Urban solid waste incineration temperature state qualitative and classification method

Abstract

The invention provides an urban solid waste incineration temperature state qualitative and classification method, which comprises the following steps of: collecting and processing field temperature process data to obtain a temperature sample data matrix with consistent characteristics; performing feature clustering based on multi-distance fusion to obtain a plurality of feature clusters with consistent ranges and trends; aiming at the feature cluster, performing change point detection by utilizing a Hoeffering inequality to obtain multiple feature change points; a Bennett inequality is utilized to purify and calculate the multiple feature change points, and multiple groups of labeled temperature state data are obtained; and constructing a state recognition model to verify the validity of the marked temperature state data. According to the method, high-quality and accurately-labeled temperature state data can be provided, automatic and efficient qualitative and data-driven grouping of the industrial temperature state is realized, a prior distribution hypothesis is not needed, the actual application scene is greatly widened, and the adaptability is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of municipal solid waste incineration analysis, in particular to a municipal solid waste incineration temperature state qualitative and classification method. BACKGROUND

[0002] With the growth of population and the improvement of living standards, the amount of municipal solid waste (MSW) is increasing year by year, and it is estimated that it will reach 256 to 333 million tons in 2050. At present, MSW incineration (MSWI) has gradually become the main way of municipal solid waste treatment due to its characteristics of harmlessness, reduction, and resource utilization. The MSWI process includes stages such as feeding, combustion, heat exchange, purification, and emission. Due to environmental protection requirements and policy regulations, MSWI power plants need to ensure sufficient combustion and stable operation. In order to effectively decompose toxic and harmful substances in flue gas, such as CO, HCl, NOx, SO2, and dioxin (DXN), the furnace temperature needs to be strictly greater than 850℃. Therefore, MSWI power plants install multiple temperature sensors in the furnace according to regulations and their own needs to ensure real-time acquisition of temperature information. However, the actual site operators cannot determine the combustion state in the furnace through numerous temperature information, and can only operate the equipment through experience, which easily causes operation fluctuations and unstable conditions. The research on intelligent monitoring of the combustion state of the MSWI process is urgently needed.

[0003] At present, artificial intelligence technology is gradually deepening its application in the industrial field due to its powerful performance, showing great application potential. Among them, the development of industrial automation and informatization promotes data-driven modeling to become an important research direction of AI technology. The basic idea of data-driven modeling is to extract information from historical data to establish a mathematical model that describes the behavior of the system. Therefore, data-driven modeling methods are widely used in the fields of prediction, diagnosis, and optimization of MSWI processes. However, existing research mainly focuses on the grate position, and lacks attention to the combustion state in the furnace. Moreover, the combustion state lacks a standardized qualitative framework and relies on manual annotation, hindering practical application.

[0004] Therefore, in the data analysis and modeling of the MSWI process, there are the following defects: 1) Traditional methods rely on statistical means such as normal, chi-square, and other parameter models, and also rely on manual grouping, experience-based determination, and individual feature analysis. 2) Data-driven methods are emerging, but they mainly focus on prediction, diagnosis, optimization, and other aspects, lacking new technologies for "intelligent grouping of combustion state itself" and "no distribution assumption". 3) AI methods appear but rely on manual annotation, lack standardization, and the qualitative framework is imperfect. SUMMARY

[0005] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a municipal solid waste incineration temperature state qualitative and classification method, which can provide high-quality, accurate labeled temperature state data, realize industrial temperature state automatic efficient qualitative and data-driven grouping, and does not require prior distribution assumption, greatly widening the actual application scene and enhancing the adaptability.

[0006] To achieve the above object, the present application provides the following scheme: a municipal solid waste incineration temperature state qualitative and classification method, comprising:

[0007] Collecting on-site temperature process data of a municipal solid waste incineration power plant, pre-processing the on-site temperature process data by using a down-sampling strategy to obtain a temperature sample data matrix, and then performing feature measurement on the temperature sample data matrix to ensure feature consistency;

[0008] Based on the temperature sample data matrix, performing feature clustering based on multi-distance fusion to obtain multiple feature clusters with consistent range and trend;

[0009] For the feature clusters, using Hoeffding inequality to detect change points to obtain multiple feature change points;

[0010] Using Bennett inequality to purify the multiple feature change points to calculate the final state change points of each feature cluster to obtain multiple sets of labeled temperature state data;

[0011] Based on decision tree, random forest, support vector machine and neural network, constructing a state recognition model to verify the effectiveness of the labeled temperature state data.

