A steam turbine vibration anomaly online identification method

Abstract

The present application relates to the technical field of steam turbine vibration test and fault diagnosis, and discloses a steam turbine vibration anomaly online identification method, comprising: collecting vibration signals of each channel and key signals, calculating basic features of each channel in a sliding window; establishing a redundancy consistency graph, calculating a consistency residual vector and a graph consistency cost for each node, calculating a comprehensive unreliable score and a reliability weight; constructing a reference feature vector, calculating a pseudo-vibration residual for each channel, and performing a pseudo-vibration priority determination; calculating a pseudo-vibration score, recalculating the reference feature vector after removing the pseudo-vibration priority channel and calculating a mechanical anomaly score; performing an anomaly determination and outputting a determination result; the present application uses a joint evaluation mechanism of the redundancy consistency graph and the channel self-checking feature to distinguish the real mechanical anomaly vibration from the pseudo-vibration caused by the measurement link online, significantly reduces the false alarm caused by the pseudo-vibration, and maintains the sensitivity to the real fault.

Description

Technical Field

[0001] This invention relates to the field of turbine vibration testing and fault diagnosis technology, and more specifically, to an online method for identifying abnormal turbine vibration. Background Technology

[0002] Steam turbine units are the core rotating equipment in power and industrial drive systems. Their rotors, bearings, and foundation systems operate continuously under high speed, high temperature, and high pressure conditions, making them prone to abnormal vibration problems such as imbalance, misalignment, unstable oil film, rubbing, and loose components. In severe cases, this can lead to unplanned shutdowns and significant production losses. To achieve real-time monitoring and fault early warning of vibration conditions, sensors for shaft vibration, bearing vibration, and casing vibration are typically installed on-site at key locations in the bearing housing and shaft system. These are paired with online monitoring and alarm systems that collect vibration signals using devices such as eddy current displacement sensors and velocity / accelerometers to conduct trend analysis and anomaly detection.

[0003] A Chinese patent with authorization announcement number CN112557039B discloses a method for diagnosing abnormal vibration faults in steam turbines by coupling operating parameters with vibration. The method includes the following steps: S1, acquiring vibration characteristic information; S2, determining whether the dominant frequency of the abnormal vibration is the fundamental frequency; S3, determining the magnitude of the 0.5 harmonic or low-frequency vibration during abnormal vibration, and determining whether the ratio of the 0.5 harmonic or low-frequency vibration value to the general frequency vibration value is greater than or equal to a set value; S4, determining whether the dominant frequency of the vibration is the second harmonic; S5, determining the magnitude of the frequency vibration peak between the fundamental frequency and the second harmonic during abnormal vibration; S6, determining the magnitude of the vibration value at a certain rotational speed; S7, determining whether the shaft vibration is significantly greater than the bearing vibration.

[0004] However, in practical applications, due to complex working conditions and harsh environments, the measurement link is prone to spurious vibration phenomena. That is, the detected vibration anomalies do not originate from the actual mechanical condition of the rotor or bearing, but are caused by faults in the sensor body, mounting structure, or signal link. The manifestations of spurious vibrations are highly similar to those of real mechanical faults, and both may exhibit characteristics such as increased amplitude, enhanced peak values ​​in specific frequency bands, and phase instability. This makes it easy for existing technologies to misjudge and fail to effectively distinguish between spurious vibrations and real mechanical anomalies. Summary of the Invention

[0005] The purpose of this invention is to provide an online method for identifying abnormal vibrations in steam turbines in order to solve the above-mentioned problems.

[0006] This invention provides an online method for identifying abnormal vibration in a steam turbine, comprising:

[0007] Vibration signals and key phase signals from each channel are collected, the quality of the key phase signals is evaluated and processed, and the basic characteristics of each channel are calculated within a sliding window.

[0008] Establish a redundancy consistency graph, calculate the consistency residual vector based on the redundancy consistency graph and basic features, and calculate the consistency residual vector for each node. Figure 1 Consistency cost, based on the Figure 1 The consistency cost is calculated by combining the unreliability score and the reliability weight.

[0009] A reference feature vector is constructed by weighted robust fusion of the basic features of multiple channels. Based on the reference feature vector and the basic features, the pseudo-vibration residual is calculated for each channel, and pseudo-vibration priority is determined.

[0010] Based on the comprehensive unreliable score and the pseudo-vibration residual, the pseudo-vibration score is calculated. After removing the pseudo-vibration priority channel, the reference feature vector is recalculated and the mechanical anomaly score is calculated.

[0011] Anomaly determination is performed based on the pseudo-vibration score and mechanical anomaly score, and the determination result is output.

[0012] Furthermore, the evaluation and processing of the bond phase signal quality includes:

[0013] The acquired vibration signals from each channel and the key phase signals are time-aligned.

[0014] The evaluation of the bond phase signal quality includes calculating the coefficient of variation of the time interval between adjacent bond phase pulses. If the coefficient of variation exceeds a preset bond phase quality threshold, the bond phase signal is deemed to have failed; otherwise, it is deemed to have passed.

[0015] When the key phase signal quality assessment passes, the time-domain signal is converted into an order-domain signal; when the key phase signal quality assessment fails, the time-domain analysis mode is maintained, and the short-time Fourier transform is used to extract the frequency-related features.

[0016] The fundamental characteristics include overall amplitude, order amplitude, band energy ratio, kurtosis, and spectral kurtosis.

[0017] Furthermore, based on the above Figure 1 The reliability weights for consistency cost calculation include:

[0018] Each channel is treated as a node, and a redundancy consistency graph is established. The redundancy consistency graph contains a set of nodes and a set of edges. The set of edges includes multiple edges. An edge is established between two channels in the same bearing orthogonal direction. An edge is established between channels of the same type of sensor on adjacent bearings. If there are cross-type channels of shaft vibration and bearing vibration or shell vibration, an edge is established.

[0019] For each edge in the redundant consistency graph, a consistency residual vector is calculated, which includes order magnitude consistency residual, coherent consistency residual, and phase difference stability residual.

[0020] The consistency residual vector is standardized.

[0021] Calculate for each node based on the redundancy consistency graph. Figure 1 Cost of conversion;

[0022] Based on the above Figure 1 Consistency cost calculation integrates unreliability score;

[0023] The overall unreliability score is mapped to a reliability weight.

[0024] Furthermore, the order magnitude consistency residual is:

[0025] Determine whether the first octave amplitude of two connected channels is greater than a preset amplitude threshold;

[0026] When the first octave amplitude of both nodes is greater than the preset amplitude threshold, the first octave amplitude of the two nodes is added to a preset small positive constant to obtain a sum. The natural logarithm of the two sums is taken respectively, and the absolute value of the difference between the two logarithm values ​​is calculated as the order amplitude consistency residual.

[0027] When the first octave amplitude of any node is lower than the preset amplitude threshold, the absolute value of the difference between the first octave amplitudes of the two nodes is divided by the amplitude term. The amplitude term is the sum of the first octave amplitudes of the two nodes plus a preset small positive constant, which is used as the order amplitude consistency residual.

[0028] Further, the coherence consistency residual is obtained by: calculating the coherence coefficients of two connected nodes at multiple characteristic frequencies, including the first harmonic, second harmonic, and half harmonic; subtracting the coherence coefficient from 1 at each characteristic frequency to obtain the coherence residual component at each frequency; and performing a weighted average of the coherence residual components at each frequency to obtain the coherence consistency residual.

