Intelligent monitoring method and system for boiler operation state

Abstract

The invention discloses a boiler operation state intelligent monitoring method and system, and the method comprises the steps: collecting multi-source heterogeneous data in a boiler operation process, carrying out the cleaning and normalization processing of the multi-source heterogeneous data, extracting key features through wavelet transform, and obtaining standard multi-source heterogeneous data; modeling time series data in the standard multi-source heterogeneous data on the basis of an LSTM (Long Short-Term Memory) network, capturing a dynamic change trend of boiler operation to obtain time series characteristics, analyzing a hearth flame image and an infrared thermal image by using a CNN (Convolutional Neural Network), and extracting combustion state characteristics and thermal distribution characteristics; and based on the comprehensive state vector, a confidence interval of each feature dimension is calculated through a GMM Gaussian mixture model, an operation state deviation degree is analyzed and evaluated in combination with an entropy value, and when the deviation degree exceeds a preset threshold value, graded early warning is triggered. And the accuracy of boiler operation state intelligent monitoring is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for industrial equipment, and in particular to an intelligent monitoring method and system for boiler operating status. Background Technology

[0002] As a core piece of equipment in industrial production, the operating status of boilers directly affects energy efficiency and safety. Traditional monitoring methods mainly rely on fixed threshold judgments or single sensor data, which are insufficient to cope with the complex operating conditions of boilers involving multiple coupled and dynamically changing variables, resulting in problems such as high false alarm rates and delayed response. Existing technologies do not fully integrate the fusion analysis of multi-source data and dynamic adaptive learning capabilities, leading to insufficient generalization ability of monitoring models and failing to meet the requirements for high precision and real-time performance. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing an intelligent monitoring method and system for boiler operation status.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned intelligent monitoring method for boiler operating status: Multi-source heterogeneous data is collected during boiler operation. The multi-source heterogeneous data is cleaned and normalized. Key features are extracted by wavelet transform to obtain standard multi-source heterogeneous data. The time series data in the standard multi-source heterogeneous data is modeled based on the LSTM long short-term memory network to capture the dynamic change trend of boiler operation and obtain time series features. The furnace flame image and infrared thermal imaging are analyzed by CNN convolutional neural network to extract combustion state features and heat distribution features. A comprehensive state vector is generated by weightedly fusing temporal features, combustion state features, and heat distribution features through an attention mechanism. Based on the comprehensive state vector, the confidence interval of each feature dimension is calculated using the Gaussian mixture model (GMM), and the deviation of the operating state is evaluated by combining entropy analysis. When the deviation exceeds a preset threshold, a graded warning is triggered.

[0005] Furthermore, in the aforementioned intelligent monitoring method for boiler operating status, the step of collecting multi-source heterogeneous data during boiler operation, cleaning and normalizing the multi-source heterogeneous data, and extracting key features through wavelet transform to obtain standard multi-source heterogeneous data includes: Collect multi-source heterogeneous data during boiler operation, including at least operating parameter data, image data, environmental data, and equipment status data; Missing values ​​in linear interpolation numerical data are processed to remove blurry and occluded data in image data. Abnormal values ​​in complete data are identified based on the 3σ criterion and replaced with the mean value of the same time period. Invalid images are filtered through image quality assessment to obtain complete data. Numerical and image data are synchronized according to timestamps, and the complete data is normalized using Min-Max normalization to obtain normalized data. Normalized data is decomposed into 3-5 levels of wavelet decomposition to obtain approximation coefficients and detail coefficients. The approximation coefficients and the first two levels of detail coefficients are retained. Key features are selected by variance contribution rate to obtain standard multi-source heterogeneous data.

[0006] Furthermore, in the aforementioned intelligent monitoring method for boiler operation status, the step of modeling the time-series data in the standard multi-source heterogeneous data based on the LSTM long short-term memory network captures the dynamic change trend of boiler operation, obtains time-series features, and uses a CNN convolutional neural network to analyze furnace flame images and infrared thermal imaging to extract combustion state features and heat distribution features, including: Numerical data in standard multi-source heterogeneous data is divided into training set, validation set and test set. The input sequence is constructed using the sliding window method, with the window size set to 30. The output is a single-time feature vector. The single-time feature vector is input into the LSTM network, and the network parameters are adjusted using the validation set. Training stops when the validation set loss does not decrease for 5 consecutive rounds, and the time-series feature vector is output.

[0007] Furthermore, in the aforementioned intelligent monitoring method for boiler operation status, the step of modeling the time-series data in the standard multi-source heterogeneous data based on the LSTM long short-term memory network captures the dynamic change trend of boiler operation, obtains time-series features, and uses a CNN convolutional neural network to analyze furnace flame images and infrared thermal imaging to extract combustion state features and heat distribution features, including: Histogram equalization is used to improve image contrast, Gaussian filtering is used to remove noise, and the temperature values ​​of the thermal image are mapped to gray values ​​to preserve temperature gradient information. The processed image data is input into the fusion flame image branch and thermal imaging branch of the CNN convolutional neural network to obtain the combustion state feature vector and heat distribution feature vector.

