Boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement

Abstract

The invention discloses a boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement. The method comprises the following steps: obtaining a preprocessed boiler operation multi-source time sequence data set; training a maximum correlation entropy convolution sparse coding model offline; generating sparse feature representation for describing the combustion dynamics of the boiler; forming a sparse feature sample set, and sorting the sparse feature sample set according to normal working condition samples and fault working condition samples to obtain a sparse feature training sample set with labels; establishing a boundary discrimination enhancement classification model to obtain boundary discrimination enhancement decision boundary parameters; outputting a working condition state label and a distance index with a decision boundary; and generating an early warning result of the corresponding grade. According to the method, the fuzzy discrimination interval between the normal working condition and the abnormal working condition is remarkably reduced, and the certainty of early abnormality recognition is improved.

Description

Technical Field

[0001] This invention relates to the field of boiler technology, and in particular to a boiler low-load operation early warning method based on sparse representation and enhanced boundary discrimination. Background Technology

[0002] With the increasing demand for flexible peak shaving in thermal power plants, low-load boiler operation has gradually become the norm. Under normal operating conditions, due to the drastic fluctuations in combustion status, frequent adjustments of auxiliary equipment, and complex electrical and control environments, the boiler operation sequence signals collected by field sensors are often mixed with a large amount of non-Gaussian impulse noise. Most existing boiler low-load early warning technologies are based on sparse representation, dictionary learning, or traditional feature engineering methods. Essentially, they usually assume that the noise follows a Gaussian distribution and take the minimum mean square error as the optimization objective. When faced with strong impulse noise interference, the algorithm is prone to misjudging noise points as fault features, resulting in weak fault signals being masked or feature reconstruction being distorted, reducing the ability to identify early risks such as combustion instability.

[0003] Under low load conditions, the automatic adjustment range of the boiler control system is large. Normal operating condition fluctuations and fault precursors are prone to high overlap in the feature space. Existing early warning methods mostly rely on fixed thresholds or standard linear classifiers, lacking a discrimination boundary enhancement mechanism for multi-label samples. This easily leads to high false alarms or false alarms, making it difficult to achieve a balance between operating condition adaptation and anomaly identification under dynamic operating conditions. Summary of the Invention

[0004] One objective of this invention is to propose a boiler low-load operation early warning method based on sparse representation and enhanced boundary discrimination. This invention significantly narrows the fuzzy discrimination interval between normal and abnormal operating conditions and improves the certainty of early anomaly identification.

[0005] A boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement according to an embodiment of the present invention includes:

[0006] Collect multi-source time-series data of boiler operation and perform preprocessing to obtain preprocessed multi-source time-series data of boiler operation;

[0007] Using a preprocessed multi-source time-series dataset of boiler operation as input, an offline maximum correlation entropy convolutional sparse coding model is trained.

[0008] The preprocessed multi-source time-series dataset of boiler operation is input into the maximum correlation entropy convolutional sparse coding model to generate a sparse feature representation describing the boiler combustion dynamics.

[0009] The sparse feature representation is normalized to form a sparse feature sample set. The sparse feature sample set is then sorted into normal operating condition samples and fault operating condition samples to obtain a labeled sparse feature training sample set.

[0010] Using a set of labeled sparse feature training samples as training data, a boundary discrimination enhancement classification model is established. The decision boundary is optimized by maximizing the inter-class margin and introducing a slack variable penalty term, thus obtaining the boundary discrimination enhancement decision boundary parameters.

[0011] Following the online inference process, the sparse feature representation generated in real time is input into the boundary discrimination enhancement classification model, and the output is the working condition label and the distance index to the decision boundary;

[0012] Based on the distance index of the decision boundary and the preset multi-level threshold mapping rules, the corresponding level of early warning results are generated.

[0013] Optionally, the process of acquiring and preprocessing the multi-source time-series dataset of boiler operation includes:

[0014] The furnace negative pressure signal, induced draft fan current signal, flame intensity signal, coal feed rate signal, and oxygen content signal are collected during the low-load operation of the boiler. Time synchronization, sampling rate unification, and format standardization are performed to construct a multi-source time series dataset of boiler operation. Non-Gaussian impulse noise preprocessing is performed on the multi-source time series dataset of boiler operation to preserve outliers with spike characteristics in the signal, resulting in a preprocessed multi-source time series dataset of boiler operation.

[0015] Optionally, the maximum correlation entropy convolutional sparse coding model includes:

[0016] The pre-processed set of boiler operation sequence signals is represented as multiple boiler operation sequence signal segments, and all boiler operation sequence signal segments constitute the set of boiler operation sequence signals.

[0017] Initialize the convolution dictionary, which consists of multiple convolution atoms. Initialize sparse activation values ​​corresponding to all convolution atoms and all sampling points for each boiler operation sequence signal segment. All sparse activation values ​​form a sparse coefficient map.

[0018] Using the boiler operating sequence signal set, convolution dictionary, and sparse coefficient map as input, an optimization objective function for the maximum correlation entropy convolutional sparse coding model is constructed.

[0019] With a fixed convolution dictionary, the proximal gradient descent algorithm is used to iteratively update each sparse activation coefficient in the sparse coefficient graph.

[0020] With a fixed sparse coefficient graph, a variant of the fast iterative shrinking threshold algorithm is used to update the gradient for each convolution atom in the convolution dictionary;

[0021] During each training iteration, the Gaussian kernel bandwidth parameter is dynamically adjusted using an annealing strategy;

[0022] When the convergence condition is met or the maximum number of iterations is reached, the final maximum correlation entropy convolutional sparse coding model is obtained.

