Fault identification analysis method based on multi-source signal analysis of hydro-generator

Abstract

The application belongs to the technical field of equipment fault identification, and specifically discloses a fault identification analysis method based on water turbine generator multi-source signal analysis, which comprises the following steps: firstly, multi-site signal acquisition and synchronization are performed, and suspected abnormal signals are screened; then, based on a fault propagation chain, a dual correlation representation analysis of fault time and trend is performed by analyzing the position of the signal in the chain, so that real fault signals are screened; when there are multiple chains, the fault dominant strength, signal density and abnormal degree of each chain are comprehensively evaluated to determine the dominant fault propagation chain; finally, the dominant chain and the abnormal degree of the fault signal thereof constitute an evidence body, which is matched with a historical fault template in terms of similarity, and the final fault type is output. The application realizes accurate tracing and diagnosis of faults, effectively eliminates isolated false alarms, thereby improving the accuracy of fault diagnosis and providing a clear direction for subsequent fault tracing and fault handling.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of equipment fault identification, and in particular, relates to a fault identification analysis method based on water turbine generator multi-source signal analysis. BACKGROUND

[0002] As the core equipment of a hydropower station, a water turbine generator is susceptible to bearing wear, rotor imbalance, stator winding short circuit and other faults due to the coupling of multiple factors such as water power, machinery and electricity during operation. Such faults not only may cause shutdown and huge economic losses, but also may cause safety accidents, thus highlighting the importance and necessity of fault identification.

[0003] The prior art 1, such as the Chinese patent application for invention with the application number 202510015766.5, discloses a water turbine abnormality detection and fault diagnosis method based on multi-sensor data fusion. The method generates a comprehensive abnormality score by weighting the results of three unsupervised algorithms, i.e. clustering, local outlier factor and autoencoder, and finally performs diagnosis by comparing with historical fault templates.

[0004] The prior art 2, such as the Chinese patent application for invention with the application number 202411326852.X, discloses a multi-strategy fusion water turbine generator set fault diagnosis method. The method constructs a comprehensive model by combining a GRU-based neural network time series model and a random forest decision model based on industrial internet platform data, fully utilizes the advantages of different algorithms, avoids the limitations of single algorithm for fault judgment, improves the accuracy of fault diagnosis, and realizes efficient fault classification.

[0005] Since there are multiple representations of equipment faults, when a fault occurs, multiple sensor alarm signals are often received, and these signals often have correlations in time and development trends. The prior art 1 and the prior art 2 do not deeply mine the correlation of different sensor signals in the time dimension and the consistency in the development trend, resulting in that the diagnosis result stays at the level of abnormality, the discrimination ability for whether the abnormal signal is transient interference is insufficient, and at the same time, the abnormality cannot be traced back, which is not convenient for subsequent abnormality processing.

[0006] Secondly, the performance of the comprehensive model constructed by the prior art 2 depends on a large amount of, complete and accurately labeled historical fault data, which limits the universality of the method. At the same time, this method assumes that the input data is reliable and synchronous. However, in the field environment, problems such as sensor drift, communication delay and data loss frequently occur, and the authenticity of the data and the authenticity of the fault are difficult to guarantee, which affects the accuracy of the final judgment result.

[0007] In addition, the prior art lacks a comprehensive analysis and comparison mechanism for multiple potential fault propagation paths when multiple sensors alarm at the same time. When a complex coupled fault occurs in a hydroelectric generator, abnormal signals are often distributed on multiple propagation chains with different correlations. The prior art is difficult to identify the dominant fault chain from these concurrent abnormal paths, resulting in unclear fault tracing direction, affecting the efficiency and accuracy of maintenance decision-making. SUMMARY

[0008] In view of this, in order to solve the above problems, a fault identification analysis method based on multi-source signal analysis of a hydroelectric generator is proposed.

[0009] The object of the present application can be achieved by the following technical solutions: The present application provides a fault identification analysis method based on multi-source signal analysis of a hydroelectric generator, which comprises: collecting multi-site signal data of the hydroelectric generator and unifying the time stamps of each signal, and screening signals that continuously deviate from the normal range as suspected abnormal signals.

[0010] For all suspected abnormal signals on the same fault propagation chain, determine the fault time correlation representation and the fault trend correlation representation according to their positions in the chain, and determine the fault signal based on the correlation representation.

[0011] Extract all fault signals, if distributed in multiple fault propagation chains, determine the dominant fault propagation chain based on the fault dominant strength of the chain, the fault signal density and the abnormality degree of the fault signal. Otherwise, the chain where the fault signal is located is taken as the dominant fault propagation chain.

[0012] The abnormality degree of each fault signal in the dominant fault propagation chain and the abnormality degree of each fault signal in the dominant fault propagation chain are combined to form an evidence body, and the cosine similarity between the evidence body and each historical fault category feature template is calculated. The historical fault category with the highest cosine similarity is taken as the final detected fault type.

[0013] Compared with the prior art, the present application has the following advantages: (1) The present application solves the problem that the prior art cannot realize abnormal tracing by establishing a fault propagation chain, effectively reduces the dependence on the number and quality of historical fault data samples, realizes the conversion from discrete alarm to continuous fault path, and thus can clearly depict the potential development path of the abnormality in the equipment, facilitating subsequent abnormal processing.

