Industrial kiln combustion state prediction method based on CNN-LSTM-Attention model

Abstract

The invention discloses an industrial kiln combustion state prediction method based on a CNN-LSTM-Attention model, and the method comprises the steps: collecting multi-dimensional operation parameters in the combustion process of an industrial kiln through multi-sensor fusion, and constructing a feature data set; performing data preprocessing on the feature data set, and constructing a training sample set; a CNN-LSTM-Attention fusion model is constructed, and the CNN-LSTM-Attention fusion model is trained by using the training sample set; and based on a dynamic threshold early warning mechanism, the trained CNN-LSTM-Attention fusion model is utilized to realize real-time prediction and fault early warning of the combustion state. The method has the advantages that experimental results show that the prediction accuracy of the model reaches 97%, and an efficient and reliable solution is provided for intelligent monitoring and predictive maintenance of the industrial kiln.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring and combustion state prediction of industrial kiln, and particularly relates to an industrial kiln combustion state prediction method based on a CNN-LSTM-Attention model. BACKGROUND

[0002] In the operation process of industrial kiln, real-time monitoring and fault early warning of combustion state are of great significance to ensure production safety, improve product quality and reduce energy consumption. Traditional kiln fault prediction mainly relies on two ways: one is to set a fixed threshold through PLC to monitor production data in real time, and alarm when the data exceeds the threshold and manually troubleshoot; the other is to arrange regular inspection and detect on site by experienced personnel. However, these methods generally have problems such as inaccurate judgment standard and untimely fault identification, which are difficult to meet the needs of modern industrial intelligent and refined management.

[0003] In recent years, related researches have tried to use different data-driven methods to improve the fault prediction ability of industrial boilers. Some researches propose a fault state prediction method based on SOM-LSTM, which first uses self-organizing mapping network to fuse multi-dimensional monitoring data, constructs a unified health index through model output error, and normalizes it to reflect the device health deterioration process, and then establishes a sequence-to-sequence regression prediction model based on LSTM, combined with threshold monitoring to achieve early warning, with an average prediction accuracy of 91.8% in the cases of superheater damage and coal economizer pipe leakage, which is better than BP network and SVR. Some researches also use wavelet packet analysis to process vibration signals and build SVM model for the fault of heating surface of power plant pulverized coal boiler, achieving a detection accuracy of 93.2%.

[0004] At the same time, the kiln operation data usually has the problem of sample imbalance, with the number of fault samples much less than that of normal samples, which further affects the generalization ability and fault recognition rate of the model. Existing solutions such as simple oversampling or weighted loss function can alleviate this problem to some extent, but there are still limitations in dealing with high-dimensional time series data in complex industrial scenarios.

[0005] Therefore, there is an urgent need for a kiln combustion state prediction method that can effectively fuse multi-source sensor data, accurately extract spatio-temporal features, adaptively adjust model attention, and handle sample imbalance problems, to achieve high-precision and real-time state monitoring and fault early warning, and provide strong guarantee for the safe and stable operation of industrial kiln. SUMMARY

[0006] In view of the deficiencies of the prior art, the purpose of the present application is to provide a CNN-LSTM-Attention model-based industrial kiln combustion state prediction method, which realizes high-precision feature extraction and real-time state recognition of kiln multi-dimensional operating parameters through multi-sensor data fusion, sample imbalance optimization and dynamic threshold early warning mechanism, and can be widely applied to high-temperature production equipment in glass, ceramic, metal processing, chemical and other industries, providing technical support for predictive maintenance and safe and stable operation of industrial kilns.

[0007] To achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0008] In a first aspect, the present application provides a CNN-LSTM-Attention model-based industrial kiln combustion state prediction method, which is characterized by comprising the following steps:

[0009] Step 1: Collecting multi-dimensional operating parameters in the combustion process of the industrial kiln through multi-sensor fusion to construct a feature data set;

[0010] Step 2: Data preprocessing of the feature data set and construction of a training sample set;

[0011] Step 3: Constructing a CNN-LSTM-Attention fusion model and training the CNN-LSTM-Attention fusion model using the training sample set;

[0012] The CNN-LSTM-Attention fusion model comprises an input layer, a CNN layer, an LSTM layer and an output layer connected in sequence, wherein the CNN layer is used to extract local spatio-temporal features in the multi-dimensional operating parameters; the LSTM layer is used to capture long-term time-dependent relationships of the multi-dimensional operating parameters according to the extracted local spatio-temporal features; the Attention layer is used to dynamically allocate feature weights; and the output layer is used to output prediction probabilities of multi-class kiln operating states according to the formaldehyde feature vector output by the Attention layer;

[0013] Step 4: Realizing real-time prediction and fault warning of the combustion state based on a dynamic threshold early warning mechanism and using the trained CNN-LSTM-Attention fusion model.