[0023] Optionally, the sparse feature representation describing boiler combustion dynamics includes:

[0024] Each boiler runtime sequence signal segment in the preprocessed boiler runtime sequence signal set is used as input, and the trained maximum correlation entropy convolutional sparse coding model is invoked.

[0025] For each boiler operation sequence signal segment, a sparse decoding process is performed under the premise of a fixed convolutional dictionary to obtain the sparse activation coefficient map corresponding to the boiler operation sequence signal segment.

[0026] The sparse activation coefficient maps obtained from all boiler operation sequence signal segments are uniformly organized to construct a sparse feature representation set.

[0027] Optionally, the step of organizing the sparse feature sample set according to normal operating condition samples and fault operating condition samples includes:

[0028] Perform a normalization transformation on each sparse feature representation to convert the sparse feature representation into a one-dimensional sparse feature vector;

[0029] Normalize each one-dimensional sparse feature vector to obtain the normalized sparse feature vector.

[0030] All normalized sparse feature vectors are combined according to the order of signal segments to obtain a normalized sparse feature sample set.

[0031] The normalized sparse feature sample set is organized, and each normalized sparse feature vector and its corresponding running state label are combined to form a labeled sparse feature training sample set.

[0032] Optionally, establishing the boundary discrimination enhancement classification model includes:

[0033] A boundary discrimination enhancement classification model is constructed. During the training process of the boundary discrimination enhancement classification model, different weights are assigned to the training errors of normal low load condition samples and fault precursor samples by setting different penalty coefficients. The optimization objective is to minimize the radius of the hypersphere, while ensuring that the normal low load condition samples are contained within the hypersphere or close to the hypersphere boundary, and the fault precursor samples are located outside the hypersphere or far from the hypersphere boundary.

[0034] An alternating optimization strategy is adopted to update the hypersphere radius, hypersphere center, and all relaxation variables in sequence, so that the value of the optimization objective function gradually decreases. The optimal values ​​of the hypersphere radius, hypersphere center, and all relaxation variables obtained together constitute the optimal decision parameter set of the boundary discrimination enhancement classification model.

[0035] Optionally, in the boundary discrimination enhancement classification model constraints, by limiting the Euclidean distance between the normalized sparse feature vector of the normal low-load operating condition sample and the center of the hypersphere to no greater than the hypersphere radius plus the relaxation variable, the control of the normal low-load operating condition sample being contained within the hypersphere or allowed to slightly exceed the boundary is adapted to the fluctuation of normal operating conditions under boiler low-load operation. By limiting the Euclidean distance between the normalized sparse feature vector of the fault precursor sample and the center of the hypersphere to no less than the hypersphere radius minus the relaxation variable, the strong constraint of the fault precursor sample falling outside the hypersphere or far from the hypersphere boundary is strengthened to enhance the boundary discrimination capability.

[0036] Optionally, the output operating condition status label and the distance index to the decision boundary include:

[0037] The normalized sparse feature vector is input into the boundary discrimination enhancement classification model, and the optimal decision parameter set is used to discriminate the real-time working condition. The Euclidean distance between the normalized sparse feature vector and the center of the optimal hypersphere is calculated.

[0038] Based on the relationship between the current Euclidean distance and the optimal hypersphere radius, output the current operating condition label:

[0039] If the current Euclidean distance is less than or equal to the optimal hypersphere radius, the operating condition label is zero, indicating that the boiler is currently operating under normal low load conditions.

[0040] If the current Euclidean distance is greater than the optimal hypersphere radius, the operating condition label is one, indicating that the current boiler operating condition is determined to be a precursor to a fault.

[0041] Optionally, the step of generating corresponding warning results based on the distance index of the decision boundary and the preset multi-level threshold mapping rule includes:

[0042] If the current distance index is less than or equal to the optimal hypersphere radius, the warning level is Level 1, indicating that the boiler is operating under normal low load conditions and there is no obvious abnormal deviation.

[0043] If the current distance index is greater than the optimal hypersphere radius and less than or equal to the first warning threshold, the warning level is Level II, indicating that the boiler operating status shows a slight deviation from the normal envelope and there is initial combustion disturbance.

[0044] If the current distance indicator is greater than the first warning threshold and less than or equal to the second warning threshold, the warning level is level three, indicating that the boiler operating status deviates significantly from the normal envelope and there is a risk of unstable combustion.

[0045] If the current distance indicator is greater than the second warning threshold, the warning level is level four, indicating that the boiler's operating status is seriously deviating from the normal envelope and is suspected to be in a state of impending failure or a high-risk operating state.

[0046] The beneficial effects of this invention are:

[0047] (1) This invention introduces the maximum correlation entropy criterion into the convolutional sparse coding process of boiler operating sequence signals, suppresses the pulse error exceeding the kernel bandwidth, and enables noise points to automatically reduce their weight during the optimization process. This effectively solves the problem of overfitting noise caused by large-scale field interference during low-load boiler operation. The dynamic annealing bandwidth parameter further enhances the model's adaptability. In the early stage, it focuses on the overall combustion structure, and in the later stage, it emphasizes local weak anomalies, enabling the maximum correlation entropy convolutional sparse coding model to accurately extract combustion oscillation and flame instability weak features from a strong noise background.

