Search

CN-122020439-A - Rotor system abnormality detection method and system based on TCN-Autoencoder and time-frequency analysis

CN122020439ACN 122020439 ACN122020439 ACN 122020439ACN-122020439-A

Abstract

The invention discloses a rotor system abnormality detection system and method based on TCN-Autoencoder and time-frequency analysis. The system is formed by sequentially connecting a data acquisition module, a data preprocessing module, a time-frequency analysis module, a TCN-Autoencoder model module, a double-channel fusion detection module, an abnormality discrimination and positioning module and an output and alarm module, wherein VMD denoising and sliding window segmentation is adopted in preprocessing, synchronous extraction transformation is utilized in time-frequency analysis to redistribute energy at an instantaneous frequency ridge line, a high-concentration SET spectrum is formed, frequency band energy distribution characteristics are extracted, the model reconstructs a normal waveform by an expanding causal convolution coding-decoding structure, channel attention is enhanced, parameter quantity is less than 30 ten thousand, single-time reasoning at the edge end is less than 1 ms, double-channel simultaneous calculation time domain DTW error and frequency domain Wasserstein deviation are weighted and fused to obtain an abnormality score, an adaptive threshold value is automatically updated along with drift, fault types are mapped according to frequency band contribution degree in abnormal conditions, and millisecond time-frequency-component triple positioning is realized.

Inventors

  • Qi Qingning
  • GAO ZHENNING
  • MA QING
  • WANG LI
  • LIU FENG
  • WANG SHOUSHUN

Assignees

  • 国能中卫发电有限公司

Dates

Publication Date
20260512
Application Date
20251203

Claims (10)