[0014] Further, the feature data set comprises instantaneous gas flow parameters of several channels, total instantaneous flow parameters, total flow parameters, gas valve opening parameters of several channels, flue gas temperature parameters, air pressure parameters, air thermocouple temperature parameters of several channels and temperature difference parameters of several channels.

[0015] Further, the data preprocessing of the feature data set in step 2 comprises the following processes:

[0016] Standardization preprocessing: converting the date string in the feature dataset into a standard time format;

[0017] Feature scaling: data cleaning, normalization and missing value processing on the data in the feature dataset, and extracting rolling statistical features and difference features;

[0018] Sliding window segmentation: constructing time series samples using a sliding time window;

[0019] Data balancing: using the Borderline-SMOTE method to oversample the minority class samples, and introducing FocalLoss as the loss function to dynamically adjust the sample weight.

[0020] Further, the expression of the Focal Loss loss function is:

[0021]

[0022] wherein, is the prediction probability of the model for the true class, and γ≥0 is the adjustment factor.

[0023] Further, the CNN layer includes 2 one-dimensional convolution structures, the first one-dimensional convolution structure is provided with 64 filters, a 3x1 convolution kernel and a ReLU activation function, and is subsequently connected to a one-dimensional maximum pooling unit and a regularization unit, the second one-dimensional convolution structure is provided with 128 filters, a 3x1 convolution kernel and a ReLU activation function, and is subsequently connected to a one-dimensional maximum pooling unit and a regularization unit.

[0024] Further, the LSTM layer includes 4 stacked LSTM units, the return sequence of the first 3 LSTM units is set to true, the return sequence of the last LSTM unit is set to false, and a batch normalization unit and a regularization unit are connected after each LSTM unit.

[0025] Further, the training of the CNN-LSTM-Attention fusion model using the training sample set in step 3 includes the following processes:

[0026] Step 3.1, divide the training sample set into training set and test set according to the proportion of 8:2, and set the training round and batch;

[0027] Step 3.2, the CNN-LSTM-Attention fusion model is trained by using the training set, and three callback mechanisms of ModelCheckpoint, EarlyStopping and ReduceLROnPlateau are introduced to dynamically control the training process, wherein the ModelCheckpoint mechanism is used to save the model parameters when the performance of the test set is optimal, the EarlyStopping mechanism is used to terminate the training in advance when the validation loss does not decrease continuously for many rounds, and the ReduceLROnPlateau mechanism is used to automatically reduce the learning rate when the validation loss is stagnant.

[0028] Step 3.3, the trained CNN-LSTM-Attention fusion model is tested by using the test set, and the CNN-LSTM-Attention fusion model at the optimal performance of the test set is obtained.

[0029] Further, the dynamic threshold early warning mechanism comprises: adjusting the classification threshold by temperature scaling; and dynamically adjusting the alarm sensitivity in combination with the characteristics of the kiln process stage.

[0030] In a second aspect, the present application provides an industrial kiln combustion state prediction system based on a CNN-LSTM-Attention model, which is used to realize the method as described in the first aspect, and comprises:

[0031] A data acquisition module is configured to acquire multi-dimensional operating parameters in the industrial kiln combustion process by multi-sensor fusion, and construct a feature data set.

[0032] A data preprocessing module is configured to preprocess the feature data set and construct a training sample set.

[0033] A model construction and training module is configured to construct a CNN-LSTM-Attention fusion model, and train the CNN-LSTM-Attention fusion model by using the training sample set.

[0034] The CNN-LSTM-Attention fusion model comprises an input layer, a CNN layer, an LSTM layer and an output layer connected in sequence, wherein the CNN layer is configured to extract local space-time features in the multi-dimensional operating parameters; the LSTM layer is configured to capture long-term time-dependent relationships of the multi-dimensional operating parameters according to the extracted local space-time features of the multi-dimensional operating parameters; the Attention layer is configured to dynamically allocate feature weights; and the output layer is configured to output prediction probabilities of multi-class kiln operating states according to a formaldehyde feature vector output by the Attention layer.

[0035] A state prediction module is configured to utilize the trained CNN-LSTM-Attention fusion model to realize real-time prediction and fault warning of the combustion state based on a dynamic threshold early warning mechanism.

[0036] In a third aspect, the present application provides a computer device comprising a processor and a memory, wherein the memory stores a computer program, and the computer program is called and executed by the processor to implement the method of the first aspect.