[0048] (2) This invention introduces a label-based supervised weighted hypersphere model, which applies different constraints to normal samples and fault precursor samples respectively. Normal samples are constrained inside the hypersphere, while fault precursor samples are forcibly excluded from the boundary of the hypersphere. The asymmetric penalty coefficient structure strengthens the rejection effect on fault samples. While maintaining the tolerance of fluctuations in normal low-load conditions, it significantly narrows the fuzzy discrimination interval between normal and abnormal conditions, and improves the certainty of early anomaly identification. Compared with traditional early warning methods based on linear classifiers or fixed thresholds, the boundary discrimination enhancement model improves the classification accuracy in the normal / abnormal overlapping area, effectively reducing the false alarm rate and false negative rate. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0050] Figure 1 This is a flowchart of a boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement proposed in this invention. Detailed Implementation

[0051] Example 1: Reference Figure 1 A boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement includes:

[0052] Collect multi-source time-series data of boiler operation and perform preprocessing to obtain preprocessed multi-source time-series data of boiler operation;

[0053] In this embodiment, multi-source time-series data of boiler operation is collected and preprocessed, including:

[0054] The furnace negative pressure signal, induced draft fan current signal, flame intensity signal, coal feed rate signal, and oxygen content signal are collected during the low-load operation of the boiler. Time synchronization, sampling rate unification, and format standardization are performed to construct a multi-source time series dataset of boiler operation. Non-Gaussian impulse noise preprocessing is performed on the multi-source time series dataset of boiler operation to preserve outliers with spike characteristics in the signal, resulting in a preprocessed multi-source time series dataset of boiler operation.

[0055] Using a preprocessed multi-source time-series dataset of boiler operation as input, an offline maximum correlation entropy convolutional sparse coding model is trained.

[0056] In this embodiment, the maximum correlation entropy convolutional sparse coding model includes:

[0057] The pre-processed set of boiler operation sequence signals is represented as multiple boiler operation sequence signal segments, and all boiler operation sequence signal segments constitute the set of boiler operation sequence signals.

[0058] Each boiler operation sequence signal segment is a sampling vector of equal length, and the number of sampling points in each boiler operation sequence signal segment is denoted as the sampling length.

[0059] Initialize the convolution dictionary, which consists of multiple convolution atoms. Initialize sparse activation values ​​corresponding to all convolution atoms and all sampling points for each boiler operation sequence signal segment. All sparse activation values ​​form a sparse coefficient map.

[0060] In Example 1, each convolution atom is a one-dimensional vector with a length equal to the length of the atom, and the number of convolution atom is equal to the number of convolution atoms. The initialization of the convolution dictionary is obtained by selecting typical waveforms from the fault-free samples in the historical boiler operation sequence signal segment after dimensionality reduction by principal component analysis. For each boiler operation sequence signal segment, a sparse activation value is initialized for each convolution atom and each sampling point in the convolution dictionary, and all sparse activation values ​​form a sparse coefficient map.

[0061] Using the boiler operating sequence signal set, convolution dictionary, and sparse coefficient map as input, an optimization objective function for the maximum correlation entropy convolutional sparse coding model is constructed.

[0062] The optimized objective function comprises three terms. First, the robust feature extraction objective function utilizes the maximum correlation entropy criterion and measures the reconstruction residual using a Gaussian kernel function. Outliers with error amplitudes exceeding the kernel bandwidth are automatically assigned extremely low weights, effectively suppressing the influence of on-site electrical interference and sensor spike noise, thus improving the extraction capability of weak combustion instability features in boilers under extreme noise environments. Second, the sparse structure discovery objective function ensures that only a very small number of key dictionary atoms are used to describe the complex boiler operating dynamics, helping to restore the essential behavioral patterns of the boiler combustion process and reducing the contamination of feature representations by irrelevant fluctuations and noise. Third, the group sparsity regularization term constrains the activation intensity of the entire group of convolutional atoms in the maximum correlation entropy convolutional sparse coding model. The aim is to introduce a group-level structural sparsity control mechanism while maintaining single-point sparsity. In low-load boiler operation scenarios, the combustion process often exhibits a dynamic pattern of local rather than global activation; that is, only a few types of combustion oscillation features are significantly activated at a certain time, corresponding to a few convolutional atoms in the convolutional dictionary having a high overall response, while most convolutional atoms are in an inactive state.

[0063] ;

[0064] in, Indicates the first The boiler operation sequence signal at time The observed values, Indicates the first The boiler operation sequence signal at time sparse reconstruction values, This represents the Gaussian kernel bandwidth parameter for the current iteration round. It is used to suppress gradient backpropagation and eliminate the effects of on-site electrical interference and spike noise when the error magnitude exceeds the bandwidth. Indicates the first The signal segment in the first Each convolution dictionary of atoms, time step The sparse activation coefficients, where D is the convolution dictionary. Indicates the first Sparse coefficient diagram corresponding to segment signals. This represents the set of all bandwidth parameters. This represents the total number of segments in the set of sequence signals during boiler operation. This indicates the number of sampling points for each signal segment. This indicates the number of atoms in the convolution dictionary. Represents the sparse regularization coefficient. represents the group sparse regularization coefficient.

[0065] With a fixed convolution dictionary, the proximal gradient descent algorithm is used to iteratively update each sparse activation coefficient in the sparse coefficient graph.

[0066] Using the boiler operating sequence signal set, the fixed convolution dictionary, and the current Gaussian kernel bandwidth parameters as input, the gradient information of each sparse activation coefficient is calculated according to the optimization objective function of the maximum correlation entropy convolutional sparse coding model. Through iterative optimization, the value of the optimization objective function is continuously reduced, and the converged sparse activation coefficient value is obtained, thus obtaining the optimal solution of the sparse coefficient map.