[0008] Furthermore, in the aforementioned intelligent monitoring method for boiler operating status, the step of generating a comprehensive state vector by weightedly fusing temporal features, combustion state features, and heat distribution features through an attention mechanism includes: Initialize the query vector, use the dot product attention mechanism to calculate the similarity between the query vector and the temporal feature vector and the image feature vector respectively, and obtain the fusion weight by normalization through the Softmax function; Based on the fusion weights, the fusion time sequence characteristics, combustion state characteristics, and heat distribution characteristics, a 3D comprehensive state vector is obtained.

[0009] Furthermore, in the aforementioned intelligent monitoring method for boiler operating status, the step of calculating the confidence interval of each feature dimension based on the comprehensive state vector using a Gaussian mixture model (GMM), and evaluating the deviation of the operating status using entropy analysis, triggering a tiered early warning when the deviation exceeds a preset threshold, includes: Collect the comprehensive state vector under normal boiler operation as training data for GMM. Set the number of mixed components K=3-5 and use the EM algorithm to estimate the mean, covariance matrix and mixing coefficient of each component. Based on the trained GMM, a 95% confidence interval is calculated for each feature dimension, and the information entropy is calculated for each dimension of the real-time integrated state vector. The deviation is defined, with D≥1 indicating the existence of deviation.

[0010] Furthermore, in the aforementioned intelligent monitoring method for boiler operating status, the step of calculating the confidence interval of each feature dimension based on the comprehensive state vector using a Gaussian mixture model (GMM), and evaluating the deviation of the operating status using entropy analysis, triggering a tiered early warning when the deviation exceeds a preset threshold, includes: The comprehensive state vector, entropy value, and deviation D of boiler operation are calculated in real time. When 1.2≤D<1.5: mild warning; when 1.5≤D<2.0: moderate warning; when D≥2.0: severe warning.

[0011] Furthermore, in a boiler operating status intelligent monitoring system, the boiler operating status intelligent monitoring system includes the following modules: The multi-source data acquisition module is used to collect multi-source heterogeneous data during boiler operation, clean and normalize the multi-source heterogeneous data, and extract key features through wavelet transform to obtain standard multi-source heterogeneous data. The feature vector extraction module is used to model the time series data in the standard multi-source heterogeneous data based on the LSTM long short-term memory network, capture the dynamic change trend of boiler operation, obtain time series features, and use CNN convolutional neural network to analyze furnace flame images and infrared thermal imaging to extract combustion state features and heat distribution features. The feature vector fusion module is used to weightedly fuse temporal features, combustion state features, and heat distribution features through the Attention mechanism to generate a comprehensive state vector; The deviation-level early warning module is used to calculate the confidence interval of each feature dimension based on the comprehensive state vector using the Gaussian mixture model (GMM), and to evaluate the deviation of the operating state by combining entropy value analysis. When the deviation exceeds a preset threshold, a level warning is triggered.

[0012] Furthermore, in a boiler operation status intelligent monitoring system, the deviation classification early warning module includes the following sub-modules: The estimation submodule is used to collect the comprehensive state vector under normal boiler operation as training data for GMM. The number of mixed components is set to K=3-5, and the EM algorithm is used to estimate the mean, covariance matrix and mixing coefficient of each component. Define a submodule for calculating the 95% confidence interval for each feature dimension based on the trained GMM, calculating the information entropy for each dimension of the real-time integrated state vector, and defining the deviation degree, where D≥1 indicates the existence of deviation.

[0013] Furthermore, in a boiler operation status intelligent monitoring system, the deviation classification early warning module includes the following sub-modules: The calculation submodule is used to calculate the comprehensive state vector, entropy value and deviation D of boiler operation in real time. When 1.2≤D<1.5: mild warning, when 1.5≤D<2.0: moderate warning, when D≥2.0: severe warning.

[0014] Its beneficial effects lie in overcoming the limitations of traditional single-point monitoring methods in adapting to complex operating conditions by integrating time-series, image, and equipment status data, and utilizing an attention mechanism to weightedly fuse key features. CNN analysis of furnace flame images can accurately capture combustion anomalies, and combined with LSTM dynamic modeling of time-series data such as pressure and temperature, the accuracy of fault identification is improved compared to traditional methods. The online incremental learning strategy combined with transfer learning technology enables the model to respond to changes in boiler operating conditions in real time, avoiding false alarms caused by historical data bias. The early warning response time is shortened in scenarios of sudden steam pressure changes, compared to the fixed threshold method. The dynamic threshold mechanism based on Gaussian mixture model and entropy algorithm can distinguish between minor anomalies and emergency faults, achieving graded early warning and effectively avoiding the over-alarm problem caused by traditional binary threshold judgment, thus improving the accuracy of intelligent monitoring of boiler operating status. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of an intelligent monitoring method for boiler operating status according to the present invention; Figure 2 This is a schematic diagram of a second embodiment of an intelligent monitoring method for boiler operating status according to the present invention; Figure 3 This is a schematic diagram of the first embodiment of an intelligent monitoring system for boiler operating status according to the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, an intelligent monitoring method for boiler operating status includes the following steps: Step 101: Collect multi-source heterogeneous data during boiler operation, clean and normalize the multi-source heterogeneous data, and extract key features through wavelet transform to obtain standard multi-source heterogeneous data. Specifically, in this embodiment, multi-source heterogeneous data is collected during boiler operation, including at least operating parameter data, image data, environmental data, and equipment status data. Missing values ​​in the linear interpolation numerical data are processed, blurry and occluded data in the image data are removed, and abnormal values ​​in the complete data are identified based on the 3σ criterion and replaced with the mean value of the same time period. Invalid images are filtered through image quality assessment to obtain complete data. The numerical data and image data are synchronized according to timestamps, and the complete data is normalized using Min-Max normalization to obtain normalized data. The normalized data is decomposed into 3-5 levels of wavelet decomposition to obtain approximation coefficients and detail coefficients. The approximation coefficients and the first two levels of detail coefficients are retained, and key features are screened through variance contribution rate to obtain standard multi-source heterogeneous data.