[0037] The present application has the following remarkable effects:

[0038] 1. The industrial furnace combustion state prediction method based on CNN-LSTM-Attention according to the present application collects 30-dimensional core combustion parameters through multi-sensor fusion, adopts Borderline-SMOTE oversampling to expand minority class samples, and FocalLoss dynamically adjusts sample weights to optimize the imbalance problem, combines the CNN-LSTM-Attention model to extract local spatiotemporal features, model long-term dependencies, and focus on key abnormal signals, and realizes end-to-end prediction; greatly improves the fault recognition rate and model generalization ability, and realizes a prediction accuracy of 97% in the industrial furnace combustion state prediction task, which is about 20 percentage points higher than the traditional judgment method relying on artificial experience, and the response time reaches seconds, which is significantly faster than the 4 to 12 hours of manual inspection, and is suitable for various industrial furnaces such as tubular annealing furnaces, and can be extended to the state monitoring and fault warning scene of other high-temperature industrial production equipment.

[0039] 2. The industrial furnace combustion state prediction method based on CNN-LSTM-Attention according to the present application extracts local spatiotemporal features through the CNN layer, captures long-term dependencies of multi-dimensional parameters through the LSTM layer, and dynamically allocates feature weights through the Attention layer, while introducing temperature scaling to adjust the prediction probability calibration, and combining process stage features to calculate alarm sensitivity; greatly improves the model's adaptability to complex working conditions, effectively reduces the false positive and false negative rates, ensures the real-time and reliability of the warning results, and can be extended to high-temperature production equipment in the glass, ceramic, and metal processing industries, providing an efficient solution for intelligent monitoring and predictive maintenance of industrial furnaces. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 The flowchart of the method of the present application is shown in the figure;

[0041] Figure 2 The TD1 temperature deviation schematic diagram of the four combustion classification categories in the present application is shown in the figure;

[0042] Figure 3 The comparison diagram before and after data balancing in the present application is shown in the figure;

[0043] Figure 4 This is a diagram showing the overall architecture of the CNN-LSTM-Attention fusion model in this invention;

[0044] Figure 5 This is a schematic diagram of the accuracy and loss function in this invention;

[0045] Figure 6 This is a schematic diagram of the confusion matrix in this invention;

[0046] Figure 7 For comparison of experimental radar images;

[0047] Figure 8 This is a schematic diagram of the system described in this invention. Detailed Implementation

[0048] The specific embodiments and working principles of the present invention will be further described in detail below with reference to the accompanying drawings.

[0049] Example 1:

[0050] like Figure 1 As shown, a method for predicting the combustion state of an industrial kiln based on a CNN-LSTM-Attention model is described, with the following specific steps:

[0051] Step 1: Collect multi-dimensional operating parameters of the industrial kiln combustion process through multi-sensor fusion to construct a feature dataset;

[0052] Generally, the feature dataset includes several instantaneous gas flow parameters, total instantaneous flow parameters, total flow parameters, several gas valve opening parameters, flue gas temperature parameters, air pressure parameters, several air thermocouple temperature parameters, and several temperature difference parameters.

[0053] In specific implementation, a tubular annealing furnace used for heat treatment of steel wire by a steel wire manufacturing enterprise is taken as the implementation object. This equipment provides a stable high-temperature environment of 800-1000℃ through gas combustion, and the combustion state directly affects the key properties of steel wire such as hardness and toughness. The data acquisition stage relies on the hardware architecture of this equipment and includes three major functional modules: First, the gas supply module is configured with 8 independent gas branches (corresponding to the instantaneous gas flow acquisition of IF1~IF8), a total flow control unit (corresponding to the acquisition of TIF total instantaneous flow and TFR total flow), and 8 gas valves (corresponding to the acquisition of VP1~VP8 opening degree) to realize the on-demand gas distribution; Second, the temperature monitoring module is equipped with 2 air thermocouples (AT1, AT2), flue gas temperature sensor (FGT), and furnace body zone temperature detection points (used to calculate the temperature difference of TD1~TD8, that is, the difference between the actual temperature and the set temperature, with a range covering -166~20℃) to capture the combustion temperature distribution; Third, the pressure monitoring module collects the air supply pressure through an air pressure sensor (AP) to ensure stable combustion air distribution. The data collection scheme adopted a 24-hour uninterrupted mode, with a period from 00:00 on March 1, 2024 to 24:00 on March 18, 2024. The sampling interval was 10 seconds per sampling, and a total of 38 raw parameters were collected. After removing redundant parameters through correlation analysis, a 30-dimensional core feature set was finally determined to provide real industrial working condition data for subsequent model training.

[0054] That is, the feature dataset corresponding to the implementation object is a 30-dimensional feature set, specifically including: 8 instantaneous gas flow rates (IF1~IF8), total instantaneous flow rate (TIF), total flow rate (TFR), 8 gas valve openings (VP1~VP8), flue gas temperature (FGT), air pressure (AP), 2 air thermocouple temperatures (AT1, AT2) and 8 temperature differences (TD1~TD8, with ranges of -130-15℃, -154-+9℃, -150-20℃, -166-+2℃, -161-+3℃, -138-+2℃, -130-+4℃, and -144-+3℃ respectively). The temperature difference parameter is obtained by calculating the difference between the actual temperature and the set temperature, and is used to reflect combustion uniformity and abnormal heat distribution.

