Hydraulic instability diagnosis method for hydraulic machinery based on joint symbodynamic entropy

Abstract

The invention discloses a method for diagnosing hydraulic instability of hydraulic machinery based on joint symbolic dynamics entropy. The method comprises the following steps: firstly, collecting a vibration signal through a vibration sensor mounted outside a system; the collected vibration signals are preprocessed, and characteristic parameters representing internal flow of the system are extracted; then establishing and training a hydraulic instability diagnosis model; and finally, collecting vibration signals on line in real time, calculating the fine composite multi-scale joint symbol dynamics entropy of unknown signals, constructing feature vectors, and inputting the feature vectors into the trained hydraulic instability diagnosis model to diagnose the internal flow state of the unit. According to the method, the problem of noise interference in traditional time domain and frequency domain analysis in the prior art is solved, the accuracy of flow instability diagnosis is improved, and meanwhile the method can be further expanded in practical engineering with the low signal-to-noise ratio.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of hydraulic machinery stability optimization, and particularly relates to a hydraulic machinery hydraulic instability diagnosis method based on joint symbolic dynamics entropy. BACKGROUND

[0002] Serious failures are often caused by fluid dynamic instability in the operation process of hydraulic systems (such as water pump water turbines, water conveying pipelines, axial flow water turbines, etc.), for example, cavitation phenomenon, two-phase flow distortion, etc. These unstable states not only lead to a sharp drop in efficiency, but also may cause mechanical vibration, structural damage and even system collapse. Taking cavitation as an example, the occurrence of water turbine cavitation is highly concealed in the early stage, and once it occurs, it will cause irreversible local cavitation and abrasion damage to the flow parts, and worsen the stability of the water turbine, which is called the "cancer" of the water turbine. Therefore, it is necessary to propose an efficient and accurate hydraulic instability diagnosis method.

[0003] The existing diagnosis technology has the following key defects: the traditional method relies on time-frequency analysis (such as FFT, wavelet transform) of pressure fluctuation signals or vibration signals, and the sensitivity to the nonlinear characteristics of hydraulic instability is insufficient, and the signal energy is weak in the early stage of instability. And the signal caused by cavitation and other hydraulic instability flow phenomena has non-stationary and low signal-to-noise ratio characteristics, so that the commonly used frequency spectrum envelope analysis method cannot distinguish effective features from background turbulent noise. In addition, the conventional entropy value algorithm (approximate entropy, sample entropy) has high computational complexity and is significantly disturbed by random noise, which easily causes false alarms or false alarms. Therefore, it is necessary to develop a diagnosis mechanism with strong noise resistance, efficient calculation and universality for multiple types of hydraulic instability, to realize early and accurate capture of hydraulic instability phenomena such as hump stall, flow distortion, cavitation inception, etc. SUMMARY

[0004] The purpose of the present application is to provide a hydraulic machinery hydraulic instability diagnosis method based on joint symbolic dynamics entropy, which solves the noise interference problem in traditional time and frequency domain analysis in the prior art, improves the accuracy of flow instability diagnosis, and can be further expanded in actual engineering with low signal-to-noise ratio.

[0005] The technical solution adopted by the present application is a hydraulic machinery hydraulic instability diagnosis method based on joint symbolic dynamics entropy, which is implemented according to the following steps: Step 1, collecting vibration signals through a vibration sensor installed outside the system x t Step 2, preprocessing the vibration signals collected in step 1 to extract characteristic parameters representing the flow inside the system; Step 3, constructing and training a hydraulic instability diagnosis model; ​​Step 4: Real-time online acquisition of vibration signals, calculation of the fine composite multi-scale joint symbolic dynamic entropy of unknown signals, construction of feature vectors, input into the trained hydraulic instability diagnostic model for diagnosis of the internal flow state of the unit.

[0006] The invention is further characterized in that, In step 2, after the vibration data acquisition is completed, the vibration sequence distribution is mean-processed according to the Laida criterion to remove abnormal data interference.

[0007] Step 2 is implemented in the following steps: Step 2.1, for vibration signals x ( t ), x ( t () is a one-dimensional vibration time series, based on Takens' delay embedding theorem to convert the vibration signal x ( t Reconstructing the higher-dimensional phase space Z ( m ), (1) in m The phase space embedding dimension is represented by N, and the length of the one-dimensional vibration time series is represented by N. z i ( m ) indicates the first i A phase space vector, i =1,2,…,N-m+1, x 1, x 2,…, x N This represents the 1st, 2nd, ..., Nth data value in a one-dimensional vibration time series. Step 2.2: Calculate the cosine similarity between adjacent trajectories in phase space. d ( z i ( m ), z i+1 ( m )), thus obtaining the similarity sequence D ( m )={ d ( z 1( m ), z 2( m )), d ( z 2( m ), z 3( m )),…, d ( z N-m( m ), z N-m+1 ( m ))}, z i ( m )and z i ( m The vectors (i) and (j) in the phase space, i.e., the i and j trajectories, are used to quantify the local pattern similarity of the system's dynamic state. (2) Cosine similarity in the formula d j = d ( z i ( m ), z j ( m The value range is (-1, 1); Step 2.3: Convert the cosine similarity sequence between trajectories into a symbol sequence. S = { s 1, s 2, …, s N-m Let the static symbol number be} e The range of cosine similarity values ​​[-1, 1] is divided into... e There are equidistant intervals, each interval having a length Δ. s =2 / e , recorded as I 1, I 2, …, I ε The cosine similarity sequence is transformed into a symbol sequence using the following formula, and then the probability that the cosine similarity falls between each cell is calculated, denoted as the state probability. p 1, p 2, …, p ε ), (3) s 1, s 2, …, s ε They are respectively e Static symbol marker, d k Let Δ be the k-th cosine similarity value calculated by the above formula. s The static interval length is 2 / e ; Step 2.4: Perform a first-order difference on the cosine similarity sequence between the tracks to obtain the similarity fluctuation trend Δ.d The dynamic rate of change of quantified similarity: (4) Δ d k d represents the calculated k-th first-order difference cosine similarity value. k+1 and d k Let the (k+1)th and kth cosine similarity values ​​calculated by the above formula be ; and let the number of dynamic fluctuation symbols be . w The trend of change Δ d The range is divided into w There are equidistant intervals, and the length of each dynamic interval is... d =4 / w The following formula transforms the fluctuation trend sequence into a symbolic sequence. T = { π 1, π 2, …, π N-m}: (5) π 1, π 2, …, π ω They are respectively w Dynamic symbolic marker, Δ d The first-order difference cosine similarity value calculated by the above formula is... d The static interval length is 4 / w ; Step 2.5: Combine the static symbols and dynamic fluctuation trend symbols at each moment into a two-dimensional joint symbol. u k =( s k , π k The size of the two-dimensional joint symbolic state space is... = w × e Statistical analysis of each joint symbol ( s i , π j The frequency of occurrence of ) yields the joint symbol probability. p ( s i , π j ): (6) Number(a) represents the number of times the combined symbol encoding 'a' appears; Step 2.6: Based on Shannon's information entropy theory and the calculated joint state probabilities... p ij The joint symbolic dynamic entropy JSFE is calculated using the following formula: (7) Where x represents the vibration time series, and m represents the phase space reconstruction dimension. e Represents static symbolic numbers, w P represents the number of signs of dynamic fluctuations. ij This represents the joint sign probability calculated by the above formula; Step 2.7: Based on the maximum scale quantity t max Construct coarse-grained sequences.

