Range hood operation state identification and grading early warning method and system based on ensemble learning

Abstract

The invention relates to the field of machine learning, in particular to a range hood operation state recognition and grading early warning method and system based on integrated learning. Comprising the steps of obtaining operation data of a range hood; training the initial integrated learning model to obtain a meta classifier; calculating a state confusion degree index of each abnormal state, and determining a training sequence of the special classifier corresponding to each abnormal state; copying the meta classifier as an initial model according to the training sequence, training the initial model, and in the training process, adjusting the meta error classification rate of the abnormal data in the abnormal state and the data weight of the sample data in combination with the meta classifier, and performing iterative training to obtain a special classifier for identifying the abnormal state; obtaining an integrated classifier; and inputting the operation data of the range hood to be identified into the integrated classifier, and outputting a state classification result through the integrated classifier. The method can improve the recognition accuracy of the operation states of the range hood, such as the operation energy efficiency of the range hood and the health state of equipment.

Description

Technical Field

[0001] This invention relates to the field of machine learning, specifically to a method and system for identifying and classifying the operating status of tobacco machinery based on ensemble learning. Background Technology

[0002] Operating status is essentially a qualitative or quantitative description of the dynamic relationship between "production process data" and "energy consumption". When the actual relationship between production process data and energy consumption stably approaches the theoretical optimal performance curve of the equipment, it is a highly efficient and optimized state; when energy consumption shows a systematic and trending increase under the same production process data demand, it is a degraded state; when the stable relationship between production process data and energy consumption is disrupted, and there are drastic fluctuations, sudden changes, or unexplained outliers, it is a pre-failure state.

[0003] In the tobacco industry, with the continuous increase in the production speed of tobacco machinery, the energy consumption demand of tobacco machinery has increased significantly. Accurate identification of its operating status is crucial for energy efficiency management and predictive maintenance. Due to the numerous factors affecting energy consumption (such as operating conditions, load, and equipment health), a single classifier is insufficient to achieve high identification accuracy. Existing technologies typically employ ensemble learning methods to identify different operating states. However, addressing the inherent imbalance in data volume across different tobacco machinery operating states (e.g., data for normal states far exceeds data for various abnormal states), existing methods often use data augmentation to increase the data volume for a few states to achieve balance. However, relying on data augmentation carries the risk of misidentification; the augmented data may not actually occur in the actual production process, leading to insufficient generalization ability of the model during actual deployment and making it difficult to reliably apply in real-world tobacco machinery operation monitoring. Summary of the Invention

[0004] This invention provides a method and system for identifying and classifying the operating status of a tobacco machine based on ensemble learning, in order to solve existing problems.

[0005] The method for identifying and classifying the operating status of a tobacco hood based on ensemble learning in this invention adopts the following technical solution:

[0006] One embodiment of the present invention provides a method for identifying and classifying the operating status of a tobacco hood based on ensemble learning. The method includes the following steps:

[0007] Acquire the operating data of the range hood, wherein the operating data includes normal data of normal state and abnormal data of at least one abnormal state;

[0008] The initial ensemble learning model is trained using normal data as positive samples and abnormal data as negative samples to obtain a meta-classifier that can distinguish between normal and abnormal states.

[0009] Based on the meta-classifier, the state confusion index of each abnormal state is calculated, and the training order of the special classifiers corresponding to each abnormal state is determined according to the state confusion index.

[0010] According to the training order, for each abnormal state, the structure and parameters of the meta-classifier are copied as the initial model. The abnormal data of the abnormal state is used as positive samples, and the data of other states are used as negative samples to train the initial model. During the training process, the data weights of the sample data are adjusted in combination with the meta-classifier's meta-misclassification rate of the abnormal data of the abnormal state. Iterative training is carried out to obtain a specialized classifier for recognizing the abnormal state.

[0011] The meta-classifier and the specialized classifier for each abnormal state are integrated to obtain the ensemble classifier;

[0012] The operating data of the smoke machine to be identified is input into the integrated classifier, which outputs the status classification results and generates graded early warning information based on the status classification results.

[0013] Optionally, the meta-classifier is an ensemble classifier based on the AdaBoost framework, wherein the base learner in the ensemble classifier based on the AdaBoost framework is a gradient boosting decision tree.

[0014] Optionally, based on the meta-classifier, a state confusion index is calculated for each abnormal state, specifically including:

[0015] The test set is input into the meta-classifier, and the discrimination results of all base learners in the meta-classifier for each data sample in the test set are obtained. The normal state in the test set is the same as the normal state in the running data, and the abnormal state types in the test set are the same as the abnormal state types in the running data. Different types of abnormal state data in the test set are distinguished by different labels. The discrimination result is the result indicating whether each data sample in the test set belongs to the normal state or the abnormal state.

[0016] For each abnormal state, a state vector is formed based on the discrimination results of all data samples under that abnormal state to characterize the discrimination features of that abnormal state.

[0017] Calculate the similarity between the state vector of each abnormal state and the state vectors of other abnormal states, and calculate the state confusion index of each abnormal state based on the similarity.

[0018] Optionally, based on the discrimination results of all data samples under this abnormal state, a state vector is formed to characterize the discrimination features of this abnormal state, specifically including:

[0019] For the i-th abnormal state, obtain the discrimination results of all base learners in the meta-classifier for the a-th abnormal data sample;

[0020] The number of base learners that classify the a-th abnormal data sample as normal in the statistical discrimination results is determined as the number of misclassifications.

[0021] The ratio of the number of misclassifications to the number of all base learners in the meta-classifier is determined as the misclassification rate of the a-th abnormal data sample.

[0022] The misclassification rates of each abnormal data sample under the i-th abnormal state are concatenated to form a state vector that characterizes the discrimination features of the i-th abnormal state.

[0023] Optionally, the similarity between the state vector of each abnormal state and the state vectors of other abnormal states is calculated, and a state confusion index for each abnormal state is calculated based on the similarity, specifically including:

[0024] For the state vector of the i-th abnormal state, calculate the cosine similarity between the state vector of the i-th abnormal state and the state vector of each other abnormal state to obtain the similarity between the state vector of the i-th abnormal state and the state vector of each other abnormal state.

[0025] The state difference of the i-th abnormal state is obtained by averaging the similarity between the state vector of the i-th abnormal state and the state vectors of all other abnormal states.