[0020] Data types must cover four core categories of information: operating parameter data, which directly reflects the boiler's operating conditions, including steam pressure, main steam pressure, reheat steam pressure, medium temperature, furnace temperature, superheater outlet temperature, economizer inlet and outlet temperatures, working fluid flow rate, feedwater flow rate, steam flow rate, drum water level, and flue gas composition, including O2, CO, and NO. xConcentration and other data are directly related to combustion efficiency and safety thresholds; image data focuses on the internal state of the furnace, with RGB images of the furnace flame used to analyze combustion stability, such as flame shape and flicker frequency, while infrared thermal imaging can quantify the temperature field distribution inside the furnace, such as local high-temperature zones and temperature gradients; environmental data needs to be collected simultaneously with the boiler room ambient temperature, relative humidity, and atmospheric pressure to correct for the impact of environmental fluctuations on boiler thermal efficiency; equipment status data includes the vibration acceleration of the induced draft fan and the forced draft fan, horizontal / vertical direction, operating noise decibel value, and motor operating current, used to assess the health status of auxiliary equipment.

[0021] The selection and frequency settings of the data acquisition equipment must match the data characteristics: Industrial-grade high-precision sensors for parameters such as pressure and temperature should be used, with an accuracy of ±0.1%FS, installed at key monitoring points such as the top of the steam drum and the flue outlet. The sampling frequency should be set to 1-10Hz, with the upper limit for rapidly changing furnace temperature and the lower limit for slowly changing steam drum water level. A 2-megapixel industrial camera equipped with a high-temperature resistant protective cover and an 850nm narrowband filter should be used for furnace flame image acquisition, with a frame rate of 10-20fps to capture dynamic changes in the flame. An infrared thermal imager with a resolution of 640×512 and an uncooled focal plane detector should be used, with a temperature range of -20℃ to 1500℃ and a frame rate of 5-10fps to balance accuracy and storage costs. A piezoelectric accelerometer should be used for vibration sensors, with a sampling frequency of 50-100Hz to capture high-frequency vibration signals. All equipment must be time-synchronized via a time synchronization server using the NTP protocol to calibrate timestamps, with an error controlled within 10ms. Data storage adopts a hierarchical architecture: numerical data is stored in a time-series database, such as InfluxDB, in CSV format. Each record contains the device ID, acquisition time, accurate to milliseconds, and parameter values. Image data is stored in a distributed file system, such as HDFS, in JPG / PNG format. The file name is associated with the corresponding timestamp and camera number to ensure data traceability.

[0022] Data preprocessing is crucial for improving feature extraction accuracy, requiring the construction of a standardized dataset through cleaning, normalization, and feature transformation. Data cleaning is implemented in three steps: For missing value handling, numerical data employs a strategy based on the duration of missing values—linear interpolation is used for missing values ​​< 5s, fitting data based on the preceding and following 10 time points; cubic spline interpolation is used for missing values ​​5-30s, preserving data trends. Image data blurred due to lens occlusion or overexposure is directly removed and marked as abnormal periods. Outlier handling identifies outliers using the 3σ criterion, calculating the feature mean μ and standard deviation σ. Values ​​exceeding the range [μ-3σ, μ+3σ] are considered abnormal and replaced with the mean of the same time period, such as the sliding window mean of the previous hour. Image data is filtered for invalid samples using two criteria: sharpness (Laplacian gradient > 50) and contrast (grayscale standard deviation > 30). Data alignment uses interpolation matching based on timestamps, mapping image data (maximum frame rate 20fps) and numerical data (maximum 10Hz) to a 10Hz time axis, ensuring that multi-source data from the same time point form related samples.