[0029] The phase difference stability residual is obtained by: acquiring the phase value sequences of two connected channels at one harmonic; calculating the phase difference sequence of the two phase value sequences; performing phase expansion processing on the phase difference sequence to eliminate the boundary jump of positive and negative pi; and calculating the variance of the expanded phase difference sequence as the phase difference stability residual.

[0030] Furthermore, calculate for each node Figure 1 The costs of consistency include:

[0031] Iterate through all edges connected to the node. For each edge, sum the absolute values ​​of the components of the consistency residual vector to obtain the residual norm of the edge. Multiply the residual norm by a preset edge weight to obtain the weighted residual norm. Accumulate the weighted residual norms of all connected edges to obtain the node's residual norm. Figure 1 Cost of life.

[0032] Furthermore, based on the reference feature vector and the basic features, the pseudo-vibration residual is calculated for each channel, and the pseudo-vibration priority determination includes:

[0033] Within the same bearing or rotor segment, the basic features of multiple channels are weighted and robustly fused to construct a reference feature vector, which is calculated using a robust aggregation method based on Huber loss.

[0034] For each channel, calculate the pseudo-vibration residual, subtract the corresponding elements of the basic feature of the channel from the reference feature vector to obtain the difference vector, and calculate the Euclidean norm of the difference vector.

[0035] When the pseudo-vibration residual of the channel exceeds the preset residual threshold and the reliability weight is lower than the preset weight threshold, the abnormality of the channel is determined to be pseudo-vibration priority.

[0036] Furthermore, the calculation of the pseudo-vibration score includes:

[0037] For each channel, the unreliable score and the pseudo-vibration residual are weighted and summed, and the weighted sum is nonlinearly transformed to obtain the pseudo-vibration tendency value of the channel.

[0038] For channels marked as pseudo-vibration priority channels, the weighted sum is amplified when calculating their pseudo-vibration tendency values;

[0039] Among all the pseudo-vibration tendency values ​​of all channels, the value of the preset quantile is taken as the pseudo-vibration score.

[0040] Furthermore, the calculation of mechanical anomaly scoring includes:

[0041] Remove the channels marked as pseudo-vibration priority channels and recalculate the reference eigenvectors using the remaining channels;

[0042] The trend increment of the reference feature vector is calculated, which is obtained by calculating the difference between the current reference feature vector and the historical baseline reference feature vector.

[0043] The reference feature vector, trend increment, and basic features of operating parameters are concatenated into a vector. The inner product of the vector and the preset weight vector is calculated. The inner product is then subjected to a nonlinear transformation to obtain the mechanical anomaly score. The basic features of operating parameters include the current values ​​and rates of change of operating parameters such as load, main steam pressure, main steam temperature, and back pressure.

[0044] Furthermore, the anomaly determination and output of the determination result based on the pseudo-vibration score and mechanical anomaly score include:

[0045] When the false vibration score is greater than or equal to the preset false vibration discrimination threshold and the mechanical anomaly score is less than the preset mechanical anomaly discrimination threshold, the output judgment result is either measurement link anomaly or false vibration.

[0046] When the mechanical anomaly score is greater than or equal to the preset mechanical anomaly discrimination threshold and the pseudo vibration score is less than the preset pseudo vibration discrimination threshold, the output judgment result is mechanical anomaly vibration.

[0047] When the false vibration score is greater than or equal to the preset false vibration discrimination threshold and the mechanical anomaly score is greater than or equal to the preset mechanical anomaly discrimination threshold, the output judgment result is suspected mechanical abnormal vibration accompanied by measurement anomaly.

[0048] The beneficial effects of this invention are as follows: This invention, through a joint evaluation mechanism of redundancy consistency diagrams and channel self-checking features, distinguishes between real mechanical abnormal vibrations and spurious vibrations caused by the measurement link online, significantly reducing false alarms caused by spurious vibrations while maintaining sensitivity to real faults. By adaptively selecting order domain or time domain feature extraction based on bond phase quality assessment, usability is enhanced under conditions of speed fluctuations and bond phase degradation. By constructing a redundancy consistency diagram with channels as nodes, utilizing multi-dimensional residuals such as order amplitude consistency, coherence consistency, and phase difference stability, and employing robust normalization and physical correlation weighting, an interpretable... Figure 1 This approach leverages the consistency cost to locate suspicious channels with isolated anomalies. By mapping the comprehensive unreliability score to reliability weights and constructing robust reference features based on Huber loss, a reliable baseline insensitive to outliers is obtained. Pseudo-vibration priority labeling is applied to channels with high residuals and low weights, eliminating their influence in mechanical anomaly scoring and increasing their contribution in pseudo-vibration scoring, achieving bidirectional suppression and enhancement. By employing a dual-gated discrimination logic of pseudo-vibration scoring and mechanical anomaly scoring, and outputting the most suspicious channel and residual evidence chain, online weight reduction, shielding, or maintenance suggestions are supported, improving the system's stability and engineering practicality under scenarios of strong interference, packet loss, and sensor degradation. Attached Figure Description

[0049] Figure 1 This is a module example diagram of an online identification method for abnormal turbine vibration according to the present invention;

[0050] Figure 2 This is an example diagram illustrating the calculation of the comprehensive unreliability score and reliability weight in an online identification method for abnormal turbine vibration according to the present invention.

[0051] Figure 3 This is an example diagram illustrating the pseudo-vibration priority determination in the online identification method for abnormal turbine vibration according to the present invention;

[0052] Figure 4 This is an example diagram illustrating the computational mechanical anomaly score of an online turbine vibration anomaly identification method according to the present invention. Detailed Implementation

[0053] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0054] A method for online identification of abnormal turbine vibration, such as Figure 1 As shown, it includes:

[0055] Step 100: Collect vibration signals and key phase signals from each channel, evaluate and process the quality of the key phase signals, and calculate the basic characteristics of each channel within a sliding window.

[0056] The acquired vibration signals and key phase signals from each channel are time-aligned, and lost segments are removed. A missing mask is recorded. The missing mask is a binary indicator function, taking a value of 1 when data is complete and a value of 0 when data is missing. A channel refers to a complete measurement link in the turbine vibration monitoring system, consisting of a single sensor and its signal acquisition and transmission path. Each channel corresponds to a physical measurement point location for acquiring vibration signals at that location. The vibration signal is a time-varying electrical signal acquired by the sensor in the channel, characterizing the vibration state of the turbine rotor or bearing housing. The amplitude of this signal reflects the vibration intensity, and the frequency components reflect the vibration characteristics. The key phase signal is a pulse signal generated by the key phase sensor mounted on the rotor shaft per revolution. This pulse signal is used to mark the rotor's angular reference position, providing speed and phase references for the vibration signal.

[0057] Before performing order domain transformation, the quality of the bond phase signal is first evaluated. This evaluation is crucial because the bond phase signal is the foundation of order domain analysis, and its quality directly determines the accuracy of subsequent feature extraction. When the bond phase signal is affected by electromagnetic interference, sensor loosening, or drastic speed fluctuations, forcing order domain transformation can lead to spectral aliasing and phase distortion, resulting in false anomalies in key features such as the first harmonic amplitude. By pre-evaluating the bond phase signal quality and selecting an appropriate analysis mode based on the evaluation results, we can avoid misjudging defects in the bond phase signal itself as mechanical faults, thereby improving the reliability of anomaly identification. The bond phase signal quality evaluation includes detecting the stability of the bond phase pulse interval and the consistency of the pulse amplitude. Specifically, the coefficient of variation of the time interval between adjacent bond phase pulses is calculated. If the coefficient of variation exceeds a preset bond phase quality threshold, the bond phase signal is considered unqualified; otherwise, it is considered qualified.