[0026] The result of subtracting the state difference of the i-th abnormal state from 1 is determined as the state confusion index of the i-th abnormal state.

[0027] Optionally, the training order of the specialized classifiers corresponding to each abnormal state is determined based on the state confusion index, specifically including:

[0028] The state confusion index of each abnormal state is sorted in descending order to obtain the sorting result;

[0029] The order of abnormal states corresponding to each element in the sorting results is used as the training order for the specialization classifiers corresponding to the abnormal states.

[0030] Optionally, during training, the data weights of the sample data are adjusted based on the meta-misclassification rate of the abnormal data in the abnormal state, specifically including:

[0031] Obtain the specific misclassification rate of the a-th abnormal data sample in the i-th abnormal state in the k-th training batch;

[0032] The misclassification rate of the meta-classifier for the a-th abnormal data sample is defined as the meta-misclassification rate of the meta-classifier for the a-th abnormal data sample in the i-th abnormal state.

[0033] Based on the specific misclassification rate and the meta-misclassification rate of the a-th abnormal data sample, calculate the data weight of the a-th abnormal data sample in the i-th abnormal state in the k+1-th training batch.

[0034] Obtain the data weight of each abnormal data sample in the i-th abnormal state in the (k+1)-th training batch, and obtain the data weight of the positive sample in the (k+1)-th training batch.

[0035] Based on the data weight of each abnormal data sample in the i-th abnormal state in the (k+1)-th training batch, the data weight of the negative sample is normalized to obtain the data weight of the negative sample in the (k+1)-th training batch. The sum of the data weight of the abnormal data sample and the data weight of the negative sample is 1, and the data weight of each negative sample is the same.

[0036] Normalize the data weights of all data samples in the (k+1)th training batch, and set the sum of the data weights of all data samples to 1;

[0037] Use the positive and negative samples from the (k+1)th training batch to train the (k+1)th training batch;

[0038] Among them, the adjustment range of data weights is positively correlated with the magnitude of the specific misclassification rate and the magnitude of the meta-misclassification rate.

[0039] Optionally, the specific misclassification rate of the a-th abnormal data sample in the i-th abnormal state in the k-th training batch is obtained, specifically including:

[0040] For the a-th abnormal data sample in the i-th abnormal state in the k-th training batch, obtain the discrimination results of all special base learners in the special classifier of the i-th abnormal state for the a-th abnormal data sample;

[0041] The number of specific base learners that classify the a-th abnormal data sample as normal in the statistical discrimination results is determined as the number of specific misclassifications.

[0042] The ratio of the number of misclassified special cases to the number of all special case base learners in the special case classifier of the i-th abnormal state is determined as the special case misclassification rate of the a-th abnormal data sample.

[0043] Optionally, the meta-classifier and the specialized classifier for each abnormal state are integrated, specifically including:

[0044] The outputs of each classifier are weighted and fused according to their respective classifier weights to obtain the output of the ensemble classifier.

[0045] The classifiers include meta-classifiers and specialized classifiers, and their weights are positively correlated with the classification accuracy of each classifier on the validation set.

[0046] This invention proposes a tobacco hood operation status identification and hierarchical early warning system based on ensemble learning, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the tobacco hood operation status identification and hierarchical early warning method based on ensemble learning.

[0047] The beneficial effects of the technical solution of the present invention are:

[0048] In this embodiment of the invention, by combining different operating states of the tobacco machine, the natural dominance of normal state data in industrial production is fully utilized. First, a meta-classifier capable of accurately representing the system's baseline operating state is trained based on all historical data. Building upon this, the traditional approach of constructing various abnormal state samples through data augmentation is abandoned. Instead, multiple base classifiers (i.e., specialized classifiers) with specialized capabilities are derived by fine-tuning the meta-classifier and reweighting feature attention. Each derived base classifier, while inheriting the meta-classifier's understanding of the overall system operation, specifically enhances its sensitivity and discrimination ability for a particular abnormal operating state (such as equipment degradation, control mismatch, etc.). Finally, by integrating these base classifiers originating from the same baseline but with different focuses, a final integrated classifier is constructed, thereby achieving accurate and robust identification of various operating states of the tobacco machine. This design achieves a high degree of unity between data distribution characteristics, model learning process, and actual physical laws in industrial settings. It not only significantly improves the model's performance under real imbalanced data, but also enhances the interpretability and engineering credibility of the identification results, laying the foundation for the reliable implementation and long-term stable operation of the system. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A flowchart illustrating a method for identifying and classifying the operating status of a tobacco machine based on ensemble learning, provided in an embodiment of the present invention;

[0051] Figure 2 This is a structural diagram of a tobacco machine operation status identification and hierarchical early warning system based on ensemble learning, provided in one embodiment of the present invention. Detailed Implementation

[0052] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the integrated learning-based method for identifying and classifying the operating status of a tobacco machine according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0053] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0054] The following description, in conjunction with the accompanying drawings, details the specific scheme of the tobacco machine operation status identification and hierarchical early warning method provided by the present invention.

[0055] This invention provides a method and system for identifying and classifying the operating status of tobacco machinery based on ensemble learning. Please refer to [link / reference]. Figure 1 The diagram illustrates a flowchart of a method for identifying and classifying the operating status of a tobacco machine based on ensemble learning, according to an embodiment of the present invention. The method includes the following steps:

[0056] S101. Obtain the operating data of the range hood, wherein the operating data includes normal data of normal state and abnormal data of at least one abnormal state.

[0057] For example, the operating states of a range hood mainly include the following:

[0058] In a highly efficient and optimized state, the characteristics of its operating data are: the range hood is in production, and the gas consumption / vehicle speed, pressure, and flow rate are within a reasonable range.

[0059] In the no-load standby state, the characteristics of its operating data are as follows: the range hood is in standby mode, the pressure is normal, and the gas consumption and flow rate are relatively reduced.

[0060] When the machine is stopped, its operating data is characterized by the following: the gas hood is stopped, and the gas consumption and flow rate are extremely low (approaching 0).

[0061] Under high-demand conditions, the operating data is characterized as follows: When the range hood is in production, gas consumption, speed, and flow rate increase, while the gas consumption / speed ratio and pressure remain within a reasonable range; when the range hood is in production, gas consumption, speed, and flow rate increase, while the gas consumption / speed ratio is relatively high, and the pressure is normal or low; when the range hood is in production, speed increases, gas consumption and flow rate decrease, pressure is normal or high, and the gas consumption / speed ratio is relatively low.

