Fan bearing fault positioning method based on self-adaptive voiceprint sensor network

Abstract

The invention discloses a fan bearing fault positioning method based on a self-adaptive voiceprint sensor network. The fan bearing fault positioning method comprises the following steps: S1, constructing a fan bearing acoustic signal set; s2, dynamically adjusting the fan bearing acoustic signal set by using a self-adaptive voiceprint sensor network; s3, preprocessing the fan bearing acoustic signal set; s4, constructing a fan bearing acoustic feature set; s5, performing fault identification on the fused fan bearing acoustic feature set by using a semi-supervised machine learning model; s6, further utilizing an integrated learning model to recheck and further classify the fault signals identified in the step S5; and S7, outputting the fan bearing fault positioning result determined in the step S6 through an alarm system, including the fault type, the occurrence position and the severity of the fault, and notifying maintenance personnel to take corresponding maintenance measures according to the fault positioning result. According to the invention, the accuracy and precision of fault classification are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fans, and in particular to a method for locating a fan bearing fault based on an adaptive soundprint sensor network. Background Art

[0002] In the prior art, fault detection of fan bearings mainly relies on regular manual inspections or vibration analysis technology based on a single sensor. The existing method requires operation and maintenance personnel to regularly manually inspect fan bearings or monitor them through vibration sensors installed near the bearings. Such traditional methods are inefficient and easily interfered by environmental noise and complex working conditions. It is often difficult to capture subtle fault signals in the early stages of a fault. The prior art usually uses a single type of sensor for data acquisition, which makes it impossible to fully and comprehensively obtain the operating status of the fan bearings, resulting in insufficient data integrity and accuracy, further affecting the timeliness and accuracy of fault detection.

[0003] In addition to the problem of detection accuracy, most of the existing fan bearing fault location technologies rely on post-analysis, that is, troubleshooting and repair are carried out only after the equipment has obvious faults. The fault location response time is long, which not only causes long-term equipment downtime and increases maintenance costs, but may also cause serious equipment damage and even affect the operating stability of the entire fan system. Traditional data analysis methods based on single-point sensors are also unable to fully utilize multi-dimensional data for comprehensive evaluation, resulting in poor robustness of fault location results. The determination of fault type and location under complex working conditions is often not accurate enough.

[0004] There is another problem with the existing technology, which is the lack of effective intelligent analysis methods when processing large amounts of real-time monitoring data. The current fault detection system usually only relies on simple signal processing technology and traditional machine learning methods, which cannot flexibly deal with the interference of data noise and redundant information, and it is difficult to adjust the sensor's acquisition strategy and analysis model in real time according to changes in the wind turbine operating environment, further limiting the accuracy and real-time performance of fault detection and positioning.

[0005] In summary, the existing technology has the following main shortcomings: first, the fault detection accuracy is low, especially in the early stage of the fault, it is difficult to accurately capture the abnormal signal; second, the fault location response is slow, resulting in long equipment downtime, increasing maintenance costs and risks; finally, the system lacks intelligence and adaptive capabilities, and cannot dynamically adjust the data collection and analysis strategies according to changes in the wind turbine operating environment, which limits the accuracy and real-time performance of fault detection and location. Summary of the invention

[0006] One object of the present invention is to provide a method for locating a fan bearing fault based on an adaptive soundprint sensor network, which improves the accuracy and precision of fault classification.

[0007] A method for locating a fan bearing fault based on an adaptive voiceprint sensor network according to an embodiment of the present invention comprises the following steps:

[0008] S1. Multiple adaptive soundprint sensor nodes are arranged at multiple key positions of the fan to collect vibration frequency, noise signal and other acoustic characteristic signals related to bearing faults during the operation of the fan bearing, and construct a fan bearing acoustic signal set;

[0009] S2. Dynamically adjust the wind turbine bearing acoustic signal set using an adaptive soundprint sensor network. The adaptive soundprint sensor network dynamically adjusts the sensitivity, operating frequency and acquisition mode of the adaptive soundprint sensor node according to the wind turbine's operating environment, wind turbine load changes, environmental noise and temperature changes.

[0010] S3, preprocessing the fan bearing acoustic signal set, removing environmental noise and redundant information in the fan bearing acoustic signal set, and retaining effective acoustic signal features related to the fan bearing operating state;

[0011] S4. Use signal processing technology to extract features from the preprocessed fan bearing acoustic signal set, extract the vibration features, energy features and abnormal noise features in the fan bearing acoustic signal, summarize the vibration features, energy features and abnormal noise features of each sensor node, eliminate redundant data and noise interference, retain the effective acoustic signal feature data related to the fan bearing operating status, and construct the fan bearing acoustic feature set;

[0012] S5. Use the semi-supervised machine learning model to identify the faults of the fused acoustic feature set of the fan bearing. By training the semi-supervised machine learning model, the vibration features, energy features and abnormal noise features are extracted and analyzed from the acoustic feature set of the fan bearing, and the potential fault signals of the fan bearing are preliminarily identified. The fault signals are classified into bearing wear fault, crack fault and fatigue fault.

[0013] S6. Further review and further classify the fault signals identified in S5 using the integrated learning model, optimize the determination of the fault type through comprehensive evaluation of vibration characteristics, energy characteristics and abnormal noise characteristics, and prioritize the faults according to their severity;

[0014] S7. Output the fan bearing fault location result determined in S6 through the alarm system, including the fault type, location and severity of the fault, and notify the maintenance personnel to take corresponding maintenance measures according to the fault location result.

[0015] Optionally, the S1 includes the following steps:

[0016] S11. Arrange multiple adaptive soundprint sensor nodes N1, N2, ..., N at key locations of the fan. n , where n is the number of sensor nodes. The sensor nodes are used to collect acoustic signals generated during the operation of the fan bearing, including the vibration frequency f v , noise signal S n and abnormal acoustic signals A related to bearing failure t ;

[0017] S12, each sensor node N i During the operation of the bearing, the vibration frequency f is collected according to different load conditions and working conditions. v,i (t), noise signal S n,i (t) and abnormal acoustic signal A t,i (t), where t is time;

[0018] S13, based on the collected vibration frequency f v,i (t), noise signal S n,i (t) and abnormal acoustic signal A t,i (t) Construct the fan bearing acoustic signal set D s :

[0019] D s ={(f v,1 (t),S n,1 (t),A t,1 (t)),(f v,2 (t),S n,2 (t),A t,2 (t)),…,(f v,n (t),S n,n (t),A t,n (t))}.

