CN-121976924-A - Fan transmission chain state monitoring method based on multiple sensors

CN121976924ACN 121976924 ACN121976924 ACN 121976924ACN-121976924-A

Abstract

The invention discloses a fan transmission chain state monitoring method based on multiple sensors, which relates to the technical field of fan state monitoring, and aims to realize multi-source data synchronous acquisition through vision-voiceprint multi-sensor accurate collaborative layout and a unified timestamp trigger mechanism, simplify or supplement point positions according to direct drive/doubly-fed fan model differences, improve layout universality, realize multi-dimensional feature refined extraction through image preprocessing and voiceprint filtering technology, eliminate dimension differences through normalization processing to form standardized feature input, construct a CNN-BiLSTM fusion model with an attention mechanism, realize self-adaptive weight distribution and depth fusion of vision space features and force acoustic time sequence features, and finish accurate quantization output of fusion feature comprehensive values by combining a full-connection-layer multi-layer cascade structure and an output-layer activation function.

Inventors

GUO SONG
LIANG JUN
ZHANG XIANG
ZHANG PU

Assignees

绵阳师范学院
东方电气新能科技(成都)有限公司
德阳经开智航科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260403

Claims (10)

1. The fan transmission chain state monitoring method based on the multiple sensors is characterized by comprising the following specific steps of: S1, based on a sensor layout basis of a direct-drive/doubly-fed fan, performing vision-voiceprint multi-sensor collaborative layout on monitoring points of a fan transmission chain connecting part, setting a unified timestamp trigger mechanism to realize multi-source data synchronous acquisition of a vision sensor and a voiceprint sensor, simplifying or supplementing the sensor points according to model differences of the direct-drive, doubly-fed/semi-direct-drive fan, and adopting a general layout frame without redesigning; S2, respectively carrying out feature extraction on the acquired visual and voiceprint data, carrying out normalization processing on the visual feature parameters and the force and sound coupling feature parameters by adopting a normalization algorithm, and uniformly mapping all the parameters to a [0,1] interval; S3, constructing a CNN-BiLSTM fusion model with an attention mechanism, inputting normalized visual characteristics and force-sound coupling characteristics into the fusion model for deep fusion, dividing the loosening state of a fan transmission chain connecting part into 3 quantization levels by combining wind power field actual measurement data and simulation experiment data, and establishing a one-to-one correspondence between fusion characteristic parameters and loosening levels; And S4, constructing a TCN time convolution network trend pre-judging model, training the TCN model by utilizing historical fusion characteristic data, realizing trend pre-judging of the looseness level of the connecting part, setting a light, medium and heavy three-level grading early warning mechanism, and linking the pre-judging result of the TCN model with an alarm platform of the intelligent inspection system of the fan and a wind power operation and maintenance management platform to realize grading early warning treatment of the looseness state.
2. The method for monitoring the state of a fan transmission chain based on multiple sensors according to claim 1 is characterized in that in the step S2, a two-dimensional coordinate system is established by taking the center of a bolt as an origin, visual characteristics are extracted, namely moment line offset, moment line offset angle, moment line integrity, bolt micro-displacement, moment line pixel gray level difference and bolt group offset consistency, wherein the moment line offset is horizontal offset and vertical offset of the tail end of the moment line relative to an initial position, the bolt micro-displacement is axial micro-displacement and radial micro-displacement of the head of the bolt relative to an installation reference plane, the moment line integrity is the breaking length/fuzzy area ratio of the moment line, and the bolt group offset consistency is the offset variance of each bolt moment line in the same group of bolts.
3. The method for monitoring the state of a fan transmission chain based on multiple sensors according to claim 1, wherein in the step S2, original voiceprint signals collected by the voiceprint sensors are analyzed, force-sound coupling characteristics including time domain and frequency domain are extracted, the time domain characteristics are root mean square value, peak factor, pulse factor and kurtosis coefficient, and the frequency domain characteristics are characteristic frequency peak value, harmonic distortion rate, frequency band energy duty ratio and formant offset.
4. The method for monitoring the state of a fan drive chain based on multiple sensors according to claim 1, wherein in S3, the framework of the CNN-BiLSTM fusion model includes a CNN branch, a BiLSTM branch, an attention mechanism layer, a full connection layer and an output layer; The CNN branches perform feature extraction and dimension reduction on the visual feature matrix to obtain visual space feature vectors; The BiLSTM branch performs feature extraction on the time sequence data of the force acoustic coupling feature to obtain a force acoustic time sequence feature vector; the attention mechanism layer dynamically distributes weights for the vision space feature vector and the force sound time sequence feature vector according to different loosening stages of the fan transmission chain connecting part, and fuses the feature vectors; the full connection layer carries out linear transformation and nonlinear activation treatment on the self-adaptive fusion feature vector F; And the output layer outputs a fusion characteristic comprehensive value between 0 and 1 as a basis for judging the loosening grade.
