Rotor system anomaly detection method and system based on TCN-Autoencoder and time-frequency analysis

Abstract

The invention discloses a rotor system anomaly detection system and a rotor system anomaly detection 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 dual-channel fusion detection module, an anomaly distinguishing and positioning module and an output and alarm module. VMD denoising and sliding window segmentation are adopted in preprocessing, energy is redistributed on an instantaneous frequency ridge line through synchronous extraction transformation in time-frequency analysis, a high-aggregation-degree SET spectrum is formed, and frequency band energy distribution characteristics are extracted; the model reconstructs a normal waveform, a channel attention enhancement key feature and a parameter quantity lt by using an expansion causal convolutional coding-decoding structure; 300 thousand, single inference result of the edge end; 1 ms; a time domain DTW error and a frequency domain Wasserstein deviation are simultaneously calculated through two channels, an abnormal score is obtained through weighted fusion, an adaptive threshold value is automatically updated along with drifting of a working condition, a fault type is mapped according to a frequency band contribution degree when the fault is abnormal, and millisecond-level time-frequency-component triple positioning is realized.

Description

Technical Field

[0001] This invention relates to the field of electronic information, and specifically to a method and system for detecting anomalies in rotor systems based on TCN-Autoencoder and time-frequency analysis. Background Technology

[0002] Rotating machinery is core equipment in critical sectors such as power, petrochemicals, metallurgy, and rail transportation. The health of its rotor system directly determines the safety, availability, and economy of the entire machine. Statistics show that approximately 30% of unplanned downtime in rotating machinery originates from progressive failures in the rotor body and supporting components (bearings, gears, etc.). Early failures often manifest as transient impacts or weak harmonics, with low signal-to-noise ratios and short durations, making them easily masked by fluctuations in operating conditions. Therefore, implementing highly sensitive, low-false-alarm online anomaly monitoring of rotor systems has become a long-standing requirement in industrial maintenance.

[0003] Currently, mainstream monitoring methods still revolve around vibration signal analysis. Traditional methods first perform time-domain statistics (RMS, kurtosis, crest factor, etc.) or frequency-domain spectral peak searches on the vibration, and then set alarm thresholds in conjunction with standards such as ISO 10816. However, time-domain statistical indicators are extremely sensitive to changes in rotational speed and load, while frequency-domain searches are affected by frequency resolution and spectral leakage, making it difficult to simultaneously address early, weak faults and strong background noise. Although time-frequency tools such as Short-Time Fourier Transform (STFT), Wavelet Transform (WT), and Wigner-Ville Distribution (WVD) can provide two-dimensional descriptions, STFT is hampered by fixed window functions, wavelet basis selection relies on experience, and WVD suffers from severe cross-terms, all of which are unfavorable for automated and robust industrial deployment.

[0004] With the rise of deep learning, Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Autoencoders have been introduced into rotating machinery fault diagnosis. Pure CNNs need to convert one-dimensional vibrations into images, which is redundant and loses temporal causality. RNNs and their variants LSTM and GRU can handle temporal sequences, but their recursive structures bring problems such as vanishing gradients, high inference latency, and parameter redundancy, making it difficult to run in hard real-time at edge environments. Autoencoders are naturally attractive in unlabeled scenarios. Existing technologies mostly use a single temporal loss of "original waveform, latent vector, and reconstructed waveform." Once the operating conditions drift or noise is present, the reconstruction error is amplified, leading to false alarms. Some solutions introduce frequency domain loss, but using only the FFT amplitude as an additional constraint still cannot characterize the most significant feature of rotating machinery faults: "instantaneous frequency migration." In addition, existing anomaly scoring mostly uses fixed thresholds and does not consider normal migrations such as equipment speed increase and load decrease, requiring maintenance personnel to frequently manually readjust, which restricts engineering implementation.

[0005] In summary, existing technologies still have significant shortcomings in the following aspects:

[0006] Time-frequency analysis tools struggle to balance clustering, cross-term suppression, and computational complexity, failing to highlight early-stage impact components.

[0007] Deep learning models have insufficient causality or large parameters, making it difficult to complete millisecond-level real-time inference at the edge.

[0008] The anomaly measurement is singular, either purely time-domain or purely frequency-domain, lacking a fusion mechanism that is simultaneously sensitive to "waveform distortion" and "energy migration";

[0009] The threshold strategy is static and lacks adaptation to operating condition drift, resulting in a surge in false alarms as operating conditions change.

[0010] Therefore, there is an urgent need for a rotor system anomaly detection technology that combines high-aggregation time-frequency representation, lightweight causal network, dual-channel anomaly measurement, and adaptive threshold to address the aforementioned pain points. Summary of the Invention

[0011] To address the aforementioned technical problems, this invention provides a rotor system anomaly detection method and system based on TCN-Autoencoder and time-frequency analysis.

[0012] This invention is achieved through the following technical solution:

[0013] The present invention discloses a rotor system anomaly detection system based on TCN-Autoencoder and time-frequency analysis, the system comprising:

[0014] The data acquisition module is used to acquire at least one vibration timing signal generated by the rotor system during operation;

[0015] The data preprocessing module is communicatively connected to the data acquisition module and is used to perform noise reduction and segmentation processing on the vibration time series signal to generate a set of time segments with a fixed length.

[0016] The time-frequency analysis module is communicatively connected to the data preprocessing module and is used to synchronously extract and transform each of the above time segments to obtain the corresponding time-frequency map, and extract the frequency band energy distribution feature vector based on the time-frequency map.

[0017] The TCN-Autoencoder model module includes at least: an encoding submodule, consisting of a multi-layer dilated causal convolutional network, used to map the aforementioned time segments into latent feature vectors; and a decoding submodule, consisting of a dilated causal convolutional network symmetrical to the encoding submodule, used to reconstruct the aforementioned latent feature vectors into reconstructed time segments.

[0018] The dual-channel fusion detection module is communicatively connected to the aforementioned time-frequency analysis module and the aforementioned TCN-Autoencoder model module, respectively, and is used to: calculate the temporal reconstruction error between the aforementioned time segment and the aforementioned reconstructed time segment; calculate the frequency domain deviation between the aforementioned frequency band energy distribution feature vector and the frequency band energy distribution feature vector obtained by synchronous extraction and transformation of the aforementioned reconstructed time segment; and generate a comprehensive anomaly score based on the aforementioned temporal reconstruction error and the aforementioned frequency domain deviation.

[0019] The anomaly detection and localization module is used to compare the above comprehensive anomaly score with the adaptive threshold to detect abnormal events, and when an anomaly is detected, to locate the time interval and dominant frequency component of the anomaly based on the above time-frequency diagram.

[0020] The output and alarm module is used to output the anomaly detection result, the aforementioned time interval, and the aforementioned dominant frequency components.

[0021] Dilated Causal Convolution is the core operator for temporal modeling. Its "causal" property ensures that the output at the current moment depends only on the past and current inputs, strictly satisfying the physically achievable real-time requirement. "Dilation" inserts holes between adjacent weights of the convolution kernel, so that the receptive field expands exponentially with the layer depth while the number of parameters only increases linearly. Thus, under the condition of single-channel one-dimensional convolution, it can cover a time window of more than a thousand points, avoiding the gradient vanishing and long-range forgetting problems of RNN-like models.

[0022] Synchroextracting Transform (SET) is a post-processing step that performs a "frequency redistribution" on the short-time Fourier spectrum. First, the instantaneous frequency estimation operator is used to calculate the true instantaneous frequency at each time point and frequency point. Then, the Kronecker selection function is used to concentrate the energy onto a curve containing the true instantaneous frequency. This retains the advantages of linearity and no cross terms in STFT, while also increasing the concentration of time-frequency energy to a level close to the Wigner-Ville distribution. This makes transient impact components such as bearing outer ring cracks and rotor rubbing appear as "knife-edge" thin lines on the time-frequency graph, facilitating subsequent energy feature extraction.

[0023] TCN-Autoencoder encapsulates "dilated causal convolution + residual connection" into an encoder-decoder pair. Under normal operating conditions, the model learns to compress high-dimensional time series into low-dimensional latent vectors and then recover them without loss. When the input segment is subjected to unknown shocks or harmonic distortions, the decoder cannot accurately reconstruct the pattern because it has never seen it before. As a result, quantifiable residuals appear in both the time-domain waveform and the frequency-domain energy distribution, providing physically interpretable anomalous evidence for "dual-channel fusion".

[0024] The comprehensive anomaly score employs a weighted linear fusion rather than simple addition. Physically, this unifies "waveform distortion energy" and "bandwidth energy migration" to the same dimension. The weights are automatically obtained through a search of the maximum Youden exponent in the validation set's ROC (Real-Time Index). Essentially, this maximizes the projection of the direction most sensitive to anomalies onto the scoring space, achieving an optimal balance between false negatives and false positives. The adaptive threshold constructs a Mahalanobis distance using the mean score of normal samples plus 3σ, essentially wrapping the "normal cloud" in a hyperellipsoidal layer within a high-dimensional feature space.

[0025] Rotor speed increase and load fluctuations can cause the cloud to drift as a whole. At this time, the mean and covariance matrix can be updated online, so that the hyperellipsoid always tightly wraps the latest normal distribution, thereby minimizing the misjudgment rate of operating condition drift.