[0014] (2) The present application determines the fault signal by double verification of the fault time correlation representation and the fault trend correlation representation, effectively eliminates isolated abnormal signals caused by accidental false alarms or external interference, can screen out real fault sources with consistency in time and space dimensions, and thus improves the credibility and anti-interference ability of fault source identification.

[0015] (3) The application can automatically identify abnormal signal misplacement problems caused by communication delay through lag time verification in fault time correlation representation analysis. Through time delay cross-correlation and maximum information coefficient calculation in fault trend correlation representation analysis, it can effectively filter pseudo-abnormal signals caused by sensor drift. Further, it can ensure high diagnostic accuracy and reliability in the working condition of scarce historical fault data, noise and asynchronous problems in the field data, effectively improving the universality and robustness of fault identification.

[0016] (4) In the presence of multiple potential fault paths, the application determines the dominant fault propagation chain by fusing the fault dominant strength, fault signal density and fault abnormality of each chain, determines the fault tracing direction, realizes the objective evaluation and comparison of the influence of multiple chains, and improves the diagnostic efficiency and accuracy of complex faults, further improving the efficiency and accuracy of subsequent maintenance decisions.

[0017] (5) The application reduces the dependence on the accuracy of single sensor data by forming an evidence body composed of the dominant fault propagation chain and the abnormality degree of each fault source inside it, and performing similarity matching with the historical fault template, and enhances the stability of the diagnostic results. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 The overall implementation flowchart of the application.

[0020] Figure 2 The overall implementation flowchart of the application.

[0021] Figure 3 The local implementation flowchart of the application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only some embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0023] Please refer to Figure 1As shown, the present application provides a fault identification analysis method based on multi-source signal analysis of a hydroelectric generator, which comprises: S1, collecting multi-site signal data of the hydroelectric generator, and unifying the time stamps of each signal, and screening signals continuously deviating from the normal range as suspected abnormal signals.

[0024] Considering that the fault occurrence and development process of the hydroelectric generator has typical spatio-temporal correlation characteristics. Single-site signal monitoring cannot fully reflect the equipment state, and multiple time-asynchronous signals will directly lead to fault timing relationship confusion and cause-effect relationship misjudgment, therefore, when collecting multi-site signal data of the hydroelectric generator, the time stamp needs to be unified.

[0025] Specifically, the multi-site includes but is not limited to bearing, blade and generator winding sites in a specific embodiment, and the signals include temperature signal, pressure signal, vibration signal, voltage signal and current signal, etc.

[0026] It should be noted that in the process of screening signals continuously deviating from the normal range, the duration of the signal deviating from the normal range needs to exceed the normal fluctuation duration threshold value marked for the signal in the factory setting of the hydroelectric generator, so as to distinguish the real fault precursor signal from the transient disturbance fluctuation.

[0027] For example, the vibration signal needs to exceed the threshold value continuously for 3 sampling periods, and the temperature signal needs to deviate continuously for 5 sampling periods, so as to be judged as a suspected abnormal signal. By introducing the continuity requirement, misjudgment caused by transient disturbance can be effectively avoided, and the reliability of the corresponding reference signal data for subsequent fault analysis is ensured.

[0028] It should also be noted that the normal range is mainly determined according to the factory setting of the equipment. Specifically, the rated value, allowable deviation value and alarm value of each signal in the technical specification provided by the manufacturer are taken as the reference, for temperature, pressure and other signals that deviate in one direction to indicate state deterioration or fault risk, the rated value is taken as the reference, and the alarm value set in the factory is taken as the upper limit (or lower limit) of the safety range to define the normal range boundary, and the reference and the upper limit (or lower limit) of the safety range form the normal range. For voltage, vibration and other signals that deviate in one direction to indicate state deterioration or fault risk, the rated value plus or minus the factory allowable deviation value is taken as the upper and lower limits of the normal range.

[0029] The application can ensure that the signal data from different acquisition units has a comparable time reference through unified timestamps, provide data consistency guarantee for subsequent fault time correlation characterization analysis and fault propagation timing inference, and can identify potential abnormalities at the initial stage of the hydroelectric generator fault, thereby providing a target signal set for subsequent fault tracing analysis. Moreover, the multi-site synchronous monitoring breaks the limitations of traditional single-point diagnosis, further improving the effectiveness of subsequent fault diagnosis.

[0030] S2, for all suspected abnormal signals on the same fault propagation chain, determining the fault time correlation characterization and the fault trend correlation characterization according to the position of the suspected abnormal signals in the chain, and determining the fault signal based on the correlation characterization.

[0031] It should be explained that the fault propagation chain is essentially a physical path representation of the transmission of fault energy among hydraulic, mechanical, electrical and other systems. In order to more clearly explain the construction logic and transmission rules of the fault propagation chain, the typical bearing fault propagation chain and the generator stator winding fault propagation chain in the hydroelectric generator will be described in detail.

[0032] Exemplarily, the construction and implementation process of the typical bearing fault propagation chain in the hydroelectric generator is as follows: when the thrust bearing part of the hydroelectric generator set occurs early wear and tear fault, the signal data of the following parts is collected through unified timestamps: signal A: thrust bearing seat radial vibration.

[0033] Signal B: thrust bearing tile temperature.

[0034] Signal C: lower guide bearing cooling water outlet temperature.