5. The method for monitoring the state of a fan transmission chain based on multiple sensors according to claim 4, wherein the specific steps of dynamically distributing weights to the visual space feature vector and the force-sound time sequence feature vector by the attention mechanism layer according to different loosening stages of the fan transmission chain connecting component are as follows: view space feature vector through full connection layer Force sound time sequence characteristic vector Respectively performing feature significance learning to obtain initial attention weight of visual space features Initial attention weight of force sound time sequence characteristic ; According to the characteristic significance difference of different loosening stages of the fan transmission chain connecting part, introducing characteristic coefficients of the loosening stages Dynamically correcting the initial attention weight to obtain a dynamic weight coefficient of the visual space characteristic Dynamic weighting coefficient of force-sound time sequence characteristic ; According to And is also provided with The modified dynamic weight coefficient is respectively multiplied with the corresponding feature vector weight, and the weighted two types of feature vectors are added element by element to obtain the self-adaptive fusion feature vector output by the attention mechanism layer And realizing the self-adaptive depth fusion of the visual-force-sound coupling characteristics.
6. The method of claim 4, wherein the full link layer is to adaptively blend feature vectors Input as first full connection layer By linear transformation formula Performing feature space mapping, wherein Is the first The weight matrix of the layer full-connection layer, Is the first The bias term of the layer full-connection layer is passed through the ReLU nonlinear activation function formula For the characteristic vector after linear transformation Processing to realize nonlinear mapping of features and invalid feature suppression, and processing feature vectors by the first layer Sequentially inputting the subsequent full connection layer, completing the iterative processing of multi-round linear transformation and non-linear activation of ReLU, and mining the deep association among features to obtain the first High-dimensional abstract fusion feature vector output by layer full-connection layer 。
7. The multi-sensor based fan drive chain state monitoring method of claim 4, wherein the output layer fuses feature vectors in a high-dimensional abstraction Substitution into linear transformation formula The feature transformation is completed, and then the function formula is activated through Sigmoid For a pair of Normalized mapping is carried out, and fusion characteristic comprehensive values in a 0-1 interval are output , wherein, The characteristic value after linear transformation of the output layer is obtained; the weight vector of the output layer is a self-adaptive updated parameter in the model training process; Is a bias term for the output layer.
8. The method for monitoring the state of a fan transmission chain based on multiple sensors according to claim 1, wherein the loosening state of the fan transmission chain connecting component is divided into 3 quantization levels, and the corresponding threshold value of each level is: The level 0 is free of looseness, Y is 0 and 0.3, the offset of a moment line is less than 0.5mm, the offset angle is less than 2 degrees, the micro-displacement of a bolt is less than 20 mu m, the moment line is free of fracture/blurring, the energy ratio of a 2000-5000Hz frequency band is less than 15%, and the peak factor is less than 5; 1-level slight looseness is that Y epsilon (0.3, 0.7], the offset of moment lines is 0.5-2.0mm, the offset angle is 2-10 degrees, the bolt is slightly displaced by 20-100 mu m, the moment lines are slightly blurred/locally broken, the energy of 2000-5000Hz frequency band accounts for 15% -50%, the peak factor is 5-12, and slight looseness abnormal sound is accompanied; the 2-level heavy looseness is that Y epsilon (0.7,1.0), the offset of moment lines is larger than 2.0mm, the offset angle is larger than 10 degrees, the micro displacement of bolts is larger than 100 mu m, the moment lines are broken/blurred in large area, the energy ratio of a 2000-5000Hz frequency band is larger than 50%, the peak factor is larger than 12, and abnormal sound is remarkable and accompanying metal collision sound.
9. The method for monitoring the state of the fan transmission chain based on the multiple sensors according to claim 1 is characterized in that in the step S4, a sliding time window method is adopted, fusion characteristic integrated values in the operation process of the fan transmission chain connecting component are collected in a time unit of 1 hour, a time sequence data set is built, the input of the data set is a fusion characteristic integrated value sequence of historical N time windows, the output of the data set is a fusion characteristic integrated value prediction sequence with the future preset duration, a TCN time convolution network trend prediction model is built, the causal convolution and expansion convolution characteristics of the TCN model are utilized, the time sequence data set is trained, corresponding submodels are trained for a direct drive fan and a double feed fan respectively in combination with fan operation and maintenance materials, and trend prediction of the looseness level of the fan transmission chain connecting component is achieved through the trained TCN model.
10. The method for monitoring the state of the fan transmission chain based on the multiple sensors is characterized in that in the step S4, a light, medium and heavy three-level grading early warning mechanism is arranged, a pre-judging result of a TCN model is linked with a fan intelligent inspection system warning platform and a wind power operation and maintenance management platform, wherein the light early warning is that a pre-judging pre-set time period connecting part enters 1 level from 0 level without looseness and slightly loosens, the warning platform is marked and the monitoring frequency of the part is improved, the medium early warning is that the pre-judging pre-set time period connecting part enters 2 level from 1 level with slight looseness and severely loosens, a warning platform popup window and a mobile terminal short message warning is triggered, the heavy early warning is that the pre-judging pre-set time period connecting part is in 2 level with severe looseness or the looseness degree is continuously aggravated, and the warning platform acousto-optic warning and mobile terminal telephone and short message double warning are triggered, and grading early warning and treatment of the looseness state are achieved.