[0020] Optionally, S3 includes the following steps:

[0021] S31. Obtaining the fan bearing acoustic signal set D s , for the noise N in the fan operating environment e (t) is modeled, and the environmental noise characteristics are extracted using frequency domain filtering technology, and the noise removal function H(N e (t)) performing noise elimination processing on the acoustic signal;

[0022] S33, removing redundant data from the processed acoustic signal, using a redundant removal algorithm to remove repeated or irrelevant acoustic data, and obtaining a valid fan bearing acoustic signal set D after cleaning. clean ;

[0023] S34, extracting effective acoustic signal features related to the operating status of the fan bearing, including the vibration frequency f v,i (t), the noise signal after cleaning S clean,i (t) and abnormal acoustic signal A t,i (t), integrated into the pre-processed fan bearing acoustic signal set D proc :

[0024] D proc ={f v,i (t),S clean,i (t),A t,i (t)}.

[0025] Optionally, S4 includes the following steps:

[0026] S41, the pre-processed fan bearing acoustic signal set D proc Vibration feature extraction is performed, and the intrinsic mode function and instantaneous frequency characteristics of the bearing vibration signal are extracted using empirical mode decomposition and Hilbert-Huang transform technology:

[0027] The vibration signal f collected by each sensor node v,i (t) Perform empirical mode decomposition and decompose it into several intrinsic mode functions IMF k,i (t) and the residual term r i (t):

[0028]

[0029] Among them, IMF k,i (t) is the kth order eigenmode function, K is the total number of modal components, r i (t) is the residual term;

[0030] For each intrinsic mode function IMF k,i (t) Perform Hilbert transform to obtain the instantaneous amplitude A k,i (t) and instantaneous frequency ω k,i (t):

[0031] IMF k,i (t) = A k,i (t)cos(∫ω k,i (t)dt);

[0032] Among them, A k,i (t) reflects the amplitude change of the vibration signal at different times, ω k,i (t) represents the instantaneous frequency, which can reflect the changes in the vibration characteristics of the bearing during operation;

[0033] The instantaneous amplitude and instantaneous frequency characteristics of all eigenmode functions constitute the vibration characteristic vector Vi , vibration eigenvector V i Describe the vibration state of the fan bearing at different times and frequencies:

[0034] V i ={(A 1,i (t),ω 1,i (t)),(A 2,i (t),ω 2,i (t)),…,(A K,i (t),ω K,i (t))};

[0035] S42, noise signal S clean,i (t) Perform wavelet packet decomposition to obtain decomposition coefficients W at different scales and positions j,l,i :

[0036]

[0037] Among them, J is the number of decomposition levels, l is the node position, ψ j,l (t) is the corresponding mother wavelet function;

[0038] Calculate the energy characteristics E of the frequency band corresponding to each decomposition coefficient j,l,i , energy characteristic E j,l,i Reflects the energy distribution of noise signals in different frequency bands:

[0039]

[0040] Construct energy eigenvector E i Normalize the energy features:

[0041]

[0042] S43. Abnormal acoustic signal A t,i (t) Extract abnormal noise features and calculate abnormal acoustic signal A t,i The bispectrum B of (t) i (f1,f2):

[0043]

[0044] in, is the third-order autocorrelation function, E[·] is the mathematical expectation;

[0045] Extract the amplitude and phase characteristics of the bispectrum and the bispectral entropy H i :

[0046]

[0047] in, is the bispectral energy density function, and the bispectral entropy H i It reflects the non-Gaussianity and nonlinearity of the signal;

[0048] Construct an abnormal noise feature vector B containing the bispectral amplitude, phase and entropy characteristics of the abnormal noise signal i :

[0049] B i ={|B i (f1,f2)|,∠B i (f1,f2),H i};

[0050] S44, extract the vibration feature vector V i , energy eigenvector E i and abnormal noise feature vector B i Integration, construct the acoustic feature vector set D of the fan bearing feat :

[0051] D feat = {V i ,E i ,B i}.

[0052] Optionally, S5 includes the following steps:

[0053] S51, the fan bearing acoustic feature set D feat Divided into labeled dataset D l ={(X l ,Y l )},X l is the feature vector of the labeled sample, Y l are the known fault type labels, including bearing wear fault, crack fault and fatigue fault, and the unlabeled dataset D u ={X u}, X u is the feature vector of the unlabeled sample;

[0054] S52. Build a semi-supervised machine learning model M semi , the similarity matrix W between samples is initially calculated using the graph-based label propagation algorithm:

[0055]

[0056] Among them, W ij Represents sample X i and sample X j The similarity between ∥X i -X j ∥ is the Euclidean distance between feature vectors, σ is the bandwidth parameter, which controls the decay speed of similarity;

[0057] Normalize the similarity matrix W to obtain the normalized Laplace matrix S;

[0058] S53, initialize the label matrix Y, the labels of the labeled samples remain unchanged, and the labels of the unlabeled samples are set to zero vectors:

[0059]

[0060] S55. Iteratively update the label matrix F through the label propagation algorithm:

[0061] F (t+1) =αSF (t) +(1-α)Y;

[0062] Among them, F (t) is the label matrix of the tth iteration, α is the smoothing parameter, 0<α<1, which controls the trade-off between new and old labels;

[0063] The iteration process continues until the convergence condition is met:

[0064] ∥F (t+1) -F (t) ∥ F <ε;

[0065] Among them, ∥·∥ F is the Frobenius norm, ε is the preset convergence threshold;

[0066] S56. When the iteration converges, use the final label matrix F * Predict the fault type and label of unlabeled samples Determined by the category corresponding to the maximum value, the prediction confidence C u It is given by the normalized values ​​of the label matrix:

[0067]

[0068] in, For sample The label value in the kth class;

[0069] S57, set the confidence threshold θ, and Unlabeled samples above the threshold and its predicted label Add labeled datasets and update labeled datasets:

[0070]

[0071] S58. Using the updated labeled dataset D′ l Semi-supervised model M semiRetrain and repeat steps S52 to S57 until all unlabeled samples are classified or the maximum number of iterations T is reached. max ;

[0072] S59, finally the optimized semi-supervised fault identification model The acoustic feature set D of the fan bearing is analyzed by using the semi-supervised fault recognition model. feat Perform fault identification and preliminarily identify potential fault signals of fan bearings And the fault signals are classified into bearing wear fault, crack fault and fatigue fault.