[0035] Signal D: lower guide bearing seat vibration.

[0036] After the data acquisition is completed, first, the signal A (vibration amplitude continuously exceeds the normal range) and the signal B (temperature value continuously exceeds the normal range) are identified as suspected abnormal signal sources. Since the abnormal vibration and temperature rise of the thrust bearing will be transmitted to the adjacent lower guide bearing through the shaft system, thereby causing the working condition of the adjacent lower guide bearing to deteriorate, which further causes the cooling water temperature of the lower guide bearing to rise and the vibration to intensify. Therefore, a fault propagation chain can be constructed: the thrust bearing fault (A, B) will cause the lower guide bearing to be affected (C, D), which accurately reflects the complete path of the fault transmission along the shaft system from the thrust bearing to the lower guide bearing.

[0037] Exemplarily, the construction and implementation process of the generator stator winding fault propagation chain includes: when the generator stator winding occurs local overheating or turn-to-turn short circuit, the signal data of the following parts is obtained through unified timestamp acquisition system, specifically including signals D, E and F.

[0038] Signal D: stator current harmonic content.

[0039] Signal E: stator core temperature.

[0040] Signal F: unit output power signal.

[0041] After the data acquisition is completed, the signals D (harmonic content continuously exceeds the normal range), E (temperature continuously exceeds the normal range) and F (power fluctuation continuously exceeds the normal range) are identified as suspected abnormal signal sources. Since the turn-to-turn short circuit fault of the stator winding first causes the magnetic field distribution distortion, when the distortion occurs, it shows that the specific harmonic content increases. When the specific harmonic content increases, it will in turn cause the electromagnetic loss to increase, and when the electromagnetic loss increases, it will cause the local temperature of the stator core to rise. Finally, the stability of the electromagnetic torque is affected, causing the unit power fluctuation to continuously exceed the normal range. Therefore, a fault propagation chain can be constructed: the stator winding fault first shows the current harmonic content anomaly of signal D, the electrical anomaly causes the stator core temperature anomaly of signal E through the increase of electromagnetic loss, and finally reflects the unit output power anomaly of signal F.

[0042] It can be known from the above fault propagation chain construction example that the fault propagation chain described in the application can accurately reflect the transmission law of the fault between different systems, and provides a reliable information foundation for subsequent fault tracing and diagnosis analysis.

[0043] The application solves the problem that the prior art cannot realize abnormal tracing, effectively reduces the dependence on the number and quality of historical fault data samples, realizes the conversion from discrete alarm to continuous fault path, and thus can clearly depict the potential development path of the abnormality in the equipment, facilitating subsequent abnormality processing.

[0044] It should be noted that the signal data in the fault propagation chain is sorted according to the structural connection relationship, physical connection relationship or energy flow / material flow transmission path of the part in the device.

[0045] Since the signal abnormalities at different positions on the same fault propagation chain may exist in the order and trend change law during the operation of the hydroelectric generator. Therefore, the application analyzes the relative position relationship of the signals in the fault propagation chain to perform subsequent fault time correlation characterization and fault trend correlation characterization analysis.

[0046] In a specific embodiment, please refer to Figure 2 According to its position in the chain, the determination of the fault time correlation characterization includes: the statistical implementation process of the fault time correlation characterization is as follows: first, for any two suspected abnormal signals on the same fault propagation chain, determine their relative positions in the chain.

[0047] If the relative position is the adjacent position, the fault time correlation characterization statistics in the adjacent position is performed, if the relative position is the non-adjacent position, the fault time correlation characterization statistics in the non-adjacent position is performed, and finally, based on the fault time correlation characterization of the two suspected abnormal signals in the adjacent position and the non-adjacent position, the fault time correlation characterization between the suspected abnormal signals is output.

[0048] Preferably, referring to (a) in the description, Figure 3 The specific implementation process of performing the fault time correlation characterization statistics in the adjacent position includes: first, calculating the interval time length between the abnormal starting points of the two suspected abnormal signals, denoted as the lag time length. Then, it is judged whether the lag time length is within the preset delay range. If the lag time length is within the preset delay range, the fault time correlation characterization of the two suspected abnormal signals is output as strong correlation, and when the lag time length exceeds the preset delay range, the fault time correlation characterization of the two suspected abnormal signals is output as weak correlation.

[0049] It should be noted that considering that when the vibration signal propagates in the metal structure, the damping characteristics inside the material will cause energy dissipation and speed change, and the reflection and refraction phenomena of the structure interface will change the propagation path, and then change the propagation time. And other signals such as temperature will change with the change of working conditions, and then cause the transmission time to fluctuate within a certain range. Therefore, the lag time length needs to be judged.

[0050] In one specific embodiment, the preset delay range is calculated based on the equipment size of the specific hydroelectric generator unit.

[0051] Exemplarily, taking the vibration signal and the temperature signal as an example, a specific construction implementation example of the preset reference lag time range is performed: 1) vibration signal: the length of the shaft section between the thrust bearing and the lower guide bearing is obtained from the equipment size, and the longitudinal wave propagation speed of the steel material is determined based on the material manual. The ratio of the shaft section length to the longitudinal wave propagation speed is taken as the theoretical delay time of the vibration signal.