[0062] The equipment is in a degraded state, characterized by the following operational data: the smoke machine is in production, with increased gas consumption and flow rate, and normal speed and pressure; the smoke machine is in production, with increased gas consumption and flow rate, and normal speed and pressure. This state lasts for more than 5 hours.

[0063] When the control mismatch is in effect, the characteristics of its operating data are as follows: when the range hood is in production, the gas consumption and flow rate are higher than the reasonable range; when the range hood is in standby mode, the gas consumption and flow rate are higher than the reasonable standby range; when the range hood is in shutdown mode, the gas consumption and flow rate are significantly higher than 0.

[0064] The abnormal / fault precursor state is characterized by the following features in the operating data: the smoke machine is in production, and the gas consumption, pressure, and flow rate are significantly abnormal.

[0065] The same state may correspond to multiple characteristics of operational data.

[0066] To construct the training dataset, each record in the historical running data is manually identified and classified into the corresponding running state, thereby obtaining a sample data set for each state.

[0067] Each sample data point contains production process data reflecting the operating status of the smoke machine, including but not limited to: outlet pressure, instantaneous flow rate, supply pressure, temperature, relative humidity, gas consumption, and machine speed (or rotation speed).

[0068] After obtaining the training dataset, the efficient optimization state is marked as 1, indicating a normal state, and its running data is recorded as normal data. All states other than the efficient optimization state are marked as 0, indicating abnormal states, and their running data is recorded as abnormal data.

[0069] S102. Train the initial ensemble learning model using normal data as positive samples and abnormal data as negative samples to obtain a meta-classifier that can distinguish between normal and abnormal states.

[0070] In this embodiment, the meta-classifier is an ensemble classifier based on the AdaBoost framework, and the base learner in the ensemble classifier based on the AdaBoost framework is a gradient boosting decision tree.

[0071] For example, to address the problem of imbalanced data volume across different operational states, this embodiment provides an improved ensemble learning structure to overcome the resulting model identification bias, rather than relying on traditional data augmentation methods. Traditional methods balance the dataset by incrementally processing data from a few states, but the generated data may not actually exist in production, leading to a decrease in the model's generalization ability during actual deployment.

[0072] The solution proposed in this embodiment is to directly improve the internal structure of the ensemble learning model so that it can adapt to the naturally imbalanced distribution of data. Specifically, considering that the efficient optimization state (i.e., the normal state) has the highest probability of occurrence and the largest amount of data among all operating states, this embodiment first uses all historical data to train a base classifier with a strong ability to identify this state, and establishes it as the meta-classifier. This meta-classifier can accurately represent the baseline operating state of the system.

[0073] Subsequently, instead of training other classifiers from scratch, multiple specialized classifiers were derived based on the meta-classifier through targeted fine-tuning and weight adjustments. Each specialized classifier, while inheriting the meta-classifier's overall understanding of the system, specifically strengthens its ability to identify and discriminate a particular abnormal operating state (such as equipment degradation, control mismatch, etc.).

[0074] Finally, the meta-classifier is integrated with all specialized classifiers through weighted voting or stacking fusion to form the final ensemble classifier. This "from benchmark to specialized, then ensemble" architecture fundamentally avoids the bias problem caused by directly training multiple classifiers due to imbalanced data.

[0075] First, data preparation and labeling are performed. Data from historical data labeled as efficient and optimized states (i.e., normal states) are used as positive samples, while data from all other abnormal operating states (such as high-demand processes, equipment degradation, etc.) are mixed together as negative samples. This constitutes the binary classification training dataset used to train the meta-classifier.

[0076] Next, the model is configured and trained. This invention preferably uses the AdaBoost ensemble learning framework, with gradient boosting decision trees as its base learners. Each GBDT model contains multiple decision trees that work together to complete the classification task. The prepared training data (features and their corresponding positive / negative sample labels) is then input into the GBDT model for training.

[0077] In terms of model structure, since the specialized classifier is obtained by copying the structure of the meta-classifier and fine-tuning it, the number of base learners of the specialized classifier is kept consistent with that of the meta-classifier during initialization.

[0078] The number of base learners is a crucial hyperparameter, and its optimal value needs to be determined experimentally based on the specific data scale and feature complexity. A typical method for determining this is to set a series of different base learner numbers during the model development phase, train the model separately, and record the final training accuracy (e.g., classification accuracy on the validation set) and training speed for each. Analysis of the generated "number-accuracy" curve reveals that as the number of base learners increases, model accuracy typically shows a rapid initial increase followed by a gradual plateau. When the number increases to a certain critical value, the improvement in accuracy becomes negligible, and may even decrease due to overfitting.

[0079] Therefore, the number of base learners corresponding to the inflection point where the accuracy growth on the quantity-precision curve flattens out can be selected as the final number of base learners for the meta-classifier and subsequent specialty classifiers. This method balances model performance with model complexity and training efficiency.

[0080] During training, the sample weights and model parameters are continuously adjusted through multiple iterations of the AdaBoost algorithm, enabling the model to highly sensitively and accurately distinguish between positive samples (normal states) and negative samples (abnormal states). After training, the resulting model is the meta-classifier. This meta-classifier ensures correct identification of the vast majority of normal state samples, laying a solid foundation of understanding the system's baseline operating state for subsequent derived specialized classifiers.

[0081] The trained meta-classifier can effectively distinguish between efficient optimized states (normal states) and other abnormal states. In industrial scenarios, the data features of various abnormal states can be regarded as some deviation or derivation from the normal state benchmark.

[0082] Based on this understanding, this embodiment does not train a classifier independently from scratch for each abnormal state. Instead, it uses a robust meta-classifier as a "base model" or "prior model". For each abnormal state that needs to be identified (such as equipment degradation), the structure and parameters of the meta-classifier are used as the initialization starting point. By making targeted fine-tuning on this basis, a specialized classifier focused on identifying that specific state is obtained.

[0083] Because the meta-classifier is trained using predominantly normal state data as positive samples, it exhibits high recall and accuracy in identifying normal states (e.g., it can correctly identify 98 out of 100 normal state samples). This ensures its reliability and stability as the common foundation for all subsequent specialized classifiers.

