A fan equipment fault soundprint monitoring and diagnosing method and system

Abstract

The application discloses a kind of fan equipment fault voiceprint monitoring diagnosis method and system, and the application relates to voiceprint monitoring technical field, based on voiceprint acquisition array real-time acquisition fan under the voiceprint signal of given power and carry out time series discretization, using short-time Fourier transform to execute time-frequency domain analysis, extract the instantaneous characteristic frequency of each time window and simultaneously calculate short-time energy, construct associated time sequence group;Further, the time sequence synchronization features and difference features between the two groups of sequences are mined, a multi-dimensional fault feature vector is constructed, input into the trained multilayer perceptron model, and the fault type recognition and severity assessment are simultaneously realized;By associating local abnormal fluctuation positioning fault time interval and frequency component, generate the diagnostic results containing fault type, fault time and frequency band and output to monitoring terminal, based on voiceprint acquisition, time-frequency analysis, fault feature extraction, fault diagnosis, abnormal positioning and output module and monitoring terminal constitute system, realize real-time voiceprint monitoring and intelligent diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of voiceprint monitoring technology, specifically, it relates to a method and system for voiceprint monitoring and diagnosis of wind turbine equipment faults. Background Technology

[0002] As a non-contact means of sensing equipment status, voiceprint monitoring technology can capture changes in acoustic characteristics of wind turbine equipment in real time, providing rich data support for equipment health management.

[0003] In existing technologies for monitoring wind turbines based on acoustic signature information, frequency or energy characteristics are typically processed in isolation during time-frequency analysis. For example, only amplitude changes or total energy fluctuations in a specific frequency band are monitored. This results in the inability to capture the dynamic coupling relationship between frequency and energy on the time axis. Furthermore, most existing technologies use single or simple combinations of statistical features as fault criteria, failing to quantify the synchronous or lagging relationship between frequency changes and energy fluctuations. This relationship is precisely a sensitive indicator of early faults such as wind turbine blade imbalance, bearing wear, or abnormal gear meshing. In addition, fault diagnosis models in existing technologies often output discrete fault categories, lacking continuous quantitative assessment of severity. They cannot correlate diagnostic results back to local abnormal fluctuations, pinpointing the specific time interval and frequency components of the fault occurrence. Alarm information from existing methods usually only provides abnormal or coarse timestamps, failing to output information that is highly instructive for on-site maintenance.

[0004] To address the aforementioned problems, this invention proposes a method and system for monitoring and diagnosing wind turbine equipment faults using acoustic signatures. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for monitoring and diagnosing wind turbine equipment faults using acoustic signatures, solving the problem that existing technologies cannot simultaneously utilize frequency and energy temporal coupling characteristics for accurate fault location and severity quantification.

[0006] The objective of this invention can be achieved through the following technical solutions: A method for acoustic monitoring and diagnosis of wind turbine equipment faults, the method comprising: Step 1: Based on the real-time monitoring of the acoustic signature signal of the target wind turbine equipment under a predetermined power operating state by the acoustic signature acquisition device, the acoustic signature signal is divided into time sequences in combination with time to determine the acoustic signature signal sequence associated with the target wind turbine equipment. Step 2: Perform time-frequency domain analysis on the acoustic signature signal sequence, lock the time-frequency signal, generate the instantaneous characteristic frequency sequence associated with the target wind turbine equipment, and simultaneously construct a short-time energy sequence, which is then associated with the instantaneous characteristic frequency sequence to form an associated time sequence group. Step 3: Extract the temporal synchronization features and difference features between the instantaneous feature frequency sequence and the short-time energy sequence in the associated time series group, and construct a multi-dimensional fault feature vector; Input multidimensional fault feature vectors into the fault diagnosis model to identify fault types and severity. Step four involves associating the identified fault type and severity with the corresponding local abnormal fluctuations in the associated time series group, locating the time interval and frequency components of the fault occurrence, generating diagnostic results that include fault type, fault time, and fault frequency band, and outputting them to the monitoring terminal.

[0007] As a further aspect of the present invention, the specific method for real-time monitoring of the acoustic signature signal of the target wind turbine equipment under a predetermined power operating state based on the acoustic signature acquisition device in step one is as follows: The target wind turbine equipment is identified and denoted as Q; Based on the pre-deployed acoustic fingerprint acquisition array associated with the target wind turbine equipment Q, the acoustic fingerprint signal of the target wind turbine equipment Q in the predetermined power operation state is collected. The predetermined power is preset by the operator in combination with the rated power of the target wind turbine equipment Q. The voiceprint acquisition array acquires voiceprint signals at a preset acquisition frequency by the operator, records the voiceprint signals as W, and stores them.

[0008] As a further aspect of the present invention, the specific method for determining the acoustic signature signal sequence associated with the target wind turbine equipment in step one is as follows: Based on the time sequence of the discrete voiceprint signal W collected by the voiceprint acquisition array of the target wind turbine equipment Q, the voiceprint signal sequence W1, W2, ..., Wj associated with the target wind turbine equipment Q is obtained, where j is the total number of discrete voiceprint signals, which increases with the acquisition time of the voiceprint acquisition array.

[0009] As a further aspect of the present invention, in step two, the specific method for performing time-frequency domain analysis on the acoustic signature signal sequence, locking the time-frequency signal, and generating the instantaneous characteristic frequency sequence associated with the target wind turbine equipment is as follows: Extract the voiceprint signal sequence W1, W2, ..., Wj, and perform a short-time Fourier transform, where the time window length is a preset length L, the window function is a Hamming window, and the window shift is L / 2; Transform the signal segments in the corresponding voiceprint signal sequence within each time window to the frequency domain to construct the time-frequency distribution matrix ST(t,f), where t is the index of the time window and f is the frequency value; For each time window t, the frequency component with the largest amplitude is extracted from the time-frequency distribution matrix ST(t,f) and used as the instantaneous characteristic frequency of the corresponding time window; The instantaneous characteristic frequencies of all time windows are sorted in chronological order to form the instantaneous characteristic frequency sequence F1, F2, ..., Fk associated with the target wind turbine equipment Q, where k is the total number of time windows.