[0055] Based on the process records and fault logs of the tubular annealing furnace, combined with Figure 2 The fluctuation pattern of the key feature "TD1 temperature difference" is analyzed, and the operating status is divided into four categories: normal (a1), process adjustment (a2), manual adjustment (a3), and fault (b1), with samples labeled accordingly.

[0056] Class A1 (normal state), corresponding to Figure 2 The curves with fluctuation range ≤ ±5℃, with stable parameters and no intervention, and the labeled time period includes 6 time periods such as March 1, 00:00:09-19:56:46, totaling 69,224 samples;

[0057] Category A2 (Process Adjustment Status) Figure 2 The temperature fluctuation of the medium curve is ≤±15℃ (caused by process switching). The labeled time period includes 4 time periods, such as March 1, 19:56:57-22:10:53, with a total of 31,504 samples.

[0058] Category A3 (Artificially regulated state), corresponding to Figure 2 The temperature fluctuation of the curve was ≤±25℃ (caused by human intervention). The time period marked was from 04:39:14 on March 8 to 16:16:58 on March 11, with a total of 25,526 samples.

[0059] Class B1 (fault state), corresponding to Figure 2 The temperature fluctuation of the medium curve is greater than ±30℃ (caused by fan shutdown due to undervoltage). The time period marked is from 16:17:09 to 18:20:39 on March 11 and from 13:10:42 to 16:01:07 on March 17, with a total of 1496 samples.

[0060] Figure 2 It intuitively demonstrates the differences in characteristics among different categories, providing a visual basis for sample labeling and ensuring labeling accuracy.

[0061] In this example, the samples and time periods for the four categories are shown in Table 1:

[0062]

[0063] Step 2: Perform data preprocessing on the feature dataset and construct a training sample set;

[0064] This embodiment preprocesses the feature dataset using a multi-step strategy. In some implementations, the data preprocessing of the feature dataset includes the following steps:

[0065] Step 2.1, Standardization Preprocessing: Convert the date strings in the feature dataset into a standard time format to match the time recording rules of the data acquisition unit;

[0066] Step 2.2, Feature Scaling: Clean, normalize, and handle missing values ​​in the feature dataset, and extract rolling statistical features and difference features;

[0067] For missing values, rows containing missing values ​​are removed from the training set, and the test set is filled with the median.

[0068] To enhance the model's representational ability, rolling statistical features such as the rolling mean and standard deviation of numerical features are extracted, and difference features (Δx(t)=x(t)-x(t-1)) are introduced to enhance the model's ability to represent parameter change trends, expanding the feature dimension from 30 to 90 dimensions and capturing them more accurately. Figure 2The changing trends of parameters such as temperature difference;

[0069] Label encoder is used to encode the label field into integers, the four types of labels are encoded as 0-3 and the encoder parameters are saved. MinMaxScaler is used to normalize all numerical features to the [0,1] interval to eliminate the difference in units.

[0070] Step 2.3, Sliding window segmentation: Construct time series samples using a sliding time window;

[0071] Specifically, a sliding time window is used to convert one-dimensional data into two-dimensional time-series samples, resulting in a training set of 102,400 data points and a test set of 25,600 data points, providing standard format data for subsequent model input.

[0072] Step 2.4, Data Balancing: The Borderline-SMOTE method is used to oversample the minority class samples, and Focal Loss is introduced as the loss function to dynamically adjust the sample weights.

[0073] To address the class imbalance issue of only 1496 fault samples in class b1 (compared to 69224 in class a1, 31504 in class a2, and 25526 in class a3), the Borderline-SMOTE algorithm was employed to oversample and augment the minority class samples. This ultimately balanced the number of samples across all classes to 13398-13840, effectively mitigating the adverse effects of class bias on model training. The results are as follows: Figure 3 As shown: Before balancing, the left bar chart shows 69,224 samples in class a1, 31,504 in class a2, 25,526 in class a3, and only 1,496 in class b1, indicating significant class bias. After balancing, the right bar chart shows that by identifying boundary samples of class b1 and generating new samples with consistent features, the number of samples in the four classes is evenly distributed. This processing solves the common class bias problem in industrial data, avoids model training bias towards the majority class, lays a data foundation for improving the fault recognition rate of class b1, and forms a synergistic optimization with the subsequent FocalLoss loss function.

[0074] It should be noted that the Borderline-SMOTE algorithm is an oversampling method for class imbalance problems, primarily used to address the issue of minority class samples being misclassified near the decision boundary. Its core idea is to oversample only the boundary samples of the minority class to improve the classifier's ability to discriminate key regions. Its main steps are as follows:

[0075] Identify boundary samples: Calculate the K-nearest neighbors for minority class samples, and select samples in the K-nearest neighbors where the majority class samples account for more than 50% as boundary samples;

[0076] Synthesized new samples: Synthetic samples are generated between boundary samples and minority class samples using linear interpolation.

