A rotating machinery fault diagnosis method based on dilated residual network

Abstract

The application discloses a rotating machinery fault diagnosis method based on an expanded residual network, relates to the technical field of mechanical faults, and comprises the following steps: firstly, collecting continuous time sequence vibration signals of rotating machinery by using an acceleration sensor; obtaining a high-order harmonic energy sequence set and determining an abnormal trend segment after fast Fourier transform; obtaining and updating a structural abnormal candidate set based on the abnormal trend segment and the running environment state of the rotating machinery; generating a first drift vector set according to the updated structural abnormal candidate set; identifying a structural deviation section according to the first drift vector set and historical running data; tracing the structural deviation section back to an original time domain signal to establish a second drift vector set; and finally, inputting the first drift vector set and the second drift vector set into an expanded residual network to analyze whether the rotating machinery is evolving into a fault state.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of mechanical faults, in particular to a rotating machine fault diagnosis method based on a dilated residual network. BACKGROUND

[0002] The application relates to the technical field of mechanical equipment operation state monitoring and intelligent fault diagnosis, in particular to the subfield of rotating machine state recognition and fault early warning in industrial equipment, and particularly relates to a fault diagnosis method for multi-scale dynamic analysis by using a deep learning structure, namely a dilated residual network (DRN). The method is mainly applied to typical rotating machine systems such as steam turbines, centrifugal compressors, motors, gearboxes and blowers. By deep recognition of high-order harmonic energy drift characteristics, it can accurately determine whether the system is in a structural degradation evolution state, so as to realize early identification and risk quantitative control of potential fault trends.

[0003] During the operation of a rotating machine, subtle misalignment, early loosening and lubrication abnormalities and other implicit structural faults often do not cause significant time-domain impacts or frequency jumps, but are manifested in the form of slow drift of high-order harmonic energy. However, the high-order harmonic characteristics are non-obvious, weakly disturbed and easily confused, and are easily disturbed by non-fault factors such as background load fluctuation, working condition change and speed disturbance, leading to misjudgment and misdiagnosis of the drift. Therefore, although a large number of fault recognition methods based on vibration signals have been proposed, there is a certain neglect of the form of slow drift of high-order harmonic energy, for example, in a wind farm, early peeling of a bearing is only manifested as a slight drift of a certain high-order harmonic energy, which is easily covered by wind speed change or blade load disturbance, and a traditional model may consider it as normal fluctuation and ignore it. SUMMARY

[0004] In view of the deficiencies of the prior art, the application provides a rotating machine fault diagnosis method based on a dilated residual network, which solves the problems in the background art.

[0005] To achieve the above object, the application is implemented by the following technical scheme: a rotating machine fault diagnosis method based on a dilated residual network, comprising the following steps: collecting continuous time sequence vibration signals of a rotating machine by using an acceleration sensor, obtaining a high-order harmonic energy sequence set after fast Fourier transform, and determining an abnormal trend segment;

[0006] Based on the abnormal trend segment and the operating environment state of the rotating machine, a structural abnormality candidate set is obtained and updated, and a first drift vector set is generated according to the updated structural abnormality candidate set;

[0007] According to the first set of drift vectors and historical operation data, identify the structural deviation section, and trace the structural deviation section back to the original time domain signal to establish the second set of drift vectors;

[0008] By inputting the first set of drift vectors and the second set of drift vectors into the inflation residual network, it is analyzed whether the rotating machinery is evolving into a fault state.

[0009] Preferably, the acceleration sensor is used to collect continuous time sequence vibration signals of the rotating machinery, and a window length is set to divide the continuous time sequence vibration signals into multiple groups of time segment signals in a sliding window manner.

[0010] The corresponding frequency spectrum is obtained by fast Fourier transform for each group of time segment signals.

[0011] The corresponding fundamental frequency is extracted from the frequency spectrum, and the amplitude of each order harmonic frequency point is calculated to form a high-order harmonic energy sequence set, which is specifically represented as: Where H is the high-order harmonic energy sequence set. 、 and are the energy time sequences on the 2nd, 3rd and vth order harmonics respectively, and v is the highest order harmonic set.

[0012] According to the energy time sequences on each order harmonic, the energy change rate of each order harmonic frequency point is calculated by the first-order difference method to analyze the dynamic change speed of each order harmonic frequency point, which is specifically represented as: Where is the energy change rate of the mth order harmonic frequency point.

[0013] Based on the energy change rate, the energy change rate sequence of each order harmonic is obtained. When the energy change rate of the corresponding order harmonic frequency point in the continuous time segment exceeds the preset change threshold, it indicates that there is an abnormal trend segment. After statistics, the abnormal trend set of the energy time sequence of each order harmonic is obtained.

[0014] Preferably, according to the abnormal trend set of the energy time sequence of each order harmonic, the synchronous drift trend between the abnormal trend segments of each order harmonic is judged. If the condition is met: Then it is judged that the energy time sequences of each order harmonic are in homologous coupling, and are included in the structural abnormal candidate set, otherwise they are not in homologous coupling and are not included in the structural abnormal candidate set; wherein, and are the energy time sequences on the mth and nth order harmonics respectively, is a proportional factor.

[0015] Preferably, according to each order harmonic frequency point, a non-integer multiple wave peak frequency in the corresponding frequency spectrum is identified, and a drift rate trajectory of each non-integer multiple wave peak frequency is obtained, which is specifically represented as: wherein, is the drift speed of the kth sub-harmonic frequency point in time, is the drift amount of the kth sub-harmonic frequency point, is the time span of the frequency change, and k is the sub-harmonic frequency point number.