[0008] The fluctuation trend Δ in step 2.4 d The value range is (-2, 2).

[0009] Step 2.7 is implemented according to the following steps: For a one-dimensional vibration time series x ( t The obtained first k Coarse-grained sequences y (τ) k ={ y (τ) k,1 , y (τ) k,2 , …}as follows: (9) in, N τ is the length of the vibration sequence; τ is the scale factor. x b The first vibration sequence b One value; Calculate the joint symbolic dynamic entropy at each scale, and obtain the following: t max Each entropy value is named the fine-grained composite multi-scale joint symbolic dynamics entropy, and the fine-grained composite multi-scale joint symbolic dynamics entropy at each scale is constructed into a feature vector. S : (10).

[0010] RCMJSFE denotes the fine composite multi-scale joint symbolic dynamic entropy, where x represents the vibration time series and m represents the phase space reconstruction dimension. e Represents static symbolic numbers, w The sign of the dynamic fluctuation is represented by τ, which is the scaling factor, and y is the number of signs. α(τ) This represents the coarse-grained sequence at scale τ obtained from the above equation.

[0011] In step 3, the hydraulic instability diagnosis model is a random forest model based on ensemble learning. This model constructs a feature vector from the refined composite multi-scale joint symbolic dynamic entropy feature parameters for each working condition calculated in step 2. S As input data to the random forest model, the coded symbols NC / IC / CD / CC corresponding to the operating state of each working condition are used as labels. Each input data and its corresponding label are used as training samples. The training samples are normalized and then input into the random forest model.

[0012] In step 3, five-fold cross-validation and grid search are used to optimize the hyperparameters of the random forest model. When the number of training iterations reaches the set maximum or the accuracy on the validation set no longer increases after multiple iterations, the random forest model is saved as a diagnostic model after training.

[0013] Step 4 is as follows: Based on real-time vibration signals acquired by sensors indicating unknown internal flow conditions, the fine-grained composite multi-scale joint symbolic dynamic entropy characteristic parameters of the real-time vibration signals at various scales are calculated, constructing the feature vector S corresponding to this unknown operating condition. The feature vector S corresponding to the unknown working condition The flow state label corresponding to the unknown operating condition is obtained by inputting the hydraulic instability diagnosis model trained in step 3. The internal flow state of the unit is diagnosed based on the flow state label. If the label corresponds to an unstable flow condition, an early warning is issued to the power plant operator to perform the corresponding follow-up operations.

[0014] The beneficial effects of this invention are that the hydraulic instability diagnosis method for hydraulic machinery based on joint symbolic kinetic entropy non-invasively collects the flow-induced oscillation signal of the system through vibration sensors installed outside the system, and extracts signal features using the proposed joint symbolic kinetic entropy method. This method captures dynamic patterns at different time scales through multi-scale coarse-grained processing; the phase space reconstruction process effectively suppresses environmental noise; and the joint symbolic entropy combining static similarity and dynamic trends comprehensively assesses the complexity and unpredictability of time series, making it more suitable for extracting flow characteristics of non-stationary processes such as hydraulic instability. Combining it with a random forest model can effectively improve the accuracy of hydraulic instability diagnosis, and it requires less operator skill. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the method of the present invention; Figure 2 This is a schematic diagram of the signal fine composite multi-scale joint symbol dynamic entropy calculation process in the method of this invention; Figure 3 This is a diagram of the gas-liquid two-phase pipe flow testing system in Example 5 of the present invention; Figure 4(a) shows the pipeline vibration signal under the slug flow condition in Example 5 of the present invention; Figure 4(b) shows the pipe vibration signal under the plug flow condition in Example 5 of the present invention; Figure 4(c) shows the pipeline vibration signal under the foam wavy flow condition in Example 5 of the present invention; Figure 5 This is a confusion matrix diagram of the diagnostic results of the method of the present invention in Example 5 of the present invention; Figure 6 This is a diagram of the axial-flow turbine testing system in Example 6 of the present invention; Figure 7(a) is a time-domain diagram of the rotor vibration signal in the non-cavitation state in Example 6 of the present invention; Figure 7(b) is a time-domain diagram of the rotor vibration signal in the initial cavitation state in Example 6 of the present invention; Figure 7(c) is a time-domain diagram of the vibration signal when the cavitation state continues to develop in Example 6 of the present invention; Figure 7(d) is a time-domain diagram of the vibration signal when the cavitation state in Example 6 of the present invention continues to develop further based on Figure 7(c); Figure 8 This is a diagram showing the diagnostic results of the method of the present invention in Example 6 of the present invention.

[0016] In the diagram, 1 is the outlet water tank; 2 is the booster pump; 3 is the inlet water tank; 4 is the electromagnetic shut-off valve; 5 is the electromagnetic flowmeter; 6 is the test section; 7 is the screw compressor; 8 is the pressure stabilizing gas tank; 9 is the gas-liquid mixer; 10 is the data acquisition card; 11 is the computer; 12 is the booster centrifugal pump; 13 is the water tank; 14 is the low-pressure water tank; 15 is the experimental axial flow turbine; and 16 is the high-pressure water tank. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0018] This invention relates to a method for diagnosing hydraulic instability in hydraulic machinery based on joint symbolic dynamic entropy, the flowchart of which is shown below. Figure 1 , Figure 2 As shown, please follow these steps: Step 1: Collect vibration signals using vibration sensors installed externally to the system. x ( t ); Step 2: The vibration signals collected in Step 1 are preprocessed, and then the characteristic parameters characterizing the internal flow of the system are extracted using the proposed fine composite multi-scale joint symbolic dynamic entropy method. Step 2 is implemented in the following steps: Step 2.1, for vibration signalsx ( t ), x ( t () is a one-dimensional vibration time series, based on Takens' delay embedding theorem to convert the vibration signal x ( t Reconstructing the higher-dimensional phase space Z ( m ), (1) in m The phase space embedding dimension is represented by N, and the length of the one-dimensional vibration time series is represented by N. z i ( m ) indicates the first i A phase space vector, i =1,2,…,N-m+1, x 1, x 2,…, x N This represents the 1st, 2nd, ..., Nth data value in a one-dimensional vibration time series. Step 2.2: Calculate the cosine similarity between adjacent trajectories in phase space. d ( z i ( m ), z i+1 ( m )), thus obtaining the similarity sequence D ( m )={ d ( z 1( m ), z 2( m )), d ( z 2( m ), z 3( m )),…, d ( z N-m ( m ), z N-m+1 ( m ))}, z i ( m )and z i ( m The vectors (i) and (j) in the phase space, i.e., the i and j trajectories, are used to quantify the local pattern similarity of the system's dynamic state. (2) Cosine similarity in the formulad j = d ( z i ( m ), z j ( m The value range is (-1, 1); Step 2.3: Convert the cosine similarity sequence between trajectories into a symbol sequence. S = { s 1, s 2, …, s N-m Let the static symbol number be} e The range of cosine similarity values ​​[-1, 1] is divided into... e There are equidistant intervals, each interval having a length Δ. s =2 / e , recorded as I 1, I 2, …, I ε The cosine similarity sequence is transformed into a symbol sequence using the following formula, and then the probability that the cosine similarity falls between each cell is calculated, denoted as the state probability. p 1, p 2, …, p ε ), (3) s 1, s 2, …, s ε They are respectively e Static symbol marker, d k Let Δ be the k-th cosine similarity value calculated by the above formula. s The static interval length is 2 / e ; Step 2.4: Perform a first-order difference on the cosine similarity sequence between the tracks to obtain the similarity fluctuation trend Δ. d The dynamic rate of change of quantified similarity: (4) Δ d k d represents the calculated k-th first-order difference cosine similarity value. k+1 and d k Let the (k+1)th and kth cosine similarity values ​​calculated by the above formula be ; and let the number of dynamic fluctuation symbols be . w The trend of change Δ d The range is divided into wThere are equidistant intervals, and the length of each dynamic interval is... d =4 / w The following formula transforms the fluctuation trend sequence into a symbolic sequence. T = { π 1, π 2, …, π N-m}: (5) π 1, π 2, …, π ω They are respectively w Dynamic symbolic marker, Δ d The first-order difference cosine similarity value calculated by the above formula is... d The static interval length is 4 / w ; The fluctuation trend Δ in step 2.4 d The value range is (-2, 2).