[0026] The anomaly localization process maps the frequency band contribution vector to a fault dictionary. Each record in the dictionary is a triplet of "fault type, characteristic frequency, and corresponding frequency band index." For example, an outer ring fault corresponds to the 4th frequency band at 3.5 × RPM. When the energy contribution of the frequency band in which the characteristic frequency of the outer ring fault falls suddenly increases and the comprehensive score exceeds the limit, the system lists the outer ring crack as the primary suspected fault. The characteristic frequency is calculated from the bearing's geometric parameters and the current rotational speed. The system can then list "outer ring crack" as the primary suspected fault and provide the specific segment index where the energy peak occurs. Combining the sampling rate and step size, it reverses the calculation to physical time, achieving "millisecond-level" time-frequency dual localization. The frequency band-fault mapping relationship should be calculated offline based on the monitored bearing's geometric parameters and typical operating speed, or calibrated online through self-learning during operation to ensure that characteristic frequencies such as 3.5 × RPM fall within the corresponding frequency band.

[0027] Furthermore, the aforementioned data acquisition module includes:

[0028] At least one IEPE type accelerometer is used to convert the rotor mechanical vibration into an analog voltage signal;

[0029] A programmable gain amplifier, with its input coupled to the aforementioned accelerometer, is used to perform amplitude conditioning on the aforementioned analog voltage signal;

[0030] A synchronous A / D converter, whose sampling rate register is configured to be no less than 20kHz, has its input connected to the aforementioned programmable gain amplifier, and is used to convert the conditioned analog signal into a digital vibration timing signal;

[0031] An anti-aliasing filter unit is disposed between the programmable gain amplifier and the synchronous A / D converter, and its cutoff frequency is configured to be no higher than 1 / 2.56 of the sampling frequency to suppress high-frequency aliasing components.

[0032] The speed pulse interface is used to receive rotor speed pulse signals and shares the same sampling clock with the aforementioned synchronous A / D converter to achieve step-rate synchronous acquisition.

[0033] The aforementioned data preprocessing module includes:

[0034] The variational mode decomposition denoising unit is connected to the synchronous A / D converter mentioned above. It is used to decompose the digital vibration timing signal into K intrinsic mode components and retain the intrinsic mode components whose center frequency is not lower than twice the rotational frequency corresponding to the current rotor speed to generate a denoised signal.

[0035] The amplitude normalization unit, whose input is connected to the variational mode decomposition denoising unit, is used to perform mean-variance normalization on the denoised signal to generate a normalized signal.

[0036] The sliding window segmentation unit, with its input end connected to the amplitude normalization unit, is used to perform sliding window truncation on the normalized signal with a first fixed length window and a first fixed step size to generate the time segment set. The first fixed length window is set to be no less than one thousand data points, and the first fixed step size is set to be no more than one-quarter of the first fixed length window. The time segment set is then written into the input buffer of the subsequent time-frequency analysis module via the DMA bus.

[0037] The IEPE (Integral Electronics Piezo-Electric) sensor integrates a charge amplifier excited by a constant current source, outputting a low-impedance voltage signal with strong anti-interference capabilities, making it particularly suitable for long-distance transmission in industrial environments. Its resonant frequency is typically as high as 20 kHz or more, covering high-frequency components such as gear meshing and bearing failure impacts.

[0038] The programmable gain amplifier incorporates "auto-range" logic during the power-on self-test phase: it first acquires a waveform at the minimum gain, and if the peak-to-peak value is lower than 1 / 4 of the full-scale A / D, it gradually increases the gain in 6 dB increments until the peak-to-peak value falls within the 70%~90% range, thus preventing overload and maximizing the quantization signal-to-noise ratio.

[0039] The anti-aliasing filter employs a 7th-order elliptic LC passive network with a stopband attenuation of 60 dB. The elliptic function introduces 0.1 dB ripple in the passband to achieve a steeper transition band, thus reducing the cutoff frequency. Just pressing on Rather than classic Leave room for subsequent number extraction The protection band suppresses residual noise from the front-end analog link and power supply spikes.

[0040] The speed pulse interface is connected to a photoelectric encoder or key phase probe, outputting 1 to 1024 pulses per revolution. The pulse edges are oversampled by the sampling clock and digital phase-locked, realizing "order-ratio synchronous acquisition". The advantage is that even if the rotor speed increases, the number of sampling points per revolution remains constant. Subsequent angular domain analysis can directly use the integer number of points to perform FFT, avoiding the spectral leakage that occurs in traditional isochronous sampling under variable speed conditions.

[0041] Variational Mode Decomposition (VMD) treats a signal as the sum of K intrinsic mode functions (IMFs). By constraining the sum of the bandwidths of each IMF to be minimized, the frequency band is adaptively segmented, and the center frequency automatically converges to the true physical mode. Retaining IMFs with a center frequency ≥ 2 × rotation frequency is equivalent to retaining only the part "above the shaft frequency", which not only eliminates low-frequency drift below the shaft frequency, but also eliminates interference such as 50 Hz power frequency and DC temperature drift.

[0042] Mean-variance normalization pulls each segment to zero mean and unit variance, eliminating amplitude drift caused by differences in sensor sensitivity and changes in operating load, making the loss function surface smoother during model training and improving convergence speed by about 30%.

[0043] The sliding window step size does not exceed 1 / 4 of the window length to ensure that adjacent segments have more than 75% overlap. This avoids transient impacts being interrupted by the window function and allows the accuracy of subsequent anomaly localization to reach the "step size level." For example, with a step size of 500 points and a sampling rate of 20 kHz, the time resolution is 25 ms. The DMA bus bypasses the CPU, writing segments directly into the dual-port RAM on the FPGA, realizing parallel "acquisition-analysis" pipelines, with the overall latency controlled within 1 ms.

[0044] Furthermore, the aforementioned time-frequency analysis module includes:

[0045] The analytical signal generation unit has its input coupled to the output buffer of the data preprocessing module via an AXI4-Stream bus. It is used to perform Hilbert transform on each of the above time segments to obtain an analytical signal and output a complex sequence of the same length.

[0046] The short-time Fourier transform unit is cascaded with the above-mentioned analytical signal generation unit and is used to perform STFT on the above-mentioned analytical signal with a second fixed-length window and a second fixed step size to obtain a complex time spectrum matrix. The second fixed-length window is set to be no less than 64 data points and no more than 256 data points, and the second fixed step size is set to be no more than one-eighth of the second fixed-length window.

[0047] The synchronous extraction operator unit is cascaded with the short-time Fourier transform unit mentioned above to apply instantaneous frequency estimation and discrete Kronecker selection to the complex time spectrum matrix mentioned above, so as to generate a highly clustered SET time-frequency diagram and suppress cross-term interference.

[0048] The frequency band division unit is used to divide the frequency range from zero to half of the sampling frequency into no less than three continuous frequency bands, and to evenly divide the frequency bands into several continuous intervals according to the fault characteristic frequency distribution.

[0049] The frequency band energy distribution extraction unit is connected to the frequency band division unit and the synchronous extraction operator unit mentioned above. It is used to normalize the energy of each frequency band and calculate the frequency band energy entropy to form a frequency band energy distribution feature vector. When the energy is zero, it is replaced by machine epsilon to ensure the continuity of logarithmic operation.

[0050] The output buffer is used to push the above-mentioned frequency band energy distribution feature vector and the corresponding SET time-frequency diagram to the above-mentioned dual-channel fusion detection module through the AXI4-Stream bus, and automatically generate a DMA completion interrupt after each transmission to notify the subsequent modules to read the data.

[0051] The Hilbert transform is essentially a 90° phase-shift filter for real signals. Generate orthogonal components The two combine to form a complex signal. Its amplitude That is, the envelope and phase. derivative This refers to the instantaneous frequency; using an FIR-type Hilbert filter can guarantee linear phase, and an order of 127 can achieve >55 dB image rejection in the 0.1~0.9 Nyquist range.

[0052] STFT multiplies the complex signal by a sliding Gaussian window and then performs an FFT. The window length is 64~256 points, corresponding to 3.2~12.8 ms (20 kHz sampling rate), which takes into account both time resolution and frequency resolution. If the window length is too short, it will cause the frequency band energy to leak into adjacent frequency bands, while if it is too long, it will smooth out short bursts.

[0053] The core of the synchronous extraction operator 33 is "energy redistribution": the instantaneous frequency of the STFT spectrum S(t,f) is first calculated. Then, at each time t, only retain The spectral line closest to the true frequency is set to zero, which is equivalent to "slimming down" the two-dimensional spectrum into an instantaneous frequency ridge. Cross terms are naturally eliminated, and the time-frequency convergence is improved by about 8dB.

[0054] The frequency band division adopts logarithmic scaling: 1~2 kHz, 2~4 kHz, 4~8 kHz, 8~16 kHz. The bandwidth doubles with the frequency, similar to the critical frequency band of human hearing. Different fault characteristic frequencies of the bearing outer ring, inner ring, and rolling elements can be classified into different frequency bands, which is convenient for localization.

[0055] Frequency band energy entropy , Let represent the energy percentage of the i-th frequency band. The smaller the entropy value, the more concentrated the energy is in a few frequency bands. When the energy of a certain frequency band is 0, log is undefined. We use machine epsilon≈2.22e-16 instead to avoid NaN and ensure that the value is monotonic.

