Intelligent fault diagnosis method for pumped storage unit based on tensor Poincare algorithm

Abstract

The invention belongs to the field of intelligent operation and maintenance of a hydroelectric power generation system, and particularly provides a tensor Poincare algorithm-based intelligent fault diagnosis method for a pumped storage unit, which comprises the following steps of: collecting data of different operation states of the pumped storage unit; based on an enhanced hierarchical decomposition method using a moving average algorithm and a moving difference algorithm, obtaining a frequency component of the operation state signal of the pumped storage unit; obtaining coarseness subsequences of different frequency components based on a coarseness processing method; converting the coarseness subsequence into a Poincare map, and obtaining a tensor Poincare index; the obtained tensor Poincare indexes are divided into a training data set and a test data set, the training data set is used for training the random forest, a classification model is obtained, and the abnormal state of the pumped storage power station unit is recognized through a fault diagnosis model. The method can improve the capability of efficiently extracting information from a large amount of data by the pumped storage unit and enhance the accuracy and efficiency of fault diagnosis so as to ensure safe and stable operation of a power station.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent operation and maintenance of hydropower generation systems, and specifically relates to an intelligent fault diagnosis method for a pumped storage unit based on a tensor Poincare algorithm. Background Art

[0002] Against the backdrop of the dual carbon goals, the proportion of intermittent energy sources such as wind power and photovoltaics has steadily increased, and hydropower has a dual role of supply and regulation in the energy system. Pumped storage is crucial to the future development of hydropower, and pumped storage includes two modes: pumping and power generation. By switching back and forth between these two modes, pumped storage hydropower stations undertake the tasks of peak regulation, frequency regulation, and phase regulation, which is of great significance for increasing the consumption of clean energy and ensuring the safety and reliability of the power system. According to statistics, by 2030, the total installed capacity of pumped storage is expected to reach 230 million kilowatts. With the gradual commissioning of pumped storage units, the focus has gradually shifted to the operation and maintenance of pumped storage units (PSUs). However, there are two problems in the operation of pumped storage units, which seriously endanger the safe and stable operation of pumped storage units. On the one hand, the probability of failure of pumped storage units is higher than that of conventional turbines. The flexible regulation of pumped storage units requires pumped storage units to have a wider operating range, and they need to operate in low load and high vibration areas during regulation. Continuous operation of pumped storage units in these areas will lead to an increased probability of failure. On the other hand, failures of pumped storage units are more dangerous than those of conventional turbine units. Pumped storage units are developing towards high head, high speed and large capacity, and once a failure occurs, the consequences are extremely serious. Traditional planned maintenance cannot quickly capture potential early abnormal information in pumped storage units. Therefore, extracting key information reflecting the abnormal state of pumped storage units and quickly detecting early failures of pumped storage units play a vital role in promoting the efficient operation of pumped storage units.

[0003] Existing research on hydraulic machinery signal feature extraction for intelligent fault diagnosis mainly focuses on the following three aspects: (1) signal processing methods; (2) manual feature extraction methods; and (3) deep learning methods. The above three feature extraction methods are helpful to expand the field of early fault diagnosis of hydraulic machinery. However, these methods have some limitations: 1) Signal processing methods require a lot of expert experience and prior knowledge, and problems such as parameter adjustment, algorithm selection, and signal window length setting will make it difficult to apply them in real-time monitoring of hydraulic machinery; 2) The features extracted by the above manual feature extraction methods are vector features, which can only be extracted from the time domain or frequency domain of the signal, and it is impossible to extract features from the signal that clearly represent the relationship between the time domain and the frequency domain; 3) The performance of the fault diagnosis model based on deep learning algorithms depends on abundant data, and it is impractical to obtain a large amount of abnormal data from actual power plants. In addition, deep learning requires a lot of learning cost to extract features from high-dimensional data, which will place high demands on the hardware infrastructure of the power plant server. 4) Due to the lack of abnormal state data from the power unit, this research mainly focuses on conventional rotating units or bearing components, resulting in a lack of comprehensive exploration of early fault detection methods unique to power units.