[0016] Preferably, by collecting the running environment state of the rotating machinery, a disturbance index vector is obtained, which is specifically: wherein, is the time series of the disturbance index vector, is the time series of the power fluctuation rate, is the time series of the rotating speed, and are respectively the time series of the spectrum expansion factor on the 2nd order harmonic and the vth order harmonic;

[0017] By comparing each component in the time series of the disturbance index vector with the corresponding preset disturbance threshold, if there is a component exceeding the corresponding preset disturbance threshold, it is considered that the corresponding time segment is disturbed, otherwise it is considered that the corresponding time segment is not disturbed.

[0018] If the time segment affected by the disturbance is located in the structural anomaly candidate set, it is removed, the structural anomaly candidate set is updated, and the energy change rate of the corresponding order harmonic frequency point in the time segment corresponding to the structural anomaly candidate set is extracted to generate a first drift vector set.

[0019] Preferably, the energy change rate of the multi-order harmonic frequency point under the healthy working condition is screened out from the historical running data to generate a normal drift vector set, and the normal drift vector set and the first drift vector set are aligned in the same time range to analyze the trend consistency of the normal drift vector set and the first drift vector set, and a difference sequence is obtained, which is specifically represented as: wherein, is the difference value, is the first drift vector set, is the normal drift vector set;

[0020] By using the ADF test method, it is tested whether the difference sequence has a unit root, if the test result is to reject the null hypothesis, it means that the difference sequence is in a stationary state as a whole and does not have a unit root, otherwise it is determined that the rotating machinery in the corresponding time window is in a cointegration destruction state, and the time window where the cointegration destruction occurs is marked as a structural deviation section.

[0021] Preferably, when the structural offset section is identified, a backtracking mechanism is started, the structural offset section is corresponded to the collection time period of the continuous time sequence vibration signal using the sliding window mapping relationship, and the vibration signal of the collection time period corresponding to the structural offset section is intercepted, which is recorded as an original signal subsequence;

[0022] The drift rate trajectories of the peak frequencies of each non-integer multiple frequency in the collection time period corresponding to the original signal subsequence are extracted, and a second drift vector set is constructed in combination with the first drift vector set in the structural offset section.

[0023] Preferably, the first drift vector set and the second drift vector set are projected into the same dimensional representation space by using an autoencoder to obtain the first drift vector set and the second drift vector set in the same dimensional representation space.

[0024] By inputting the first drift vector set and the second drift vector set in the same dimensional representation space into the dilated residual network, after the dilated convolution layer, the residual connection structure, the local attention layer and the full connection layer, a risk assessment value is output, specifically: wherein, is the risk assessment value of the corresponding time, is an activation function, is a weight matrix of the full connection layer, is a residual feature extraction block, and the first drift vector set and the second drift vector set in the same dimensional representation space, respectively, is a bias term.

[0025] Preferably, in the structural offset section, the evaluation difference is calculated in sequence, specifically: wherein, is the evaluation difference, is the risk assessment value of time t in the structural offset section, is the risk assessment value corresponding to the time after time t;

[0026] The maximum evaluation difference is extracted, and the maximum evaluation difference is compared with a preset evaluation threshold value. If the maximum evaluation difference exceeds the evaluation threshold value, it is judged that the current rotating machine has a fault risk and a fault instruction is uploaded to an operation background to perform shutdown processing.

[0027] The present application provides a rotating machine fault diagnosis method based on a dilated residual network, which has the following beneficial effects:

[0028] (1) By obtaining a high-order harmonic energy sequence set in the spectrum analysis stage, and combining the windowed difference method to identify abnormal trend segments, the weak and continuous energy drift signals exhibited by rotating machinery in the early stage of operation can be captured, effectively enhancing the early perception ability of structural evolution faults, and further overcoming the problem of traditional dependence on energy mutation points prone to missing gradual faults. In combination with the operating environment state, a disturbance index vector is constructed and dynamically compared and eliminated with the abnormal trend segments, thereby significantly improving the purity of the structural abnormal candidate set. Through this mechanism, the system can automatically exclude pseudo-abnormal segments caused by background load fluctuations, speed disturbances and other non-fault factors, improving the accuracy and robustness of subsequent diagnosis. By comparing the first drift vector set with the historical normal data, the cointegration destruction area is extracted and backtracked to the original time domain signal to establish the second drift vector set, combined with the self-encoder projection mechanism and the inflation residual network, the nonlinear structure comparison of the drift trend is realized to output the structural trend risk score index R(t), which is used to judge whether the system is evolving from a healthy state to a structural instability state, providing a quantitative criterion and feedback mechanism for online fault warning of rotating machinery. In summary, the present application constructs an integrated closed-loop process of abnormal trend identification, disturbance stripping, multi-round backtracking reconstruction and networked risk scoring, not only improves the identification accuracy and real-time performance of complex nonlinear structural faults, but also enhances the disturbance immunity in dynamic operating environment, and can be applied to intelligent operation and maintenance and reliability protection systems of key rotating machinery such as wind power, motors, compressors and pumps.

[0029] (2) By setting a sliding window to segment the original acceleration signal, and performing fast Fourier transform on each time segment, a set of multi-order harmonic energy sequences that can be continuously tracked in the frequency domain is formed, which can effectively amplify the change trend of small harmonic components, especially in the early stage of structural loosening, bearing micro-cracking or coupling deformation defects, the energy fluctuation of specific order harmonics can be observed, providing a signal basis for early fault prediction. By implementing first-order difference processing on each harmonic energy time sequence, the energy change rate index is obtained, and the preset change threshold is used as the judgment basis to identify abnormal trend segments that continuously exceed the threshold. Further, by judging the synchronous drift trend among multi-order harmonics, a homologous coupling relationship is constructed, and a structural abnormal candidate set is formed, which can effectively exclude non-structural abnormalities caused by accidental noise or short-term disturbances, improving the robustness and credibility of trend judgment.