[0012] Optionally, collecting temperature process data samples of a municipal solid waste incineration power plant, pre-processing the temperature process data samples by using a down-sampling strategy to obtain a temperature sample data matrix, and then performing feature measurement on the temperature sample data matrix to ensure feature consistency, comprising:

[0013] Collecting temperature process data samples of a municipal solid waste incineration power plant, for the temperature process data samples, using a deep learning down-sampling strategy to segment the temperature process data samples into multiple segments at a set interval along the sample direction, and then calculating the mean value of each segment to obtain a temperature sample data matrix;

[0014] Based on the temperature sample data matrix, calculating the Euclidean distance between each pair of features to obtain multiple feature Euclidean distance vectors that measure the range consistency between features, and then based on the feature Euclidean distance vectors, using Wasserstein distance to extract trend information that measures the trend consistency of features to ensure feature consistency.

[0015] Optionally, based on the temperature sample data matrix, multi-distance fusion-based feature clustering is performed based on the temperature sample data matrix to obtain a plurality of feature clusters with consistent range and trend, including:

[0016] S1, based on the temperature sample data matrix with consistent features, a Euclidean distance vector is randomly selected as a first initial center point, the Wasserstein distance between other Euclidean distance vectors and the first initial center point is calculated to obtain the selected probability of other Euclidean distance vectors, and the Euclidean distance vector with the maximum selected probability is selected as a second initial center point;

[0017] S2, repeating step S1 until the feature clusters are all selected to obtain a plurality of center points;

[0018] S3, calculating the Wasserstein distance between the Euclidean distance vector and the center point, and distributing the Euclidean distance vector to the feature cluster with the minimum Wasserstein distance;

[0019] S4, after the end of the distribution of the Euclidean distance vector, the mean of the Euclidean distance vectors in the feature cluster is calculated to update the center point;

[0020] S5, repeating step S4 until the Euclidean distance vectors in the feature cluster do not change, and then distributing the Euclidean distance vectors to the corresponding feature clusters to obtain a plurality of feature clusters with consistent range and trend.

[0021] Optionally, for the feature cluster, Hoeffding inequality is used for change point detection to obtain a plurality of feature change points, including:

[0022] For each of the feature clusters, a statistical window is preset, the maximum value, the minimum value, the mean value and the mean error limit of the data in the statistical window are counted, and Hoeffding inequality is used to determine whether the mean value change of the current temperature sample data matrix exceeds the mean error limit. If yes, it is determined as an initial state change point;

[0023] According to the initial state change point, the null hypothesis is set as unchanged and the alternative hypothesis is set as changed, and the state change is determined by using the statistical significance level to count a plurality of change points of each feature in the feature cluster to obtain a plurality of feature change points.

[0024] Optionally, according to the state probability, Bennett inequality is used to purify the plurality of feature change points to calculate the final state change point of each feature cluster to obtain a plurality of labeled temperature state data, including:

[0025] Based on the multi-feature variable point, a state probability variance and window mean limit are calculated in the statistical window, whether the mean exceeds the window mean limit is judged by using Bennett inequality, if yes, the final state change point of the statistical window is judged;

[0026] The interval mean of all the feature cluster final state change points is calculated, and the interval mean is set as the final state division point to divide the statistical window into multiple combustion state sections, and multiple groups of labeled temperature state data are obtained.

[0027] Optionally, according to the state probability, a state recognition model is constructed based on a decision tree, a random forest, a support vector machine and a neural network to verify the effectiveness of the labeled temperature state data, including:

[0028] The number of leaf node samples of the decision tree is set to 1, and the number of feature selections is set to 5;

[0029] The number of forests of the random forest is set to 20, the number of leaf node samples is set to 1, and the number of feature selections is set to 5;

[0030] The kernel function of the support vector machine is set to a linear kernel function, and the kernel size is set to 1;

[0031] The learning rate of the neural network is set to 0.001, the training algebra is set to 2000, and the network structure is set to a four-layer feedforward neural network;

[0032] According to the state probability, the state recognition model is constructed based on the set decision tree, random forest, support vector machine and neural network;

[0033] The temperature state data is divided into three groups of different proportions of samples, the three groups of samples are input into the state recognition model, and the accuracy, precision, recall and F1 score are used to verify the effectiveness of the labeled temperature state data.