[0067] With a fixed sparse coefficient graph, a variant of the fast iterative shrinking threshold algorithm is used to update the gradient for each convolution atom in the convolution dictionary;

[0068] In Example 1, the fast iterative shrinking threshold algorithm variant is as follows: For each convolutional atom, a prediction step is performed. Based on the current value of the convolutional atom and the update result of the previous round, a predicted value is obtained through weighted combination. The gradient of the objective function with respect to the convolutional atom is calculated. The predicted value of the convolutional atom is updated using the gradient and the set step size parameter. A shrinking threshold operation is performed on the updated convolutional atom, shrinking each element of the convolutional atom element by element according to the threshold function, so that the sparsity and constraint conditions of the updated convolutional atom are satisfied. The shrinking convolutional atom is normalized to the unit norm. The prediction, gradient update and shrinking operations are repeated until the change amplitude of the convolutional dictionary is lower than the set convergence threshold or the maximum number of iterations is reached, and the optimal convolutional dictionary representing the boiler operation signal structure is obtained.

[0069] During each training iteration, the Gaussian kernel bandwidth parameter is dynamically adjusted using an annealing strategy;

[0070] In the early stages of iteration, the Gaussian kernel bandwidth parameter is set to an initial bandwidth value greater than the threshold to enhance the ability to capture the global operating trend of the boiler. As the number of training rounds increases, the Gaussian kernel bandwidth parameter is gradually reduced using an exponential decay strategy until it reaches the preset minimum bandwidth threshold. The annealing strategy enables the maximum correlation entropy convolutional sparse coding model to focus on the overall structure of the signal in the early stages and on details and weak anomalies in the later stages.

[0071] ;

[0072] in, This is the annealing attenuation coefficient. The preset minimum bandwidth threshold ultimately makes Focus on the local detailed features required for anomaly detection.

[0073] When the convergence condition is met or the maximum number of iterations is reached, the final maximum correlation entropy convolutional sparse coding model is obtained.

[0074] In this embodiment, by introducing a robust reconstruction term based on Gaussian kernel function, single-point and group sparse regularization constraints, and adopting a dynamic annealing bandwidth adjustment strategy, the maximum correlation entropy convolutional sparse coding model can effectively suppress non-Gaussian impulse noise and on-site electrical interference under low boiler load conditions. This enables accurate extraction of weak combustion anomaly features in boiler operation signals and high-fidelity restoration of essential dynamic behavior, thereby improving the noise resistance of feature extraction and the physical consistency of the representation structure.

[0075] The preprocessed multi-source time-series dataset of boiler operation is input into the maximum correlation entropy convolutional sparse coding model to generate a sparse feature representation describing the boiler combustion dynamics.

[0076] In this embodiment, the sparse feature representation describing the boiler combustion dynamics includes:

[0077] Each boiler runtime sequence signal segment in the preprocessed boiler runtime sequence signal set is used as input, and the trained maximum correlation entropy convolutional sparse coding model is invoked.

[0078] For each boiler operation sequence signal segment, a sparse decoding process is performed under the premise of a fixed convolutional dictionary to obtain the sparse activation coefficient map corresponding to the boiler operation sequence signal segment.

[0079] In Example 1, the sparse activation coefficient map is the set of sparse activation coefficient values ​​that minimize the total loss function of the maximum correlation entropy convolutional sparse coding model. The total loss function of the maximum correlation entropy convolutional sparse coding model includes a robust reconstruction term, an L1 regularization term, and a group sparsity regularization term. The robust reconstruction term measures the error between the boiler operation sequence signal segment and the convolutional reconstruction signal of the convolution dictionary and the sparse activation coefficient map, and adjusts the magnitude of the error using the Gaussian kernel bandwidth parameter. When the error magnitude is greater than the Gaussian kernel bandwidth parameter, the influence weight of the corresponding time step is reduced. The L1 regularization term constrains the number of non-zero elements in the sparse activation coefficient map. The group sparsity regularization term constrains the overall activation intensity of each convolution atom, improving the group-level sparsity of the feature structure. By searching among all possible values ​​of the sparse activation coefficient map, the sparse activation coefficient map corresponding to the boiler operation sequence signal segment is determined when the total loss function is minimized.

[0080] The sparse activation coefficient maps obtained from all boiler operation sequence signal segments are uniformly organized to construct a sparse feature representation set.

[0081] In Example 1, the sparse activation coefficient value corresponding to the minimum value of the total loss function is the sparse feature representation of the corresponding boiler operation sequence signal segment. The sparse feature representation set is an ordered set of all sparse activation coefficient maps obtained after mapping all preprocessed boiler operation sequence signal segments under the low load operation state through the maximum correlation entropy convolution sparse coding model. The total number of signal segments is equal to the number of sparse activation coefficient maps in the sparse feature representation set. The sparse feature representation set is used to distinguish between the normal low load operation state and the combustion abnormality precursor state.

[0082] Each element in the sparse feature representation set is a sparse feature representation. A sparse feature representation is a two-dimensional matrix determined by the number of atoms in the convolution dictionary and the sampling length of each signal segment. The number of sparse feature representations contained in the sparse feature representation set is equal to the total number of sequence signal segments during boiler operation.

[0083] The sparse feature representation is normalized to form a sparse feature sample set. The sparse feature sample set is then sorted into normal operating condition samples and fault operating condition samples to obtain a labeled sparse feature training sample set.