[0004] Therefore, establishing a fault diagnosis method that can more efficiently and accurately extract the operating characteristics of pumped-storage units is of great significance for the safe operation of pumped-storage units and ensuring the safety and reliability of power systems. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide an intelligent fault diagnosis method for a pumped-storage unit based on the tensor Poincare algorithm, so as to improve the ability of the pumped-storage unit to efficiently extract information from massive data representing abnormal conditions, enhance the accuracy and efficiency of fault diagnosis, and ensure the safe and stable operation of the power station.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: an intelligent fault diagnosis method for a pumped storage unit based on a tensor Poincare algorithm, comprising the following steps: S1. Collect and store data of different operating states of pumped storage units and build an operating state database of pumped storage units; S2, obtaining the frequency component of the operating status signal of the pumped storage unit based on an enhanced hierarchical decomposition method using a moving average algorithm and a moving difference algorithm; S3, obtaining coarse-grained subsequences of different frequency components based on a coarse-grained processing method; S4, convert all coarse-grained subsequences into Poincare maps, and obtain tensor Poincare indicators as features to describe the signal; S5. Divide the tensor Poincare index obtained in step S4 into a training data set and a test data set, use the training data set to train the random forest, obtain a fully trained pumped-storage unit fault diagnosis model, and finally accurately and quickly identify the abnormal state of the pumped-storage power station unit through the trained pumped-storage unit fault diagnosis model.

[0007] In the preferred solution, in step S1, the specific process of constructing the pumped storage unit operation status database is: First, various sensors are installed on the power supply equipment to detect the real-time dynamic changes of key indicators; second, the data is transmitted to the data acquisition instrument through the signal line, and the data is sent to the industrial personal computer; then, the SQLserver database and the time series database are used to build the operation database of the pumped storage unit; finally, the data in the database is normalized to obtain the required abnormal operation data.

[0008] In a preferred embodiment, step S2 includes the following steps: S201. Define an average operator Q 0 and a difference operator Q 1 to decompose the signal , get the low frequency component of the signal and high frequency components , the specific expression is as follows: ; in, Q 0 and Q 1 represents low-pass and high-pass filters respectively, and Characterizes the low-frequency and high-frequency information of the signal; J Indicates signal x Index point ，N Representative signal x The total number of samples; S202, Different number of layers k Operator Solve by the following formula: ; Among them: Index number s is 0 or 1; S203, at different frequencies The components of are obtained by the following formula: ; in, The number of layers is k hour, No. e Quantity, The number of layers is k hour, The index value in ; e Represents the index of the frequency component, and its expression is as follows: ; in, , .

[0009] In a preferred solution, in step S3, obtaining coarse-grained subsequences of different frequency components based on a coarse-grained processing method includes the following steps: S301, obtained by the following calculation formula At different scales The coarse-grained subsequence under : ; in, express No. Sample points; S302, repeat the above formula until Reaching the maximum scale factor , at this time, all scale subsequences The collection is considered The coarse-grained subsequence of .

[0010] In a preferred solution, in step S4, all coarse-grained subsequences are converted into Poincare maps, and Poincare map indicators are obtained as features to describe the signal, including the following steps: S401, based on phase space reconstruction theory, the signal Convert to two vectors with a delay difference and , which is expressed as follows: ; in, It's a signal No. J sampling points, , N Indicates the length of the signal, and The time interval between them is 1; S402, and The sampling points are plotted as the horizontal and vertical coordinates to form a scatter plot, which is the signal The typical indicators SD1 and SD2 that characterize the short-term and long-term fluctuations of the signal are extracted from different Poincare maps as indicators of various time and frequency scales. and , the relevant expressions are as follows: ; ; in, represents the standard deviation of the signal difference, Represents the standard deviation of the signal. SD1 and SD2 are used to characterize the information of signals at different scales and frequencies; S403, all frequencies and scales and Aggregation as signal The tensor Poincare index of : ; ; in, Represents the Poincare index of the tensor at all frequencies and scales.

[0011] The invention provides a method for intelligent fault diagnosis of pumped storage units based on tensor Poincare algorithm, which has the following beneficial effects: 1. Perform hierarchical decomposition and coarse-grained processing on the running data set to obtain components at different time and frequency scales. Convert different components into Poincare maps in turn, and extract the key indicators of these maps respectively to achieve feature extraction for quantifying the operation of pumped storage units at multiple time and frequency scales.