[0073] Optionally, the S6 includes the following steps:

[0074] S61, the potential fault signal initially identified Input ensemble learning model M ensemble Review, the integrated learning model includes multiple basic models M j An integrated framework composed of each basic model independently predicts the fault type;

[0075] S62, for each basic model M j For training, the basic model includes the random forest model M RF , Gradient Boosting Tree Model M GB and the support vector machine model M SVM , each model input vibration characteristic V i , Energy characteristics E i and abnormal noise characteristics B i , and output the corresponding fault type prediction results

[0076] S63. Prediction results of each basic model Perform voting fusion and ensemble learning model’s final prediction results Determined by voting results:

[0077]

[0078] in, is the indicator function, when It takes 1 when it is category k, otherwise it takes 0;

[0079] S64. Final prediction results based on ensemble learning model The fault types are further classified and prioritized according to the severity of the fault signal, which is determined by analyzing the amplitude A of the vibration signature. k,i (t) and the bispectral entropy H of the abnormal noise characteristics i To conduct an assessment:

[0080]

[0081] Among them, P i is the fault priority score, β1 and β2 are weight coefficients, reflecting the influence of vibration amplitude and bispectral entropy on fault severity;

[0082] S65. Score P according to fault priority i The fault types are sorted and divided into three priority levels: high, medium and low. The high-priority fault types will be processed first and alarms will be output, while the low-priority fault types will be recorded as monitoring items.

[0083] Optionally, the S7 includes the following steps:

[0084] S71. Identify and classify the fault types and the fault priority score P i As input, generate wind turbine bearing fault location results, including fault type, fault location, and fault severity;

[0085] S72, based on fault type and the fault priority score P i , generate alarm signal A through preset alarm rules i , the alarm signal is related to the severity of the fault, and the alarm threshold θ is set a , when P i When the value is higher than the alarm threshold, a high priority alarm is generated. high , otherwise a medium or low priority alarm is generated A mid or A low :

[0086]

[0087] Among them, θ a and θ b These are the thresholds for high, medium, and low priority alarms, respectively;

[0088] S73, outputting the complete fault location result through the alarm system and sending the fault information to the maintenance personnel according to the priority of the alarm signal;

[0089] S74, when receiving a high priority alarm signal A high When the system sends an emergency maintenance notice to the maintenance personnel, the content includes the specific fault type, location and severity of the fan bearing, and requires immediate maintenance measures. mid and A low Reminder notifications will be sent according to the preset schedule and regular inspections will be arranged.

[0090] The beneficial effects of the present invention are:

[0091] (1) The adaptive voiceprint sensor network in the present invention can adaptively adjust according to multiple factors such as the fan operating environment, load changes, environmental noise and temperature changes by dynamically adjusting the sensitivity, operating frequency and acquisition mode of the sensor nodes, effectively solving the problem that traditional fixed sensors cannot cope with inaccurate data acquisition under complex working conditions, so that the collected acoustic signals can accurately reflect the actual operating status of the fan bearing, thereby greatly improving the accuracy of fault detection. The dynamic adjustment mechanism ensures that the sensor can still maintain efficient monitoring even in harsh environments, avoiding the problem of missed detection or false detection caused by interference from environmental factors in traditional methods.

[0092] (2) The present invention introduces a semi-supervised machine learning model, which can significantly improve the fault recognition performance of the model by utilizing unlabeled data with a small amount of labeled data. By constructing an acoustic feature set and training it using a semi-supervised learning method, the model can adaptively extract and analyze the vibration characteristics, energy characteristics and abnormal noise characteristics in the acoustic signal of the fan bearing and accurately identify the wear, crack and fatigue fault types of the bearing. Compared with traditional supervised learning methods, semi-supervised learning not only improves the accuracy of fault recognition, but also effectively reduces the dependence on large-scale labeled data and improves the robustness of the system under different working conditions.

[0093] (3) The present invention constructs an integrated learning model by integrating multiple basic models such as random forest, gradient boosting tree and support vector machine to review and further classify the initially identified fault signals. The integrated model comprehensively evaluates the characteristics of different faults. In terms of the priority classification of fault severity, the fault type is optimized by the vibration amplitude and bispectral entropy, and the fault is prioritized according to its severity, thereby improving the accuracy and precision of fault classification, ensuring that maintenance personnel can handle high-priority faults in a timely manner, significantly shortening the fault response time and reducing the downtime risk of the fan system. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0095] Figure 1 A flow chart of a method for locating a fan bearing fault based on an adaptive soundprint sensor network proposed by the present invention;

[0096] Figure 2 This is a schematic diagram of fault identification using a semi-supervised machine learning model in a fan bearing fault location method based on an adaptive voiceprint sensor network proposed in the present invention. DETAILED DESCRIPTION

[0097] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0098] refer to Figure 1-2 , a method for locating a fan bearing fault based on an adaptive voiceprint sensor network, comprising the following steps:

[0099] S1. Multiple adaptive soundprint sensor nodes are arranged at multiple key positions of the fan to collect vibration frequency, noise signal and other acoustic characteristic signals related to bearing faults during the operation of the fan bearing, and construct a fan bearing acoustic signal set;

[0100] S2. Dynamically adjust the wind turbine bearing acoustic signal set using an adaptive soundprint sensor network. The adaptive soundprint sensor network dynamically adjusts the sensitivity, operating frequency and acquisition mode of the adaptive soundprint sensor node according to the wind turbine's operating environment, wind turbine load changes, environmental noise and temperature changes.

