CN-122020023-A - Equipment fault prediction and safety early warning system based on deep learning for thermal power plant

CN122020023ACN 122020023 ACN122020023 ACN 122020023ACN-122020023-A

Abstract

The invention relates to the technical field of thermal power plant fault prediction, in particular to a thermal power plant equipment fault prediction and safety early warning system based on deep learning. The method comprises a data acquisition and processing unit, a characteristic construction unit, a wear characteristic identification unit, a fault prediction evaluation unit and a bearing failure risk level, wherein the data acquisition and processing unit is used for acquiring bearing operation data, the characteristic construction unit is used for carrying out time-frequency explicit processing on a preprocessed bearing high-frequency vibration signal, the wear characteristic identification unit is used for extracting weak transient characteristics in a multi-channel time-frequency characteristic tensor, weak transient wear probability of each bearing node is obtained through a unit weak coupling characteristic decoder, and the fault prediction evaluation unit is used for constructing a bearing degradation evolution model based on the corrected bearing wear probability and outputting the bearing failure risk level. According to the invention, by constructing a device operation topological graph structure and utilizing a message transmission and feature aggregation mechanism of a graph neural network, cross-node transmission and fusion of operation state information among nodes are realized, and isolated abnormal response caused by single sensor noise or local interference is restrained.

Inventors

ZHANG BIXIAO
LI SHIBO
MENG FANLIANG
GAO YONG
YANG CHUNLEI
LIN HAIBO
CHEN YONGKANG
ZHANG XIAORONG
JIANG FENGLIN
ZHANG YIKAI

Assignees

华电济南章丘热电有限公司

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (9)