[0030] (3) When an index anomaly is detected in a certain time segment, the structural anomaly candidate corresponding to the segment can be eliminated, reducing the misdiagnosis rate caused by background fluctuations, speed variations and other non-structural disturbances, and improving the false alarm suppression capability and the purity of structural anomaly identification of the system from the root. By extracting the energy change rate of the multi-order harmonic frequency points under healthy working conditions from the historical running data, a normal drift vector set is constructed, and a difference sequence is formed by time alignment and difference operation between the drift vector set generated by the current structural anomaly candidate set and the first drift vector set. The difference sequence can be regarded as the deviation index of the current system running state relative to the healthy state, and directly reflects whether the system is in a trend section of drift intensification, frequency imbalance or nonlinear response amplification, providing basic data for subsequent trend consistency detection. By performing ADF test on the above difference sequence, it is determined whether it has a unit root, so as to determine whether the current trend is stable. If the difference sequence is non-stationary, that is, the co-integration relationship is destroyed, it is determined that the rotating machinery system has fallen into a structural deviation evolution state, and the trend backtracking mechanism needs to be triggered. The mechanism can automatically identify whether the current system has evolved from a normal state to an early structural abnormal state, and provide an accurate time anchor point for subsequent high-resolution backtracking diagnosis and risk score analysis, realizing adaptive identification and early warning of mechanical structural faults.

[0031] (4) When the system identifies the co-integration destruction area, the backtracking mechanism is automatically started, and the structural deviation section is mapped back to the corresponding time sequence original vibration signal based on the sliding window mapping relationship, so as to intercept the original signal subsequence of the corresponding section. The backtracking method avoids the problems of feature breakage and stage information loss, can accurately locate the source area of structural abnormal evolution under unsupervised conditions, and thus reconstructs the original physical feature trajectory of the abnormal occurrence, improving the fault root cause identification accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 A flowchart of a rotating machinery fault diagnosis method based on an expansion residual network according to the present application;

[0033] Figure 2 A logic diagram of a rotating machinery fault diagnosis method based on an expansion residual network according to the present application. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0035] Example 1: Please refer to Figure 1 and Figure 2The application provides a rotating machinery fault diagnosis method based on an expanded residual network, comprising the following steps:

[0036] S1: acquiring continuous time sequence vibration signals of the rotating machinery by using an acceleration sensor, obtaining a high-order harmonic energy sequence set after fast Fourier transform, and determining an abnormal trend segment;

[0037] In the technical system of rotating machinery operation state monitoring and fault identification, the acquisition and processing of vibration signals are the starting link of the entire diagnosis system, and the accuracy and feature expression capability thereof directly affect the effectiveness and response speed of subsequent diagnosis indexes. Therefore, the core target of step S1 is to extract a high-order harmonic energy spectrum capable of expressing the running state of the machinery from the original time domain vibration data through multi-channel data acquisition and time-frequency joint transformation means, and to provide data support and feature coordinate basis for subsequent drift modeling and feature stripping.

[0038] S2: acquiring and updating a structural abnormality candidate set based on the abnormal trend segment and the running environment state of the rotating machinery, and generating a first drift vector set according to the updated structural abnormality candidate set;

[0039] S3: identifying a structural deviation section according to the first drift vector set and historical running data, and tracing the structural deviation section back to the original time domain signal to establish a second drift vector set;

[0040] S4: inputting the first drift vector set and the second drift vector set into the expanded residual network to analyze whether the rotating machinery is evolving into a fault state.

[0041] In this embodiment, after acquiring the continuous time sequence vibration signals and performing high-order harmonic analysis, the modeling of the structural trend behavior is realized by identifying the abnormal trend segment instead of a single-point abnormality. Unlike the traditional method based on threshold judgment, this method emphasizes the trend abnormality on the continuous segment instead of the instantaneous abnormality, thereby effectively avoiding false alarms caused by background disturbances, such as motor start fluctuation and load mutation. For example, in a wind power generation system, when the fluctuation of wind speed causes the temporary increase of harmonic energy, the traditional method may misjudge as bearing fault, but the present method will avoid false positives by judging whether there is a continuous energy growth trend.

[0042] By jointly analyzing the abnormal trend segment and the running environment state, such as speed change and power fluctuation, a structural abnormality candidate set is generated and updated, so as to eliminate disturbance segments caused by non-structural factors, thereby forming a more pure and more explicit physical meaning first drift vector set.

[0043] The structural anomaly candidate set represents a set of segments that may truly reflect internal defects of the machine, such as cracks and wear. The disturbance indicator vector is extracted from the operating environment signal to identify anomalies related to non-structural disturbances, helping to achieve disturbance stripping. By backtracking the structural deviation segments to the original time-domain signal, the present application realizes the ability to track from the frequency domain anomaly to the original sampling point, further extracts the second drift vector set, and thus covers deep features that are difficult to identify by traditional methods, such as sub-harmonic disturbances. For example, in the early stage of motor insulation damage, only a slight shift in some sub-harmonic frequencies will occur, and this signal is easily overlooked in the overall spectrum. By using the present method, the anomaly can be backtracked to a specific time period and focused on, thereby enhancing feature extraction.

[0044] After aligning the first and second drift vector sets through the autoencoder, the DRN (Dilated Residual Network) is input, which can perceive the trend residual between the two rounds of diagnostic features in a unified structural space, and finally outputs the risk score R(t). This score has the advantage of strong trend perception, supporting gradual tracking and dynamic risk judgment of the structural degradation process.

[0045] The DRN (Dilated Residual Network) is a combination of multi-scale perception and residual learning structure, used to model the fault trend evolution process. The output value of the risk score R(t) is between 0 and 1, representing whether the current structural trend is strengthening, whether it needs to be warned or intervened.

[0046] In summary, this method improves the sensitivity (early detection) and specificity (few false positives) of rotating machinery fault diagnosis, and is particularly suitable for early fault warning tasks in wind power, motors, petrochemical compressors, and other scenarios.

