A method for detecting abnormality of hydropower equipment

Through the combination of variational modal decomposition and hierarchical time memory model, data complexity and noise problems in abnormal detection of hydropower equipment are solved, more accurate abnormal detection is achieved, and the reliability of the equipment is improved.

CN114065802BActive Publication Date: 2024-11-12HUADIAN ELECTRIC POWER SCI INST CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111205888.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-10-15
Publication Date
2024-11-12
Estimated Expiration
2041-10-15

AI Technical Summary

Technical Problem

The prior art has poor results in abnormal detection of hydropower equipment, mainly because the physical quantity and signal measured by the sensor are subject to multiple influencing factors, resulting in complex data and difficulty in accurately detecting abnormal phenomena.

Method used

The sensor data is decomposed into sub-signals by using the variational modal decomposition method, and an exception detection is performed using an independent hierarchical time memory (HTM) model to identify the abnormality through sparse distributed expression and prediction error calculation.

Benefits of technology

It improves the accuracy of abnormal detection and inclusiveness of noise, can more accurately identify abnormal data of hydropower equipment, and enhances the reliability of the equipment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114065802B_ABST
    Figure CN114065802B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for detecting abnormality of hydropower equipment. The time length is set to N, and the abnormality detection of hydropower equipment is completed according to the following steps: S1, obtaining the time series data f(t) of the monitoring physical quantity of the hydropower equipment within the set time length through a sensor, 1≤t≤N; S2, performing variational mode decomposition on the time series data f(t) to obtain K sub-signals f 1 (t), f 2 (t),…,f K (t); S3, apply an independent HTM model to each sub-signal for anomaly detection; if the detection result of a sub-signal is abnormal, output abnormality, otherwise output normal; S4, go to the first step and perform anomaly detection on the next time series with a duration of N'. The present invention adopts a variational mode decomposition method to decompose the complex signal measured by the sensor to obtain sub-signals. The sub-signals are relatively simple. By applying anomaly detection methods to sub-signals, the accuracy of anomaly detection is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for detecting anomalies of hydropower equipment, and belongs to the field of fault prediction and health management (PHM) of hydropower equipment. Background Art

[0002] At present, human society has an increasing demand for energy, and hydropower is still the main source of electricity. As the main body of hydropower energy, the single-unit capacity and total installed capacity of hydropower stations are getting larger and larger, and the reliability requirements of hydropower equipment are getting higher and higher. Anomaly detection is an important means to ensure the reliability of hydropower equipment. In order to realize the anomaly detection of hydropower equipment, the real-time measurement of various physical quantities reflecting the current state of hydropower equipment is realized through sensor data, so as to discover the abnormal phenomenon that the measured object deviates from the normal working state. Anomaly detection of hydropower equipment has received more and more attention and application.

[0003] However, due to the harsh operating environment of hydropower equipment, many factors affect the physical quantities and signals measured by sensors, and the complex data obtained by sensors, the usual anomaly detection methods are not effective. Therefore, if an anomaly detection method based on sensor data can be provided, which can adapt to the harsh operating environment of hydropower equipment, the above problems can be solved, and the anomaly detection of hydropower equipment can be accurately realized to ensure the reliability of hydropower equipment. Summary of the invention

[0004] The purpose of the present invention is to overcome the above-mentioned deficiencies in the prior art and to provide a method for detecting abnormalities in hydropower equipment.

[0005] The technical solution adopted by the present invention to solve the above problem is: a method for detecting abnormality of hydropower equipment, characterized in that the time length is set to N, and the abnormality detection of hydropower equipment is completed according to the following steps:

[0006] S1, obtain the time series data f(t) of the monitoring physical quantity of the hydropower equipment within a set time period through sensors, 1≤t≤N;

[0007] S2, perform variational mode decomposition on the time series data f(t) to obtain K sub-signals f1(t), f2(t), ..., f K (t);

[0008] S3, an independent HTM model is applied to each sub-signal for anomaly detection; if the detection result of a sub-signal is abnormal, the output is abnormal, otherwise the output is normal;

[0009] S4, go to the first step and perform anomaly detection on the next time series with a duration of N'.

[0010] In step S1, the monitored physical quantity is preprocessed, and the time series data f(t) of the monitored physical quantity of the hydropower equipment within a set time period is obtained through sensors.

[0011] In step S2, variational mode decomposition is performed on the time series data f(t), including:

[0012] S2-1, let the signal to be decomposed be x(t), t is time, and pre-decompose it into K mode functions x k (t), k = 1, 2, ..., K; use Hilbert transform to transform each mode x k (t) changes from real signal to analytical signal s k (t).