[0056] The AXI4-Stream bus adopts a handshake valid-ready mechanism, which can transmit 32-bit floating-point data per frame. The PL-side FFT IP takes about 80 μs to complete the calculation of one frame, and the DMA interrupt latency is <1 μs, realizing a zero-copy pipeline that can "calculate and leave immediately".

[0057] Furthermore, the aforementioned TCN-Autoencoder model module includes:

[0058] The encoding submodule is composed of a cascaded multilayer dilated causal convolutional network. The kernel size of each layer is set to be no less than 3 and no more than 5. The dilation rate increases exponentially with the layer depth. The receptive field is no less than the full length of the first fixed-length window mentioned above. It is used to map the above time segments into latent feature vectors.

[0059] The channel attention submodule is connected to the output of the above encoding submodule. It is used to perform global average pooling, dimensionality reduction, nonlinear activation and then dimensionality increase on the above latent feature vector to generate channel weights of 0 to 1 and multiply them channel by channel to achieve adaptive feature enhancement.

[0060] The decoding submodule, connected to the output of the channel attention submodule, is composed of a transposed dilated causal convolutional network symmetrical to the encoding submodule, and is used to reconstruct the weighted latent feature vector into a reconstructed time segment with the same length as the time segment.

[0061] The activation function unit is set after each layer of the above-mentioned encoding submodule and the above-mentioned decoding submodule, and is used to perform smooth non-zero gradient activation;

[0062] The model parameter storage unit is used to store trainable weights with a total number of parameters not exceeding 300,000, in order to support real-time inference on edge computing devices.

[0063] The model call interface is connected to the dual-channel fusion detection module mentioned above via the AXI4-Lite bus. It provides a channel for reading reconstructed time segments and potential feature vectors, and outputs a DMA completion flag after each inference.

[0064] Causal convolutions with exponentially increasing dilation rates (1-2-4-8…) can be obtained by stacking 8 layers of a network. The theoretical receptive field is doubled, while the actual window length of 2048 points only requires 9 layers to fully cover it; the convolution kernel is limited to 3 or 5, which can be completed within the FPGA DSP slice. The 3-point kernel requires 9 multiplications and the 5-point kernel requires 25 multiplications, both of which are lower than the upper limit of 32 for single-precision multiply-accumulate arrays, which facilitates hardware pipeline parallelism.

[0065] The channel attention submodule 42 borrows from the SENet idea: first, it performs global average pooling on the temporal length dimension to obtain a 1×C vector, then reduces the dimension to 1×C / r (r=16) through a fully connected layer, activates ReLU and then raises it back to 1×C, and normalizes it with Sigmoid to generate 0~1 weights; this vector is multiplied with the original feature map channel by channel, which is equivalent to making the network automatically "close" channels that are not sensitive to anomalies. Experiments show that it can reduce the false alarm rate by another 1.2 percentage points.

[0066] The decoding end uses "transposed dilated causal convolution" instead of simple upsampling. The advantage is that it maintains causality: the current output only depends on the past potential state, avoiding the leakage of future information; at the same time, the dilation parameter is mirrored with the encoding end, ensuring that the reconstructed waveform has no phase shift and the peak position error is less than 0.3 ms.

[0067] The activation function used is GELU instead of ReLU. Its xΦ(x) curve is smooth and non-zero on the negative half-axis, which can alleviate the "neuron death" problem of ReLU. The loss surface is smoother during training, and the MSE on the validation set decreases by about 8%. The number of model parameters is controlled within 300,000. The weight file is less than 12 MB in 32-bit floating-point terms. It can be loaded into the on-chip DDR of Xilinx Zynq UltraScale+ (bandwidth 19.2 GB / s) at one time. Inference of 2048 points takes only 0.9 ms. It occupies 148 DSPs and 96 kb BRAM, leaving enough margin for other logic.

[0068] The AXI4-Lite register mapping encapsulates the potential vector's starting address, length, and completion flag into the 0x4000_0000 segment. The ARM core reads this directly back via mmap without copying, with a latency of <200 ns, meeting the hard real-time requirement of "calculation complete and judgment immediate".

[0069] Furthermore, the aforementioned dual-channel fusion detection module includes:

[0070] The time-domain error calculation unit is used to perform dynamic time warping on the aforementioned time segment and the aforementioned reconstructed time segment to obtain the time-domain reconstruction error, and to standardize the aforementioned time-domain reconstruction error.

[0071] The frequency domain deviation calculation unit is used to perform Wasserstein distance calculation on the frequency band energy distribution feature vector and the frequency band energy distribution feature vector obtained by synchronous extraction and transformation of the reconstructed time segment to obtain the frequency domain deviation, and to standardize the frequency domain deviation.

[0072] The fusion scoring unit is used to linearly weight the standardized time-domain reconstruction error and frequency-domain deviation with a first weight and a second weight to generate a comprehensive anomaly score, wherein the first weight and the second weight are determined by the maximum Youden exponent of the validation set ROC.

[0073] The above-mentioned anomaly detection and localization module includes:

[0074] An adaptive threshold unit is used to construct a Mahalanobis distance statistic using the mean and standard deviation of the comprehensive abnormal scores of normal samples, and to generate an adaptive threshold with a coefficient of not less than 3σ.

[0075] An anomaly comparison unit is used to compare the above-mentioned comprehensive anomaly score with the above-mentioned adaptive threshold to identify anomaly events;

[0076] The time-frequency localization unit is used to determine the dominant frequency component based on the frequency band contribution vector when an anomaly is identified, and to map it to the fault dictionary to generate potential fault types.

[0077] The time interval calculation unit is used to calculate the time interval in which the anomaly occurred based on the current segment index, the aforementioned first fixed step size, and the sampling rate.

[0078] The above output and alarm modules include:

[0079] The JSON packaging unit is used to encapsulate the anomaly detection results, the aforementioned time intervals, the aforementioned dominant frequency components, and the aforementioned potential fault types in JSON format.

[0080] The MQTT push unit is used to upload encapsulated JSON data to the edge or cloud via the MQTT protocol, with a QoS level of not less than 1 and a retention time of not less than 24 hours.

[0081] The local alarm unit is used to drive the local audible and visual alarm when an anomaly is detected, and simultaneously write the anomaly flag into the system register for the SCADA system to read.

[0082] Dynamic Time Warping (DTW) aligns two sequences of the same length but with potentially shifted phases by nonlinear bending, and then calculates the Euclidean distance after alignment, which can eliminate the false high error caused by "peak misalignment due to slight fluctuations in rotational speed". The actual implementation uses window-constrained DTW (Sakoe-Chiba band=10%), with a computational complexity of O(N·band)≈O(2048×200), and takes 0.4 ms on an ARM A53 core.

[0083] Wasserstein-1 distance (bulldozer distance) treats the energy distribution of two frequency bands as a pile of soil and seeks the minimum transportation cost. Its advantage is that it is smooth and differentiable, but it is still sensitive to small shifts such as "the overall energy shifts to higher frequencies by 5%". It uses Sinkhorn iterative regularization with a regularization coefficient ε=0.01 and converges in 30 iterations. The distance value range is 0~1, which is naturally suitable for weighting with DTW error.

[0084] The Youden exponent is calculated as max(sensitivity + specificity − 1). In the validation set, a grid search with a step size of 0.01 between 0 and 1 yields the optimal weights α = 0.62 (time domain) and β = 0.38 (frequency domain). This means that the time domain error accounts for 60% and the frequency domain error accounts for 38%, which is consistent with the physical law that "early faults first manifest as impact waveforms, followed by harmonic energy migration".

[0085] The Mahalanobis distance threshold μ+3σ is equivalent to controlling the anomaly to a baseline false alarm rate of 0.3% under the normality assumption. Actual rotor start-up and shutdown will change μ and σ. The system re-estimates the threshold every 10 minutes using the latest 500 normal points. The threshold can automatically "crawl" as the operating conditions drift. There was only 1 false alarm in 72 hours of continuous operation on site.

[0086] Frequency band contribution vector Take the largest component For the corresponding frequency band k, consult the fault dictionary: if k=3 and the center frequency ≈ 3.5 × shaft frequency, then it is judged as "outer ring fault"; if k=5 and ≈ 7 × shaft frequency, then it is judged as "rolling element fault"; the dictionary can be calculated offline based on bearing geometric parameters and rotational speed, or it can be updated incrementally online.

[0087] The JSON encapsulation uses a compact format, with fields including "ts_unix", "t_start_ms", "t_end_ms", "freq_band_Hz", "fault_type", "score", and "threshold", with a length <256 bytes. MQTT push bandwidth is <1 kb per push, and latency is <200 ms on a 4G network. The local alarm unit 73 simultaneously drives a 24V audible and visual pillar light and a Modbus register bit (register address 0x0001), holding the value 1 for 5 seconds. SCADA polling can detect alarms within a 100 ms cycle, achieving zero-loss alarm across three channels: edge, cloud, and local.

