Flame image combustion stability intelligent identification method and system

Abstract

The present application relates to the field of combustion diagnosis and image intelligent analysis, and provides a kind of flame image combustion stability intelligent identification method and system, with the flame image sequence of combustion process as data source, utilize multi-scale convolution self-encoding to extract the deep feature of flame image, and introduce adaptive structure similarity index to cluster deep feature, realize the adaptive division of combustion state, construct flame dynamic fluctuation index on time domain through dynamic permutation entropy, to quantitatively characterize the complexity and stability of combustion timing, the deep feature sequence and timing stability label are jointly input into the model of fusion long short-term memory network and timing attention mechanism, realize the trend prediction of combustion stability.The present application can realize non-contact, real-time combustion stability identification only by relying on visible light image, with high precision, strong robustness and deployability, can carry out online monitoring and early warning to the combustion process of combustion device, and provide reliable technical support for energy efficiency optimization and safe operation.

Description

Technical Field

[0001] This invention relates to the field of combustion diagnosis and intelligent image analysis technology, and in particular to an intelligent identification method and system for combustion stability of flame images. Background Technology

[0002] In large-scale combustion devices (such as boilers, gas turbines, industrial kilns, etc.), combustion stability is directly related to the system's thermal efficiency, safety, and pollution emission levels. Traditional combustion stability monitoring mainly relies on sensor measurement methods such as pressure pulsation, ion current, or acoustic signals. These methods have the following shortcomings: (1) Response lag and insufficient spatial resolution: Pressure and acoustic signals only reflect local airflow disturbances and cannot accurately depict the overall flame morphology changes, making it difficult to achieve multi-dimensional combustion status identification; (2) Susceptible to noise and interference: In high-temperature and high-noise environments, sensor signals are easily distorted, affecting the reliability of stability diagnosis; (3) Reliance on manual experience and threshold setting: Traditional methods usually require manual setting of characteristic thresholds or spectral boundaries, lacking adaptive capabilities and making it difficult to cope with dynamic changes in different fuels, loads, and combustion conditions.

[0003] In recent years, combustion state recognition methods based on flame images have gradually attracted attention. Flame images can intuitively reflect the spatial distribution and dynamic characteristics of the combustion process, and have advantages such as rich information, strong visualization, and non-contact operation. However, existing combustion state recognition methods based on flame images still have certain limitations. For example, most methods only focus on the static features of the image (such as brightness, texture, and morphology), lacking modeling of the temporal evolution of the combustion process; feature extraction often relies on manual design or single-scale convolutional networks, making it difficult to take into account both the local details and global structural information of the flame; and temporal modeling methods generally fail to fully consider the importance of features at different time steps, resulting in insufficient prediction accuracy and stability. Summary of the Invention

[0004] The present invention aims to solve at least one of the problems existing in the prior art, and provides a method and system for intelligent identification of combustion stability in flame images. It can integrate multi-scale spatial features and time-series dynamic information of flame images to achieve high-precision, real-time and interpretable identification of combustion stability, and provide support for combustion optimization control and safe operation.

[0005] One aspect of the present invention provides a method for intelligent recognition of combustion stability in flame images, the method comprising the following steps: Step S1: Acquire a sequence of flame images during the combustion process, and preprocess each flame image in the sequence; Step S2: Based on the preprocessed flame image, construct image samples at multiple time scales using a sliding window, and determine the optimal window size based on the sample entropy of the image data within the sliding window; Step S3: Based on the sliding window corresponding to the optimal window size, a multi-scale convolutional autoencoder structure is used to extract deep features from the preprocessed flame image, and a new loss function is set for the multi-scale convolutional autoencoder structure. Step S4: Introduce an adaptive structural similarity index to perform cluster analysis on the extracted deep features, forming two independent clusters; Step S5: Calculate the dynamic permutation entropy of two independent clusters based on the time series window. Define the cluster with dynamic permutation entropy less than a set threshold as a stable combustion cluster, and define the cluster with dynamic permutation entropy greater than a set threshold as an unstable combustion cluster. Step S6: Calculate the centroids of stable and unstable combustion clusters respectively, and calculate the combustion stability probability index based on the extracted deep features to quantify the probability of instability in the flame image. Step S7: Input the extracted deep features and their corresponding combustion stability probabilities into the fusion model of long short-term memory network and temporal attention mechanism to perform combustion stability identification and trend prediction.

[0006] Optionally, in step S1, preprocessing is performed on each flame image in the flame image sequence, including: A noise reduction algorithm combining Gaussian filtering and median filtering is adopted. The pixel intensity of each flame image is mapped to the [0, 1] interval by linear normalization, and data enhancement processing is performed on each flame image by random rotation, mirroring and brightness perturbation.