[0052] Subsequently, the historical vibration propagation time sequence of the hydroelectric generator in the normal operating state is called, and for each measured propagation time in the sequence, the deviation of the measured propagation time from the theoretical delay time is calculated. All these deviation values are divided into two groups according to positive and negative, and the average value of the positive deviation group ( ) and the average value of the negative deviation group ( ) are calculated respectively. The sum of the theoretical delay time and is taken as the lower limit of the preset range, and the sum of the theoretical delay time and The preset reference lag time range of the vibration signal is outputted, and when the measured lag time is within the range, it is determined that the correlation is strong.

[0053] 2) Temperature signal: the oil path length from the thrust bearing to the lower guide bearing is obtained from the equipment size, and the oil reference flow rate and oil density are extracted from the equipment instruction manual, and the convective heat transfer coefficient and specific heat capacity of the oil are determined based on the material manual, and the ratio of the oil path length and the oil reference flow rate is taken as the flow delay time, and then the heat transfer delay time is calculated according to the heat balance equation , , wherein is the oil density, is the specific heat capacity of the oil, is the oil reference flow rate, is the convective heat transfer coefficient of the oil, is the heat transfer area, and the combination of the flow delay time and the heat transfer delay time is taken as the theoretical delay time of the temperature signal.

[0054] The historical temperature transfer time sequence under the normal operating state is called, and the deviation of each measured value from the theoretical value is calculated. Similarly, all these deviation values are divided into two groups according to positive and negative, the average values of each group are calculated, and the preset reference lag time range of the temperature signal is determined in the same way as the determination of the preset reference lag time range of the vibration signal, and when the measured lag time is within the range, it is determined that the correlation is strong.

[0055] Further preferably, referring to (b) shown in Figure 3 The specific implementation process of the fault time correlation representation statistics in the non-adjacent positions includes: identifying all adjacent signal pairs between the two suspected abnormal signals, and for each adjacent signal pair, the fault time correlation representation of the adjacent signal is statistically determined based on the lag time of the adjacent signal according to the statistical method of the fault time correlation representation in the adjacent position.

[0056] If the fault time correlation representations of all adjacent signal pairs are strong correlation, the fault time correlation representation of the two non-adjacent suspected abnormal signals is finally outputted as strong correlation. Otherwise, it is finally outputted as weak correlation.

[0057] It should be noted that for long fault propagation chains, parallel computing can be used to improve real-time performance, and the subsequent fault trend correlation representation determination implementation process is also processed in the same way.

[0058] For better understanding of the fault time correlation characterization of the present application, the present application takes the bearing fault propagation chain as an example to statistically demonstrate the fault time correlation characterization: first, based on the above collected signals A, B, C and D, assuming that signals A and D and signals B and C are two groups of suspected abnormal signals, signals A and D are taken as the first group of signal pairs, signals B and C are taken as the second group of signal pairs, and it is assumed that the preset delay range is 5 to 10 seconds.

[0059] Next, based on the construction process of the above bearing fault propagation chain, it can be known that the thrust bearing fault (A, B) will cause the lower guide bearing to be affected (C, D), that is, in this fault propagation chain, signal A and signal D are non-adjacent positions, signal B and signal C are adjacent positions, that is, the first group of signal pairs are non-adjacent positions, and the second group of signal pairs are adjacent positions.

[0060] For the second group of signal pairs, the starting time of the first deviation of signal B from the normal range is T1=14:20:15.000, and the starting time of the first deviation of signal C from the normal range is T2=14:20:22.000. The lag time of the two is 7 seconds. Since 7 seconds is within the preset range, the fault time correlation characterization of the two suspected abnormal signals (B, C) in the second group is strong correlation.

[0061] For the first group of signal pairs, since they are non-adjacent positions, all adjacent signal pairs (A-B, B-C, C-D) in the middle are identified. The fault time correlation characterization of all adjacent signals in the two groups is determined according to the determination method of a group. It is determined that the fault time characterization of A-B, B-C and C-D is strong correlation. Therefore, the fault time correlation characterization of the two suspected abnormal signals (A, D) in the first group is strong correlation.

[0062] In one specific embodiment, the specific implementation process of determining the fault trend correlation characterization includes: selecting any two signals from each suspected abnormal signal in the same fault propagation chain to form a target signal pair.

[0063] If the positions of the two signals in the fault propagation chain are adjacent, the data sequences of the two signals in the preset time window are extracted. If they are not adjacent, the data sequences of all adjacent signal pairs on the shortest fault propagation path connecting the two signals in the preset time window are extracted.

[0064] Based on the extracted data sequences, the time-lag cross-correlation index between the two signals and the maximum information coefficient between the data sequences of the two signal pairs are calculated. The time-lag cross-correlation index includes a time-lag correlation coefficient and a target time-lag value.

[0065] If the positions of the two signals are adjacent and simultaneously satisfy all the following conditions, it is judged that the fault trend correlation characterization is strong correlation. Otherwise, it is weak correlation: the time-lag correlation coefficient is higher than the first preset threshold.

[0066] The target time delay value is within a preset delay range.

[0067] The maximum information coefficient is higher than a second preset threshold.

[0068] For non-adjacent signals, if the fault trend correlation characteristics of all adjacent signal pairs on the path connecting them are strong correlation, it is judged that the fault trend correlation characteristics between the two non-adjacent signals are strong correlation. If there is any adjacent signal pair that is weak correlation, it is judged as weak correlation.