[0010] As a further aspect of the present invention, the specific method for constructing the associated time series group in step two is as follows: Extract signal segments from the voiceprint signal sequences corresponding to all time windows, calculate the short-time energy of the signal segments in each time window, and arrange them in chronological order to form a short-time energy sequence E1, E2, ..., Ek; The short-time energy sequences E1, E2, ..., Ek are extracted synchronously and matched with the instantaneous feature frequency sequences F1, F2, ..., Fk according to the time sequence to form the associated time sequence group {Eo, Fo}, where o is the counting index, with a value range from 1 to k.

[0011] As a further aspect of the present invention, the specific method for constructing the multidimensional fault feature vector in step three is as follows: Extract the associated time series group {Eo, Fo}; Based on the correlation, the Pearson correlation coefficient between the short-time energy sequence E1, E2, ..., Ek and the instantaneous characteristic frequency sequence F1, F2, ..., Fk is calculated and used as the time-series synchronous correlation coefficient feature TZ1; Then, based on the correlation, the short-time energy sequence and the instantaneous characteristic frequency sequence are subjected to first-order difference, and the proportion of time windows with the same difference sign is statistically analyzed and denoted as the synchronization rate feature TZ2 of the change direction. Determine the time window indices of all peak points for the short-time energy sequence and the instantaneous characteristic frequency sequence respectively, and calculate the mean of the absolute values ​​of the differences between the peak point time window indices, which is denoted as the peak offset difference feature TZ3; For short-time energy sequences and instantaneous characteristic frequency sequences, the ratio of local fluctuation amplitudes of short-time energy and instantaneous characteristic frequency within the same time window is calculated. The local fluctuation amplitude ratio sequence is obtained in chronological order, and the coefficient of variation of the local fluctuation amplitude ratio sequence is calculated. This is denoted as the energy-frequency coupling difference feature TZ4. The timing synchronization correlation coefficient feature TZ1, the change direction synchronization rate feature TZ2, the peak offset difference feature TZ3, and the energy frequency coupling difference feature TZ4 are combined to form a multidimensional fault feature vector [TZ1,TZ2,TZ3,TZ4].

[0012] As a further aspect of the present invention, in step three, the specific method for inputting the multidimensional fault feature vector into the fault diagnosis model to identify the fault type and severity is as follows: A fault diagnosis model based on a multilayer perceptron is constructed, which includes an input layer, a first hidden layer, a second hidden layer, and an output layer; The input layer contains four neurons for receiving multidimensional fault feature vectors [TZ1, TZ2, TZ3, TZ4]. The first hidden layer and the second hidden layer contain 32 and 16 neurons respectively, both of which use the ReLU activation function; The output layer includes a parallel fault type classification branch and a severity regression branch. The fault type classification branch contains K neurons and uses the Softmax activation function, where K is the total number of fault types to be identified. The severity regression branch contains 1 neuron and uses a linear activation function. Input the multidimensional fault feature vector [TZ1,TZ2,TZ3,TZ4] into the trained fault diagnosis model, take the fault type corresponding to the maximum probability value in the output of the Softmax activation function as the diagnosed fault type, take the output value of the linear activation function as the diagnosed severity, and output the diagnosed fault type and the diagnosed severity.

[0013] As a further aspect of the present invention, the training method for the fault diagnosis model in step three specifically includes: Acquire historical acoustic signature signals of several wind turbines of the same specifications as the target wind turbine equipment Q under different fault types and severity states. Extract multidimensional fault feature vectors based on steps one to three, and have the operators label the fault type and severity to form a training sample set. The multidimensional fault feature vector in the training sample set is used as input, and the forward propagation is performed through a multilayer perceptron. The fault type classification branch outputs the predicted fault type probability distribution, and the severity regression branch outputs the predicted severity value. The classification loss between the predicted fault type probability distribution and the fault type label is calculated using the classification cross-entropy loss function, and the regression loss between the predicted severity value and the severity label is calculated using the mean squared error loss function. The classification loss and regression loss are weighted and summed according to preset weights to obtain the total loss function. The Adam optimization algorithm is adopted to minimize the total loss function. The network weights and bias parameters of the multilayer perceptron are updated through backpropagation, and the training is iterated until the total loss function converges, thus completing the training of the fault diagnosis model.

[0014] As a further aspect of the present invention, in step four, the specific method for generating diagnostic results including fault type, fault time, and fault frequency band, and outputting them to the monitoring terminal is as follows: Extract the short-time energy sequences E1, E2, ..., Ek and the instantaneous feature frequency sequences F1, F2, ..., Fk corresponding to the associated time series group {Eo, Fo}; The average of the historical short-time energy series constructed in step two under normal and fault-free operation of the target wind turbine equipment Q is used as the normal energy baseline value. For short-time energy sequences, calculate the energy deviation between the short-time energy and the normal energy baseline value for each time window, and mark the time windows with energy deviations greater than a preset energy deviation threshold as abnormal energy windows; Obtain the rated frequency of the target wind turbine equipment Q. For the instantaneous characteristic frequency sequence, calculate the frequency difference between the instantaneous characteristic frequency and the rated frequency for each time window. Mark the time window where the absolute value of the frequency difference is greater than the preset frequency difference threshold as an abnormal frequency window. The fault time window interval is obtained by taking the union of the abnormal energy window and the abnormal frequency window; Extract all instantaneous characteristic frequency values ​​within the fault time window interval, calculate their maximum and minimum frequencies, and form a fault frequency band; The fault time window interval is converted into actual time based on the time window length L and the window shift L / 2 to form the fault time. The diagnostic results are combined by combining the fault type, severity, time of fault, and frequency band, and then output to the monitoring terminal.