[0023] Data normalization requires eliminating dimensional differences: Numerical data uses Min-Max normalization, calculating the minimum value x_min and maximum value x_max of each feature based on historical normal operation data, and mapping it to the [0,1] interval through x_norm=(x-x_min) / (x_max-x_min) to avoid interference from extreme values, such as normalizing the furnace temperature of 300-1500℃ to 0-1; Image data is first normalized by pixel value, divided by 255, and then unified in size through center cropping and resizing, with flame images being 256×256 pixels, and infrared thermal images retaining temperature gradient information. Wavelet transform is used to extract multi-scale features from numerical data: the db4 wavelet basis is selected, which combines temporal locality and frequency resolution. The normalized time series data is decomposed into 3-5 levels, with the number of levels adjusted according to the data fluctuation characteristics. For example, pressure data is decomposed into 3 levels, and vibration data into 5 levels. This yields approximate coefficients reflecting the overall trend, low-frequency and detail coefficients reflecting abrupt changes, and high-frequency data. Key features are selected by variance contribution rate, and the variance proportion of each coefficient is calculated. Approximate coefficients with a cumulative contribution rate ≥85% and the detail coefficients of the first two levels are retained. Finally, the data is combined into standard multi-source heterogeneous data containing low-frequency trend and high-frequency abrupt change information, laying the foundation for subsequent feature modeling.

[0024] Step 102: Model the time series data in the standard multi-source heterogeneous data based on the LSTM long short-term memory network, capture the dynamic change trend of boiler operation, obtain time series features, and use CNN convolutional neural network to analyze furnace flame images and infrared thermal imaging to extract combustion state features and heat distribution features. Specifically, in this embodiment, the numerical data in the standard multi-source heterogeneous data is divided into a training set, a validation set, and a test set. A sliding window method is used to construct the input sequence, with a window size of 30, and the output is a single-time feature vector. This single-time feature vector is input into an LSTM network, and the network parameters are adjusted using the validation set. Training stops when the validation set loss does not decrease for five consecutive rounds, and the time-series feature vector is output. Histogram equalization is used to improve image contrast, Gaussian filtering is used to remove noise, and the temperature values ​​of the thermal imaging image are mapped to grayscale values ​​while preserving temperature gradient information. The processed image data is input into the fusion branch of the flame image and the thermal imaging image in a CNN convolutional neural network to obtain the combustion state feature vector and the heat distribution feature vector.

[0025] Step 103: Use the Attention mechanism to weightedly fuse temporal features, combustion state features, and heat distribution features to generate a comprehensive state vector; Specifically, in this embodiment, the query vector is initialized, and the dot product attention mechanism is used to calculate the similarity between the query vector and the temporal feature vector and the image feature vector, respectively. The fusion weight is obtained by normalization through the Softmax function. The temporal features, combustion state features and heat distribution features are weighted and fused according to the fusion weight to obtain the 3D comprehensive state vector.

[0026] 1LSTM time-series data modeling is a core step in capturing the dynamic trends of boiler operation, requiring meticulous data preprocessing and network design to uncover temporal features. Time-series data preprocessing must balance data integrity and temporal correlation: data partitioning employs a time-series-specific strategy, splitting the training, validation, and test sets in a 7:2:1 ratio, strictly adhering to the chronological order. For example, the first 70% of the time period is the training set, the middle 20% is the validation set, and the last 10% is the test set, avoiding future data leakage issues caused by random partitioning and ensuring the model learns true temporal dependencies. Sequence construction is based on the sliding window method, with a window size of 30 time points, combined with a 10Hz sampling frequency corresponding to a 3-second duration. This covers both short-term fluctuations in boiler parameters, such as instantaneous pressure changes, and captures fundamental trends, such as slow temperature increases. The input consists of multi-dimensional data from the 30 time points within the window, such as pressure, temperature, and flow rate, and the output is the feature vector at the end of the window, forming a historical-current mapping relationship. Adjusting the data dimension requires adapting to the LSTM input format. The original two-dimensional data, sample number × feature number, is reshaped into a three-dimensional tensor [sample number, time step, 30, feature dimension, such as 12-dimensional running parameters] to ensure that the network can process both time series length and feature dimension information simultaneously.

[0027] The construction of the LSTM network needs to balance feature extraction capability and generalization performance: The network adopts a two-layer stacked structure. The first LSTM layer has 64 hidden units and the activation function tanh to capture local dynamic features in the original time series data, such as pressure pulse fluctuations. A Dropout layer is introduced with a dropout rate of 0.2 to randomly deactivate some neurons, suppressing overfitting, especially for noisy patterns in the training set. The second LSTM layer has 32 hidden units to further refine high-level time series patterns, such as the lag effect of temperature changing with load. Finally, a fully connected layer compresses the features to 16 dimensions, preserving key information while reducing the difficulty of subsequent fusion. The training parameters are specifically optimized: the Adam optimizer with a learning rate of 0.001 is suitable for handling non-stationary time series data and can adaptively adjust the learning step size; the mean squared error (MSE) loss function focuses on the accuracy of parameter prediction, driving the model's ability to learn trend fitting; 50 iterations with a batch size of 32 ensure training efficiency while making gradient updates more stable. The model validation adopts an early stopping strategy: training is terminated when the validation set loss does not decrease for 5 consecutive iterations to avoid overfitting and save the optimal model. The final output 16-dimensional time-series feature vector contains dynamic information such as pressure change rate, temperature fluctuation period, and flow coupling relationship, providing a quantitative basis for boiler operation trend analysis.