[0084] In this embodiment, the sparse feature sample set is organized into normal operating condition samples and fault operating condition samples, including:

[0085] Perform a normalization transformation on each sparse feature representation to convert the sparse feature representation into a one-dimensional sparse feature vector;

[0086] The normalization transformation is performed by squaring all activation coefficients corresponding to each convolution dictionary atom, summing them, and then taking the square root of the summation result to obtain the aggregated response intensity of the convolution dictionary atom on the corresponding signal segment. The aggregated response intensities of all convolution dictionary atoms are arranged in order to form the one-dimensional sparse feature vector corresponding to the sparse feature representation.

[0087] Normalize each one-dimensional sparse feature vector to obtain the normalized sparse feature vector.

[0088] The normalization process is as follows: square all elements of the one-dimensional sparse feature vector, sum them, and then take the square root. Divide each element of the original one-dimensional sparse feature vector by the square root result to obtain the normalized sparse feature vector.

[0089] All normalized sparse feature vectors are combined according to the order of signal segments to obtain a normalized sparse feature sample set.

[0090] Each element of the normalized sparse feature sample set is a normalized sparse feature vector. The number of elements in the normalized sparse feature sample set is equal to the total number of sequence signal segments during boiler operation. The normalized sparse feature sample set is a two-dimensional matrix with the number of rows equal to the total number of sequence signal segments during boiler operation and the number of columns equal to the number of atoms in the convolution dictionary.

[0091] The normalized sparse feature sample set is organized, and each normalized sparse feature vector and its corresponding running state label are combined to form a labeled sparse feature training sample set.

[0092] Each labeled sparse feature training sample includes a normalized sparse feature vector and an operating status label. The operating status label takes the value of zero or one, where zero represents a normal low-load operating condition sample and one represents a fault precursor sample. The operating status label is obtained from historical manual annotation or expert system feedback.

[0093] Using a set of labeled sparse feature training samples as training data, a boundary discrimination enhancement classification model is established. The decision boundary is optimized by maximizing the inter-class margin and introducing a slack variable penalty term, thus obtaining the boundary discrimination enhancement decision boundary parameters.

[0094] In this embodiment, a boundary discrimination enhancement classification model is established, including:

[0095] A boundary discrimination enhancement classification model is constructed. During the training process of the boundary discrimination enhancement classification model, different weights are assigned to the training errors of normal low load condition samples and fault precursor samples by setting different penalty coefficients. The optimization objective is to minimize the radius of the hypersphere, while ensuring that the normal low load condition samples are contained within the hypersphere or close to the hypersphere boundary, and the fault precursor samples are located outside the hypersphere or far from the hypersphere boundary.

[0096] ;

[0097] ;

[0098] in, denoted by the hypersphere radius, it is used to limit the maximum envelope range of the sparse feature space under normal boiler operating conditions. Its value determines the maximum feature distance that samples under normal low-load conditions are included in the boundary discrimination enhancement classification model. c is the center of the hypersphere. Indicates the first Each training sample has a relaxation variable, and all relaxation variables take values ​​no less than zero. This represents the penalty coefficient corresponding to all normal samples with a running status label of zero. This represents the penalty coefficient corresponding to all fault precursor samples with an operating status label of one. The penalty coefficient for normal low-load operating condition samples is less than that for fault precursor samples, reflecting a higher penalty for false fault diagnosis. Indicates the first The sparse feature vector of the boiler's operating sequence signal segment after normalization Indicates the first The running status labels of sparse feature samples are used to guide the boundary discrimination direction of the boundary discrimination enhancement classification model. This represents the total number of training samples with labeled sparse features. The number of atoms in the convolution dictionary is represented by the decision boundary hypersphere, which is a multidimensional sphere uniquely determined by its radius and center in the normalized sparse feature space. The envelope surface of the normal operating condition feature distribution is used to distinguish between normal and abnormal states.

[0099] In this embodiment, the boundary discrimination enhancement classification model's constraints restrict the Euclidean distance between the normalized sparse feature vector of the normal low-load operating condition sample and the center of the hypersphere to no greater than the hypersphere radius plus the relaxation variable. This controls whether the normal low-load operating condition sample is contained within the hypersphere or allowed to slightly exceed the boundary, adapting to the fluctuations in normal operating conditions under boiler low-load operation. Furthermore, by restricting the Euclidean distance between the normalized sparse feature vector of the fault precursor sample and the center of the hypersphere to no less than the hypersphere radius minus the relaxation variable, this strongly constrains whether the fault precursor sample falls outside the hypersphere or is far from the hypersphere boundary, thus enhancing the boundary discrimination capability.

[0100] An alternating optimization strategy is adopted to update the hypersphere radius, hypersphere center, and all relaxation variables in sequence, so that the value of the optimization objective function gradually decreases. The optimal values ​​of the hypersphere radius, hypersphere center, and all relaxation variables obtained together constitute the optimal decision parameter set of the boundary discrimination enhancement classification model.

[0101] In practical applications, the boundary discrimination enhancement classification model outputs, for each normalized sparse feature vector, a discrimination result indicating whether it is within the normal operating condition decision boundary, a minimum Euclidean distance index between the sparse feature vector and the decision boundary, and a decision level indicator for multi-level early warning.

[0102] The optimal set of decision parameters includes the optimal hypersphere radius, the optimal hypersphere center, and the optimal values ​​of all slack variables.