[0015] A wind turbine equipment fault acoustic signature monitoring and diagnostic system, the system comprising: The voiceprint acquisition module monitors the voiceprint signal of the target wind turbine equipment in real time under a predetermined power operating state, and divides the voiceprint signal into time sequence based on time to determine the voiceprint signal sequence associated with the target wind turbine equipment. The time-frequency analysis module performs time-frequency domain analysis on the acoustic signal sequence, locks the time-frequency signal, generates the instantaneous characteristic frequency sequence associated with the target wind turbine equipment, and simultaneously constructs a short-time energy sequence, which is then associated with the instantaneous characteristic frequency sequence to form an associated time sequence group. The fault feature extraction module extracts the temporal synchronization and difference features between the instantaneous feature frequency sequence and the short-time energy sequence in the associated time series group, and constructs a multi-dimensional fault feature vector. The fault diagnosis module has a built-in trained fault diagnosis model, receives multi-dimensional fault feature vectors, and identifies the fault type and severity. The anomaly localization and output module associates the identified fault type and severity with the corresponding local anomaly fluctuations in the associated time series group, locates the time interval and frequency components of the fault occurrence, generates diagnostic results including fault type, fault time, and fault frequency band, and outputs them to the monitoring terminal. The monitoring terminal interacts with operators in real time, coordinating data transmission and operation between various modules.

[0016] The beneficial effects of this invention are: This invention achieves non-contact monitoring through acoustic signature acquisition, avoiding physical modifications and interference to wind turbine equipment. Simultaneously, it combines time-frequency domain analysis with the temporal correlation of instantaneous characteristic frequencies and short-time energy sequences to capture weak acoustic signature fluctuations caused by faults. It also extracts synchronization and difference features to construct multi-dimensional vectors, improving the sensitivity and accuracy of fault diagnosis. Based on a built-in fault diagnosis model, it automatically identifies the fault type and severity, further correlates local abnormal fluctuations, and precisely locates the time interval and frequency components of the fault occurrence, providing a clear basis for rapid repair. The final diagnostic results are intuitive and comprehensive, supporting real-time interaction and early warning, ensuring the safe and stable operation of wind turbine equipment and reducing unplanned downtime losses. This invention monitors the operating status of a target wind turbine in real time at a predetermined power level using a voiceprint acquisition array. It discretizes the voiceprint signal using a preset acquisition frequency and forms an ordered voiceprint signal sequence over time, which facilitates time-series analysis, trend identification, and anomaly comparison of voiceprint features. This enhances the sensitivity and accuracy of fault early warning and realizes non-intrusive and automated monitoring of the wind turbine's operating status, reducing the cost of manual inspection. This invention uses short-time Fourier transform with Hamming window and overlapping window shift to suppress spectral leakage and improve time-frequency resolution, thereby locking the frequency component with the largest amplitude within each time window and generating an instantaneous characteristic frequency sequence. Simultaneously, it calculates the short-time energy of each time window and aligns it with the instantaneous characteristic frequencies in time sequence to construct a correlated time series group, achieving synchronous monitoring of energy and frequency. Based on the combination of short-time energy reflecting signal strength fluctuations and instantaneous characteristic frequencies reflecting rotational or vibrational characteristics, it can more sensitively identify abnormal wind turbine states, avoiding the one-sidedness of single-parameter analysis and improving the accuracy and robustness of state assessment. This invention combines the construction of multi-dimensional feature vectors with a multi-task neural network model. Based on four features—time-series synchronization correlation coefficient, change direction synchronization rate, peak offset difference, and energy-frequency coupling difference—it characterizes the dynamic relationship between energy and frequency in the voiceprint signal from the dimensions of correlation, directional consistency, peak alignment, and fluctuation coupling. This can capture anomalies caused by faults and avoid the shortcomings of single features being susceptible to noise interference. At the same time, it adopts a multilayer perceptron model that includes classification and regression branches to synchronously output the fault type and severity, achieving end-to-end diagnosis. By combining the energy baseline and rated frequency, it accurately calibrates the specific time and frequency band of the fault occurrence, providing maintenance personnel with clearly located diagnostic results and supporting rapid response and precise repair. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating the method described in this invention; Figure 2 This is a schematic diagram of the system described in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, this application provides a method and system for monitoring and diagnosing wind turbine equipment faults using acoustic signatures; As an embodiment 1 of this application, it specifically includes: Step 1: Based on the real-time monitoring of the acoustic signature signal of the target wind turbine equipment under a predetermined power operating state by the acoustic signature acquisition device, the acoustic signature signal is divided into time sequences in combination with time to determine the acoustic signature signal sequence associated with the target wind turbine equipment. Step 2: Perform time-frequency domain analysis on the acoustic signature signal sequence, lock the time-frequency signal, generate the instantaneous characteristic frequency sequence associated with the target wind turbine equipment, and simultaneously construct a short-time energy sequence, which is then associated with the instantaneous characteristic frequency sequence to form an associated time sequence group. Step 3: Extract the temporal synchronization features and difference features between the instantaneous feature frequency sequence and the short-time energy sequence in the associated time series group, and construct a multi-dimensional fault feature vector; Input multidimensional fault feature vectors into the fault diagnosis model to identify fault types and severity. Step four involves associating the identified fault type and severity with the corresponding local abnormal fluctuations in the associated time series group, locating the time interval and frequency components of the fault occurrence, generating diagnostic results that include fault type, fault time, and fault frequency band, and outputting them to the monitoring terminal.

[0021] Example 2

[0022] This embodiment is based on Embodiment 1, and provides a detailed description of the modules involved in a wind turbine equipment fault acoustic signature monitoring and diagnosis system. It also discloses a basic real-time process applied to a wind turbine equipment fault acoustic signature monitoring and diagnosis method and system, such as... Figure 1 , Figure 2 As shown, it specifically includes the following: First, a wind turbine equipment fault acoustic monitoring and diagnosis system is constructed. The system includes an acoustic acquisition module, a time-frequency analysis module, a fault feature extraction module, a fault diagnosis module, an anomaly location and output module, and a monitoring terminal.