[0019] Step 2.5: Combine the static symbols and dynamic fluctuation trend symbols at each moment into a two-dimensional joint symbol. u k =( s k , π k The size of the two-dimensional joint symbolic state space is... = w × e Statistical analysis of each joint symbol ( s i , π j The frequency of occurrence of ) yields the joint symbol probability. p ( s i , π j ): (6) Number(a) represents the number of times the combined symbol encoding 'a' appears; Step 2.6: Based on Shannon's information entropy theory and the calculated joint state probabilities... p ij The joint symbolic dynamic entropy JSFE is calculated using the following formula: (7) Where x represents the vibration time series, and m represents the phase space reconstruction dimension. e Represents static symbolic numbers, w P represents the number of signs of dynamic fluctuations.ij This represents the joint sign probability calculated by the above formula; Step 2.7: Based on the maximum scale quantity t max Construct coarse-grained sequences.

[0020] Step 2.7 is implemented according to the following steps: For a one-dimensional vibration time series x ( t The obtained first k Coarse-grained sequences y (τ) k ={ y (τ) k,1 , y (τ) k,2 , …}as follows: (9) in, N τ is the length of the vibration sequence; τ is the scale factor. x b The first vibration sequence b One value; Calculate the joint symbolic dynamic entropy at each scale, and obtain the following: t max Each entropy value is named the fine-grained composite multi-scale joint symbolic dynamics entropy, and the fine-grained composite multi-scale joint symbolic dynamics entropy at each scale is constructed into a feature vector. S : (10).

[0021] RCMJSFE denotes the fine composite multi-scale joint symbolic dynamic entropy, where x represents the vibration time series and m represents the phase space reconstruction dimension. e Represents static symbolic numbers, w The sign of the dynamic fluctuation is represented by τ, which is the scaling factor, and y is the number of signs. α (τ) This represents the coarse-grained sequence at scale τ obtained from the above equation.

[0022] Step 3: Construct and train the hydraulic instability diagnostic model; normalize the feature parameters extracted in Step 2 and use them as a sample set. Divide the sample set into a training set and a validation set according to a certain ratio. Input the training set into the random forest training module. Adjust the model parameters in real time through validation during the training process. Finally, the hydraulic instability diagnostic model is trained. In step 3, the hydraulic instability diagnosis model is a random forest model based on ensemble learning. This model constructs a feature vector from the refined composite multi-scale joint symbolic dynamic entropy feature parameters for each working condition calculated in step 2. S As input data for the random forest model, the coded symbols NC / IC / CD / CC corresponding to the operating state of each working condition are used as labels. Each input data and its corresponding label are used as training samples. After normalization, the training samples are input into the random forest model. Five-fold cross-validation and grid search are used to optimize the hyperparameters of the model. When the number of training iterations reaches the set maximum or the accuracy on the validation set no longer increases after multiple iterations, the random forest model is saved as the trained diagnostic model.

[0023] Step 4: Collect vibration signals in real time online using sensors, calculate the fine composite multi-scale joint symbolic dynamic entropy of the unknown signal, construct a feature vector, and input it into the trained hydraulic instability diagnostic model to diagnose the internal flow state of the unit.

[0024] Step 4 is as follows: Based on real-time vibration signals acquired by sensors indicating unknown internal flow conditions, the fine-grained composite multi-scale joint symbolic dynamic entropy characteristic parameters of the real-time vibration signals at various scales are calculated, constructing the feature vector S corresponding to this unknown operating condition. The feature vector S corresponding to the unknown working condition The flow state label corresponding to the unknown operating condition is obtained by inputting the hydraulic instability diagnosis model trained in step 3. The internal flow state of the unit is diagnosed based on the flow state label. If the label corresponds to an unstable flow condition, an early warning is issued to the power plant operator to perform the corresponding follow-up operations.

[0025] Example 1 This invention relates to a method for diagnosing hydraulic instability in hydraulic machinery based on joint symbolic dynamic entropy, the flowchart of which is shown below. Figure 1 As shown, please follow these steps: Step 1: Collect vibration signals using vibration sensors installed externally to the system. x ( t ); Step 2: The vibration signals collected in Step 1 are preprocessed, and then the characteristic parameters characterizing the internal flow of the system are extracted using the proposed fine composite multi-scale joint symbolic dynamic entropy method. Step 3: Construct and train the hydraulic instability diagnostic model; normalize the feature parameters extracted in Step 2 and use them as a sample set. Divide the sample set into a training set and a validation set according to a certain ratio. Input the training set into the random forest training module. Adjust the model parameters in real time through validation during the training process. Finally, the hydraulic instability diagnostic model is trained. Step 4: Collect vibration signals in real time online using sensors, calculate the fine composite multi-scale joint symbolic dynamic entropy of the unknown signal, construct a feature vector, and input it into the trained hydraulic instability diagnostic model to diagnose the internal flow state of the unit.

[0026] This invention extracts signal features based on the proposed joint symbolic dynamic entropy method, and combines it with a random forest model to effectively improve the accuracy of hydraulic instability diagnosis. By capturing dynamic patterns at different time scales through multi-scale coarsening, the phase space reconstruction process effectively suppresses environmental noise. The joint symbolic entropy, combining static similarity and dynamic trends, comprehensively assesses the complexity and unpredictability of time series, making it more suitable for extracting flow characteristics of non-stationary processes such as hydraulic instability. Mapping vibration signals to a high-dimensional space greatly avoids noise interference problems in traditional time and frequency domain analyses, improving the accuracy of flow instability diagnosis. Furthermore, it can be further expanded in practical engineering applications with lower signal-to-noise ratios, showing great application potential.