[0103] This implementation introduces a supervised boundary discrimination enhanced classification model, which can fully utilize the label information of normal operating condition samples and fault precursor samples in the scenario of boiler low load operation. It uses an asymmetric penalty term to distinguish the influence of the two types of samples on the decision boundary, thereby improving the separability of fault precursor samples and normal low load operating condition samples in the sparse feature space. The boundary discrimination enhanced classification model can not only adapt to the fluctuation of normal operating conditions under boiler low load operation and improve the sensitivity of abnormal state identification, but also effectively reduce the false alarm and false alarm rates, enhance the discrimination and robustness of the early warning system to the actual operating conditions, and realize early and accurate early warning of boiler combustion anomalies.

[0104] Following the online inference process, the sparse feature representation generated in real time is input into the boundary discrimination enhancement classification model, and the output is the working condition label and the distance index to the decision boundary;

[0105] In this embodiment, the output operating condition status label and the distance index from the decision boundary include:

[0106] The normalized sparse feature vector is input into the boundary discrimination enhancement classification model, and the optimal decision parameter set is used to discriminate the real-time working condition. The Euclidean distance between the normalized sparse feature vector and the center of the optimal hypersphere is calculated.

[0107] Euclidean distance is used to measure the distance between the current real-time signal segment and the center of the optimal hypersphere in the sparse feature space.

[0108] Based on the relationship between the current Euclidean distance and the optimal hypersphere radius, output the current operating condition label:

[0109] If the current Euclidean distance is less than or equal to the optimal hypersphere radius, the operating condition label is zero, indicating that the boiler is currently operating under normal low load conditions.

[0110] If the current Euclidean distance is greater than the optimal hypersphere radius, the operating condition label is one, indicating that the current boiler operating condition is determined to be a precursor to a fault.

[0111] Based on the distance index of the decision boundary and the preset multi-level threshold mapping rules, the corresponding level of early warning results are generated.

[0112] In this embodiment, based on the distance index of the decision boundary and the preset multi-level threshold mapping rules, a warning result of the corresponding level is generated, including:

[0113] If the current distance index is less than or equal to the optimal hypersphere radius, the warning level is Level 1, indicating that the boiler is operating under normal low load conditions and there is no obvious abnormal deviation.

[0114] If the current distance index is greater than the optimal hypersphere radius and less than or equal to the first warning threshold, the warning level is Level II, indicating that the boiler operating status shows a slight deviation from the normal envelope and there is initial combustion disturbance.

[0115] If the current distance indicator is greater than the first warning threshold and less than or equal to the second warning threshold, the warning level is level three, indicating that the boiler operating status deviates significantly from the normal envelope and there is a risk of unstable combustion.

[0116] If the current distance indicator is greater than the second warning threshold, the warning level is level four, indicating that the boiler's operating status is seriously deviating from the normal envelope and is suspected to be in a state of impending failure or a high-risk operating state.

[0117] In Example 1, Euclidean distance is used as the decision boundary distance index, denoted as the current distance index, to measure the degree of deviation of the real-time boiler operating condition from the normal operating condition boundary. The current distance index represents the distance difference between the normalized sparse feature vector and the center of the optimal hypersphere, and truly reflects the degree of closeness or deviation of the boiler combustion state from the normal operating envelope in the sparse feature space.

[0118] Based on the relationship between the current distance index and the optimal hypersphere radius, and combined with the preset multi-level threshold mapping rules, the boiler operating status is classified into multiple levels of early warning. The preset multi-level threshold mapping rules include multiple early warning thresholds. Each early warning threshold is determined based on historical boiler operating data, manually marked early warning levels, and statistical feature boundaries. It is used to classify different boiler operating risk levels. The multiple early warning thresholds are ordered in sequence to distinguish between normal state, slightly abnormal state, moderately abnormal state, and severely abnormal state.

[0119] Example 2: During a deep peak-shaving operation of the boiler, the system collected multi-source signals including furnace negative pressure, flame intensity, induced draft fan current, oxygen content, and coal feed rate. The original timing signal sampling period was 1 second, and the continuous operation duration was 10 minutes. The collected data is as follows (partial excerpt):

[0120] Furnace negative pressure signal (Pa): [-730, -722, -718, -980, -721, -723, -715, -1130, -718, -719, ...];

[0121] Flame intensity signal (au): [258, 260, 262, 935, 259, 257, 256, 1380, 261, 260, ...];

[0122] Exhaust fan current signal (A): [185, 187, 189, 340, 186, 185, 185, 390, 188, 186,...];

[0123] Oxygen levels (%): [4.4, 4.5, 4.3, 5.7, 4.5, 4.4, 4.4, 5.9, 4.3, 4.5, ...];

[0124] Coal feed rate signal (kg / min): [27, 27, 27, 41, 27, 27, 27, 39, 28, 27, ...];

[0125] All signals are automatically aligned, sampled at a uniform rate, and formatted. The system detects abnormal spikes in each sensor at the 4th and 8th seconds, retains the original pulses without filtering, and inputs them completely into the subsequent process.

[0126] The system divides the aforementioned 10-minute signal into 600 sampling segments (each segment being 60 seconds long). Each segment is convolved pointwise with 80 atoms from the convolution dictionary to initialize sparse activation coefficients. The training dictionary is initialized using PCA-reduced waveforms of 1000 historical normal signal segments, with signal segment y as the basis. n Convolution dictionary D and coefficient graph x n With the input as the maximum correlation entropy loss + L1 regularization + group sparse regularization objective, the furnace negative pressure anomaly and convolution reconstruction residual reached 326 Pa at the 4th second during training, and the Gaussian kernel bandwidth σ=0.72. Since it is much larger than σ, this point is automatically assigned a very low weight. After training convergence, the pulse-like anomaly at the 8th second has no dominant influence on the model, and the obtained sparse coefficients are only significantly activated at the combustion dynamic anomaly.