[0028] The 2CNN image feature extraction uses a dual-branch network to analyze the visual information of the furnace, accurately capturing the combustion state and heat distribution characteristics. Preprocessing and network design must be adapted to the specific characteristics of the image data. Image preprocessing focuses on quality improvement and sample expansion: Flame image enhancement employs global histogram equalization, stretching the grayscale value distribution range to increase the grayscale difference between the high-temperature core area and the low-temperature edge area of ​​the flame by more than 30%, strengthening the flame outline and brightness gradient; 3×3 Gaussian filtering with a standard deviation of 0.8 effectively filters out dust noise and sensor noise in the industrial environment while preserving the subtle dynamics of flame flickering. Thermal imaging processing needs to consider both temperature quantization and visual features. First, the original temperature values ​​(-20℃~1500℃) are linearly mapped to grayscale values ​​of 0~255, using the formula: grayscale value = 255×(T-T_min) / (T_max-T_min). Then, Jet pseudo-color annotation is used, with red-yellow-blue corresponding to high-medium-low temperatures, enhancing the spatial distribution characteristics of the temperature field through color gradients. Data augmentation employs diverse transformations on the training set: random horizontal / vertical flipping to simulate camera installation angle deviation, ±10° rotation to adapt to asymmetric combustion of furnace flames, and ±5 pixel translation to compensate for image shift caused by fan vibration, thereby expanding the sample size to twice the original and significantly improving the model's generalization ability.

[0029] The CNN dual-branch network design needs to match the feature attributes of different images: The flame image branch uses 3-channel RGB input to focus on extracting dynamic combustion features. The first layer uses 32 3×3 convolutional kernels to capture local textures, such as the jagged shape of the flame edges. After dimensionality reduction via 2×2 max pooling, the second layer uses 64 convolutional kernels to extract the overall flame contour, such as the conical shape of stable combustion and the scattered shape of unstable combustion. The third layer uses 128 convolutional kernels to deeply analyze the flame brightness distribution, the difference between the high-temperature brightness of the core area and the oxidation brightness of the edges. Finally, global average pooling replaces the traditional fully connected layer, reducing parameters by 70% and outputting 8-dimensional combustion state features, including indicators such as flame stability and combustion completeness. The thermal imaging branch uses 1-channel grayscale input to focus on temperature field features. The network structure is consistent with the flame branch. It extracts local high-temperature areas step by step through convolutional layers, such as potential coking points and temperature gradients, such as the heat transfer efficiency of the heated surface and the global thermal balance state, outputting 8-dimensional thermal distribution features. Training parameters were optimized for image data: the Adam optimizer learning rate was set to 0.0001, lower than LSTM, to avoid drastic fluctuations in image features. The cross-entropy loss function, combined with manually labeled combustion state tags (e.g., normal, partially burned, locally high temperature), drove parameter updates. Thirty iterations with a batch size of 16 balanced training time and accuracy, ensuring the validation set accuracy remained stable above 92%. Finally, the outputs of the two branches were concatenated into a 16-dimensional joint feature vector, enabling the visualization and quantification of combustion state and heat distribution features.

[0030] Attention-based weighted fusion is a core component for achieving efficient collaboration between temporal and image features. It highlights key information through dynamic weight allocation, providing accurate comprehensive feature support for state assessment. Feature standardization is a prerequisite for fusion. Since the 16-dimensional temporal features output by LSTM (e.g., pressure fluctuation period, temperature change rate) and the 16-dimensional image features output by CNN (e.g., flame stability, heat distribution uniformity) have heterogeneous origins and significant dimensional differences, and the temporal features are standardized trend coefficients while the image features are convolutional feature values, Z-score standardization must be used to eliminate interference. Specifically, based on historical normal operation data, the mean μ and standard deviation σ of each type of feature are calculated. The formula x_std=(x-μ) / σ is used to map all features to a standard distribution with a mean of 0 and a variance of 1, ensuring that all features are equally competitive in subsequent weight calculations and avoiding feature weight bias due to differences in numerical ranges.

[0031] Weight calculation and feature fusion employ a dot product attention mechanism for dynamic adaptation: The query vector, Query, is initialized as a 1×32-dimensional random vector with a value range of [-0.1, 0.1]. This vector is not a fixed value but is used to construct a joint training framework with the subsequent GMM model, continuously optimizing it through backpropagation to adapt to the boiler's operating characteristics. During attention score calculation, the time-series feature vector V_t (16 dimensions) and the image feature vector V_i (16 dimensions) are first padded with zeros to 32 dimensions, and then a dot product operation is performed with the query vector to obtain two original scores, S1 and S2. The score directly reflects the correlation between the feature and the current monitoring task; for example, under high load conditions, the heat distribution feature score is higher. The Softmax function is used to normalize S1 and S2 to obtain weights ω1 and ω2, satisfying ω1 + ω2 = 1, achieving adaptive enhancement of key features. The fusion phase performs element-wise weighted summation using the formula V=ω1×V_t+ω2×V_i to generate a 32-dimensional comprehensive state vector. This vector incorporates both the dynamic trends of the time-series data and the operational details of the visual data. Model optimization is achieved through end-to-end training. The output of the fusion module is directly input into the GMM model, using the deviation error of the state assessment as the objective function. Backpropagation is used to synchronously adjust the query vector parameters and weight allocation logic, improving the information utilization rate of the fused features by over 40% and significantly enhancing the accuracy of subsequent assessments.

