CN-122020278-A - Boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement

CN122020278ACN 122020278 ACN122020278 ACN 122020278ACN-122020278-A

Abstract

The invention discloses a boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement, which comprises the steps of obtaining a preprocessing boiler operation multisource time sequence data set, offline training a maximum correlation entropy convolution sparse coding model, generating sparse feature representation describing boiler combustion dynamics, forming a sparse feature sample set, sorting the sparse feature sample set according to normal working condition samples and fault working condition samples to obtain a labeled sparse feature training sample set, establishing a boundary discrimination enhancement classification model to obtain boundary discrimination enhancement decision boundary parameters, outputting working condition state labels and distance indexes with decision boundaries, and generating early warning results of corresponding grades. The invention obviously shortens the fuzzy discrimination interval between the normal working condition and the abnormal working condition and improves the certainty of early abnormal recognition.

Inventors

BAI HUANLIN
HAN XIONG
SONG SHUANGSHUANG

Assignees

陕西华电榆横煤电有限责任公司榆横发电厂
清贺蓝(上海)环保科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251224

Claims (9)

1. A boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement is characterized by comprising the following steps: Collecting a boiler operation multi-source time sequence data set and performing pretreatment to obtain a pretreatment boiler operation multi-source time sequence data set; Taking a preprocessing boiler operation multisource time sequence data set as input, and offline training a maximum correlation entropy convolution sparse coding model; inputting a preprocessing boiler operation multisource time sequence data set into a maximum correlation entropy convolution sparse coding model to generate sparse feature representation describing boiler combustion dynamics; performing normalization conversion on the sparse feature representation to form a sparse feature sample set, and sorting the sparse feature sample set according to the normal working condition sample and the fault working condition sample to obtain a sparse feature training sample set with labels; Taking a labeled sparse feature training sample set as training data, establishing a boundary discrimination enhancement classification model, and optimizing a decision boundary by maximizing inter-class intervals and introducing a relaxation variable penalty term to obtain boundary discrimination enhancement decision boundary parameters; According to an online reasoning flow, inputting the sparse feature representation generated in real time into a boundary discrimination enhancement classification model, and outputting a working condition state label and a distance index with a decision boundary; And generating an early warning result of a corresponding grade based on the distance index of the decision boundary and a preset multi-grade threshold mapping rule.
2. The boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement according to claim 1, wherein the collecting boiler operation multisource time series data sets and performing preprocessing comprises: Collecting a hearth negative pressure signal, an induced draft fan current signal, a flame light intensity signal, a coal supply signal and an oxygen content signal during low-load operation of a boiler, performing time synchronization, sampling rate unification and format standardization, constructing a boiler operation multi-source time sequence data set, performing non-Gaussian pulse noise preprocessing on the boiler operation multi-source time sequence data set, and maintaining abnormal points with peak characteristics in the signals to obtain a preprocessed boiler operation multi-source time sequence data set.
3. The boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement according to claim 1, wherein the maximum correlation entropy convolution sparse coding model comprises the following steps: representing the pretreatment boiler operation time sequence signal set as a plurality of boiler operation time sequence signal segments, wherein all the boiler operation time sequence signal segments form the boiler operation time sequence signal set; initializing a convolution dictionary, wherein the convolution dictionary consists of a plurality of convolution atoms, and initializing sparse activation values corresponding to all the convolution atoms and all sampling points for each boiler operation time sequence signal segment, wherein all the sparse activation values form a sparse coefficient graph; Taking a boiler operation time sequence signal set, a convolution dictionary and a sparse coefficient graph as inputs to construct an optimization objective function of a maximum correlation entropy convolution sparse coding model; On the premise of fixing a convolution dictionary, iteratively updating each sparse activation coefficient in the sparse coefficient map by using a near-end gradient descent algorithm; on the premise of fixing the sparse coefficient graph, gradient updating is carried out on each convolution atom in the convolution dictionary by adopting a rapid iteration shrinkage threshold algorithm variant; in each round of training iteration process, dynamically adjusting Gaussian kernel bandwidth parameters by using an annealing strategy; And when the convergence condition is met or the maximum iteration number is reached, obtaining a final maximum correlation entropy convolution sparse coding model.