[0023] The voiceprint acquisition module includes a voiceprint acquisition array, specifically a multi-channel microelectromechanical system microphone array with a frequency response range of 20 Hz to 20 kHz and a signal-to-noise ratio of 65 dB. The voiceprint acquisition array is uniformly arranged in a circular pattern on the inner wall of the nacelle of the target wind turbine equipment Q, 0.8 meters away from the main shaft bearing seat, and has a total of 6 acquisition channels. The signals from each microphone are connected to a synchronous data acquisition card via shielded audio cables, supporting 24-bit analog-to-digital conversion, with a sampling frequency set to 51.2 kHz. The data acquisition card is connected to an industrial embedded computer via a USB 2.0 interface. The time-frequency analysis module, fault feature extraction module, fault diagnosis module, and anomaly location and output module are all deployed in the industrial embedded computer as software functional components. The monitoring terminal is an industrial touch screen human-machine interface, which is connected to the industrial embedded computer via Ethernet to display diagnostic results and receive operator input.

[0024] In step one, the operator inputs the rated power of the target wind turbine equipment Q, for example, 1.5 MW, through the monitoring terminal, and sets the threshold for determining the predetermined power operating state to 85% of the rated power, i.e., 1.275 MW. The soundprint acquisition module reads the sound pressure waveform of the multi-channel microelectromechanical system microphone in real time through the data acquisition card. When the output power of the target wind turbine equipment Q is detected to be within ±5% of 1.275 MW, the wind turbine is identified as being in the predetermined power operating state, and the sound pressure signal collected during this period is recorded as the soundprint signal W. The voiceprint acquisition module discretizes the continuous voiceprint signal W into signal segments according to the time sequence, generating a voiceprint signal sequence W1, W2, ..., Wj, where j is an index that increases with the acquisition time. In step two, the time-frequency analysis module is used to perform time-frequency domain analysis on the voiceprint signal sequence. In this embodiment, both short-time Fourier transform and continuous wavelet transform are supported. Here, short-time Fourier transform is used as an example. The time window length L is set to 1024 sampling points, the window function is Hamming window, and the window shift is 512 sampling points, i.e., L / 2. For each signal segment Wj, the signal segment is truncated by sliding window, and the fast Fourier transform of each signal segment is calculated to obtain the time-frequency distribution matrix ST(t,f), where t represents the time window index, such as 1, 2, 3, ..., and f is the discrete frequency value. For each time window t, the frequency component with the largest amplitude is retrieved from ST(t,f), and the corresponding frequency value is taken as the instantaneous characteristic frequency of the corresponding time window. This process is repeated to determine the instantaneous characteristic frequencies of all time windows and arrange them in time order to form the instantaneous characteristic frequency sequence F1, F2, ..., Fk, where k is the total number of time windows. Simultaneously, based on the time-frequency analysis module, the sum of squares of the amplitude of the sampling points of the signal segments within each time window is calculated as the short-time energy of that time window. The short-time energy sequences E1, E2, ..., Ek are constructed in time sequence, and the short-time energy sequences are matched one-to-one with the instantaneous characteristic frequency sequences according to the same time window index to form an associated time series group {Eo, Fo}, where o is the counting index and the value of o ranges from 1 to k. In step three, the fault feature extraction module extracts the temporal synchronization and difference features between the instantaneous feature frequency sequence and the short-time energy sequence in the associated time series group {Eo,Fo}. Specifically, the feature construction method is to calculate the Pearson correlation coefficient between the instantaneous feature frequency sequence F1,F2,...,Fk and the short-time energy sequence E1,E2,...,Ek to obtain the temporal synchronization correlation coefficient feature. Determine the time window indices corresponding to the peak points of the top 10% amplitude in the instantaneous characteristic frequency sequences F1, F2, ..., Fk and the short-time energy sequences E1, E2, ..., Ek, respectively. Calculate the average of the absolute values ​​of the differences between the two types of peak point indices as the peak offset difference feature. Then calculate the Euclidean distance between the instantaneous characteristic frequency sequences F1, F2, ..., Fk and the short-time energy sequences E1, E2, ..., Ek at each time window, and calculate the variance of the Euclidean distance under all time windows as the energy-frequency coupling difference feature value. Then perform first-order differences between the short-time energy sequences E1, E2, ..., Ek and the instantaneous characteristic frequency sequences F1, F2, ..., Fk respectively, and calculate the proportion of time windows with the same difference sign, which is denoted as the synchronization rate feature of the change direction. The above features are combined to form a multidimensional fault feature vector. The fault diagnosis module is equipped with a pre-trained three-layer backpropagation neural network model (fault diagnosis model). Its input layer contains 4 neurons, corresponding to the four features; the hidden layer contains 8 neurons, using the Sigmoid activation function; and the output layer contains K+1 neurons, where K is the total number of fault types to be identified. The first K neurons output the probability of each fault type through the Softmax function, and the K+1th neuron outputs the severity value. The multidimensional fault feature vector is input into the fault diagnosis model, and the fault type corresponding to the highest probability is taken as the diagnosed fault type. The output value of the linear neuron is taken as the diagnosed severity.

[0025] In step four, the anomaly location and output module pre-stores the mean value of the short-time energy sequence obtained in step two under the fault-free normal operation state of the target wind turbine equipment Q, and uses it as the normal energy baseline value; for the current short-time energy sequence, the deviation between the energy of each time window and the baseline value is calculated, and the time window with a deviation exceeding 20% ​​of the baseline value is marked as an abnormal energy window; Simultaneously, the rated frequency of the target wind turbine equipment Q is obtained, the absolute difference between the instantaneous characteristic frequency and the rated frequency of each time window is calculated, and the time windows with a difference exceeding 1 Hz are marked as abnormal frequency windows. The union of the abnormal energy window and the abnormal frequency window is taken to obtain the fault time window interval. The maximum and minimum values ​​of all instantaneous characteristic frequency values ​​in the fault time window interval are extracted as the fault frequency band. Based on the time window length L, the window shift L / 2 and the sampling frequency, the fault time window index is converted into the actual physical time to form the fault time. Finally, the determined fault type, severity, time of fault, and frequency band are combined into a diagnostic result, which is then encapsulated in JSON data format and sent to the monitoring terminal via Ethernet. The monitoring software then displays the result in text and waveform format.

[0026] This embodiment achieves automatic extraction of acoustic signature features and fault diagnosis of wind turbine equipment under given operating conditions through a distributed microphone array and collaborative time-frequency analysis, feature extraction and neural network model. It also provides fault location capabilities at specific times and frequencies. Compared with traditional vibration monitoring methods, it has the advantages of being non-contact and integrating multiple information dimensions.