[0007] Optionally, step S2 specifically includes the following steps: Step S2.1: Initialize the size of a series of sliding windows and step length ; Step S2.2: Based on the preprocessed flame image, determine each and The combination of flame image time series within the corresponding sliding window is calculated and recorded according to the following formula. and The sample entropy of the flame image time series within the corresponding sliding window is: ; in, This represents the sample entropy of the flame image time series within the sliding window. The embedding dimension during phase space reconstruction is . When the distance between two phase space vectors reconstructed from the flame image time series within the same sliding window, but with different start times, is less than a distance threshold. The probability of pattern matching. The embedding dimension during phase space reconstruction is . When the distance between two phase space vectors reconstructed from the flame image time series within the same sliding window, but with different start times, is less than a distance threshold. The probability of pattern matching; Step S2.3: Plot the sample entropy as a function of... and A two-dimensional surface plot of combined variations, with a threshold set. Find samples whose entropy is greater than a threshold. The minimum sliding window is used as the optimal window size, and the size and step size of the minimum sliding window are used as the optimal window size.

[0008] Optionally, the multi-scale convolutional autoencoder structure includes multiple convolutional encoders and multiple convolutional decoders connected in sequence, with the last convolutional encoder connected to the first convolutional decoder. In step S3, based on the sliding window corresponding to the optimal window size, a multi-scale convolutional autoencoder structure is used to extract deep features from the preprocessed flame image, including the following steps: Step S3.1: Use a sliding window with the optimal window size to convert the preprocessed flame image into matrix data as the input image for the multi-scale convolutional autoencoder structure. Let the input image The dimension is O×P , O For the input image height, P For the input image The width; Step S3.2: Input image x The data is sent to a convolutional encoder for processing. Each convolutional encoder uses several windows of size [missing information]. c×c Step size is q The convolutional filter extracts features from the input, and then... ReLU The activation function performs feature activation and uses a window size of [missing value]. o×o Step size is The max-pooling layer performs feature dimensionality reduction and outputs the corresponding deep features. The deep features output by the previous convolutional encoder serve as the input to the next convolutional encoder, and the input to the first convolutional encoder is the input image. x ; Step S3.3: The deep features output by the last convolutional encoder are sent to the convolutional decoder for processing. All convolutional decoders except the last one use a window size of... g×g The upsampling layer upscales the input features, and then sequentially applies convolutional filters, ReLU The activation function processes the data and outputs the corresponding deep features. The final convolutional decoder uses a window size of... g×g The upsampling layer upscales the input features, and then sequentially applies convolutional filters, Sigmoid The function processes the data, and the output dimension is... O×P Reconstructed image x rec In this process, the deep features output by the previous convolutional decoder are used as the input to the next convolutional decoder, and the input to the first convolutional decoder is the deep features output by the last convolutional encoder.

[0009] Alternatively, the novel loss function for the multi-scale convolutional autoencoder structure is expressed as: ; in, This represents the total loss of the multi-scale convolutional autoencoder structure. Indicates the first loss term and , Indicates the second loss term and , Indicates the third loss item and , Indicates the input image. Indicates the reconstructed image. O Indicates the height of the input image. P Indicates the width of the input image. and These represent the input image and the reconstructed image, respectively. The value at the location, Indicates the image height index. Indicates the image width index. and represent the mean values ​​of the input image and the reconstructed image, respectively. and These represent the variances of the input image and the reconstructed image, respectively. This represents the covariance between the input image and the reconstructed image. and All are constants. , , They represent The weighting coefficients.

[0010] Optionally, the adaptive structural similarity index is labeled as Its numerical range is [0,1], and it is specifically defined as: ; in, Represents the luminance component. Represents the contrast component. Represents structural components, , , They represent The weights, and , , Adaptive optimization using genetic algorithms can improve clustering accuracy. Indicates the first Deep features of the flame image within a sliding window. Indicates the first Deep features of the flame image within a sliding window; Step S4 specifically includes the following steps: Step S4.1: Randomly select two deep features extracted from the multi-scale convolutional autoencoder structure as the initial centroids of the two clusters; Step S4.2: Calculate the adaptive structural similarity index between all deep features extracted by the multi-scale convolutional autoencoder structure and the two initial centroids, and reassign the deep features extracted by the multi-scale convolutional autoencoder structure to the two clusters accordingly; Step S4.3: Calculate the mean of all deep features in each dimension for the two clusters respectively, and use the calculated mean as the centroid of the corresponding cluster; Step S4.4: Repeat steps S4.2 and S4.3 until the centroids of the two clusters no longer change or the limit number of updates is reached, and finally two independent clusters are obtained.

[0011] Optionally, the formula for calculating the dynamic permutation entropy in step S5 is: ; in, For time window Inner The probability of occurrence of a pixel gradient pattern; Indicates the number of pixel gradient modes; For time window The permutation entropy within a time window is used to characterize the time window. Internal combustion fluctuations, when Exceeding the threshold At that time, determine the time window The combustion state inside is unstable.

[0012] Optionally, step S6 specifically includes: Calculate the average value of all deep features in each dimension in the stable combustion cluster and the unstable combustion cluster respectively, and use the calculated average value as the centroid of the corresponding stable combustion cluster and unstable combustion cluster; The combustion stability probability index is calculated using the following formula: ; in, For the first The deep feature vector is composed of the deep features of each sliding window. , These are the centroids of the stable combustion cluster and the unstable combustion cluster, respectively. For the first The probability of combustion stability in a sliding window. The numerical range is 0 to 1, where 0 represents stable combustion, 1 represents unstable combustion, and 0.5 is the dividing value between the defect-free state and the defective state.