[0058] The coefficient of variation (COP) for adjacent bond phase pulse time intervals is calculated as follows: Within a preset sliding window, the arrival time of each bond phase pulse is detected from the bond phase signal. The difference between the arrival times of two adjacent bond phase pulses is used to obtain the adjacent bond phase pulse time interval sequence. Only valid time intervals that do not cross packet loss segments and whose pulse pairs are complete are retained. The mean of this valid time interval sequence is calculated, and the standard deviation is obtained by calculating the dispersion of each time interval relative to the mean. The ratio of this standard deviation to the mean is then used as the COP. When it needs to be expressed as a percentage, this ratio is multiplied by 100% to obtain the percentage value of the COP. This COP characterizes the fluctuation level of the bond phase pulse interval relative to the mean; a larger value indicates a more unstable bond phase pulse interval. When the number of valid time intervals is insufficient to support stable statistics, the quality of the bond phase signal within the preset sliding window can be judged as failing to pass, thus avoiding misjudgments caused by a small number of samples.

[0059] When the key phase signal quality assessment passes, the vibration signal is synchronously resampled with rotational speed as the independent variable, converting the time-domain signal into an order-domain signal. The order-domain signal uses rotational angle as the independent variable to reduce spectral drift caused by rotational speed fluctuations. When the key phase signal quality assessment fails, the time-domain analysis mode is maintained, and short-time Fourier transform is used to extract frequency-related features. The residual calculation method is adjusted accordingly in subsequent consistency analysis; in this case, the order amplitude consistency residual is calculated using the energy ratio of frequency bands near the rotational frequency in the time-domain spectrum. The preset key phase quality threshold has a default value of 5% and a range of 3% to 10%, determined based on the accuracy level of the rotational speed measurement system and the unit's rotational speed fluctuation characteristics. The frequency-related features refer to the frequency components in the vibration signal spectrum that have a multiple relationship with the rotor rotational frequency. These include the first harmonic component (vibration amplitude and phase at the rotor's fundamental frequency), the second harmonic component (vibration amplitude at twice the rotor's fundamental frequency), and the half-harmonic component (vibration amplitude at half the rotor's fundamental frequency). These features can reflect typical mechanical fault states such as rotor imbalance, misalignment, and rubbing.

[0060] Within a sliding window, the fundamental characteristics of each channel are calculated, including overall amplitude, order amplitude, band energy ratio, kurtosis, and spectral kurtosis. These five types of fundamental characteristics were chosen because they can characterize the vibration signal from different dimensions, providing multi-dimensional criteria for subsequent redundancy consistency analysis and pseudo-vibration identification. Overall amplitude reflects the overall intensity of the vibration; order amplitude reflects specific frequency components related to rotational speed; band energy ratio reflects the energy distribution characteristics in the frequency domain; kurtosis reflects the impact characteristics of the time-domain waveform; and spectral kurtosis reflects the impact characteristics in the frequency domain. The combination of these five types of characteristics can effectively distinguish between real mechanical faults and measurement link anomalies, because real mechanical faults typically exhibit consistent characteristic change patterns across multiple channels, while measurement link anomalies often manifest as isolated characteristic anomalies only in a single channel.

[0061] The overall amplitude is represented by the root mean square (RMS) value. The calculation method is as follows: first, the vibration signal values ​​at each sampling point within the preset sliding window are squared; then, the sum of all squared values ​​is divided by the total number of sampling points to obtain the mean; finally, the square root of this mean is taken. The RMS value is used instead of the peak value or average value because it comprehensively reflects the energy level of the signal, has good characterization ability for both periodic and random vibrations, and is less susceptible to interference from occasional spikes. The default length of the preset sliding window is 1024 sampling points, determined based on the sampling frequency and rotational speed range, preferably including at least two complete rotational speed cycles. Specifically, the overall amplitude of a specified channel is equal to the square root of the sum of the squares of the vibration signal values ​​at all sampling points within the preset sliding window, divided by the total number of sampling points.

[0062] The order amplitude refers to the amplitude of the frequency component corresponding to a specific frequency multiple, including the first harmonic amplitude, second harmonic amplitude, and half-harmonic amplitude. It is obtained by extracting the amplitude of the corresponding order component after performing a Fourier transform on the order domain signal. The reason for choosing the first, second, and half-harmonic orders is that they correspond to the three most common mechanical fault modes in steam turbines. An increase in the first harmonic amplitude usually indicates rotor imbalance, an increase in the second harmonic amplitude usually indicates rotor misalignment or abnormal ellipticity, and an increase in the half-harmonic amplitude usually indicates fluid excitation phenomena such as oil film eddy or oil film oscillation. The combination of these three order amplitudes can cover the frequency domain characteristics of most typical faults.

[0063] The frequency band energy ratio is calculated as follows: First, the Fourier transform of the vibration signal is calculated. Then, the sum of the squares of the amplitudes of each frequency component within a preset specified frequency band is used as the numerator, and the sum of the squares of the amplitudes of all frequency components is used as the denominator. Dividing the two yields the energy ratio of the frequency band. The preset specified frequency band includes a low-frequency band, a power frequency band, a second harmonic band, and a high-frequency band. Specifically, the low-frequency band is from 0.1 to 0.8 harmonics, the power frequency band is from 0.8 to 1.2 harmonics, the second harmonic band is from 1.8 to 2.2 harmonics, and the high-frequency band is from 2.5 to 10 harmonics. The boundaries of each frequency band are determined based on the unit's speed range and typical fault characteristic frequencies. This energy ratio represents the proportion of energy within the preset specified frequency band to the total energy of the entire frequency band. The reason for using the frequency band energy ratio instead of a single frequency amplitude is that the energy ratio can reflect the distribution pattern of vibration energy in the frequency domain, providing better differentiation for identifying special vibration modes such as broadband vibration and high-frequency resonance, while also reducing the impact of single-frequency amplitude fluctuations caused by insufficient frequency resolution.

[0064] The kurtosis is calculated as follows: First, the mean of the vibration signal within a preset sliding window is calculated. Then, the average of the fourth power of the difference between each sampling point and the mean is calculated as the numerator, and the square of the average of the square of the difference between each sampling point and the mean is calculated as the denominator. Dividing the two values ​​yields the kurtosis value. Kurtosis is used to characterize the sharpness of the signal distribution and is sensitive to impact signals. Kurtosis is used as a fundamental feature because it can effectively identify faults that cause impact vibrations, such as rubbing and bearing damage. Furthermore, kurtosis is highly sensitive to measurement link anomalies such as probe loosening and poor contact, as these anomalies often manifest as intermittent spike signals.

[0065] The spectral kurtosis is obtained by calculating the ratio of the fourth moment to the square of the second moment of each frequency component amplitude after performing a short-time Fourier transform on the signal. This ratio is used to identify intermittent noise and impulse signals. The reason for using spectral kurtosis is that it can locate the frequency band of the impulse signal in the frequency domain, providing richer fault characteristic information compared to time-domain kurtosis. It is particularly effective in identifying faults such as bearing faults and gear faults that generate periodic impulses in specific frequency bands. Simultaneously, spectral kurtosis also provides good indication of measurement link anomalies such as electromagnetic interference and probe resonance exhibited in specific frequency bands.