[0012] 2. The operating characteristics of the pumped-storage unit are extracted at multiple time and frequency scales, and the fault diagnosis model is constructed using the extracted features. This has the advantages of accurate recognition and fast recognition speed, which helps to ensure the safe and stable operation of the pumped-storage unit and promote the intelligent construction of pumped-storage power stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the tensor Poincare index algorithm; Figure 3 Methods for monitoring vortex zones in pumped storage; Figure 4 The typical states of flow in the tailwater pipe at different stages; Figure 5 It is the time domain waveform of pressure pulsation corresponding to different flow stages in the tailwater pipe; Figure 6 are the characteristics of pressure pulsation signals under different conditions; Figure 7 To obtain low-dimensional feature distribution through t-SNE; Figure 8 is the confusion matrix of the proposed diagnostic model. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0015] A technical roadmap of an intelligent fault diagnosis method for pumped storage units based on the tensor Poincare algorithm is as follows Figure 1 As shown, the following steps are included: S1. Collect and store data on different operating states of pumped-storage units and build an operating state database of pumped-storage units.

[0016] The specific process of constructing the pumped storage unit operation status database is as follows: First, various sensors are installed on the power supply equipment to detect the real-time dynamic changes of key indicators. Second, the data is transmitted to the data acquisition instrument (DAI) through the signal line, and the data is sent to the industrial personal computer (IPC) using a combination of the transmission control protocol / Internet protocol (TCP / IP) and the communication protocol. Then, the SQLserver database and the time series database (InfluxDB) are used to build the operation database of the pumped storage unit. Finally, the data in the database is normalized to obtain the required abnormal operation data.

[0017] S2. Obtaining the frequency component of the operating status signal of the pumped storage unit based on an enhanced hierarchical decomposition method using a moving average algorithm and a moving difference algorithm, including the following steps: S201. Define an average operator Q 0 and a difference operator Q 1 to decompose the signal , get the low frequency component of the signal and high frequency components , the specific expression is as follows: ; in, Q 0 and Q 1 represents low-pass and high-pass filters respectively, and Characterizes the low-frequency and high-frequency information of the signal; J Indicates signalx Index point ，N Representative signal x The total number of samples.

[0018] S202, Different number of layers k Operator Solve by the following formula: ; Among them: Index number s Is 0 or 1.

[0019] S203, at different frequencies The components of are obtained by the following formula:

[0020] in, The number of layers is k hour, No. e Quantity, The number of layers is k hour, The index value in .

[0021] e Represents the index of the frequency component, and its expression is as follows:

[0022] in, , .

[0023] S3, obtaining coarse-grained subsequences of different frequency components based on a coarse-grained processing method, comprising the following steps: S301, obtained by the following calculation formula At different scales The coarse-grained subsequence under : ; in, express No. Sample points; S302, repeat the above formula until Reaching the maximum scale factor , at this time, all scale subsequences The set of coarse-grained subsequences of The coarse-grained subsequence of .

[0024] S4, converting all coarse-grained subsequences into Poincare maps and obtaining tensor Poincare indicators as features to describe the signal, including the following steps: S401, based on phase space reconstruction theory, the signal Convert to two vectors with a delay difference and , the expression is as follows:

[0025] in, It's a signal No. J sampling points, , N Indicates the length of the signal, and The time interval between them is 1.

[0026] S402, and The sampling points are plotted as the horizontal and vertical coordinates to form a scatter plot, which is the signal The Poincare map (PP) is used to extract the typical indicators SD1 and SD2 that characterize the short-term and long-term fluctuations of the signal from different Poincare maps as indicators of various time and frequency scales. and , the relevant expressions are as follows: ; ; in, represents the standard deviation of the signal difference, Represents the standard deviation of the signal. SD1 and SD2 are used to characterize the information of signals at different scales and frequencies; S403, all frequencies and scales and Aggregation as signal The tensor Poincare index (TPPI) is: ; ; in, Represents the Poincare index of the tensor at all frequencies and scales.

[0027] S5, dividing the features extracted using the tensor Poincare index (TPPI) algorithm into a training data set and a test data set, that is, the tensor Poincare index obtained by steps S2 to S4, such as Figure 2 shown.

[0028] The obtained tensor Poincare index is divided into a training data set and a test data set. The training data set is used to train the random forest to obtain a fully trained pumped-storage unit fault diagnosis model. Finally, the trained pumped-storage unit fault diagnosis model is used to accurately and quickly identify the abnormal state of the pumped-storage power station unit.

[0029] Example: In pumped storage units, hydraulic anomalies are usually caused by two main reasons. On the one hand, the unit often needs to deviate from the rated operating conditions, which makes its internal flow characteristics extremely complicated. On the other hand, as the operating time of the pumped storage unit increases, the aging degree of the unit gradually deepens. In this case, the rated operating conditions are no longer the optimal operating state of the unit, resulting in abnormal flow characteristics during operation. Due to the above factors, vortex belts in the tailwater pipe have become a common hydraulic anomaly, which poses a significant challenge to the safe and stable operation of the unit.

