A method and system for evaluating the health of a cut tobacco machine based on data augmentation and ensemble learning

Abstract

The application discloses a kind of based on data enhancement and integrated learning's cut tobacco machine health degree evaluation method and system, comprising: cut tobacco machine data preprocessing and feature analysis, realize data pseudo-label marking and data screening, the calculation of data feature importance and the sorting and screening of feature.Cut tobacco machine single model health degree value calculation, through respectively establishing based on long short-term memory network and based on the health degree value calculation model of antagonsitic self-encoder, realize the data enhancement of monitoring data.Cut tobacco machine combination model health degree calculation and evaluation, through index correlation weighting method and fuzzy evaluation method, establish cut tobacco machine health degree value calculation integrated model, to realize the health degree evaluation of cut tobacco machine.The application can solve the problem that the health condition of cut tobacco machine is difficult to effectively evaluate under the condition that sample label is unbalanced and data quantity is insufficient.

Description

Technical Field

[0001] The present invention belongs to the fields of electronic engineering and computer science, and particularly relates to a tofu drying machine health assessment method and system based on data enhancement and ensemble learning. Background Art

[0002] The tobacco cut drying process, a crucial step in tobacco production, directly impacts the quality of the cut tobacco. The cut tobacco dryer is a key piece of equipment in the intelligent production of the cut tobacco process. The dryer adjusts the moisture content and other quality indicators of the dried cut tobacco by adjusting process parameters, directly impacting the quality of the cut tobacco. However, malfunctions during the dryer's operation can lead to equipment malfunction or even downtime, resulting in reduced or even substandard product quality, safety incidents, and financial losses.

[0003] To assess the health of complex equipment, such as tofu dryers, various technologies and methods are required. These include advanced technologies such as data mining, machine learning, and artificial intelligence. These technologies can analyze large amounts of historical data, uncovering the inherent connections and patterns within the data, and thus enabling health assessment of complex equipment.

[0004] Equipment health assessment has been applied, but challenges persist, such as high data feature dimensionality, insufficient data labels, and difficulty in accurately and real-time fault monitoring. Existing methods are less than ideal for concrete tofu drying equipment. This is because these models do not fully match the monitoring data and actual conditions of the tofu drying equipment. Therefore, a method with higher matching accuracy and the ability to address these issues, such as high data feature dimensionality and insufficient data labels, is needed. Summary of the Invention

[0005] To address technical issues such as the difficulty in accurately assessing the health of a tofu-strip drying machine, the present invention first preprocesses the data and performs pseudo-label extraction and feature analysis. A health value calculation model based on a bidirectional long-short-term memory network and an adversarial autoencoder are proposed, respectively. Through ensemble learning, a fuzzy evaluation of the tofu-strip drying machine's health is performed based on these models.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The present invention provides a method for evaluating the health of a tofu drying machine based on data enhancement and ensemble learning, comprising the following steps:

[0008] Step (1), preprocessing and feature analysis of the tofu drying machine data; realizing data pseudo-label marking and data screening, calculation of data feature importance, and feature sorting and screening;

[0009] Step (2), calculating the health value of a single model of the tofu drying machine, by respectively establishing a health value calculation model based on a long short-term memory network and a health value calculation model based on an adversarial autoencoder, thereby achieving data enhancement of the monitoring data;

[0010] Step (3): Calculate and evaluate the health of the combined model of the tofu drying machine, establish an integrated model for calculating the health value of the tofu drying machine, and thus realize the health evaluation of the tofu drying machine.

[0011] The present invention provides a tofu drying machine health assessment system based on data enhancement and ensemble learning, comprising:

[0012] The tofu drying machine data preprocessing and feature analysis module implements data pseudo-labeling and data screening, calculates the importance of data features, and sorts and screens features;

[0013] The single-model health value calculation module for the tofu drying machine achieves data enhancement of monitoring data by establishing health value calculation models based on long-short-term memory networks and adversarial autoencoders.

[0014] The tofu drying machine combined model health calculation and evaluation module establishes an integrated model for tofu drying machine health value calculation, thereby realizing tofu drying machine health evaluation.

[0015] The present invention has the following beneficial effects:

[0016] This paper proposes a method for pseudo-label extraction and data screening based on semi-supervised learning for tofu drying machine monitoring data with insufficient label information. This method solves the problem of insufficient data set labels to a certain extent and improves the calculation speed of the health assessment model.