[0066] Step 200: Establish a redundancy consistency graph, calculate the consistency residual vector based on the redundancy consistency graph and basic features, and calculate the consistency residual vector for each node. Figure 1 Consistency cost, based on the Figure 1 Consistency cost calculation combines unreliability score and reliability weight, specifically as follows: Figure 2 As shown.

[0067] Each channel is treated as a node in a graph, and a redundant consistency graph is established. This graph includes a node set and an edge set. The node set contains multiple nodes, each corresponding to a channel, and the edge set contains multiple edges, each corresponding to a connection between two nodes. The reason for using this redundant consistency graph structure is that turbine vibration monitoring systems typically deploy redundant sensors at multiple measurement points. These sensors have physical coupling and measurement correlations. Real mechanical vibration will exhibit consistent characteristic changes across multiple physically coupled channels, while measurement link anomalies usually only affect a single channel or a few channels, leading to inconsistencies between that channel and its adjacent channels. By constructing the redundant consistency graph and calculating the consistency residuals between nodes, suspicious channels exhibiting inconsistencies with adjacent channels can be effectively identified, thus distinguishing measurement link anomalies from real mechanical anomalies. The rules for establishing edges include: establishing an edge between two channels in the same orthogonal direction to the bearing; establishing an edge between channels of the same type of sensor on adjacent bearings; and establishing an edge between cross-type channels of shaft vibration and bearing vibration or shell vibration if there are such channels. The rules for establishing these edges are based on the physical structural characteristics of the turbine rotor system. Two channels orthogonal to the bearing measure vibrations in different directions of the same bearing housing, which have a strong coupling relationship. Channels of adjacent bearings transmit vibration energy through the rotor shaft system, which have a moderate coupling relationship. Although shaft vibration and bearing vibration or shell vibration are different objects of measurement, they reflect the vibration state at the same location and have a certain correlation.

[0068] For each edge in the redundancy consistency graph, a consistency residual vector is calculated. This vector includes order amplitude consistency residual, coherence consistency residual, and phase difference stability residual. These three types of residuals are chosen because they characterize the degree of consistency between two channels from three different dimensions: amplitude, coherence, and phase, comprehensively reflecting the measurement consistency between channels. The order amplitude consistency residual reflects the amplitude difference between the two channels at their main frequency components; the coherence consistency residual reflects the degree of linear correlation between the signals of the two channels; and the phase difference stability residual reflects the time-varying characteristics of the phase relationship between the two channels. When a measurement link anomaly occurs in a channel, it often exhibits a significant deviation from adjacent channels in one or more of these three dimensions, while real mechanical vibrations typically maintain good consistency between physically coupled channels.

[0069] The method for calculating the order amplitude consistency residual is as follows: First, determine whether the first octave amplitude of both connected nodes is greater than a preset amplitude threshold. The preset amplitude threshold is determined based on the sensor noise floor, and its default value is one-thousandth of the sensor range. When the first octave amplitude of both nodes is greater than the preset amplitude threshold, add a preset small positive constant to the first octave amplitude of each node to obtain a sum. Then, take the natural logarithm of each sum. Finally, calculate the absolute value of the difference between the two logarithmic values ​​as the order amplitude consistency residual. The default value of the preset small positive constant is 0.1 times the preset amplitude threshold, used to prevent zero or negative values ​​from occurring in the logarithmic operation. When the first octave amplitude of any node is lower than the preset amplitude threshold, the relative difference calculation method is used instead. That is, divide the absolute value of the difference between the first octave amplitudes of the two nodes by the amplitude term. The amplitude term is the sum of the first octave amplitudes of the two nodes plus the preset small positive constant, which is used as the order amplitude consistency residual. The residual reflects the relative deviation of the amplitudes of the two nodes at a first octave. The amplitude threshold avoids the distortion problem of logarithmic calculations in the low-amplitude region. Logarithmic transformation is used because vibration amplitudes often span multiple orders of magnitude; it maps relative deviations at different amplitude levels to a similar scale range, giving the residual calculation reasonable sensitivity in both low and high amplitude regions. Relative difference is used in the low-amplitude region because when the amplitude is close to the noise floor, logarithmic transformation amplifies the noise's influence; in this case, relative difference provides a more robust measure of consistency.

[0070] The calculation method for the coherence consistency residual is as follows: First, calculate the coherence coefficients of two connected nodes at multiple characteristic frequencies, including the first harmonic, second harmonic, and half harmonic. Then, subtract 1 from the coherence coefficient at each characteristic frequency to obtain the coherence residual component for each characteristic frequency. Finally, perform a weighted average of the coherence residual components at each characteristic frequency to obtain the coherence consistency residual. In the weighted average, the default weight for the first harmonic is 0.5, the default weight for the second harmonic is 0.3, and the default weight for the half harmonic is 0.2, with a sum of weights of 1.0. The weight values ​​are determined based on the importance of each characteristic frequency component in fault diagnosis. The coherence coefficient is the ratio of the square of the cross-power spectral density modulus to the product of the two self-power spectral densities, reflecting the degree of linear correlation between two signals at a specific characteristic frequency, and its value ranges from 0 to 1. This multi-frequency coherence consistency residual design can identify sensor anomalies that occur in frequency bands other than the first harmonic. The coherence coefficient is used as a consistency measure because it reflects the degree of linear correlation between two signals at a specific frequency, unaffected by amplitude scale, and is effective in identifying measurement link anomalies such as sensor sensitivity drift and frequency response changes. The first, second, and half-harmonic frequencies are chosen because they cover the main fault characteristic frequency bands of the turbine, allowing for the verification of consistency between channels across different frequency ranges. A weighted average is used instead of a simple average because different characteristic frequencies have varying importance in fault diagnosis; the first harmonic, as the most dominant vibration component, is given the highest weight, while the second and half-harmonic frequencies, as auxiliary discriminative features, are given lower weights.

[0071] The method for calculating the phase difference stability residual is as follows: First, obtain the phase value sequences of two connected nodes at the first harmonic. Then, calculate the phase difference sequence of the two phase value sequences. Perform phase expansion processing on the phase difference sequence to eliminate the boundary jumps between positive and negative pi. The phase expansion processing identifies jump points by detecting whether the change in phase difference between adjacent time points exceeds pi and performs cumulative compensation. Finally, calculate the variance of the expanded phase difference sequence as the phase difference stability residual. This residual reflects the stability of the phase relationship between the two nodes, and the phase expansion processing avoids the problem of artificially high variance caused by boundary jumps.

[0072] The aforementioned boundary jump between positive and negative pi refers to a situation where the phase value is typically represented within a finite interval bounded by negative and positive pi. When the actual phase changes continuously over time and crosses this boundary, the representation result exhibits a range reversal, jumping from a side closer to positive pi to a side closer to negative pi, or vice versa. This jump originates from the periodic representation of phase and does not imply a physical abrupt change in the oscillating phase. However, it creates a large instantaneous change in the phase difference sequence, amplifying stability statistics such as variance and leading to misjudgments of phase instability.