[0101] S3, preprocessing the fan bearing acoustic signal set, removing environmental noise and redundant information in the fan bearing acoustic signal set, and retaining effective acoustic signal features related to the fan bearing operating state;

[0102] S4. Use signal processing technology to extract features from the preprocessed fan bearing acoustic signal set, extract the vibration features, energy features and abnormal noise features in the fan bearing acoustic signal, summarize the vibration features, energy features and abnormal noise features of each sensor node, eliminate redundant data and noise interference, retain the effective acoustic signal feature data related to the fan bearing operating status, and construct the fan bearing acoustic feature set;

[0103] S5. Use the semi-supervised machine learning model to identify the faults of the fused acoustic feature set of the fan bearing. By training the semi-supervised machine learning model, the vibration features, energy features and abnormal noise features are extracted and analyzed from the acoustic feature set of the fan bearing, and the potential fault signals of the fan bearing are preliminarily identified. The fault signals are classified into bearing wear fault, crack fault and fatigue fault.

[0104] S6. Further review and further classify the fault signals identified in S5 using the integrated learning model, optimize the determination of the fault type through comprehensive evaluation of vibration characteristics, energy characteristics and abnormal noise characteristics, and prioritize the faults according to their severity;

[0105] S7. Output the fan bearing fault location result determined in S6 through the alarm system, including the fault type, location and severity of the fault, and notify the maintenance personnel to take corresponding maintenance measures according to the fault location result.

[0106] In this implementation, S1 includes the following steps:

[0107] S11. Arrange multiple adaptive soundprint sensor nodes N1, N2, ..., N at key locations of the fan. n , where n is the number of sensor nodes. The sensor nodes are used to collect acoustic signals generated during the operation of the fan bearing, including the vibration frequency f v , noise signal S n and abnormal acoustic signals A related to bearing failure t ;

[0108] S12, each sensor node N i During the operation of the bearing, the vibration frequency f is collected according to different load conditions and working conditions. v,i (t), noise signal S n,i (t) and abnormal acoustic signal A t,i (t), where t is time;

[0109] S13, based on the collected vibration frequency f v,i (t), noise signal S n,i (t) and abnormal acoustic signal A t,i (t) Construct the fan bearing acoustic signal set D s :

[0110] D s ={(f v,1 (t),S n,1 (t),A t,1 (t)),(f v,2 (t),S n,2 (t),A t,2 (t)),…,(f v,n (t),S n,n (t),A t,n (t))}.

[0111] In this implementation, S3 includes the following steps:

[0112] S31. Obtaining the fan bearing acoustic signal set D s , for the noise N in the fan operating environment e (t) is modeled, and the environmental noise characteristics are extracted using frequency domain filtering technology, and the noise removal function H(N e (t)) performing noise elimination processing on the acoustic signal;

[0113] S33, removing redundant data from the processed acoustic signal, using a redundant removal algorithm to remove repeated or irrelevant acoustic data, and obtaining a valid fan bearing acoustic signal set D after cleaning. clean ;

[0114] S34, extracting effective acoustic signal features related to the operating status of the fan bearing, including the vibration frequency f v,i (t), the noise signal after cleaning S clean,i (t) and abnormal acoustic signal A t,i (t), integrated into the pre-processed fan bearing acoustic signal set D proc :

[0115] D proc ={f v,i (t),S clean,i (t),A t,i (t)}.

[0116] In this implementation, S4 includes the following steps:

[0117] S41, the pre-processed fan bearing acoustic signal set D proc Vibration feature extraction is performed, and the intrinsic mode function and instantaneous frequency characteristics of the bearing vibration signal are extracted using empirical mode decomposition and Hilbert-Huang transform technology:

[0118] The vibration signal f collected by each sensor node v,i (t) Perform empirical mode decomposition and decompose it into several intrinsic mode functions IMF k,i (t) and the residual term r i (t):

[0119]

[0120] Among them, IMF k,i (t) is the kth order eigenmode function, K is the total number of modal components, r i (t) is the residual term;

[0121] For each intrinsic mode function IMF k,i (t) Perform Hilbert transform to obtain the instantaneous amplitude A k,i (t) and instantaneous frequency ω k,i (t):

[0122] IMF k,i (t) = A k,i (t)cos(∫ω k,i (t)dt);

[0123] Among them, A k,i (t) reflects the amplitude change of the vibration signal at different times, ωk,i (t) represents the instantaneous frequency, which can reflect the changes in the vibration characteristics of the bearing during operation;

[0124] The instantaneous amplitude and instantaneous frequency characteristics of all eigenmode functions constitute the vibration characteristic vector V i , vibration eigenvector V i Describe the vibration state of the fan bearing at different times and frequencies:

[0125] V i ={(A 1,i (t),ω 1,i (t)),(A 2,i (t),ω 2,i (t)),…,(A K,i (t),ω K,i (t))};

[0126] S42, noise signal S clean,i (t) Perform wavelet packet decomposition to obtain decomposition coefficients W at different scales and positions j,l,i :

[0127]

[0128] Among them, J is the number of decomposition levels, l is the node position, ψ j,l (t) is the corresponding mother wavelet function;

[0129] Calculate the energy characteristics E of the frequency band corresponding to each decomposition coefficient j,l,i , energy characteristic E j,l,i Reflects the energy distribution of noise signals in different frequency bands:

[0130]

[0131] Construct energy eigenvector E i Normalize the energy features:

[0132]

[0133] S43. Abnormal acoustic signal A t,i (t) Extract abnormal noise features and calculate abnormal acoustic signal A t,i The bispectrum B of (t) i (f1,f2):

[0134]

[0135] in, is the third-order autocorrelation function, E[·] is the mathematical expectation;

[0136] Extract the amplitude and phase characteristics of the bispectrum and the bispectral entropy Hi :

[0137]

[0138] in, is the bispectral energy density function, and the bispectral entropy H i It reflects the non-Gaussianity and nonlinearity of the signal;

[0139] Construct an abnormal noise feature vector B containing the bispectral amplitude, phase and entropy characteristics of the abnormal noise signal i :

[0140] B i ={|B i (f1,f2)|,∠B i (f1,f2),H i};

[0141] S44, extract the vibration feature vector V i , energy eigenvector E i and abnormal noise feature vector B i Integration, construct the acoustic feature vector set D of the fan bearing feat :

[0142] D feat = {V i ,E i ,B i}.

