set of dryer status categories and define the judgment criteria for each category; collect dryer operation monitoring data; segment the data according to a fixed time window, and label the samples according to the judgment criteria to form a sample set; the feature engineering module is used to extract features from the sample set to form feature vectors, and integrate the sample set and feature vectors to obtain a new sample set; divide the new sample set into a training set, a validation set, and a test set; the benchmark model module is used to build a multi-classification model based on the training set to obtain a benchmark model; define core indicators to quantify the performance of the benchmark model; determine whether the benchmark model's performance meets the set requirements based on the core indicators; when the benchmark model is determined to have failed to meet the set requirements, determine whether the root cause is sample class imbalance; the imbalance handling module is used to calculate the class weights based on the number of samples in each category of the new sample set when the root cause is determined to be class imbalance; adjust the weighted loss function based on the class weights; the optimization model module is used to retrain the multi-classification model based on the training set and the validation set, using the adjusted weighted loss function to obtain an optimized model; verify the effect of the optimized model based on the test set; when the optimized model meets the set requirements, make predictions based on the optimized model and provide feedback.

[0011] The data preprocessing module includes a category definition unit, a data acquisition and cleaning unit, and a sample segmentation and labeling unit. The category definition unit is used to clarify the set of dryer status categories and define the criteria for each category; the data acquisition and cleaning unit is used to collect the operation monitoring data of the dryer under all working conditions and to fill in missing values ​​and remove outliers; the sample segmentation and labeling unit is used to segment the processed continuous data according to a fixed time window, form samples and label them with category tags to form a sample set.

[0012] The feature engineering module includes a feature extraction and filtering unit, a feature processing unit, and a dataset construction unit; The feature extraction and filtering unit is used to extract basic features in the time domain, frequency domain, or time series, and to filter the features; the feature processing unit is used to perform standardization and dimensionality reduction on the filtered features; the dataset construction unit is used to integrate the sample set and feature vectors to obtain a new sample set, and to divide it into training set, validation set, and test set.

[0013] The baseline model module includes a model training unit, a model evaluation unit, and a root cause analysis unit; The model training unit is used to construct a multi-classification model to obtain a baseline model; the model evaluation unit is used to define core indicators and determine whether the performance of the baseline model meets the set requirements; the root cause analysis unit is used to determine whether the root cause is sample class imbalance when the baseline model fails to meet the set requirements.

[0014] The imbalance processing module includes a weight calculation unit and a loss adjustment unit; The weight calculation unit is used to calculate the category weights based on the number of samples in each category; the loss adjustment unit is used to adjust the weighted loss function based on the category weights. The optimized model module includes a model retraining unit, a model validation unit, and a model deployment unit; The model retraining unit is used to retrain the multi-classification model using the adjusted weighted loss function; the model validation unit is used to validate the effect of the optimized model; and the model deployment unit is used to make predictions and provide feedback based on the optimized model to achieve real-time fault monitoring.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention uses a four-dimensional root cause determination mechanism of "minority class proportion threshold + majority class proportion threshold + double recall threshold + weighted F1 threshold" to identify sample class imbalance as the core cause of deterioration in fault identification performance, avoiding the ineffective optimization of existing technologies by blindly applying imbalance handling methods, and ensuring that technical resources are focused on the core problem; This invention uses a dynamic class weight formula derived based on the total number of samples, the number of classes, and the number of samples in each class, combined with a double-weighted loss function (weighted cross-entropy / weighted focus loss) adapted to the identification of difficult minority class samples, to effectively correct the core bias of the loss function being dominated by the majority class, ensuring the comprehensiveness of fault identification. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a machine learning-based dryer operation fault monitoring system according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example: Figure 1 As shown, the present invention provides a technical solution, a method for monitoring dryer operation faults based on machine learning, the method comprising the following steps: Step 1: Define the set of dryer status categories and define the criteria for each category; collect dryer operation monitoring data; divide the data into fixed time windows, and label the samples according to the criteria to form a sample set; Step 2: Extract features from the sample set to form feature vectors, and integrate the sample set and feature vectors to obtain a new sample set; based on the new sample set, divide it into a training set, a validation set, and a test set; Step 3: Construct a multi-class classification model based on the training set to obtain the benchmark model; define core metrics to quantify the performance of the benchmark model; determine whether the benchmark model's performance meets the set requirements based on the core metrics; when the benchmark model is determined to have failed to meet the set requirements, determine whether the root cause is class imbalance. Step 4: When the root cause is determined to be class imbalance, calculate the class weights based on the number of samples in each category of the new sample set; adjust the weighted loss function based on the class weights. Step 5: Based on the training set and validation set, retrain the multi-classification model using the adjusted weighted loss function to obtain the optimized model; verify the effect of the optimized model based on the test set; when the optimized model meets the set requirements, make predictions based on the optimized model and provide feedback.