[0027] Example 3

[0028] This embodiment is based on Embodiment 2, but differs in that it further discloses the specific implementation methods for voiceprint signal acquisition and voiceprint signal sequence determination, and focuses on explaining the deployment of the voiceprint acquisition array, the preset method of the predetermined power, and the discretization rules. The specific content is as follows: The target wind turbine equipment Q is a doubly fed asynchronous wind turbine with a rated power and a rated frequency corresponding to the grid frequency of 50 Hz. The acoustic fingerprint acquisition array consists of four microphones, which are respectively installed 0.3 meters above the high-speed shaft bearing housing of the gearbox, 0.5 meters to the side of the front bearing housing of the generator, 0.4 meters below the main shaft bearing housing, and 0.6 meters from the rear of the nacelle at the yaw system. All microphones are fixed with special clamps and equipped with wind noise suppression foam covers. The sampling frequency of the data acquisition card is set to 48 kHz with a resolution of 24 bits. The predetermined power is set by the operator in the human-machine interface of the monitoring terminal according to the operation and maintenance plan. For example, the rated power value is input as 2000 kW, and the monitoring power percentage is selected as 90%. The system automatically calculates the monitoring power window as 1800 kW ± 5%. When the real-time power value returned by the converter of the target wind turbine equipment Q falls into the monitoring power window and lasts for more than 10 seconds, the acoustic fingerprint acquisition module triggers valid acquisition, and all sound pressure data streams collected during this period are recorded as acoustic fingerprint signals W.

[0029] The specific method for dividing the voiceprint signal into time sequences is as follows: taking the start time of acquisition as the origin, the voiceprint signal W is cut into continuous non-overlapping segments with a fixed duration of 1 second, and recorded as W1, W2, ..., Wj in chronological order, where j is the total number of segments accumulated since the start of acquisition, which automatically increases as the voiceprint acquisition array continues to monitor. Each time a new voiceprint signal segment is generated, it is transmitted to the time-frequency analysis module in real time through a shared memory queue.

[0030] This embodiment improves the signal-to-noise ratio and operating condition consistency of acoustic signature signals for specific wind turbine equipment by using a multi-point, near-sound-source microphone arrangement and a precise power window triggering mechanism, thereby enhancing the separability of subsequent fault characteristics.

[0031] Example 4

[0032] This embodiment is based on Embodiment 3 and further discloses a method for time-frequency domain analysis to lock time-frequency signals, generate instantaneous feature frequency sequences, and construct associated time series groups. The specific content is as follows: The time-frequency analysis module extracts each signal segment from the voiceprint signal sequence W1, W2, ..., Wj one by one and performs a short-time Fourier transform. The time window length L is set to 2048 sampling points, which corresponds to a duration of approximately 42.7 milliseconds at a sampling rate of 48 kHz. The window function uses a Hamming window to suppress spectral leakage. The window shift is set to 1024 sampling points, which means the time window overlap rate is 50%. For the 2048 data points in each time window, zeros are padded to 4096 points and then the fast Fourier transform is calculated to obtain the spectral vector of the corresponding time window. Finally, the spectral vectors of all time windows are determined and arranged in chronological order to form the time-frequency distribution matrix ST(t,f) of each signal segment, where t is the time window index and f is the discrete frequency value between 0 Hz and 24 kHz. It should be noted that, considering the main information of the acoustic signature of the wind turbine equipment is concentrated in the low and mid frequency bands, the search range for extracting instantaneous feature frequencies is limited to 10 Hz to 2000 Hz. Within this range, for each time window t, the peak value of the ST(t,f) amplitude spectrum is retrieved, and the frequency value corresponding to the peak value is recorded as the instantaneous feature frequency of the corresponding time window. The instantaneous feature frequencies of all time windows in the current signal segment are arranged in chronological order to form the instantaneous feature frequency subsequence corresponding to the signal segment. As the signal segments are processed sequentially, several instantaneous feature frequency subsequences are obtained, and they are spliced ​​and combined in chronological order to form the global instantaneous feature frequency sequence F1, F2, ..., Fk, where k is the total number of time windows accumulated since the start of acquisition.

[0033] Next, a short-time energy sequence is constructed. For each time window t, 2048 original sampling point data points within that time window are extracted, and the sum of squares of the amplitudes of each sampling point is calculated. This sum is used as the short-time energy of that window. The short-time energies are arranged in the order of the time windows to form a short-time energy sequence E1, E2, ..., Ek of the same length as the instantaneous characteristic frequency sequence F1, F2, ..., Fk. Then, according to the same time window index o, the short-time energy Eo is paired with the instantaneous characteristic frequency Fo to form an associated time series group {Eo,Fo}, where o=1,2,...,k. The associated time series group {Eo,Fo} is stored in the memory of the industrial embedded computer in the form of a structure array for the fault feature extraction module to call.