[0073] The phase unrolling process aims to restore the phase jumps caused by the interval reversal to a continuously changing phase difference sequence. This is achieved by performing integer-cycle phase accumulation compensation on the subsequent phase difference sequence after detecting a boundary jump, ensuring the phase difference sequence maintains continuity on the time axis and aligns with the actual phase change trend. Phase difference stability is used as a consistency measure because the phase difference between two physically coupled channels should remain relatively stable under normal conditions, changing only slowly with operating conditions. However, abnormalities in the measurement link, such as poor contact or inadequate cable shielding, often lead to phase drift or phase jumps, manifesting as phase difference instability. Variance is used as a stability measure because it comprehensively reflects the fluctuation degree of the phase difference sequence and has good sensitivity to both intermittent phase jumps and persistent phase drift. Phase unrolling is performed because the phase value itself is periodic, jumping at the positive and negative pi boundaries. Without unrolling, these mathematical boundary jumps would be misinterpreted as physical phase instability, leading to distorted residual calculations.

[0074] The three residual components mentioned above are arranged in sequence to form a consistent residual vector. The consistent residual vector is then standardized using a robust scaling normalization method based on a historical sliding window. Specifically, the median absolute deviation of each residual component within a preset historical time window is maintained as a robust scaling estimate. The standardized residual component is obtained by subtracting the historical median from the current residual component and then dividing by 1.4826 times the median absolute deviation. The robust scaling normalization method is used instead of the traditional mean-variance standardization method because outliers and abnormal values ​​frequently appear in vibration monitoring data. Traditional mean and standard deviation are sensitive to outliers, leading to distorted standardization results. The median absolute deviation, as a robust statistic, is insensitive to outliers and can still provide a stable scaling estimate even in the presence of outliers. The historical sliding window is used because vibration characteristics drift slowly with operating conditions and seasonal changes. Using a historical window can adapt to this drift while maintaining sensitivity to sudden anomalies. The default value for the preset historical time window length is 72 hours, with a range of 24 to 168 hours, determined based on the unit's operating condition variation cycle and statistical stability requirements. The method for calculating the median absolute deviation is as follows: first, calculate the absolute value of the difference between each residual value and the median within the historical window; then, take the median of these absolute values. This standardization method is insensitive to outliers, ensuring that different residual components are within the range of... Figure 1 Comparability and threshold stability in consistency cost calculation. Real mechanical anomalies exhibit interpretable consistency across redundant channels, while pseudo-vibrations such as probe resonance, loose installation, and poor contact often manifest as isolated local anomalies. This difference between consistency and isolation is the core basis for distinguishing between real mechanical anomalies and measurement link anomalies; standardization ensures that this difference can be quantified and compared on a unified scale.

[0075] Calculate for each node based on the redundancy consistency graph. Figure 1 The consistency cost is calculated by iterating through all edges connected to the node. For each edge, the absolute values ​​of the components of the consistency residual vector are summed to obtain the residual norm of that edge. This norm is then multiplied by a preset edge weight to obtain the weighted residual norm. Finally, the weighted residual norms of all connected edges are summed to obtain the node's consistency cost. Figure 1 Consistency cost. The preset edge weights are given by the physical correlation of the channels. The default value of the edge weight for channels of the same type and bearing is 1.0, the default value of the edge weight for channels of the same type and adjacent bearings is 0.6, and the default value of the edge weight for channels of different types is 0.4. The weight values ​​range from 0.1 to 1.0 and are determined based on the physical coupling strength and historical correlation statistics between channels.

[0076] The comprehensive unreliability score is calculated by combining the channel self-test characteristics. The specific method is as follows: calculate... Figure 1The product of the normalized value of consistency cost and a preset first weighting coefficient, plus the product of the saturation or clipping indicator value and a preset second weighting coefficient, plus the product of the normalized value of spectral kurtosis and a preset third weighting coefficient, plus the product of the normalized value of the absolute value of the zero-point bias change and a preset fourth weighting coefficient, are summed to obtain the comprehensive unreliability score. The default values ​​for the preset first weighting coefficient are 0.4, the preset second weighting coefficient, the preset third weighting coefficient, and the preset fourth weighting coefficient are 0.1, with each weighting coefficient ranging from 0 to 1, and the sum of the four weighting coefficients is 1.0. These values ​​are determined through historical data statistical analysis based on the contribution of each feature to pseudo-vibration identification. The normalized value is obtained through a normalization function, which employs a quantile mapping method based on historical cumulative distribution. Specifically, it maintains the empirical cumulative distribution function of each feature within a preset historical time window, maps the current feature value to its quantile in the historical distribution, and uses this as the normalized value, with a value ranging from 0 to 1. The preset historical time window length is consistent with the historical window length in the standardization process, ensuring uniform statistical standards across channels and across time. When the number of monitoring channels is less than four, the sorting normalization uses the quantile mapping of the historical distribution across time to avoid the coarsening of sorting between channels at the same moment. The saturation or clipping indicator value is 1 when saturation or clipping occurs, and 0 otherwise. The zero-point bias change is obtained by regression estimation from the low-frequency band mean or steady-state segment.

[0077] The overall unreliability score is mapped to a reliability weight. Specifically, an exponential function is calculated using the natural constant as the base and the negative value of the overall unreliability score as the exponent, and the result is used as the reliability weight. The reliability weight is then truncated to a value between a preset minimum and 1. The preset minimum defaults to 0.1, and its range is from 0.05 to 0.3, determined based on the system's tolerance for low-reliability channels. A smaller value indicates a more severe penalty for suspicious channels. A smaller reliability weight indicates a channel is more likely to produce spurious vibrations.

[0078] Step 300: A weighted robust fusion of the basic features from multiple channels is performed to construct a reference feature vector. Based on the reference feature vector and the basic features, the pseudo-vibration residual is calculated for each channel, and a pseudo-vibration priority determination is performed, specifically as follows: Figure 3 As shown.

[0079] Within the same bearing or rotor segment, the fundamental features of multiple channels are weighted and robustly fused to construct a reference feature vector. This reference feature vector is calculated using a robust aggregation method based on Huber loss. This method takes the reliability weights and fundamental features of each channel in the channel set as input and outputs a robustly aggregated reference feature vector. The reason for using the Huber loss robust aggregation method is that traditional weighted averaging methods are sensitive to outliers. When a measurement link anomaly occurs in a channel, even if the channel's reliability weight is low, its abnormal fundamental feature value will still have a significant impact on the averaging result. The Huber loss function uses squared loss for small deviations and linear loss for large deviations, which can reduce the impact weight of outliers while maintaining statistical efficiency for normal data. This robust aggregation method can still obtain a reliable reference feature vector even with a few abnormal channels, providing a robust benchmark for subsequent pseudo-vibration identification. The fundamental features of each channel include overall amplitude, order amplitude, band energy ratio, kurtosis, and spectral kurtosis, which have already been calculated in step 100. The robust aggregation of Huber loss is implemented using iterative weighted averaging. For a given weighted sample set, the weights are updated iteratively to reduce the impact of outliers, eventually converging to a robust aggregation result. The specific steps of this iterative process are as follows: First, the reliability weights of each channel are used as initial weights, and the weighted average is calculated as the initial reference feature vector. Then, the residuals between the basic features of each channel and the reference feature vector are calculated. The weights are adjusted according to the magnitude of the residuals; channels with larger residuals have lower weights, while channels with smaller residuals have their weights maintained or increased. The reference feature vector is recalculated using the adjusted weights. This process is repeated until the reference feature vector converges or the preset maximum number of iterations is reached. The preset maximum number of iterations defaults to 10, but ranges from 5 to 20, determined based on computational resources and convergence accuracy requirements. This iterative process typically converges after 3 to 5 iterations. The convergence criterion is that the relative change of each component of the reference feature vector between two consecutive iterations is less than 0.01.