[0077] Add to training set: Add new samples to the original training set to balance the data distribution.

[0078] The Borderline-SMOTE algorithm oversamples minority class samples, enhancing only the boundary samples that play a crucial role in classification, thus avoiding the blind oversampling of "safe samples" by the traditional SMOTE algorithm. This also reduces the risk of noise being introduced from samples far from the boundary.

[0079] To further address model training bias caused by uneven distribution of sample classes, this invention introduces FocalLoss as the core loss function. Its core principle is to dynamically adjust the sample loss weights through a modulation factor, guiding the model to focus on feature learning of hard-to-classify samples (such as fault class b1). The mathematical expression for FocalLoss is:

[0080]

[0081] in, The probability of the model predicting the true class is defined as:

[0082]

[0083] In the formula, For the true labels of the samples, Predict the probability that a sample belongs to class 1 for the model; parameters As a moderating factor, it is used to adjust the degree of attention given to easy and difficult samples, i.e., the weight allocation. When, FocalLoss is equivalent to standard cross-entropy loss; when At this time, the loss weight of low-confidence samples is amplified, while the loss contribution of high-confidence samples is suppressed, thereby increasing the model's attention to minority and hard-to-classify samples.

[0084] In actual calculations, the predicted probabilities are clipped by a value (limiting the probability values ​​to the range [1e-8, 1-1e-8]) to avoid gradient anomalies caused by extreme probabilities. For multi-class tasks, one-hot encoding is used to process the labels, and the total model loss is obtained by weighted summation of the loss values ​​along the class dimension.

[0085]

[0086] in, Number of categories (in this embodiment) ), The sample belongs to the category The predicted probability.

[0087] Step 3: Construct a CNN-LSTM-Attention fusion model to achieve multi-level extraction and dynamic attention of key features in industrial time series data, and train the CNN-LSTM-Attention fusion model using the training sample set;

[0088] This invention achieves deep feature learning of industrial kiln time-series data by constructing a fusion architecture of "local feature extraction - long-term dependency modeling - key feature focusing". The specific implementation logic of this architecture can be obtained through... Figure 4 Intuitively, during implementation, the first step is to adapt the input features to the architecture. The original 30-dimensional features are expanded to 90 dimensions through rolling statistical features and difference features, and then normalized to the [0,1] interval by the MinMaxScaler to form the standard time series sample format required by the InputLayer.

[0089] The functional layers are then implemented sequentially. The CNN layer contains two one-dimensional convolutional structures. The first Conv1D layer has 64 filters, a 3×1 convolutional kernel, and a ReLU activation function, followed by MaxPooling1D and Dropout to capture local spatiotemporal correlations. The second Conv1D layer has 128 filters, a 3×1 convolutional kernel, and a ReLU activation function, also followed by MaxPooling1D and Dropout to deepen feature extraction. The LSTM layer has four stacked units. The first three layers have return_sequences=True, with the number of neurons being 128, 64, and 32 respectively. The final layer has `return_sequences=False` and 16 neurons. Each LSTM layer is followed by BatchNormalization and Dropout to model long-term time dependencies. The Attention layer is located between the LSTM layer and the fully connected layer. It calculates the weight distribution of the hidden state in the last LSTM layer through an additive attention mechanism, dynamically amplifying the weights of key fault features such as temperature difference and air pressure. The output layer has two fully connected layers. It outputs the predicted probabilities of four states, a1 to b1, through a softmax activation function. Temperature scaling is also introduced to optimize the probability calibration.

[0090] Therefore, the CNN-LSTM-Attention fusion model described in this embodiment includes an input layer, a CNN layer, an LSTM layer, and an output layer connected in sequence. The CNN layer adopts a one-dimensional design and a multi-layer convolutional structure, using a local window of the time-series sample as input, and is used to extract local spatiotemporal features from the multidimensional operating parameters through sliding calculation of the convolutional kernel. The LSTM layer is used to capture the long-term temporal dependencies of the multidimensional operating parameters based on the extracted local spatiotemporal features of the multidimensional operating parameters. The Attention layer is used to dynamically allocate feature weights. The output layer is used to output the predicted probabilities of the operating states of multiple kilns based on the formaldehyde feature vector output by the Attention layer.

[0091] The specific processing procedure of the CNN-LSTM-Attention fusion model for the input data is as follows:

[0092] First, the CNN layer employs a one-dimensional design and a multi-layer convolutional structure, using local windows of temporal samples as input. Local spatiotemporal features are extracted through sliding convolutional kernel computation. The first Conv1D layer, with 64 filters, 3×1 convolutional kernels, and a ReLU activation function, captures local temporal features through sliding convolution, introducing non-linear representation. A MaxPooling1D layer (pool_size=2) is then connected to retain key features and reduce dimensionality and computation; a Dropout layer randomly discards 40% of neuron connections to suppress overfitting and improve generalization ability. Finally, a second Conv1D layer (128 filters, 3×1 convolutional kernels, ReLU activation) is connected to deepen feature mining and combination based on the initial feature extraction, strengthening the capture of complex temporal patterns and continuously optimizing training.