[0034] Example 5

[0035] This embodiment builds upon Embodiment 4, detailing the specific construction method of the multidimensional fault feature vector, the structure of the fault diagnosis model, and the training method. It achieves refined fault feature extraction and diagnosis based on the associated time series groups generated in Embodiment 4, as detailed below: The fault feature extraction module extracts the associated time series group {Eo, Fo} and calculates and constructs a four-dimensional fault feature vector according to the following method: Based on the correlation relationship of the associated time series group {Eo,Fo}, the instantaneous feature frequency sequence F1,F2,...,Fk and the short-time energy sequence E1,E2,...,Ek are extracted; Calculate the Pearson correlation coefficient between the instantaneous characteristic frequency sequence F1, F2, ..., Fk and the short-time energy sequence E1, E2, ..., Ek, using the formula TZ1=Σ[(Eo-μE)(Fo-μF)] / [√Σ(Eo-μE)]. 2 √Σ(Fo-μF) 2 ], where μF and μE are the mean values ​​of the instantaneous characteristic frequency sequence F1, F2, ..., Fk and the short-time energy sequence E1, E2, ..., Ek, respectively. The calculated Pearson correlation coefficient TZ1 is denoted as the time-series synchronous correlation coefficient feature TZ1; First-order differences are performed on the instantaneous characteristic frequency sequences F1, F2, ..., Fk and the short-time energy sequences E1, E2, ..., Ek respectively: ΔEo = Eo - Er and ΔFo = Fo - Fr. The number of time windows with the same sign for ΔEo and ΔFo is counted and divided by the total number of differences k-1 to obtain the synchronization rate feature TZ2 of the direction of change, where r = 0-1 and the minimum value of r is 1. In this algorithm, Eo does not include E1, and similarly, Fo does not include F1. Next, retrieve all local peak points of the short-time energy sequences E1, E2, ..., Ek, and similarly retrieve all local peak points of the instantaneous characteristic frequency sequences F1, F2, ..., Fk. Obtain the time window index set PE of the local peak points of all short-time energy sequences and the time window index set PF of the local peak points of the instantaneous characteristic frequency sequences. In PE and PF, match the nearest neighbor peak point pairs in chronological order, calculate the absolute value of the difference between each pair of time window indices, and take the average value to obtain the peak offset difference feature TZ3. For any time window o, calculate the local fluctuation amplitude of short-term energy ΔEo_ca=|Eo-μE| / μE, and the local fluctuation amplitude of instantaneous characteristic frequency ΔFo_ca=|Fo-μF| / μF. Calculate the ratio Ro=ΔEo_ca / ΔFo_ca to form a ratio sequence R1,R2,...,Rk. Then calculate the coefficient of variation of the ratio sequence, which is the standard deviation σR of the ratio sequence R1,R2,...,Rk divided by the mean μR, denoted as the energy-frequency coupling difference feature TZ4. The above four features are combined into a multidimensional fault feature vector [TZ1,TZ2,TZ3,TZ4]; The fault diagnosis module incorporates a fault diagnosis model based on a multilayer perceptron (MLP). Its network structure includes an input layer with four neurons, corresponding to a multidimensional fault feature vector [TZ1, TZ2, TZ3, TZ4]. The first hidden layer contains 32 neurons, and the second hidden layer contains 16 neurons. The activation function for both hidden layers is the linear rectified function ReLU. The output layer is divided into two parallel branches: a fault type classification branch with K neurons (K = 6) corresponding to normal, rotor imbalance, shaft misalignment, bearing inner ring wear, gear tooth breakage, and blade crack, respectively. The fault type classification branch uses the Softmax activation function to output the probability of each category. The severity regression branch contains one neuron, using a linear activation function to output a severity value between 0 and 1, where 0 represents no fault and 1 represents complete failure.

[0036] The training process of the fault diagnosis model is as follows: First, the acoustic signature signals generated by several wind turbines of the same specifications as the target wind turbine Q during historical operation are collected, including data on 6 fault types and different degrees of fault status, for a total of 2000 sets of samples. For each set of samples, the same operations as in steps one to three are performed to extract the four-dimensional fault feature vector [TZ1, TZ2, TZ3, TZ4]. Professional analysts then label the corresponding fault type and severity labels to form the training sample set.

[0037] The training sample set was randomly divided into a training set and a validation set in a 4:1 ratio. During the training of the fault diagnosis model, the multidimensional feature vector of the training set was input into the multilayer perceptron. The predicted probability distribution of the classification branch and the predicted severity of the regression branch were obtained through forward propagation. The classification loss was calculated using the classification cross-entropy loss function, and the regression loss was calculated using the mean squared error loss function. The total loss function was set as Loss_total = 0.6 × Loss_class + 0.4 × Loss_reg. The Adam optimizer was used with a learning rate of 0.01 and a mini-batch size of 32. The weights and bias parameters of each layer of the network were updated through backpropagation. Each iteration of the training set is considered an iteration round, with a total of 200 training rounds. When the total loss of the validation set does not decrease for 10 consecutive rounds, early stopping is triggered, and the model parameters with the minimum validation loss are selected as the final trained fault diagnosis model.

[0038] When the fault diagnosis model performs diagnosis, the multi-dimensional fault feature vector extracted in real time is input into the trained fault diagnosis model. The fault type corresponding to the highest probability index in the softmax output of the classification branch is taken as the diagnosed fault type, and the floating-point value of the regression branch output is taken as the diagnosed severity. In this way, through four feature vectors with clear physical meaning, the synchronicity and difference between energy and frequency are fully quantified. Combined with the multi-task multilayer perceptron model, the synchronous high-precision identification of fault type and severity is achieved, which has better generalization performance than the single-task model.

[0039] Example 6

[0040] This embodiment is implemented based on embodiment 5, and further discloses the complete process of associating fault type and severity with local abnormal fluctuations, locating fault time intervals and frequency components, generating diagnostic results and outputting them to the monitoring terminal, as detailed below: The anomaly localization and output module extracts the short-time energy sequences E1, E2, ..., Ek and the instantaneous characteristic frequency sequences F1, F2, ..., Fk corresponding to the associated time series group {Eo, Fo} used for the current diagnosis from the storage area; The anomaly location and output module pre-stores the arithmetic mean of the short-time energy sequences of all time windows calculated according to the method in step two during 12 hours of continuous normal and fault-free operation of the target wind turbine equipment Q, which is used as the normal energy baseline value E_base; Next, the energy deviation between the short-term energy Eo and E_base for each time window o is calculated as Dev_Eo = |Eo - E_base|. At the same time, a preset energy deviation threshold of E_base × 15% is extracted, and time windows with energy deviation Dev_Eo greater than the energy deviation threshold are marked as abnormal energy windows. Next, the theoretical characteristic frequency F_rated of the target wind turbine equipment Q under rated operating conditions is obtained. The theoretical characteristic frequency F_rated is determined by the rated speed and bearing geometric parameters, for example, the characteristic frequency of the main bearing is 63.2 Hz. For the instantaneous characteristic frequency sequence, the frequency difference Dev_Fo=|Fo-F_rated| for each time window is calculated, and the preset frequency difference threshold = F_rated×0.8%, i.e. 0.5 Hz, is extracted. The time windows in which Dev_Fo is greater than the frequency difference threshold are marked as abnormal frequency windows. The abnormal energy window and the abnormal frequency window are combined to obtain the fault time window interval. The fault time window interval is composed of several continuous or discrete time window index segments. Extract the instantaneous characteristic frequency values ​​corresponding to all time window indices within the fault time window interval, and iterate through them to find the maximum frequency value F_max and the minimum frequency value F_min, forming the fault frequency band [F_min, F_max]. Convert the start and end time window indices of the fault time window interval according to the time window length L=2048 points, window shift S=1024 points and the known sampling rate to obtain the fault start time and end time, thus forming the fault time interval. Finally, the diagnostic fault type, diagnostic severity, and the aforementioned fault time interval and fault frequency band obtained in Example 5 are combined into a diagnostic result record, which is sent to the monitoring terminal in JSON string form via TCP / IP protocol. After being parsed by the monitoring terminal software, the record is highlighted in the alarm list on the interface for the operator.