  1. 1. A rotor system anomaly detection system based on TCN-Autoencoder and time-frequency analysis, the system comprising: The data acquisition module is used for acquiring at least one path of vibration time sequence signal generated in the running process of the rotor system; the data preprocessing module is in communication connection with the data acquisition module and is used for performing denoising and segmentation processing on the vibration time sequence signal so as to generate a time segment set with a fixed length; The time-frequency analysis module is in communication connection with the data preprocessing module and is used for synchronously extracting and transforming each time segment to obtain a corresponding time-frequency diagram and extracting a frequency band energy distribution characteristic vector based on the time-frequency diagram; The TCN-Autoencoder model module at least comprises a coding submodule, a decoding submodule and a decoding submodule, wherein the coding submodule is formed by a multi-layer expansion causal convolution network and is used for mapping the time segment into a potential feature vector; The double-channel fusion detection module is respectively in communication connection with the time-frequency analysis module and the TCN-Autoencoder model module and is used for calculating a time domain reconstruction error between the time segment and the reconstruction time segment, calculating a frequency domain deviation between the frequency band energy distribution characteristic vector and the frequency band energy distribution characteristic vector obtained by synchronously extracting and transforming the reconstruction time segment again, and generating a comprehensive anomaly score based on the time domain reconstruction error and the frequency domain deviation; The abnormality judging and positioning module is used for comparing the comprehensive abnormality score with an adaptive threshold value to judge an abnormality event and positioning an abnormality occurrence time interval and a dominant frequency component based on the time-frequency diagram when the abnormality is judged; and the output and alarm module is used for outputting an abnormality judgment result, the time interval and the dominant frequency component.
  2. 2. The system of claim 1, wherein the data acquisition module comprises: at least one path of IEPE type acceleration sensor is used for converting mechanical vibration of the rotor into an analog voltage signal; The input end of the programmable gain amplifier is coupled with the acceleration sensor and is used for carrying out amplitude conditioning on the analog voltage signal; a synchronous A/D converter, the sampling rate register of which is configured to be not lower than 20kHz, and the input end of which is connected with the programmable gain amplifier and is used for converting the conditioned analog signal into a digital vibration time sequence signal; an anti-aliasing filter unit disposed between the programmable gain amplifier and the synchronous a/D converter, the cutoff frequency of which is configured to be not higher than 1/2.56 of the sampling frequency for suppressing high-frequency aliasing components; The rotating speed pulse interface is used for receiving a rotor rotating speed pulse signal and sharing the same sampling clock with the synchronous A/D converter so as to realize order synchronous acquisition; The data preprocessing module comprises: The input end of the variable-mode decomposition denoising unit is connected with the synchronous A/D converter and is used for decomposing the digital vibration time sequence signal into K intrinsic mode components and reserving the intrinsic mode components with the center frequency not lower than twice the corresponding rotation frequency of the current rotor speed so as to generate a denoising signal; The input end of the amplitude normalization unit is connected with the variation modal decomposition denoising unit and is used for performing mean-variance normalization on the denoising signal so as to generate a normalization signal; And the input end of the sliding window segmentation unit is connected with the amplitude normalization unit and is used for carrying out sliding window interception on the normalization signal by using a first fixed length window and a first fixed step length to generate the time segment set, wherein the first fixed length window is set to be not less than one thousand data points, the first fixed step length is set to be not more than one fourth of the first fixed length window, and the time segment set is written into an input buffer area of a subsequent time frequency analysis module through a DMA bus.
  3. 3. The system of claim 1, wherein the time-frequency analysis module comprises: The input end of the analytic signal generating unit is coupled with the output buffer of the data preprocessing module through an AXI4-Stream bus, and is used for executing Hilbert transformation on each time segment to obtain analytic signals and outputting complex sequences with the same length; A short-time fourier transform unit, cascaded with the analytic signal generation unit, for STFT of the analytic signal with a second fixed length window and a second fixed step size, to obtain a complex time spectrum matrix, wherein the second fixed length window is set to be not less than 64 data points and not more than 256 data points, and the second fixed step size is set to be not more than one eighth of the second fixed length window; The synchronous extraction operator unit is cascaded with the short-time Fourier transform unit and is used for applying instantaneous frequency estimation and discrete Kronecker selection on the complex time spectrum matrix to generate a SET time-frequency diagram with high aggregation degree and inhibiting cross term interference; The frequency band dividing unit is used for dividing the frequency range from zero to half of the sampling frequency into at least three continuous frequency bands and uniformly dividing the frequency bands into a plurality of continuous intervals according to fault characteristic frequency distribution; the frequency band energy distribution extraction unit is connected with the frequency band dividing unit and the synchronous extraction operator unit and is used for normalizing the energy of each frequency band and calculating the frequency band energy entropy so as to form a frequency band energy distribution characteristic vector, and when the energy is zero, the frequency band energy distribution characteristic vector is replaced by a machine epsilon so as to ensure logarithmic operation continuity; And the output buffer is used for pushing the frequency band energy distribution characteristic vector and the corresponding SET time-frequency diagram to the dual-channel fusion detection module through an AXI4-Stream bus, and automatically generating a DMA completion interrupt after each transmission is completed so as to inform a subsequent module to read data.