[0013] S2-2, estimate each modal analytical signal s k (t) The center frequency ω k , by multiplying by e-jω k t modulates the corresponding spectrum to the baseband, where j is the imaginary sign of the complex number to achieve frequency mixing; calculate the L2 norm of the demodulated signal gradient, and minimize the sum of the bandwidths of each mode as a constraint condition to estimate the bandwidth of each mode and construct a constrained variational model, whose optimization function is:

[0014]

[0015]

[0016] Among them, δ(t) is the pulse signal, is the partial differential with respect to time t, and x(t) is the source signal.

[0017] S2-3, the optimization problem in S2-2 can be solved by using alternating direction multiplication in the Fourier domain. When the error meets the requirements, the iteration is stopped; the result obtained by the iteration is subjected to inverse Fourier transform, and its actual part is the modal component u in the time domain k (t).

[0018] The first two steps are constrained variational model steps, and the last step is an iterative solution step.

[0019] In step S3, an independent HTM model is applied to each sub-signal for anomaly detection, including:

[0020] S3-1, HTM input and output:

[0021] The input data is fed into the HTM neural model and expressed by the encoder in a sparse distributed manner with 0 and 1 as the basic units; then it is pooled in the sparse space pool to obtain a sparse binary vector a(x) representing the current input, a(x) at time t-1, and t-1) is predicted by learning the sequence memory, and another sparse vector π(x t-1 ) form to output the prediction, and directly output a(x t ), that is, x t The previous sequence is represented and encoded in a sparse distributed manner, and the HTM model predicts the future in the form of sparse vectors by learning these representations.

[0022] S3-2, calculate the prediction deviation:

[0023] a(x t ) is the current input x t Sparse coding, sparse vector π(x t-1 ) represents the HTM network model for a(x t ), a(x t )、π(x t-1 ) is equal to the number of columns in the HTM network (2048 is usually used as a standard value); set S t is the prediction error, S t The value is equal to a(x t ) and π(x t-1 ) is a scalar value inversely proportional to the number of common bits between them, where |a(x t )| is the scalar norm, that is, a(x t ) in the total number of 1 digits;

[0024]

[0025] Where: If the current a(x t ) completely matches the prediction, then the error S t is 0; if a(x t ) and π(x t-1 ) The two vectors are orthogonal, then the error S t is 1.

[0026] S3-3, calculate the likelihood anomaly:

[0027] Prediction error S t It represents the instantaneous measure of the predictability of the current input stream. However, due to the inherent noise of the HTM underlying structure system, the instantaneous error is usually not accurate enough, so the likelihood anomaly L is introduced. t ; Establish S t The distribution model of the error value is used to check whether the current state is likely to be abnormal; in order to calculate the likelihood of abnormality L t , maintain a window containing the last W error values, and use the rolling normal distribution model as the distribution model of the likelihood anomaly, where the sample mean μ t and variance σ 2 tContinuously update based on previous error values;

[0028]

[0029]

[0030] Then, the short-term average of the most recent forecast errors is calculated and a threshold is applied to the Gaussian tail probability, i.e., the Q function, to determine whether it is an anomaly; the likelihood anomaly is defined as the complement of the Q function:

[0031]

[0032]

[0033] Where: W' is the window of the short-term moving average, which is much smaller than the duration W for calculating the forecast error distribution; based on the custom parameter ∈ (∈ is usually set to 10 -5 ), L t Used to report anomalies at time t, the forecast conditions are as follows:

[0034] L1≥1-∈ (6)

[0035] The HTM-based time series data anomaly detection model can perform real-time online detection, is highly tolerant of noise data, and can continuously adapt to changes in time series data, thereby more accurately identifying abnormal data.

[0036] Compared with the prior art, the present invention has the following advantages and effects: the present invention adopts a variational mode decomposition method to decompose the complex signal measured by the sensor to obtain sub-signals, which are relatively simple. By applying an HTM-based anomaly detection model to the sub-signals for anomaly detection, the accuracy of anomaly detection is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a diagram of anomaly detection model based on HTM in the present invention.

[0038] Figure 2 It is a flow chart of a method for detecting abnormality of hydropower equipment in an embodiment of the present invention.

[0039] Figure 3 1 is a diagram of original sensor signals collected in an embodiment of the present invention.