[0017] The present invention proposes a data enhancement method for tofu drying machine equipment. It uses an adversarial autoencoder to generate real-time monitoring data after tofu drying machine screening, and performs real-time calculation of reconstruction error, which improves the diversity of data set patterns and the generalization performance of the health assessment model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a structural block diagram of the tofu drying machine health assessment method based on data enhancement and ensemble learning of the present invention. DETAILED DESCRIPTION

[0019] In order to make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other. To achieve the above-mentioned objectives, the present invention adopts the following technical solutions.

[0020] The present invention discloses a tow-bread drying machine health assessment method based on data enhancement and ensemble learning, comprising: data preprocessing, which processes missing values, outliers and preliminary processing with strong interpretability for collected data; pseudo-label marking, which uses a clustering algorithm to perform preliminary label marking on the data; a feature extraction module uses extreme gradient boosting to calculate, sort and screen the importance of data features; tow-bread drying machine health value calculation, which calculates the health value based on a bidirectional long short-term memory network and an adversarial autoencoder model according to specific tow-bread drying machine monitoring data; a tow-bread drying machine health assessment block, which obtains an integrated health value calculation method through ensemble learning, and performs fuzzy evaluation based on the health value, so as to realize real-time assessment of the health status of the tow-bread drying machine.

[0021] The structural diagram of the tofu drying machine health evaluation method based on data enhancement and ensemble learning of the present invention is as follows: Figure 1 As shown, it includes data preprocessing and feature analysis of the tofu drying machine health assessment, health value calculation of the tofu drying machine health assessment, and health assessment of the tofu drying machine. The specific steps are as follows:

[0022] Step (1), data preprocessing and feature analysis of the tofu drying machine, is specifically implemented as follows:

[0023] Step (1.1), data preprocessing: Data preprocessing specifically includes missing value processing, outlier processing, preliminary feature processing, data standardization and normalization, and filtering.

[0024] Among them, missing value processing is to fill the missing data with the average of the five data before and after the missing data. Outlier processing is to check whether there are abnormal data in the original data. Its method is to replace the missing data with the average of the five data before and after the missing data. The specific processing formulas for missing values ​​and outlier values ​​are as follows:

[0025] (1)

[0026] In the above formula, N is the total number of data. is the i+jth data, is the data used to supplement or replace the i-th position.

[0027] On this basis, the original features are preliminarily merged, added, and deleted according to prior knowledge. For the monitoring of the tofu drying machine, the "hot air speed setting value" and the "hot air speed actual value" are merged into the feature "hot air speed working deviation" which is of practical significance for health assessment. The "BBHS_area 1 barrel temperature setting value" and the "BBHS_area 1 barrel temperature actual value" are merged into "BBHS_area 1 barrel temperature", and the "BBHS_area 2 barrel temperature setting value" and the "BBHS_area 2 barrel temperature actual value" are merged into "BBHS_area 2 Cylinder temperature", "BBHS_Pressure setting value at discharge hood" and "BBHS_Pressure actual value at discharge hood" are merged into "BBHS_Pressure at discharge hood", "BBHS_Final cut tobacco moisture setting value" and "BBHS_Final cut tobacco moisture actual value" are merged into "BBHS_Final cut tobacco moisture working deviation", "HT_Steam pressure preferred setting value" and "HT_Steam pressure preferred actual value" are merged into "HT_Steam pressure preferred working deviation", "Drying tube wall 2 temperature preferred (set) value" and "Drying tube wall 2 temperature preferred (set) value" are merged into "Drying tube wall 2 temperature preferred working deviation", "Drying tube wall 1 temperature preferred (set) value" and "Drying tube wall 1 temperature preferred (set) value" are merged into "Drying tube wall 1 temperature preferred working deviation", and the feature "Cumulative amount of cut tobacco flow" that has little relationship with the actual fault is deleted. Data standardization and normalization are the processing of different data features so that the data are all on the same scale and prevent certain features from occupying too dominant a position. The specific formula is as follows:

[0028] (2)

[0029] In the above formula, is the i-th data, and are the minimum and maximum data values, respectively. is the normalized result of the i-th data.