[0084] S103. Based on the meta-classifier, calculate the state confusion index for each abnormal state, and determine the training order of the specialized classifiers corresponding to each abnormal state according to the state confusion index.

[0085] In this embodiment, based on the meta-classifier, the state confusion index for each abnormal state is calculated, specifically including:

[0086] The test set is input into the meta-classifier, and the discrimination results of all base learners in the meta-classifier for each data sample in the test set are obtained. The normal state in the test set is the same as the normal state in the running data, and the abnormal state types in the test set are the same as the abnormal state types in the running data. Different types of abnormal state data in the test set are distinguished by different labels. The discrimination result is the result indicating whether each data sample in the test set belongs to the normal state or the abnormal state.

[0087] For each abnormal state, a state vector is formed based on the discrimination results of all data samples under that abnormal state to characterize the discrimination features of that abnormal state.

[0088] Calculate the similarity between the state vector of each abnormal state and the state vectors of other abnormal states, and calculate the state confusion index of each abnormal state based on the similarity.

[0089] Based on the discrimination results of all data samples under this abnormal state, a state vector is formed to characterize the discrimination features of this abnormal state, specifically including:

[0090] For the i-th abnormal state, obtain the discrimination results of all base learners in the meta-classifier for the a-th abnormal data sample;

[0091] The number of base learners that classify the a-th abnormal data sample as normal in the statistical discrimination results is determined as the number of misclassifications.

[0092] The ratio of the number of misclassifications to the number of all base learners in the meta-classifier is determined as the misclassification rate of the a-th abnormal data sample.

[0093] The misclassification rates of each abnormal data sample under the i-th abnormal state are concatenated to form a state vector that characterizes the discrimination features of the i-th abnormal state.

[0094] Calculate the similarity between the state vector of each abnormal state and the state vectors of other abnormal states, and calculate the state confusion index for each abnormal state based on the similarity, specifically including:

[0095] For the state vector of the i-th abnormal state, calculate the cosine similarity between the state vector of the i-th abnormal state and the state vector of each other abnormal state to obtain the similarity between the state vector of the i-th abnormal state and the state vector of each other abnormal state.

[0096] The state difference of the i-th abnormal state is obtained by averaging the similarity between the state vector of the i-th abnormal state and the state vectors of all other abnormal states.

[0097] The result of subtracting the state difference of the i-th abnormal state from 1 is determined as the state confusion index of the i-th abnormal state.

[0098] The training order of the specialized classifiers corresponding to each abnormal state is determined based on the state confusion index, specifically including:

[0099] The state confusion index of each abnormal state is sorted in descending order to obtain the sorting result;

[0100] The order of abnormal states corresponding to each element in the sorting results is used as the training order for the specialization classifiers corresponding to the abnormal states.

[0101] For example, each specialized classifier is not trained from scratch, but rather fine-tuned based on a pre-trained meta-classifier. The core capability of the meta-classifier lies in distinguishing between normal states and a set of "abnormal" states. The training objective of the specialized classifier is to fine-tune this macroscopic decision boundary, making it highly sensitive to a specific abnormal state (e.g., specifically for "equipment degradation").

[0102] To achieve orderly and efficient training, it is necessary to quantify the degree of confusion between different abnormal states to determine the training order. Therefore, a key concept is introduced: the state confusion index. Since the meta-classifier is a binary classification model (output 0 represents abnormal, 1 represents normal), it cannot directly give the probability of a sample belonging to "which specific abnormal state". Therefore, the confusion between states cannot be calculated using the traditional classification error rate.

[0103] Therefore, the specific process for calculating the state confusion index for each abnormal state in this embodiment can be as follows:

[0104] Obtain the base learner's discrimination results: Input the labeled test set (whose state structure is consistent with the training data) into the trained meta-classifier. Record the discrimination result of each base learner (i.e., each decision tree) in the meta-classifier for each data sample in the test set, with the result being "1" (indicating a normal state) or "0" (indicating an abnormal state).

[0105] Forming the State Vector: For each abnormal state to be evaluated (e.g., "equipment degradation state"): Take all test samples belonging to that state. For each sample in that state, calculate its misclassification rate, i.e., divide the number of base learners in the meta-classifier that classify the sample as "1" (normal) by the total number of base learners. Arrange the misclassification rates of all samples in that state in a fixed order to form a vector. This vector is the state vector representing the discriminative feature of that state. It reflects the tendency of the decision-making collective within the meta-classifier to "misclassify" the sample in that state as a normal state.

[0106] Calculating the State Confusion Index: Calculating Inter-State Similarity: Select the state vectors of two different abnormal states and calculate their cosine similarity. The higher the value, the more similar the meta-classifier views the two states, and the harder it is to distinguish them. Calculating Average Confusion: For a specific state, calculate the average cosine similarity between its state vector and the state vectors of all other abnormal states. This average reflects the average similarity between this state and other states. Generating the Final Index: Subtract the above average from 1; the result is defined as the state confusion index for that state. The smaller the index value, the easier it is to confuse this state with other states (lower distinguishability); the larger the value, the more unique the state's features (higher distinguishability).

[0107] To ensure efficient and stable training, the training of the specialized classifiers is not arbitrary, but follows a scientific sequence. Specifically, the training order of the specialized classifiers for each abnormal state is determined by ranking the state confusion index values ​​from largest to smallest.

[0108] This sequence means that specialized classifiers for abnormal states with higher state confusion index values ​​(i.e., higher distinguishability from other states and more unique features) should be trained first; then specialized classifiers for abnormal states with lower state confusion index values ​​(i.e., easier to confuse with other states) should be trained sequentially. This "easy-to-difficult" strategy helps to quickly establish a stable discrimination benchmark in the early stages of training.

[0109] Optionally, regarding the order of establishing the specialized classifiers, since the training of each specialized classifier is unrelated and there is no dependency in the training, the specialized classifiers can also be trained in descending order of state confusion index value. There is no specific restriction here, and it can be determined according to the actual computing resources, and the actual training can be carried out in the training order.

[0110] S104. Following the training order, for each abnormal state, copy the structure and parameters of the meta-classifier as the initial model. Use the abnormal data of that abnormal state as positive samples and the data of other states as negative samples to train the initial model. During the training process, combine the meta-classifier with the meta-misclassification rate of the abnormal data of that abnormal state to adjust the data weights of the sample data and perform iterative training to obtain a specialized classifier for recognizing that abnormal state.