[0019] In step 1, a high-quality, structured labeled sample set is constructed to provide a data foundation for subsequent model training and avoid the model performance being affected by ambiguous category definitions or data quality issues. Based on the dryer's operating mechanism, historical maintenance records, and industrial application scenarios, the overall category set C = {F0, F1, ..., F...} is defined. kDefine the criteria for each category; for example: the criteria for F1 (motor failure) is "motor current fluctuation exceeds ±15% and lasts for ≥30s", the criteria for F2 (heating system failure) is "the difference between actual temperature and set temperature is ≥10℃ and lasts for ≥1min", and the criteria for F0 (normal state) is "all operating parameters are within the preset normal range". Where F0 represents the normal operating state, F1~F k The set of fault categories F is formed (e.g., F1 is motor fault, F2 is heating system fault, F3 is sensor fault, F4 is transmission system fault, F5 is control system fault, i.e., k=5), where k is the total number of fault categories; the total number of categories m=k+1; Collect operational monitoring data of the dryer under all operating conditions, including sensor output signals (such as real-time output data of temperature sensor, humidity sensor, current sensor, and vibration sensor) and equipment operating parameter records (such as start-up time, shutdown time, load, set drying temperature, and wind speed). The operational monitoring data is processed, including missing value imputation (using conventional data processing methods such as linear interpolation and mean imputation, for example: for missing values ​​caused by temporary sensor interruption, the mean of the five valid data points before and after is used to imput them) and outlier removal (using the 3σ criterion, box plot method, etc., for example: when a sensor data exceeds the range of μ±3σ (μ is the mean of the sensor data, σ is the standard deviation), it is judged as an outlier and removed). The continuous data after processing is divided into fixed time windows (the window duration can be set according to the operating characteristics of the dryer, such as 5 minutes / window, to ensure that each window can fully reflect the characteristics of a period of operation). Each dataset corresponding to a time window is defined as 1 sample. Each sample's Y is labeled by professionals according to the criteria of category set C. i All labeled samples constitute the sample set S; Where i represents any sample index; Y i ∈C represents the sample label and its corresponding category.