[0027] Example 2 This invention relates to a method for diagnosing hydraulic instability in hydraulic machinery based on joint symbolic dynamic entropy, the flowchart of which is shown below. Figure 1 As shown, please follow these steps: Step 1: Collect vibration signals using vibration sensors installed externally to the system. x ( t ); Step 2: The vibration signals collected in Step 1 are preprocessed, and then the characteristic parameters characterizing the internal flow of the system are extracted using the proposed fine composite multi-scale joint symbolic dynamic entropy method. Step 2 is implemented in the following steps: Step 2.1, for vibration signals x ( t ), x ( t () is a one-dimensional vibration time series, based on Takens' delay embedding theorem to convert the vibration signal x ( t Reconstructing the higher-dimensional phase space Z ( m ), (1) in m The phase space embedding dimension is represented by N, and the length of the one-dimensional vibration time series is represented by N. z i ( m ) indicates the first i A phase space vector, i =1,2,…,N-m+1, x 1, x2,…, x N This represents the 1st, 2nd, ..., Nth data value in a one-dimensional vibration time series. Step 2.2: Calculate the cosine similarity between adjacent trajectories in phase space. d ( z i ( m ), z i+1 ( m )), thus obtaining the similarity sequence D ( m )={ d ( z 1( m ), z 2( m )), d ( z 2( m ), z 3( m )),…, d ( z N-m ( m ), z N-m+1 ( m ))}, z i ( m )and z i ( m The vectors (i) and (j) in the phase space, i.e., the i and j trajectories, are used to quantify the local pattern similarity of the system's dynamic state. (2) Cosine similarity in the formula d j = d ( z i ( m ), z j ( m The value range is (-1, 1); Step 2.3: Convert the cosine similarity sequence between trajectories into a symbol sequence. S = { s 1, s 2, …, s N-m Let the static symbol number be} e The range of cosine similarity values ​​[-1, 1] is divided into... e There are equidistant intervals, each interval having a length Δ. s =2 / e , recorded asI 1, I 2, …, I ε The cosine similarity sequence is transformed into a symbol sequence using the following formula, and then the probability that the cosine similarity falls between each cell is calculated, denoted as the state probability. p 1, p 2, …, p ε ), (3) s 1, s 2, …, s ε They are respectively e Static symbol marker, d k Let Δ be the k-th cosine similarity value calculated by the above formula. s The static interval length is 2 / e ; Step 2.4: Perform a first-order difference on the cosine similarity sequence between the tracks to obtain the similarity fluctuation trend Δ. d The dynamic rate of change of quantified similarity: (4) Δ d k d represents the calculated k-th first-order difference cosine similarity value. k+1 and d k Let the (k+1)th and kth cosine similarity values ​​calculated by the above formula be ; and let the number of dynamic fluctuation symbols be . w The trend of change Δ d The range is divided into w There are equidistant intervals, and the length of each dynamic interval is... d =4 / w The following formula transforms the fluctuation trend sequence into a symbolic sequence. T = { π 1, π 2, …, π N-m}: (5) π 1, π 2, …, π ω They are respectively w Dynamic symbolic marker, Δ d The first-order difference cosine similarity value calculated by the above formula is... d The static interval length is 4 / w ; Step 2.5: Combine the static symbols and dynamic fluctuation trend symbols at each moment into a two-dimensional joint symbol. u k =( s k , π k The size of the two-dimensional joint symbolic state space is... = w × e Statistical analysis of each joint symbol ( s i , π j The frequency of occurrence of ) yields the joint symbol probability. p ( s i , π j ): (6) Number(a) represents the number of times the combined symbol encoding 'a' appears; Step 2.6: Based on Shannon's information entropy theory and the calculated joint state probabilities... p ij The joint symbolic dynamic entropy JSFE is calculated using the following formula: (7) Where x represents the vibration time series, and m represents the phase space reconstruction dimension. e Represents static symbolic numbers, w P represents the number of signs of dynamic fluctuations. ij This represents the joint sign probability calculated by the above formula; Step 2.7: Based on the maximum scale quantity t max Construct coarse-grained sequences.

[0028] Step 3: Construct and train the hydraulic instability diagnostic model; normalize the feature parameters extracted in Step 2 and use them as a sample set. Divide the sample set into a training set and a validation set according to a certain ratio. Input the training set into the random forest training module. Adjust the model parameters in real time through validation during the training process. Finally, the hydraulic instability diagnostic model is trained. Step 4: Collect vibration signals in real time online using sensors, calculate the fine composite multi-scale joint symbolic dynamic entropy of the unknown signal, construct a feature vector, and input it into the trained hydraulic instability diagnostic model to diagnose the internal flow state of the unit.

[0029] Example 3 This invention relates to a method for diagnosing hydraulic instability in hydraulic machinery based on joint symbolic dynamic entropy, the flowchart of which is shown below. Figure 1As shown, please follow these steps: Step 1: Collect vibration signals using vibration sensors installed externally to the system. x ( t ); Step 2: The vibration signals collected in Step 1 are preprocessed, and then the characteristic parameters characterizing the internal flow of the system are extracted using the proposed fine composite multi-scale joint symbolic dynamic entropy method. Step 2 is implemented in the following steps: Step 2.1, for vibration signals x ( t ), x ( t () is a one-dimensional vibration time series, based on Takens' delay embedding theorem to convert the vibration signal x ( t Reconstructing the higher-dimensional phase space Z ( m ), (1) in m The phase space embedding dimension is represented by N, and the length of the one-dimensional vibration time series is represented by N. z i ( m ) indicates the first i A phase space vector, i =1,2,…,N-m+1, x 1, x 2,…, x N This represents the 1st, 2nd, ..., Nth data value in a one-dimensional vibration time series. Step 2.2: Calculate the cosine similarity between adjacent trajectories in phase space. d ( z i ( m ), z i+1 ( m )), thus obtaining the similarity sequence D ( m )={ d ( z 1( m ), z 2( m )), d ( z 2( m ), z 3( m )),…, d ( z N-m ( m ),z N-m+1 ( m ))}, z i ( m )and z i ( m The vectors (i) and (j) in the phase space, i.e., the i and j trajectories, are used to quantify the local pattern similarity of the system's dynamic state. (2) Cosine similarity in the formula d j = d ( z i ( m ), z j ( m The value range is (-1, 1); Step 2.3: Convert the cosine similarity sequence between trajectories into a symbol sequence. S = { s 1, s 2, …, s N-m Let the static symbol number be} e The range of cosine similarity values ​​[-1, 1] is divided into... e There are equidistant intervals, each interval having a length Δ. s =2 / e , recorded as I 1, I 2, …, I ε The cosine similarity sequence is transformed into a symbol sequence using the following formula, and then the probability that the cosine similarity falls between each cell is calculated, denoted as the state probability. p 1, p 2, …, p ε ), (3) s 1, s 2, …, s ε They are respectively e Static symbol marker, d k Let Δ be the k-th cosine similarity value calculated by the above formula. s The static interval length is 2 / e ; Step 2.4: Perform a first-order difference on the cosine similarity sequence between the tracks to obtain the similarity fluctuation trend Δ. dThe dynamic rate of change of quantified similarity: (4) Δ d k d represents the calculated k-th first-order difference cosine similarity value. k+1 and d k Let the (k+1)th and kth cosine similarity values ​​calculated by the above formula be ; and let the number of dynamic fluctuation symbols be . w The trend of change Δ d The range is divided into w There are equidistant intervals, and the length of each dynamic interval is... d =4 / w The following formula transforms the fluctuation trend sequence into a symbolic sequence. T = { π 1, π 2, …, π N-m}: (5) π 1, π 2, …, π ω They are respectively w Dynamic symbolic marker, Δ d The first-order difference cosine similarity value calculated by the above formula is... d The static interval length is 4 / w ; The fluctuation trend Δ in step 2.4 d The value range is (-2, 2).