1. A thermal power plant equipment fault prediction and safety early warning system based on deep learning is characterized by comprising: the data acquisition and processing unit (1), the said data acquisition and processing unit (1) is used for gathering the bearing operation data, and adopt wavelet packet to make an uproar, sample and synchronize and signal normalization process the bearing operation data gathered, get the preprocessed dataset; the bearing operation data at least comprises a bearing high-frequency vibration signal, a shaft temperature, a bearing rotating speed, load data, a lubricating oil temperature and a lubricating oil pressure; the characteristic construction unit (2) is used for performing time-frequency explicit processing on the preprocessed bearing dither signal, and respectively constructing a multichannel time-frequency characteristic tensor by extracting Hilbert envelope, STFT time-frequency diagram and wavelet transient peak characteristics ; A wear feature recognition unit (3), wherein the wear feature recognition unit (3) extracts multichannel time-frequency feature tensors based on a weak transient recognition network The weak transient characteristics of the branches with different physical mechanisms are subjected to trans-scale and trans-modal decoding through a unit weak coupling characteristic decoder, and finally the weak transient wear probability of each bearing node is obtained; The weak transient identification network is constructed by introducing a multi-class discriminator based on a full-connection layer and a Softmax activation function based on a multi-scale convolutional coding network, a unit coupling enhancement network and a multi-branch fusion network; the equipment structure diagram association unit (4), the equipment structure diagram association unit (4) constructs an equipment operation topological diagram structure based on vibration, temperature and load, adopts a graph neural network to perform cross-node feature fusion, and outputs bearing wear probability after structural constraint correction; The failure prediction evaluation unit (5), the failure prediction evaluation unit (5) constructs a bearing degradation evolution model based on the corrected bearing wear probability, and outputs a bearing failure risk level.
2. The deep learning-based equipment fault prediction and safety early warning system of the thermal power plant according to claim 1 is characterized in that the data acquisition and processing unit (1) preprocesses bearing operation data, and comprises the following specific steps: synchronously acquiring multi-source data of a target bearing under actual operation conditions to obtain an original input data set Wherein the original input data set Comprising bearing vibration signals Temperature of shaft Rotational speed Load(s) Temperature of lubricating oil Lubricating oil pressure ; The collected bearing vibration signal Separating the high-frequency sub-bands through wavelet packet decomposition to obtain a noise-reduced high-frequency vibration signal ; Noise-reduced dither signal With the temperature of the shaft Rotational speed Load(s) Temperature of lubricating oil Lubricating oil pressure Synchronizing time axis to obtain synchronous data set ; For the synchronized shaft temperature Rotational speed Load(s) Temperature of lubricating oil Lubricating oil pressure Normalization processing is carried out to obtain a working condition state vector ; Meanwhile, the natural structural frequency of a unit rotating part is introduced to construct a structural feature vector The natural structural frequency of the unit rotating component comprises a bearing rolling body passing frequency, a rotor fundamental frequency and higher harmonics; based on the noise-reduced dither signal Working condition state vector Structural feature vector Constructing a preprocessed dataset 。
3. The deep learning-based equipment fault prediction and safety early warning system of a thermal power plant according to claim 2, wherein the feature construction unit (2) comprises a channel feature extraction module (21) and a multi-channel feature combination module (22); The channel characteristic extraction module (21) respectively reduces noise of the high-frequency vibration signals Hilbert transformation is carried out to obtain envelope energy As a first channel feature; For the high-frequency vibration signal after noise reduction Performing short-time Fourier transform to obtain a time-frequency spectrum matrix As a second channel feature; For the high-frequency vibration signal after noise reduction Extracting transient peak values under different scales by Morlet continuous wavelet transformation As a third channel feature; The multi-channel feature combination module (22) is used for combining the first channel feature, the second channel feature, the third channel feature and the working condition state vector Structural feature vector Performing channel-level splicing to construct multi-channel time-frequency characteristic tensors 。
4. The deep learning-based equipment fault prediction and safety early warning system for the thermal power plant according to claim 1 is characterized in that in the wear characteristic recognition unit (3), probability distribution of each type of weak transient state is obtained through the weak transient recognition network, and the specific steps involved are as follows: tensor of multi-channel time-frequency characteristics Inputting the multi-scale convolution coding network, extracting weak transient primitive features with different durations through a parallel convolution kernel group, and obtaining multi-scale weak transient coding features ; Based on structural feature vectors And operating mode state vector The method comprises the steps of constructing a unit coupling enhancement network, wherein the unit coupling enhancement network comprises a fundamental frequency resonance detection branch, an oil film disturbance detection branch and a thermal coupling impact detection branch, and the fundamental frequency resonance detection branch, the oil film disturbance detection branch and the thermal coupling impact detection branch are respectively used for extracting weak transient characteristics caused by natural frequency excitation, lubricating oil film disturbance and thermal unbalance of a structure in the running process of a unit, and a unit weak coupling characteristic decoder is introduced and used for carrying out time sequence alignment, signal modulation, semantic fusion and coupling structure decoding on the weak transient characteristics to obtain unit-level weak transient coupling characteristics ; Characterizing multi-scale weak transient encodings Weak transient coupling features to unit level Inputting the final weak transient characteristic vector into a multi-branch fusion network, and obtaining the final weak transient characteristic vector through residual error splicing, cross-channel attention fusion and characteristic scaling ; Eventually, the weak transient characteristic vector is calculated And inputting the data to a multi-class discriminator formed by the full connection layer and the Softmax, and obtaining the weak transient wear probability of each bearing node.
5. The deep learning-based equipment fault prediction and safety early warning system for a thermal power plant according to claim 4, wherein weak transient characteristics are extracted through the unit coupling enhancement network, and the specific steps involved are as follows: fundamental frequency resonance detection branch circuit is based on structural feature vector Constructing bandpass convolution in defined natural frequency interval for extracting weak resonance transient to obtain characteristic of weak transient of fundamental frequency resonance ; Oil film disturbance detection branch circuit is based on operating mode state vector The method comprises the steps of constructing a weak circulation disturbance extraction model, wherein the weak circulation disturbance extraction model is constructed by a small-scale convolution network and a gating circulation unit, an oil film disturbance detection branch is used for extracting oil film low-frequency disturbance primitives based on the small-scale convolution network, modeling disturbance periodicity by the gating circulation unit and introducing a working condition state vector The extracted features are weighted by the relevance of the working conditions as amplitude modulation factors to obtain the oil film disturbance weak transient features ; Thermal coupling impact detection branch circuit based on structural feature vector And operating mode state vector The thermodynamic coupling weak impact extraction model of the multi-head time sequence attention mechanism is constructed, and impact response intensities under different time windows are adaptively adjusted through attention weights, so that thermodynamic coupling weak transient characteristics are obtained 。