[0111] In this embodiment, during the training process, the data weights of the sample data are adjusted based on the meta-misclassification rate of the abnormal data in the abnormal state, specifically including:

[0112] Obtain the specific misclassification rate of the a-th abnormal data sample in the i-th abnormal state in the k-th training batch;

[0113] The misclassification rate of the meta-classifier for the a-th abnormal data sample is defined as the meta-misclassification rate of the meta-classifier for the a-th abnormal data sample in the i-th abnormal state.

[0114] Based on the specific misclassification rate and the meta-misclassification rate of the a-th abnormal data sample, calculate the data weight of the a-th abnormal data sample in the i-th abnormal state in the k+1-th training batch.

[0115] Obtain the data weight of each abnormal data sample in the i-th abnormal state in the (k+1)-th training batch, and obtain the data weight of the positive sample in the (k+1)-th training batch.

[0116] Based on the data weight of each abnormal data sample in the i-th abnormal state in the (k+1)-th training batch, the data weight of the negative sample is normalized to obtain the data weight of the negative sample in the (k+1)-th training batch. The sum of the data weight of the abnormal data sample and the data weight of the negative sample is 1, and the data weight of each negative sample is the same.

[0117] Normalize the data weights of all data samples in the (k+1)th training batch, and set the sum of the data weights of all data samples to 1;

[0118] Use the positive and negative samples from the (k+1)th training batch to train the (k+1)th training batch;

[0119] Among them, the adjustment range of data weights is positively correlated with the magnitude of the specific misclassification rate and the magnitude of the meta-misclassification rate.

[0120] Obtain the specific misclassification rate of the a-th abnormal data sample in the i-th abnormal state in the k-th training batch, specifically including:

[0121] For the a-th abnormal data sample in the i-th abnormal state in the k-th training batch, obtain the discrimination results of all special base learners in the special classifier of the i-th abnormal state for the a-th abnormal data sample;

[0122] The number of specific base learners that classify the a-th abnormal data sample as normal in the statistical discrimination results is determined as the number of specific misclassifications.

[0123] The ratio of the number of misclassified special cases to the number of all special case base learners in the special case classifier of the i-th abnormal state is determined as the special case misclassification rate of the a-th abnormal data sample.

[0124] For example, the training process begins with model initialization when training a specialized classifier for the first abnormal state (here, "state A").

[0125] The model structure and parameters of the pre-trained meta-classifier are completely copied and used as the initial model for the "State A Specialized Classifier" to be trained. At this point, the behavior of the initial model is almost identical to that of the meta-classifier, but its task objective has been redefined: to identify data of "State A" as positive samples, and to identify data of all other states (including normal states and other abnormal states) as negative samples.

[0126] After initialization, the decision boundary of this model needs to be fine-tuned through training from the macroscopic boundary of the meta-classifier to the local boundary that adapts to the new task.

[0127] The decision boundary of a meta-classifier aims to distinguish between "normal states" and "the set of abnormal states." It can be understood as a macro-level boundary that divides the normal and abnormal regions.

[0128] The target decision boundary for the State A specialized classifier needs to form two key boundaries. The first is the boundary between "State A and normal states" that is inherited and maintained; the second is the new boundary between "State A and other abnormal states" that needs to be learned from scratch.

[0129] Therefore, the core objective of training is to drive the model's decision boundary from its initial macroscopic form to a local fine boundary that can accurately define "state A" and all other states (especially other abnormal states).

[0130] To achieve the aforementioned boundary adjustments, the sample weight update strategy needs to be highly targeted during subsequent iterative training. The model needs to pay special attention to samples that are misclassified by the initial model, especially: samples that belong to "state A" but are misclassified as other states (improving the model's ability to capture the core features of state A); and samples that belong to "other abnormal states" but are misclassified as "state A" (strengthening the model's ability to distinguish state A from other anomalies).

[0131] By dynamically adjusting the weights of these key samples, the model is guided to optimize its decision boundary, ultimately obtaining a specialized classifier that can accurately identify "state A".

[0132] For the *a*th abnormal data sample in the target state (taking state A as an example), after one round of training, count the number of decision trees that misclassify it within the current specialty classifier. Divide this number by the total number of decision trees in the specialty classifier to obtain the specialty misclassification rate for that sample. The higher this value, the weaker the current model's ability to recognize that sample.

[0133] For the *a*th abnormal data sample, input it into the meta-classifier and count the number of decision trees that classify it as normal. Divide this number by the total number of decision trees in the meta-classifier to obtain the meta-misclassification rate of the sample. The higher this value, the more the meta-classifier (as the system's baseline perception) tends to consider the sample "like a normal state," meaning the meta-classifier has a lower degree of certainty that the sample belongs to the "abnormal" attribute, and the more ambiguous the classification is.

[0134] Based on the specific misclassification rate and the meta-misclassification rate of the a-th abnormal data sample, the data weight of the a-th abnormal data sample in the i-th abnormal state in the k+1-th training batch is calculated. In a specific embodiment, this can be:

[0135] For the *a*th abnormal data sample, calculate the arithmetic mean of its specific misclassification rate and its meta-misclassification rate to obtain a comprehensive difficulty index *b*. The smaller the value of *b*, the more confident the current specific classifier is in its judgment (few misclassifications), and the more clearly the meta-classifier also considers it to be in an abnormal state (clear discrimination). Conversely, the larger the value of *b*, the more "difficult" the sample is in a dual sense: the current model struggles to identify it, and the meta-classifier's discrimination is also ambiguous.

[0136] Based on the principle that "the adjustment magnitude of data weights is positively correlated with both the specific misclassification rate and the meta-misclassification rate," difficult samples (with large b-values) should receive higher attention. Therefore, the weight adjustment magnitude of such samples should be positively correlated with the b-value. One specific implementation is to multiply their current weights by a b-based amplification factor (e.g., (1+b) or other monotonically increasing functions), then normalize the weights of all training samples in this round to ensure the stability of the weight distribution, and then use them for the next round of iterative training.

[0137] Optionally, after obtaining a comprehensive difficulty index b, a monotonically increasing function can be designed to generate the weight amplification coefficient. Specifically, in this embodiment, the following formula can be used for calculation:

[0138]

[0139] in, For an adjustable hyperparameter (e.g.) =0.5), used to control the overall strength of weight updates. This coefficient ensures that hard samples receive a larger weight increase.