[0047] Embodiment 2: Please refer to Figure 1 Specifically, the acceleration sensor is used to collect continuous time series vibration signals of the rotating machinery, and a window length is set to divide the continuous time series vibration signals into multiple groups of time segment signals in a sliding window manner.

[0048] Here, the acceleration sensor is used to collect continuous time series vibration signals of the rotating machinery, which are the original time-domain signals.

[0049] Each group of time segment signals is subjected to fast Fourier transform to obtain the corresponding frequency spectrum.

[0050] The corresponding fundamental frequency is extracted from the frequency spectrum, and the amplitude of each order harmonic frequency point is calculated to form a high-order harmonic energy sequence set, which is specifically represented as: Where H is the high-order harmonic energy sequence set, containing energy time sequences of all selected orders. 、 And are the energy time sequences on the 2nd, 3rd, and vth order harmonics, respectively, and v is the highest order harmonic set.

[0051] In the operation of rotating machinery, the vibration signal contains multiple frequency components, including not only the fundamental frequency (usually corresponding to the rotational speed or the frequency of one revolution of the shaft), but also integer multiples of the fundamental frequency, referred to as harmonic frequency points.

[0052] When the vibration signal in a time period is subjected to fast Fourier transform (FFT), a frequency spectrum is obtained, i.e., the amplitude of each frequency point (corresponding to the strength of the frequency component). For example, the amplitude of the 2nd harmonic frequency point is the peak height of the spectrum at this frequency. The greater the amplitude, the stronger the frequency component in the current vibration.

[0053] According to the energy time series on each harmonic, the energy change rate of each harmonic frequency point is calculated by the first-order difference method to analyze the dynamic change speed of each harmonic frequency point. Specifically, wherein, is the energy change rate of the mth harmonic frequency point.

[0054] Based on the energy change rate, the energy change rate sequence on each harmonic is obtained. When the energy change rate of the corresponding harmonic frequency point in a continuous time segment exceeds the preset change threshold, it indicates that there is an abnormal trend segment. After statistics, the abnormal trend set of the energy time series on each harmonic is obtained. When the energy change rate of the corresponding harmonic frequency point in a continuous time segment does not exceed the preset change threshold, it indicates that there is no abnormal trend segment at present.

[0055] wherein, the continuous time segment refers to more than two continuous time segments.

[0056] According to the abnormal trend set of the energy time series on each harmonic, the synchronous drift trend between the abnormal trend segments on each harmonic is judged. If the condition is met: then it is judged that the energy time series on each harmonic is in homogenous coupling, and is included in the structural abnormality candidate set, otherwise it is not in homogenous coupling and is not included in the structural abnormality candidate set; wherein, and are the energy time series on the mth harmonic and the nth harmonic, respectively, is a proportional factor used to measure the drift trend intensity ratio between two different harmonics, for example, the change amplification or reduction degree of the drift trend of the 3rd harmonic relative to the drift trend of the 2nd harmonic.

[0057] If a number of harmonic sequences constitute homogenous coupling, it indicates that there may be a cross-band physical abnormal mode in the system, and its excitation appears as a synchronous drift of multi-harmonic response.

[0058] According to each order harmonic frequency point, the non-integer multiple frequency peak frequency in the corresponding frequency spectrum is identified, and the drift rate trajectory of each non-integer multiple frequency peak frequency is obtained, which is specifically represented as: wherein, is the drift speed (i.e. frequency change rate) of the kth sub-harmonic frequency point in time, is the drift amount of the kth sub-harmonic frequency point, is the time span of the frequency change, and k is the sub-harmonic frequency point number.

[0059] The drift rate trajectory of each non-integer multiple frequency peak frequency can be used to assist in judging whether the abnormal source is from a nonlinear disturbance;

[0060] In the embodiment, by setting the sliding window length and performing slicing processing on the acceleration signal, the system can accurately analyze the local dynamic change in the long-time running process, and overcome the problem that the traditional global analysis method is not sensitive to short-time disturbance. Each sliding window segment is converted to the frequency domain by fast Fourier transform (FFT) to ensure the timeliness and distinguishability of the frequency spectrum feature extraction. For example, during the running process of an industrial compressor, when the load of the compressor suddenly increases, a short-term high-frequency disturbance peak will appear in the frequency spectrum. If windowing processing is not performed, the peak will be easily submerged by the overall energy and cannot be identified.

[0061] By extracting the high-order harmonic frequency amplitude and analyzing the energy change rate, multi-order frequency drift monitoring and abnormal trend identification are realized. Specifically, the fundamental frequency and its multiple integer harmonic positions are extracted from each group of frequency spectra, and a high-order harmonic energy sequence set is constructed. Then, first-order difference calculation is applied to these sequences to obtain the energy change rate at each harmonic point, and a representation of the dynamic behavior of the frequency point is formed. This strategy enables the system not only to identify sudden changes in frequency point energy, but also to capture gradual growth type faults, such as the early energy slowly changing behavior trend in bearing wear. A multi-layer abnormal mechanism for trend segment discrimination, synchronous drift clustering and nonlinear sub-harmonic disturbance identification is constructed, which is specifically:

[0062] a) Trend segment extraction: if the energy change rate of a certain order harmonic continuously exceeds the preset threshold in multiple windows, it is marked as a possible fault evolution behavior, forming an abnormal trend set.

[0063] b) Synchronous drift judgment: among multiple harmonics, if their abnormal segments show similar trends (i.e. synchronous drift behavior), it is considered that these signals have homologous coupling, i.e. they come from the same structural instability mechanism, and enter the structural abnormal candidate set. For example, if the 2nd and 4th order harmonics simultaneously show slow rising and synchronous jitter, it may indicate that the coupling is loose; on the contrary, only a sudden mutation in a certain order is caused by external impact.