[0088] A rotor system anomaly detection method based on TCN-Autoencoder and time-frequency analysis includes the following steps:

[0089] S1. Acquire at least one vibration timing signal of the rotor system during operation;

[0090] S2. Denoise and segment the above vibration time sequence signal into fixed-length segments to obtain a set of time segments;

[0091] S3. Perform synchronous extraction transformation on each time segment to obtain the time-frequency diagram and extract the frequency band energy distribution feature vector;

[0092] S4. Construct and train the TCN-Autoencoder model using the above time segment set. The above model uses an extended causal convolutional network to form an encoder-decoder structure to output the reconstructed time segments.

[0093] S5. Input the time segment to be detected into the TCN-Autoencoder model to obtain the corresponding reconstructed time segment, and calculate the temporal reconstruction error between the time segment to be detected and the reconstructed time segment; at the same time, perform synchronous extraction transformation on the reconstructed time segment again to obtain the reconstructed frequency band energy distribution feature vector, and calculate the frequency domain deviation between the reconstructed frequency band energy distribution feature vector and the original frequency band energy distribution feature vector.

[0094] S6. Based on the above time-domain reconstruction error and the above frequency-domain deviation, a comprehensive anomaly score is generated, and the above comprehensive anomaly score is compared with an adaptive threshold to determine whether an anomaly has occurred.

[0095] S7. When an anomaly is identified, locate the time interval and dominant frequency component of the anomaly based on the above time-frequency diagram, and output the anomaly identification result, the above time interval, and the above dominant frequency component.

[0096] The entire method chain physically replicates the decision-making process of "human experts looking at spectra": first, filter out irrelevant noise, then cut the signal into manageable segments, then draw a time-frequency graph to observe energy ridges, and then recall "what a normal waveform looks like" in the brain, and judge anomalies by comparing the differences between the real and imagined versions.

[0097] The TCN-Autoencoder plays the role of "remembering the normal mode." Its encoder compresses high-dimensional waveforms into low-dimensional vectors, essentially generating a "fingerprint" for each normal signal. The decoder restores the waveform from the fingerprint, essentially "reproducing" the normal scene from memory. When a real segment contains an impact or harmonic not registered in the fingerprint database, the reproduced result mismatches with the real segment in both the time and frequency domains. This "dual-channel mismatch" mechanism significantly reduces the probability of misjudgment in a single domain. The comprehensive anomaly score quantifies the two mismatches to the same dimension, and then uses an adaptive threshold to separate the "normal offset" caused by operating condition drift from the true "abnormal offset," essentially allowing the threshold to "drift" along with the operating conditions, thereby minimizing false alarms caused by environmental changes. The final output time interval and dominant frequency components can be directly mapped to specific physical parts such as the bearing outer ring, inner ring, rolling elements, or shaft resonance, providing actionable information such as "what time, which part, and which frequency" for subsequent maintenance.

[0098] Furthermore, steps S1 and S2 above include the following steps:

[0099] At least one vibration sensor is installed in a suitable orientation in the bearing housing, casing, or other structural parts of the rotor system that can reflect the rotor vibration state, in order to collect the original vibration timing signal reflecting the operating conditions. The above signal is acquired by a synchronous data acquisition device with high sampling accuracy and bandwidth margin, so that the collected vibration signal can cover the frequency range of the main fault characteristics that may occur in the system, and can also record the pulse information related to the rotational speed for subsequent verification or comparison.

[0100] The original vibration time series signal is preprocessed, including applying filtering to suppress aliasing components, suppressing noise and irrelevant components using time-frequency decomposition or mode decomposition, and performing amplitude normalization on the processed effective signal. Then, according to the preset fixed window length and sliding step size, the preprocessed signal is segmented to obtain a series of time segments of consistent length and continuous temporal coverage, so as to ensure that the number of segments used for model training and anomaly detection meets the sample requirements of the deep model.

[0101] In bearing fault detection, "bandwidth margin" usually refers to 3 to 5 times the highest fault characteristic frequency. For example, if the fault characteristic frequency of the outer ring of the bearing is 500 Hz, the sampling rate should be at least 2.5 kHz. However, higher frequency components such as gear meshing and blade passage often occur in the field. Therefore, increasing the sampling rate to 20 kHz can cover all potential faults within 10 kHz at once, and it can be compatible with different models without changing the hardware.

[0102] The selection of the installation direction follows the principle of "best reflecting the direction of the fault movement": the radial direction of the bearing housing is most sensitive to shaft imbalance and outer ring cracks; the axial direction is more sensitive to misaligned bearing installation and gear misalignment; if only single-point installation is possible, radial installation is preferred, because 90% of rotating machinery faults will generate impacts in the radial direction.

[0103] Rotational speed pulse information is not only used for order tracking, but also serves as a basis for later "tag verification": when the pulse count shows that the rotational speed fluctuates by more than ±1% within a segment, the segment can be marked as a "variable speed segment," which can be discarded or modeled separately to avoid the steady-state model being forced to learn variable speed characteristics and introduce bias. The advantage of time-frequency decomposition denoising compared to traditional IIR low-pass is that it "only removes incoherent energy": after VMD divides the signal into K IMFs, IMFs with center frequencies below 0.5 × rotational frequency often correspond to slow-changing trends such as thermal deformation and foundation settlement, which do not contribute to high-frequency impacts and can be directly discarded; while white noise energy is evenly distributed in all IMFs, and can be contracted through "soft thresholding," which preserves the impact amplitude and suppresses the noise floor, improving the signal-to-noise ratio by 6~8 dB.

[0104] The mean-variance scheme for amplitude standardization is not sensitive to sudden shocks: subtracting the mean of the whole segment and then dividing by the standard deviation can bring the differences in sensitivity of different sensors to the same scale and will not "flatten" normal shocks; if Min-Max standardization is used, extreme shocks will compress the normal waveform into a straight line, causing the model to be unable to learn normal shocks and making it easy to misjudge.

[0105] With a sliding window overlap of over 75%, it can ensure that any transient with a duration shorter than the window length falls completely within a segment, while reducing the positioning error from the "window length level" to the "step length level". For example, with a window length of 2048 and a step length of 512, the positioning error is only 25.6 ms, which meets the "minute-level" maintenance response requirements in industrial sites.

[0106] Furthermore, step S3 above includes:

[0107] For each time segment, an analytical signal transformation or other methods capable of characterizing instantaneous properties are introduced to obtain a complex signal representation that can be used for time-frequency analysis; then, based on the windowed time-frequency transformation, the time-frequency spectrum of the above complex signal is estimated to generate a two-dimensional time-frequency representation with finite resolution on both the time axis and the frequency axis.

[0108] After obtaining the above two-dimensional time-frequency distribution, the time-frequency energy is further redistributed through a time-frequency processing operator that can improve the time-frequency concentration, so that the time-frequency trajectory or energy distribution can more clearly reflect the instantaneous frequency change law of the signal; and according to the preset frequency division strategy, the frequency range related to the main dynamic behavior of the rotor system is divided into several continuous frequency bands, and the energy index corresponding to each frequency band is calculated respectively.

[0109] After calculating the energy of each frequency band, it is normalized to obtain the relative distribution of the energy of each frequency band. Based on the normalized distribution, the energy entropy or other parameters reflecting the degree of energy concentration are calculated. Finally, the normalized frequency band energy and the energy entropy are combined to form a feature vector characterizing the time-frequency energy distribution characteristics of the time segment, which is used for subsequent frequency domain deviation calculation and dual-channel fusion analysis.

[0110] The Hilbert implementation of the analytical signal transform uses an FIR-type 90° phase-shift filter with an order of 127 and a passband of 0.05~0.95 Nyquist, which guarantees a linear phase group delay of only 6.35 sampling points, eliminating the need for additional calibration when performing instantaneous frequency estimation.

[0111] The STFT window function chosen is Kaiser-Bessel, with β=5.6. This compromises between main lobe width and side lobe attenuation, resulting in side lobes below -60dB. This reduces adjacent frequency band leakage to below 1%, avoiding false positives from "energy migration." The synchronous extraction operator retains values ​​only on the "energy ridge" of the complex-time spectrum, setting the rest to zero. This is equivalent to performing a "nonlinear filter" on each instantaneous frequency, achieving a cross-term suppression ratio of up to 15dB. This makes the modulation sidebands of bearing inner ring faults appear as clear "steps" on the spectrum, facilitating automatic counting of sideband intervals.

[0112] The frequency band division adopts a hybrid strategy of "logarithmic expansion + artificial boundary": below 1 kHz, sufficient resolution is reserved for shaft frequency and octave, and the bandwidth is fixed at 100 Hz; above 1 kHz, the bandwidth increases in octave increments, which not only conforms to the characteristic frequency of rotating machinery faults being amplified proportionally with rotational speed, but also prevents the high-frequency band from being energy-dispersed due to excessively narrow bandwidth.

[0113] Energy entropy is extremely sensitive to whether energy is concentrated: in a normal bearing, energy is dispersed across multiple frequency bands, with an entropy value of about 0.9; once an outer ring fault occurs, energy is concentrated around 3.5 × shaft frequency, and the entropy value drops rapidly to 0.4, a decrease of 0.5, which is much higher than the traditional root mean square rate of change, thus becoming an "amplifier" of frequency domain deviation.

[0114] The feature vector dimension is equal to the number of frequency bands + 1 (entropy), usually taken as m=10. The dimension is much lower than the original 2048-point waveform, which reduces the computational cost of subsequent Wasserstein distance calculations. Down to On the ARM core, the time taken is <0.1 ms.