  4. 4. The system of claim 2, wherein the TCN-Autoencoder model module comprises: The coding submodule is formed by cascading a plurality of layers of expansion causal convolution networks, the size of each layer of convolution kernel is set to be not less than 3 and not more than 5, the expansion rate increases gradually with the layer depth index, and the receptive field is not less than the whole length of the first fixed length window and is used for mapping the time slices into potential feature vectors; The channel attention sub-module is connected with the output end of the coding sub-module and is used for carrying out global average pooling, dimension reduction and nonlinear activation on the potential feature vectors and then dimension increasing so as to generate channel weights of 0 to 1 and multiplying the channel weights by channel to realize self-adaptive feature enhancement; The decoding submodule is connected with the output end of the channel attention submodule and is composed of a transposition expansion causal convolution network symmetrical to the coding submodule, and the decoding submodule is used for reconstructing the weighted potential characteristic vector into a reconstruction time segment with the same length as the time segment; An activation function unit, which is arranged behind each layer of the coding submodule and the decoding submodule and is used for executing smooth non-zero gradient activation; A model parameter storage unit for storing trainable weights having a total parameter number of no more than three million to support real-time reasoning on the edge computing device; and the model calling interface is connected with the dual-channel fusion detection module through an AXI4-Lite bus and is used for providing a reading channel for reconstructing the time slices and the potential feature vectors and outputting a DMA completion mark after each reasoning is completed.
  5. 5. The system of claim 3, wherein the dual channel fusion detection module comprises: a time domain error calculation unit, configured to perform dynamic time warping on the time segment and the reconstructed time segment, so as to obtain a time domain reconstruction error, and normalize the time domain reconstruction error; A frequency domain deviation calculation unit, configured to perform wasperstein distance calculation on the frequency band energy distribution feature vector and the frequency band energy distribution feature vector obtained by performing synchronous extraction and transformation on the reconstructed time segment again, so as to obtain a frequency domain deviation, and normalize the frequency domain deviation; The fusion scoring unit is used for linearly weighting the normalized time domain reconstruction error and the normalized frequency domain deviation by a first weight and a second weight so as to generate a comprehensive anomaly score, wherein the first weight and the second weight are determined by a verification set ROC maximum Youden index; the abnormality discrimination and positioning module includes: The self-adaptive threshold unit is used for constructing a Markov distance statistic by using the mean value and the standard deviation of the normal sample comprehensive abnormal score and generating a self-adaptive threshold by using a coefficient not lower than 3 sigma; an anomaly comparison unit for comparing the integrated anomaly score with the adaptive threshold to determine an anomaly event; The time-frequency positioning unit is used for determining dominant frequency components based on the frequency band contribution vector when the frequency band contribution vector is judged to be abnormal, and mapping the dominant frequency components to the fault dictionary to generate potential fault types; the time interval calculation unit is used for calculating the time interval when the abnormality occurs according to the current fragment index, the first fixed step length and the sampling rate; the output and alarm module comprises: The JSON packaging unit is used for packaging the abnormal judgment result, the time interval, the dominant frequency component and the potential fault type according to a JSON format; The MQTT pushing unit is used for uploading the packaged JSON data to an edge end or a cloud end through an MQTT protocol, the QoS grade is not lower than 1, and the retention time is not less than twenty-four hours; and the local alarm unit is used for driving the local audible and visual alarm when the abnormality is judged, and synchronously writing the abnormality mark into the system register for the SCADA system to read.
  6. 6. The rotor system abnormality detection method based on TCN-Autoencoder and time-frequency analysis is characterized by comprising the following steps: S1, collecting at least one vibration time sequence signal of a rotor system in the running process; S2, denoising and segmenting the vibration time sequence signal with a fixed length to obtain a time segment set; s3, synchronous extraction transformation is carried out on each time segment, a time-frequency diagram is obtained, and frequency band energy distribution feature vectors are extracted; S4, constructing and adopting the time segment set to train a TCN-Autoencoder model, wherein the model forms an encoder-decoder structure by an expansion causal convolution network and is used for outputting a reconstructed time segment; S5, inputting a time segment to be detected into the TCN-Autoencoder model to obtain a corresponding reconstruction time segment, and calculating a time domain reconstruction error between the time segment to be detected and the reconstruction time segment; S6, generating a comprehensive anomaly score based on the time domain reconstruction error and the frequency domain deviation, and comparing the comprehensive anomaly score with an adaptive threshold value to judge whether anomaly occurs; And S7, when the abnormality is judged, locating a time interval and a dominant frequency component of the occurrence of the abnormality according to the time-frequency diagram, and outputting an abnormality judgment result, the time interval and the dominant frequency component.