[0013] Optionally, the weight calculation formula for the temporal attention mechanism is: ; in, The temporal attention mechanism is assigned to Attention weights; This indicates that the Long Short-Term Memory (LSTM) network is in the first... The hidden state features output at each time step; This indicates that the Long Short-Term Memory (LSTM) network is in the first... The hidden state feature vector output at each time step; Number the time steps and ; Indicates the total number of time steps; This is the context query vector in the attention mechanism; This is the attention weight mapping matrix; This indicates a vector concatenation operation.

[0014] Another aspect of the present invention provides an intelligent recognition system for the combustion stability of flame images, the intelligent recognition system for the combustion stability of flame images comprising: The preprocessing module is used to acquire a sequence of flame images during the combustion process and to preprocess each flame image in the sequence. The determination module is used to construct multi-timescale image samples based on the preprocessed flame image using a sliding window, and determine the optimal window size based on the sample entropy of the image data within the sliding window; The feature extraction module is used to perform deep feature extraction on the preprocessed flame image based on the sliding window corresponding to the optimal window size and a multi-scale convolutional autoencoder structure, and sets a new loss function for the multi-scale convolutional autoencoder structure. The clustering analysis module is used to introduce an adaptive structural similarity index to perform clustering analysis on the extracted deep features, forming two independent clusters; The permutation entropy calculation module is used to calculate the dynamic permutation entropy of two independent clusters based on a time series window. Clusters with dynamic permutation entropy less than a set threshold are defined as stable combustion clusters, and clusters with dynamic permutation entropy greater than a set threshold are defined as unstable combustion clusters. The probability calculation module is used to calculate the centroids of stable and unstable combustion clusters respectively, and calculate the combustion stability probability index based on the extracted deep features to quantify the probability of instability in the flame image. The identification and prediction module is used to input the extracted deep features and their corresponding combustion stability probabilities into the fusion model of long short-term memory network and temporal attention mechanism to identify combustion stability and predict trends.

[0015] Compared to existing technologies, this invention features online inference capabilities, enabling real-time inference on edge computing devices. It achieves high-precision monitoring and trend prediction of the combustion process, offering significant advantages in accuracy, real-time performance, and deployability. Specifically, firstly, a multi-scale convolutional autoencoder network is used to fuse multi-layer spatial features of flame images, simultaneously extracting local texture and global morphological information, significantly improving the accuracy and robustness of combustion state recognition. Secondly, an adaptive structural similarity index is introduced for feature clustering, which adaptively adjusts weights based on flame brightness, contrast, and structural consistency, improving the accuracy and physical consistency of combustion state classification. Thirdly, dynamic permutation entropy quantifies the temporal complexity of the flame, enabling early detection of combustion fluctuations and unstable trends. Combining a long short-term memory network with a temporal attention mechanism highlights key moment features, achieving accurate prediction of stable trends. Furthermore, this invention features a lightweight structure, high computational efficiency, and can run in real time on an edge computing platform. It can perform non-contact, real-time combustion stability monitoring using only visible light images, offering advantages in high reliability and low cost. It also boasts high precision, strong robustness, and deployability, making it widely applicable to combustion systems such as coal-fired boilers, gas turbines, and industrial kilns. This enables online monitoring of the combustion process, fault early warning, and optimized combustion control, providing reliable technical support for energy efficiency optimization and safe operation. It has outstanding engineering application value and promising prospects for widespread adoption. Attached Figure Description

[0016] One or more embodiments are illustrated by way of example with the corresponding pictures in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0017] Figure 1 A flowchart illustrating an intelligent recognition method for combustion stability in flame images, provided by an embodiment of the present invention; Figure 2 A flowchart for determining the optimal window size is provided as another embodiment of the present invention; Figure 3 A schematic diagram of a multi-scale convolutional autoencoder structure provided in another embodiment of the present invention; Figure 4 A flowchart of clustering analysis based on adaptive structural similarity index provided for another embodiment of the present invention; Figure 5 A schematic diagram of the computation process of a long short-term memory network provided for another embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details are presented in the various embodiments of the present invention to facilitate a better understanding of the invention. However, the technical solutions claimed in the present invention can be implemented even without these technical details and with various variations and modifications based on the following embodiments. The division of the various embodiments below is for ease of description and should not constitute any limitation on the specific implementation of the present invention. The various embodiments can be combined with and referenced by each other without contradiction.

[0019] One embodiment of the present invention provides a method for intelligent recognition of combustion stability in flame images, the process of which is as follows: Figure 1 As shown, the process includes steps S1 to S7.

[0020] Step S1: Acquire a sequence of flame images during the combustion process and preprocess each flame image in the sequence.

[0021] Specifically, step S1 can use an image acquisition device to acquire a high-speed flame image sequence of the combustion process. For example, the image acquisition device can be used to acquire a flame image sequence of the furnace in a large combustion device such as a boiler, gas turbine, or industrial kiln.