[0030] Then, the noise interference is eliminated by using the data of the five previous and next time steps for sliding average filtering;

[0031] Step (1.2), pseudo label extraction: This step is to solve the problem of insufficient label marking. A small number of features are selected based on prior knowledge, and the labels are initially calibrated and updated through the k-means clustering algorithm. Clusters n_clusters=2 and n_clusters=4 are selected as coarse labels and fine labels respectively. Each cluster is set to a different label value according to its degree of deviation from the healthy operation state. The greater the degree of deviation of the cluster center from the healthy operation state, the larger the label value is set. Judgment, where d is the degree of deviation, , They are clusters and data corresponding to healthy operation. Each cluster corresponds to a different label size. The coarse label is set to 0 (healthy) and 1 (faulty). The label sizes of the four fine label clusters are set to 0, 0.4, 0.7, and 1. Run the k-means model training results under the two cluster parameters to obtain the corresponding two sets of labels. Then, the confidence evaluation and screening are performed based on the distance between the cluster sample point and the cluster center. The confidence is set to α=0.05, and the clustering results are screened. The monitoring data corresponding to the clustering results after screening are used as available data and recorded as AD.

[0032] Step (1.3), feature analysis and screening: For the problem of high feature dimension of monitoring data and unclear impact on health, feature screening method is used to solve it. Data feature mining analysis and feature importance calculation are performed through the extreme gradient boosting algorithm. This algorithm has better performance and accuracy than algorithms such as gradient boosting tree, and can perform data feature mining analysis and feature importance calculation. First, input the corresponding features of the original tofu drying machine monitoring data, and set the tree depth max_depth, number of trees (ie, number of iterations) n_estimators, and learning rate learning_rate related parameters. For the XGBoost model with set parameters, perform model pre-training and parameter adjustment, and save the training results. Calculate the feature importance of the trained model and sort it from large to small according to the feature importance. Select the feature screening number n, and select the features corresponding to the largest n feature importances as the filtered features based on this, and retain the corresponding data of the filtered features as the subsequent model training data based on the feature screening results;

[0033] Step (2), the single model health value calculation module of the tofu drying machine is specifically implemented as follows:

[0034] Step (2.1), calculate the health value of the bidirectional long short-term memory network model: to calculate the health value of the tofu drying machine. A Bi-LSTM method is used to take advantage of the historical temporal nature of the data. First, the specific network structure and hyperparameters of the Bi-LSTM need to be analyzed and set. The Bi-LTSM mainly consists of a three-layer structure: input layer, hidden layer, and output layer. The hidden layer contains three gated unit structures: forget gate, input gate, and output gate, which can be regarded as three fully connected layers. The input gate is used to control the proportion of information flowing into the hidden layer, the forgetting layer is used to control the proportion of information to be forgotten, and the output layer is used to calculate the output value at the current moment. The Bi-LSTM network structure is pre-set based on empirical knowledge, and the model is pre-trained using the available data AD obtained in step (1.2) to obtain the corresponding model results and parameters. Then, based on the comparison of the training results with the available data AD, the specific network structure and hyperparameters of the Bi-LSTM are further adjusted. Ultimately, a four-layer Bi-LSTM structure and one fully connected layer were selected as the basic structure for the training model (32 layers, 16 layers, 16 layers, and 8 layers, respectively). The input feature dimension was set to 14 to match the dimensionality of the monitoring data. By training the Bi-LSTM model, the temporal features and patterns of the raw data were learned, thereby learning the correspondence between the tofu drying machine monitoring data and its health value. By training the Bi-LSTM model, the temporal features and patterns of the raw data were learned, thereby learning the correspondence between the tofu drying machine monitoring data and its health value. The input was the filtered feature-corresponding data, and the output was the tofu drying machine's real-time health value. For the tofu drying machine's time-series monitoring data, each Bi-LSTM layer captured the data's time series characteristics. The model hyperparameters were then determined and modified as appropriate based on the training results. The specific structure is shown in Table 1. The Bi-LSTM model was retrained, and the trained model parameters were saved, denoted as M1. Finally, the available data AD was used as the model input, and the health value was calculated based on the trained model parameters M1, denoted as h1.

[0035] Table 1 Bi-LSTM network structure and hyperparameter settings

[0036]