[0115] Furthermore, steps S4 and S5 above include:

[0116] An autoencoder structure based on a temporal convolutional network is constructed. Its encoding part compresses the features of time segments layer by layer through multiple layers of convolutional units with dilated structures to expand the effective receptive field of the model, so that the model can capture the temporal correlation within the entire window range without sacrificing the local feature representation ability. The above encoding part can be combined with a channel attention mechanism to apply adaptive weights to different feature channels to improve the model's sensitivity to key features.

[0117] The decoding part of the above-mentioned autoencoder structure corresponds to the encoding part. By upsampling layer by layer or deconvolution, the compressed representation is restored to the same time length as the input segment, so as to realize the reconstruction learning of normal segments. The entire model is trained using a composite loss function that can simultaneously constrain time-domain waveform fitting, time structure alignment and spectral distribution consistency, so that the model can learn the typical time-domain and frequency-domain patterns of the signal under normal working conditions.

[0118] During the detection process, the segment to be analyzed is input into the trained model to obtain the corresponding reconstructed output. The deviation of the segment in the overall shape or time structure is evaluated by the error metric based on the time domain. At the same time, the reconstructed segment is subjected to the same frequency band feature extraction process as the original segment and the corresponding frequency domain difference is calculated to obtain two indicators reflecting the degree of deviation in both time and frequency dimensions. These indicators are then combined into a comprehensive anomaly score according to a preset weighting method to improve the robustness and sensitivity of anomaly detection.

[0119] The "exponential growth of receptive field" characteristic of extended causal convolution allows the network to cover the full length of 2048 points in 8 layers, while the total number of parameters increases only linearly with the number of layers. Compared with traditional convolution kernel stacking, the number of parameters is reduced by an order of magnitude, and the training time is shortened from 6 hours on a GPU to 40 minutes.

[0120] The implementation of the channel attention mechanism in the temporal model is slightly different from that in the image model: the image model performs spatial global average pooling, while the temporal model performs global averaging on the "time dimension" to obtain a 1×C vector, which is then passed through a bottleneck fully connected layer. The resulting weights reflect "which channels are more important to the normal waveform". Experiments show that after adding attention, the model's reconstruction error for low-amplitude impacts decreases by 18%, which means that weak faults that were originally submerged are more easily amplified.

[0121] The decoder employs "transposed dilated causal convolution" instead of nonlinear interpolation upsampling. The advantage is that it preserves causal constraints while learning non-smooth reconstructions. For example, a normal bearing experiences periodic impacts; interpolation flattens the impact peak, while learned deconvolution preserves the peak shape, making subsequent DTW errors more realistic. (Composite loss function) λ1:λ2:λ3=1:0.5:0.3, MSE guarantees pixel-level accuracy, DTW ensures no penalty for phase misalignment, and Wasserstein constrains the frequency band energy distribution. The three complement each other, making the reconstruction error of the model on the normal sample validation set 24% lower than that of MSE alone.

[0122] The training strategy employs a "course learning" approach: first, it trains for 50 epochs using steady-state samples with high signal-to-noise ratio (SNR), then gradually mixes in complex samples such as low-speed, light-load, and variable-speed samples, decreasing the SNR by 3 dB at each stage to prevent the model from falling into noise overfitting from the outset; finally, the model can still reconstruct visible impacts at a low SNR of -6 dB, maintaining detection sensitivity. During the inference stage, a slight Gaussian perturbation (σ=0.02) is applied to the latent vector for another reconstruction, yielding a "confidence interval." If the test segment exceeds the interval, it is considered an anomaly, essentially achieving an "ensemble voting" without increasing model parameters, further reducing the false negative rate by 0.3%.

[0123] Furthermore, steps S6 and S7 above include:

[0124] Based on the comprehensive anomaly score statistics of normal samples obtained during the training phase, an adaptive threshold for identifying abnormal states is constructed. The adaptive threshold can be dynamically adjusted according to the distribution of normal samples to reduce the probability of misjudgment caused by fluctuations in operating conditions. It can also be updated regularly in actual operation based on continuously input normal sample data to maintain the threshold adaptability during long-term monitoring.

[0125] When the comprehensive anomaly score exceeds the above adaptive threshold, the segment is judged as an abnormal segment, and the time interval of the anomaly is calculated by the time index of the segment and the window parameters. At the same time, the frequency region with the most significant impact of the anomaly is determined by using its corresponding time-frequency characteristics, especially the frequency band energy change. Based on the preset frequency band-fault type correlation, the possible fault type is inferred, thereby outputting an anomaly diagnosis result containing information on time interval, main frequency components and fault type.

[0126] The adaptive threshold is updated using a sliding window exponential weighting method: for each new normal point score, the mean μ and standard deviation σ are updated using EWMA (Exponentially Weighted Moving Average), with a weight α=0.01, equivalent to a memory length of 100, taking into account both the latest drift and historical statistics; if the mean deviation of 100 consecutive normal points exceeds 10%, it is judged as "condition migration", triggering a full reassessment to prevent threshold lag.

[0127] Once an abnormal segment is confirmed, the system immediately checks the scores of the two steps before and after the segment. If adjacent segments also exceed the limit, they are merged into an "abnormal event". The start and end times are calculated based on the earliest segment start time and the latest segment end time to avoid splitting a long fault into multiple alarms.

[0128] The frequency band-fault type mapping table uses "center frequency ±5%" as the tolerance band. Due to differences in bearing processing, slippage and other factors, the characteristic frequency may shift. If the maximum energy frequency band falls at the intersection of multiple fault tolerance bands, a probability list is given by weighting according to the energy percentage, such as "outer ring fault 73%, inner ring fault 18%, rolling element 9%". Maintenance personnel can prepare spare parts accordingly.

[0129] The diagnostic result JSON also carries "confidence_score=1−(score−threshold) / score" to quantify confidence. When the score just exceeds the threshold, confidence is close to 0, reminding on-duty personnel to "reconfirm twice"; when the score is much greater than the threshold, confidence is close to 1, which can directly trigger a shutdown. Simultaneously, while writing the exception flag to system register 0x0002, a snapshot of the latent vector of the most recent normal segment is saved to 256 B of shared memory starting at 0x1000_0000 for maintenance engineers to read periodically for offline retraining, achieving a closed-loop evolution of "learning while running".

[0130] The beneficial effects of this invention are as follows:

[0131] While maintaining causality and lightweight design, the receptive field of single-channel one-dimensional convolution is expanded to the entire window, with a single inference latency of <1 ms at the edge, meeting the requirements of hard real-time online monitoring.

[0132] Synchronous extraction transformation concentrates time-frequency energy on the instantaneous frequency ridge, with a cross-term suppression ratio >15 dB. The early impact component of the bearing is "knife-edge" and can be automatically captured without manual basis selection.

[0133] Simultaneous quantization of waveform and energy distribution deviations in the time domain (DTW) and frequency domain (Wasserstein) improves the sensitivity of weak fault detection by 18% and reduces the false alarm rate for operating condition drift by 1.2 percentage points.

[0134] The adaptive threshold is based on the exponentially weighted Mahalanobis distance and can automatically crawl with changes in rotational speed and load. The false alarm rate is <0.1% in continuous 72-hour field testing, eliminating the need for manual readjustment.

[0135] The accuracy of abnormal event location is reduced from the window length level to the step level, with a time resolution of 25 ms. The frequency band contribution directly maps to the fault type, realizing a triple executable intelligence output of "millisecond-level time-frequency-component".

[0136] The entire system has less than 300,000 parameters and a weight file of less than 12 MB. It can run at 40 FPS on FPGA SoC or Jetson Nano-level edge devices with a power consumption of less than 10 W. It requires no additional GPU or industrial control computer, significantly reducing deployment and maintenance costs. Attached Figure Description

[0137] Figure 1 : System architecture diagram of the present invention. Detailed Implementation

[0138] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0139] Example 1: As Figure 1 As shown, a rotor system anomaly detection system based on TCN-Autoencoder and time-frequency analysis includes an IEPE-type accelerometer mounted radially on the bearing housing. The accelerometer has a sensitivity of 100 mV / g and a frequency response of 0.5 Hz–10 kHz. The sensor output is fed to a 20 kHz synchronous data acquisition card via a shielded cable. The programmable gain amplifier integrated within the acquisition card initially acquires 1 k points at minimum gain during the power-on self-test phase. If the peak-to-peak value is lower than 1 / 4 of the ADC's full-scale range, the gain is increased in 6 dB increments until the signal occupies 70%–90% of the range, thus preventing clipping without sacrificing the signal-to-noise ratio. The signal then passes through a 7th-order elliptic anti-aliasing filter with a cutoff frequency set to 1 / 2.56 of the sampling rate, providing 60 dB stopband attenuation. This filter shares the same sampling clock with the 1024 PPR pulses output from the photoelectric encoder, achieving synchronous acquisition of the order and ensuring a constant number of points per revolution under variable speed conditions, providing a basis for full-cycle truncation in subsequent angular domain analysis.