[0143] In this implementation, S5 includes the following steps:

[0144] S51, the fan bearing acoustic feature set D feat Divided into labeled dataset D l ={(X l ,Y l )},X l is the feature vector of the labeled sample, Y l are the known fault type labels, including bearing wear fault, crack fault and fatigue fault, and the unlabeled dataset D u ={X u}, X u is the feature vector of the unlabeled sample;

[0145] S52. Build a semi-supervised machine learning model M semi , the similarity matrix W between samples is initially calculated using the graph-based label propagation algorithm:

[0146]

[0147] Among them, W ij Represents sample Xi and sample X j The similarity between ∥X i -X j ∥ is the Euclidean distance between feature vectors, σ is the bandwidth parameter, which controls the decay speed of similarity;

[0148] Normalize the similarity matrix W to obtain the normalized Laplace matrix S;

[0149] S53, initialize the label matrix Y, the labels of the labeled samples remain unchanged, and the labels of the unlabeled samples are set to zero vectors:

[0150]

[0151] S55. Iteratively update the label matrix F through the label propagation algorithm:

[0152] F (t+1) =αSF (t) +(1-α)Y;

[0153] Among them, F (t) is the label matrix of the tth iteration, α is the smoothing parameter, 0<α<1, which controls the trade-off between new and old labels;

[0154] The iteration process continues until the convergence condition is met:

[0155] ∥F (t+1) -F (t) ∥ F <ε;

[0156] Among them, ∥·∥ F is the Frobenius norm, ε is the preset convergence threshold;

[0157] S56. When the iteration converges, use the final label matrix F * Fault type prediction for unlabeled samples

[0158] Test, predict labels Determined by the category corresponding to the maximum value, the prediction confidence C u It is given by the normalized values ​​of the label matrix:

[0159]

[0160] in, For sample The label value in the kth class;

[0161] S57, set the confidence threshold θ, and Unlabeled samples above the threshold and its predicted label Add labeled datasets and update labeled datasets:

[0162]

[0163] S58. Using the updated labeled dataset D′ l Semi-supervised model M semi Retrain and repeat steps S52 to S57 until all unlabeled samples are classified or the maximum number of iterations T is reached. max ;

[0164] S59, finally the optimized semi-supervised fault identification model The acoustic feature set D of the fan bearing is analyzed by using the semi-supervised fault recognition model. feat Perform fault identification and preliminarily identify potential fault signals of fan bearings And the fault signals are classified into bearing wear fault, crack fault and fatigue fault.

[0165] In this implementation, S6 includes the following steps:

[0166] S61, the potential fault signal initially identified Input ensemble learning model M ensemble Review, the integrated learning model includes multiple basic models M j An integrated framework composed of each basic model independently predicts the fault type;

[0167] S62, for each basic model M j For training, the basic model includes the random forest model M RF , Gradient Boosting Tree Model M GB and the support vector machine model M SVM , each model input vibration characteristic V i 、Energy characteristics E i and abnormal noise characteristics B i , and output the corresponding fault type prediction results

[0168] S63. Prediction results of each basic model Perform voting fusion and ensemble learning model’s final prediction results Determined by voting results:

[0169]

[0170] in, is the indicator function, when It takes 1 when it is category k, otherwise it takes 0;

[0171] S64. Final prediction results based on ensemble learning model The fault types are further classified and prioritized according to the severity of the fault signal, which is determined by analyzing the amplitude A of the vibration signature. k,i (t) and the bispectral entropy H of the abnormal noise characteristics i To conduct an assessment:

[0172]

[0173] Among them, P i is the fault priority score, β1 and β2 are weight coefficients, reflecting the influence of vibration amplitude and bispectral entropy on fault severity;

[0174] S65. Score P according to fault priority i The fault types are sorted and divided into three priority levels: high, medium and low. The high-priority fault types will be processed first and alarms will be output, while the low-priority fault types will be recorded as monitoring items.

[0175] In this implementation, S7 includes the following steps:

[0176] S71. Identify and classify the fault types and the fault priority score P i As input, generate wind turbine bearing fault location results, including fault type, fault location, and fault severity;

[0177] S72, based on fault type and the fault priority score P i , generate alarm signal A through preset alarm rules i , the alarm signal is related to the severity of the fault, and the alarm threshold θ is set a , when P i When the value is higher than the alarm threshold, a high priority alarm is generated. high , otherwise a medium or low priority alarm is generated A mid or A low :

[0178]

[0179] Among them, θ a and θ b These are the thresholds for high, medium, and low priority alarms, respectively;

[0180] S73, outputting the complete fault location result through the alarm system and sending the fault information to the maintenance personnel according to the priority of the alarm signal;

[0181] S74, when receiving a high priority alarm signal A highWhen the system sends an emergency maintenance notice to the maintenance personnel, the content includes the specific fault type, location and severity of the fan bearing, and requires immediate maintenance measures. mid and A low Reminder notifications will be sent according to the preset schedule and regular inspections will be arranged.

[0182] Embodiment 1:

[0183] In this embodiment 1, a large wind farm is located in the coastal area of ​​A and has 100 large wind turbines. The bearing system of each wind turbine is a key component in the operation of the wind turbine. The operation and maintenance team of the power farm found in past maintenance records that due to the complex operating environment of the bearings (including high humidity, high wind speed, and large temperature difference between day and night), the bearing failure rate is high, and wear and crack failures are more frequent. The power farm decided to adopt the wind turbine bearing fault location method based on the adaptive voiceprint sensor network of the present invention to improve the accuracy and real-time performance of fault detection.