[0037] Step (2.2), data augmentation of the adversarial autoencoder model: the adversarial autoencoder (AAE) is a data generation model with high training stability and rich generated data, which is suitable for generating high-dimensional and nonlinear monitoring data of the cut tobacco drying machine. At the same time, the problem of insufficient data is solved by data augmentation. The structure specifically includes an encoder, a decoder, and a discriminator. The encoder is used to reduce the dimension of the original data and encode it into abstract features. The decoder is used to decode and restore the encoded abstract features. The discriminator inputs the Gaussian random distribution sampling and the original data dimension reduction result respectively, and needs to distinguish whether the input data is the original data. The structure of the three parts of the model is determined respectively, and a linear layer is used as the basic structure of the three modules. Based on the empirical knowledge, the AAE network structure is pre-set, and the model is pre-trained based on the available data AD obtained in step (1.2) to obtain the corresponding results and parameters of the model. Then, according to the comparison between the training results and the available data AD, the specific network structure and hyperparameters of the AAE are further adjusted to obtain the final structure of the AAE: a four-layer linear layer encoder, a four-layer decoder structure, and a four-layer discriminator structure, with an input feature dimension of 14 (corresponding to the monitoring data dimension). Based on the selected loss function and learning rate, the model is trained and the parameters are adjusted, and the trained results are saved. The specific model structure is shown in Table 2. The AAE model with determined hyperparameters is trained again to establish the correspondence between the cut tobacco drying machine monitoring data and the health degree value of the cut tobacco drying machine, and the trained model parameters are saved. Through the trained AAE model, normal data ND and fault data FD are generated to supplement and enhance the original data;

[0038] Step (2.3), health degree value calculation based on the adversarial autoencoder model: the data reconstruction error e of the normal data ND and the available data AD obtained in step (2.2) is calculated according to the formula , and the reconstruction error is normalized to obtain a quantitative value representing the health degree of the cut tobacco drying machine (i.e., the health degree value), denoted as h2. Through the gradient descent algorithm, the 14-dimensional feature data is taken as the input and the data reconstruction error is taken as the output, and the health degree value is linearly regressed to obtain the quantitative correspondence between the available data AD and the health degree value, denoted as M2.

[0039] Table 2 AAE network structure and hyperparameter setting

[0040]

[0041] Step (3), health degree calculation and evaluation of the cut tobacco drying machine combination model, which is implemented as follows:

[0042] Step (3.1), health degree value calculation based on integrated model: adopt Stacking integrated model, and use the index correlation weighting method (CRITIC method) to objectively weight. This method can reflect the relationship between data and comprehensively consider the relationship between various indicators and data volatility of each single observation method.

[0043] The specific calculation method is as follows:

[0044] For M1, M22 single models (evaluation objects) corresponding to root mean square error MSE, mean absolute error MAE, decision coefficient R 2 , and fuzzy evaluation error rate J, set the index evaluation matrix:

[0045] (3)

[0046] Where y ij represents the evaluation value of the i-th evaluation index of the j-th single model, n is the total number of models (n=2), and m is the number of evaluation indexes (m=4).

[0047] For cost-type indexes (the smaller the better), normalize and calculate the mean value:

[0048] (4)

[0049] (5)

[0050] u ij and are the standardized data and data mean value, respectively.

[0051] For standard deviation S j and correlation coefficient r jj' (j and j' between two indexes), calculate:

[0052] (6)

[0053] (7)

[0054] Then calculate the conflict f j between the j-th index and all indexes:

[0055] (8)

[0056] Calculate the information amount C j contained in the j-th evaluation index:

[0057] (9)

[0058] Weight the amount of information and calculate the objective weight w j :

[0059] (10)

[0060] This method assigns objective weights based on the degree of indicator difference and the correlation between the indicators. The larger the standard deviation of an indicator, the more information it provides when distinguishing different evaluation objects, and the stronger the corresponding objective weight. The greater the degree of negative correlation between two indicators, the stronger their conflict, the greater the difference in information provided when distinguishing different evaluation models, and the stronger the corresponding objective weight. The calculation formulas for some evaluation indicators are as follows:

[0061] Root mean square error (MSE):

[0062] (11)

[0063] Mean absolute error MAE:

[0064] (12)

[0065] Decision coefficient R 2 :

[0066] (13)

[0067] In the above formulas (11), (12), and (13), is the true output value of the data corresponding to the i-th time point, is the predicted output value of the data corresponding to the i-th time point, and N is the total amount of data.

[0068] According to the above method, the corresponding models are calculated as follows:

[0069] (14)

[0070] Among them, H is the total health value (i.e. the health value of the combined model), is the i-th model weight, Predict the health value for the i-th model, where n is the total number of models. Through the above steps, the overall health value of the tofu drying machine is obtained;

[0071] Step (3.2), health assessment based on fuzzy evaluation: Using a fuzzy evaluation method, the health value is divided into four levels: healthy, subhealthy, degraded, and faulty. Based on the constructed health levels, a membership function is established, and the health value range is determined to be 0-1. When the health value is set between 0 and 0.2, the tofu dryer is considered to be in a healthy state; when the health value is between 0.2 and 0.5, the tofu dryer is considered to be in a subhealthy state; when the health value is between 0.5 and 0.8, the tofu dryer is considered to be in a degraded state; and when the health value is between 0.8 and 1.0, the tofu dryer is considered to be in a faulty state. The combined model health value H calculated in step (3.1) is placed in the membership function to obtain the qualitative health status of the tofu dryer as the health assessment result.