[0140] Multiply the current weight by the amplification factor to obtain the initial unnormalized weight d of the a-th abnormal data sample in the i-th abnormal state in the (k+1)-th training batch:

[0141]

[0142] in, The weight of the a-th abnormal data sample in the i-th abnormal state in the (k+1)-th training batch.

[0143] The data weights of each abnormal data sample in the i-th abnormal state in the (k+1)-th training batch are calculated using the above formula. Based on this, the total weight of the negative samples is determined and evenly distributed to each negative sample. The weights of all samples (updated positive samples + equally weighted negative samples) are globally normalized so that the sum of the data weights of all data samples in the (k+1)-th training batch is 1. Finally, the normalized weights that can be used in the (k+1)-th training batch are obtained.

[0144] After updating and normalizing the weights of the current batch of samples, this new weight distribution is used for the next round of iterative training. This cycle of "calculating weights - training the model" continues, and the decision boundary of the model is continuously fine-tuned and optimized.

[0145] When the model's performance on the validation set stabilizes or reaches a preset stopping condition (such as reaching the maximum number of iterations or the error rate being below a threshold), the specialized classifier for the current target state (such as state A) is considered to have completed training. At this point, a specialized classifier capable of accurately identifying this specific abnormal state has been obtained.

[0146] Subsequently, the system selects the next abnormal state to be trained according to the predetermined training order (i.e., the state confusion index in ascending order), and repeats the above complete process:

[0147] Initialization: Duplicate the meta-classifier as the initial model for the new specialty classifier.

[0148] Data labeling: Data from the new target state is used as positive samples, and all other state data is used as negative samples.

[0149] Iterative training: Combining the meta-misclassification rate information provided by the meta-classifier, the sample weights are dynamically adjusted and iterative fine-tuning is performed.

[0150] Complete training: until the model performance converges, and obtain the specialized classifier for that state.

[0151] S105. Integrate the meta-classifier and the specialized classifier for each abnormal state to obtain the integrated classifier.

[0152] In this embodiment, the meta-classifier and the specialized classifier for each abnormal state are integrated, specifically including:

[0153] The outputs of each classifier are weighted and fused according to their respective classifier weights to obtain the output of the ensemble classifier.

[0154] The classifiers include meta-classifiers and specialized classifiers, and their weights are positively correlated with the classification accuracy of each classifier on the validation set.

[0155] For example, after the meta-classifier and all specialized classifiers have been trained, they together constitute the set of base classifiers used for the final decision. Subsequently, these specialized classifiers need to be integrated into a unified, stronger ensemble classifier.

[0156] In this embodiment, the data is logically and functionally divided into three independent parts: runtime data, test set, and verification set.

[0157] Running dataset: This is the core data used to directly train the model parameters. It includes: Normal data: Specifically refers to all data samples in the running dataset that are marked as being in an efficient, optimized state. Abnormal data: Specifically refers to all data samples in the running dataset that are marked as being in an abnormal state other than the efficient, optimized state. Special note: In this embodiment, unless otherwise specified, references to normal data, abnormal data, normal data samples, and abnormal data samples all refer specifically to the corresponding data in the running dataset.

[0158] Test set: Used for model analysis and evaluation, it does not participate in any model parameter training. Its data samples are independent of the running data. The state categories it contains (normal states and various abnormal states) are consistent with the running data, but the specific data samples are different. It is specifically used to calculate the state confusion index to determine the training order of the specialized classifiers.

[0159] Validation set: Used for model hyperparameter tuning and final performance evaluation. Its data samples are independent of the running data and test set. It is specifically used to evaluate the performance of each trained classifier (meta-classifier and each specialty classifier) ​​to determine their classifier weights in the final ensemble classifier.

[0160] The core training and building process is as follows:

[0161] Meta-classifier training phase: Only the running data is used. Normal data is used as positive samples, and the mixture of all abnormal data is used as negative samples to train a binary meta-classifier.

[0162] Training order determination phase: Only the test set is used. The test set is input into the trained meta-classifier to obtain the discrimination results of its internal base learners. Based on these results, a state confusion index is calculated for each abnormal state. The training order of each specialty classifier is determined by sorting the classifiers from largest to smallest based on this index value.

[0163] Specialized classifier training phase: The running data is reused (re-labeled according to the new task). Following the determined training order, specialized classifiers for each specific abnormal state are trained sequentially. For the current target state, its data in the running data is used as positive samples, and the data from all other states in the running data are used as negative samples. The structure and parameters of the meta-classifier are copied as the initial model for the current specialized classifier, and iterative fine-tuning is performed based on this.

[0164] Ensemble classifier construction phase: Use only the validation set. Evaluate the classification accuracy of the meta-classifier and all trained specialty classifiers on the validation set. Based on the evaluation results, assign classifier weights to each classifier (weights are positively correlated with accuracy). Bind all weights to their corresponding classifiers to construct the final ensemble classifier.

[0165] Determining the weights of base classifiers: The ensemble employs a weighted fusion strategy. The classifier weights of each base classifier (including the meta-classifier and each specialty classifier) ​​are determined based on their respective classification performance. Specifically:

[0166] The performance of each base classifier is evaluated on an independent validation set, typically using classification accuracy as the core metric. The weights of base classifiers are positively correlated with their classification accuracy on the validation set. That is, the higher the classification accuracy of a base classifier, the greater its weight is assigned in the ensemble decision, and the greater its impact on the final result. Binding all base classifiers and their corresponding classifier weights constitutes the final ensemble classifier.

[0167] S106. Input the operating data of the smoke machine to be identified into the integrated classifier, output the status classification result through the integrated classifier, and generate hierarchical early warning information based on the status classification result.

[0168] For example, in practical applications, the real-time production process data of the smoke machine (which comprehensively reflects the operating status of the smoke machine and mainly includes production process data (such as outlet pressure, instantaneous flow rate, supply pressure, temperature, etc.) and energy consumption parameters (such as power, current, voltage, etc.)) can be input into the integrated classifier to obtain the identification result of the current operating status of the smoke machine, that is, the status classification result.

[0169] Based on the above identification results, the system automatically triggers corresponding tiered warnings to guide on-site operation and maintenance. Warning levels can be defined as follows:

[0170] Blue (Alert Level): The system is in an acceptable non-optimal state, or experiencing reasonable temporary high energy consumption. This level only requires attention and no immediate action is needed.