  7. 7. The method according to claim 6, wherein the steps S1 and S2 include the steps of: The vibration sensor is arranged in at least one proper installation direction selected from a bearing seat, a shell or other structural parts which can reflect the vibration state of the rotor, so as to acquire an original vibration time sequence signal which reflects the operation working condition, the signal is acquired through a synchronous data acquisition device with higher sampling precision and bandwidth allowance, the acquired vibration signal can cover the possible main fault characteristic frequency range of the system, and pulse information related to the rotating speed can be recorded for subsequent verification or comparison. The method comprises the steps of preprocessing an original vibration time sequence signal, including applying filtering operation for suppressing aliasing components to the signal, suppressing noise and irrelevant components in a mode based on time-frequency decomposition or modal decomposition, and performing amplitude normalization to the processed effective signal, and then segmenting the preprocessed signal according to a preset fixed window length and a sliding step length to obtain a series of time segment sets with consistent length and continuous coverage in time so as to ensure that the number of segments for model training and anomaly detection meets the sample requirement of a depth model.
  8. 8. The method according to claim 6, wherein the step S3 includes: the method comprises the steps of introducing analytic signal transformation or other methods capable of representing transient characteristics to each time segment to obtain a complex signal representation which can be used for time-frequency analysis, and then carrying out time-frequency spectrum estimation on the complex signal based on windowed time-frequency transformation to generate a two-dimensional time-frequency representation with limited resolution on a time axis and a frequency axis; After the two-dimensional time-frequency distribution is obtained, the time-frequency energy is further redistributed through a time-frequency processing operator capable of improving the time-frequency aggregation degree, so that the time-frequency track or the energy distribution reflects the instantaneous frequency change rule of the signal more clearly; After the energy of each frequency band is calculated, normalization processing is carried out on the energy of each frequency band to obtain the relative distribution of the energy of each frequency band, the energy entropy or other parameters reflecting the energy concentration degree are calculated based on the normalization distribution, and finally, the normalized frequency band energy and the energy entropy are combined to form a characteristic vector representing the time-frequency energy distribution characteristic of the time segment for later frequency domain deviation calculation and double-channel fusion analysis.
  9. 9. The method according to claim 6, wherein the steps S4 and S5 include: Constructing a self-coding structure based on a time convolution network, wherein a coding part realizes layer-by-layer characteristic compression of time slices through a plurality of convolution units with expansion structures so as to enlarge an effective receptive field of a model, thereby enabling the model to capture time sequence association within the whole window range on the premise of not sacrificing the expression capacity of local characteristics; the whole model is trained by adopting a composite loss function which can simultaneously restrict time domain waveform fitting, time structure alignment and spectrum distribution consistency, so that the model can learn typical time domain and frequency domain modes of signals under normal working conditions; In the detection process, inputting the segment to be analyzed into a trained model to obtain corresponding reconstruction output, evaluating the deviation of the segment on the whole shape or time structure based on the error measurement of the time domain, simultaneously executing the same frequency band characteristic extraction flow as the original segment on the reconstruction segment and calculating the corresponding frequency domain difference to obtain two indexes reflecting the time-frequency double-dimension deviation degree, and combining the two indexes into a comprehensive anomaly score according to a preset weighting mode to improve the robustness and the sensitivity of anomaly detection.
  10. 10. The method according to claim 6, wherein the steps S6 and S7 include: Based on the comprehensive abnormal score statistical result of the normal sample obtained in the training stage, constructing an adaptive threshold for identifying an abnormal state, wherein the adaptive threshold can be dynamically adjusted according to the distribution condition of the normal sample so as to reduce the misjudgment probability caused by the fluctuation of the operation working condition, and can be regularly updated according to the continuously input normal sample data in the actual operation so as to maintain the threshold adaptability in the long-term monitoring process; And simultaneously utilizing the corresponding time-frequency characteristics, especially the frequency band energy change condition, to determine the frequency area with the most obvious abnormal influence, and deducing possible fault types based on the preset frequency band-fault type association relation, thereby outputting an abnormality diagnosis result comprising the time interval, the main frequency components and fault type information.