[0040] Figure 4 yes Figure 3 Modal decomposition diagram of the signal K=3. DETAILED DESCRIPTION

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and by way of examples. The following examples are intended to explain the present invention but the present invention is not limited to the following examples.

[0042] In this embodiment, a method for detecting abnormality of hydropower equipment is provided. The duration is set to N, and the abnormality detection of hydropower equipment is completed according to the following steps:

[0043] S1, obtain the time series data f(t) of the monitoring physical quantity of the hydropower equipment within a set time period through sensors, 1≤t≤N;

[0044] S2, perform variational mode decomposition on the time series data f(t) to obtain K sub-signals f1(t), f2(t), ..., f K (t);

[0045] S3, an independent HTM model is applied to each sub-signal for anomaly detection; if the detection result of a sub-signal is abnormal, the output is abnormal, otherwise the output is normal;

[0046] S4, go to the first step and perform anomaly detection on the next time series with a duration of N'.

[0047] In this method, the signal is first decomposed by VMD, and each sub-signal is subjected to anomaly detection using HTM.

[0048] 1. Variational Mode Decomposition

[0049] Signal adaptive decomposition is a common signal processing transformation method. These methods can be used to adaptively represent the original signal as a superposition of multiple different sub-band signals without the need to preset the benchmark and size of the transformation. It can also closely link the time-frequency domain detail information and is widely used in practical engineering.

[0050] The variational mode decomposition (VMD) method is an adaptive decomposition method with important applications in the field of signal processing. It is a non-recursive adaptive signal decomposition method proposed on the basis of traditional Wiener filtering. Compared with the EMD (Empirical Mode Decomposition) method and the LMD (Local Mean Decomposition) method, the variational mode decomposition of the signal decomposed by this method has obvious advantages in signal decomposition accuracy and noise robustness, and has the characteristics of high accuracy, fast convergence and good robustness.

[0051] The VMD algorithm assumes that each decomposed mode has a finite bandwidth with a corresponding center frequency, converts the optimal decomposition problem into a variational constraint problem, and finally uses the alternating direction multiplier method to solve each mode and its center frequency. The algorithm steps can be divided into two steps: constructing a constrained variational model and iterative solution.

[0052] The steps of the VMD algorithm are as follows:

[0053] (1) Let the signal to be decomposed be x(t), where t is time, and pre-decompose it into K mode functions x k (t), k = 1, 2, ..., K; use Hilbert transform to transform each mode x k (t) changes from real signal to analytical signal s k (t).

[0054] (2) Estimate each modal analytical signal s k (t) The center frequency ω k , by multiplying by e-jω k t modulates the corresponding spectrum to the baseband, where j is the imaginary sign of the complex number to achieve frequency mixing; calculate the L2 norm of the demodulated signal gradient, and minimize the sum of the bandwidths of each mode as a constraint condition to estimate the bandwidth of each mode and construct a constrained variational model, whose optimization function is:

[0055]

[0056]

[0057] Among them, δ(t) is the pulse signal, is the partial differential with respect to time t, and x(t) is the source signal.

[0058] (3) The optimization problem in S2-2 can be solved by alternating direction multiplication in the Fourier domain. When the error meets the requirements, the iteration is stopped. The inverse Fourier transform of the iterative result is the modal component u in the time domain. k (t).

[0059] The first two steps are constrained variational model steps, and the last step is an iterative solution step.

[0060] 2. Hierarchical time memory

[0061] The Hierarchical Temporal Memory (HTM) model is a neural network-based machine learning technology that simulates the function of the cerebral cortex proposed by computational neuroscience. It realizes information processing by imitating the structure and algorithmic characteristics of the neocortex.

[0062] The vision, hearing, touch, movement, language and planning of mammals are all completed by the neocortex, and these different intelligent functions are all realized by the neocortex through a highly unified set of neural circuits. HTM is a theoretical framework developed by imitating the neocortex and its functions. The framework model includes hierarchical structure, sparse discrete representation, cortical learning algorithm, spatial data pool and temporal data pool. The HTM model is mainly trained by input data stream, and it can be understood as a neural network. The HTM model simulates a layer of cortex, which is composed of a group of small columns, each of which contains multiple neurons. Each neuron is mainly composed of a cell body, proximal dendrites, terminal dendrites, synapses and neuronal outputs.

[0063] HTM-based anomaly detection models such as Figure 1 As shown, the steps are as follows:

[0064] Step 1: HTM input and output.