[0171] Yellow (Optimized Level): The system energy efficiency has shown a measurable decline, and there is a clear need for optimization or planned maintenance. It is recommended to arrange for its handling within the plan.

[0172] Red (Emergency Level): The equipment is in an abnormal or pre-fault state that may immediately affect safety or cause damage, requiring immediate intervention.

[0173] The specific warning triggering logic and content examples are as follows:

[0174] When identified as "high-efficiency optimized state": Sample characteristics: The range hood is in production, and the gas consumption / vehicle speed, pressure, and flow rate are within a reasonable range. Warning content: No warning; system feedback: "Current operation is at optimal energy efficiency."

[0175] When identified as a "high-demand process state": Blue warning sample characteristics: The range hood is in production, gas consumption, speed, and flow rate are increased, while the gas consumption / speed ratio and pressure are within a reasonable range. Blue warning content: Automatically triggered, prompting: "Currently in a high-load production period, increased energy consumption is normal." Yellow warning sample characteristics: The range hood is in production, gas consumption, speed, and flow rate are increased, with the gas consumption / speed ratio relatively high, and pressure normal or low; or the speed is increased, gas consumption and flow rate are decreased, pressure is normal or high, and gas consumption / speed is relatively low. Yellow warning content: Upgraded to a yellow warning, prompting: "Energy efficiency is low during high-load production; it is recommended to check process coordination" or "The range hood's gas path may be blocked; it is recommended to check and clean it."

[0176] When identified as "equipment degradation state": Yellow warning sample characteristics: The range hood is in production, gas consumption and flow rate are increased, while speed and pressure are normal. Yellow warning content: Automatically triggered, prompting: "Range hood energy efficiency has decreased; it is recommended to arrange for gas line leak investigation." Red warning sample characteristics: The range hood is in production, gas consumption and flow rate are increased, while speed and pressure are normal, and this state lasts for more than 5 hours. Red warning content: Upgraded to a red warning, prompting: "Range hood energy efficiency continues to decrease; immediately conduct gas line leak investigation."

[0177] When identified as "control mismatch state": Yellow warning sample characteristics: The smoke machine is in production state, and the gas consumption and flow rate are higher than the reasonable range; or the smoke machine is in standby state, and the gas consumption and flow rate are higher than the reasonable standby range; or the smoke machine is in shutdown state, and the gas consumption and flow rate are significantly higher than 0. Yellow warning content: Automatically triggered, prompting: "Gas consumption is not economically matched with production speed, it is recommended to adjust the gas supply settings", "Gas consumption of the smoke machine in standby is not economical, it is recommended to adjust the gas supply settings", or "Gas supply to the smoke machine is not turned off when it is shut down, it is recommended to adjust the gas supply settings".

[0178] When identified as an "abnormal / precursor state": Red alert sample characteristics: The range hood is in production, and the gas consumption, pressure, and flow rate are significantly abnormal. Red alert content: It will be automatically triggered immediately, accompanied by a light alarm, prompting: "Abnormal gas consumption of the range hood has been detected. Please stop the machine immediately for troubleshooting."

[0179] In summary, in this embodiment of the invention, a meta-classifier capable of accurately representing the baseline operating state of the system is first trained based on all historical data. Then, by fine-tuning this meta-classifier and reweighting the feature attention, multiple base classifiers (i.e., specialized classifiers) with their own specializations are derived. Finally, by integrating these base classifiers, which originate from the same baseline but each have their own focus, a final ensemble classifier is constructed, thereby achieving accurate and robust identification of various operating states of the tobacco machine.

[0180] This invention also proposes a system for identifying and classifying the operating status of tobacco machinery based on ensemble learning. Please refer to [link / reference]. Figure 2 The diagram shows a structural diagram of a tobacco machine operation status identification and hierarchical early warning system based on ensemble learning provided by an embodiment of the present invention. The system includes: a data acquisition module 101, a model training module 102, and an early warning hierarchical module 103.

[0181] The data acquisition module 101 is used to acquire the operating data of the range hood, wherein the operating data includes normal data of normal state and abnormal data of at least one abnormal state;

[0182] The model training module 102 is used to train the initial ensemble learning model with normal data as positive samples and abnormal data as negative samples to obtain a meta-classifier that can distinguish between normal and abnormal states.

[0183] Based on the meta-classifier, the state confusion index of each abnormal state is calculated, and the training order of the special classifiers corresponding to each abnormal state is determined according to the state confusion index.

[0184] Following the training sequence, for each abnormal state, the structure and parameters of the meta-classifier are replicated as the initial model. The abnormal data from that abnormal state are used as positive samples, and data from other states are used as negative samples to train the initial model. During training, the data weights of the sample data are adjusted based on the meta-classifier's misclassification rate for the abnormal data in that abnormal state. Iterative training is then performed to obtain a specialized classifier for identifying that abnormal state.

[0185] The meta-classifier and the specialized classifier for each abnormal state are integrated to obtain the ensemble classifier;

[0186] The early warning classification module 103 is used to input the operating data of the smoke machine to be identified into the integrated classifier, output the status classification result through the integrated classifier, and generate classified early warning information based on the status classification result.

[0187] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the tobacco machine operation status identification and hierarchical early warning system based on ensemble learning and the tobacco machine operation status identification and hierarchical early warning method based on ensemble learning provided in the above embodiments belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0188] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0189] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0190] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying and hierarchically warning about the operating status of a tobacco machine based on ensemble learning, characterized in that, include: Acquire the operating data of the range hood, wherein the operating data includes normal data of normal state and abnormal data of at least one abnormal state; The initial ensemble learning model is trained using normal data as positive samples and abnormal data as negative samples to obtain a meta-classifier that can distinguish between normal and abnormal states. Based on the meta-classifier, the state confusion index of each abnormal state is calculated, and the training order of the special classifiers corresponding to each abnormal state is determined according to the state confusion index. According to the training order, for each abnormal state, the structure and parameters of the meta-classifier are copied as the initial model. The abnormal data of the abnormal state is used as positive samples, and the data of other states are used as negative samples to train the initial model. During the training process, the data weights of the sample data are adjusted in combination with the meta-classifier's meta-misclassification rate of the abnormal data of the abnormal state. Iterative training is carried out to obtain a specialized classifier for recognizing the abnormal state. The meta-classifier and the specialized classifier for each abnormal state are integrated to obtain the ensemble classifier; The operating data of the smoke machine to be identified is input into the integrated classifier, which outputs the status classification results and generates graded early warning information based on the status classification results.