[0184] In April 2024, the operation and maintenance team installed an adaptive soundprint sensor network on 20 wind turbines. Four adaptive soundprint sensor nodes were arranged at the bearings, shaft and blade base of each wind turbine. After the system is started, the sensor network monitors the operation of the wind turbine bearings in real time, collects the vibration frequency, noise signal and abnormal acoustic signal during bearing operation, and automatically adjusts the sensitivity and operating frequency of the sensor through an adaptive algorithm to ensure that the abnormal characteristic signals of the wind turbine bearings can be captured under various environmental conditions.

[0185] During a full-load operation phase in May 2024, the system detected an abnormality in the bearing of the fan numbered "FJ-45". Specifically, the amplitude of the vibration frequency suddenly increased in a short period of time, far exceeding the average value during normal operation and reaching 1.5 times the set threshold. At the same time, the noise signal also showed obvious abnormal fluctuations, and the energy characteristics in the noise spectrum increased rapidly during the high-load period of the bearing operation.

[0186] After the sensor nodes monitored these signals, the system immediately started fault analysis. The semi-supervised machine learning model system preliminarily classified these abnormal signals and preliminarily identified that the bearing might have a wear fault. Further verification was performed through an integrated learning model. The results of the integrated random forest, gradient boosting tree and support vector machine models showed that in addition to wear faults, the abnormal vibration characteristics of the bearing were also highly consistent with the bearing crack characteristics. The model gave the fault signal a priority score of 85 points, which exceeded the high-priority fault alarm threshold of 70 points.

[0187] The system generated a detailed fault report at 15:30 on May 18, 2024, recording the specific fault type (wear and cracks), fault location (main bearing area) and fault severity of the fan "FJ-45". The report showed that the vibration frequency detected by the sensor increased sharply in a short period of time, and the energy spectrum characteristics of the noise signal showed nonlinear changes, suggesting that small cracks may appear on the bearing surface. The system sent a high-priority alarm signal to the maintenance team through the alarm module, and sent emergency maintenance notices via SMS and emails requiring technicians to go and check immediately.

[0188] After receiving the alarm notification, the maintenance team shut down the fan to inspect it at 16:15 on May 18, 2024. Through on-site inspection, the technicians confirmed that the fan main bearing did have wear and tiny cracks. If not replaced in time, it might cause more serious bearing damage. The technicians completed the replacement and maintenance of the bearing within two hours after the shutdown, and the fan was restarted at 18:30 on the same day, avoiding prolonged shutdown and larger-scale equipment damage.

[0189] In order to further verify the effectiveness of the method of the present invention, the operation and maintenance team compared the performance of the traditional fault detection method and the method of the present invention in the same time period. The following Table 1 is a specific data comparison:

[0190] Table 1 Data comparison between the method of the present invention and the traditional method

[0191] project Method of the present invention Traditional methods Fault detection times 18 times 9 times Detection fault type Wear, cracks, fatigue Wear only Fault detection accuracy 97％ 75％ Average fault response time (minutes) 12 minutes 50 minutes High priority fault location accuracy 96％ 60％ Maintenance completion time (hours) 2 hours 6 hours System downtime (hours) 2 hours 8 hours Maintenance cost (yuan) About 10,000 yuan About 25,000 yuan

[0192] It can be clearly seen from the two-month test data that the method of the present invention has significant advantages in fault detection accuracy and response time. The traditional method can only detect bearing wear problems through manual or single sensors when obvious problems occur in the fan and it usually takes 50 minutes to respond. The present invention monitors the operation of the fan bearings in real time through an adaptive voiceprint sensor network, and can identify faults within 12 minutes. It can also identify more fault types through semi-supervised learning and integrated learning models, thereby more comprehensively monitoring the health status of the fan bearings.

[0193] In addition, traditional methods also have obvious disadvantages in fault handling and maintenance costs. Taking the fan "FJ-45" ​​as an example, if the traditional detection method is used, the crack fault will not be discovered in the early stage until the bearing is seriously damaged and maintenance is not taken, resulting in long-term shutdown and increased risk of equipment damage. The present invention can issue an early warning at the beginning of the crack and promptly notify maintenance personnel to carry out repairs, avoiding high maintenance costs and long-term equipment shutdown.

[0194] This embodiment fully demonstrates the application effect of the wind turbine bearing fault location method based on the adaptive voiceprint sensor network of the present invention in an actual wind farm. Through real-time monitoring of the adaptive sensor network, fault identification of semi-supervised machine learning and fault review of the integrated learning model, the present invention not only improves the accuracy of fault detection but also greatly shortens the fault response time, reduces system downtime and reduces maintenance costs, and provides a strong guarantee for the safe operation of the wind power generation system.

[0195] The adaptive voiceprint sensor network in the present invention can adaptively adjust according to multiple factors such as the fan operating environment, load changes, environmental noise and temperature changes by dynamically adjusting the sensitivity, operating frequency and acquisition mode of the sensor nodes, effectively solving the problem that traditional fixed sensors cannot cope with inaccurate data collection under complex working conditions, so that the collected acoustic signals can accurately reflect the actual operating status of the fan bearing, thereby greatly improving the accuracy of fault detection. The dynamic adjustment mechanism ensures that the sensor can still maintain efficient monitoring even in harsh environments, avoiding the problems of missed detection or false detection caused by interference from environmental factors in traditional methods.

[0196] The present invention introduces a semi-supervised machine learning model, which can significantly improve the fault recognition performance of the model by utilizing unlabeled data with a small amount of labeled data. By constructing an acoustic feature set and training it using a semi-supervised learning method, the model can adaptively extract and analyze the vibration characteristics, energy characteristics and abnormal noise characteristics in the acoustic signals of the fan bearings and accurately identify the wear, cracks and fatigue fault types of the bearings. Compared with traditional supervised learning methods, semi-supervised learning not only improves the accuracy of fault recognition, but also effectively reduces the dependence on large-scale labeled data and improves the robustness of the system under different working conditions.