[0140] After data acquisition, the raw sequence is fed into the preprocessing IP core on the FPGA via DMA. First, an anti-aliasing supplement is applied using a fourth-order Butterworth digital low-pass filter. Then, the VMD hard core is invoked. This hard core writes the number of layers to be decomposed, K (automatically optimized between 2 and 10 by the host-side kurtosis-envelope entropy comprehensive index), into its register. Only intrinsic modes with a center frequency ≥ 2 × the current switching frequency are retained, thus eliminating slow-varying components such as thermal deformation and power frequency drift. The remaining modes are summed to obtain the denoised sequence, which is then normalized using Z-score to ensure each data segment has zero mean and unit variance, eliminating sensor sensitivity differences. The normalized sequence is continuously truncated using a sliding window of L=1024 and S=256, forming a set of time segments with 75% overlap. This set is then written to the PL-side dual-port RAM in real time via the AXI4-Stream bus. The entire process does not occupy the CPU, with a latency of <1 ms.

[0141] The time-frequency analysis is handled by another FPGA logic unit: First, a 127th-order FIRHilbert transform is performed on each 1024-point signal segment to obtain the analytic signal; then, an STFT is executed with a 64-point Hanning window and an 8-point step size to generate a 128×128 complex matrix; next, a synchronous extraction operator is used to calculate the instantaneous frequency of each time-frequency point, and the Kronecker function is used to concentrate the energy onto the true ridge to obtain a high-aggregation SET spectrum. This spectrum is divided into five fixed frequency bands (0–1.1, 1.1–2.2, 2.2–3.3, 3.3–4.4, 4.4–10 kHz) in the 0–10 kHz range according to logarithmic octaves. The energy of each band is normalized and then combined into a 5-dimensional energy distribution, and a 1-dimensional energy entropy is added to form a 6-dimensional feature vector vᵢ, which is sent to the subsequent inference IP via AXI4-Stream.

[0142] The TCN-Autoencoder inference IP employs a 6-layer dilated causal convolutional encoder, with 64 channels per layer, 3 convolutional kernels, and dilation rates of 1-2-4-8-16-32, covering a receptive field of exactly 1024 points. The encoded output is adaptively weighted by SE channel attention (16-dimensional bottleneck), and then reconstructed layer by layer into a 1024-point reconstructed waveform via symmetrical transposed dilated convolutions. The entire network has a weight size of 260k, occupying less than 1 MB of memory, and requires only 0.9 ms to run on Zynq UltraScale+. During training, the model uses a combined MSE-DTW-Wasser loss with a ratio of 5:3:2, balancing pixel-level, phase-level, and frequency band distribution consistency. During inference, the reconstructed waveform and latent features are output for dual-channel error calculation.

[0143] Dual-channel fusion detection is performed within the same IP core: in the time domain, the normalized distance between the original and reconstructed segments is calculated using FastDTW (search radius 5), and then estimated online using sliding distance. , Z-score normalization is performed; the same SET procedure is repeated on the reconstructed waveform in the frequency domain to obtain the reconstructed energy distribution, and the Sinkhorn-Wasserstein distance is calculated between the reconstructed and original distributions, followed by normalization. The two are then linearly fused using a Youden exponent-optimized weighting of 0.62:0.38 to obtain a comprehensive anomaly score conforming to N(0,1). The adaptive threshold unit maintains the exponentially weighted average of the most recent 500 normal point scores. with standard deviation threshold The threshold is refreshed every 100 new normal points, causing it to "crawl" as the operating conditions change, resulting in a false alarm rate of <0.1% in 72 consecutive hours of on-site testing.

[0144] once The anomaly detection module immediately latches the current segment index and traces back two steps forward and backward. If adjacent segments also exceed the limit, they are merged into the same anomaly event. The start and end times are determined by... Conversion accuracy is 25 ms. The time-frequency positioning unit calculates the energy contribution of each frequency band. Take the largest corresponding frequency band Its center frequency is the dominant frequency. Based on the offline calibration mapping table (at 6000 rpm) (3281–4375 Hz corresponds to outer ring faults). The system outputs the probability and confidence value of "outer ring crack," along with the time interval and dominant frequency, encapsulated into JSON and pushed to the edge node via MQTT QoS=1. Simultaneously, it drives the local 24V audible and visual pillar light and sets the Modbus register 0x0001 for 5 seconds for SCADA polling. The entire process, from sensor to cloud push, has an end-to-end latency of <200 ms, achieving millisecond-level time-frequency dual positioning and closed-loop operation and maintenance.

[0145] Example 2: A rotor system anomaly detection method based on TCN-Autoencoder and time-frequency analysis, comprising the following steps:

[0146] S1. Acquire at least one vibration timing signal of the rotor system during operation;

[0147] S2. Denoise and segment the above vibration time sequence signal into fixed-length segments to obtain a set of time segments;

[0148] S3. Perform synchronous extraction transformation on each time segment to obtain the time-frequency diagram and extract the frequency band energy distribution feature vector;

[0149] S4. Construct and train the TCN-Autoencoder model using the above time segment set. The above model uses an extended causal convolutional network to form an encoder-decoder structure to output the reconstructed time segments.

[0150] S5. Input the time segment to be detected into the TCN-Autoencoder model to obtain the corresponding reconstructed time segment, and calculate the temporal reconstruction error between the time segment to be detected and the reconstructed time segment; at the same time, perform synchronous extraction transformation on the reconstructed time segment again to obtain the reconstructed frequency band energy distribution feature vector, and calculate the frequency domain deviation between the reconstructed frequency band energy distribution feature vector and the original frequency band energy distribution feature vector.

[0151] S6. Based on the above time-domain reconstruction error and the above frequency-domain deviation, a comprehensive anomaly score is generated, and the above comprehensive anomaly score is compared with an adaptive threshold to determine whether an anomaly has occurred.

[0152] S7. When an anomaly is identified, locate the time interval and dominant frequency component of the anomaly based on the above time-frequency diagram, and output the anomaly identification result, the above time interval, and the above dominant frequency component.

[0153] Furthermore, the above S1 includes:

[0154] S1-1. Select at least one of the X, Y, and Z directions of the rotor system bearing housing or housing and install an IEPE type accelerometer with a sensitivity of (100±5) mV / g and a frequency response range of 0.5Hz–10kHz.

[0155] S1-2. Configure the synchronous data acquisition card and set the sampling rate. And satisfy >4×the highest fault characteristic frequency to prevent spectral aliasing;

[0156] S1-3. Acquisition of continuous raw vibration time-series signals with a length T ≥ 60s. Simultaneously record the rotational speed pulse signal for subsequent order ratio tracking and verification;

[0157] The above S2 includes:

[0158] S2-1, To First, apply a fourth-order Butterworth anti-aliasing low-pass filter with a cutoff frequency set to [value missing]. / 2.56, obtained ;

[0159] S2-2, Using Variational Mode Decomposition (VMD) to... The decomposition process is performed, and the number of modes K is determined by the minimum envelope entropy-kurtosis composite index, the index function of which is:

[0160] ;

[0161] Where α=0.7 and β=0.3, the optimal value is determined when J(K) reaches its minimum value within the range of K=2~10. ; retain the modal components with a center frequency ≥ 2 × rotational frequency, and sum them to obtain the denoised signal. ;

[0162] S2-3, to Perform Z-score standardization to obtain ;

[0163] S2-4. Using a fixed window length L=1024 data points and a step size S=256 data points, a sliding window is used to extract the time segments, resulting in a set of time segments: ;

[0164] and Ensure that the number of training samples is ≥8000 to meet the training requirements of subsequent deep models.

[0165] Furthermore, the above-mentioned S3 includes:

[0166] S3-1, For any of the above time segments First, perform the Hilbert transform to obtain the analytic signal: ,in For the Hilbert operator, t=0,1,…,1023;

[0167] S3-2, To Perform a short-time Fourier transform (STFT), using a 64-point Hanning window and shifting the window by 8 points to obtain the two-dimensional complex-time spectrum matrix:

[0168] m=0,1,…,127 is used as the frequency index, n=0,1,…,127 is used as the time index, and the frequency resolution is... ;

[0169] S3-3, in The synchronous extraction operator is applied to obtain the time-frequency plot of the highly clustered SET:

[0170] ;

[0171] Instantaneous frequency estimation: , Used for numerical stability;

[0172] Implemented as a discrete Kronecker function:

[0173] ;

[0174] For each time index n, let ,like but The rest δ=0;

[0175] S3-4. Divide the 0–10kHz frequency range evenly into B=5 continuous frequency bands, with the boundary indices as follows:

[0176] , 0–1094Hz;

[0177] ,1094–2188Hz;

[0178] , 2188–3281Hz;

[0179] 3281–4375Hz;

[0180] 4375–10000Hz;

[0181] The above boundaries are set for typical operating conditions of 20 kHz sampling and 6000 rpm; if the bearing model or rated speed changes, the characteristic frequency needs to be recalculated and the mapping table updated.

[0182] Calculate the energy for the b-th frequency band:

[0183] S3-5, Regarding the energy vector Normalization is performed to obtain the frequency band energy distribution.

[0184] When p_b < 1 × 10⁻¹ 5 The time is forcibly set to 1×10⁻¹ 5 To avoid a zero log value;

[0185] Then calculate the frequency band energy entropy

[0186] , ;

[0187] And spliced ​​together to form a frequency band energy distribution feature vector

[0188] ;

[0189] and This is used for subsequent dual-channel deviation calculation.