[0022] Preprocessing of each flame image mainly involves grayscale conversion, normalization, and enhancement to form a time-series image dataset. For example, in step S1, preprocessing of each flame image in the flame image sequence includes: employing a noise reduction algorithm combining Gaussian filtering and median filtering; mapping the pixel intensity of each flame image to the [0, 1] interval through linear normalization; and performing data enhancement processing on each flame image through random rotation, mirroring, and brightness perturbation.

[0023] Step S2: Based on the preprocessed flame image, construct image samples at multiple time scales using a sliding window, and determine the optimal window size based on the sample entropy of the image data within the sliding window.

[0024] For example, combined Figure 2 Step S2 specifically includes the following steps: Step S2.1: For the preprocessed flame image, initialize the size of a series of sliding windows. and step length ; Step S2.2: Based on the preprocessed flame image, determine each and The combination of flame image time series within the corresponding sliding window is calculated and recorded according to the following formula. and The sample entropy of the flame image time series within the corresponding sliding window is: ; in, This represents the sample entropy of the flame image time series within the sliding window. The embedding dimension during phase space reconstruction is . When the distance between two phase space vectors reconstructed from the flame image time series within the same sliding window, but with different start times, is less than a distance threshold. The probability of pattern matching. The embedding dimension during phase space reconstruction is . When the distance between two phase space vectors reconstructed from the flame image time series within the same sliding window, but with different start times, is less than a distance threshold. The probability of pattern matching; Step S2.3: Plot the sample entropy as a function of... and A two-dimensional surface plot of combined variations, with a threshold set. Find samples whose entropy is greater than a threshold. The minimum sliding window is used as the optimal window size, and the size and step size of the minimum sliding window are taken as the optimal window size.

[0025] Step S3: Based on the sliding window corresponding to the optimal window size, a multi-scale convolutional autoencoder (MS-CAE) structure is used to extract deep features from the preprocessed flame image, and a novel loss function is set for the multi-scale convolutional autoencoder structure.

[0026] Specifically, the multi-scale convolutional autoencoder architecture includes multiple convolutional encoders and multiple convolutional decoders connected in sequence, with the last convolutional encoder connected to the first convolutional decoder. For example, they are combined... Figure 3 A multi-scale convolutional autoencoder architecture may include four convolutional encoders connected in sequence, namely... e 1. e 2. e 3. e 4, and the four convolutional decoders connected in sequence are respectively d 1. d 2. d 3. d 4. Convolutional Encoder e 4. Convolutional Decoder d 1. Connected.

[0027] In step S3, based on the sliding window corresponding to the optimal window size, a multi-scale convolutional autoencoder structure is used to extract deep features from the preprocessed flame image, including the following steps: Step S3.1: Use a sliding window with the optimal window size to convert the preprocessed flame image into matrix data as the input image for the multi-scale convolutional autoencoder structure. Let the input image The dimension is O×P , O For the input image height, P For the input image The width; Step S3.2: Input image x The data is sent to a convolutional encoder for processing. Each convolutional encoder uses several windows of size [missing information]. c×c Step size is q The convolutional filter extracts features from the input, and then... ReLU The activation function performs feature activation and uses a window size of [missing value]. o×o Step size is The max-pooling layer performs feature dimensionality reduction and outputs the corresponding deep features. The deep features output by the previous convolutional encoder serve as the input to the next convolutional encoder, and the input to the first convolutional encoder is the input image. x ; Step S3.3: The deep features output by the last convolutional encoder are sent to the convolutional decoder for processing. All convolutional decoders except the last one use a window size of... g×g The upsampling layer upscales the input features, and then sequentially applies convolutional filters, ReLU The activation function processes the data and outputs the corresponding deep features. The final convolutional decoder uses a window size of... g×g The upsampling layer upscales the input features, and then sequentially applies convolutional filters, Sigmoid The function processes the data, and the output dimension is... O×P Reconstructed image x rec In this process, the deep features output by the previous convolutional decoder are used as the input to the next convolutional decoder, and the input to the first convolutional decoder is the deep features output by the last convolutional encoder.

[0028] For example, the dimension is O×P Input image x Sent to Figure 3 When the multi-scale convolutional autoencoder structure shown is processed by the convolutional encoder, it first... e 1 pair of input images x Processing, convolutional encoder e 1. Firstly adopt Q One window size is c×c Step size is q Convolutional filters C 1 pair of input images x Feature extraction is performed, and then... ReLU The activation function performs feature activation and uses a window size of [missing value]. o×o Step size is Maximum pooling layer P 1. Perform feature dimensionality reduction, and the final output dimension is... a 1 ×b 1 ×c Deep features of 1 h 1, among which, ReLU The activation function is expressed as , Similarly, by representing the input neurons, we can obtain the convolutional encoder. e 2 pairs of deep features h The dimension of the output after processing is 1. a 2 ×b 2 ×c 2 Deep features h 2. Convolutional Encoder e 3 pairs of deep features h 2. The dimension of the output after processing is a 3 ×b 3×c 3 Deep features h 3. Convolutional Encoder e 4 pairs of deep features h 3. The dimension of the output after processing is a 4 ×b 4 ×c 4 deep features h 4. Next, deep features h 4 is fed to the convolutional decoder d 1. Processing, convolutional decoder d 1. First, set the window size to... g×g upsampling layer U 1 (g×g) For deep features h 4. Perform feature upscaling, and then apply convolutional filters sequentially. C 5 (Q 5 @c×c+q) , ReLU The activation function is used for processing, and the output dimension is... a 3 ×b 3 ×c 3 Deep features h 5. Similarly, a convolutional decoder can be obtained. d 2 pairs of deep features h 5. The dimension of the output after processing is a 2 ×b 2 ×c 2 Deep features h 6. Convolutional Decoder d 3 pairs of deep features h 6. The dimension of the output after processing is a 1 ×b 1 ×c Deep features of 1 h 7. The last convolutional decoder is the convolutional decoder. d 4. Use a window size of g×g Upsampling layers for deep features h 7. Perform feature upscaling, and then sequentially apply convolutional filters, Sigmoid The function processes the data, and the output dimension is... O×P Reconstructed image x rec .