[0069] It can be understood that the first preset threshold and the second preset threshold can be determined according to historical data or experiments of the specific hydroelectric generating set.

[0070] The setting and implementation process of the first preset threshold is: the historical time delay correlation coefficient distribution of the two signals of the hydroelectric generating set in the normal operation state is counted, and the 95% quantile of the historical time delay correlation coefficient distribution is taken as the first preset threshold.

[0071] The setting and implementation process of the second preset threshold is: the historical maximum information distribution of the two signals of the hydroelectric generating set in the normal operation state is extracted, and the 99% quantile of the historical maximum information distribution is taken as the second preset threshold.

[0072] In view of the uncertain time delay of the fault signal in the actual propagation process, direct calculation of the signal correlation may underestimate the real correlation degree due to the misalignment of the time delay. Therefore, the time delay cross-correlation index is calculated to make up for the fault trend correlation characteristic determination error caused by the misalignment of the time delay.

[0073] In a preferred embodiment of the present application, the time delay cross-correlation index calculation implementation process includes: first, setting the time delay search range to -50ms to +50ms, with a sampling interval of 10ms as the step, then, within the preset time delay search range, with the data sampling interval as the step, the time offset between the corresponding data sequences of the two signals is changed to realign, such as starting from -50ms, increasing the time offset by 10ms each time, and realigning the corresponding data sequences of the two signals. Each time the time offset is increased, the current time delay position is recorded, and the Pearson correlation coefficient of the aligned data sequences of the two signals at each time delay position is calculated.

[0074] The maximum Pearson correlation coefficient is selected from the corresponding calculated Pearson correlation coefficients at each time delay position, and the maximum Pearson correlation coefficient is taken as the time delay correlation coefficient between the two signals, and then the time offset corresponding to the time delay correlation coefficient is taken as the target time delay value between the two signals.

[0075] It should be noted that the time delay search range and the sampling interval are only a preferred example of the present application, and can be set by the implementer in specific actual scenarios. The Pearson correlation coefficient is a known function, and its expression formula is not described again.

[0076] The present application can accurately capture the optimal correlation degree between signals and the corresponding time delay through time delay search and correlation coefficient calculation, thereby providing a quantitative basis for judging the fault propagation trend.

[0077] In view of the fact that there is often a complex nonlinear relationship between actual fault signals, the traditional correlation coefficient mainly reflects the linear relationship and is difficult to fully represent the dependence between signals. It is necessary to use a nonlinear statistical quantity to supplement the analysis, and the present application preferably selects the maximum information coefficient as the nonlinear statistical quantity.

[0078] In a preferred embodiment of the present application, the calculation process of the maximum information coefficient includes: constructing a scatter plot of the data sequence, and performing grid division in the X and Y axis directions with different densities, for example, dividing into 2 rows and 2 columns, 2 rows and 3 columns, 3 rows and 2 columns, etc., until a maximum grid number is reached according to the size of the data amount, and the density range of grid division can be preferably exemplified as 10x10 to 30x30.

[0079] For each grid division method, the data points are assigned to the corresponding grid cells, and the mutual information value of the two signals under the grid division is calculated according to the distribution of the data points in each cell. The mutual information value is subjected to minimum-maximum normalization processing.

[0080] All grid division methods are traversed, and the normalized mutual information value obtained at each step is recorded. The maximum value is selected from all recorded normalized mutual information values as the maximum information coefficient.

[0081] It should be noted that the calculation of the mutual information value is a known technology, and its specific calculation process is not described again. By comparing the calculated maximum information coefficient with the second preset threshold, data services are provided for the fault trend correlation determination of the fault propagation chain development.

[0082] By calculating the maximum information coefficient, the present application can effectively capture various dependence relationships between signals, including nonlinear relationships, which complement the aforementioned linear time delay correlation coefficient, and together ensure the comprehensiveness and accuracy of the fault trend correlation trend judgment.

[0083] Secondly, the application can automatically identify abnormal signal misplacement problems caused by communication delay through lag time verification in fault time correlation characterization analysis. Through time lag cross-correlation and maximum information coefficient calculation in fault trend correlation characterization analysis, it can effectively filter pseudo abnormal signals caused by sensor drift. Further, it can ensure high diagnostic accuracy and reliability under the condition of scarce historical fault data, noise and asynchronous problems in field data, effectively improving the universality and robustness of fault identification.

[0084] In another specific embodiment, the determining fault signals comprises: for each suspected abnormal signal, extracting its fault time correlation characterization and fault trend correlation characterization with at least one other suspected abnormal signal in the same fault propagation chain.

[0085] If both the fault time correlation characterization and the fault trend correlation characterization of the suspected abnormal signal and the at least one other suspected abnormal signal satisfy strong correlation at the same time, the suspected abnormal signal is determined as a fault signal.

[0086] S3, extracting all fault signals, if distributed in multiple fault propagation chains, determining the dominant fault propagation chain based on the fault dominant strength, fault signal density and abnormal degree of the fault signal of the chain. Otherwise, the chain where the fault signal is located is taken as the dominant fault propagation chain.

[0087] In actual operation, multiple faults may occur at the same time or there may be multiple potential propagation paths. Based on this, the application considers whether it is distributed in multiple fault propagation chains, and when it is distributed in multiple fault propagation chains, the dominant fault propagation chain is determined from the aspects of fault dominant strength, fault signal density and fault abnormal degree.