[0032] Step 104: Based on the comprehensive state vector, calculate the confidence interval of each feature dimension using the Gaussian mixture model (GMM), and evaluate the deviation of the operating state by combining entropy analysis. When the deviation exceeds the preset threshold, trigger a graded warning.

[0033] Specifically, in this embodiment, a comprehensive state vector under normal boiler operation is collected as training data for the Gaussian Mixture Model (GMM). The number of mixed components is set to K=3-5. The EM algorithm is used to estimate the mean, covariance matrix, and mixing coefficients of each component. Based on the trained GMM, a 95% confidence interval is calculated for each feature dimension, and the information entropy is calculated for each dimension of the real-time comprehensive state vector. A deviation is defined, with D≥1 indicating a deviation. The comprehensive state vector, entropy value, and deviation D of boiler operation are calculated in real time. When 1.2≤D<1.5: mild warning; when 1.5≤D<2.0: moderate warning; when D≥2.0: severe warning.

[0034] The combination of Gaussian Modeling (GMM) and entropy analysis enables quantitative assessment and precise early warning of boiler operating status. A closed-loop monitoring system is constructed through statistical modeling and information theory analysis. GMM model training focuses on normal operating data. Training data requires collecting comprehensive state vectors of the boiler under different loads (50%-100% rated load) and environmental conditions, such as low temperatures in winter and high temperatures in summer. The sample size must be no less than 1000 sets, and the data must undergo deduplication and noise reduction preprocessing to ensure coverage of all normal operating modes. During parameter estimation, the number of mixture components K is determined using the Bayesian information criterion (BIC). The BIC value is calculated when K = 3-5: BIC = ln(n)k - 2ln(L), where n is the number of samples, k is the number of parameters, and L is the likelihood function value. The K value with the smallest BIC is selected; typically, K = 4 best fits the multimodal normal state of the boiler. The expectation-maximization (EM) algorithm is used to iteratively estimate the mean, covariance matrix, and mixing coefficients of each component. The iteration stops when the change in the log-likelihood function is less than 1e-6, ensuring that the parameters converge to the optimal solution. The confidence interval calculation was carried out separately for the 32-dimensional comprehensive features. Based on the mean μ and standard deviation σ of each dimension in the normal samples, the interval [μ-1.96σ, μ+1.96σ] was determined at a 95% confidence level to provide a benchmark threshold for subsequent deviation judgment.

[0035] Entropy analysis enables the quantification and graded early warning of deviation: When calculating feature entropy, the probability density p_ij of each dimension of the real-time comprehensive state vector is first calculated using a trained GMM model, reflecting the probability that the feature of that dimension is in a normal state. Then, the uncertainty is calculated using the information entropy formula H_j=-Σp_ij×ln(p_ij). The larger H_j is, the higher the probability that the dimension deviates from the normal state. The calculation of deviation D introduces feature weight w_j, which is determined by the variance contribution rate of each dimension. The higher the contribution rate, the greater the weight. For example, the weight of the dimension reflecting furnace temperature can reach 0.15. Combined with the baseline entropy value H_j0 under normal conditions, it is calculated from historical normal samples and the global deviation is quantified using the formula D=Σ(H_j / H_j0)×w_j. D≥1 indicates that there is a deviation in operating conditions. The tiered early warning thresholds are set based on historical fault data statistics: D1=1.2 corresponds to a mild warning, such as parameter fluctuations exceeding the normal range but not affecting safety, triggering a system pop-up prompting inspection personnel to focus on checking sensor status; D2=1.5 corresponds to a moderate warning, such as uneven combustion leading to uneven heat distribution, activating audible and visual alarms and pushing parameter adjustment suggestions, such as adjusting the air supply volume; D3=2.0 corresponds to a severe warning, such as local overheating potentially causing pipe rupture, automatically triggering a load shedding procedure and sending an emergency shutdown notification to maintenance personnel. After an early warning is triggered, the system automatically records the comprehensive feature vector at the time of the warning, the original monitoring data, and the high-contribution deviation dimension, generating an early warning report including fault location suggestions, reducing fault investigation time by more than 60%.