Description

Rotor system abnormality detection method and system based on TCN-Autoencoder and time-frequency analysis Technical Field The invention relates to the field of electronic information, in particular to a rotor system abnormality detection method and system based on TCN-Autoencoder and time-frequency analysis. Background The rotating machinery is core equipment in key fields of electric power, petrifaction, metallurgy, rail transit and the like, and the health state of a rotor system directly determines the safety, usability and economy of the whole machine. It is counted that about 30% of the unplanned stops of the rotating machine originate from progressive faults of the rotor body and the supporting parts (bearings, gears, etc.), while early faults often appear as transient impacts or weak harmonics, have a low signal-to-noise ratio and a short duration, and are very prone to being submerged by operating condition fluctuations. Therefore, implementing high-sensitivity, low-false-alarm online anomaly monitoring for rotor systems has become a long-felt need in the field of industrial maintenance. The current mainstream monitoring means still uses vibration signal analysis as a core. The traditional method is to firstly perform time domain statistics (RMS, kurtosis, crest factor and the like) or frequency domain spectrum peak search on vibration, and then set an alarm threshold value by combining with ISO 10816 and other standards. However, the time domain statistical index is extremely sensitive to the change of the rotating speed and the load, the frequency domain search is influenced by the frequency resolution and the spectrum leakage, and it is difficult to simultaneously consider the early weak fault and the strong background noise. While Short Time Fourier Transform (STFT), wavelet Transform (WT) and Wigner-Ville distribution (WVD) time-frequency tools can provide two-dimensional descriptions, STFT suffers from fixed window functions, wavelet basis selection depends on experience, and WVD cross terms are severe, which is not beneficial to automatic and robust industrial deployment. With the advent of deep learning, convolutional Neural Networks (CNNs), recurrent Neural Networks (RNNs), and self-encoders (Autoencoder) were introduced into rotary machine fault diagnostics. The RNN and its variants LSTM and GRU can process time sequence, but the recursive structure brings problems of gradient disappearance, high reasoning delay, redundant parameters, etc., and is hard to run on the edge side in real time. The self-encoder has natural attraction under the non-label scene, the prior art mostly adopts single time domain loss of 'original waveform, potential vector and reconstruction waveform', and once working condition drifts or noise exists, reconstruction errors are amplified, so that false alarms are caused. There are also schemes to introduce frequency domain loss, but only FFT amplitude is used as an additional constraint, and the most obvious feature of the rotating machinery failure, namely "instantaneous frequency migration", cannot be characterized. In addition, the existing abnormal scoring mostly adopts a fixed threshold value, normal migration such as equipment speed up and load down is not considered, maintenance personnel need to readjust frequently and manually, and engineering landing is restricted. In summary, the prior art still has significantly shorter plates in several respects: The time-frequency analysis tool is difficult to balance between the aggregation degree, cross term inhibition and computational complexity, and cannot highlight early impact components; the deep learning model has insufficient causality or huge parameters, and millisecond real-time reasoning is difficult to complete at the edge end; the anomaly measure is single, or pure time domain or pure frequency domain, and a fusion mechanism sensitive to both waveform distortion and energy migration is lacked; The threshold strategy is static, and the lack of self-adaption to working condition drift causes false alarms to proliferate with the change of the operating conditions. Therefore, a rotor system anomaly detection technique with a combination of high-concentration time-frequency representation, lightweight causal network, dual-channel anomaly metrics, and adaptive thresholds is needed to address the pain issues. Disclosure of Invention In order to solve the technical problems, the invention provides a rotor system abnormality detection method and system based on TCN-Autoencoder and time-frequency analysis. The invention is realized by the following technical scheme: the invention relates to a rotor system abnormality detection system based on TCN-Autoencoder and time-frequency analysis, which comprises: The data acquisition module is used for acquiring at least one path of vibration time sequence signal generated in the running process of the rotor system; the data preprocessing module is in communication connectio