[0088] In one specific embodiment, the specific implementation process of determining the dominant fault propagation chain comprises: for each fault propagation chain, traversing the remaining fault propagation chains to check whether there is a shared fault signal or a preset strong correlation signal with the current chain.

[0089] If there is, the correlation mark number of the current chain is increased by 1, the fault dominant strength of each fault propagation chain is counted based on the correlation mark number, and if there is not, the next fault propagation chain is traversed.

[0090] The number of fault signals identified on all fault propagation chains is divided by the total number of signal sources of the corresponding chain to obtain the fault signal density of the chain.

[0091] For each fault signal on the fault propagation chain, the absolute value of the relative deviation of the real-time data value from the corresponding median of the normal range is calculated as the abnormal degree, and the fault abnormal degree of the fault propagation chain is calculated based on the abnormal degree.

[0092] The fault dominant strength variation coefficient of the fault propagation chain is calculated by comparing the fault dominant strength of each fault propagation chain.

[0093] If the variation coefficient is less than or equal to a preset threshold, a linear weighted sum of the fault dominant strength, the fault signal density and the fault anomaly degree is taken as a final anomaly dominant score, and the fault propagation chain with the highest score is taken as a dominant fault propagation chain.

[0094] If the variation coefficient is greater than the preset threshold, the fault propagation chain with the maximum fault dominant strength is taken as the dominant fault propagation chain.

[0095] Further, since a single association marking number cannot distinguish between extensive influence and repeated association, the present application introduces a repeated association ratio and a marking ratio to perform fault dominant strength statistics, and the specific statistical process of the fault dominant strength of each fault propagation chain is as follows: for each fault propagation chain, the association marking number is counted, and all fault propagation chains producing the marking are recorded to form an association chain list of the chain.

[0096] For each chain in the association chain list, the repeated association marking number between the chain and the current chain is counted, and the repeated association marking number refers to the multiple association markings between the current chain and a certain association chain due to multiple shared fault sources or multiple strong association signals in the traversal inspection process.

[0097] For each fault propagation chain, the number of chains in the association chain list with a repeated association marking number of zero is counted, and the number is recorded as the effective association chain number.

[0098] The effective association chain number corresponding to each fault propagation chain is divided by the difference between the total number of fault propagation chains and 1 to obtain a repeated association ratio, and the repeated association ratio represents how many other chains in the network have a deep association with the current chain. The larger the value of the repeated association ratio, the more extensive the influence of the chain.

[0099] From the association marking numbers of all fault propagation chains, the maximum value is selected as the highest marking number.

[0100] The ratio of the association marking number of each fault propagation chain to the highest marking number is calculated and recorded as a marking ratio, and the marking ratio represents how large the total amount of association of the current chain with the entire network is. The larger the value of the marking ratio, the stronger the overall activity and influence of the chain.

[0101] For each fault propagation chain, the repeated association ratio and the marking ratio are multiplied to obtain the fault dominant strength.

[0102] It should be noted that two interference situations commonly exist in the fault propagation network, one is that a certain chain is repeatedly associated with only a few other chains, which is manifested as high association depth but insufficient breadth. The other is that a certain chain is only associated with a large number of other chains, which is manifested as large association breadth but insufficient depth. If a simple linear weighted sum is used, both of these two interference situations may cause the final result to be larger due to a higher score of a certain indicator. By introducing multiplication operation, only when the repeated association ratio and the marking ratio reach a high value at the same time, the dominant strength score of the chain will be correspondingly improved. Further, the fault propagation chain with both extensive influence and depth coupling in the fault propagation network can be screened out, thereby improving the accuracy and reliability of fault tracing analysis.

[0103] In the scene of the present application with multiple potential fault paths, the dominant fault propagation chain is determined by fusing the fault dominant strength, the fault signal density and the fault abnormality degree of each chain, thereby determining the fault tracing direction, realizing objective evaluation and comparison of the influence of multiple chains, and further improving the diagnosis efficiency and accuracy of complex faults, and further improving the efficiency and accuracy of subsequent maintenance decision.

[0104] Further, considering that the field data is filled with various transient interference and sensor false positives, if only individual signal is abnormal, it is likely to be an occasional fault or false positive, that is, the abnormality degree of a single fault signal cannot represent the abnormality level of the entire chain. Therefore, the present application determines the final fault abnormality degree of the entire chain based on the abnormality degree of the fault signal, and the calculation process of the fault abnormality degree of the fault propagation chain comprises: comparing the abnormality degrees of all fault signals, and taking the maximum value as the reference fault abnormality degree of the corresponding fault propagation chain.

[0105] For the same fault propagation chain, the average value of the abnormality degrees of all fault signals is counted, and the proportion of fault signals whose abnormality degrees exceed the average value is counted.

[0106] For each fault propagation chain, the fault signal proportion is multiplied by the reference fault abnormality degree to obtain the final fault abnormality degree.

[0107] The present application combines the reference abnormality degree with the proportion of fault signals exceeding the average value, which not only considers the influence of the most serious fault point, but also reflects the distribution breadth of the abnormality in the chain, and comprehensively reflects the overall abnormality severity of the fault propagation chain.