[0034] Optionally, the four-layer feedforward neural network includes an input layer composed of 14 neurons, a first hidden layer composed of 50 neurons and a ReLU activation function, a second hidden layer composed of 100 neurons and a ReLU activation function, a third hidden layer composed of 150 neurons and a ReLU activation function, and an output layer composed of 4 neurons and a Softmax activation function.

[0035] The present application provides a kind of urban solid waste incineration temperature state qualitative and classification method, and the following technical effects are disclosed:

[0036] 1. High data quality: By collecting and processing temperature process data, 1) it can retain the main features while reducing the computational complexity, weakening the influence of noise. 2) Compared with traditional single-point or mean processing, down-sampling retains the sequence trend and improves the measurement representativeness. 3) Avoid high-dimensional redundant data directly involved in modeling, improve the overall processing efficiency.

[0037] 2. Efficient feature analysis: Through feature clustering, 1) fusion range and trend consistency clustering not only distinguifies the spatial level (range) of the original temperature data, but also detects similar features (trends), effectively avoiding missed detection and misclassification. 2) Using Wasserstein distance, unlike traditional Euclidean distance clustering or manual grouping, it takes into account the subtle change trend of the feature. 3) The traditional method only relies on physical structure, artificial experience or single distance, and this method can realize standardized automatic clustering, with strong compatibility and good generalization.

[0038] 3. Wide range of practical applications: Through state qualitative analysis, 1) by non-parametric Hoeffding inequality, the dependence on distribution of parameterized model is completely eliminated, and the complex, multi-source and variable actual industrial data is adapted. 2) It can capture state changes in real time and dynamically, and is suitable for strong disturbance actual scene, reducing false alarms.

[0039] 4. Accurate temperature state judgment: 1) By combining multiple feature change points through Bennett inequality and sliding window method, the robustness and accuracy of state judgment can be improved, and accidental misjudgment caused by single feature can be reduced. 2) State probability integration and multi-channel information fusion can significantly increase the credibility of state judgment.

[0040] The technical solutions of the present application will be further described in detail below with the help of the accompanying drawings and examples. DETAILED DESCRIPTION

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0042] Figure 1 The method flowchart provided for embodiment 1 of the present application is shown in the figure;

[0043] Figure 2 The state data acquisition strategy based on non-parametric method provided for embodiment 2 of the present application is shown in the figure;

[0044] Figure 3 The temperature state quantization result provided for embodiment 2 of the present application is shown in the figure. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0046] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] Example 1

[0048] like Figure 1 As shown, the present invention provides a method for characterizing and classifying the temperature state of municipal solid waste incineration, comprising:

[0049] 1. Collect on-site temperature process data from municipal solid waste incineration power plants, pre-process the data using a downsampling strategy to obtain a temperature sample data matrix, and then perform feature measurement on the temperature sample data matrix to ensure feature consistency. Specifically, this includes:

[0050] Temperature process data samples of a municipal solid waste incineration power plant are collected. For the temperature process data samples, a deep learning downsampling strategy is used to segment the temperature process data samples into multiple segments at set intervals along the sample direction. The mean of each segment is then calculated to obtain a temperature sample data matrix.

[0051] Based on the temperature sample data matrix, the Euclidean distance between each pair of features is calculated to obtain multiple feature Euclidean distance vectors that measure the range consistency between features. Then, based on the feature Euclidean distance vectors, the Wasserstein distance is used to extract trend information that measures the consistency of feature trends to ensure feature consistency.