[0041] This embodiment uses energy and frequency dual-dimensional deviation detection and time window-physical time conversion to map the fault diagnosis results to specific time points and frequency ranges in a fine-grained manner, enabling maintenance personnel to quickly confirm the precise time and frequency characteristics of the fault occurrence and shorten the fault troubleshooting time.

[0042] All data in the formulas described above have been calculated with dimensions removed. Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0043] The above description is merely an example and illustration of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the invention or exceed the scope defined in the claims, they should all fall within the protection scope of the present invention.

[0044] It should be stated that all user data collected in this application was collected with the user's consent and authorization. Furthermore, the uses of user data are legal and compliant, and the use and processing of user data comply with the relevant laws, regulations, and standards of the relevant regions.

Claims

1. A method for monitoring and diagnosing faults in wind turbine equipment using acoustic signatures, characterized in that, The method includes: Step 1: Based on the real-time monitoring of the acoustic signature signal of the target wind turbine equipment under a predetermined power operating state by the acoustic signature acquisition device, the acoustic signature signal is divided into time sequences in combination with time to determine the acoustic signature signal sequence associated with the target wind turbine equipment. Step 2: Perform time-frequency domain analysis on the acoustic signature signal sequence, lock the time-frequency signal, generate the instantaneous characteristic frequency sequence associated with the target wind turbine equipment, and simultaneously construct a short-time energy sequence, which is then associated with the instantaneous characteristic frequency sequence to form an associated time sequence group. Step 3: Extract the temporal synchronization features and difference features between the instantaneous feature frequency sequence and the short-time energy sequence in the associated time series group, and construct a multi-dimensional fault feature vector; Input multidimensional fault feature vectors into the fault diagnosis model to identify fault types and severity. Step four involves associating the identified fault type and severity with the corresponding local abnormal fluctuations in the associated time series group, locating the time interval and frequency components of the fault occurrence, generating diagnostic results that include fault type, fault time, and fault frequency band, and outputting them to the monitoring terminal.

2. The method according to claim 1, characterized in that, In step one, the specific method for real-time monitoring of the acoustic signature signal of the target wind turbine under a predetermined power operating state based on the acoustic signature acquisition device is as follows: The target wind turbine equipment is identified and denoted as Q; Based on the pre-deployed acoustic fingerprint acquisition array associated with the target wind turbine equipment Q, the acoustic fingerprint signal of the target wind turbine equipment Q in the predetermined power operation state is collected. The predetermined power is preset by the operator in combination with the rated power of the target wind turbine equipment Q. The voiceprint acquisition array acquires voiceprint signals at a preset acquisition frequency by the operator, records the voiceprint signals as W, and stores them.

3. The method according to claim 2, characterized in that, In step one, the specific method for determining the acoustic signature signal sequence associated with the target wind turbine equipment is as follows: Based on the time sequence of the discrete voiceprint signal W collected by the voiceprint acquisition array of the target wind turbine equipment Q, the voiceprint signal sequence W1, W2, ..., Wj associated with the target wind turbine equipment Q is obtained, where j is the total number of discrete voiceprint signals, which increases with the acquisition time of the voiceprint acquisition array.

4. The method according to claim 3, characterized in that, In step two, the specific method for performing time-frequency domain analysis on the acoustic signature signal sequence, locking the time-frequency signal, and generating the instantaneous characteristic frequency sequence associated with the target wind turbine equipment is as follows: Extract the voiceprint signal sequence W1, W2, ..., Wj, and perform a short-time Fourier transform, where the time window length is a preset length L, the window function is a Hamming window, and the window shift is L / 2; Transform the signal segments in the corresponding voiceprint signal sequence within each time window to the frequency domain to construct the time-frequency distribution matrix ST(t,f), where t is the index of the time window and f is the frequency value; For each time window t, the frequency component with the largest amplitude is extracted from the time-frequency distribution matrix ST(t,f) and used as the instantaneous characteristic frequency of the corresponding time window; The instantaneous characteristic frequencies of all time windows are sorted in chronological order to form the instantaneous characteristic frequency sequence F1, F2, ..., Fk associated with the target wind turbine equipment Q, where k is the total number of time windows.

5. The method according to claim 4, characterized in that, In step two, the specific method for constructing the associated time series group is as follows: Extract signal segments from the voiceprint signal sequences corresponding to all time windows, calculate the short-time energy of the signal segments in each time window, and arrange them in chronological order to form a short-time energy sequence E1, E2, ..., Ek; The short-time energy sequences E1, E2, ..., Ek are extracted synchronously and matched with the instantaneous feature frequency sequences F1, F2, ..., Fk according to the time sequence to form the associated time sequence group {Eo, Fo}, where o is the counting index, with a value range from 1 to k.