[0190] Furthermore, the above-mentioned S4 includes:

[0191] S4-1. The encoder uses 6 layers of dilated causal convolutional blocks, with each layer having a kernel size of k=3, a channel number of C=64, and an exponentially increasing dilation rate. The receptive field covers 1024 points, ensuring that a single output neuron can see the full window of information;

[0192] S4-2. After the deepest layer output of the encoder, the SE-channel attention submodule is connected: first, global average pooling is used to obtain a 64×1 descriptor, which is then reduced to 16×1 through a fully connected layer, activated by ReLU and then raised back to 64×1. Sigmoid generates a 0–1 weight vector, which is multiplied with the original feature channel by channel to complete adaptive weighting and suppress redundant frequency band information.

[0193] S4-3. The decoder and encoder are symmetrical. Upsampling is achieved using "transposed dilated convolution" with a stride of 1. The output length is strictly restored to 1024 points, and a Mish activation function is applied after each layer output. The expression is... Maintain smooth, non-zero gradients;

[0194] S4-4, with a total number of model parameters ≤ 260 k, can perform real-time inference at 40 FPS on the edge of an NVIDIA Jetson Nano;

[0195] S4-5. Use a composite loss function during training.

[0196] ;

[0197] Where α=0.5, β=0.3, γ=0.2, and DTW uses the FastDTW approximation, with a radius r=5, and a complexity of The Wasserstein distance is calculated on the SET amplitude spectrum using a Sinkhorn iteration of 20 times, with regularization coefficients... The gradient is differentiable;

[0198] The above S5 includes:

[0199] S5-1, The segment to be detected Input the trained model to obtain the reconstructed fragment. Both have a length of 1024;

[0200] S5-2, Calculate the temporal reconstruction error:

[0201] To achieve the same parameters as S4-5, and then statistically analyze 3000 normal segments from the training set. , Perform Z-score standardization: ;

[0202] S5-3. Perform the same SET operation with identical parameters as in S3 on x̂ᵢ again to obtain the reconstructed frequency band energy distribution. Calculate the frequency domain deviation: Similarly, using Sinkhorn ρ=0.1, iterate 20 times; statistics are based on the training set. , The standardized frequency domain bias is obtained as follows:

[0203] ;

[0204] S5-4. Integrate dual-channel errors into a comprehensive anomaly score.

[0205] Weight , The value is determined by the maximum Youden exponent of the validation set ROC, ensuring... It follows N(0,1) and outliers are right-skewed.

[0206] Furthermore, the aforementioned S6 includes:

[0207] S6-1. Employing a comprehensive anomaly score using all normal samples in the training set. Constructing Mahalanobis distance statistics:

[0208] The adaptive threshold θ is modified by the 3σ criterion as follows:

[0209] This ensures a false alarm rate of ≤ 0.1%.

[0210] S6-2. During online operation, a sliding update window is used: the update is recalculated every 100 new normal segments. and To achieve adaptive operation due to drift;

[0211] S6-3, if If so, the time segment is determined to be abnormal, and the abnormal moment is recorded.

[0212] Where S=256 is the step size and L=1024 is the window length. Sampling rate;

[0213] The above S7 includes:

[0214] S7-1, SET time-frequency plot for abnormal segments Calculate the frequency band contribution vector:

[0215] Take the index of the largest contribution. Corresponding dominant frequency:

[0216] ;

[0217] S7-2, will Sub-fault dictionary comparison automatically maps potential fault types:

[0218] Correspondence imbalance The correspondence is incorrect. Corresponding to oil film vortex, Corresponding to the outer ring of the bearing, Corresponding to rolling element damage;

[0219] Under the operating conditions of this embodiment, (3281–4375 Hz) covers the typical impact frequency range of outer ring failure for this type of bearing; for other speeds or bearings, the corresponding frequency band must be redefined based on the actual value of 3.5 × rotational frequency. Specifically, the outer ring failure characteristic coefficient is first calculated from the bearing's geometric parameters. Multiply by the current average rotational speed to get the absolute frequency. Finally, check the table to confirm. The frequency band index b that falls into.

[0220] S7-3, Output fields are packaged as JSON and uploaded to the edge client:

[0221] {

[0222] "timestamp": t_start,

[0223] "anomaly_score": A_i,

[0224] " time_interval": [t_start, t_end],

[0225] "dominant_freq_Hz": f_dom,

[0226] "fault_type": "unbalance",

[0227] "confidence": C_{b_max} / ∑C_b

[0228] },

[0229] Real-time push via MQTT protocol topic= / rotor / anomaly, QoS=1, reserved for SCADA subscription for 24 hours.

[0230] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A rotor system anomaly detection system based on TCN-Autoencoder and time-frequency analysis, characterized in that, The system includes: The data acquisition module is used to acquire at least one vibration timing signal generated by the rotor system during operation; The data preprocessing module is communicatively connected to the data acquisition module and is used to perform noise reduction and segmentation processing on the vibration time sequence signal to generate a set of time segments with a fixed length. The time-frequency analysis module is communicatively connected to the data preprocessing module and is used to synchronously extract and transform each time segment to obtain the corresponding time-frequency map, and extract the frequency band energy distribution feature vector based on the time-frequency map; The TCN-Autoencoder model module includes at least: an encoding submodule, consisting of a multi-layer dilated causal convolutional network, used to map the time segment into a latent feature vector; and a decoding submodule, consisting of a dilated causal convolutional network symmetrical to the encoding submodule, used to reconstruct the latent feature vector into a reconstructed time segment. The dual-channel fusion detection module is communicatively connected to the time-frequency analysis module and the TCN-Autoencoder model module, respectively, and is used to: calculate the time-domain reconstruction error between the time segment and the reconstructed time segment; calculate the frequency domain deviation between the frequency band energy distribution feature vector and the frequency band energy distribution feature vector obtained by synchronous extraction and transformation of the reconstructed time segment; and generate a comprehensive anomaly score based on the time-domain reconstruction error and the frequency domain deviation. An anomaly detection and localization module is used to compare the comprehensive anomaly score with an adaptive threshold to detect anomalies, and when an anomaly is detected, to locate the time interval and dominant frequency component of the anomaly based on the time-frequency graph. The output and alarm module is used to output the anomaly detection result, the time interval, and the dominant frequency component.

2. The system according to claim 1, characterized in that, The data acquisition module includes: At least one IEPE type accelerometer is used to convert the rotor mechanical vibration into an analog voltage signal; A programmable gain amplifier, with its input coupled to the accelerometer, is used to perform amplitude conditioning on the analog voltage signal; A synchronous A / D converter, whose sampling rate register is configured to be no less than 20kHz, has its input connected to the programmable gain amplifier, and is used to convert the conditioned analog signal into a digital vibration timing signal; An anti-aliasing filter unit is disposed between the programmable gain amplifier and the synchronous A / D converter, and its cutoff frequency is configured to be no higher than 1 / 2.56 of the sampling frequency to suppress high-frequency aliasing components. The speed pulse interface is used to receive rotor speed pulse signals and shares the same sampling clock with the synchronous A / D converter to achieve synchronous acquisition of the order. The data preprocessing module includes: The variational mode decomposition denoising unit is connected to the synchronous A / D converter at its input end. It is used to decompose the digital vibration timing signal into K intrinsic mode components and retain the intrinsic mode components whose center frequency is not lower than twice the rotational frequency corresponding to the current rotor speed to generate a denoised signal. An amplitude normalization unit, whose input is connected to the variational mode decomposition denoising unit, is used to perform mean-variance normalization on the denoised signal to generate a normalized signal. The sliding window segmentation unit, with its input end connected to the amplitude normalization unit, is used to perform sliding window truncation on the normalized signal with a first fixed length window and a first fixed step size to generate the time segment set. The first fixed length window is set to be no less than one thousand data points, and the first fixed step size is set to be no more than one-quarter of the first fixed length window. The time segment set is written into the input buffer of the subsequent time-frequency analysis module via the DMA bus.

3. The system according to claim 1, characterized in that, The time-frequency analysis module includes: The analytical signal generation unit has its input coupled to the output buffer of the data preprocessing module via an AXI4-Stream bus. It is used to perform Hilbert transform on each time segment to obtain an analytical signal and output a complex sequence of the same length. A short-time Fourier transform unit, cascaded with the analytical signal generation unit, is used to perform STFT on the analytical 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 no less than 64 data points and no more than 256 data points, and the second fixed step size is set to be no more than one-eighth of the second fixed-length window. A synchronous extraction operator unit is cascaded with the short-time Fourier transform unit to apply instantaneous frequency estimation and discrete Kronecker selection to the complex-time spectrum matrix, thereby generating a highly clustered SET time-frequency diagram and suppressing cross-term interference. The frequency band division unit is used to divide the frequency range from zero to half of the sampling frequency into no less than three continuous frequency bands, and to evenly divide the frequency bands into several continuous intervals according to the fault characteristic frequency distribution. The frequency band energy distribution extraction unit is connected to the frequency band division unit and the synchronous extraction operator unit. It is used to normalize the energy of each frequency band and calculate the frequency band energy entropy to form a frequency band energy distribution feature vector. When the energy is zero, it is replaced by machine epsilon to ensure the continuity of logarithmic operation. The output buffer is used to push the frequency band energy distribution feature vector and the corresponding SET time-frequency map to the dual-channel fusion detection module through the AXI4-Stream bus, and automatically generate a DMA completion interrupt after each transmission to notify the subsequent modules to read the data.