[0197] The present invention constructs an integrated learning model by integrating multiple basic models such as random forest, gradient boosting tree and support vector machine to review and further classify the initially identified fault signals. The integrated model comprehensively evaluates different fault characteristics. In terms of fault severity priority classification, the fault type determination is optimized by vibration amplitude and bispectral entropy and priority is divided according to the severity of the fault, thereby improving the accuracy and precision of fault classification, ensuring that maintenance personnel can handle high-priority faults in a timely manner, significantly shortening the fault response time and reducing the downtime risk of the fan system.

[0198] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for locating fan bearing faults based on an adaptive soundprint sensor network, characterized in that: The steps include: S1. Multiple adaptive soundprint sensor nodes are arranged at multiple key positions of the fan to collect vibration frequency, noise signal and other acoustic characteristic signals related to bearing faults during the operation of the fan bearing, and construct a fan bearing acoustic signal set; S2. Dynamically adjust the wind turbine bearing acoustic signal set using an adaptive soundprint sensor network. The adaptive soundprint sensor network dynamically adjusts the sensitivity, operating frequency and acquisition mode of the adaptive soundprint sensor node according to the wind turbine's operating environment, wind turbine load changes, environmental noise and temperature changes. S3, preprocessing the fan bearing acoustic signal set, removing environmental noise and redundant information in the fan bearing acoustic signal set, and retaining effective acoustic signal features related to the fan bearing operating state; S4. Use signal processing technology to extract features from the preprocessed fan bearing acoustic signal set, extract the vibration features, energy features and abnormal noise features in the fan bearing acoustic signal, summarize the vibration features, energy features and abnormal noise features of each sensor node, eliminate redundant data and noise interference, retain the effective acoustic signal feature data related to the fan bearing operating status, and construct the fan bearing acoustic feature set; S5. Use the semi-supervised machine learning model to identify the faults of the fused acoustic feature set of the fan bearing. By training the semi-supervised machine learning model, the vibration features, energy features and abnormal noise features are extracted and analyzed from the acoustic feature set of the fan bearing, and the potential fault signals of the fan bearing are preliminarily identified. The fault signals are classified into bearing wear fault, crack fault and fatigue fault. S6. Further review and further classify the fault signals identified in S5 using the integrated learning model, optimize the determination of the fault type through comprehensive evaluation of vibration characteristics, energy characteristics and abnormal noise characteristics, and prioritize the faults according to their severity; S7. Output the fan bearing fault location result determined in S6 through the alarm system, including the fault type, location and severity of the fault, and notify the maintenance personnel to take corresponding maintenance measures according to the fault location result.