6. The deep learning-based equipment fault prediction and safety pre-warning system of a thermal power plant according to claim 5, wherein the weak transient characteristics of the branches with different physical mechanisms are decoded by the unit weak coupling characteristic decoder to obtain unit-level weak transient coupling characteristics The specific steps involved are: Weak transient characteristic of fundamental frequency resonance by unit weak coupling characteristic decoder Weak transient characteristics of oil film disturbance Thermally coupled weak transient features Performing time domain alignment decoding to obtain an alignment weak transient characteristic set; the alignment weak transient characteristic is subjected to dynamic amplitude modulation and credibility weighting by introducing a joint modulation mechanism of a structural characteristic vector and a working condition state vector, so that a modulation weak transient characteristic is obtained; after being modulated by a joint modulation mechanism, a unit weak coupling characteristic decoder carries out cross-branch interactive weak coupling modeling on the modulation weak transient characteristic under different physical mechanisms through a bilinear gating fusion mechanism to obtain a cross-mechanism coupling weak transient characteristic; And introducing a characteristic attention mechanism at the decoding end for carrying out correlation enhancement and noise suppression on the cross-mechanism coupling weak transient characteristic to obtain a unit-level weak transient coupling characteristic 。
7. The deep learning-based equipment fault prediction and safety early warning system of the thermal power plant according to claim 1, wherein the equipment structure diagram association unit (4) comprises an equipment structure topology modeling module (41), a cross-node diagram feature fusion module (42) and a correction output module (43); The equipment operation structure topology modeling module (41) constructs an equipment structure node set based on a bearing, a rotor, a coupler, a load transmission part and a supporting structure in the unit, constructs an equipment operation topology graph structure based on a mechanical transmission relation, a thermal coupling relation and a load association relation among vibration, temperature and load, and generates an adjacent matrix used for representing the coupling strength of the equipment nodes; the cross-node diagram feature fusion module (42) is used for outputting unit-level weak transient coupling features output by the wear feature identification unit (3) Mapping the corresponding vibration, temperature and load characteristics into node characteristic vectors, and under the structural constraint of the equipment operation topological graph, realizing cross-node propagation and fusion of running state information among nodes based on a message transmission and characteristic aggregation mechanism of a graph neural network to obtain node-level structural fusion characteristic vectors under the structural constraint; The correction output module (43) is used for receiving the weak transient wear probability of each bearing node and the node level structure fusion feature vector output by the cross-node diagram feature fusion module (42), and based on the equipment operation topological graph structure as a structural constraint condition for wear state correction, performing topology consistency correction on the node level structure fusion feature vector based on the weak transient wear probability, and performing probability mapping on node response subjected to structural constraint correction to obtain a bearing wear probability set subjected to structural constraint correction, wherein the bearing wear probability set is used for representing the real wear state of each bearing in a turbo generator set shafting of a thermal power plant.
8. The deep learning-based equipment fault prediction and safety early warning system for a thermal power plant according to claim 1, wherein the fault prediction evaluation unit (5) comprises a wear probability time sequence modeling module (51), a working condition constraint degradation evolution module (52), a segmentation trend prediction module (53) and a failure risk evaluation module (54); The wear probability time sequence modeling module (51) is used for correcting the bearing wear probability set after structural constraint correction Performing time sequence modeling, and constructing the wear probability of each bearing node in continuous time as a wear probability time sequence state matrix, wherein the wear probability is used as basic state input of bearing degradation evolution analysis; The working condition constraint degradation evolution module (52) is used for introducing the operation working condition parameters of the steam turbine generator unit of the thermal power plant on the basis of the abrasion probability time sequence state matrix, carrying out working condition differentiation and evolution constraint modeling on the bearing abrasion degradation process, and outputting a bearing abrasion degradation evolution state sequence set corresponding to different operation working condition intervals ; The operating condition parameters at least comprise a unit load, a load change rate, a start-stop state and a thermal condition change parameter, and are used for distinguishing differences of bearing wear evolution mechanisms of the unit in rated load operation, peak regulation operation and start-stop process; The sectional trend prediction module (53) is used for receiving the bearing wear degradation evolution state sequence, and based on the division result of the operation working condition interval, adopting a time sequence prediction model to conduct sectional prediction on the bearing wear degradation evolution state sequence of each bearing node in different operation working condition intervals, outputting a wear probability trend prediction set of the bearing in different operation working condition intervals, and describing the evolution trend change condition of the bearing wear state in different operation phases; The failure risk assessment module (54) is used for assessing the failure risk of each bearing in the future operation stage based on the abrasion probability trend prediction set of the bearing in different operation working condition intervals, outputting a corresponding failure risk level and reaching a preset safety early warning threshold time window, and is used for quantitatively representing the potential failure risk level of the bearing in different operation stages.
9. The deep learning-based equipment fault prediction and safety early warning system for a thermal power plant according to claim 8, wherein the working condition constraint degradation evolution module (52) outputs a bearing wear degradation evolution state sequence corresponding to different operation working condition intervals, and the specific steps involved are as follows: in successive time steps Simultaneously acquiring the abrasion probability time sequence state matrix, synchronously acquiring the operation condition parameters of the steam turbine generator unit of the thermal power plant, and constructing an operation condition parameter time sequence vector ; Based on the operating condition parameter timing vector The operation states on the time axis are distinguished through an operation condition judging rule, and the operation states are used for realizing automatic division of operation condition intervals on the time axis; Based on the division result of the operation condition interval, dividing the abrasion probability time sequence state matrix along the time dimension according to the operation condition interval to generate abrasion probability subsequences corresponding to different operation conditions; aiming at each operation working condition category, a working condition constraint rule is introduced, a permissible change interval is set for different working conditions, and a time step change quantity sequence under the working condition constraint is output; Constraint mapping is carried out on the wear probability state, and finally a bearing wear degradation evolution state sequence set corresponding to different operation working condition intervals is generated 。