4. The method for early warning of low-load operation of a boiler based on sparse representation and boundary discrimination enhancement according to claim 1, wherein the sparse feature representation describing the combustion dynamics of the boiler comprises: Taking each boiler operation time sequence signal segment in the pretreatment boiler operation time sequence signal set as input, and calling a trained maximum correlation entropy convolution sparse coding model; performing a sparse decoding process on the premise of fixing a convolution dictionary aiming at each boiler operation time sequence signal segment to obtain a sparse activation coefficient graph corresponding to the corresponding boiler operation time sequence signal segment; and uniformly organizing the sparse activation coefficient graphs obtained by all the boiler operation time sequence signal segments 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, wherein the sorting of the sparse feature sample set according to the normal working condition sample and the fault working condition sample comprises the following steps: performing normalization conversion on each sparse feature representation, and converting the sparse feature representation into a one-dimensional sparse feature vector; carrying out normalization processing on each one-dimensional sparse feature vector to obtain normalized sparse feature vectors; combining all normalized sparse feature vectors according to the sequence of the signal segments to obtain a normalized sparse feature sample set; and sorting the normalized sparse feature sample set, and forming a labeled sparse feature training sample set by each normalized sparse feature vector and the corresponding running state label.
6. The method for early warning of low-load operation of a boiler based on sparse representation and boundary discrimination enhancement according to claim 1, wherein the establishing a boundary discrimination enhancement classification model comprises: the method comprises the steps of constructing a boundary discrimination enhancement classification model, respectively giving different weights to training errors of a normal low-load working condition sample and a fault precursor sample by setting different punishment coefficients in the boundary discrimination enhancement classification model training process, wherein an optimization target is to minimize the radius of a hypersphere, and simultaneously, the normal low-load working condition sample is contained in or close to the boundary of the hypersphere, and the fault precursor sample is outside or far from the boundary of the hypersphere; And sequentially updating the radius of the hypersphere, the center of the hypersphere and all relaxation variables by adopting an alternative optimization strategy, so that the optimization objective function value is gradually reduced, and the obtained optimal values of the radius of the hypersphere, the center of the hypersphere and all relaxation variables jointly form an 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 is characterized in that in constraint conditions of the boundary discrimination enhancement classification model, the Euclidean distance between a sparse feature vector normalized by a normal low-load working condition sample and the center of a hypersphere is not larger than the radius of the hypersphere plus a relaxation variable, the normal low-load working condition sample is contained in the hypersphere or allowed to slightly cross the boundary is controlled to adapt to normal working condition fluctuation under the boiler low-load operation, and the discrimination capability of the boundary is enhanced by limiting the Euclidean distance between the sparse feature vector normalized by a fault precursor sample and the center of the hypersphere to be not smaller than the radius of the hypersphere minus the relaxation variable, and the fault precursor sample falls outside the hypersphere or is far away from the strong constraint of the hypersphere boundary.
8. The boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement according to claim 1, wherein the output working condition state label and the distance index from the decision boundary comprise: inputting the normalized sparse feature vector into a boundary discrimination enhancement classification model, discriminating the real-time working condition state by adopting an optimal decision parameter set, and calculating the Euclidean distance between the normalized sparse feature vector and the center of an optimal hypersphere; Outputting a current working condition state label according to the relation between the current Euclidean distance and the optimal hypersphere radius: if the current Euclidean distance is smaller than or equal to the optimal hypersphere radius, the working condition state label is zero, and the current working condition of the boiler is judged to be a normal low-load working condition; If the current Euclidean distance is larger than the optimal hypersphere radius, the working condition state label is one, which indicates that the current working condition of the boiler is judged to be the failure precursor.
9. The method for early warning of low-load operation of a boiler based on sparse representation and boundary discrimination enhancement according to claim 1, wherein the generating the early warning result of the corresponding level based on the distance index of the decision boundary and the preset multi-level threshold mapping rule comprises the following steps: if the current distance index is smaller than or equal to the optimal hypersphere radius, the early warning level is a first-level early warning, which means that the running state of the boiler is in a normal low-load working condition and no obvious abnormal offset exists; If the current distance index is larger than the optimal hypersphere radius and smaller than or equal to the first early warning threshold value, the early warning level is a second early warning, and the early warning level shows that the running state of the boiler has a trend of slightly deviating from the normal envelope surface and has initial combustion disturbance; if the current distance index is larger than the first early warning threshold value and smaller than or equal to the second early warning threshold value, the early warning level is three-level early warning, and the running state of the boiler is obviously deviated from the normal envelope surface, so that the risk of unstable combustion exists; if the current distance index is larger than the second early warning threshold, the early warning level is four-level early warning, and the early warning level indicates that the running state of the boiler deviates from the normal envelope surface seriously and is suspected to be in a fault aura state or a high risk running state.