[0030] Step 2.5: Combine the static symbols and dynamic fluctuation trend symbols at each moment into a two-dimensional joint symbol. u k =( s k , π k The size of the two-dimensional joint symbolic state space is... = w × e Statistical analysis of each joint symbol ( s i , π j The frequency of occurrence of ) yields the joint symbol probability. p ( s i , π j ): (6) Number(a) represents the number of times the combined symbol encoding 'a' appears; Step 2.6: Based on Shannon's information entropy theory and the calculated joint state probabilities... p ij The joint symbolic dynamic entropy JSFE is calculated using the following formula: (7) Where x represents the vibration time series, and m represents the phase space reconstruction dimension. e Represents static symbolic numbers, w P represents the number of signs of dynamic fluctuations. ij This represents the joint sign probability calculated by the above formula; Step 2.7: Based on the maximum scale quantity t max Construct coarse-grained sequences.

[0031] Step 2.7 is implemented according to the following steps: For a one-dimensional vibration time series x ( t The obtained first k Coarse-grained sequences y (τ) k ={ y (τ) k,1 , y (τ) k,2 , …}as follows: (9) in, N τ is the length of the vibration sequence; τ is the scale factor. x b The first vibration sequence b One value; Calculate the joint symbolic dynamic entropy at each scale, and obtain the following: t max Each entropy value is named the fine-grained composite multi-scale joint symbolic dynamics entropy, and the fine-grained composite multi-scale joint symbolic dynamics entropy at each scale is constructed into a feature vector. S : (10).

[0032] RCMJSFE denotes the fine composite multi-scale joint symbolic dynamic entropy, where x represents the vibration time series and m represents the phase space reconstruction dimension. e Represents static symbolic numbers, w The sign of the dynamic fluctuation is represented by τ, which is the scaling factor, and y is the number of signs. α (τ)This represents the coarse-grained sequence at scale τ obtained from the above equation.

[0033] Step 3: Construct and train the hydraulic instability diagnostic model; normalize the feature parameters extracted in Step 2 and use them as a sample set. Divide the sample set into a training set and a validation set according to a certain ratio. Input the training set into the random forest training module. Adjust the model parameters in real time through validation during the training process. Finally, the hydraulic instability diagnostic model is trained. In step 3, the hydraulic instability diagnosis model is a random forest model based on ensemble learning. This model constructs a feature vector from the refined composite multi-scale joint symbolic dynamic entropy feature parameters for each working condition calculated in step 2. S As input data for the random forest model, the coded symbols NC / IC / CD / CC corresponding to the operating state of each working condition are used as labels. Each input data and its corresponding label are used as training samples. After normalization, the training samples are input into the random forest model. Five-fold cross-validation and grid search are used to optimize the hyperparameters of the model. When the number of training iterations reaches the set maximum or the accuracy on the validation set no longer increases after multiple iterations, the random forest model is saved as the trained diagnostic model.

[0034] Step 4: Collect vibration signals in real time online using sensors, calculate the fine composite multi-scale joint symbolic dynamic entropy of the unknown signal, construct a feature vector, and input it into the trained hydraulic instability diagnostic model to diagnose the internal flow state of the unit.

[0035] Example 4 This invention relates to a method for diagnosing hydraulic instability in hydraulic machinery based on joint symbolic dynamic entropy, the flowchart of which is shown below. Figure 1 As shown, please follow these steps: Step 1: Collect vibration signals using vibration sensors installed externally to the system. x ( t ); Step 2: The vibration signals collected in Step 1 are preprocessed, and then the characteristic parameters characterizing the internal flow of the system are extracted using the proposed fine composite multi-scale joint symbolic dynamic entropy method. Step 3: Construct and train the hydraulic instability diagnostic model; normalize the feature parameters extracted in Step 2 and use them as a sample set. Divide the sample set into a training set and a validation set according to a certain ratio. Input the training set into the random forest training module. Adjust the model parameters in real time through validation during the training process. Finally, the hydraulic instability diagnostic model is trained. In step 3, the hydraulic instability diagnosis model is a random forest model based on ensemble learning. This model constructs a feature vector from the refined composite multi-scale joint symbolic dynamic entropy feature parameters for each working condition calculated in step 2.S As input data for the random forest model, the coded symbols NC / IC / CD / CC corresponding to the operating state of each working condition are used as labels. Each input data and its corresponding label are used as training samples. After normalization, the training samples are input into the random forest model. Five-fold cross-validation and grid search are used to optimize the hyperparameters of the model. When the number of training iterations reaches the set maximum or the accuracy on the validation set no longer increases after multiple iterations, the random forest model is saved as the trained diagnostic model.

[0036] Step 4: Collect vibration signals in real time online using sensors, calculate the fine composite multi-scale joint symbolic dynamic entropy of the unknown signal, construct a feature vector, and input it into the trained hydraulic instability diagnostic model to diagnose the internal flow state of the unit.

[0037] Example 5 Step 1: Collect vibration signals using vibration sensors installed outside the system. x ( t ) Specifically: such as Figure 3 As shown, by adjusting the power of the booster pump 2 and the screw compressor 7 to regulate the gas-liquid flow rate, the gas-liquid two-phase velocity ratio is continuously changed, realizing the evolution process of different flow patterns such as plug flow, slug flow, and foam wavy flow. Vibration signals are collected using a vibration acceleration sensor under different flow conditions. x ( t ).

[0038] In this example, vibration signals were acquired under conditions where the apparent liquid phase flow velocity was 0.6-1.6 m / s, and the gas phase flow velocity was adjustable within the range of 0-2.5 m / s. The vibration signal sampling frequency was 51200 Hz. The vibration signal acquisition instrument used in this example was a TEB120 series high-precision accelerometer, with a sampling frequency set to 10.24 kHz, a sensitivity of 100 mV / g, and a measurement error of less than 1% FS.

[0039] In this example, a total of 642 sets of vibration data for each of plug flow (PG), slug flow (SG), and foam wave flow (EW) were collected. The sampling time for each set of data was 0.32s. The training set and validation set were divided according to the 5-fold cross-validation method.

[0040] Step 2: After the vibration data acquisition is completed, the vibration sequence distribution is mean-processed according to the Laida criterion to remove abnormal data interference. Figures 4(a) to 4(c) show the time-domain waveforms of the vibration sequences under different flow patterns. Figure 4(a) shows the pipeline vibration signal under slug flow conditions, revealing a significant high-amplitude transient impact, corresponding to the strong impact of the liquid slug on the pipeline. Furthermore, the signal shows alternating low-amplitude and high-amplitude segments, reflecting the periodic flow structure of the liquid slug and gas spring in slug flow. Figure 4(b) shows the pipeline vibration signal under plug flow conditions, revealing similar vibration characteristics to slug flow conditions, with a significant high-amplitude transient impact, corresponding to the strong impact of the liquid slug on the pipeline. Figure 4(c) shows the pipeline vibration signal under emulsion wavy flow conditions, revealing that the vibration signal amplitude is at a moderate level with no significant abrupt changes, reflecting the structural characteristics of uniform emulsification of gas and liquid with small bubbles. In summary, it is difficult to distinguish between slug flow and plug flow, two intermittent flow patterns, based solely on time-domain characteristics. Subsequently, the proposed refined composite multi-scale joint symbolic kinetic entropy method is used to extract characteristic parameters that can characterize the internal flow of the system. The specific implementation steps are as follows: Based on Takens' delayed embedding theorem, the vibration signal sequence is reconstructed into a high-dimensional phase space. Z ( m ).