[0065] The input data is fed into the HTM neural model and expressed by the encoder in a sparse distributed manner with 0 and 1 as the basic units; then it is pooled in the sparse space pool to obtain a sparse binary vector a(x) representing the current input, a(x) at time t-1, and t-1 ) is predicted by learning the sequence memory, and another sparse vector π(x t-1 ) form to output the prediction, and directly output a(x t ), that is, x t The previous sequence is represented and encoded in a sparse distributed manner, and the HTM model predicts the future in the form of sparse vectors by learning these representations.

[0066] Step 2: Calculate the prediction deviation:

[0067] a(x t ) is the current input x t Sparse coding, sparse vector π(x t-1 ) represents the HTM network model for a(x t ), a(x t )、π(x t-1 ) is equal to the number of columns in the HTM network (2048 is usually used as a standard value); set S t is the prediction error, S t The value is equal to a(x t ) and π(x t-1 ) is a scalar value inversely proportional to the number of common bits between them, where |a(x t )| is the scalar norm, that is, a(x t ) in the total number of 1 digits;

[0068]

[0069] Where: If the current a(x t ) completely matches the prediction, then the error S t is 0; if a(x t ) and π(x t-1 ) The two vectors are orthogonal, then the error S t is 1.

[0070] Step 3: Calculate the likelihood anomaly:

[0071] Prediction error S t It represents the instantaneous measure of the predictability of the current input stream. However, due to the inherent noise of the HTM underlying structure system, the instantaneous error is usually not accurate enough, so the likelihood anomaly L is introduced. t ; Establish S t The distribution model of the error value is used to check whether the current state is likely to be abnormal; in order to calculate the likelihood of abnormality L t , maintain a window containing the last W error values, and use the rolling normal distribution model as the distribution model of the likelihood anomaly, where the sample mean μ t and variance σ 2 t Continuously update based on previous error values;

[0072]

[0073]

[0074] Then, the short-term average of the most recent forecast errors is calculated and a threshold is applied to the Gaussian tail probability, i.e., the Q function, to determine whether it is an anomaly; the likelihood anomaly is defined as the complement of the Q function:

[0075]

[0076]

[0077] Where: W' is the window of the short-term moving average, which is much smaller than the duration W for calculating the forecast error distribution; based on the custom parameter ∈ (∈ is usually set to 10- 5 ), L t Used to report anomalies at time t, the forecast conditions are as follows:

[0078] L1≥1-∈ (6)

[0079] The HTM-based time series data anomaly detection model can perform real-time online detection, is highly tolerant of noise data, and can continuously adapt to changes in time series data, thereby more accurately identifying abnormal data. Based on these characteristics, the HTM detection model can be applied in the field of fault prediction and health management (PHM) of hydropower equipment, and can more effectively discover equipment problems hidden in sensor data than commonly used threshold detection.

[0080] Example of VMD decomposition of sensor signal:

[0081] The data comes from a hydropower station. Among the sensors connected to the hydropower unit, three types of sensor data are used. These three types of sensor data include: generator set flow, guide vane opening and oil tank oil level. The flow of the generator set is the amount of water passing through the turbine per unit time, which directly affects the power generation. The guide vane opening reflects the direct control of the water flow during the operation of the turbine. The generator guide bearing adopts a block structure and uses lubricating oil. The reduction of the oil level in the oil tank will reduce the amount of oil involved in the circulation, affect the lubrication of the bearing, and cause the bearing temperature to rise. Monitor the oil level of four oil tanks, including the upper guide tank, the leakage tank, the water guide leakage tank and the oil collecting tank. In this way, a total of 6 sensor data are collected and used, including the generator set flow, the guide vane opening and the oil level of 4 oil tanks. The data is collected once per second, and the data length is determined to be 5 minutes, that is, N=300. In order to evaluate the abnormality detection method proposed in the present invention, other abnormality detection methods are also used for comparison. These methods include clustering method (Ahmed Imtiaz, Dagnino Aldo, and Ding Yu. Unsupervised anomaly detection based on minimum spanning tree approximated distance measures and its application to hydropower. IEEE Transactions on Automation Science and Engineering, 2019, 654-667), HTM method (Hawkins J, Ahmad S. Why neurons have thousands of synapses, a theory of sequence memory in neocortex, Frontiers in Neural Circuits, 2016, 10: 1-13), using recall rate, precision rate, accuracy rate, specificity and F1 score to evaluate the results, and these indicators are the smaller the better. 500 groups of normal and abnormal data were used for testing, and the results are shown in Table 1. It can be seen that the method of the present invention has the highest index.