[0029] For example, the novel loss function for a multi-scale convolutional autoencoder structure is expressed as follows: ; in, This represents the total loss of the multi-scale convolutional autoencoder structure. Indicates the first loss term and , Indicates the second loss term and , Indicates the third loss item and , Indicates the input image. Indicates the reconstructed image. O Indicates the height of the input image. P Indicates the width of the input image. and These represent the input image and the reconstructed image, respectively. The value at the location, Indicates the image height index. Indicates the image width index. and represent the mean values ​​of the input image and the reconstructed image, respectively. and These represent the variances of the input image and the reconstructed image, respectively. This represents the covariance between the input image and the reconstructed image. and All are constants, such as all are extremely small constants. , , They represent The weighting coefficients.

[0030] Step S4: Introduce the Adaptive Structural Similarity (ASSIM) index to perform cluster analysis on the extracted deep features, forming two independent clusters.

[0031] For example, the adaptive structural similarity index is labeled as Its numerical range is [0,1]. The specific definition is: ; in, Represents the luminance component. Represents the contrast component. Represents structural components, , , They represent The weights, and , , Adaptive optimization using genetic algorithms can improve clustering accuracy. Indicates the first Deep features of the flame image within a sliding window. Indicates the first Deep features of the flame image within a sliding window. It can reflect the consistency of brightness, contrast and structure of feature vectors.

[0032] Combined Figure 4 Step S4 specifically includes the following steps: Step S4.1: Randomly select two deep features extracted from the multi-scale convolutional autoencoder structure, i.e., two MS-CAE deep features, as the initial centroids of the two clusters respectively; Step S4.2: Calculate the adaptive structural similarity index between all deep features extracted by the multi-scale convolutional autoencoder structure, i.e. all MS-CAE deep features, and the two initial centroids, and accordingly reassign the deep features extracted by the multi-scale convolutional autoencoder structure, i.e. MS-CAE deep features, to the two clusters. Step S4.3: Calculate the mean of all deep features (MS-CAE deep features) in each dimension in the two clusters respectively, and use the calculated mean as the centroid of the corresponding cluster to update the centroid of the two clusters. Step S4.4: Repeat steps S4.2 and S4.3 until the termination condition is met, that is, the centroids of the two clusters no longer change or the limit of the number of updates is reached, and finally the clustering result is obtained, that is, two independent clusters.

[0033] Step S5: Calculate the dynamic permutation entropy (DPE) of two independent clusters based on the time series window. Define the cluster with dynamic permutation entropy less than a set threshold as a stable combustion cluster, and define the cluster with dynamic permutation entropy greater than the set threshold as an unstable combustion cluster.

[0034] For example, the formula for calculating the dynamic permutation entropy in step S5 is: ; in, For time window Inner The probability of occurrence of a pixel gradient pattern; Indicates the number of pixel gradient modes; For time window The permutation entropy within a time window is used to characterize the time window. Internal combustion fluctuations; when Exceeding the threshold At that time, determine the time window The combustion state inside is unstable; when Less than the threshold At that time, determine the time window The combustion state within the cluster is stable. If the combustion state of an independent cluster is determined to be unstable, then that independent cluster is defined as an unstable combustion cluster; conversely, if the combustion state of an independent cluster is determined to be stable, then that independent cluster is defined as a stable combustion cluster.

[0035] Step S6: Calculate the centroids of stable and unstable combustion clusters respectively, and calculate the combustion stability probability index based on the extracted deep features to quantify the probability of instability in the flame image.

[0036] For example, step S6 specifically includes: Calculate the average value of all deep features in each dimension in both the stable combustion cluster and the unstable combustion cluster. Use the calculated average value as the centroid of the corresponding stable combustion cluster and unstable combustion cluster. That is, the centroid of the stable combustion cluster is the average value of all deep features in each dimension in the stable combustion cluster, and the centroid of the unstable combustion cluster is the average value of all deep features in each dimension in the unstable combustion cluster. The combustion stability probability index is calculated using the following formula: ; in, For the first The deep feature vector is composed of the deep features of each sliding window. , These are the centroids of the stable combustion cluster and the unstable combustion cluster, respectively. For the first The probability of combustion stability in a sliding window. The numerical range is 0 to 1, where 0 represents stable combustion, 1 represents unstable combustion, and 0.5 is the dividing value between the defect-free state and the defective state.