[0080] For each channel, a pseudo-vibration residual is calculated. Specifically, the corresponding elements of the channel's fundamental eigenvector and the reference eigenvector are subtracted, and then the Euclidean norm of the difference vector is calculated. The Euclidean norm is the square root of the sum of the squares of the components of the difference vector. A larger pseudo-vibration residual indicates a greater deviation between the channel's fundamental eigenvector and the reference eigenvector obtained by fusing other channels in the same group; in other words, a higher degree of inconsistency between the channel's vibration behavior and that of other channels.

[0081] For each channel, a pseudo-vibration priority determination is performed. When a channel exhibits a combination of high pseudo-vibration residuals and low reliability weights, it is marked as a pseudo-vibration priority channel. High pseudo-vibration residuals refer to channels whose pseudo-vibration residuals exceed a preset residual threshold, indicating a significant deviation between the channel's vibration characteristics and the fused reference characteristics of other channels in the same group. Low reliability weights refer to channels whose reliability weights are below a preset weight threshold, indicating a high overall unreliability score calculated in step 200, meaning the channel has [a certain degree of unreliability]. Figure 1 Indicators of measurement link problems include high consistency costs, saturation clipping, spectral kurtosis anomalies, or zero-point offset variations. The default value of the preset residual threshold is 1.5 times the sum of the standard deviations of each component of the reference eigenvector, with a range of 1.0 to 2.5 times, determined based on the statistical distribution of normal differences between channels. The default value of the preset weight threshold is 0.4, with a range of 0.3 to 0.6, determined based on the system's tolerance for suspicious channels. When a channel simultaneously meets the conditions that the pseudo-vibration residual is greater than the preset residual threshold and the reliability weight is less than the preset weight threshold, it indicates that the channel exhibits both isolated anomalies inconsistent with other channels and a low score in its own measurement link quality assessment. This combination strongly suggests that the anomaly of the channel originates from a measurement link failure rather than actual mechanical vibration.

[0082] The purpose of prioritizing pseudo-vibration is to differentiate suspicious channels in subsequent scoring calculations, specifically in two aspects. Firstly, when calculating the mechanical anomaly score in step 400, channels marked as pseudo-vibration priority channels are removed to prevent their measurement link anomalies from interfering with the assessment of the true mechanical condition. Secondly, when calculating the pseudo-vibration score in step 400, channels marked as pseudo-vibration priority channels are given higher weights, making their contribution to the overall pseudo-vibration score more significant, thereby improving the system's sensitivity to identifying measurement link anomalies. Through this two-way adjustment mechanism, the false alarm rate of mechanical anomalies caused by pseudo-vibration is reduced while retaining the ability to detect real mechanical faults.

[0083] Step 400: Calculate the pseudo-vibration score based on the comprehensive unreliable score and pseudo-vibration residuals. After removing the pseudo-vibration priority channel, recalculate the reference feature vector and calculate the mechanical anomaly score, as detailed below. Figure 4 As shown.

[0084] The pseudo-vibration score is calculated as follows: For each channel, the product of the overall unreliability score and a preset first weighting coefficient is added to the product of the pseudo-vibration residual and a preset second weighting coefficient to obtain the weighted sum for that channel. Then, a Sigmoid function transformation is applied to this weighted sum to obtain the pseudo-vibration tendency value for that channel. For channels marked as pseudo-vibration priority channels in step 300, the weighted sum is multiplied by a preset amplification factor when calculating their pseudo-vibration tendency value. The default value of this preset amplification factor is 1.5, and its range is from 1.2 to 2.0, used to increase the contribution weight of the pseudo-vibration priority channel to the overall pseudo-vibration score. Among all the pseudo-vibration tendency values ​​of all channels, the value in the top 80% quantile after sorting is taken as the pseudo-vibration score, rather than directly taking the maximum value. The default value of the preset first weighting coefficient is 0.6, and the default value of the preset second weighting coefficient is 0.4. Both weighting coefficients range from 0 to 1, and their sum is 1.0, determined based on the trade-off between the channel's own reliability and the importance of its relative deviation. The Sigmoid function is calculated as follows: using the natural constant as the base and the negative value of the input as the exponent, the value of the exponential function is calculated, and then 1 is divided by the sum of the exponent value and 1. The Sigmoid function maps any real number input to the interval between 0 and 1. The design employs the 80th percentile instead of the maximum value, maintaining sensitivity to anomalies in the measurement link while reducing the excessive influence of occasional disturbances in a single channel on the overall pseudo-vibration score, avoiding the problem of excessive suppression of mechanical anomaly judgment caused by single-point triggering. When the total number of monitoring channels is less than 5, the pseudo-vibration score is taken as the median of the pseudo-vibration tendency values ​​of all channels, further enhancing robustness under small-scale channel configurations.

[0085] Before calculating the mechanical anomaly score, channels marked as pseudo-vibration priority channels in step 300 are first removed, leaving only unmarked channels for subsequent calculations. This removal is achieved by forcibly setting the reliability weight of pseudo-vibration priority channels to zero, ensuring they do not contribute to the robust aggregation of the reference feature vector. After removing pseudo-vibration priority channels, the reference feature vector is recalculated using the remaining channels. Then, this reference feature vector is used in conjunction with the operating conditions to calculate the mechanical anomaly score. Specifically, the reference feature vector, its trend increment, and the basic features of the operating parameters are concatenated into a single vector. The inner product of this concatenated vector and the preset weight vector is then calculated, and finally, the inner product is transformed using the Sigmoid function to obtain the mechanical anomaly score. The trend increment of the reference feature vector is obtained by calculating the difference between the current reference feature vector and the historical baseline reference feature vector, reflecting the trend of vibration characteristics over time. The reason for introducing the trend increment is that the development of mechanical faults typically manifests as a gradual change in vibration characteristics. The trend increment can capture this change process and provides richer information on fault evolution compared to using only the current feature value. The reason for introducing basic characteristics of operating parameters is that vibration characteristics are closely related to operating conditions. The same vibration amplitude has different degrees of abnormality under different loads or different operating conditions. Combining operating parameters can achieve abnormal scoring that is adaptive to operating conditions and reduce false alarms caused by changes in operating conditions.

[0086] The historical baseline reference feature vector refers to a benchmark value of the reference feature vector established through long-term statistics under normal unit operation. This benchmark value represents the typical vibration characteristic pattern of the unit under healthy conditions. The historical baseline reference feature vector is obtained during system initialization by collecting vibration data for at least 72 consecutive hours under stable unit operation conditions, weighting and robustly fusing the basic features of each channel, and then taking the time average. During system operation, the historical baseline reference feature vector is dynamically updated according to update conditions to adapt to the slow drift of vibration characteristics caused by seasonal changes, maintenance, or break-in. The difference between the current reference feature vector and the historical baseline reference feature vector reflects the degree of deviation of the current vibration state from the healthy benchmark, and this deviation is an important basis for judging mechanical anomalies. The basic features of the operating parameters include the current values ​​and rates of change of operating parameters such as load, main steam pressure, main steam temperature, and back pressure. The preset weight vector is obtained through supervised learning of historical fault samples and normal samples, trained using logistic regression or support vector machine methods. The reason for using supervised learning to train the weight vector is that mechanical anomaly scoring is essentially a binary classification problem, requiring the distinction between normal and abnormal states. Supervised learning can learn the optimal combination of feature weights from historical labeled samples, enabling the scoring results to achieve the best classification performance on the training samples. Logistic regression or support vector machines are chosen because these two methods have good generalization ability and interpretability. The weight coefficients of logistic regression directly reflect the direction and strength of each feature's contribution to the anomaly probability, while support vector machines can find the optimal classification hyperplane in the feature space by maximizing the classification margin.