[0093] Then, the local feature maps output by the CNN layers are flattened into one-dimensional feature vectors and input into LSTM layers to capture the long-term temporal dependencies of multi-dimensional working condition parameters. Four LSTM layers are stacked sequentially, with the first three layers retaining sequence outputs to support multi-layer feature propagation, and the last layer outputting a single vector. Each LSTM segment is followed by BatchNormalization to accelerate training and enhance stability, and Dropout operations to suppress overfitting. Subsequent layers are fully connected with ReLU activation to adapt to the needs of multi-class tasks.

[0094] Finally, to ensure the model focuses on features crucial for predicting combustion states, an attention mechanism is introduced after the LSTM layer to dynamically calculate the weight distribution of the LSTM hidden states. The weighted feature vector output from the attention layer is then input into the fully connected layer of the output layer, and after passing through the softmax activation function, the predicted probability distributions for each category (a1, a2, a3, b1) are output.

[0095]

[0096] in, This is the weight matrix of the fully connected layer. For bias terms, For the model's final predicted probability, satisfying .

[0097] In this embodiment, to improve the model's generalization ability and training stability, the present invention designs a multi-dimensional training optimization strategy. Specifically, the steps for training the CNN-LSTM-Attention fusion model using the training sample set are as follows:

[0098] Step 3.1: Randomly divide the training sample set into a training set and a test set in an 8:2 ratio to ensure that the sample distribution characteristics of the training set and the test set are consistent; set the training epoch to 100 and the batch size to 128, and use the Adam optimizer (initial learning rate 1e-4) to minimize FocalLoss.

[0099] Step 3.2: Train the CNN-LSTM-Attention fusion model using the training set, and introduce three callback mechanisms—ModelCheckpoint, EarlyStopping, and ReduceLROnPlateau—to dynamically control the training process. The ModelCheckpoint mechanism monitors the model's prediction accuracy on the validation set in real time. When the accuracy reaches its current optimum, it automatically saves the model parameters to avoid performance regression in later training. The EarlyStopping mechanism sets the validation loss as the monitoring metric. When the validation loss does not decrease in 10 consecutive training rounds, the training process is terminated early to effectively prevent overfitting. The ReduceLROnPlateau mechanism automatically reduces the learning rate to 1 / 10 of its original value when the validation loss stagnates (no decrease in 5 consecutive rounds), refining the parameter optimization process in the later stages of training and helping the model converge to a better solution.

[0100] Step 3.3: Test the trained CNN-LSTM-Attention fusion model using the test set to obtain the CNN-LSTM-Attention fusion model with the best performance on the test set.

[0101] Training effect passed Figure 5 Quantitative presentation. From Figure 6The test data shown indicates that the prediction accuracy for class a1 (normal state) is 98%, for class a2 (process adjustment) it is 96%, for class a3 (manual adjustment) it is 95%, and the critical prediction accuracy for class b1 (fault state) is 96.5%, with only 3.5% misclassified as class a3. This result is far superior to the traditional SOM-LSTM model, proving that the technical solution of amplifying key fault features such as temperature difference and air pressure in the Attention layer is effective and can lay the foundation for subsequent low false alarm rate early warning.

[0102] Step 4: Based on the dynamic threshold warning mechanism, the trained CNN-LSTM-Attention fusion model is used to realize real-time prediction of combustion status and fault warning.

[0103] In practical implementation, to address the problem that traditional fixed-threshold early warning methods are difficult to adapt to kiln process fluctuations and prone to false alarms / missed alarms, this invention proposes a dynamic threshold early warning mechanism to improve the calibrability of prediction probabilities. Specifically, the dynamic threshold early warning mechanism includes: adjusting the classification threshold by temperature scaling to improve the calibrability of prediction probabilities, and dynamically adjusting the alarm sensitivity in conjunction with the characteristics of kiln process stages, effectively reducing false alarm and missed alarm rates and ensuring the reliability and real-time nature of early warning results.

[0104] It should be noted that temperature scaling is a post-processing technique used to calibrate the output probabilities of machine learning models. By introducing a temperature parameter (T), it adjusts the distribution of the model's logits, making the prediction confidence more accurately reflect actual reliability. Temperature scaling technology is common knowledge in the field, and therefore will not be elaborated upon here.