Description

Equipment fault prediction and safety early warning system based on deep learning for thermal power plant Technical Field The invention relates to the technical field of thermal power plant fault prediction, in particular to a thermal power plant equipment fault prediction and safety early warning system based on deep learning. Background The thermal power plant is used as a core component of an electric power system, the long-period safe and stable operation of the thermal power plant is crucial to the guarantee of the reliability of a power grid, large rotary machines such as a steam turbine generator unit and the like are key equipment of the thermal power plant, the operation environment of the thermal power plant is bad, the thermal power plant bears high temperature, high pressure and heavy load for a long time, the rolling bearing is used as a core component for supporting a rotor in the rotary machine, the rolling bearing is one of the components with the highest failure rate, once the bearing fails, the equipment vibration exceeds standard and the efficiency is reduced due to light weight, and catastrophic accidents such as tile burning, machine destruction and the like are caused due to heavy weight. The method mainly depends on a threshold value alarming or shallow layer characteristic analysis method based on vibration signals aiming at fault monitoring and diagnosis of rotating equipment of a thermal power plant, and is forced to frequently participate in deep peak shaving along with large-scale grid connection of new energy, and the thermal power unit runs under non-rated load for a long time and is accompanied with frequent lifting load and start-stop operation. In the unsteady state process, the vibration amplitude of equipment, the bearing bush temperature and the lubricating oil pressure can naturally fluctuate along with the change of working conditions, the existing fault prediction model is mostly based on steady state working condition design, and the lack of automatic identification and self-adaptive constraint mechanisms for different operation working conditions (rated load, variable load peak regulation and start-stop process) leads to the fact that the system is extremely easy to misjudge normal working condition response fluctuation (such as vibration climbing during variable load) as equipment performance degradation or faults, and the early warning accuracy is greatly reduced, so that a deep learning-based equipment fault prediction and safety early warning system for a thermal power plant is needed. Disclosure of Invention The invention aims to provide a deep learning-based equipment fault prediction and safety early warning system for a thermal power plant, which aims to solve the problems that in the background art, the existing fault prediction model is mostly based on steady-state working condition design, and the system is extremely easy to misjudge normal working condition response fluctuation (such as vibration climbing during load change) as equipment performance degradation or fault due to lack of automatic identification and self-adaptive constraint mechanisms for different operation working conditions (rated load, load change peak shaving and start-stop processes), so that early warning accuracy is greatly reduced. In order to achieve the above object, the present invention provides a deep learning based equipment failure prediction and safety early warning system for a thermal power plant, comprising: The data acquisition and processing unit is used for acquiring bearing operation data and adopting wavelet packet noise reduction, sampling synchronization and signal normalization to process the acquired bearing operation data to obtain a preprocessing data set; the bearing operation data at least comprises a bearing high-frequency vibration signal, a shaft temperature, a bearing rotating speed, load data, a lubricating oil temperature and a lubricating oil pressure; the characteristic construction unit is used for carrying out time-frequency explicit processing on the bearing high-frequency vibration signals after pretreatment, and respectively constructing a multichannel time-frequency characteristic tensor by extracting Hilbert envelope, STFT time-frequency diagram and wavelet transient peak characteristics ; The abrasion characteristic identification unit is used for extracting multichannel time-frequency characteristic tensors based on a weak transient identification networkThe weak transient characteristics of the branches with different physical mechanisms are subjected to trans-scale and trans-modal decoding through a unit weak coupling characteristic decoder, and finally the weak transient wear probability of each bearing node is obtained; The weak transient identification network is constructed by introducing a multi-class discriminator based on a full-connection layer and a Softmax activation function based on a multi-scale convolutional coding network, a