[0030] This example uses a pumped storage unit as a case study to identify hydraulic anomalies. Figure 3 The specific process of monitoring the vortex flow pattern of a pumped storage unit is demonstrated. Figure 3 As shown in the figure, we use a high-speed camera to quickly capture images of the flow pattern in the straight cone section of the tailwater pipe. These images are an important basis for annotating different sample data. Subsequently, the MPM489 pressure sensor installed in the straight cone section of the tailwater pipe is used to detect the dynamic changes of the pressure pulsation signal under different flow conditions. These signals are effective carriers for characterizing the vortex flow state. Finally, the signal is transmitted to the server through the data acquisition interface for storage.

[0031] By adjusting key parameters such as guide vane opening and load, and conducting a large number of modeling experiments, this example collected images of abnormal flow patterns of the unit and their corresponding pressure pulsation signals. After analysis, this example divides the flow state in the unit's tailwater pipe into five stages: normal stage (NS), vortex band generation stage (VRBS), vortex band development stage (VRDS), vortex band evolution stage (VRES) and columnar vortex band evolution stage (CVRES). Figure 4 The typical flow patterns in the tailwater pipe in these stages and the flow changes in different abnormal flow patterns are shown. As shown in the figure, these five stages reflect the evolution of the vortex belt in the tailwater pipe, and each stage has obvious flow characteristics. In the NS stage, the flow in the tailwater pipe is relatively stable, and there is no obvious cavitation vortex belt; in terms of vortex belt morphology, VRBS, VRDS, VRES and CVRES show a gradual development trend. Among them, VRBS, VRDS and VRES are spiral vortex belts, which pose a greater threat to the safety of the unit; CVRES is a columnar vortex belt, which lasts for a short time and is relatively less harmful.

[0032] According to the annotations of the above images, this example constructs a hydraulic anomaly test dataset to verify the practicality of the proposed model. The dataset contains 250 sets of pressure pulsation signals, covering different flow stages, with 50 sets of data in each stage. Each set of signals contains 4096 sampling points, and the sampling frequency is 1200 Hz. Figure 5 The pressure pulsation time domain waveforms corresponding to different flow states in the tailwater pipe are shown. Figure 5 As shown in the figure, the pressure pulsation in the normal stage (NS) is significantly lower than that in the abnormal stage, mainly because the flow in this stage is relatively stable. With the increase of cavitation degree, the pressure pulsation amplitude of VRBS, VRDS and VRES gradually increases. The pressure pulsation amplitude of CVRES is lower than that of VRES because the spiral vortex is more harmful. In addition, the pressure pulsation signals at different stages have noise due to the complex working environment, which brings challenges to the manual detection of abnormal flow states in the tailwater pipe of the pumped storage unit.

[0033] In this example, the multi-scale indicators of different level components (including SD1 and SD2) are flattened in sequence to form a 1×128 feature vector. Figure 6 The characteristics of pressure pulsation signals under different flow conditions are shown, which mainly include two parts: TPPI-SD1 and TPPI-SD2. Figure 6 As shown in Figure 2, the characteristic value fluctuations of the pressure pulsation signal on different indicators under the same flow state are small, indicating that TPPI has high stability in reflecting the same state of the signal. In addition, there are still significant differences in the characteristic distribution of pressure pulsation signals at different stages. For example, by observing Figure 6 The feature values ​​with indexes ranging from 1 to 10 in (a) can distinguish NS from other states, mainly because the values ​​of this feature block are significantly lower than those of other states. Figure 6 In (b), VRES can be distinguished from other states by the feature values ​​with indexes ranging from 1 to 10, because the value of this feature block is significantly higher than that of other states.

[0034] High-dimensional features are difficult to intuitively show the degree of distinction between different state features, so this example uses t-distributed stochastic neighbor embedding (t-SNE) to compress high-dimensional features into three-dimensional low-dimensional features containing rich state information. Figure 13 shows the distribution of these low-dimensional features. Figure 7 As shown in (a), the distribution of three-dimensional features shows that the features of NS, VRBS and VRES are clearly distinguished from other state signals, while the low-dimensional features of VRDS and CVRES may partially overlap. To further demonstrate the degree of mixing of the two state features, Figure 7(b) provides a scatter matrix of low-dimensional features. From the relationship between feature 1 and feature 3, the mixing degree of VRDS and CVRES is very low, which may be due to the misjudgment caused by the 3D display angle problem. Therefore, based on the above results, it can be concluded that the features extracted by the TPPI algorithm have strong discrimination ability.