[0105] In specific implementation, the dynamic adjustment of alarm sensitivity based on the characteristics of different kiln process stages involves: optimizing the dynamic threshold in stages by combining the parameter characteristics of different kiln process stages (such as kiln drying, normal production, and process adjustment). By pre-marking the time intervals of each process stage, the 95th quantile threshold is calculated for each stage, ensuring that the early warning strategy matches the process characteristics and further reducing the false alarm rate. The final early warning decision rule is:

[0106]

[0107] in, For the first The time-matter model predicts the probability that a sample belongs to the fault class (b1). This indicates that a fault warning has been triggered. This indicates that the current status is normal.

[0108] To highlight the advancements of this invention compared to existing technologies, this invention compares the CLA model with traditional manual maintenance, a single CNN model, and a CNN+LSTM model using six evaluation dimensions: accuracy, fault response time, false alarm rate, false negative rate, annual average maintenance cost optimization rate, and generalization ability. The results are presented through... Figure 7 The larger the radar chart radius, the better the performance. The CLA model leads the way in all dimensions, verifying the comprehensive advantages of this invention in terms of accuracy, efficiency, cost, and versatility. It solves the core pain points of traditional methods, such as low accuracy, slow response, and poor generalization.

[0109] The prediction model established through the embodiments of the present invention was verified on 58,008 sets of data from February 9 to February 16, 2025, and accurately identified three time intervals of state b1, which highly matched the actual fault situation.

[0110] As can be seen, the embodiments of the present invention provide an efficient and reliable solution for intelligent monitoring and predictive maintenance of industrial kilns, which can be widely applied to various high-temperature industrial production scenarios.

[0111] Example 2:

[0112] See appendix Figure 8 This invention provides an industrial kiln combustion state prediction system based on a CNN-LSTM-Attention model, used to implement the method described in Example 1, including:

[0113] The data acquisition module is used to collect multi-dimensional operating parameters during the combustion process of industrial kilns through multi-sensor fusion and construct a feature dataset;

[0114] The data preprocessing module is used to preprocess the feature dataset and construct a training sample set;

[0115] The model building and training module is used to build a CNN-LSTM-Attention fusion model and train the CNN-LSTM-Attention fusion model using the training sample set;

[0116] The CNN-LSTM-Attention fusion model comprises an input layer, a CNN layer, an LSTM layer, and an output layer connected in sequence. The CNN layer is used to extract local spatiotemporal features from multidimensional operating parameters; the LSTM layer is used to capture the long-term temporal dependencies of multidimensional operating parameters based on the extracted local spatiotemporal features; the Attention layer is used to dynamically allocate feature weights; and the output layer is used to output the predicted probabilities of multi-category kiln operating states based on the formaldehyde feature vector output by the Attention layer.

[0117] The state prediction module is used to achieve real-time prediction of combustion status and fault warning based on the dynamic threshold warning mechanism and the trained CNN-LSTM-Attention fusion model.

[0118] Example 3:

[0119] This invention provides a computer device including a processor and a memory, wherein the memory stores a computer program that, when called and executed by the processor, implements the method described in Embodiment 1.

[0120] In summary, the CNN-LSTM-Attention-based method for predicting the combustion state of industrial kilns described in this invention extracts local spatiotemporal features using CNN, captures long-term dependencies of multidimensional parameters using LSTM, and dynamically allocates feature weights using the Attention layer. It also introduces temperature scaling to adjust the prediction probability calibration and calculates a dynamic 95th quantile threshold based on process stage features. This significantly improves the model's adaptability to complex operating conditions, effectively reduces false alarms and false negatives, ensures the real-time performance and reliability of early warning results, and can be extended to high-temperature production equipment in industries such as glass, ceramics, and metal processing, providing an efficient solution for intelligent monitoring and predictive maintenance of industrial kilns.

[0121] The technical solution provided by this invention has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. It should be noted that those skilled in the art can make several improvements and modifications to this invention without departing from the principles of this invention, and these improvements and modifications also fall within the protection scope of the claims of this invention.

Claims

1. A method for predicting the combustion state of industrial kilns based on a CNN-LSTM-Attention model, characterized in that, Includes the following steps: Step 1: Collect multi-dimensional operating parameters of the industrial kiln combustion process through multi-sensor fusion to construct a feature dataset; Step 2: Perform data preprocessing on the feature dataset and construct a training sample set; Step 3: Construct a CNN-LSTM-Attention fusion model and train the CNN-LSTM-Attention fusion model using the training sample set; The CNN-LSTM-Attention fusion model comprises an input layer, a CNN layer, an LSTM layer, and an output layer connected in sequence. The CNN layer is used to extract local spatiotemporal features from multidimensional operating parameters; the LSTM layer is used to capture the long-term temporal dependencies of multidimensional operating parameters based on the extracted local spatiotemporal features; the Attention layer is used to dynamically allocate feature weights; and the output layer is used to output the predicted probabilities of multi-category kiln operating states based on the formaldehyde feature vector output by the Attention layer. Step 4: Based on the dynamic threshold warning mechanism, the trained CNN-LSTM-Attention fusion model is used to realize real-time prediction of combustion status and fault warning.