[0020] In step 2, the raw data is transformed into low-dimensional, high-discrimination feature vectors that the model can recognize, and the dataset is reasonably divided to avoid data leakage and ensure the effectiveness of model training and validation. For each sample, basic features in the time domain, frequency domain, or time series are extracted based on statistical analysis (such as mean, variance, peak value, kurtosis, skewness, etc., used to extract time-domain features; for example, the peak value of the motor current signal can reflect abnormal motor load) and signal transformation (such as Fast Fourier Transform (FFT), wavelet transform, etc., used to extract frequency-domain features; for example, by converting the vibration signal into a frequency-domain signal through FFT, peak features at a specific frequency can be extracted to reflect the bearing wear state). Features are then screened based on variance analysis and mutual information (the F-statistic of each feature is calculated through variance analysis, and features with F-statistics greater than a preset threshold are screened to remove redundant features with low discriminative power). The screened features are then standardized (using Z-score standardization). Feature dimensionality reduction techniques (such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), etc.) are used to remove redundant information between features and reduce the computational complexity of the model, resulting in the feature vector X for each sample. i ; Integrate the sample set S and the feature vector X of each sample i The new sample set S'={(X1,Y1),(X2,Y2),...,(X N ,Y N )}; Where N represents the total number of samples; i∈{1,2,…,N}; Based on the new sample set S', a hierarchical partitioning strategy is adopted to divide S' into a training set S'. tr Validation set S val With test set S te Maintain the proportion of each category of samples in the training set, validation set, and test set consistent with S', and avoid the absence of samples of a certain category in a certain dataset; In step 3, the problems of the original dataset are verified through the benchmark model, and the core root cause of sample class imbalance is accurately located, providing a basis for subsequent adjustment of the loss function; Based on a multi-class classification model, using the training set S tr Perform initial training to obtain the baseline model M0; Choose mature models in the field of fault diagnosis, such as tree models (XGBoost, LightGBM, suitable for small sample and high-dimensional feature scenarios) or deep learning models (CNN, LSTM, suitable for large data volume and time-series feature scenarios), and select the model according to the actual number of samples and data types. The training process does not introduce any class imbalance handling strategies, and is used to simulate the performance of regular machine learning models under imbalanced samples. Define core metrics to quantitatively evaluate M0 performance, specifically including minority class recall, weighted F1 score, and confusion matrix: Confusion matrix: Visually displays the prediction accuracy for each class of samples; for example: rows represent the true class, columns represent the predicted class, and matrix elements CM c1,c2 This indicates the number of samples whose true class is c1 but are predicted as c2, clearly showing whether minority class samples are being misclassified as majority class samples. Minority recall rate R min :R min =TP / (TP+FN); R min It directly reflects the model's ability to identify minority class faults; the higher the value, the lower the false negative rate of minority class faults. Among them, the minority class refers to the class whose sample size is less than a% of the total sample size, where a is set by professionals; TP represents the number of true samples in this class that are correctly predicted, and FN represents the number of true samples in this class that are incorrectly predicted; Weighted F1 score F w :F w =Σ c=0 m-1 n c / N·F1 c ; F w Taking into account the proportion of samples in each category, we avoid the problem of inflated accuracy caused by a high proportion of samples in the majority class; Among them, F1 c Let represent the F1 score of class c; the number of samples in class c is denoted as n. c ;c∈{0,1,...,m-1},Σ c=0 m-1 n c =N; When R min <θ1 and F w When the value is less than θ2, the baseline model is deemed to have failed to meet the set requirements. When the baseline model is determined to fail to meet the set requirements, the root cause is determined to be class imbalance: when the sample set S' contains both a majority class and a minority class, and the confusion matrix shows that the majority class recall is ≥α1 and the minority class recall is ≥α1. min When α < 2, the root cause is determined by class imbalance in the sample. Among them, the majority class refers to the class whose sample size accounts for more than b% of the total sample size, where b is set by professionals; θ1, θ2, α1, and α2 represent the thresholds preset based on the actual needs of industrial fault diagnosis.