[0064] c) Nonlinear disturbance identification: further extract the wave peak position of non-integer multiple frequency points (sub-harmonic) in the spectrum, and calculate its drift rate trajectory to determine whether it is from background disturbance, such as load fluctuation, instantaneous speed change, etc. If the sub-harmonic frequency point drifts frequently, but the high-order integer harmonic drifts stably, it can be judged that the abnormal source of the system is non-structural disturbance, and misdiagnosis is avoided. Take a wind turbine gearbox as an example, when the high-speed shaft bearing begins to shed slightly, the energy sequence of the 3rd and 5th order harmonics may show gradual rise and similar fluctuation frequency (synchronous drift); At this time, due to the continuous fluctuation of wind speed, the spectrum in the sub-harmonic region appears to drift, but the 2nd order harmonic has no obvious change. Through the above method, the system can accurately identify that the phenomenon is a false alarm caused by structural abnormal evolution rather than disturbance, improve the diagnosis accuracy and avoid unnecessary shutdown, at the same time, as part of the construction of the second drift vector set, after the backtracking mechanism is started, the drift trajectory of these non-integer multiples is extracted as a marker signal of deep nonlinear behavior, and is integrated into the second drift vector set to participate in risk score calculation;

[0065] The application constructs a logical closed loop link in signal modeling, abnormality identification, disturbance stripping and coupling judgment, innovatively combines multi-order energy trend identification and nonlinear frequency response modeling, effectively improves the early identification ability and diagnosis accuracy of structural abnormalities in rotating machinery running state, and provides a key algorithm basis for intelligent monitoring system.

[0066] Embodiment 3: please refer to Figure 1 , specifically: by collecting the running environment state of the rotating machinery, a disturbance index vector is obtained, specifically: , wherein, is the time series of the disturbance index vector, is the time series of the power fluctuation rate, is the time series of the speed, and are respectively the time series of the spectral expansion factor on the 2nd order harmonic and the vth order harmonic;

[0067] The spectral expansion factor on the corresponding order harmonic refers to the modal disturbance index of the corresponding order harmonic frequency point at each time, which reflects the widening degree of a certain frequency peak in the spectrum, and its specific acquisition method is: spectral expansion factor = center frequency of corresponding order harmonic Corresponding to the half width of the spectral peak of the order harmonic;

[0068] When the spectral peak is very sharp and concentrated, the half width of the spectral peak of the corresponding order harmonic is smaller, at this time, the spectral expansion factor on the corresponding order harmonic will be larger, indicating that the modal disturbance is small, and the system is stable, otherwise, it indicates that the modal disturbance is enhanced, which may be structural coupling or fault triggering.

[0069] If the half-height width of the spectral peak on the corresponding harmonic is significantly widened, it indicates that the frequency energy is divergent, which may be caused by the following two reasons:

[0070] 1. Cross-modal excitation: multiple adjacent modes are activated simultaneously under vibration coupling, causing energy to be dispersed in multiple frequency points.

[0071] 2. Double peak merging and nonlinear broadening: nonlinear coupling, looseness or crack evolution, etc., can cause the peak shape to change from single peak to double peak or wide peak.

[0072] The center frequency on the corresponding harmonic, i.e. the frequency position of the spectral peak;

[0073] The half-height width of the spectral peak on the corresponding harmonic refers to the frequency width corresponding to half the maximum amplitude of the spectral peak. The steps for obtaining it are as follows: first, find the maximum amplitude corresponding to the spectral peak, then calculate half the height, then find the two frequency points that intersect with the maximum amplitude corresponding to the spectral peak to the left and right, and finally obtain the half-height width of the spectral peak on the corresponding harmonic by subtracting the two frequency points.

[0074] Compare each component in the time series of the disturbance index vector with its corresponding preset disturbance threshold. If there is a component that exceeds the corresponding preset disturbance threshold, it is considered that the corresponding time segment is disturbed, otherwise it is considered that the corresponding time segment is not disturbed.

[0075] After constructing the high-order harmonic energy and extracting the energy change rate sequence, the rotating machinery vibration signal not only contains the real fault excitation characteristics caused by structural defects such as bearing wear, eccentricity, cracks, etc., but is also inevitably disturbed by background working condition disturbances such as load fluctuation, instantaneous speed change, external environment induced modal coupling, etc. These disturbance sources may induce transient drift of energy, leading to misdiagnosis or missed diagnosis. In order to achieve high reliability of fault recognition, these disturbances must be identified and removed from the target feature vector through systematic modeling methods.

[0076] If the time segment affected by the disturbance is located in the structural anomaly candidate set, it is removed, the structural anomaly candidate set is updated, and the energy change rate of the corresponding harmonic frequency point in the time segment corresponding to the structural anomaly candidate set is extracted to generate a first drift vector set.

[0077] From the historical running data, the energy change rate of the multi-order harmonic frequency point under healthy working condition is screened out to generate a normal drift vector set, and the normal drift vector set and the first drift vector set are aligned in the same time range to analyze the trend consistency of the normal drift vector set and the first drift vector set, and the difference sequence is obtained, which is specifically represented as: , wherein, is the difference value, which represents the deviation of the current running state from the normal state, is a first set of drift vectors, representing the energy rate of change of the n-th harmonic frequency point currently monitored, is a normal set of drift vectors, representing the reference energy rate of change of the same order under normal state;

[0078] To achieve accurate positioning of fault causes, it is necessary to trace back to the time domain signal. The original frequency disturbance behavior corresponding to the time window in the original signal is extracted;

[0079] The historical operation data refers to the operation state of the rotating machinery in the historical period;

[0080] Aligning the normal set of drift vectors and the first set of drift vectors in the same time range means extracting corresponding vector segments of the same length, the same time interval, and the same harmonic frequency point. The purpose is to compare the deviation between the current state and the reference state at each time to identify abnormal trends;

[0081] By using the ADF test method, it is tested whether the difference sequence has a unit root (i.e. whether it is non-stationary). If the test result rejects the null hypothesis, it means that the difference sequence as a whole is in a stationary state and does not have a unit root. Otherwise, it is determined that the rotating machinery in the corresponding time window has entered a cointegration destruction state, and the time window where the cointegration destruction occurs is marked as a structural deviation section.