[0036] Its beneficial effects lie in overcoming the limitations of traditional single-point monitoring methods in adapting to complex operating conditions by integrating time-series, image, and equipment status data, and utilizing an attention mechanism to weightedly fuse key features. CNN analysis of furnace flame images can accurately capture combustion anomalies, and combined with LSTM dynamic modeling of time-series data such as pressure and temperature, the accuracy of fault identification is improved compared to traditional methods. The online incremental learning strategy combined with transfer learning technology enables the model to respond to changes in boiler operating conditions in real time, avoiding false alarms caused by historical data bias. The early warning response time is shortened in scenarios of sudden steam pressure changes, compared to the fixed threshold method. The dynamic threshold mechanism based on Gaussian mixture model and entropy algorithm can distinguish between minor anomalies and emergency faults, achieving graded early warning and effectively avoiding the over-alarm problem caused by traditional binary threshold judgment, thus improving the accuracy of intelligent monitoring of boiler operating status.

[0037] Please see Figure 2 In a method for intelligent monitoring of boiler operating status, multi-source heterogeneous data is collected during boiler operation. This data is then cleaned and normalized, and key features are extracted using wavelet transform to obtain standard multi-source heterogeneous data. The process includes the following steps: Step 201: Collect multi-source heterogeneous data during boiler operation, including at least operating parameter data, image data, environmental data, and equipment status data; Step 202: Process missing values ​​in linear interpolation numerical data, remove blurry and occluded data in image data, identify abnormal values ​​in complete data based on the 3σ criterion, replace them with the mean of the same time period, filter invalid images through image quality assessment, and obtain complete data; Step 203: Synchronize numerical data and image data according to timestamps, and normalize the complete data using Min-Max normalization to obtain normalized data; Step 204: Perform 3-5 level wavelet decomposition on the normalized data to obtain approximation coefficients and detail coefficients. Retain the approximation coefficients and the first 2 level detail coefficients. Filter key features through variance contribution rate to obtain standard multi-source heterogeneous data.

[0038] The above describes an embodiment of the intelligent monitoring method for boiler operating status according to the present invention. Please refer to [link / reference]. Figure 3 In a boiler operation status intelligent monitoring system, the boiler operation status intelligent monitoring system includes the following modules: The multi-source data acquisition module is used to collect multi-source heterogeneous data during boiler operation, clean and normalize the multi-source heterogeneous data, and extract key features through wavelet transform to obtain standard multi-source heterogeneous data. The feature vector extraction module is used to model time-series data in standard multi-source heterogeneous data based on LSTM long short-term memory network, capture the dynamic change trend of boiler operation, obtain time-series features, and use CNN convolutional neural network to analyze furnace flame images and infrared thermal imaging to extract combustion state features and heat distribution features. The feature vector fusion module is used to weightedly fuse temporal features, combustion state features, and heat distribution features through the Attention mechanism to generate a comprehensive state vector; The deviation-level early warning module is used to calculate the confidence interval of each feature dimension based on the comprehensive state vector through the Gaussian mixture model (GMM), and to evaluate the deviation of the operating state by combining entropy value analysis. When the deviation exceeds the preset threshold, a level warning is triggered.

[0039] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent monitoring of boiler operating status, characterized in that, The intelligent monitoring method for boiler operating status includes the following steps: Multi-source heterogeneous data is collected during boiler operation. The multi-source heterogeneous data is cleaned and normalized. Key features are extracted by wavelet transform to obtain standard multi-source heterogeneous data. The time series data in the standard multi-source heterogeneous data is modeled based on the LSTM long short-term memory network to capture the dynamic change trend of boiler operation and obtain time series features. The combustion state features and heat distribution features are extracted by analyzing furnace flame images and infrared thermal imaging using CNN convolutional neural network. A comprehensive state vector is generated by weightedly fusing temporal features, combustion state features, and heat distribution features through an attention mechanism. Based on the comprehensive state vector, the confidence interval of each feature dimension is calculated using the Gaussian mixture model (GMM), and the deviation of the operating state is evaluated by combining entropy analysis. When the deviation exceeds a preset threshold, a graded warning is triggered.

2. The intelligent monitoring method for boiler operating status as described in claim 1, characterized in that, The process involves collecting multi-source heterogeneous data during boiler operation, cleaning and normalizing the data, and extracting key features using wavelet transform to obtain standard multi-source heterogeneous data, including: Collect multi-source heterogeneous data during boiler operation, including at least operating parameter data, image data, environmental data, and equipment status data; Missing values ​​in linear interpolation numerical data are processed to remove blurry and occluded data in image data. Abnormal values ​​in complete data are identified based on the 3σ criterion and replaced with the mean value of the same time period. Invalid images are filtered through image quality assessment to obtain complete data. Numerical and image data are synchronized according to timestamps, and the complete data is normalized using Min-Max normalization to obtain normalized data. Normalized data is decomposed into 3-5 levels of wavelet decomposition to obtain approximation coefficients and detail coefficients. The approximation coefficients and the first two levels of detail coefficients are retained. Key features are selected by variance contribution rate to obtain standard multi-source heterogeneous data.