[0041] (1) in m This represents the embedding dimension of the phase space. Z i ( m ) indicates the first i A phase space vector.

[0042] Calculate the cosine similarity between adjacent trajectories in phase space. d ( z i ( m ), z i+1 ( m ), to obtain the similarity sequence D ( m )={ d ( z 1( m ), z 2( m )), d ( z 2( m ), z 3( m )),…, d ( z N-m+1 ( m ), zN-m ( m ))} is used to quantify the local pattern similarity of the system's dynamic state.

[0043] (2) The cosine similarity value in the formula ranges from (-1, 1).

[0044] The calculated cosine similarity sequence between trajectories is converted into a symbol sequence. S = { s 1, s 2, …, s N-m Let the static symbol number be... e The range of cosine similarity values ​​[-1, 1] is divided into... e There are equidistant intervals, each interval having a length Δ. s =2 / e , denoted as ( I 1, I 2, …, I ε The cosine similarity sequence is transformed into a symbol sequence using the following formula. Then, the probability that the cosine similarity falls between each cell is calculated and denoted as the state probability. p 1, p 2, …, p ε ).

[0045] (3) Performing a first-order difference on the cosine similarity sequence between orbits yields the similarity fluctuation trend Δ. d This is used to quantify the dynamic rate of change of similarity.

[0046] (4) Fluctuation trend Δ d The value range of is (-2, 2). Let the number of dynamic fluctuation symbols be . w The trend of change Δ d The range is divided into w There are equidistant intervals, each interval having a length of . d =4 / w The following formula transforms the fluctuation trend sequence into a symbolic sequence. T = { π 1, π 2, …, π N-m}

[0047] (5) The static symbol and dynamic fluctuation trend symbol at each moment are combined into a two-dimensional joint symbol.u k =( s k , π k The size of the two-dimensional joint symbolic state space is = w × e Statistical analysis of each joint symbol ( s i , t j The frequency of occurrence of ) yields the joint probability: (6) Based on Shannon's information entropy theory and the calculated joint state probability p ij The joint symbolic dynamic entropy JSFE is calculated using the following formula: (7) (8) Based on the maximum scale number t max Construct coarse-grained sequences. For time series... x ( t The obtained first k Coarse-grained sequences y (τ) k ={ y (τ) k,1 , y (τ) k,2 , …}as follows: (9) Calculate the joint symbolic dynamic entropy at each scale, and obtain the following: t max Each entropy value is used to construct a feature vector: (10) Step 3: Normalize the feature vectors calculated in Step 2 and use them as a sample set. Divide the sample set into a training set and a validation set using the five-fold cross-validation method. Input the training set into the random forest training module. During the training process, adjust the model parameters in real time through validation. When the number of training iterations reaches the set maximum or the accuracy on the validation set no longer increases after multiple iterations, save the random forest model as the diagnostic model after training.

[0048] In this example, the number of ensemble trees is set to 100, and the other parameters remain at their default settings.

[0049] Step 4: Real-time online acquisition of vibration signals by sensors, calculation of the fine composite multi-scale joint symbolic dynamic entropy of unknown signals, construction of feature vectors and input into the trained hydraulic instability diagnostic model for diagnosis of the internal flow state of the unit.

[0050] In this example, after the diagnostic model was trained, a total of 276 sets of vibration test data for slug flow, plug flow, and foam wave flow were collected, with each set of data having a sampling length of 0.32 s. The refined composite multi-scale joint symbolic dynamic entropy of the test data was calculated, and the resulting feature vector was input into the trained hydraulic instability diagnostic model. The results are shown below. Figure 5 As shown, the test results show that the accuracy rate for each flow pattern state diagnosis is 95.8%; the accuracy rates for identifying slug flow, foamy wavy flow, and slug flow are 95.3%, 97.1%, and 94.9%, respectively; the recall rates for identifying slug flow, foamy wavy flow, and slug flow are 96.4%, 96.7%, and 94.2%, respectively; and the F1 scores for identifying slug flow, foamy wavy flow, and slug flow are 95.8%, 96.9%, and 94.5%, respectively.

[0051] To further illustrate the superiority of the method of this invention, the diagnostic results of several other methods, such as the Refined Composite Multiscale Spread Entropy (RCMDE), were also calculated and compared, as shown in Table 1. To avoid the randomness of the test results, the data were tested 10 times under each working condition, and the mean and standard deviation were calculated.

[0052] Table 1 Performance Comparison of Example 5 with the Same Method

[0053] Table 1 shows that the accuracy, precision, recall, and average F1 score of the proposed hydraulic instability diagnosis method are all above 95%. Compared to the traditional entropy algorithm combined with a random forest model, which generally achieves an accuracy of around 92%, this invention improves the diagnostic accuracy by more than 3%. Among traditional methods, the refined composite multiscale wave spread entropy (RCMFDE) combined with a random forest model has the best precision and recall, around 92.3%, but this is still more than 2.5% lower than the method proposed in this invention, fully demonstrating the effectiveness and superiority of the method proposed in this patent.

[0054] Example 6 Step 1: Collect vibration signals using vibration sensors installed outside the system. x ( t ) Specifically, taking a prototype axial-flow turbine with 5 rotor blades and 24 guide vanes as an example, such as... Figure 6As shown, the axial-flow turbine test bench is started. Water is drawn from the water tank by the booster centrifugal pump 1 and gradually flows through the axial-flow turbine 4 and the low-pressure water tank 3 from the high-pressure water tank 5. This puts the turbine in a cavitation-free state. A laser vibrometer is used to collect the time series of vibration signals of the turbine runner. The cavitation coefficient of the axial-flow turbine is continuously reduced to realize the evolution of the cavitation state in the runner chamber.

[0055] In this example, the vibration signal sampling frequency was set to 10.24 kHz, and the measurement resolution was 0.02 μm / (s). Hz 0.5 The measurement error is less than 1% FS.

[0056] The turbine operating parameters for the experimental conditions in this example are shown in Table 2: Table 2 Operating parameters of axial-flow turbine under Example 6

[0057] In this example, a total of 216 sets of vibration data were collected for the non-cavitation state (NC) (cavitation coefficients of 1.63, 1.51, 1.41, and 1.31), 54 sets of vibration data for the initial cavitation state (IC) (cavitation coefficient of 1.21), 54 sets of vibration data for the cavitation development state (CD) (cavitation coefficient of 1.11), and 108 sets of vibration data for the fully cavitation state (OC) (cavitation coefficients of 1.01 and 0.91). Each set of data has a length of 4096. The training set and validation set were divided according to the 5-fold cross-validation method.