[0087] The training steps for the preset weight vector are as follows: First, collect historical fault samples and normal samples. Fault samples are derived from confirmed mechanical fault cases in historical maintenance records, recording vibration characteristics and operating parameters before the fault occurred. Normal samples are derived from vibration characteristics and operating parameters during long-term stable operation of the unit. Second, extract a reference feature vector, the trend increment of the reference feature vector, and basic operating parameter features for each sample. Concatenate these three parts into a feature vector, and perform zero-mean, unit-variance standardization on each feature component. Record the mean and standard deviation parameters used for standardization for subsequent online applications. Third, label fault samples as positive classes and normal samples as negative classes to construct a training dataset. Fourth, train the dataset using logistic regression or support vector machine algorithms. For logistic regression, maximum likelihood estimation is used to solve for the weight vector; for support vector machine, sequential minimum optimization is used to solve for the support vector and weight vector. Fifth, evaluate the classification performance of the trained weight vector on an independent validation dataset, calculating metrics such as accuracy, recall, and F1 score. If the performance does not meet the requirements, adjust the regularization parameters or feature selection and retrain. The sixth step is to save the trained weight vector and standardized parameters for use by the online anomaly identification system. Before training, each basic feature component is standardized with zero mean and unit variance to ensure the comparability of basic features with different dimensions. The training samples include at least 50 confirmed fault cases and at least 200 normal operating condition samples. Sample labeling is determined by domain experts in conjunction with maintenance records and vibration analysis reports. The values ​​of each component of the trained weight vector range from -5.0 to +5.0. The sign and magnitude of the weights reflect the direction and importance of the corresponding basic feature in indicating mechanical anomalies. In a typical embodiment, the weight of the first harmonic growth basic feature is +2.5, the weight of the second harmonic growth basic feature is +1.8, the weight of the phase drift basic feature is +1.2, and the weight of the spectral kurtosis increase basic feature is +1.5. These values ​​are only examples; in actual applications, they are determined through supervised learning based on the specific unit type and fault mode.

[0088] The update conditions are defined as follows: the mechanical anomaly score is less than the preset baseline update threshold, the spurious vibration score is less than the preset baseline update threshold, and the operating condition is stable. The default value of the preset baseline update threshold is 0.30, and the range is 0.2 to 0.4. This value is lower than the discrimination threshold to ensure that the baseline is updated only under clearly normal conditions. The determination condition for the stable operating condition is that the absolute values ​​of the speed change rate and the load change rate are both lower than their respective preset stability thresholds. The default value of the preset stability threshold for the speed change rate is 0.5% / minute, and the default value of the preset stability threshold for the load change rate is 2% / minute. The range of values ​​is determined according to the unit's regulation characteristics and steady-state operation standards.

[0089] Within a preset sliding window that meets the update conditions, the historical baseline reference feature vector is updated using an exponential sliding method. Specifically, the current historical baseline reference feature vector is multiplied by 1 (the difference between this and the preset learning rate), and then added to the product of the current window's reference feature vector and the preset learning rate. The sum of these two values ​​yields the updated historical baseline reference feature vector. The preset learning rate has a default value of 0.05 and ranges from 0.01 to 0.2. A smaller value indicates a more conservative baseline update, while a larger value indicates faster adaptation to new data. The value is determined based on a trade-off between the system's tracking requirements for slow drift and its anti-interference capabilities. This update mechanism allows the system to adapt to seasonal changes or slow drift after maintenance, while avoiding oversensitivity.

[0090] Step 500: Based on the pseudo-vibration score and mechanical anomaly score, perform anomaly determination and output the determination result.

[0091] A dual-gated discrimination logic is used for anomaly detection. When the pseudo-vibration score is greater than or equal to a preset pseudo-vibration threshold and the mechanical anomaly score is less than a preset mechanical anomaly threshold, the output result is either a measurement link anomaly or pseudo-vibration. When the mechanical anomaly score is greater than or equal to a preset mechanical anomaly threshold and the pseudo-vibration score is less than a preset pseudo-vibration threshold, the output result is mechanical abnormal vibration. When both the pseudo-vibration score and the mechanical anomaly score are greater than or equal to a preset mechanical anomaly threshold, the output result is suspected mechanical abnormal vibration accompanied by measurement anomaly, triggering channel switching and verification suggestions. The reason for using the dual-gated discrimination logic is that a single score is difficult to accurately distinguish between genuine mechanical anomalies and measurement link anomalies. By simultaneously calculating the pseudo-vibration score and the mechanical anomaly score, and making a judgment based on the combination pattern of the two scores, a more accurate anomaly classification can be achieved. When the false vibration score is high and the mechanical anomaly score is low, it indicates that the anomaly mainly originates from the measurement link. When the mechanical anomaly score is high and the false vibration score is low, it indicates that the anomaly mainly originates from a real mechanical fault. When both scores are high, it indicates a complex situation where mechanical and measurement anomalies may coexist, requiring further verification. This dual-gating logic can effectively reduce the false alarm rate while maintaining the sensitivity to detect real faults. The default value of the preset false vibration discrimination threshold is 0.65, and the value range is 0.5 to 0.8, determined based on the tolerance for false vibration false alarms. The default value of the preset mechanical anomaly discrimination threshold is 0.70, and the value range is 0.6 to 0.85, determined based on the trade-off between the risk of missed mechanical faults and the cost of false alarms.

[0092] When the vibration is determined to be spurious, the most suspicious channel and its supporting evidence are output. This evidence includes: the channel's overall unreliability score, the consistency residual vector with adjacent channels, the presence of saturation or packet loss, and the suspected probe resonance frequency band. The purpose of outputting this evidence is to provide maintenance personnel with interpretable diagnostic information, helping them quickly pinpoint the specific cause and location of the measurement link anomaly. The overall unreliability score reflects the channel's overall reliability level, the consistency residual vector reflects the degree and dimension of deviation between the channel and adjacent channels, saturation or packet loss indicates hardware failure in the signal acquisition stage, and the suspected probe resonance frequency band indicates mechanical installation problems. Suggested actions are also provided, including checking the installation preload, probe length, grounding shielding, channel replacement, or temporarily downweighting and shielding the channel. These suggested actions are based on typical handling methods for different types of measurement link anomalies, providing operational guidance for maintenance personnel.

[0093] When the determination result is abnormal mechanical vibration, the term contributing the most in the output reference feature vector is used as evidence. This includes changes in basic features such as first-harmonic growth, second-harmonic growth, phase drift, and increased spectral kurtosis, along with the anomaly confidence level under the current operating conditions, for operational decision-making reference. The purpose of outputting the feature term contributing the most is to provide maintenance personnel with a preliminary basis for judging the fault type. First-harmonic growth usually indicates rotor imbalance, second-harmonic growth usually indicates rotor misalignment, phase drift usually indicates changes in support stiffness or cracks, and increased spectral kurtosis usually indicates rubbing or bearing damage. The purpose of providing the anomaly confidence level is to quantitatively assess the severity of the anomaly and provide a reference for maintenance personnel to formulate response strategies. High-confidence anomalies require immediate action, while low-confidence anomalies can continue to be observed.