[0082] During the operation of the rotating machinery, structural damage often does not instantly trigger significant vibration abnormalities, but gradually affects the frequency domain response characteristics of the system through long-term weak changes. Especially for high-order harmonic energy, its change usually manifests as continuous energy center drift, local bandwidth expansion, or multi-frequency synchronous excitation. To identify these "non-mutant" evolutionary trends, it is difficult to accurately extract them by relying solely on traditional spectrum comparison or threshold judgment. Therefore, this step introduces a local time cointegration analysis method to identify whether the long-term stable relationship between each order of high-order harmonic energy sequence and the historical reference state has been destroyed.

[0083] ADF test method is a classic statistical test method for testing whether a time series is a stationary sequence (i.e. whether the mean and variance remain unchanged over time). It is mainly used to determine whether a sequence has a unit root, thereby determining whether it has features such as trend drift and unstable fluctuations.

[0084] Unit root represents the long-term unstable random trend of a time series, which is generally an indicator of abnormal state;

[0085] Healthy working condition refers to the high-order harmonic drift characteristics in the time segment not affected by disturbance;

[0086] Cointegration destruction state refers to a state indicating that the system may enter a structural change state, and the backtracking mechanism must be started;

[0087] In this embodiment, by collecting the operating environment state parameters of rotating machinery, such as power fluctuation rate, speed change rate and spectrum spread factor of multi-order harmonics, a disturbance index vector is constructed and compared with the preset disturbance threshold corresponding to each component. This mechanism can distinguish whether there is interference behavior from external excitation sources such as load variation, power fluctuation, controller intervention, etc.

[0088] The spectrum spread factor is used to measure the spread degree of the spectrum near a certain harmonic point, which can identify the spectrum blur or peak broadening caused by non-structural excitation;

[0089] The power fluctuation rate is used to reflect the instantaneous disturbance intensity of the driving power; the speed is used to correct the influence of working condition change on harmonic drift. If any component in the disturbance index vector of a certain time segment exceeds the set threshold, it is determined that there is external disturbance in this time segment, and further judgment is made whether this segment belongs to the structural abnormality candidate set at the same time. If not, it is excluded, thereby effectively avoiding the confusion between external disturbance and real structural evolution.

[0090] By healthy comparison difference and cointegration test, the credible discrimination ability for structural deterioration is significantly improved. After completing the disturbance segment rejection, the system selects the harmonic energy change rate sequence confirmed as healthy working condition from the historical operating data to form a normal drift vector set. The first drift vector set obtained by current monitoring is time-aligned with the normal drift vector set, and a difference sequence is constructed;

[0091] The difference value is used to quantify the deviation degree of the current operating state relative to the healthy state;

[0092] Trend consistency determines whether the current state deviates from the normal structural response trajectory;

[0093] ADF test method is used to judge whether the time series is stable. If there is a unit root, it means that there is an irreversible trend in the sequence, i.e. there may be structural abnormalities. If the difference sequence is found to be a non-stationary sequence (failure to reject the original hypothesis) by ADF test, it means that the current monitoring data has produced an uncoordinated trend damage compared with the healthy benchmark, and it is determined that the current system may have entered the structural abnormality evolution stage, and the time segment is marked as a structural deviation segment.

[0094] Taking the main bearing monitoring of a wind turbine as an example, during the strong wind mutation period, the system power fluctuates sharply, and the blade speed is temporarily unstable due to wind shear effect, resulting in multi-order high-frequency harmonic drift in the system vibration spectrum. At this time, if the disturbance causes are not distinguished, it is easy to misjudge the spectrum disturbance as structural crack evolution.

[0095] In summary, the method constructs a structural deviation discrimination mechanism with clear structure, sensitive response and strong engineering applicability by introducing disturbance index comparison, health reference alignment and cointegration stability test, effectively improving the accuracy, forward-looking and adaptability of rotating machinery fault identification.

[0096] Embodiment 4: Please refer to Figure 1 Specifically, when the structural deviation section is identified, a backtracking mechanism is started, the structural deviation section is mapped into the collection time period of the continuous time sequence vibration signal using the sliding window mapping relationship, and the vibration signal corresponding to the structural deviation section is intercepted as the original signal subsequence;

[0097] The drift rate trajectories of the peak frequencies of each non-integer multiple frequency in the collection time period corresponding to the original signal subsequence are extracted, and a second drift vector set is constructed in combination with the first drift vector set in the structural deviation section.

[0098] The first drift vector set and the second drift vector set are projected into the same dimensional representation space by using the self-encoder to obtain the first drift vector set and the second drift vector set in the same dimensional representation space.

[0099] By inputting the first drift vector set and the second drift vector set in the same dimensional representation space into the dilated residual network, after the dilated convolution layer, the residual connection structure, the local attention layer and the full connection layer, the risk assessment value is output, specifically: wherein, is the risk assessment value of the corresponding time, represents a nonlinear function of the current system trend state change, and the value range is (0, 1), the closer to 1 indicates the more intense the trend, is an activation function, is the weight matrix of the full connection layer, which plays a weighted sum role to compress the multi-dimensional residual features into a single numerical output, which is obtained by network training and learning, is a residual feature extraction block, and are the first drift vector set and the second drift vector set in the same dimensional representation space, respectively, is a bias term, which constitutes an affine transformation with the weight matrix of the full connection layer to form the final full connection layer.

[0100] Wherein, the dilated residual network includes:

[0101] The dilated convolution layer is used to extract the cross-scale trend change;

[0102] The residual connection structure is used to avoid gradient disappearance and learn trend mutation;

[0103] The local attention layer is used to enhance important trend perception;

[0104] The full connection layer is to obtain the risk assessment value.