[0058] Step 2: After the vibration data acquisition is completed, the vibration sequence distribution is mean-processed according to the Laida criterion to remove abnormal data interference. Figures 7(a) to 7(d) show the time-domain waveforms of the rotor vibration sequence under various cavitation states. Figure 7(a) shows the rotor vibration signal in the non-cavitation state, where the vibration signal amplitude is relatively small. Figure 7(b) shows the rotor vibration signal in the initial cavitation state, where a large number of random, high-frequency fine burrs appear, representing the microscopic impacts generated when cavitation bubbles collapse on the blade surface. As the cavitation state continues to develop, the vibration signal amplitude in Figures 7(c) and 7(d) increases rapidly, and the signal waveform becomes very irregular, with the original periodicity being submerged by a large number of dense random impacts. It can be seen that a quantitative index to describe the system flow characteristics cannot be obtained solely from the time-domain waveform analysis. Subsequently, the proposed refined composite multi-scale joint symbolic dynamic entropy method is used to extract characteristic parameters that characterize the internal flow of the system. The specific implementation steps are as follows: Based on Takens' delayed embedding theorem, the vibration signal sequence is reconstructed into a high-dimensional phase space. Z ( m ).

[0059] (1) in m This represents the embedding dimension of the phase space. Z i ( m ) indicates the first i A phase space vector.

[0060] Calculate the cosine similarity between adjacent trajectories in phase space. d ( z i ( m ), z i+1 ( m ), to obtain the similarity sequence D ( m )={ d ( z 1( m ), z 2( m )), d ( z 2( m ), z 3( m )),…, d ( z N-m+1 ( m ), z N-m ( m ))} is used to quantify the local pattern similarity of the system's dynamic state.

[0061] (2) The cosine similarity value in the formula ranges from (-1, 1).

[0062] The calculated cosine similarity sequence between trajectories is converted into a symbol sequence. S = { s 1, s 2, …, s N-m Let the static symbol number be... e The range of cosine similarity values ​​[-1, 1] is divided into... e There are equidistant intervals, each interval having a length Δ. s =2 / e , denoted as ( I 1, I 2, …, I ε The cosine similarity sequence is transformed into a symbol sequence using the following formula. Then, the probability that the cosine similarity falls between each cell is calculated and denoted as the state probability. p 1,p 2, …, p ε ).

[0063] (3) Performing a first-order difference on the cosine similarity sequence between orbits yields the similarity fluctuation trend Δ. d This is used to quantify the dynamic rate of change of similarity.

[0064] (4) Fluctuation trend Δ d The value range of is (-2, 2). Let the number of dynamic fluctuation symbols be . w The trend of change Δ d The range is divided into w There are equidistant intervals, each interval having a length of . d =4 / w The following formula transforms the fluctuation trend sequence into a symbolic sequence. T = { π 1, π 2, …, π N-m}

[0065] (5) The static symbol and dynamic fluctuation trend symbol at each moment are combined into a two-dimensional joint symbol. u k =( s k , π k The size of the two-dimensional joint symbolic state space is = w × e Statistical analysis of each joint symbol ( s i , t j The frequency of occurrence of ) yields the joint probability: (6) Based on Shannon's information entropy theory and the calculated joint state probability p ij The joint symbolic dynamic entropy JSFE is calculated using the following formula: (7) (8) Based on the maximum scale number t max Construct coarse-grained sequences. For time series... x ( t The obtained firstk Coarse-grained sequences y (τ) k ={ y (τ) k,1 , y (τ) k,2 , …}as follows: (9) Calculate the joint symbolic dynamic entropy at each scale, and obtain the following: t max Each entropy value is used to construct a feature vector: (10) Step 3: Normalize the feature vectors calculated in Step 2 and use them as a sample set. Divide the sample set into a training set and a validation set using the five-fold cross-validation method. Input the training set into the random forest training module. During the training process, adjust the model parameters in real time through validation. When the number of training iterations reaches the set maximum or the accuracy on the validation set no longer increases after multiple iterations, save the random forest model as the diagnostic model after training.

[0066] In this example, the number of ensemble trees is set to 100, and the other parameters remain at their default settings.

[0067] Step 4: Real-time online acquisition of vibration signals by sensors, calculation of the fine composite multi-scale joint symbolic dynamic entropy of unknown signals, construction of feature vectors and input into the trained hydraulic instability diagnostic model for diagnosis of the internal flow state of the unit.

[0068] In this example, after the diagnostic model was trained, a total of 92 sets of vibration test data were collected in the non-cavitation state, 23 sets in the initial cavitation state, 23 sets in the cavitation development state, and 46 sets in the fully cavitation state, with each set of data being 4096 bytes long. The refined composite multi-scale joint symbolic dynamic entropy of the test data was calculated, and the resulting feature vector was input into the trained hydraulic instability diagnostic model. The results are shown below. Figure 8 As shown, the test results show that the diagnostic accuracy for each cavitation state is 97.3%; the accuracy for identifying the absence of cavitation and the cavitation development state is 95.8%, and the accuracy for identifying initial cavitation and complete cavitation is 100%; the recall rate for identifying the absence of cavitation and the cavitation development state is 100%, the recall rate for identifying initial cavitation is 91.3%, and the recall rate for identifying complete cavitation is 93.5%; the F1 score for identifying the absence of cavitation and the cavitation development state is 97.9%, the F1 score for identifying initial cavitation is 95.5%, and the F1 score for identifying complete cavitation is 96.6%.

[0069] To further illustrate the superiority of the method of this invention, the diagnostic results of several other methods, such as Refined Composite Multiscale Spread Entropy (RCMDE), were also calculated and compared, as shown in Table 3. Since cavitation data is limited, accuracy alone cannot be used to evaluate this type of imbalanced sample dataset. Therefore, in addition to accuracy, precision, recall, and F1 score were also calculated.

[0070] Table 3 Performance Comparison of Different Methods in Example 6

[0071] Table 3 shows that the proposed hydraulic instability diagnosis method achieves an average accuracy, precision, recall, and F1 score of over 93%. This is significantly higher than the traditional entropy algorithm combined with a random forest model, which typically achieves around 90% accuracy. The accuracy of the refined composite multiscale symbolic dynamic entropy (RCMSDE) combined with a random forest model is only 83.3%. Therefore, the accuracy of this invention is more than 5% higher than traditional methods. Among traditional methods, the refined composite multiscale wave spread entropy (RCMFDE) combined with a random forest model achieves the best precision and recall at around 87%, but this is still more than 7% lower than the method proposed in this invention, fully demonstrating the effectiveness and superiority of the proposed method.

[0072] The operating environment of generator units is usually complex. In order to further verify the noise resistance performance of the method of the present invention, the diagnostic performance of different methods under low signal-to-noise ratio (0dB) was calculated and compared. The results are shown in Table 4.

[0073] Table 4. Performance comparison of different methods under low signal-to-noise ratio in Example 6.

[0074] Table 4 shows that the proposed hydraulic instability diagnosis method maintains an accuracy of approximately 89%-90% even under low signal-to-noise ratio conditions, while traditional methods have an accuracy below 80%. The average precision, recall, and F1 score are all above 85%. Compared to the traditional entropy algorithm combined with a random forest model, which typically achieves only 60-70% accuracy in hydraulic instability diagnosis, this invention generally improves diagnostic performance by more than 10% under low signal-to-noise ratio conditions. This fully demonstrates the superior noise resistance of the proposed method.