[0127] For a newly sampled 60-second signal segment, the system uses the trained D and parameter set to infer and obtain a sparse coefficient map. Taking a certain segment as an example, the peak value of the flame intensity coefficient (48th atom) is 0.062, and most of the other atoms are below 0.007. After aggregation, the feature vector x3 = [0.011, 0.008, ..., 0.062, ..., 0.007] is obtained.

[0128] For all x n After normalization, the square root of the sum of squares of the 80-dimensional features of the third segment in Example 2 is 0.106. After normalization, the 48th dimension is 0.584, and most of the others are below 0.07. All normalized features correspond one-to-one with manual or historical labels (normal / fault precursors) and are combined into a training sample set. In Example 2, the label y3 of the third segment is 1 (the on-site ignition system inspection record shows that an abnormal flame appears at this time).

[0129] The boundary discrimination enhancement classification model takes the labeled training sample set as input and constructs a boundary discrimination enhancement hypersphere with a final radius R=1.29 and center c=[0.010, 0.007, ..., 0.011]. For normal samples with y=0, the distance to the center is 1.12 (less than 1.29); for fault precursor samples with y=1, the distance to the center is 1.36 (greater than 1.29). C0=0.9, C1=4.2. Five normal samples are relaxed by the boundary, and 13 fault samples are forcibly penalized outside the boundary. After optimization and convergence, the in-training set discrimination accuracy is 96.3%.

[0130] The system infers from the real-time data stream: if a segment of normalized sparse feature vector xᵣ has a distance dᵣ=1.24 from the center, which is lower than R=1.29, the operating condition is judged as normal; if a new segment has dᵣ=1.41, the system directly judges it as a precursor to a fault. Each inference records the Euclidean distance and label discrimination.

[0131] Based on the difference between dᵣ and R from the decision boundary, a three-level early warning threshold is set: dᵣ≤1.16 is safe (green light), 1.16<dᵣ≤1.29 is a first-level early warning (yellow light), and dᵣ>1.29 is a second-level early warning (red light). During the actual monitoring period, the system issued 27 first-level early warnings (23 of which were confirmed on-site as normal fluctuations in the regulating condition, with no false alarms), and 7 second-level early warnings (all related to abnormal boiler flames; 3 were identified as pre-extinguishing signs within 30 minutes before on-site ignition, all of which were manually confirmed). In a comparative experiment, using the traditional L2 norm SVM, there were 19 false alarms and 7 missed alarms out of 600 data segments, while the method of this invention only had 4 false alarms and no missed alarms, improving the average early warning time by 16 minutes.

[0132] Table 1. Examples of partial comparative data (taking flame light intensity signal segment as an example)

[0133] Signal segment number Operating condition label Traditional method distance Traditional discrimination The present invention is far from dᵣ This invention distinguishes Early warning level of this invention 107 0 1.11 normal 1.12 normal Safety 118 1 1.28 normal 1.36 Early signs of failure Level II Warning 121 0 1.22 Warning 1.14 normal Level 1 124 1 1.27 normal 1.31 Early signs of failure Level II Warning

[0134] In actual operation and maintenance decision-making scenarios, the system significantly improves the safe operation level of low-load boilers. Example 2 fully demonstrates the innovation and effectiveness of the present invention in suppressing impulse noise, extracting abnormal features with high precision, and enhancing boundary discrimination and early warning capabilities for low-load boiler operation.

[0135] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A boiler low-load operation early warning method based on sparse representation and enhanced boundary discrimination, characterized in that, include: Collect multi-source time-series data of boiler operation and perform preprocessing to obtain preprocessed multi-source time-series data of boiler operation; Using a preprocessed multi-source time-series dataset of boiler operation as input, an offline maximum correlation entropy convolutional sparse coding model is trained. The preprocessed multi-source time-series dataset of boiler operation is input into the maximum correlation entropy convolutional sparse coding model to generate a sparse feature representation describing the boiler combustion dynamics. The sparse feature representation is normalized to form a sparse feature sample set. The sparse feature sample set is then sorted into normal operating condition samples and fault operating condition samples to obtain a labeled sparse feature training sample set. Using a set of labeled sparse feature training samples as training data, a boundary discrimination enhancement classification model is established. The decision boundary is optimized by maximizing the inter-class margin and introducing a slack variable penalty term, thus obtaining the boundary discrimination enhancement decision boundary parameters. Following the online inference process, the sparse feature representation generated in real time is input into the boundary discrimination enhancement classification model, and the output is the working condition label and the distance index to the decision boundary; Based on the distance index of the decision boundary and the preset multi-level threshold mapping rules, the corresponding level of early warning results are generated.

2. The boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement according to claim 1, characterized in that, The process of collecting multi-source time-series data on boiler operation and performing preprocessing includes: The furnace negative pressure signal, induced draft fan current signal, flame intensity signal, coal feed rate signal, and oxygen content signal are collected during the low-load operation of the boiler. Time synchronization, sampling rate unification, and format standardization are performed to construct a multi-source time series dataset of boiler operation. Non-Gaussian impulse noise preprocessing is performed on the multi-source time series dataset of boiler operation to preserve outliers with spike characteristics in the signal, resulting in a preprocessed multi-source time series dataset of boiler operation.