[0072] In summary, this invention discloses a tofu dryer health assessment method based on data augmentation and ensemble learning, including data preprocessing and feature analysis, health value calculation, and health assessment. This method effectively addresses the difficulty in effectively assessing the health of a tofu dryer during operation when sample labels are imbalanced and data volume is insufficient.

[0073] The contents not described in detail in the specification of the present invention belong to the prior art known to those skilled in the art.

[0074] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A tofu drying machine health assessment method based on data enhancement and ensemble learning, characterized in that: The steps include: Step (1), preprocessing and feature analysis of the tofu drying machine data; realizing data pseudo-label marking and data screening, calculation of data feature importance, and feature sorting and screening; Step (2), calculating the health value of a single model of the tofu drying machine, by respectively establishing a health value calculation model based on a long short-term memory network and a health value calculation model based on an adversarial autoencoder, thereby achieving data enhancement of the monitoring data; Step (3): Calculate and evaluate the health of the combined model of the tofu drying machine, establish an integrated model for calculating the health value of the tofu drying machine, and thus realize the health evaluation of the tofu drying machine; Step (1) is specifically implemented as follows: Step (1.1), data preprocessing: Data preprocessing includes missing value processing, outlier processing, and preliminary feature processing steps; Among them, missing value processing is to fill the gap with the average value of multiple data before and after the missing data as the value of the missing data; Outlier processing is to check whether there is abnormal data in the original data. The method is to replace the missing data with the average value of multiple data before and after the missing data as the replacement data; The initial feature processing is to merge, add and delete the original features of the tofu drying machine. The "hot air speed setting value" and "hot air speed actual value" are merged into the feature "hot air speed working deviation", "BBHS_area 1 barrel temperature setting value" and "BBHS_area 1 barrel temperature actual value" are merged into "BBHS_area 1 barrel temperature", "BBHS_area 2 barrel temperature setting value" and "BBHS_area 2 barrel temperature actual value" are merged into "BBHS_area 2 barrel temperature". "BBHS_Pressure at discharge hood set value" and "BBHS_Pressure at discharge hood actual value" are merged into "BBHS_Pressure at discharge hood", "BBHS_Final cut tobacco moisture set value" and "BBHS_Final cut tobacco moisture actual value" are merged into "BBHS_Final cut tobacco moisture working deviation", "HT_Steam pressure optimization set value" and "HT_Steam pressure optimization actual value" are merged into "HT_Steam pressure optimization working deviation", "Drying tube wall 2 temperature optimization set value" and "Drying tube wall 2 temperature optimization set value" are merged into "Drying tube wall 2 temperature optimization working deviation", "Drying tube wall 1 temperature optimization set value" and "Drying tube wall 1 temperature optimization set value" are merged into "Drying tube wall 1 temperature optimization working deviation"; delete the feature "Cumulative amount of cut tobacco flow" and perform normalization on the filtered data; Step (1.2), pseudo-label extraction: Select features that are highly correlated with the health of the tofu drying machine, and preliminarily label the monitoring data using a clustering algorithm; then select clusters n_clusters=2 as coarse labels and n_clusters=4 as fine labels, and each cluster is based on the formula: , calculate the degree of deviation from the healthy operating state, where d is the degree of deviation, , The corresponding data of clusters and healthy operation are respectively clustered, and then the corresponding label values ​​are set according to the label category according to the degree of deviation of the cluster center. The coarse label is set to 0, which means healthy, and 1 means fault. The values ​​of the fine label are set to 0, 0.4, 0.7, and 1. Finally, the confidence is set to α = 0.05, that is, the confidence interval is set. The confidence evaluation is performed by the distance between the cluster sample point and the cluster center. The data of each cluster is screened in turn. The monitoring data corresponding to the clustering result after screening is used as the available data and recorded as AD; Step (1.3), feature analysis and feature screening: Use the available data of the tofu drying machine n0 dimension as the model input, and the labels obtained by pseudo-label extraction as the model output, and perform XGBoost model training. Use the trained XGBoost model to calculate and sort the feature importance, and select the top features with higher feature importance as the input features of the health value calculation model; Step (2) is specifically implemented as follows: Step (2.1), calculation of the health value of the bidirectional long short-term memory network model: first, the Bi-LSTM network structure is preset based on empirical knowledge, and the model is pre-trained using the available data AD obtained in step (1.2) to obtain the corresponding results and parameters of the model; then, based on the comparison of the training results with the available data AD, the specific network structure and hyperparameters of the Bi-LSTM network structure are further adjusted, and the obtained four-layer Bi-LSTM network structure and one fully connected layer structure are used as the final model structure. The four-layer Bi-LSTM structures have 32 layers, 16 layers, 16 layers, and 8 layers respectively, and the input feature dimension is 14. Then, the extracted coarse labels and fine labels are used as the health values ​​of the tofu drying machine, and the Bi-LSTM network structure is retrained to establish a corresponding relationship between the tofu drying machine monitoring data and the health value of the tofu drying machine, and the trained model parameters are saved, recorded as M1. Finally, the available data AD is used as the model input, and the health value is calculated based on the trained model parameters M1, recorded as h1; Step (2.2), data enhancement of the adversarial autoencoder model, namely AAE: first, the AAE network structure is preset based on empirical knowledge, and the model is pre-trained using the available data AD obtained in step (1.2) to obtain the corresponding results and parameters of the model. Then, based on the comparison between the training results and the available data AD, the specific network structure and hyperparameters of AAE are further adjusted to obtain the final structure of AAE: a four-layer linear encoder, a four-layer decoder structure, and a four-layer discriminator structure. The input feature dimension is 14. For the AAE model with determined hyperparameters, training is performed to establish the corresponding relationship between the drying machine monitoring data and the drying machine health value, and the trained model parameters are saved. Through the trained AAE model, normal data ND and fault data FD are generated to supplement and enhance the original data. Step (2.3), calculate the health value based on the adversarial autoencoder model: according to the formula , calculate the data reconstruction error e between the normal data ND obtained in step (2.2) and the available data AD, and normalize the reconstruction error to obtain the quantitative value characterizing the health of the tofu drying machine, namely the health value, recorded as h2. Through the gradient descent algorithm, take the 14-dimensional feature data as input and the data reconstruction error as output, perform linear regression on the health value calculation formula, and obtain the quantitative correspondence between the available data AD and the health value, recorded as M2.