Description

Fan transmission chain state monitoring method based on multiple sensors Technical Field The invention relates to the technical field of fan state monitoring, in particular to a fan transmission chain state monitoring method based on multiple sensors. Background The driving chain connecting part of the wind driven generator is used as a core stress part for the operation of the fan, and the loosening of bolts is a typical operation fault. If the loosening state is not timely and accurately monitored and treated, serious problems such as bolt falling, driving chain clamping stagnation and the like are caused, so that the fan is not planned to stop and even equipment is damaged, the operation safety and economic benefits of the wind power plant are directly influenced, and along with the development of the fan to the high-capacity and high-reliability direction, the requirements on the accuracy, instantaneity and intellectualization of driving chain state monitoring are increasingly improved. The prior fan transmission chain connecting part looseness monitoring related technology has various inherent defects that a monitoring means mainly comprises single visual detection or single voiceprint/vibration detection, a plurality of sensors are used for cooperatively acquiring and information complementation capability, monitoring dimensions are single and judging basis is single, visual monitoring can only realize qualitative judgment of a looseness state, a standardized quantitative feature extraction system is not formed, the detection stability and precision are easy to be interfered by on-site wind sand, greasy dirt, cabin vibration and the like, voiceprint vibration monitoring is easy to be covered by voiceprint signals of normal operation working conditions of a gearbox, a generator and the like, fault sensitive feature extraction accuracy is low, multi-feature fusion adopts a fixed weight distribution strategy, self-adaptive dynamic adjustment cannot be carried out according to feature significance differences of different development stages of looseness, feature fusion characterization capability is limited, a standardized linear transformation and nonlinear activation processing flow cannot be easily carried out on depth excavation fusion features, and stable and reliable quantitative judgment results cannot be output. Disclosure of Invention The invention aims to overcome the defects of the prior art, provides a fan transmission chain state monitoring method based on multiple sensors, which can realize multi-source data synchronous acquisition through a vision-voiceprint multi-sensor accurate collaborative layout and a unified timestamp trigger mechanism, reduce or supplement point positions according to direct drive/doubly-fed fan model differences, improve layout universality, realize multi-dimensional feature refined extraction by adopting an image preprocessing and voiceprint filtering technology, eliminate dimension differences through normalization processing to form standardized feature input, construct a CNN-BiLSTM fusion model with an attention mechanism, realize self-adaptive weight distribution and depth fusion of vision space features and force acoustic time sequence features, complete accurate quantization output of fusion feature comprehensive values by combining a full-connection layer multi-layer cascade structure and an output layer activation function, establish a 3-level quantization grade system, realize loosening grade trend pre-judgment by combining a TCN time convolution network, set a light/medium/heavy three-level grading mechanism and form service closed loop in linkage with a fan intelligent inspection and operation maintenance management platform, and realize accurate quantization judgment and pre-judgment of a loosening state of a fan transmission chain connecting part. The invention provides a fan transmission chain state monitoring method based on multiple sensors, which comprises the following specific steps: S1, based on a sensor layout basis of a direct-drive/doubly-fed fan, performing vision-voiceprint multi-sensor collaborative layout on monitoring points of a fan transmission chain connecting part, setting a unified timestamp trigger mechanism to realize multi-source data synchronous acquisition of a vision sensor and a voiceprint sensor, simplifying or supplementing the sensor points according to model differences of the direct-drive, doubly-fed/semi-direct-drive fan, and adopting a general layout frame without redesigning; S2, respectively carrying out feature extraction on the acquired visual and voiceprint data, carrying out normalization processing on the visual feature parameters and the force and sound coupling feature parameters by adopting a normalization algorithm, and uniformly mapping all the parameters to a [0,1] interval; S3, constructing a CNN-BiLSTM fusion model with an attention mechanism, inputting normalized visual characteristics and force-sound c