[0035] In this example, the features extracted by TPPI are evenly divided into a training set and a test set. The training set is input into the random forest (RF) model for training to obtain a trained model, and the test set is used to evaluate the diagnostic effect of the model. Figure 8 The model's diagnostic results are presented in the form of a confusion matrix. As shown in the figure, after effective training, the model did not misclassify the test set samples (the accuracy of the training set reached 100%), which shows that the model has a good application effect in identifying hydraulic anomalies.

[0036] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent fault diagnosis method for pumped storage units based on tensor Poincare algorithm, characterized in that: The following steps are involved: S1. Collect and store data of different operating states of pumped storage units and build an operating state database of pumped storage units; S2, obtaining the frequency component of the operating status signal of the pumped storage unit based on an enhanced hierarchical decomposition method using a moving average algorithm and a moving difference algorithm; S3, obtaining coarse-grained subsequences of different frequency components based on a coarse-grained processing method; S4, convert all coarse-grained subsequences into Poincare maps, and obtain tensor Poincare indicators as features to describe the signal; S5. Divide the tensor Poincare index obtained in step S4 into a training data set and a test data set, use the training data set to train the random forest, obtain a fully trained pumped-storage unit fault diagnosis model, and finally accurately and quickly identify the abnormal state of the pumped-storage power station unit through the trained pumped-storage unit fault diagnosis model.

2. The intelligent fault diagnosis method for pumped storage units based on the tensor Poincare algorithm according to claim 1 is characterized in that: In step S1, the specific process of constructing the pumped storage unit operation status database is as follows: First, various sensors are installed on the power supply equipment to detect the real-time dynamic changes of key indicators; second, the data is transmitted to the data acquisition instrument through the signal line, and the data is sent to the industrial personal computer; then, the SQLserver database and the time series database are used to build the operation database of the pumped storage unit; finally, the data in the database is normalized to obtain the required abnormal operation data.

3. The intelligent fault diagnosis method for pumped storage units based on the tensor Poincare algorithm according to claim 1 is characterized in that: The step S2 includes the following steps: S201. Define an average operator Q 0 and a difference operator Q 1 to decompose the signal , get the low frequency component of the signal and high frequency components , the specific expression is as follows: ； in, Q 0 and Q 1 represents low-pass and high-pass filters respectively, and Characterizes the low-frequency and high-frequency information of the signal; J Indicates signal x Index point ，N Representative signal x The total number of samples; S202, Different number of layers k Operator Solve by the following formula: ； Where: Index number s is 0 or 1; S203, at different frequencies The components of are obtained by the following formula: ； in, The number of layers is k hour, No. e Quantity, The number of layers is k hour, The index value in ; e Represents the index of the frequency component, and its expression is as follows: ； in, , .

4. The intelligent fault diagnosis method for pumped storage units based on the tensor Poincare algorithm according to claim 1 is characterized in that: In step S3, obtaining coarse-grained subsequences of different frequency components based on a coarse-grained processing method includes the following steps: S301, obtained by the following calculation formula At different scales The coarse-grained subsequence under : ； in, express No. Sample points; S302, repeat the above formula until Reaching the maximum scale factor , at this time, all scale subsequences The collection is considered The coarse-grained subsequence of .

5. The intelligent fault diagnosis method for pumped storage units based on tensor Poincare algorithm according to claim 1 is characterized in that: In step S4, all coarse-grained subsequences are converted into Poincare maps, and Poincare map indicators are obtained as features to describe the signal, including the following steps: S401, based on phase space reconstruction theory, the signal Convert to two vectors with a delay difference and , which is expressed as follows: ； in, It's a signal No. J sampling points, , N Indicates the length of the signal, and The time interval between them is 1; S402, and The sampling points are plotted as the horizontal and vertical coordinates to form a scatter plot, which is the signal The Poincare maps are used to extract the typical indicators SD1 and SD2 that characterize the short-term and long-term fluctuations of the signal from different Poincare maps as indicators at various time and frequency scales. and , the relevant expressions are as follows: ； ； in, represents the standard deviation of the signal difference, Represents the standard deviation of the signal. SD1 and SD2 are used to characterize the information of signals at different scales and frequencies; S403, all frequencies and scales and Aggregation as signal The tensor Poincare index of : ； ； in, Represents the Poincare index of the tensor at all frequencies and scales.

Images