[0021] In step 4, the loss function is adjusted by class weights so that the model training process focuses on minority class samples, thus solving the problem of majority class dominating the loss. When the root cause is determined to be class imbalance, the class weight w for each class is calculated based on the number of samples nc in each class of the sample set S'. c :w c =N / (m·n c ); The weight calculation logic is as follows: the smaller the number of minority class samples, the larger the weight, ensuring that the loss from minority class prediction errors is amplified; Based on w c Choose one of the following two suitable weighted loss functions to correct the bias: Weighted cross-entropy loss L WCE :L WCE =-Σ c=0 m-1 w c ·y c log(p c ); Among them, y c This represents the one-hot encoding of the sample's true label; y c =1 indicates that the sample belongs to class c; otherwise, it is 0. c This indicates the probability that the model predicts the sample belongs to class c; Weighted focus loss L WFL :L WFL =-Σ c=0 m-1 w c ·(1-p c ) γ ·y c log(p c ); Where γ≥0 represents the focusing parameter, with a default value of 2, which is used to reduce the loss proportion of easily classified samples, further strengthen the model's attention to difficult-to-classify samples such as minority classes, and avoid the model being overly biased towards the majority class; In step 5, the optimized model is trained and validated to ensure that it can meet the actual fault monitoring needs, and ultimately achieve real-time and accurate fault identification and alarm. Based on training set S tr and validation set S val The multi-class classification model was retrained using the adjusted weighted loss function, and during training, S... val Real-time monitoring R min With F w Optimize model hyperparameters (using conventional hyperparameter optimization methods such as grid search and random search; for example, XGBoost hyperparameters include learning rate, tree depth, regularization coefficient, etc., through S...) val The optimal hyperparameter combination is selected based on the feedback of the indicators, resulting in the optimized model M1. If M1 does not meet the above criteria, repeat the training and validation process in steps 4-5 until the model meets the criteria or reaches the preset termination condition.

[0022] Based on test set S te Verify the effect of M1 by comparing the R values ​​of M1 and M0. min F w and confusion matrix, when R min <θ1 and F w When θ < 2, deploy M1 to the dryer real-time monitoring system: collect real-time operation monitoring data and extract feature vector X. real (Ensure that the real-time feature extraction method is consistent with the training set to avoid data distribution drift), predict the output class Y using M1. pred And the confidence level P (P represents the model's confidence in the prediction results); When Y pred When ∈F and P≥ preset threshold θ, the system immediately triggers an alarm and records the fault type, occurrence time and corresponding feature vector to achieve real-time fault monitoring.

[0023] A machine learning-based dryer operation fault monitoring system includes a data preprocessing module, a feature engineering module, a baseline model module, an imbalance processing module, and an optimization model module. The data preprocessing module is used to clarify the set of dryer status categories and define the judgment criteria for each category; collect dryer operation monitoring data; segment the data according to a fixed time window, and label the samples according to the judgment criteria to form a sample set; the feature engineering module is used to extract features from the sample set to form feature vectors, and integrate the sample set and feature vectors to obtain a new sample set; divide the new sample set into a training set, a validation set, and a test set; the benchmark model module is used to build a multi-classification model based on the training set to obtain a benchmark model; define core indicators to quantify the performance of the benchmark model; determine whether the benchmark model's performance meets the set requirements based on the core indicators; when the benchmark model is determined to have failed to meet the set requirements, determine whether the root cause is sample class imbalance; the imbalance handling module is used to calculate the class weights based on the number of samples in each category of the new sample set when the root cause is determined to be class imbalance; adjust the weighted loss function based on the class weights; the optimization model module is used to retrain the multi-classification model based on the training set and the validation set, using the adjusted weighted loss function to obtain an optimized model; verify the effect of the optimized model based on the test set; when the optimized model meets the set requirements, make predictions based on the optimized model and provide feedback.

[0024] The data preprocessing module includes a category definition unit, a data acquisition and cleaning unit, and a sample segmentation and labeling unit. The category definition unit is used to clarify the set of dryer status categories and define the criteria for each category; the data acquisition and cleaning unit is used to collect the operation monitoring data of the dryer under all working conditions and to fill in missing values ​​and remove outliers; the sample segmentation and labeling unit is used to segment the processed continuous data according to a fixed time window, form samples and label them with category tags to form a sample set.