[0105] The residual feature extraction block is a structural functional module in the dilated residual network DRN, responsible for nonlinear feature extraction and residual signal reservation of the input feature vector (such as the first or second drift vector). The specific acquisition steps are as follows:

[0106] Step 1: Design the residual feature extraction block structure: when building the neural network architecture, the developer explicitly defines a residual block module (code level): composed of dilated convolution + activation + BatchNorm + residual connection, and the example framework can use PyTorch to define the model structure;

[0107] Step 2: input the drift vector into the structure and perform forward propagation (feature extraction): the input of the network is the preprocessed and projected first drift vector and second drift vector. These vectors pass through the residual feature extraction block and output a new feature, which contains local changes (such as nonlinear drift), multi-scale time patterns (modeled by dilated convolution), and a straight-through path that preserves the input feature (residual connection);

[0108] Step 3: In the training phase, all parameters in the residual block are optimized through error backpropagation: the convolution weights, biases, and normalization factors in the residual feature extractor RFE are trainable parameters. The system defines a loss function (such as the error between the final risk assessment value and the labeled target), then uses an optimizer (such as Adam, SGD) to continuously iterate and update these parameters until the network converges. The residual feature extraction block is trained, i.e., the parameters are learned.

[0109] The risk assessment value is the structural trend evolution score output by the dilated residual network, which is a nonlinear structural score of the trend residual between the potential state expression and the The risk assessment value is the structural trend evolution score output by the dilated residual network, which is a nonlinear structural score of the trend residual between the potential state expression and the The risk assessment value is the structural trend evolution score output by the dilated residual network, which is a nonlinear structural score of the trend residual between the potential state expression and the

[0110] In the structural deviation section, the evaluation difference is calculated in sequence, specifically: wherein, is the evaluation difference, is the risk assessment value at time t in the structural deviation section, is the risk assessment value corresponding to the time after time t;

[0111] extracting a maximum evaluation difference, comparing the maximum evaluation difference with a preset evaluation threshold, if the maximum evaluation difference exceeds the evaluation threshold, judging that the current rotating machinery has a risk of failure and triggering a failure instruction uploaded to an operation background to stop running;

[0112] If the maximum evaluation difference does not exceed the evaluation threshold, it is judged that the current rotating machinery does not have a risk of failure at present, and the running of the rotating machinery is continued.

[0113] In this embodiment, when the system identifies a structural deviation section by means such as cointegration test, the method automatically starts a trend backtracking mechanism, maps the deviation section back to the corresponding original time domain data window, restores the source vibration behavior that may induce the abnormality through sliding window mapping, called original signal subsequence, and the logical meaning of this step is to not only rely on surface frequency domain abnormal index, but also track the real starting point of structural abnormality through original physical signal backtracking, to ensure the interpretability and accuracy of diagnosis.

[0114] After obtaining the original signal subsequence, the system extracts the non-integer multiple frequency wave peak drift rate trajectory in it, which reflects the nonlinear response or excitation characteristics in the system, such as looseness, friction, and staggered harmonics. At the same time, combined with the existing first drift vector set (representing structural abnormal characteristics) of this section, a second drift vector set is formed as a higher dimensional and more complete abnormal behavior representation set.

[0115] The first drift vector set is based on the initial abnormal segment and represents structural trend drift; the second drift vector set is combined with nonlinear frequency spectrum behavior and represents deep dynamic evolution; the non-integer multiple frequency drift trajectory refers to the moving trajectory of subharmonic frequency affected by nonlinear factors in the system over time;

[0116] Since the normal drift vector set and the first drift vector set differ in dimension or physical properties, in order to make a reasonable comparison, the method projects the two drift vector sets into the same dimensional latent representation space through the self-encoder coding mechanism, so that they can be effectively trend-aligned and semantically fused. Subsequently, the system extracts the feature vector of the unified representation space as the input of the neural network. 、 As input, it is sent to the dilated residual network DRN for nonlinear structural residual learning, and outputs a risk evaluation value R(t), which indicates whether the current system has a trend structural abnormality at time t. The closer to 1, the more likely the system is on the path of deterioration and evolution.

[0117] For example, in a wind turbine, when a bearing appears early peeling, the first drift vector set may only show a slight high-order harmonic energy disturbance, while the second drift vector set in the backtracking signal appears non-integer multiple frequency peak drift. In the representation space, the two are mapped as 、 If the R(t) output by the inflation residual network DRN is 0.89, it indicates that the system has strong trend structure deviation, and the fault early warning response mechanism should be immediately introduced. The application further introduces an evaluation difference calculation mechanism in the structural deviation section, that is, comparing the risk score difference at different time points, extracting the maximum difference, and comparing it with the preset threshold. This mechanism avoids excessive alarm due to short-term fluctuations, and can also quickly respond to sudden deterioration, ensuring the safety of the equipment.

[0118] Part of the data in the above formula is dimensionless for numerical calculation, and the contents not described in detail in the specification all belong to the prior art known to those skilled in the art.