2. A method for locating a fan bearing fault based on an adaptive soundprint sensor network according to claim 1, characterized in that: The S1 comprises the following steps: S11. Arrange multiple adaptive soundprint sensor nodes N1, N2, ..., N at key locations of the fan. n , where n is the number of sensor nodes. The sensor nodes are used to collect acoustic signals generated during the operation of the fan bearing, including the vibration frequency f v , noise signal S n and abnormal acoustic signals A related to bearing failure t ; S12, each sensor node N i During the operation of the bearing, the vibration frequency f is collected according to different load conditions and working conditions. v,i (t), noise signal S n,i (t) and abnormal acoustic signal A t,i (t), where t is time; S13, based on the collected vibration frequency f v,i (t), noise signal S n,i (t) and abnormal acoustic signal A t,i (t) Construct the fan bearing acoustic signal set D s : D s ={(f v,1 (t),S n,1 (t),A t,1 (t)),(f v,2 (t),S n,2 (t),A t,2 (t)),…,(f v,n (t),S n,n (t),A t,n (t))}。 3. A method for locating a fan bearing fault based on an adaptive soundprint sensor network according to claim 1, characterized in that: The S3 comprises the following steps: S31. Obtaining fan bearing acoustic signal set D s , for the noise N in the fan operating environment e (t) is modeled, and the environmental noise characteristics are extracted using frequency domain filtering technology, and the noise removal function H(N e (t)) performing noise elimination processing on the acoustic signal; S33, removing redundant data from the processed acoustic signal, using a redundant removal algorithm to remove repeated or irrelevant acoustic data, and obtaining a valid fan bearing acoustic signal set D after cleaning. clean ; S34, extracting effective acoustic signal features related to the operating status of the fan bearing, including the vibration frequency f v,i (t), the noise signal after cleaning S clean,i (t) and abnormal acoustic signal A t,i (t), integrated into the pre-processed fan bearing acoustic signal set D proc : D proc ={f v,i (t),S clean,i (t),A t,i (t)}。 4. The method for locating a fan bearing fault based on an adaptive soundprint sensor network according to claim 1, characterized in that: The S4 comprises the following steps: S41, the pre-processed fan bearing acoustic signal set D proc Vibration feature extraction is performed, and the intrinsic mode function and instantaneous frequency characteristics of the bearing vibration signal are extracted using empirical mode decomposition and Hilbert-Huang transform technology: The vibration signal f collected by each sensor node v,i (t) Perform empirical mode decomposition and decompose it into several intrinsic mode functions IMF k,i (t) and the residual term r i (t): Among them, IMF k,i (t) is the kth order eigenmode function, K is the total number of modal components, r i (t) is the residual term; For each intrinsic mode function IMF k,i (t) Perform Hilbert transform to obtain the instantaneous amplitude A k,i (t) and instantaneous frequency ω k,i (t): IMF k,i (t)=A k,i (t)cos(∫ω k,i (t)dt); Among them, A k,i (t) reflects the amplitude change of the vibration signal at different times, ω k,i (t) represents the instantaneous frequency, which can reflect the changes in the vibration characteristics of the bearing during operation; The instantaneous amplitude and instantaneous frequency characteristics of all eigenmode functions constitute the vibration characteristic vector V i , vibration eigenvector V i Describe the vibration state of the fan bearing at different times and frequencies: V i ={(A 1,i (t),ω 1,i (t)),(A 2,i (t),ω 2,i (t)),…,(A K,i (t),ω K,i (t))}; S42, noise signal S clean,i (t) Perform wavelet packet decomposition to obtain decomposition coefficients W at different scales and positions j,l,i : Among them, J is the number of decomposition levels, l is the node position, ψ j,l (t) is the corresponding mother wavelet function; Calculate the energy characteristics E of the frequency band corresponding to each decomposition coefficient j,l,i , energy characteristic E j,l,i Reflects the energy distribution of noise signals in different frequency bands: Construct energy eigenvector E i Normalize the energy features: S43. Abnormal acoustic signal A t,i (t) Extract abnormal noise features and calculate abnormal acoustic signal A t,i The bispectrum B of (t) i (f1,f2): in, is the third-order autocorrelation function, E[·] is the mathematical expectation; Extract the amplitude and phase characteristics of the bispectrum and the bispectral entropy H i : in, is the bispectral energy density function, and the bispectral entropy H i It reflects the non-Gaussianity and nonlinearity of the signal; Construct an abnormal noise feature vector B containing the bispectral amplitude, phase and entropy characteristics of the abnormal noise signal i : B i ={|B i (f1,f2)|,∠B i (f1,f2),H i }; S44, extract the vibration feature vector V i , energy eigenvector E i and abnormal noise feature vector B i Integration, construct the acoustic feature vector set D of the fan bearing feat : D feat ={V i ,E i ,B i }。 5. The method for locating a fan bearing fault based on an adaptive soundprint sensor network according to claim 1, characterized in that: The S5 comprises the following steps: S51, the fan bearing acoustic feature set D feat Divided into labeled dataset D l ={(X l ,Y l )},X l is the feature vector of the labeled sample, Y l are the known fault type labels, including bearing wear fault, crack fault and fatigue fault, and the unlabeled dataset D u ={X u }, X u is the feature vector of the unlabeled sample; S52. Build a semi-supervised machine learning model M semi , the similarity matrix W between samples is initially calculated using the graph-based label propagation algorithm: Among them, W ij Represents sample X i and sample X j The similarity between ∥X i -X j ∥ is the Euclidean distance between feature vectors, σ is the bandwidth parameter, which controls the decay speed of similarity; Normalize the similarity matrix W to obtain the normalized Laplace matrix S; S53, initialize the label matrix Y, the labels of the labeled samples remain unchanged, and the labels of the unlabeled samples are set to zero vectors: S55. Iteratively update the label matrix F through the label propagation algorithm: F (t+1) =αSF (t) +(1-α)Y; Among them, F (t) is the label matrix of the tth iteration, α is the smoothing parameter, 0<α<1, which controls the trade-off between new and old labels; The iteration process continues until the convergence condition is met: ∥F (t+1) -F (t) ∥ F <ε; Among them, ∥·∥ F is the Frobenius norm, ε is the preset convergence threshold; S56. When the iteration converges, use the final label matrix F * Predict the fault type and label of unlabeled samples Determined by the category corresponding to the maximum value, the prediction confidence C u It is given by the normalized values ​​of the label matrix: in, For sample The label value in the kth class; S57, set the confidence threshold θ, and Unlabeled samples above the threshold and its predicted label Add labeled datasets and update labeled datasets: S58. Using the updated labeled dataset D′ l Semi-supervised model M semi Retrain and repeat steps S52 to S57 until all unlabeled samples are classified or the maximum number of iterations T is reached. max ; S59, finally the optimized semi-supervised fault identification model Using semi-supervised fault identification model to identify the acoustic feature set D of fan bearings feat Perform fault identification and preliminarily identify potential fault signals of fan bearings And the fault signals are classified into bearing wear fault, crack fault and fatigue fault.

6. The method for locating a fan bearing fault based on an adaptive soundprint sensor network according to claim 1, characterized in that: The S6 comprises the following steps: S61, the potential fault signal initially identified Input ensemble learning model M ensemble Review, the integrated learning model includes multiple basic models M j An integrated framework composed of each basic model independently predicts the fault type; S62, for each basic model M j For training, the basic model includes the random forest model M RF , Gradient Boosting Tree Model M GB and the support vector machine model M SVM , each model input vibration characteristic V i , Energy characteristics E i and abnormal noise characteristics B i , and output the corresponding fault type prediction results S63. Prediction results of each basic model Perform voting fusion and ensemble learning model’s final prediction results Determined by voting results: in, is the indicator function, when It takes 1 when it is category k, otherwise it takes 0; S64. Final prediction results based on ensemble learning model The fault types are further classified and prioritized according to the severity of the fault signal, which is determined by analyzing the amplitude A of the vibration signature. k,i (t) and the bispectral entropy H of the abnormal noise characteristics i To conduct an assessment: Among them, P i is the fault priority score, β1 and β2 are weight coefficients, reflecting the influence of vibration amplitude and bispectral entropy on fault severity; S65. Score P according to fault priority i The fault types are sorted and divided into three priority levels: high, medium and low. The high-priority fault types will be processed first and alarms will be output, while the low-priority fault types will be recorded as monitoring items.

7. The method for locating a fan bearing fault based on an adaptive soundprint sensor network according to claim 1, characterized in that: The S7 comprises the following steps: S71. Identify and classify the fault types and the fault priority score P i As input, generate wind turbine bearing fault location results, including fault type, fault location, and fault severity; S72, based on fault type and the fault priority score P i , generate alarm signal A through preset alarm rules i , the alarm signal is related to the severity of the fault, and the alarm threshold θ is set a , when P i When the value is higher than the alarm threshold, a high priority alarm is generated. high , otherwise a medium or low priority alarm is generated A mid or A low : Among them, θ a and θ b These are the thresholds for high, medium, and low priority alarms, respectively; S73, outputting the complete fault location result through the alarm system and sending the fault information to the maintenance personnel according to the priority of the alarm signal; S74, when receiving a high priority alarm signal A high When the system sends an emergency maintenance notice to the maintenance personnel, the content includes the specific fault type, location and severity of the fan bearing, and requires immediate maintenance measures. mid and A low Reminder notifications will be sent according to the preset schedule and regular inspections will be arranged.

Images