[0025] The feature engineering module includes a feature extraction and filtering unit, a feature processing unit, and a dataset construction unit; The feature extraction and filtering unit is used to extract basic features in the time domain, frequency domain, or time series, and to filter the features; the feature processing unit is used to perform standardization and dimensionality reduction on the filtered features; the dataset construction unit is used to integrate the sample set and feature vectors to obtain a new sample set, and to divide it into training set, validation set, and test set.

[0026] The baseline model module includes a model training unit, a model evaluation unit, and a root cause analysis unit; The model training unit is used to construct a multi-classification model to obtain a baseline model; the model evaluation unit is used to define core indicators and determine whether the performance of the baseline model meets the set requirements; the root cause analysis unit is used to determine whether the root cause is sample class imbalance when the baseline model fails to meet the set requirements.

[0027] The imbalance processing module includes a weight calculation unit and a loss adjustment unit; The weight calculation unit is used to calculate the category weights based on the number of samples in each category; the loss adjustment unit is used to adjust the weighted loss function based on the category weights. The optimized model module includes a model retraining unit, a model validation unit, and a model deployment unit; The model retraining unit is used to retrain the multi-classification model using the adjusted weighted loss function; the model validation unit is used to validate the effect of the optimized model; and the model deployment unit is used to make predictions and provide feedback based on the optimized model to achieve real-time fault monitoring.

[0028] In this embodiment, a dryer in a food processing plant needs to achieve intelligent fault monitoring, and the method of the present invention is implemented as follows: The categories are clearly defined as normal state (F0), motor failure (F1), heating system failure (F2), sensor failure (F3), and transmission system failure (F4), and the judgment criteria for each category are defined based on operating parameters. Collect sensor signals and operation logs of the dryer, including temperature, current, and vibration. After missing value filling and outlier removal, the data is divided into 10-minute time windows and labeled to form a sample set. The time-domain and frequency-domain features of each sample were extracted, and the feature vectors were obtained by variance analysis, standardization and PCA dimensionality reduction. The samples were then divided into training set, validation set and test set. XGBoost was used to build a baseline model. The evaluation found that the recall rate of sensor failure (F3) was low and there were majority class (F0) and minority class (F3), which was judged as sample imbalance. The class weights are calculated based on the number of samples in each class, and the model is retrained using weighted cross-entropy loss to obtain the optimized model. Test set validation showed a significant improvement in minority class recall. After the model met the standards, it was deployed, and data was collected and features were extracted in real time. When the prediction was a fault category and the confidence level met the standards, an alarm was triggered and relevant data was recorded.

[0029] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A machine learning-based dryer operation failure monitoring method, characterized by: The method comprises the following steps: Step 1, defining a set of dryer state categories, defining category determination criteria; collecting dryer operation monitoring data; dividing the data into fixed time windows, labeling the samples based on the determination criteria to form a sample set; Step 2, extracting features of the sample set to form a feature vector, integrating the sample set and the feature vector to obtain a new sample set; dividing the training set, validation set and test set based on the new sample set; Step 3, constructing a multi-classification model based on the training set to obtain a benchmark model; defining a core index to quantify the performance of the benchmark model; determining whether the benchmark model meets the set requirements based on the core index; when the benchmark model is determined to not meet the set requirements, determining whether the root cause is sample category imbalance; Step 4, when the root cause is determined to be category imbalance, calculating the category weight based on the number of samples of each category in the new sample set; adjusting the weighted loss function based on the category weight; Step 5, retraining the multi-classification model using the adjusted weighted loss function based on the training set and the validation set to obtain an optimized model; verifying the effect of the optimized model based on the test set; when the optimized model meets the set requirements, making a prediction based on the optimized model and feeding back. 2.The machine learning-based dryer operation failure monitoring method of claim 1, wherein: In step 1, based on the operation mechanism of the dryer, historical maintenance records and industrial application scenarios, the overall category set C = {F0, F1,..., F k} is determined; the determination criteria of each category are defined; Wherein, F0 represents normal operation state, F1~F k F, k is the total number of fault categories; the total number of categories m=k+1; Collecting operation monitoring data of the dryer under all working conditions, including sensor output signals and device operation parameter records; Processing the operation monitoring data, including missing value filling and outlier removal; Dividing the processed continuous data into fixed time windows, and defining the data set corresponding to each time window as one sample; Y is annotated for each sample by a professional according to the judgment standard of the category set C i All the samples with completed annotation constitute a sample set S where i denotes any sample index; Y i ∈ C, denotes sample label, corresponding to the category to which it belongs. 3.The machine learning-based dryer operation failure monitoring method of claim 2, wherein: In step 2, for each sample, extract time domain, frequency domain or time series basic features based on statistical analysis and signal transformation; Filtering features based on variance analysis and mutual information method; Standardizing the filtered features; Using feature dimension reduction technique, get the feature vector X of each sample i ; Integrate the sample set S and the feature vector X of each sample i Obtain a new sample set S' = {(X1, Y1), (X2, Y2),..., (X N , Y N )}; Wherein, N represents the total number of samples; i∈{1,2,…,N}; Based on the new sample set S', the stratified division strategy is adopted to divide S' into a training set S tr , a validation set S val , and a test set S te .