[0052] 2. Based on the temperature sample data matrix, perform feature clustering based on multi-distance fusion to obtain multiple feature clusters with consistent ranges and trends; specifically, the following steps are included:

[0053] S1. Based on the temperature sample data matrix with feature consistency, randomly select a Euclidean distance vector as a first initial center point, calculate the Wasserstein distances between other Euclidean distance vectors and the first initial center point, obtain the selection probabilities of other Euclidean distance vectors, and select the Euclidean distance vector with the largest selection probability as the second initial center point;

[0054] S2, repeating step S1 until all the feature clusters are selected, obtaining a plurality of center points;

[0055] S3, calculating the Wasserstein distance between the Euclidean distance vector and the center point, and assigning the Euclidean distance vector to the feature cluster with the minimum Wasserstein distance;

[0056] S4, after the Euclidean distance vector assignment is completed, calculating the mean of the Euclidean distance vector in the feature cluster to update the center point;

[0057] S5, repeating step S4 until the Euclidean distance vector in the feature cluster does not change, and then assigning the Euclidean distance vector to the corresponding feature cluster, obtaining a plurality of feature clusters with consistent range and trend.

[0058] 3, for the feature cluster, using Hoeffding inequality to detect change points, obtaining multiple feature change points. Specifically including:

[0059] For each of the feature clusters, a statistical window is preset, and the maximum value, minimum value, mean value and mean error limit of the data in the statistical window are counted, and the Hoeffding inequality is used to determine whether the mean value change of the current temperature sample data matrix exceeds the mean error limit, if so, it is determined as an initial state change point.

[0060] According to the initial state change point, the null hypothesis is set as unchanged and the alternative hypothesis is set as changed, and the state change is determined by using the statistical significance level, so as to count a plurality of change points of each feature in the feature cluster, and obtain multiple feature change points.

[0061] 4, using Bennett inequality to purify the multiple feature change points to calculate the final state change point of each feature cluster, obtaining a plurality of labeled temperature state data. Specifically including:

[0062] Based on the multiple feature change points, the state probability variance and window mean limit are calculated in the statistical window, and the Bennett inequality is used to determine whether the mean value exceeds the window mean limit, if so, it is determined as the final state change point of the statistical window.

[0063] The interval mean of all the feature cluster final state change points is calculated, and the interval mean is set as the final state division point, so as to divide the statistical window into a plurality of combustion state sections, obtaining a plurality of labeled temperature state data.

[0064] 5, based on decision tree, random forest, support vector machine and neural network, a state recognition model is constructed to verify the effectiveness of the labeled temperature state data. Specifically including:

[0065] Set the number of leaf node samples of the decision tree to 1 and the number of feature selections to 5;

[0066] Set the number of forests of the random forest to 20, the number of leaf node samples to 1, and the number of feature selections to 5;

[0067] Set the kernel function of the support vector machine to a linear kernel function and the inner kernel size to 1;

[0068] Set the learning rate of the neural network to 0.001, the number of training iterations to 2000, and the network structure to a four-layer feedforward neural network; wherein the four-layer feedforward neural network includes an input layer composed of 14 neurons, a first hidden layer composed of 50 neurons and a ReLU activation function, a second hidden layer composed of 100 neurons and a ReLU activation function, a third hidden layer composed of 150 neurons and a ReLU activation function, and an output layer composed of 4 neurons and a Softmax activation function.

[0069] Incremental hidden layer neuron number 50 -> 100 -> 150: can let the network gradually abstract from low-level features to high-level features;

[0070] ReLU activation function: high computational efficiency, avoids gradient disappearance, and is suitable for deep networks;

[0071] Softmax output layer: suitable for multi-classification problems, and the output probability distribution is convenient for model prediction and evaluation.

[0072] According to the state probability, based on the set decision tree, random forest, support vector machine and neural network, the state recognition model is constructed.

[0073] The temperature state data is divided into 3 groups of different proportion samples, and the 3 groups of samples are input into the state recognition model, and the accuracy, precision, recall and F1 score are used to verify the effectiveness of the labeled temperature state data.

[0074] Embodiment 2

[0075] As shown in Figure 2 The present application provides a city solid waste incineration temperature state qualitative and classification method, mainly including feature clustering based on multi-distance fusion and state qualitative analysis based on non-parametric detection.