Description

Boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement Technical Field The invention relates to the technical field of boilers, in particular to a boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement. Background Along with the continuous improvement of the flexibility peak regulation requirement of the thermal power plant, the low-load operation of the boiler gradually becomes a normal working condition, under the normal working condition, due to severe fluctuation of the combustion state, frequent regulation of auxiliary machines and complex electrical and control environments, a great amount of non-Gaussian pulse noise is often mixed in boiler operation time sequence signals acquired by on-site sensors, the existing low-load early warning technology of the boiler is mostly based on sparse representation, dictionary learning or a traditional characteristic engineering method, the noise is generally assumed to be subjected to Gaussian distribution basically, and the minimum mean square error is taken as an optimization target, when the interference of the strong pulse noise is faced, the algorithm easily misjudges noise points as fault characteristics, so that weak fault signals are covered or characteristic reconstruction distortion is caused, and the recognition capability of early risks such as combustion instability is reduced. The automatic regulation amplitude of the boiler control system under the low-load working condition is large, the high aliasing easily occurs in the characteristic space between the fluctuation of the normal working condition and the premonition of the fault, the traditional early warning method is mostly dependent on a fixed threshold value or a standard linear classifier, a discrimination boundary enhancement mechanism aiming at a multi-label sample is lacking, high false alarm or false alarm easily occurs, and the balance between the working condition self-adaption and the abnormal recognition is difficult to realize under the dynamic working condition. Disclosure of Invention The invention aims to provide a boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement, the invention obviously shortens the fuzzy discrimination interval between the normal working condition and the abnormal working condition and improves the certainty of early abnormal recognition. According to the embodiment of the invention, the boiler low-load operation early warning method based on sparse representation and boundary discrimination enhancement comprises the following steps: Collecting a boiler operation multi-source time sequence data set and performing pretreatment to obtain a pretreatment boiler operation multi-source time sequence data set; Taking a preprocessing boiler operation multisource time sequence data set as input, and offline training a maximum correlation entropy convolution sparse coding model; inputting a preprocessing boiler operation multisource time sequence data set into a maximum correlation entropy convolution sparse coding model to generate sparse feature representation describing boiler combustion dynamics; performing normalization conversion on the sparse feature representation to form a sparse feature sample set, and sorting the sparse feature sample set according to the normal working condition sample and the fault working condition sample to obtain a sparse feature training sample set with labels; Taking a labeled sparse feature training sample set as training data, establishing a boundary discrimination enhancement classification model, and optimizing a decision boundary by maximizing inter-class intervals and introducing a relaxation variable penalty term to obtain boundary discrimination enhancement decision boundary parameters; According to an online reasoning flow, inputting the sparse feature representation generated in real time into a boundary discrimination enhancement classification model, and outputting a working condition state label and a distance index with a decision boundary; And generating an early warning result of a corresponding grade based on the distance index of the decision boundary and a preset multi-grade threshold mapping rule. Optionally, the collecting boiler runs a multi-source time series data set and performs preprocessing, including: Collecting a hearth negative pressure signal, an induced draft fan current signal, a flame light intensity signal, a coal supply signal and an oxygen content signal during low-load operation of a boiler, performing time synchronization, sampling rate unification and format standardization, constructing a boiler operation multi-source time sequence data set, performing non-Gaussian pulse noise preprocessing on the boiler operation multi-source time sequence data set, and maintaining abnormal points with peak characteristic