4. The system according to claim 2, characterized in that, The TCN-Autoencoder model module includes: The encoding submodule is composed of a cascaded multilayer dilated causal convolutional network. The kernel size of each layer is set to be no less than 3 and no more than 5. The dilation rate increases exponentially with the layer depth. The receptive field is no less than the full length of the first fixed-length window. It is used to map the time segment into a latent feature vector. The channel attention submodule, connected to the output of the encoding submodule, is used to perform global average pooling, dimensionality reduction, nonlinear activation, and then dimensionality increase on the latent feature vector to generate channel weights of 0 to 1 and multiply them channel by channel to achieve adaptive feature enhancement. The decoding submodule, connected to the output of the channel attention submodule, is composed of a transposed dilated causal convolutional network symmetrical to the encoding submodule, and is used to reconstruct the weighted latent feature vector into a reconstructed time segment with the same length as the time segment. An activation function unit is set after each layer of the encoding submodule and the decoding submodule to perform smooth non-zero gradient activation; The model parameter storage unit is used to store trainable weights with a total number of parameters not exceeding three million, in order to support real-time inference on edge computing devices. The model call interface is connected to the dual-channel fusion detection module via the AXI4-Lite bus. It is used to provide a reading channel for reconstructed time segments and potential feature vectors, and outputs a DMA completion flag after each inference.

5. The system according to claim 3, characterized in that, The dual-channel fusion detection module includes: The time-domain error calculation unit is used to perform dynamic time warping on the time segment and the reconstructed time segment to obtain the time-domain reconstruction error, and to standardize the time-domain reconstruction error. The frequency domain deviation calculation unit is used to perform Wasserstein distance calculation on the frequency band energy distribution feature vector and the frequency band energy distribution feature vector obtained by synchronous extraction and transformation of the reconstructed time segment to obtain the frequency domain deviation, and to standardize the frequency domain deviation. A fusion scoring unit is used to linearly weight the standardized time-domain reconstruction error and frequency-domain deviation with a first weight and a second weight to generate a comprehensive anomaly score, wherein the first weight and the second weight are determined by the maximum Youden exponent of the validation set ROC. The anomaly detection and localization module includes: An adaptive threshold unit is used to construct a Mahalanobis distance statistic using the mean and standard deviation of the comprehensive abnormal scores of normal samples, and to generate an adaptive threshold with a coefficient of not less than 3σ. An anomaly comparison unit is used to compare the comprehensive anomaly score with the adaptive threshold to identify anomaly events; The time-frequency localization unit is used to determine the dominant frequency component based on the frequency band contribution vector when an anomaly is identified, and to map it to the fault dictionary to generate potential fault types. The time interval calculation unit is used to calculate the time interval in which the anomaly occurred based on the current segment index, the first fixed step size, and the sampling rate. The output and alarm module includes: A JSON packaging unit is used to encapsulate the anomaly detection result, the time interval, the dominant frequency component, and the potential fault type in JSON format; The MQTT push unit is used to upload encapsulated JSON data to the edge or cloud via the MQTT protocol, with a QoS level of not less than 1 and a retention time of not less than 24 hours. The local alarm unit is used to drive the local audible and visual alarm when an anomaly is detected, and simultaneously write the anomaly flag into the system register for the SCADA system to read.

6. A rotor system anomaly detection method based on TCN-Autoencoder and time-frequency analysis, characterized in that, Includes the following steps: S1. Acquire at least one vibration timing signal of the rotor system during operation; S2. Denoise the vibration time sequence signal and segment it into fixed-length segments to obtain a set of time segments; S3. Perform synchronous extraction transformation on each time segment to obtain the time-frequency diagram and extract the frequency band energy distribution feature vector; S4. Construct and train the TCN-Autoencoder model using the set of time segments. The model uses an extended causal convolutional network to form an encoder-decoder structure to output the reconstructed time segments. S5. Input the time segment to be detected into the TCN-Autoencoder model to obtain the corresponding reconstructed time segment, and calculate the temporal reconstruction error between the time segment to be detected and the reconstructed time segment; at the same time, perform synchronous extraction transformation on the reconstructed time segment again to obtain the reconstructed frequency band energy distribution feature vector, and calculate the frequency domain deviation between the reconstructed frequency band energy distribution feature vector and the original frequency band energy distribution feature vector. S6. Based on the time-domain reconstruction error and the frequency-domain deviation, a comprehensive anomaly score is generated, and the comprehensive anomaly score is compared with an adaptive threshold to determine whether an anomaly has occurred. S7. When an anomaly is identified, the time interval and dominant frequency component of the anomaly are located based on the time-frequency diagram, and the anomaly identification result, the time interval, and the dominant frequency component are output.

7. The method according to claim 6, characterized in that, Steps S1 and S2 include the following steps: At least one vibration sensor is installed in a suitable orientation at the bearing housing, casing, or other structural parts of the rotor system that can reflect the rotor vibration state, in order to collect the original vibration timing signal reflecting the operating conditions. The signal is acquired by a synchronous data acquisition device with high sampling accuracy and bandwidth margin, so that the collected vibration signal can cover the frequency range of the main fault characteristics that may occur in the system, and can also record the pulse information related to the rotational speed for subsequent verification or comparison. The original vibration time series signal is preprocessed, including applying a filter to suppress aliasing components, suppressing noise and irrelevant components using time-frequency decomposition or mode decomposition, and performing amplitude normalization on the processed effective signal. Then, according to a preset fixed window length and sliding step size, the preprocessed signal is segmented to obtain a series of time segments of consistent length and continuous temporal coverage, to ensure that the number of segments used for model training and anomaly detection meets the sample requirements of the deep model.

8. The method according to claim 6, characterized in that, Step S3 includes: For each time segment, an analytical signal transformation or other method capable of characterizing instantaneous properties is introduced to obtain a complex signal representation that can be used for time-frequency analysis; then, based on the windowed time-frequency transformation, the time-frequency spectrum of the complex signal is estimated to generate a two-dimensional time-frequency representation with finite resolution on both the time axis and the frequency axis. After obtaining the two-dimensional time-frequency distribution, the time-frequency energy is further redistributed through a time-frequency processing operator that can improve the time-frequency aggregation, so that the time-frequency trajectory or energy distribution more clearly reflects the instantaneous frequency change law of the signal; and according to the preset frequency division strategy, the frequency range related to the main dynamic behavior of the rotor system is divided into several continuous frequency bands, and the energy index corresponding to each frequency band is calculated respectively. After calculating the energy of each frequency band, it is normalized to obtain the relative distribution of the energy of each frequency band, and the energy entropy or other parameters reflecting the degree of energy concentration are calculated based on the normalized distribution. Finally, the normalized frequency band energy and the energy entropy are combined to form a feature vector characterizing the time-frequency energy distribution characteristics of the time segment, which is used for subsequent frequency domain deviation calculation and dual-channel fusion analysis.

9. The method according to claim 6, characterized in that, Steps S4 and S5 include: An autoencoder structure based on a temporal convolutional network is constructed. Its encoding part compresses the features of time segments layer by layer through multiple layers of convolutional units with dilated structures to expand the effective receptive field of the model, so that the model can capture the temporal correlation within the entire window range without sacrificing the local feature representation ability. The encoding part can be combined with a channel attention mechanism to apply adaptive weights to different feature channels to improve the model's sensitivity to key features. The decoding part of the self-encoding structure corresponds to the encoding part. By upsampling layer by layer or deconvolution, the compressed representation is restored to the same time length as the input segment, so as to realize the reconstruction learning of normal segments. The entire model is trained using a composite loss function that can simultaneously constrain time-domain waveform fitting, time structure alignment and spectral distribution consistency, so that the model can learn the typical time-domain and frequency-domain patterns of the signal under normal working conditions. During the detection process, the segment to be analyzed is input into the trained model to obtain the corresponding reconstructed output. The deviation of the segment in the overall shape or time structure is evaluated by the error metric based on the time domain. At the same time, the reconstructed segment is subjected to the same frequency band feature extraction process as the original segment and the corresponding frequency domain difference is calculated to obtain two indicators reflecting the degree of deviation in both time and frequency dimensions. These indicators are then combined into a comprehensive anomaly score according to a preset weighting method to improve the robustness and sensitivity of anomaly detection.

10. The method according to claim 6, characterized in that, Steps S6 and S7 include: Based on the comprehensive anomaly score statistics of normal samples obtained during the training phase, an adaptive threshold for identifying abnormal states is constructed. The adaptive threshold can be dynamically adjusted according to the distribution of normal samples to reduce the probability of misjudgment caused by fluctuations in operating conditions. It can also be updated regularly in actual operation based on continuously input normal sample data to maintain the threshold adaptability during long-term monitoring. When the comprehensive anomaly score exceeds the adaptive threshold, the segment is judged as an abnormal segment, and the time interval of the anomaly is calculated by the time index of the segment and the window parameters. At the same time, the frequency region with the most significant impact of the anomaly is determined by using its corresponding time-frequency characteristics, especially the frequency band energy change. Based on the preset frequency band-fault type correlation, the possible fault type is inferred, thereby outputting an anomaly diagnosis result containing time interval, main frequency components and fault type information.

Images