[0037] Step S7: Input the extracted deep features and their corresponding combustion stability probabilities into the fusion model of Long Short-Term Memory (LSTM) network and temporal attention mechanism to perform combustion stability identification and trend prediction.

[0038] For example, this embodiment uses a Long Short-Term Memory (LSTM) network from the prior art. An LSM network consists of an input gate, a forget gate, and an output gate, which are combined together... Figure 5 The specific calculation process is as follows: ; ; ; ; ; ; in, These represent the input gate, forget gate, and output gate at the [number]th ... The activation vector at each time step express Sigmoid function, Indicates the first Deep features of each time step Indicates the first Temporal characteristics of each time step Indicates the first Temporal characteristics of each time step Let represent the temporal feature weight matrices of the input gate, forget gate, and output gate, respectively. These represent the deep feature weight matrices for the input gate, forget gate, and output gate, respectively. These represent the biases of the input gate, forget gate, and output gate, respectively. express Candidate storage state at any given time. Represents a hyperbolic sine function. The temporal feature weight matrix representing the candidate storage state. The deep feature weight matrix represents the candidate storage state. Indicates the bias of the candidate storage state. Indicates the first The internal state at each time step Indicates the first The internal state of each time step; This represents the dot product of matrices.

[0039] The weight calculation formula for the temporal attention mechanism is as follows: .

[0040] in, The temporal attention mechanism is assigned to Attention weights are used to measure the attention weights of the first... The relative importance of each time step in the overall temporal feature representation, with values ​​ranging from [0, 1]; This indicates that the Long Short-Term Memory (LSTM) network is in the first... The hidden state features output at each time step are used as input features for the temporal attention mechanism; This indicates that the Long Short-Term Memory (LSTM) network is in the first... The hidden state feature vector output at each time step; Number the time steps and , used to participate in the normalization calculation of attention weights in the temporal attention mechanism; Indicates the total number of time steps; This is the context query vector in the attention mechanism, used to characterize the focus of the current combustion stability discrimination task; This is the attention weight mapping matrix; This indicates a vector concatenation operation.

[0041] It should be noted that, It does not participate in the gating state update process within the Long Short-Term Memory (LSTM) network. Instead, it is introduced based on the hidden state feature vector output by the LSTM network and used to weight and fuse the hidden state feature vectors output by the LSTM network at different time steps. Through this attention weighting mechanism, the moment that has a key impact on combustion stability can be highlighted, thus achieving an effective fusion of the LSTM network and the temporal attention mechanism.

[0042] The intelligent combustion stability recognition method for flame images provided by this invention has online inference capabilities and can realize real-time inference on edge computing devices. It can achieve non-contact, real-time combustion stability recognition by relying solely on visible light images, i.e., flame images of the combustion process. It has high accuracy, strong robustness and deployability, and can realize online monitoring and early warning of the combustion process in combustion devices such as coal-fired boilers, gas turbines and industrial kilns, providing reliable technical support for energy efficiency optimization and safe operation.

[0043] In another aspect, the present invention provides an intelligent recognition system for the combustion stability of flame images, which includes a preprocessing module, a determination module, a feature extraction module, a cluster analysis module, a permutation entropy calculation module, a probability calculation module, and a recognition and prediction module.

[0044] The preprocessing module is used to acquire a sequence of flame images during the combustion process and to preprocess each flame image in the sequence. The determination module is used to construct multi-timescale image samples based on the preprocessed flame image using a sliding window, and determine the optimal window size based on the sample entropy of the image data within the sliding window; The feature extraction module is used to perform deep feature extraction on the preprocessed flame image based on the sliding window corresponding to the optimal window size and a multi-scale convolutional autoencoder structure, and sets a new loss function for the multi-scale convolutional autoencoder structure. The clustering analysis module is used to introduce an adaptive structural similarity index to perform clustering analysis on the extracted deep features, forming two independent clusters; The permutation entropy calculation module is used to calculate the dynamic permutation entropy of two independent clusters based on a time series window. Clusters with dynamic permutation entropy less than a set threshold are defined as stable combustion clusters, and clusters with dynamic permutation entropy greater than a set threshold are defined as unstable combustion clusters. The probability calculation module is used to calculate the centroids of stable and unstable combustion clusters respectively, and calculate the combustion stability probability index based on the extracted deep features to quantify the probability of instability in the flame image. The identification and prediction module is used to input the extracted deep features and their corresponding combustion stability probabilities into the fusion model of long short-term memory network and temporal attention mechanism to identify combustion stability and predict trends.

[0045] The specific implementation method of the intelligent recognition system for combustion stability of flame images provided in the embodiments of the present invention can be found in the description of the intelligent recognition method for combustion stability of flame images provided in the embodiments of the present invention, and will not be repeated here.