[0076] 1. Feature clustering based on multi-distance fusion

[0077] To obtain reasonable feature clusters, the application proposes two feature clustering criteria of range consistency and trend consistency. Among them, the range consistency ensures that the features in the same group belong to the same range and can reflect the regional characteristics of the features; the trend consistency ensures that the change trends of the features in the same group are the same and can reflect the change characteristics of the features. Based on the two criteria, the multi-distance fusion feature clustering method aims to obtain multiple feature clusters with the same characteristics and provide support for the subsequent acquisition and analysis of state data.

[0078] First, the process data is recorded as N and M are the sample number and feature number respectively. To better extract range and trend information, referring to the downsampling idea of deep learning, the samples along X = [x1, …, xN] are divided into N segments at intervals of θ x , and the data in the segments are averaged to form new samples

[0079] Next, the Euclidean distances between two features are calculated to form the Euclidean distance vector of the features.

[0080] If the Euclidean distance between two features is smaller, it means that the range is closer. Therefore, the Euclidean distance vector can well represent the range information between the features to meet the requirement of range consistency. Then, the trend information is extracted from the Euclidean distance vector using the Wasserstein distance to realize feature clustering. The roulette method is used to determine the initial center points of K feature clusters as follows:

[0081] First, a Euclidean distance vector is randomly selected as the first initial center point c1.

[0082] Next, the Wasserstein distances of other distance vectors to c1 are calculated to obtain the probabilities of other features being selected as the next feature.

[0083] Then, the distance vectors with larger Wasserstein distances to c1 are more likely to be selected, and the roulette method is used to select the second center point c2.

[0084] Finally, the above operation is repeated using c2 until K center points are selected.

[0085] The Wasserstein distances of the M Euclidean distance vectors to the K center points are calculated, and the Euclidean distance vectors are assigned to the cluster with the smallest Wasserstein distance. After the assignment is completed, the Euclidean distance vectors in the K clusters are averaged to update the center points. The above operation is repeated until the distance vectors in the K clusters do not change.

[0086] ​​Finally, the Euclidean distance vector is mapped to the feature and assigned to the corresponding feature cluster.

[0087] 2. State qualitative analysis based on non-parametric detection

[0088] Accurate qualitative analysis of temperature state helps practitioners understand the operation of the incinerator. In addition, high-quality combustion state data is essential for building reliable data-driven state recognition models. To this end, the Hoeffding inequality and Bennett inequality are used to effectively extract state information from temperature features.

[0089] For the K feature clusters after clustering The non-parametric method is used for analysis and the corresponding state data is obtained. Take the kth feature cluster data as an example to introduce the proposed method.

[0090] The sample mean of the same state should change less than the error given by the Hoeffding inequality. Therefore, the proposed non-parametric method is used to analyze the M features of k , aiming to obtain the state change point of each feature. Take the first feature data of as an example to introduce how to determine the change point using the Hoeffding inequality.

[0091] First, the statistical quantity needs to be determined within the statistical interval η stat , i.e. the maximum value b max and the minimum value a min of the data.

[0092] Next, detection starts from the η stat +1th data, and the data stream is recorded as As time increases, the increased data is recorded as The split point is The overall data is recorded as

[0093]

[0094] and The respective mean values are and

[0095] Then, the error bounds of and are calculated as follows:

[0096]

[0097] where ε basis , ε add and εall respectively, n and error bounds, n basis , n add and n all are the sample sizes, δ H is the significance level. is the significance level.

[0098] is less than is less than is less than is less than is less than is less than

[0099]

[0100] where ε basis and ε all are the error bounds.

[0101] Finally, it is determined whether the current split point is a change point. The error bound ε cond is calculated as

[0102]

[0103] where δ H is the significance level, n basis is the sample size, and n add is the sample size.

[0104] The state change determination is based on the following:

[0105]

[0106] If is a change point, the parameters are reset and the above operations are repeated from until all data points are detected. The null hypothesis is the alternative hypothesis is rejected when

[0107] In summary, a total of state change points are obtained. The number of change points from features can obtain the number of states.​​​​The change points of the same state are not the same, and need to be determined based on the state probability of the current sample The state change point of the whole state is calculated as follows:

[0108]

[0109] wherein, is the probability that the nth sample belongs to the th state; is the number of features of the nth sample belonging to the th state.