4. The machine learning-based dryer operation failure monitoring method of claim 3, wherein: In step 3, based on the multi-classification model, the training set S is used to train the base model M0. tr An initial training is performed to obtain a base model M0. Defining a core index to quantify the performance of M0, including minority class recall rate, weighted F1 score and confusion matrix: minority class recall R min : R min = TP / (TP + FN); Wherein, the minority class refers to a category whose sample quantity accounts for less than a% of the total number of samples, and a is set by professionals; TP represents the number of true samples of this category that are correctly predicted, and FN represents the number of true samples of this category that are incorrectly predicted; Weighted F1 score F w : F w =∑ c=0 m-1 n c / N·F1 c ; wherein F1 c represents the F1 score of the cth class; the number of samples of the cth class is denoted as n c ; c e {0, 1,..., m-1},∑ c=0 m- 1 n c =N; When R min <θ1and F w <θ2, it is determined that the reference model effect does not meet the set requirements; When the reference model is determined to not meet the set requirements, it is determined whether the root cause is sample class imbalance: when there are both majority classes and minority classes in the sample set S', and the confusion matrix shows that the recall rate of the majority class is greater than or equal to α1, the recall rate of the minority class is less than α2, and the recall rate of the minority class is less than α2, it is determined that the root cause is sample class imbalance. min <α2, it is determined that the root cause is sample class imbalance; Wherein, the majority class refers to a category whose sample quantity accounts for more than b% of the total number of samples, and b is set by professionals; θ1, θ2, α1, α2 represent preset thresholds.

5. The machine learning based dryer operation failure monitoring method of claim 4, wherein: In step 4, when the root cause is determined as class imbalance, the class weight w of each class is calculated based on the number of samples nc of each class in the sample set S' c : w c = N / (m · n c ) Based on w c , the following two adapted weighted loss functions are selected to correct the bias: weighted cross-entropy loss L WCE : L WCE = -∑ c=0 m-1 w c ·y c log(p c ) where y c is the one-hot encoding of the true label of the sample; y c = 1 indicates that the sample belongs to the c-th class, otherwise 0; p c is the probability that the model predicts that the sample belongs to the c-th class; Weighted focal loss L WFL : L WFL = -∑ c=0 m-1 w c ·(1-p c ) γ ·y c log(p c ) Wherein, γ≥0 represents the focus parameter; In step 5, based on the training set S tr and the validation set S val , the multi-classification model is retrained using the adjusted weighted loss function, and the training process is monitored in real time through S val R min and F w , the model hyperparameters are optimized, and an optimized model M1 is obtained; Based on the test set S te Verify the effect of M1, compare R of M1 and M0 min , F w and confusion matrix, when R min <θ1 and F w <θ2, deploy M1 to the real-time monitoring system of the dryer: real-time acquisition of operation monitoring data, extraction of feature vector X real , prediction of output category Y pred and confidence P by M1; When Y pred ∈F and P is greater than or equal to a preset threshold θ, the system immediately triggers an alarm, and records the fault category, the occurrence time and the corresponding feature vector, realizing real-time fault monitoring.