[0046] The intelligent combustion stability recognition system for flame images provided by this invention has online inference capabilities and can perform real-time inference on edge computing devices. It can achieve non-contact, real-time combustion stability recognition by relying solely on visible light images, i.e., flame images of the combustion process. It has high precision, strong robustness, and deployability, and can realize online monitoring and early warning of the combustion process in combustion devices such as coal-fired boilers, gas turbines, and industrial kilns, providing reliable technical support for energy efficiency optimization and safe operation.

[0047] Those skilled in the art will understand that the above embodiments are specific implementations of the present invention, and in practical applications, various changes can be made in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A method for intelligent recognition of combustion stability in flame images, characterized in that, The intelligent recognition method for combustion stability of flame images includes the following steps: Step S1: Acquire a sequence of flame images during the combustion process, and preprocess each flame image in the sequence; Step S2: Based on the preprocessed flame image, construct image samples at multiple time scales using a sliding window, and determine the optimal window size based on the sample entropy of the image data within the sliding window; Step S3: Based on the sliding window corresponding to the optimal window size, a multi-scale convolutional autoencoder structure is used to extract deep features from the preprocessed flame image, and a new loss function is set for the multi-scale convolutional autoencoder structure. Step S4: Introduce an adaptive structural similarity index to perform cluster analysis on the extracted deep features, forming two independent clusters; Step S5: Calculate the dynamic permutation entropy of two independent clusters based on the time series window. Define the cluster with dynamic permutation entropy less than a set threshold as a stable combustion cluster, and define the cluster with dynamic permutation entropy greater than a set threshold as an unstable combustion cluster. Step S6: Calculate the centroids of stable and unstable combustion clusters respectively, and calculate the combustion stability probability index based on the extracted deep features to quantify the probability of instability in the flame image. Step S7: Input the extracted deep features and their corresponding combustion stability probabilities into the fusion model of long short-term memory network and temporal attention mechanism to perform combustion stability identification and trend prediction.

2. The intelligent recognition method for combustion stability of flame images according to claim 1, characterized in that, In step S1, preprocessing is performed on each flame image in the flame image sequence, including: A noise reduction algorithm combining Gaussian filtering and median filtering is adopted. The pixel intensity of each flame image is mapped to the [0, 1] interval by linear normalization, and data enhancement processing is performed on each flame image by random rotation, mirroring and brightness perturbation.

3. The intelligent recognition method for combustion stability of flame images according to claim 2, characterized in that, Step S2 specifically includes the following steps: Step S2.1: Initialize the size of a series of sliding windows and step length ; Step S2.2: Based on the preprocessed flame image, determine each and The combination of flame image time series within the corresponding sliding window is calculated and recorded according to the following formula. and The sample entropy of the flame image time series within the corresponding sliding window is: ； in, This represents the sample entropy of the flame image time series within the sliding window. The embedding dimension during phase space reconstruction is . When the distance between two phase space vectors reconstructed from the flame image time series within the same sliding window, but with different start times, is less than a distance threshold. The probability of pattern matching. The embedding dimension during phase space reconstruction is . When the distance between two phase space vectors reconstructed from the flame image time series within the same sliding window, but with different start times, is less than a distance threshold. The probability of pattern matching; Step S2.3: Plot the sample entropy as a function of... and A two-dimensional surface plot of combined variations, with a threshold set. Find samples whose entropy is greater than a threshold. The minimum sliding window is used as the optimal window size, and the size and step size of the minimum sliding window are used as the optimal window size.

4. The intelligent recognition method for combustion stability of flame images according to claim 1, characterized in that, The multi-scale convolutional autoencoder structure includes multiple convolutional encoders and multiple convolutional decoders connected in sequence, with the last convolutional encoder connected to the first convolutional decoder. In step S3, based on the sliding window corresponding to the optimal window size, a multi-scale convolutional autoencoder structure is used to extract deep features from the preprocessed flame image, including the following steps: Step S3.1: Use a sliding window with the optimal window size to convert the preprocessed flame image into matrix data as the input image for the multi-scale convolutional autoencoder structure. Let the input image The dimension is O×P , O For the input image height, P For the input image The width; Step S3.2: Input image x The data is sent to a convolutional encoder for processing. Each convolutional encoder uses several windows of size [missing information]. c×c Step size is q The convolutional filter extracts features from the input, and then... ReLU The activation function performs feature activation and uses a window size of [missing value]. o×o Step size is The max-pooling layer performs feature dimensionality reduction and outputs the corresponding deep features. The deep features output by the previous convolutional encoder serve as the input to the next convolutional encoder, and the input to the first convolutional encoder is the input image. x ; Step S3.3: The deep features output by the last convolutional encoder are sent to the convolutional decoder for processing. All convolutional decoders except the last one use a window size of... g×g The upsampling layer upscales the input features, and then sequentially applies convolutional filters, ReLU The activation function processes the data and outputs the corresponding deep features. The final convolutional decoder uses a window size of... g×g The upsampling layer upscales the input features, and then sequentially applies convolutional filters, Sigmoid The function processes the data, and the output dimension is... O×P Reconstructed image x rec In this process, the deep features output by the previous convolutional decoder are used as the input to the next convolutional decoder, and the input to the first convolutional decoder is the deep features output by the last convolutional encoder.