[0110] Assuming that the expectation of the state probability is 0, the state change point is determined by using Bennett inequality and sliding window. The Jth window is denoted as J∈[1,N-φ cond +1], and the limit of the window mean is calculated as follows:

[0111]

[0112] wherein, σ J 2 is the variance of the state probability in the Jth window, φ cond is the window size, and δ B is the significance level.

[0113] Therefore, when the window mean is greater than δ cond , the (J+φ -1)th sample of is the change point of the th state.

[0114] Finally, the change points in the range η stat are averaged to obtain the final result. Accordingly, N final state change points can obtain N final +1 sets of state data.

[0115] 3. Feature clustering result

[0116] Firstly, the 14 temperature features are clustered by using the proposed multi-distance fusion feature clustering method, the interval size θ x is set to 50, and the feature group number K is set to 3, that is, the 14 features are clustered into 3 feature clusters.

[0117] ​Then, the Euclidean distance vector of 14 temperature features and the Wasserstein distance between each pair of vectors are calculated. The more similar the range and trend of the feature vectors are, the smaller the Wasserstein distance is; on the contrary, the larger the Wasserstein distance is. Therefore, the proposed range consistency criterion and trend consistency criterion can provide effective range and trend information for feature clustering.

[0118] Finally, the proposed feature clustering method based on multi-distance fusion divides the 14 temperature features into 3 feature clusters, and the clustering results are shown in Table 1 as follows:

[0119] Table 1 Feature clustering results based on multi-distance fusion

[0120]

[0121]

[0122] As shown in Table 1, the 7th to 9th features are divided into the 1st feature cluster, the 10th and 11th features are divided into the 2nd feature cluster, and the 1st to 6th features and the 12th to 14th features are divided into the 3rd feature cluster.

[0123] 4. State qualification and data acquisition results

[0124] First, the Hoeffding inequality is used to detect the change points of individual features in the 3 feature clusters to obtain the change points of individual features, the significance level δ H is set to 0.001, the statistical interval η stat is set to 90. The Bennett inequality is used to comprehensively judge the change points of features in the group to obtain the state change points of feature clusters, the significance level δ B is set to 0.001, and the window size φ cond is set to 50. The state change points of the 3 feature clusters are shown in Table 2.

[0125] Table 2 State change point positions of different feature clusters

[0126]

[0127] Finally, the 215th sample is compared with η statThe first state position is the average of the 236th sample, which is the average of the 253rd and 240th samples in the range. The second state position is the average of the 347th sample, which is the average of the 327th and 366th samples. The third state position is obtained by averaging the 472nd and 513th samples, i.e., the 493rd sample. The fourth state position is obtained by averaging the 753rd and 774th samples, i.e., the 764th sample. The fifth state position corresponds to the 866th sample. To visually display the detection results, feature 7 (first feature cluster), feature 9 (second feature cluster), and feature 6 (third feature cluster) are selected, as shown in Figure 3 .

[0128] Figure 2 The results show that the proposed method effectively detects the change positions between various states, and the six states are different, reflecting different combustion conditions at different times in the furnace. This provides strong support for operators to adjust operating parameters and enables researchers to effectively build data-driven models. For this purpose, the six sets of state data obtained are shown in Table 3:

[0129] Table 3 State data information

[0130] Status Sample number 1 235 2 111 3 146 4 271 5 102 6 125

[0131] 5. State recognition results of classical algorithms

[0132] To verify the effectiveness of the proposed method, four classical algorithms, including decision tree (DT), random forest (RF), support vector machine (SVM), and neural network (NN), are used to build state recognition models. The parameter settings of each model are as follows: 1) DT, leaf node sample size is 1, and feature selection number is 5; 2) RF, forest number is 20, leaf node sample size is 1, and feature selection number is 5; 3) SVM, kernel function is linear kernel function, and kernel size is 1; 4) NN, learning rate is 0.001, training iteration is 2000, and network structure is 14-50(ReLu)-100(ReLu)-150(ReLu)-4(Softmax). Since DT, RF, and NN have randomness, they are trained 10 times repeatedly to take the average. The state data is randomly divided into training set and test set according to different proportions, as shown in Table 4:

[0133] Table 4 Dataset division results

[0134]

[0135] The accuracy, precision, recall, and F1-score are used to verify the performance of the state recognition model, and are calculated as follows:

[0136]

[0137]

[0138] where TP is the true class; TN is the true negative class; FP is the false positive class; and FN is the false negative class.