[0094] Through the above steps, the present invention achieves online identification of abnormal turbine vibration, effectively distinguishes between real abnormal mechanical vibration and spurious vibration caused by the measurement link, provides an interpretable chain of evidence, and improves the reliability and practicality of the vibration monitoring system.

[0095] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for online identification of abnormal vibration in a steam turbine, characterized in that, include: Vibration signals and key phase signals from each channel are collected, the quality of the key phase signals is evaluated and processed, and the basic characteristics of each channel are calculated within a sliding window. A redundant consistency graph is established. Based on the redundant consistency graph and basic features, a consistency residual vector is calculated, and a graph consistency cost is calculated for each node. Based on the graph consistency cost, a comprehensive unreliability score and a reliability weight are calculated, including: Each channel is treated as a node, and a redundancy consistency graph is established. The redundancy consistency graph contains a set of nodes and a set of edges. The set of edges includes multiple edges. An edge is established between two channels in the same bearing orthogonal direction. An edge is established between channels of the same type of sensor on adjacent bearings. If there are cross-type channels of shaft vibration and bearing vibration or shell vibration, an edge is established. For each edge in the redundant consistency graph, a consistency residual vector is calculated, which includes order magnitude consistency residual, coherent consistency residual, and phase difference stability residual. The consistency residual vector is standardized. The graph consistency cost is calculated for each node based on the redundant consistency graph, including traversing all edges connected to the node, summing the absolute values ​​of each component of the consistency residual vector for each edge to obtain the residual norm of the edge, multiplying the residual norm by a preset edge weight to obtain the weighted residual norm, and summing the weighted residual norms of all connected edges to obtain the graph consistency cost of the node. Calculate the overall unreliability score based on the graph consistency cost; Map the overall unreliability score to a reliability weight; A reference feature vector is constructed by weighted robust fusion of the basic features of multiple channels. Based on the reference feature vector and the basic features, the pseudo-vibration residual is calculated for each channel, and pseudo-vibration priority is determined. Based on the comprehensive unreliable score and the pseudo-vibration residual, the pseudo-vibration score is calculated. After removing the pseudo-vibration priority channel, the reference feature vector is recalculated and the mechanical anomaly score is calculated. Anomaly determination is performed based on the pseudo-vibration score and mechanical anomaly score, and the determination result is output.

2. The online identification method for abnormal turbine vibration according to claim 1, characterized in that, The evaluation and processing of the bond phase signal quality includes: The acquired vibration signals from each channel and the key phase signals are time-aligned. The evaluation of the bond phase signal quality includes calculating the coefficient of variation of the time interval between adjacent bond phase pulses. If the coefficient of variation exceeds a preset bond phase quality threshold, the bond phase signal is deemed to have failed; otherwise, it is deemed to have passed. When the key phase signal quality assessment passes, the time-domain signal is converted into an order-domain signal; when the key phase signal quality assessment fails, the time-domain analysis mode is maintained, and the short-time Fourier transform is used to extract the frequency-related features. The fundamental characteristics include overall amplitude, order amplitude, band energy ratio, kurtosis, and spectral kurtosis.

3. The online identification method for abnormal turbine vibration according to claim 1, characterized in that, The order magnitude consistency residual is: Determine whether the first octave amplitude of two connected channels is greater than a preset amplitude threshold; When the first octave amplitude of both nodes is greater than the preset amplitude threshold, the first octave amplitude of the two nodes is added to a preset small positive constant to obtain a sum. The natural logarithm of the two sums is taken respectively, and the absolute value of the difference between the two logarithm values ​​is calculated as the order amplitude consistency residual. When the first octave amplitude of any node is lower than the preset amplitude threshold, the absolute value of the difference between the first octave amplitudes of the two nodes is divided by the amplitude term. The amplitude term is the sum of the first octave amplitudes of the two nodes plus a preset small positive constant, which is used as the order amplitude consistency residual.

4. The online identification method for abnormal turbine vibration according to claim 1, characterized in that, The coherence consistency residual is obtained by: calculating the coherence coefficients of two connected nodes at multiple characteristic frequencies, including first harmonics, second harmonics and half harmonics; and subtracting the coherence coefficient from 1 at each characteristic frequency to obtain the coherence residual components at each frequency. The coherent consistency residual is obtained by weighted averaging of the coherent residual components at each frequency. The phase difference stability residual is obtained by: acquiring the phase value sequences of the two connected channels at the first harmonic; and calculating the phase difference sequence between the two phase value sequences. The phase difference sequence is subjected to phase expansion processing to eliminate the boundary jump of positive and negative pi; the variance of the expanded phase difference sequence is calculated as the phase difference stability residual.

5. The online identification method for abnormal turbine vibration according to claim 1, characterized in that, Based on the reference feature vector and the basic features, the pseudo-vibration residual is calculated for each channel, and the pseudo-vibration priority determination includes: Within the same bearing or rotor segment, the basic features of multiple channels are weighted and robustly fused to construct a reference feature vector, which is calculated using a robust aggregation method based on Huber loss. For each channel, calculate the pseudo-vibration residual, subtract the corresponding elements of the basic feature of the channel from the reference feature vector to obtain the difference vector, and calculate the Euclidean norm of the difference vector. When the pseudo-vibration residual of the channel exceeds the preset residual threshold and the reliability weight is lower than the preset weight threshold, the abnormality of the channel is determined to be pseudo-vibration priority.

6. The online identification method for abnormal turbine vibration according to claim 1, characterized in that, The calculation of the pseudo-vibration score includes: For each channel, the unreliable score and the pseudo-vibration residual are weighted and summed, and the weighted sum is nonlinearly transformed to obtain the pseudo-vibration tendency value of the channel. For channels marked as pseudo-vibration priority channels, the weighted sum is amplified when calculating their pseudo-vibration tendency values; Among all the pseudo-vibration tendency values ​​of all channels, the value of the preset quantile is taken as the pseudo-vibration score.

7. The online identification method for abnormal turbine vibration according to claim 1, characterized in that, The computational mechanical anomaly score includes: Remove the channels marked as pseudo-vibration priority channels and recalculate the reference eigenvectors using the remaining channels; The trend increment of the reference feature vector is calculated, which is obtained by calculating the difference between the current reference feature vector and the historical baseline reference feature vector. The reference feature vector, trend increment, and basic features of operating parameters are concatenated into a vector. The inner product of the vector and the preset weight vector is calculated. The inner product is then subjected to a nonlinear transformation to obtain the mechanical anomaly score. The basic features of operating parameters include the current values ​​and rates of change of operating parameters such as load, main steam pressure, main steam temperature, and back pressure.

8. The online identification method for abnormal turbine vibration according to claim 1, characterized in that, Anomaly determination and output of determination results based on the pseudo-vibration score and mechanical anomaly score include: When the false vibration score is greater than or equal to the preset false vibration discrimination threshold and the mechanical anomaly score is less than the preset mechanical anomaly discrimination threshold, the output judgment result is either measurement link anomaly or false vibration. When the mechanical anomaly score is greater than or equal to the preset mechanical anomaly discrimination threshold and the pseudo vibration score is less than the preset pseudo vibration discrimination threshold, the output judgment result is mechanical anomaly vibration. When the false vibration score is greater than or equal to the preset false vibration discrimination threshold and the mechanical anomaly score is greater than or equal to the preset mechanical anomaly discrimination threshold, the output judgment result is suspected mechanical abnormal vibration accompanied by measurement anomaly.

Images