6. The method according to claim 5, characterized in that, In step three, the specific method for constructing the multidimensional fault feature vector is as follows: Extract the associated time series group {Eo, Fo}; Based on the correlation, the Pearson correlation coefficient between the short-time energy sequence E1, E2, ..., Ek and the instantaneous characteristic frequency sequence F1, F2, ..., Fk is calculated and used as the time-series synchronous correlation coefficient feature TZ1; Then, based on the correlation, the short-time energy sequence and the instantaneous characteristic frequency sequence are subjected to first-order difference, and the proportion of time windows with the same difference sign is statistically analyzed and denoted as the synchronization rate feature TZ2 of the change direction. Determine the time window indices of all peak points for the short-time energy sequence and the instantaneous characteristic frequency sequence respectively, and calculate the mean of the absolute values ​​of the differences between the peak point time window indices, which is denoted as the peak offset difference feature TZ3; For short-time energy sequences and instantaneous characteristic frequency sequences, the ratio of local fluctuation amplitudes of short-time energy and instantaneous characteristic frequency within the same time window is calculated. The local fluctuation amplitude ratio sequence is obtained in chronological order, and the coefficient of variation of the local fluctuation amplitude ratio sequence is calculated. This is denoted as the energy-frequency coupling difference feature TZ4. The timing synchronization correlation coefficient feature TZ1, the change direction synchronization rate feature TZ2, the peak offset difference feature TZ3, and the energy frequency coupling difference feature TZ4 are combined to form a multidimensional fault feature vector [TZ1,TZ2,TZ3,TZ4].

7. The method according to claim 6, characterized in that, In step three, the specific method for inputting the multidimensional fault feature vector into the fault diagnosis model to identify the fault type and severity is as follows: A fault diagnosis model based on a multilayer perceptron is constructed, which includes an input layer, a first hidden layer, a second hidden layer, and an output layer; The input layer contains four neurons for receiving multidimensional fault feature vectors [TZ1, TZ2, TZ3, TZ4]. The first hidden layer and the second hidden layer contain 32 and 16 neurons respectively, both of which use the ReLU activation function; The output layer includes a parallel fault type classification branch and a severity regression branch. The fault type classification branch contains K neurons and uses the Softmax activation function, where K is the total number of fault types to be identified. The severity regression branch contains 1 neuron and uses a linear activation function. Input the multidimensional fault feature vector [TZ1,TZ2,TZ3,TZ4] into the trained fault diagnosis model, take the fault type corresponding to the maximum probability value in the output of the Softmax activation function as the diagnosed fault type, take the output value of the linear activation function as the diagnosed severity, and output the diagnosed fault type and the diagnosed severity.

8. The method according to claim 7, characterized in that, In step three, the training method for the fault diagnosis model specifically includes: Acquire historical acoustic signature signals of several wind turbines of the same specifications as the target wind turbine equipment Q under different fault types and severity states. Extract multidimensional fault feature vectors based on steps one to three, and have the operators label the fault type and severity to form a training sample set. The multidimensional fault feature vector in the training sample set is used as input, and the forward propagation is performed through a multilayer perceptron. The fault type classification branch outputs the predicted fault type probability distribution, and the severity regression branch outputs the predicted severity value. The classification loss between the predicted fault type probability distribution and the fault type label is calculated using the classification cross-entropy loss function, and the regression loss between the predicted severity value and the severity label is calculated using the mean squared error loss function. The classification loss and regression loss are weighted and summed according to preset weights to obtain the total loss function. The Adam optimization algorithm is adopted to minimize the total loss function. The network weights and bias parameters of the multilayer perceptron are updated through backpropagation, and the training is iterated until the total loss function converges, thus completing the training of the fault diagnosis model.

9. The method according to claim 7, characterized in that, In step four, the specific method for generating diagnostic results containing fault type, fault time, and fault frequency band and outputting them to the monitoring terminal is as follows: Extract the short-time energy sequences E1, E2, ..., Ek and the instantaneous feature frequency sequences F1, F2, ..., Fk corresponding to the associated time series group {Eo, Fo}; The average of the historical short-time energy series constructed in step two under normal and fault-free operation of the target wind turbine equipment Q is used as the normal energy baseline value. For short-time energy sequences, calculate the energy deviation between the short-time energy and the normal energy baseline value for each time window, and mark the time windows with energy deviations greater than a preset energy deviation threshold as abnormal energy windows; Obtain the rated frequency of the target wind turbine equipment Q. For the instantaneous characteristic frequency sequence, calculate the frequency difference between the instantaneous characteristic frequency and the rated frequency for each time window. Mark the time window where the absolute value of the frequency difference is greater than the preset frequency difference threshold as an abnormal frequency window. The fault time window interval is obtained by taking the union of the abnormal energy window and the abnormal frequency window; Extract all instantaneous characteristic frequency values ​​within the fault time window interval, calculate their maximum and minimum frequencies, and form a fault frequency band; The fault time window interval is converted into actual time based on the time window length L and the window shift L / 2 to form the fault time. The diagnostic results are combined by combining the fault type, severity, time of fault, and frequency band, and then output to the monitoring terminal.

10. A wind turbine equipment fault acoustic signature monitoring and diagnostic system, characterized in that, The system includes: The voiceprint acquisition module monitors the voiceprint signal of the target wind turbine equipment in real time under a predetermined power operating state, and divides the voiceprint signal into time sequence based on time to determine the voiceprint signal sequence associated with the target wind turbine equipment. The time-frequency analysis module performs time-frequency domain analysis on the acoustic signal sequence, locks the time-frequency signal, generates the instantaneous characteristic frequency sequence associated with the target wind turbine equipment, and simultaneously constructs a short-time energy sequence, which is then associated with the instantaneous characteristic frequency sequence to form an associated time sequence group. The fault feature extraction module extracts the temporal synchronization and difference features between the instantaneous feature frequency sequence and the short-time energy sequence in the associated time series group, and constructs a multi-dimensional fault feature vector. The fault diagnosis module has a built-in trained fault diagnosis model, receives multi-dimensional fault feature vectors, and identifies the fault type and severity. The anomaly localization and output module associates the identified fault type and severity with the corresponding local anomaly fluctuations in the associated time series group, locates the time interval and frequency components of the fault occurrence, generates diagnostic results including fault type, fault time, and fault frequency band, and outputs them to the monitoring terminal. The monitoring terminal interacts with operators in real time, coordinating data transmission and operation between various modules.

Images