3. The intelligent monitoring method for boiler operating status as described in claim 1, characterized in that, The LSTM (Long Short-Term Memory) network is used to model the time-series data in the standard multi-source heterogeneous data, capturing the dynamic changes in boiler operation and obtaining time-series features. A CNN (Convolutional Neural Network) is then used to analyze furnace flame images and infrared thermal imaging to extract combustion state features and heat distribution features, including: Numerical data in standard multi-source heterogeneous data is divided into training set, validation set and test set. The input sequence is constructed using the sliding window method, with the window size set to 30. The output is a single-time feature vector. The single-time feature vector is input into the LSTM network, and the network parameters are adjusted using the validation set. Training stops when the validation set loss does not decrease for 5 consecutive rounds, and the time-series feature vector is output.

4. The intelligent monitoring method for boiler operating status as described in claim 1, characterized in that, The LSTM (Long Short-Term Memory) network is used to model the time-series data in the standard multi-source heterogeneous data, capturing the dynamic changes in boiler operation and obtaining time-series features. A CNN (Convolutional Neural Network) is then used to analyze furnace flame images and infrared thermal imaging to extract combustion state features and heat distribution features, including: Histogram equalization is used to improve image contrast, Gaussian filtering is used to remove noise, and the temperature values ​​of the thermal image are mapped to gray values ​​to preserve temperature gradient information. The processed image data is input into the fusion flame image branch and thermal imaging branch of the CNN convolutional neural network to obtain the combustion state feature vector and heat distribution feature vector.

5. The intelligent monitoring method for boiler operating status as described in claim 1, characterized in that, The process of generating a comprehensive state vector by weightedly fusing temporal features, combustion state features, and heat distribution features through an attention mechanism includes: Initialize the query vector, use the dot product attention mechanism to calculate the similarity between the query vector and the temporal feature vector and the image feature vector respectively, and obtain the fusion weight by normalization through the Softmax function; Based on the fusion weights, fusion time sequence characteristics, combustion state characteristics, and heat distribution characteristics, a 3D integrated state vector is obtained.

6. The intelligent monitoring method for boiler operating status as described in claim 1, characterized in that, Based on the comprehensive state vector, the confidence intervals of each feature dimension are calculated using a Gaussian mixture model (GMM), and the deviation of the operating state is evaluated by combining entropy analysis. When the deviation exceeds a preset threshold, a graded warning is triggered, including: Collect the comprehensive state vector under normal boiler operation as training data for GMM. Set the number of mixed components K=3-5 and use the EM algorithm to estimate the mean, covariance matrix and mixing coefficient of each component. Based on the trained GMM, a 95% confidence interval is calculated for each feature dimension, and the information entropy is calculated for each dimension of the real-time integrated state vector. The deviation is defined, with D≥1 indicating the existence of deviation.

7. The intelligent monitoring method for boiler operating status as described in claim 1, characterized in that, Based on the comprehensive state vector, the confidence intervals of each feature dimension are calculated using a Gaussian mixture model (GMM), and the deviation of the operating state is evaluated by combining entropy analysis. When the deviation exceeds a preset threshold, a graded warning is triggered, including: The comprehensive state vector, entropy value, and deviation D of boiler operation are calculated in real time. When 1.2≤D<1.5: mild warning; when 1.5≤D<2.0: moderate warning; when D≥2.0: severe warning.

8. A boiler operating status intelligent monitoring system, characterized in that, The intelligent monitoring system for boiler operation status includes the following modules: The multi-source data acquisition module is used to collect multi-source heterogeneous data during boiler operation, clean and normalize the multi-source heterogeneous data, and extract key features through wavelet transform to obtain standard multi-source heterogeneous data. The feature vector extraction module is used to model the time series data in the standard multi-source heterogeneous data based on the LSTM long short-term memory network, capture the dynamic change trend of boiler operation, obtain time series features, and use CNN convolutional neural network to analyze furnace flame images and infrared thermal imaging to extract combustion state features and heat distribution features. The feature vector fusion module is used to weightedly fuse temporal features, combustion state features, and heat distribution features through the Attention mechanism to generate a comprehensive state vector; The deviation-level early warning module is used to calculate the confidence interval of each feature dimension based on the comprehensive state vector using the Gaussian mixture model (GMM), and to evaluate the deviation of the operating state by combining entropy value analysis. When the deviation exceeds a preset threshold, a level warning is triggered.

9. The intelligent monitoring system for boiler operating status as described in claim 8, characterized in that, The deviation-level early warning module includes the following sub-modules: The estimation submodule is used to collect the comprehensive state vector under normal boiler operation as training data for GMM. The number of mixed components is set to K=3-5, and the EM algorithm is used to estimate the mean, covariance matrix and mixing coefficient of each component. Define a submodule for calculating the 95% confidence interval for each feature dimension based on the trained GMM, calculating the information entropy for each dimension of the real-time integrated state vector, and defining the deviation degree, where D≥1 indicates the existence of deviation.

10. The intelligent monitoring system for boiler operating status as described in claim 8, characterized in that, The deviation-level early warning module includes the following sub-modules: The calculation submodule is used to calculate the comprehensive state vector, entropy value and deviation D of boiler operation in real time. When 1.2≤D<1.5: mild warning, when 1.5≤D<2.0: moderate warning, when D≥2.0: severe warning.

Images