2. The method for identifying and hierarchically warning about the operating status of a tobacco machine based on ensemble learning according to claim 1, characterized in that, The meta-classifier is an ensemble classifier based on the AdaBoost framework, and the base learner in the ensemble classifier based on the AdaBoost framework is a gradient boosting decision tree.

3. The method for identifying and hierarchically warning about the operating status of a tobacco machine based on ensemble learning according to claim 1, characterized in that, The calculation of the state confusion index for each abnormal state based on the meta-classifier specifically includes: The test set is input into the meta-classifier, and the discrimination results of all base learners in the meta-classifier for each data sample in the test set are obtained. The normal state in the test set is the same as the normal state in the running data, and the abnormal state types in the test set are the same as the abnormal state types in the running data. Different types of abnormal state data in the test set are distinguished by different labels. The discrimination result is the result indicating whether each data sample in the test set belongs to the normal state or the abnormal state. For each abnormal state, a state vector is formed based on the discrimination results of all data samples under that abnormal state to characterize the discrimination features of that abnormal state. Calculate the similarity between the state vector of each abnormal state and the state vectors of other abnormal states, and calculate the state confusion index of each abnormal state based on the similarity.

4. The method for identifying and hierarchically warning about the operating status of a tobacco machine based on ensemble learning according to claim 3, characterized in that, The process of forming a state vector based on the discrimination results of all data samples under this abnormal state to characterize the discrimination features of this abnormal state specifically includes: For the i-th abnormal state, obtain the discrimination results of all base learners in the meta-classifier for the a-th abnormal data sample; The number of base learners that classify the a-th abnormal data sample as normal in the statistical discrimination results is determined as the number of misclassifications. The ratio of the number of misclassifications to the number of all base learners in the meta-classifier is determined as the misclassification rate of the a-th abnormal data sample. The misclassification rates of each abnormal data sample under the i-th abnormal state are concatenated to form a state vector that characterizes the discrimination features of the i-th abnormal state.

5. The method for identifying and hierarchically warning about the operating status of a tobacco machine based on ensemble learning according to claim 3, characterized in that, The calculation of the similarity between the state vector of each abnormal state and the state vectors of other abnormal states, and the calculation of the state confusion index of each abnormal state based on the similarity, specifically includes: For the state vector of the i-th abnormal state, calculate the cosine similarity between the state vector of the i-th abnormal state and the state vector of each other abnormal state to obtain the similarity between the state vector of the i-th abnormal state and the state vector of each other abnormal state. The state difference of the i-th abnormal state is obtained by averaging the similarity between the state vector of the i-th abnormal state and the state vectors of all other abnormal states. The result of subtracting the state difference of the i-th abnormal state from 1 is determined as the state confusion index of the i-th abnormal state.

6. The method for identifying and hierarchically warning about the operating status of a tobacco machine based on ensemble learning according to claim 1, characterized in that, The step of determining the training order of the specialized classifiers corresponding to each abnormal state based on the state confusion index specifically includes: The state confusion index of each abnormal state is sorted in descending order to obtain the sorting result; The order of abnormal states corresponding to each element in the sorting results is used as the training order for the specialization classifiers corresponding to the abnormal states.

7. The method for identifying and hierarchically warning about the operating status of a tobacco machine based on ensemble learning according to claim 1, characterized in that, During training, the data weights of the sample data are adjusted based on the meta-misclassification rate of the abnormal data in the abnormal state, according to the meta-classifier. Specifically, this includes: Obtain the specific misclassification rate of the a-th abnormal data sample in the i-th abnormal state in the k-th training batch; The misclassification rate of the meta-classifier for the a-th abnormal data sample is defined as the meta-misclassification rate of the meta-classifier for the a-th abnormal data sample in the i-th abnormal state. Based on the specific misclassification rate and the meta-misclassification rate of the a-th abnormal data sample, calculate the data weight of the a-th abnormal data sample in the i-th abnormal state in the k+1-th training batch. Obtain the data weight of each abnormal data sample in the i-th abnormal state in the (k+1)-th training batch, and obtain the data weight of the positive sample in the (k+1)-th training batch. Based on the data weight of each abnormal data sample in the i-th abnormal state in the (k+1)-th training batch, the data weight of the negative sample is normalized to obtain the data weight of the negative sample in the (k+1)-th training batch. The sum of the data weight of the abnormal data sample and the data weight of the negative sample is 1, and the data weight of each negative sample is the same. Normalize the data weights of all data samples in the (k+1)th training batch, and set the sum of the data weights of all data samples to 1; Use the positive and negative samples from the (k+1)th training batch to train the (k+1)th training batch; Among them, the adjustment range of data weights is positively correlated with the magnitude of the specific misclassification rate and the magnitude of the meta-misclassification rate.

8. The method for identifying and hierarchically warning about the operating status of a tobacco machine based on ensemble learning according to claim 7, characterized in that, The specific misclassification rate of the a-th abnormal data sample in the i-th abnormal state in the k-th training batch includes: For the a-th abnormal data sample in the i-th abnormal state in the k-th training batch, obtain the discrimination results of all special base learners in the special classifier of the i-th abnormal state for the a-th abnormal data sample; The number of specific base learners that classify the a-th abnormal data sample as normal in the statistical discrimination results is determined as the number of specific misclassifications. The ratio of the number of misclassified special cases to the number of all special case base learners in the special case classifier of the i-th abnormal state is determined as the special case misclassification rate of the a-th abnormal data sample.

9. The method for identifying and hierarchically warning about the operating status of a tobacco machine based on ensemble learning according to claim 1, characterized in that, The integration of the meta-classifier and the specialized classifier for each abnormal state specifically includes: The outputs of each classifier are weighted and fused according to their respective classifier weights to obtain the output of the ensemble classifier. The classifiers include meta-classifiers and specialized classifiers, and their weights are positively correlated with the classification accuracy of each classifier on the validation set.

10. A tobacco machine operation status identification and hierarchical early warning system based on ensemble learning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for identifying and classifying the operating status of a tobacco machine based on ensemble learning as described in any one of claims 1-9.

Images