3. The boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement according to claim 1, characterized in that, The maximum correlation entropy convolutional sparse coding model includes: The pre-processed set of boiler operation sequence signals is represented as multiple boiler operation sequence signal segments, and all boiler operation sequence signal segments constitute the set of boiler operation sequence signals. Initialize the convolution dictionary, which consists of multiple convolution atoms. Initialize sparse activation values ​​corresponding to all convolution atoms and all sampling points for each boiler operation sequence signal segment. All sparse activation values ​​form a sparse coefficient map. Using the boiler operating sequence signal set, convolution dictionary, and sparse coefficient map as input, an optimization objective function for the maximum correlation entropy convolutional sparse coding model is constructed. With a fixed convolution dictionary, the proximal gradient descent algorithm is used to iteratively update each sparse activation coefficient in the sparse coefficient graph. With a fixed sparse coefficient graph, a variant of the fast iterative shrinking threshold algorithm is used to update the gradient for each convolution atom in the convolution dictionary; During each training iteration, the Gaussian kernel bandwidth parameter is dynamically adjusted using an annealing strategy; When the convergence condition is met or the maximum number of iterations is reached, the final maximum correlation entropy convolutional sparse coding model is obtained.

4. The boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement according to claim 1, characterized in that, The sparse feature representation describing boiler combustion dynamics includes: Each boiler runtime sequence signal segment in the preprocessed boiler runtime sequence signal set is used as input, and the trained maximum correlation entropy convolutional sparse coding model is invoked. For each boiler operation sequence signal segment, a sparse decoding process is performed under the premise of a fixed convolutional dictionary to obtain the sparse activation coefficient map corresponding to the boiler operation sequence signal segment. The sparse activation coefficient maps obtained from all boiler operation sequence signal segments are uniformly organized to construct a sparse feature representation set.

5. The boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement according to claim 1, characterized in that, The process of organizing the sparse feature sample set into normal operating condition samples and fault operating condition samples includes: Perform a normalization transformation on each sparse feature representation to convert the sparse feature representation into a one-dimensional sparse feature vector; Normalize each one-dimensional sparse feature vector to obtain the normalized sparse feature vector. All normalized sparse feature vectors are combined according to the order of signal segments to obtain a normalized sparse feature sample set. The normalized sparse feature sample set is organized, and each normalized sparse feature vector and its corresponding running state label are combined to form a labeled sparse feature training sample set.

6. The boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement according to claim 1, characterized in that, The establishment of the boundary discrimination enhanced classification model includes: A boundary discrimination enhancement classification model is constructed. During the training process of the boundary discrimination enhancement classification model, different weights are assigned to the training errors of normal low load condition samples and fault precursor samples by setting different penalty coefficients. The optimization objective is to minimize the radius of the hypersphere, while ensuring that the normal low load condition samples are contained within the hypersphere or close to the hypersphere boundary, and the fault precursor samples are located outside the hypersphere or far from the hypersphere boundary. An alternating optimization strategy is adopted to update the hypersphere radius, hypersphere center, and all relaxation variables in sequence, so that the value of the optimization objective function gradually decreases. The optimal values ​​of the hypersphere radius, hypersphere center, and all relaxation variables obtained together constitute the optimal decision parameter set of the boundary discrimination enhancement classification model.

7. The boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement according to claim 1, characterized in that, In the boundary discrimination enhancement classification model, the constraints restrict the Euclidean distance between the normalized sparse feature vector of the normal low-load operating condition sample and the center of the hypersphere to no greater than the hypersphere radius plus the relaxation variable. This controls whether the normal low-load operating condition sample is contained within the hypersphere or allowed to slightly exceed the boundary, adapting to the fluctuations in normal operating conditions under boiler low-load operation. The constraints also restrict the Euclidean distance between the normalized sparse feature vector of the fault precursor sample and the center of the hypersphere to no less than the hypersphere radius minus the relaxation variable. This provides a strong constraint that the fault precursor sample falls outside the hypersphere or is far from the hypersphere boundary, thus enhancing the boundary discrimination capability.

8. The boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement according to claim 1, characterized in that, The output operating condition status label and the distance index from the decision boundary include: The normalized sparse feature vector is input into the boundary discrimination enhancement classification model, and the optimal decision parameter set is used to discriminate the real-time working condition. The Euclidean distance between the normalized sparse feature vector and the center of the optimal hypersphere is calculated. Based on the relationship between the current Euclidean distance and the optimal hypersphere radius, output the current operating condition label: If the current Euclidean distance is less than or equal to the optimal hypersphere radius, the operating condition label is zero, indicating that the boiler is currently operating under normal low load conditions. If the current Euclidean distance is greater than the optimal hypersphere radius, the operating condition label is one, indicating that the current boiler operating condition is determined to be a precursor to a fault.

9. The boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement according to claim 1, characterized in that, The distance index based on the decision boundary and the preset multi-level threshold mapping rule generate corresponding warning results, including: If the current distance index is less than or equal to the optimal hypersphere radius, the warning level is Level 1, indicating that the boiler is operating under normal low load conditions and there is no obvious abnormal deviation. If the current distance index is greater than the optimal hypersphere radius and less than or equal to the first warning threshold, the warning level is Level II, indicating that the boiler operating status shows a slight deviation from the normal envelope and there is initial combustion disturbance. If the current distance indicator is greater than the first warning threshold and less than or equal to the second warning threshold, the warning level is level three, indicating that the boiler operating status deviates significantly from the normal envelope and there is a risk of unstable combustion. If the current distance indicator is greater than the second warning threshold, the warning level is level four, indicating that the boiler's operating status is seriously deviating from the normal envelope and is suspected to be in a state of impending failure or a high-risk operating state.

Images