[0139] where the precision, recall, and F1-score are calculated using the macro-averaging method as follows:

[0140]

[0141] where N c is the number of states; and are the precision, recall, and F1-score of the i-th state.

[0142] Finally, the state recognition results of the test set for DT, RF, SVM, and NN are shown in Table 5:

[0143] Table 5 State recognition results of the test set

[0144]

[0145]

[0146] From Table 5, we can see that:

[0147] 1) Based on the obtained state data set, SVM performs outstandingly in the state recognition task. In all data set proportions, SVM achieves the best scores (accuracy, precision, recall, and F1 value) among all algorithms. In addition, SVM is not very sensitive to the change in sample size, and can effectively recognize even in the case of 1:3 (small sample size).

[0148] 2) DT, as a weak learner, is very sensitive to the sample size. As the number of training samples increases, the recognition accuracy continuously improves. For example, when the proportion of the data set increases from 3:1 to 1:1 and then to 1:3, the accuracy increases from 0.9426 to 0.9654 and finally to 0.9924.

[0149] 3) RF, as an ensemble learning algorithm based on DT, is more suitable for small sample modeling and has good performance under different training sample proportions. It is worth noting that RF achieves a near-perfect score under the 3:1 proportion.

[0150] 4) NN showed good performance in state recognition task, and achieved the highest score at the ratio of 3:1. However, compared with other algorithms, the structure and parameter design of neural network is more complex, and the stability is poor.

[0151] In summary, the more training data available, the more advantageous to data-driven algorithms. Therefore, high-quality state data sets are essential for building accurate and robust perception models. The recognition results of the four algorithms show that the proposed state qualitative and data acquisition method has good modeling performance, obvious separability and significant differences between different states. It provides valuable support for subsequent data analysis and modeling tasks.

[0152] Therefore, the present application can provide high-quality, accurate temperature state data by providing a city solid waste incineration temperature state qualitative and classification method, realize automatic and efficient qualitative and data-driven grouping of industrial temperature state, and do not need prior distribution assumption, greatly widen the actual application scene and enhance the adaptability.

[0153] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts between the various embodiments can be mutually referred to.

[0154] The principles and implementation modes of the present application are described by applying specific examples in this paper. The above description of the embodiments is only used to help understand the method of the present application and its core idea. At the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for characterizing and classifying the temperature state of municipal solid waste incineration, characterized in that: include: On-site temperature process data from a municipal solid waste incineration power plant is collected and preprocessed using a downsampling strategy to obtain a temperature sample data matrix. Feature measurement is then performed on the temperature sample data matrix to ensure feature consistency. Based on the temperature sample data matrix, feature clustering based on multi-distance fusion is performed to obtain multiple feature clusters with consistent ranges and trends; For the feature cluster, the Hoeffding inequality is used to perform change point detection to obtain multiple feature change points; Purifying the multiple feature change points using Bennett's inequality to calculate the final state change point of each feature cluster, thereby obtaining multiple sets of labeled temperature state data; Based on decision tree, random forest, support vector machine and neural network, a state recognition model is constructed to verify the validity of the labeled temperature state data.

2. A method for characterizing and classifying the temperature state of municipal solid waste incineration according to claim 1, characterized in that: Temperature process data samples from a municipal solid waste incineration power plant are collected and preprocessed using a downsampling strategy to obtain a temperature sample data matrix. Feature measurement is then performed on the temperature sample data matrix to ensure feature consistency, including: Collect temperature process data samples from a municipal solid waste incineration power plant, use a deep learning downsampling strategy to segment the temperature process data samples into multiple segments at set intervals along the sample direction, and then calculate the mean of each segment to obtain a temperature sample data matrix; Based on the temperature sample data matrix, the Euclidean distance between each pair of features is calculated to obtain multiple feature Euclidean distance vectors that measure the range consistency between features. Then, based on the feature Euclidean distance vectors, the Wasserstein distance is used to extract trend information that measures the consistency of feature trends to ensure feature consistency.