2. The method for evaluating the health of a tofu drying machine based on data enhancement and ensemble learning according to claim 1, characterized in that: Step (3) is implemented as follows: Step (3.1), calculation of health value based on integrated model: Based on the model M1 obtained in step (2.1) and the model M2 obtained in step (2.3), the root mean square error MSE, mean absolute error MAE, and decision coefficient R 2 , fuzzy evaluation after judgment error rate J There are 4 evaluation indicators in total, and the evaluation indicator matrix of the health calculation model is determined: , where y ij It represents the evaluation value of the i-th evaluation index of the j-th single model. The weights w1 and w2 corresponding to the models M1 and M2 are calculated respectively by the index correlation weighting method. The stacking method is used to integrate the models. Calculate the health value of the combined model, where H is the health value of the combined model, w j is the j-th model weight, h j Predict the health value for the jth model, where n is the total number of models, n=2, and m is the number of evaluation indicators, m=4; Step (3.2), health assessment based on fuzzy evaluation: using the fuzzy evaluation method, the health value is divided into four levels: healthy, sub-healthy, degraded, and faulty; Based on the constructed health level, a membership function is established to determine the health value range as 0-1. When the health value is set between 0-0.2, the tofu drying machine is judged to be in a healthy state. When the health value is 0.2~0.5, the tofu drying machine is judged to be in a sub-healthy state; When the health value is between 0.5 and 0.8, the tofu drying machine is judged to be in a deteriorated state; When the health value is between 0.8 and 1.0, the tofu drying machine is determined to be in a faulty state. The health value H of the combined model calculated in step (3.1) is put into the membership function to obtain the qualitative health state of the tofu drying machine as the health assessment result.

3. An evaluation system using the tofu drying machine health evaluation method based on data enhancement and ensemble learning according to any one of claims 1-2, characterized in that: include: The tofu drying machine data preprocessing and feature analysis module implements data pseudo-labeling and data screening, calculates the importance of data features, and sorts and screens features; The single-model health value calculation module for the tofu drying machine achieves data enhancement of monitoring data by establishing health value calculation models based on long-short-term memory networks and adversarial autoencoders. The tofu drying machine combined model health calculation and evaluation module establishes an integrated model for tofu drying machine health value calculation, thereby realizing tofu drying machine health evaluation.

Images