Claims

1. A method for diagnosing hydraulic instability in hydraulic machinery based on joint symbol dynamics entropy, characterized by, Specifically, the following steps are implemented: Step 1, collect vibration signal by vibration sensor installed outside the system x ( t ) Step 2, pre-processing the vibration signal collected in step 1, extracting characteristic parameters representing the internal flow of the system; Step 3, constructing and training a hydraulic instability diagnosis model; Step 4, real-time online acquisition of vibration signals, calculation of fine composite multi-scale joint symbolic dynamics entropy of unknown signals, construction of feature vectors input into the trained hydraulic instability diagnosis model for internal flow state diagnosis of the unit.

2. The joint symbol dynamics entropy based hydraulic machinery hydraulic instability diagnostic method according to claim 1, characterized in that, In step 2, after the vibration data collection is completed, the mean value of the vibration sequence distribution is processed according to the Lyapunov criterion, and the abnormal data interference is removed.

3. The joint symbol dynamics entropy based hydraulic machinery hydraulic instability diagnostic method of claim 2, wherein, The step 2 is implemented according to the following steps: Step 2.1, for the vibration signal x ( t ), x ( t ) is a one-dimensional vibration time series, the vibration signal x ( t ) is reconstructed to a high-dimensional phase space Z ( m ) based on the Takens delay embedding theorem, (1) wherein m represents the phase space embedding dimension, N represents the length of the one-dimensional vibration time series, z i ( m ) represents the i th phase space vector, i =1,2,…,N-m+1, x 1, x 2,…, x N represent the 1st, 2nd, …, Nth data values of the one-dimensional vibration time series; Step 2.

2. Calculate the cosine similarity between adjacent trajectories in the phase space d z i m z i+1 m ) to get the similarity sequence D m = { d z 1 m z 2 m ) d z 2 m z 3 m ), d z N-m m z N-m+1 m ) z i m and z i m are the i-th and j-th vectors in the phase space, i.e. the i-th and j-th trajectories, quantifying the local pattern similarity of the dynamical state of the system,​​​​​​​​​​​​​​ (2) wherein the cosine similarity d j = d z i m z j m ranges from (-1, 1);​​​​ Step 2.

3. Transform the cosine similarity sequence between trajectories into a symbol sequence S = { s 1, s 2, …, s N-m}, let the static symbol number be Epsilon , divide the value range [-1, 1] of the cosine similarity into Epsilon equal intervals, and the length of each interval is s =2 / Epsilon , denoted as I 1, I 2, …, I ε , transform the cosine similarity sequence into a symbol sequence by the following formula, and then count the probability of the cosine similarity falling in each small interval, denoted as the state probability p 1, p 2, …, p ε , (3) s 1, s 2, …, s ε respectively Epsilon A static symbol marker, d k is the kth cosine similarity value calculated by the above formula, Δ s is the length of the static interval interval 2 / Epsilon ; Step 2.

4. First-order difference of the cosine similarity sequence between trajectories, to get the fluctuation trend Δ d quantify the dynamic change rate of similarity: (4) Δ d k is the kth first-order difference cosine similarity value calculated; d k+1 and d k are the k+1th and kth cosine similarity values calculated by the above formula; let the number of dynamic fluctuation symbols be Omega , the change trend Δ d is divided into Omega equidistant intervals, and the length of each dynamic interval is Delta =4 / Omega , and the fluctuation trend sequence is converted into a symbol sequence by the following formula T = { π 1, π 2, …, π N-m} : (5) π 1, π 2, …, π ω respectively Omega A dynamic symbol marker, Δ d is a first-order differential cosine similarity value calculated by the above formula, Delta is a static interval interval length 4 / Omega ; Step 2.5, combine the static symbol and dynamic fluctuation trend symbol at each time into a two-dimensional joint symbol u k ( s k , π k ) and the size of the two-dimensional joint symbol state space is = Omega × Epsilon , count the frequency of each joint symbol ( s i , π j ) and obtain the joint symbol probability p ( s i , π j ): (6) Number (a) represents the number of times a joint symbolic code appears; Step 2.

6. Calculate the joint state probabilities from the Shannon information entropy theory and the calculated joint state probabilities p ij The joint symbol dynamics entropy JSFE is calculated by the following equation: (7) where x denotes the vibration time series, m denotes the phase space reconstruction dimension, Epsilon denotes the static symbol number, Omega denotes the dynamic fluctuation symbol number, P ij denotes the joint symbol probability calculated in the above equation; Step 2.7, according to the maximum number of dimensions Tau max Constructing a coarsened sequence.

4. The joint symbol dynamics entropy based hydraulic machinery hydraulic instability diagnostic method of claim 3, wherein, The fluctuation trend Δ in step 2.4 d is in the range (-2, 2).

5. The joint symbol dynamics entropy based hydraulic machinery hydraulic instability diagnostic method of claim 3, wherein, The step 2.7 is implemented according to the following steps: For a one-dimensional vibration time series x ( t ), the resulting first k coarse-grained sequence y (τ) k = y (τ) k,1 , y (τ) k,2 ,…} is as follows: (9) wherein, N is the vibration sequence length; τ is a scale factor, x b denotes the value of the vibration sequence at the b th position. The joint symbolic dynamics entropy of each scale is calculated, and a total of Tau max entropy values are obtained, which are named fine composite multi-scale joint symbolic dynamics entropy. The fine composite multi-scale joint symbolic dynamics entropy of each scale is constructed into a feature vector S : (10)； RCMJSFE represents refined composite multiscale joint symbolic dynamics entropy, x represents a vibration time series, m represents a phase space reconstruction dimension, Epsilon represents static symbols, Omega represents dynamic fluctuation symbols, τ is a scale factor, y α (τ) represents a coarse-grained sequence at scale τ obtained from the above equation.

6. The joint symbol dynamics entropy based hydraulic machinery hydraulic instability diagnostic method of claim 5, wherein, The hydraulic instability diagnosis model in the step 3 is a random forest model based on the integrated learning idea, and a feature vector is constructed by using each fine composite multi-scale joint symbolic dynamics entropy feature parameter calculated in the step 2 S As the input data of the random forest model, the encoding symbol NC / IC / CD / CC corresponding to the running state of each working condition is used as a label, each input data and the corresponding label are used as a training sample, and the random forest model is input after normalization processing of the training sample.

7. The joint symbol dynamics entropy based hydraulic machinery hydraulic instability diagnostic method of claim 6, wherein, In step 3, five-fold cross-validation and grid search are used to optimize the hyperparameters of the random forest model. When the training times reach the set maximum number or the accuracy on the validation set does not increase after multiple iterations, the random forest model is saved as the trained diagnosis model.

8. The joint symbol dynamics entropy based hydraulic machinery hydraulic instability diagnostic method of claim 7, wherein, The step 4 is as follows: Based on the real-time vibration signals of the unknown flow state in the unit collected by the sensor, the fine composite multi-scale joint symbol dynamics entropy feature parameters of the real-time vibration signals in each scale are calculated, and the feature vector S corresponding to the unknown working condition is constructed The feature vector S corresponding to the unknown working condition is obtained The flow state label corresponding to the unknown working condition is obtained by inputting the trained hydraulic machinery hydraulic instability diagnosis model in step 3, and the internal flow state of the unit at this time is diagnosed according to the flow state label; if the label corresponds to an unstable flow working condition, an early warning notice is sent to the power plant operator for corresponding subsequent operation.

Images