2. The method for predicting the combustion state of industrial kilns based on the CNN-LSTM-Attention model according to claim 1, characterized in that: The feature dataset includes several instantaneous gas flow parameters, total instantaneous flow parameters, total flow parameters, several gas valve opening parameters, flue gas temperature parameters, air pressure parameters, several air thermocouple temperature parameters, and several temperature difference parameters.

3. The method for predicting the combustion state of industrial kilns based on the CNN-LSTM-Attention model according to claim 1, characterized in that: Step 2 involves data preprocessing of the feature dataset, including the following steps: Standardization preprocessing: Converting date strings in the feature dataset into a standard time format; Feature scaling: The data in the feature dataset is cleaned, normalized, and missing value is handled. Rolling statistical features and difference features are then extracted. Sliding window segmentation: Time series samples are constructed using a sliding time window; Data balancing: The Borderline-SMOTE method is used to oversample minority class samples, and Focal Loss is introduced as the loss function to dynamically adjust sample weights.

4. The method for predicting the combustion state of industrial kilns based on the CNN-LSTM-Attention model according to claim 3, characterized in that: The expression for the Focal Loss function is as follows: in, γ represents the model's predicted probability of the true class, and γ≥0 is a modulating factor.

5. The method for predicting the combustion state of industrial kilns based on the CNN-LSTM-Attention model according to claim 1, characterized in that: The CNN layer includes two one-dimensional convolutional structures. The first one-dimensional convolutional structure has 64 filters, a 3×1 convolutional kernel, and a ReLU activation function, followed by a one-dimensional max pooling unit and a regularization unit. The second one-dimensional convolutional structure has 128 filters, a 3×1 convolutional kernel, and a ReLU activation function, followed by a one-dimensional max pooling unit and a regularization unit.

6. The method for predicting the combustion state of industrial kilns based on the CNN-LSTM-Attention model according to claim 1, characterized in that: The LSTM layer consists of four stacked LSTM units. The return sequence of the first three LSTM units is set to true, and the return sequence of the last LSTM unit is set to false. Each LSTM unit is followed by a batch normalization unit and a regularization unit.

7. The method for predicting the combustion state of industrial kilns based on the CNN-LSTM-Attention model according to claim 1, characterized in that: Step 3, training the CNN-LSTM-Attention fusion model using the training sample set, includes the following process: Step 3.1: Randomly divide the training sample set into a training set and a test set in an 8:2 ratio, and set the training rounds and batches; Step 3.2: Train the CNN-LSTM-Attention fusion model using the training set, and introduce three callback mechanisms: ModelCheckpoint, EarlyStopping, and ReduceLROnPlateau to dynamically control the training process. The ModelCheckpoint mechanism is used to save the model parameters when the performance on the test set is optimal. The EarlyStopping mechanism is used to terminate the training early when the validation loss no longer decreases for several consecutive rounds. The ReduceLROnPlateau mechanism is used to automatically reduce the learning rate when the validation loss stagnates. Step 3.3: Test the trained CNN-LSTM-Attention fusion model using the test set to obtain the CNN-LSTM-Attention fusion model with the best performance on the test set.

8. The method for predicting the combustion state of industrial kilns based on the CNN-LSTM-Attention model according to claim 1, characterized in that: The dynamic threshold early warning mechanism includes: adjusting the classification threshold by temperature scaling; and dynamically adjusting the alarm sensitivity based on the characteristics of the kiln process stage.

9. A combustion state prediction system for industrial kilns based on a CNN-LSTM-Attention model, used to implement the method described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to collect multi-dimensional operating parameters during the combustion process of industrial kilns through multi-sensor fusion and construct a feature dataset; The data preprocessing module is used to preprocess the feature dataset and construct a training sample set; The model building and training module is used to build a CNN-LSTM-Attention fusion model and train the CNN-LSTM-Attention fusion model using the training sample set; The CNN-LSTM-Attention fusion model comprises an input layer, a CNN layer, an LSTM layer, and an output layer connected in sequence. The CNN layer is used to extract local spatiotemporal features from multidimensional operating parameters; the LSTM layer is used to capture the long-term temporal dependencies of multidimensional operating parameters based on the extracted local spatiotemporal features; the Attention layer is used to dynamically allocate feature weights; and the output layer is used to output the predicted probabilities of multi-category kiln operating states based on the formaldehyde feature vector output by the Attention layer. The state prediction module is used to achieve real-time prediction of combustion status and fault warning based on the dynamic threshold warning mechanism and the trained CNN-LSTM-Attention fusion model.

10. A computer device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when called and executed by the processor, implements the method as described in any one of claims 1-8.

Images