6. A machine learning based dryer operation fault monitoring system applied to a machine learning based dryer operation fault monitoring method according to any one of claims 1-5, characterized in that: The system comprises a data preprocessing module, a feature engineering module, a benchmark model module, an imbalance processing module and an optimized model module; The data preprocessing module is used to define a set of dryer state categories, define category determination criteria, collect dryer operation monitoring data, divide the data into fixed time windows, label the samples based on the determination criteria to form a sample set; the feature engineering module is used to extract features of the sample set to form a feature vector, integrate the sample set and the feature vector to obtain a new sample set; divide the training set, validation set and test set based on the new sample set; the benchmark model module is used to construct a multi-classification model based on the training set to obtain a benchmark model; define a core index to quantify the performance of the benchmark model; determine whether the benchmark model meets the set requirements based on the core index; When the benchmark model is determined to not meet the set requirement, it is determined whether the root cause is sample category imbalance; the imbalance processing module is configured to calculate category weights based on the number of samples of each category in the new sample set when the root cause is determined to be category imbalance; The weighted loss function is adjusted based on the category weights; the optimization model module is configured to retrain the multi-classification model based on the training set and the validation set using the adjusted weighted loss function to obtain an optimized model; the effect of the optimized model is verified based on the test set; when the optimized model meets the set requirement, prediction is performed based on the optimized model and feedback is provided.

7. A machine learning based dryer operation fault monitoring system as claimed in claim 6, wherein: The data preprocessing module includes a category definition unit, a data acquisition and cleaning unit, and a sample segmentation and labeling unit; The category definition unit is configured to define a set of dryer state categories and define determination criteria for each category; the data acquisition and cleaning unit is configured to acquire running monitoring data of the dryer under full working conditions, and perform missing value filling and outlier removal; the sample segmentation and labeling unit is configured to segment the processed continuous data according to a fixed time window, form samples and label category labels, and form a sample set.

8. The machine learning based dryer operation fault monitoring system of claim 7, wherein: The feature engineering module includes a feature extraction and screening unit, a feature processing unit, and a data set construction unit; The feature extraction and screening unit is configured to extract time domain, frequency domain, or time series basic features and screen features; the feature processing unit is configured to perform standardization processing and dimensionality reduction processing on the screened features; the data set construction unit is configured to integrate the sample set and the feature vector to obtain a new sample set, and divide the new sample set into a training set, a validation set, and a test set.

9. The machine learning based dryer operation fault monitoring system of claim 8, wherein: The benchmark model module includes a model training unit, a model evaluation unit, and a root cause analysis unit; The model training unit is configured to construct a multi-classification model to obtain a benchmark model; the model evaluation unit is configured to define core indicators and determine whether the effect of the benchmark model meets the set requirement; The root cause analysis unit is configured to determine whether the root cause is sample category imbalance when the benchmark model does not meet the set requirement.

10. The machine learning based dryer operation failure monitoring system of claim 9, wherein: The imbalance processing module includes a weight calculation unit and a loss adjustment unit; The weight calculation unit is configured to calculate category weights based on the number of samples of each category; the loss adjustment unit is configured to adjust the weighted loss function based on the category weights; The optimization model module includes a model retraining unit, a model verification unit, and a model deployment unit; The model retraining unit is configured to retrain the multi-classification model using the adjusted weighted loss function; the model verification unit is configured to verify the effect of the optimized model; the model deployment unit is configured to perform prediction based on the optimized model and provide feedback to realize real-time fault monitoring.

Images