[0108] S4, the dominant fault propagation chain and the abnormality degree of each fault signal inside it are combined to form an evidence body, the cosine similarity between the evidence body and each historical fault category feature template is calculated, and the historical fault category with the highest cosine similarity is taken as the final detected fault type.

[0109] Specifically, the unique identifier of the determined dominant fault propagation chain and the abnormality degrees of all fault signals inside the chain are combined in a preset order to form a multi-dimensional vector, denoted as evidence body E, , denotes the identification of the dominant fault propagation chain, and can be represented by a specific number for subsequent analysis, denotes the abnormality degree of each fault signal in the dominant fault propagation chain, denotes the total number of fault signals.

[0110] According to the historical fault case library, a corresponding feature template is constructed for each known fault category, such as bearing wear, winding overheating, etc. Each template is a vector with the same dimension as the evidence body, and the components in the vector are composed of the average values of the corresponding dominant chain identifier and the fault signal abnormality degree in the historical fault case.

[0111] The evidence body E is calculated with each historical fault feature template, and the cosine similarity calculation results of the evidence body E and all historical fault templates are compared. The historical fault category with the highest similarity value is selected, and the historical fault category is determined as the final fault type of this diagnosis.

[0112] It should be noted that the data of the historical fault case library of the application is derived from the real fault events recorded and finally confirmed in the past operation of the hydro-generator unit or the same type unit. Each fault event should contain complete multi-source signal data sequence and fault type label confirmed by post-analysis, such as thrust bearing wear and stator winding interlayer short circuit, and for each historical fault event, the same analysis process as the fault recognition of the application is adopted, that is, the fault propagation chain is constructed, the dominant fault propagation chain is identified, and the abnormality degree of each fault signal on the chain is calculated.

[0113] In specific implementation, the identification of the dominant fault propagation chain and the average value of the abnormality degree of each fault signal on the chain can be combined in a preset order to form a feature vector. The feature vector constitutes a feature template of this type of fault. For the same type of fault, multiple templates can be generated by collecting multiple historical events, or the average value of the abnormality degree of each fault signal in each historical event can be taken as the final abnormality degree of each fault signal, and the feature template is formed.

[0114] The application reduces the dependence on the accuracy of single sensor data by forming the evidence body with the dominant fault propagation chain and the abnormality degree of each fault source inside the chain, and matching the historical fault template with the similarity.

[0115] ​The above merely illustrates and describes the concept of the present application, and those skilled in the art can make various modifications or supplements to the described specific embodiments or adopt similar ways to replace, as long as the modifications or supplements do not deviate from the concept of the present application or exceed the defined scope of the present application, and should belong to the protection scope of the present application.

Claims

1. A fault identification and analysis method based on multi-source signal analysis of hydro-generators, characterized in that, The method includes: Multiple signal data were collected from the hydro-generator, and the timestamps of each signal were standardized. Signals that continuously deviated from the normal range were screened as suspected abnormal signals. For all suspected abnormal signals on the same fault propagation chain, fault time correlation characterization and fault trend correlation characterization are determined according to their position in the chain, and the fault signal is determined based on the correlation characterization. Extract all fault signals. If they are distributed across multiple fault propagation chains, determine the dominant fault propagation chain based on the chain's fault dominance strength, fault signal density, and the degree of fault signal anomaly. Otherwise, take the chain containing the fault signal as the dominant fault propagation chain. The abnormality of the dominant fault propagation chain and each fault signal within it is used to form an evidence body. The cosine similarity between the evidence body and the feature templates of each historical fault category is calculated. The historical fault category with the highest cosine similarity is taken as the final detected fault type. For each fault propagation chain, traverse the remaining fault propagation chains and check if they share fault signals or have preset strong correlation signals with the current chain. If so, increment the correlation mark count of the current chain by 1, and calculate the fault dominance strength of each fault propagation chain based on the correlation mark count. Divide the number of identified fault signals on all fault propagation chains by the total number of signal sources in the corresponding chain to obtain the fault signal density of that chain. For each fault signal on the fault propagation chain, calculate the absolute value of the relative deviation between its real-time data value and the median of the normal range as the anomaly degree. Based on the anomaly degree, calculate the fault anomaly degree of the fault propagation chain. Compare the fault dominance strength of each fault propagation chain and calculate the coefficient of variation of the fault dominance strength of the fault propagation chain. If the coefficient of variation is less than or equal to a preset threshold, use the linear weighted sum of the fault dominance strength, fault signal density, and fault anomaly degree as the final anomaly dominance score, and select the fault propagation chain with the highest score as the dominant fault propagation chain. If the coefficient of variation is greater than the preset threshold, select the fault propagation chain with the largest fault dominance strength as the dominant fault propagation chain.

2. The fault identification and analysis method based on multi-source signal analysis of hydro-generators as described in claim 1, characterized in that: The signal data in the fault propagation chain are sorted according to the structural connection relationship, physical connection relationship or energy / material flow transmission path of the part of the device from which the data is collected.