[0082] Table 1 Comparison of indicators of different methods

[0083]

[0084]

[0085] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

[0086] Although the present invention has been disclosed as above by way of embodiments, it is not intended to limit the protection scope of the present invention. Any changes and modifications made by any technician familiar with the technology without departing from the concept and scope of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for detecting abnormality of hydropower equipment, characterized in that: Set the duration to N and follow the steps below to complete the abnormal detection of hydropower equipment: S1, obtain the time series data f(t) of the monitoring physical quantity of the hydropower equipment within a set time period through sensors, 1≤t≤N; S2, perform variational mode decomposition on the time series data f(t) to obtain K sub-signals f1(t), f2(t), ..., f K (t); S3, an independent HTM model is applied to each sub-signal for anomaly detection; if the detection result of a sub-signal is abnormal, the output is abnormal, otherwise the output is normal; S4, go to the first step and perform anomaly detection on the next time series with a duration of N'; In step S1, the monitored physical quantity is preprocessed, and the time series data f(t) of the monitored physical quantity of the hydropower equipment within a set time period is obtained through the sensor; In step S2, variational mode decomposition is performed on the time series data f(t), including: S2-1, let the signal to be decomposed be x(t), t is time, and pre-decompose it into K mode functions x k (t), k = 1, 2, ..., K; use Hilbert transform to transform each mode x k (t) changes from real signal to analytical signal s k (t); S2-2, estimate each modal analytical signal s k (t) The center frequency ω k , by multiplying The corresponding spectrum is modulated to the baseband, where j is the sign of the imaginary part of the complex number to achieve frequency mixing; the L2 norm of the demodulated signal gradient is calculated, and the sum of the bandwidths of each mode is minimized as a constraint condition to estimate the bandwidth of each mode and construct a constrained variational model, whose optimization function is: Among them, δ(t) is the pulse signal, is the partial differential with respect to time t, x(t) is the source signal; S2-3, use alternating direction multiplication in the Fourier domain to solve the optimization problem in S2-2. When the error meets the requirements, stop the iteration; perform inverse Fourier transform on the iterative result, and its actual part is the modal component u in the time domain. k (t); The first two steps are constrained variational model steps, and the last step is an iterative solution step; In step S3, an independent HTM model is applied to each sub-signal for anomaly detection, including: S3-1, HTM input and output: The input data is fed into the HTM neural model and expressed by the encoder in a sparse distributed manner with 0 and 1 as the basic units; then it is pooled in the sparse space pool to obtain a sparse binary vector a(x) representing the current input, a(x) at time t-1, and t-1 ) is predicted by learning the sequence memory, and another sparse vector π(x t-1 ) form to output the prediction, and directly output a(x t ), that is, x t The previous sequence is represented and encoded in a sparse distributed manner, and the HTM model predicts the future in the form of sparse vectors by learning these representations; S3-2, calculate the prediction deviation: a(x t ) is the current input x t Sparse coding, sparse vector π(x t-1 ) represents the HTM network model for a(x t ), a(x t ),π(x t-1 ) is equal to the number of columns in the HTM network; set S t is the prediction error, S t The value is equal to a(x t ) and π(x t-1 ) is a scalar value inversely proportional to the number of common bits between them, where |a(x t )| is the scalar norm, that is, a(x t ) in the total number of 1 digits; Where: If the current a(x t ) completely matches the prediction, then the error S t is 0; if a(x t ) and π(x t-1 ) The two vectors are orthogonal, then the error S t is 1; S3-3, calculate the likelihood anomaly: Prediction error S t Represents the instantaneous measure of the predictability of the current input stream, introducing the likelihood anomaly L t ; Establish S t The distribution model of the error value is used to check whether the current state is likely to be abnormal; the rolling normal distribution model is used as the distribution model of the likelihood anomaly, where the sample mean μ t and variance σ 2 t Continuously update based on previous error values; Then, the short-term average of the most recent forecast errors is calculated and a threshold is applied to the Gaussian tail probability, i.e., the Q function, to determine whether it is an anomaly; the likelihood anomaly is defined as the complement of the Q function: Where: W' is the window of the short-term moving average, which is less than the duration W of calculating the forecast error distribution; based on the custom parameter ∈, L t Used to report anomalies at time t, the forecast conditions are as follows: L t ≥1-∈ (6) The HTM-based time series data anomaly detection model performs real-time online detection, has strong tolerance for noise data, and can continuously adapt to changes in time series data, thereby accurately identifying abnormal data.