3. A method for characterizing and classifying the temperature state of municipal solid waste incineration according to claim 2, characterized in that: Based on the temperature sample data matrix, feature clustering based on multi-distance fusion is performed to obtain multiple feature clusters with consistent ranges and trends, including: S1. Based on the temperature sample data matrix with feature consistency, randomly select a Euclidean distance vector as a first initial center point, calculate the Wasserstein distances between other Euclidean distance vectors and the first initial center point, obtain the selection probabilities of other Euclidean distance vectors, and select the Euclidean distance vector with the largest selection probability as the second initial center point; S2, repeat step S1 until all the feature clusters are selected to obtain multiple center points; S3, calculating the Wasserstein distance between the Euclidean distance vector and the center point, and assigning the Euclidean distance vector to the feature cluster with the smallest Wasserstein distance; S4, after the Euclidean distance vector is allocated, the mean of the Euclidean distance vectors in the feature cluster is calculated to update the center point; S5. Repeat step S4 until the Euclidean distance vector in the feature cluster does not change, and then distribute the Euclidean distance vector to the corresponding feature cluster to obtain multiple feature clusters with consistent ranges and trends.

4. A method for characterizing and classifying the temperature state of municipal solid waste incineration according to claim 3, characterized in that: For the feature cluster, the Hoeffding inequality is used to perform change point detection to obtain multiple feature change points, including: For each of the feature clusters, a statistical window is preset, and the maximum value, minimum value, mean value, and mean error limit of the statistical data within the statistical window are calculated. The Hoeffding inequality is used to determine whether the mean change of the current temperature sample data matrix exceeds the mean error limit. If so, it is determined to be an initial state change point; According to the initial state change point, the null hypothesis is set to be unchanged, and the alternative hypothesis is set to be changed, and then the state change is determined using the statistical significance level to count multiple change points of each feature in the feature cluster to obtain multi-feature change points.

5. A method for characterizing and classifying the temperature state of municipal solid waste incineration according to claim 4, characterized in that: According to the state probability, the multiple feature change points are purified using Bennett's inequality to calculate the final state change point of each feature cluster, and multiple groups of labeled temperature state data are obtained, including: Based on the multiple feature change points, the state probability variance and the window mean limit are calculated within the statistical window, and Bennett's inequality is used to determine whether the mean exceeds the window mean limit. If so, it is determined to be the final state change point of the statistical window; The interval mean of the final state change points of all the feature clusters is calculated, and the interval mean is set as the final state division point to divide the statistical window into multiple combustion state segments, thereby obtaining multiple groups of labeled temperature state data.

6. A method for characterizing and classifying the temperature state of municipal solid waste incineration according to claim 5, characterized in that: According to the state probability, a state recognition model is constructed based on decision trees, random forests, support vector machines, and neural networks to verify the validity of the labeled temperature state data, including: Set the number of leaf node samples of the decision tree to 1 and the number of feature selections to 5; The number of random forests is set to 20, the number of leaf node samples is set to 1, and the number of feature selections is set to 5; Set the kernel function of the support vector machine to a linear kernel function and the kernel size to 1; The learning rate of the neural network is set to 0.001, the number of training generations is set to 2000, and the network structure is set to a four-layer feedforward neural network; According to the state probability, the state recognition model is constructed based on the set decision tree, random forest, support vector machine and neural network; The temperature state data is divided into three groups of samples with different proportions, and the three groups of samples are input into the state recognition model. The accuracy, precision, recall rate and F1 score are used to verify the validity of the labeled temperature state data.

7. A method for characterizing and classifying the temperature state of municipal solid waste incineration according to claim 6, characterized in that: The four-layer feedforward neural network includes an input layer consisting of 14 neurons, a first hidden layer consisting of 50 neurons and a ReLU activation function, a second hidden layer consisting of 100 neurons and a ReLU activation function, a third hidden layer consisting of 150 neurons and a ReLU activation function, and an output layer consisting of 4 neurons and a Softmax activation function.

Images