3. The fault identification and analysis method based on multi-source signal analysis of hydro-generators as described in claim 1, characterized in that: The process for determining the fault time correlation characterization is as follows: For any two suspected abnormal signals on the same fault propagation chain, determine their relative positions in the chain; If the relative positions are adjacent, calculate the time interval between the abnormal starting points of these two suspected abnormal signals, and record it as the lag time. When the lag time is within the preset delay range, the fault time correlation of the two suspected abnormal signals is characterized as a strong correlation. When the lag time exceeds the preset delay range, the fault time correlation between the two suspected abnormal signals is characterized as a weak correlation. If the relative positions are not adjacent, identify all adjacent signal pairs between the two suspected abnormal signals. For each pair of adjacent signals, determine its fault time correlation characterization based on the lag time of the adjacent signals. If the fault time correlation representations of all adjacent signal pairs are strong correlations, the fault time correlation representations of these two suspected abnormal signals are output as strong correlations; otherwise, the output is a weak correlation. Thus, the fault time correlation representations between each suspected abnormal signal are obtained.

4. The fault identification and analysis method based on multi-source signal analysis of hydro-generators as described in claim 1, characterized in that: The process for determining the fault trend correlation characterization is as follows: From all suspected abnormal signals in the same fault propagation chain, select any two signals to form a target signal pair; If the target signal pairs are adjacent in position on the fault propagation chain, the data sequence of the two signals within a preset time window is extracted; if they are not adjacent, the data sequence of all adjacent signal pairs on the shortest fault propagation path connecting the two signals within a preset time window is extracted. Based on the extracted data sequence, the time-delay cross-correlation index between two signals and the maximum information coefficient between the two signal pairs data sequences are calculated. The time-delay cross-correlation index includes the time-delay correlation coefficient and the target time-delay value. If two signals are adjacent and simultaneously meet all of the following conditions, their fault trend correlation is considered strong; otherwise, it is considered weak: The time-delay correlation coefficient is higher than the first preset threshold; The target time delay value is within the preset delay range; The maximum information coefficient is higher than the second preset threshold; For non-adjacent signals, if the fault trend correlation characteristics of all adjacent signal pairs on the path connecting them are strong, then the fault trend correlation characteristics between the two non-adjacent signals are determined to be strong; if any adjacent signal pair is weakly correlated, then it is determined to be weakly correlated.

5. The fault identification and analysis method based on multi-source signal analysis of hydro-generators as described in claim 4, characterized in that: The calculation process for the time-delay cross-correlation index includes: Within the preset time delay search range, the time offset between the data sequences corresponding to the two signals is changed with the data sampling interval as the step size. Each time the time offset is increased, the data sequences corresponding to the two signals are realigned. Calculate the Pearson correlation coefficient of the aligned data sequences of the two signals at each time delay position, and select the maximum value as the time delay correlation coefficient between the two signals; The time offset corresponding to the time delay correlation coefficient is taken as the target time delay value between the two signals.

6. The fault identification and analysis method based on multi-source signal analysis of hydro-generators as described in claim 4, characterized in that: The calculation process for the maximum information coefficient includes: The data sequence is plotted as a scatter plot, and the grid is divided into multiple grids with different densities in the X and Y directions; For each grid division method, data points are assigned to corresponding grid cells. Based on the distribution of data points in each cell, the mutual information value of the two signals under the grid division is calculated, and the mutual information value is normalized. Iterate through all grid partitioning methods, record the normalized mutual information value obtained at each step, and select the maximum value from all recorded normalized mutual information values ​​as the maximum information coefficient.

7. The fault identification and analysis method based on multi-source signal analysis of hydro-generators as described in claim 1, characterized in that: The fault signal to be determined includes: For each of the suspected abnormal signals, extract its fault time correlation characterization and fault trend correlation characterization with at least one other suspected abnormal signal in the same fault propagation chain; If a suspected abnormal signal is strongly correlated with both the fault time correlation characterization and the fault trend correlation characterization of at least one other suspected abnormal signal, then the suspected abnormal signal is determined to be a fault signal.

8. The fault identification and analysis method based on multi-source signal analysis of hydro-generators as described in claim 7, characterized in that: The statistical analysis of the fault dominance strength of each fault propagation chain includes: For each fault propagation chain, count the number of times it is associated with a marker, and record all fault propagation chains that generated that marker to form a list of associated chains for that chain; For each chain in the list of associated chains, count the number of times it is repeatedly associated with the current chain; For each fault propagation chain, count the number of chains in its associated chain list where the number of repeated association markers is not zero, and record this as the number of valid associated chains; The duplicate association ratio is obtained by dividing the number of valid association chains corresponding to each fault propagation chain by the difference between the total number of fault propagation chains and 1. From the number of associated markers in all fault propagation chains, select the maximum value as the highest number of markers; The ratio of the number of associated markers for each fault propagation chain to the highest number of markers is denoted as the marker ratio. For each fault propagation chain, the fault dominance strength is obtained by multiplying its repetition correlation ratio by its labeling ratio.

9. The fault identification and analysis method based on multi-source signal analysis of hydro-generators as described in claim 8, characterized in that: The calculation process for the degree of fault anomaly in the fault propagation chain includes: Compare the anomalies of all fault signals and take the maximum value as the baseline anomaly degree of the corresponding fault propagation chain. For the same fault propagation chain, the average anomaly degree of all fault signals is calculated, and the proportion of fault signals with anomaly degree exceeding the average value is calculated. For each fault propagation chain, the fault signal ratio is multiplied by the baseline fault anomaly level to obtain the final fault anomaly level.

Images