[0119] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made thereto without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for rotating machinery fault diagnosis based on dilated residual network, characterized in that: The method comprises the following steps: Collecting continuous time series vibration signals of the rotating machinery by using an acceleration sensor, obtaining a high-order harmonic energy sequence set after fast Fourier transform, and determining abnormal trend segments; Based on the abnormal trend segments and the running environment state of the rotating machinery, a structural abnormality candidate set is obtained and updated, and a first drift vector set is generated according to the updated structural abnormality candidate set; According to the first drift vector set and historical running data, a structural deviation section is identified, the structural deviation section is traced back to the original time domain signal to establish a second drift vector set; The first drift vector set and the second drift vector set are input into the inflation residual network to analyze whether the rotating machinery is evolving into a fault state; Collecting continuous time series vibration signals of the rotating machinery by using an acceleration sensor, and setting a window length, the continuous time series vibration signals are divided into multiple groups of time segment signals in a sliding window manner; Each group of time segment signals is subjected to fast Fourier transform to obtain corresponding frequency spectrum; The corresponding fundamental frequency is extracted from the spectrum, and the amplitude of each harmonic frequency point is calculated to form a high-order harmonic energy sequence set, which is specifically represented as: wherein H is the high-order harmonic energy sequence set; , and are the energy time sequences on the 2nd harmonic, the 3rd harmonic, and the vth harmonic, respectively, and v is the highest order harmonic set. According to the energy time sequence on each order harmonic, the energy change rate of each order harmonic frequency point is calculated by the first-order difference method to analyze the dynamic change speed of each order harmonic frequency point, specifically: wherein, is the energy change rate of the mth order harmonic frequency point, is the energy time sequence of the mth order harmonic. Based on the energy change rate, the energy change rate sequence on each order harmonic is obtained, and when the energy change rate of the corresponding order harmonic frequency point in the continuous time segment exceeds the preset change threshold, it indicates that there is an abnormal trend segment, and the abnormal trend set of the energy time sequence on each order harmonic is obtained through statistics.

2. The rotating machinery fault diagnosis method based on the inflation residual network according to claim 1, characterized in that: According to the abnormal trend set of the energy time series on each order harmonic, the synchronous drift trend between the abnormal trend segments on each order harmonic is judged. If the condition is met: then it is judged that the energy time series on each order harmonic is in homologous coupling, and is included in the structural abnormal candidate set, otherwise it is not in homologous coupling, and is not included in the structural abnormal candidate set; wherein, is the energy time series on the nth order harmonic, is a proportional factor.

3. The rotating machinery fault diagnosis method based on the inflation residual network according to claim 2, characterized in that: According to each order harmonic frequency point, a non-integer multiple wave peak frequency in the corresponding frequency spectrum is identified, and a drift rate trajectory of each non-integer multiple wave peak frequency is obtained, which is specifically represented as: wherein, is the drift speed of the kth sub-harmonic frequency point in time, is the drift amount of the kth sub-harmonic frequency point, is the time span in which the frequency change occurs, and k is the sub-harmonic frequency point number.

4. The rotating machinery fault diagnosis method based on the inflation residual network according to claim 3, characterized in that: The disturbance index vector is obtained by collecting the running environment state of the rotating machinery, specifically: wherein, is a time sequence of the disturbance index vector, is a time sequence of the power fluctuation rate, is a time sequence of the rotating speed, and are respectively a time sequence of the spectrum expansion factor on the 2nd harmonic and a time sequence of the spectrum expansion factor on the vth harmonic. Comparing each component in the time sequence of the disturbance index vector with the corresponding preset disturbance threshold, if there is a component exceeding the corresponding preset disturbance threshold, it is considered that the corresponding time segment is disturbed, otherwise it is considered that the corresponding time segment is not disturbed; If the disturbed time segment is located in the structural abnormality candidate set, it is removed, the structural abnormality candidate set is updated, and the energy change rate of the corresponding order harmonic frequency point in the time segment corresponding to the structural abnormality candidate set is extracted to generate the first drift vector set.

5. The rotating machinery fault diagnosis method based on the inflation residual network according to claim 4, characterized in that: The energy change rate of the multi-order harmonic frequency points under the healthy working condition is screened from the historical operation data to generate a normal drift vector set, and the normal drift vector set and the first drift vector set are aligned in the same time range to analyze the trend consistency of the normal drift vector set and the first drift vector set, and a difference sequence is obtained, which is specifically shown as: wherein, is a difference value, is the first drift vector set, is the normal drift vector set. ADF test method is used to test whether the difference sequence has a unit root, if the test result is to reject the null hypothesis, it means that the difference sequence is in a stationary state as a whole and does not have a unit root, otherwise it is determined that the rotating machinery in the corresponding time window is in a cointegration destruction state, and the time window where the cointegration destruction occurs is marked as a structural deviation section.

6. The rotating machinery fault diagnosis method based on the inflation residual network according to claim 5, characterized in that: When the structural deviation section is identified, a tracing mechanism is started, the structural deviation section is mapped to the collection time period of the continuous time series vibration signals by using the sliding window mapping relationship, and the vibration signal corresponding to the structural deviation section is intercepted as an original signal subsequence. The drift rate trajectories of the wave peak frequencies of each non-integer multiple frequency in the acquisition time period corresponding to the original signal subsequence are extracted, and a second drift vector set is constructed in combination with the first drift vector set in the structural offset section.

7. The rotating machinery fault diagnosis method based on the dilated residual network according to claim 6, characterized in that: The first drift vector set and the second drift vector set are projected into the same dimensional representation space by using the autoencoder to obtain the first drift vector set and the second drift vector set in the same dimensional representation space. The first drift vector set and the second drift vector set in the same dimension representation space are input into the dilated residual network, and after the dilated convolution layer, the residual connection structure, the local attention layer and the full connection layer, a risk evaluation value is output, specifically: wherein, is the risk evaluation value of the corresponding time, is an activation function, is a weight matrix of the full connection layer, is a residual feature extraction block, and respectively are the first drift vector set and the second drift vector set in the same dimension representation space, is a bias term.

8. The rotating machinery fault diagnosis method based on the dilated residual network according to claim 7, characterized in that: In the structural offset section, the evaluation difference is calculated in sequence, specifically: wherein, is the evaluation difference, is the risk evaluation value at time t in the structural offset section, is the risk evaluation value corresponding to the time after time t; The maximum evaluation difference is extracted, and the maximum evaluation difference is compared with a preset evaluation threshold value. If the maximum evaluation difference exceeds the evaluation threshold value, it is judged that the current rotating machinery has a fault risk and a fault instruction is uploaded to an operation background to perform shutdown processing.

Images