5. The intelligent recognition method for combustion stability of flame images according to claim 1, characterized in that, The novel loss function for the multi-scale convolutional autoencoder structure is expressed as: ； in, This represents the total loss of the multi-scale convolutional autoencoder structure. Indicates the first loss term and , Indicates the second loss term and , Indicates the third loss item and , Indicates the input image. Indicates the reconstructed image. O Indicates the height of the input image. P Indicates the width of the input image. and These represent the input image and the reconstructed image, respectively. The value at the location, Indicates the image height index. Indicates the image width index. and represent the mean values ​​of the input image and the reconstructed image, respectively. and These represent the variances of the input image and the reconstructed image, respectively. This represents the covariance between the input image and the reconstructed image. and All are constants. , , They represent The weighting coefficients.

6. The intelligent recognition method for combustion stability of flame images according to claim 1, characterized in that, The adaptive structural similarity index is labeled as Its numerical range is [0, 1], and it is specifically defined as: ； in, Represents the luminance component. Represents the contrast component. Represents structural components, , , They represent The weights, and , , Adaptive optimization using genetic algorithms can improve clustering accuracy. Indicates the first Deep features of the flame image within a sliding window. Indicates the first Deep features of the flame image within a sliding window; Step S4 specifically includes the following steps: Step S4.1: Randomly select two deep features extracted from the multi-scale convolutional autoencoder structure as the initial centroids of the two clusters; Step S4.2: Calculate the adaptive structural similarity index between all deep features extracted by the multi-scale convolutional autoencoder structure and the two initial centroids, and reassign the deep features extracted by the multi-scale convolutional autoencoder structure to the two clusters accordingly; Step S4.3: Calculate the mean of all deep features in each dimension for the two clusters respectively, and use the calculated mean as the centroid of the corresponding cluster; Step S4.4: Repeat steps S4.2 and S4.3 until the centroids of the two clusters no longer change or the limit number of updates is reached, and finally two independent clusters are obtained.

7. The intelligent recognition method for combustion stability of flame images according to claim 1, characterized in that, The formula for calculating the dynamic permutation entropy in step S5 is: ； in, For time window Inner The probability of occurrence of a pixel gradient pattern; Indicates the number of pixel gradient modes; For time window The permutation entropy within a time window is used to characterize the time window. Internal combustion fluctuations, when Exceeding the threshold At that time, determine the time window The combustion state inside is unstable.

8. The intelligent recognition method for combustion stability of flame images according to claim 1, characterized in that, Step S6 specifically includes: Calculate the average value of all deep features in each dimension in the stable combustion cluster and the unstable combustion cluster respectively, and use the calculated average value as the centroid of the corresponding stable combustion cluster and unstable combustion cluster; The combustion stability probability index is calculated using the following formula: ； in, For the first The deep feature vector is composed of the deep features of each sliding window. , These are the centroids of the stable combustion cluster and the unstable combustion cluster, respectively. For the first The probability of combustion stability in a sliding window. The numerical range is 0 to 1, where 0 represents stable combustion, 1 represents unstable combustion, and 0.5 is the dividing value between the defect-free state and the defective state.

9. The intelligent recognition method for combustion stability of flame images according to claim 1, characterized in that, The weight calculation formula for the temporal attention mechanism is as follows: ； in, The temporal attention mechanism is assigned to Attention weights; This indicates that the Long Short-Term Memory (LSTM) network is in the first... The hidden state features output at each time step; This indicates that the Long Short-Term Memory (LSTM) network is in the first... The hidden state feature vector output at each time step; Number the time steps and ; Indicates the total number of time steps; This is the context query vector in the attention mechanism; This is the attention weight mapping matrix; This indicates a vector concatenation operation.

10. A flame image combustion stability intelligent recognition system, characterized in that, The intelligent recognition system for combustion stability of flame images includes: The preprocessing module is used to acquire a sequence of flame images during the combustion process and to preprocess each flame image in the sequence. The determination module is used to construct multi-timescale image samples based on the preprocessed flame image using a sliding window, and determine the optimal window size based on the sample entropy of the image data within the sliding window; The feature extraction module is used to perform deep feature extraction on the preprocessed flame image based on the sliding window corresponding to the optimal window size and a multi-scale convolutional autoencoder structure, and sets a new loss function for the multi-scale convolutional autoencoder structure. The clustering analysis module is used to introduce an adaptive structural similarity index to perform clustering analysis on the extracted deep features, forming two independent clusters; The permutation entropy calculation module is used to calculate the dynamic permutation entropy of two independent clusters based on a time series window. Clusters with dynamic permutation entropy less than a set threshold are defined as stable combustion clusters, and clusters with dynamic permutation entropy greater than a set threshold are defined as unstable combustion clusters. The probability calculation module is used to calculate the centroids of stable and unstable combustion clusters respectively, and calculate the combustion stability probability index based on the extracted deep features to quantify the probability of instability in the flame image. The identification and prediction module is used to input the extracted deep features and their corresponding combustion stability probabilities into the fusion model of long short-term memory network and temporal attention mechanism to identify combustion stability and predict trends.

Images