An industrial equipment running state monitoring method based on voiceprint recognition

Abstract

The application provides an industrial equipment running state monitoring method based on voiceprint recognition, and relates to the technical field of fault monitoring. The application comprises the following steps: collecting a voiceprint signal of equipment running containing background noise, then performing wavelet transform and peak extraction, and performing noise suppression; constructing a normal running state voiceprint library of the equipment, combining a multi-index fusion equipment fault monitoring method, judging the current running state of the equipment as normal or fault; constructing an equipment fault running state voiceprint library, labeling different labels for different fault categories, training a multi-classification fault identification model based on joint entropy features and support vector machines, and realizing fault identification of equipment running. The application can effectively suppress complex background noise and has the effect of signal enhancement, effectively improves the accuracy and robustness of equipment fault monitoring and fault identification based on the multi-index fusion equipment fault monitoring method and the joint entropy feature and support vector machine based equipment fault identification method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault monitoring, in particular to an industrial equipment operation state monitoring method based on voiceprint recognition. BACKGROUND

[0002] In the field of equipment fault diagnosis, the internal state of the equipment and the external performance information can be combined together as the basis for diagnosis, and the equipment operation state is evaluated based thereon. Based on different selected signal sources, the fault diagnosis of the equipment can be roughly divided into fault diagnosis based on vibration signals and fault diagnosis based on voiceprint signals.

[0003] The fault diagnosis based on vibration signals, i.e. the fault detection of the equipment is realized by analyzing the narrowband characteristic spectrum of the vibration signals of the equipment. Although certain achievements have been made, the vibration signals cannot fully represent the fault information, i.e. the processing bandwidth and the fault information pickup capability are insufficient, so that it can only diagnose the faults that have occurred and are relatively serious, and the diagnosis capability for some hidden faults of the equipment is insufficient, and due to the limitation of the information pickup capability, the fault diagnosis accuracy is also relatively low. In addition, the disadvantage of this contact type measurement is that the sensors attached to the surface of the machine will vibrate with the machine vibration, and the vibration of the collection equipment itself will not only interfere with the collection process, but also easily cause damage to the sensors, so it cannot meet the demand of the main equipment operation state monitoring in the industrial field.

[0004] The fault diagnosis based on voiceprint signals mainly includes auscultation method, near-field measurement method and blind signal processing method.

[0005] The auscultation method is that an experienced maintenance personnel uses an iron stick to hit the equipment to be detected, and judges the equipment operation state and troubleshoots the fault cause according to the sound emitted. This method is still used in many operation sites at present, but it is based on the intuition of the maintenance personnel with years of maintenance experience, and has no scientific theory support, so that its reliability is not high and it is difficult to master.

[0006] The near-field measurement method is that a collector is placed near the equipment, the voiceprint signals are measured at a close distance, the fault position is found and located, and the specific fault position is located by comparing the sound signals at different positions. This method solves the disadvantage of the auscultation method that only relies on experience without scientific theory, but the voiceprint signals generated by the mechanical equipment have serious interference, and even if the collector is placed near the sound source, the interference influence of other sound sources cannot be excluded, so this method can only be used for simple fault identification or positioning, and the effect is poor under noise interference.

[0007] Blind signal processing method, that is, separating target signals from mixed signals to avoid cross interference, combined with fault detection and diagnosis technology, can realize real-time monitoring of the running state of equipment. The blind signal processing method solves the problem of poor signal-to-noise ratio in the field of acoustic fault diagnosis for a long time to a certain extent, and has been greatly improved compared with the near-field measurement method, but because the linear instantaneous model used is too idealized, it can only have good effect in special indoor environment and cannot meet the actual application conditions of industrial field.

[0008] Therefore, the present application provides an industrial equipment running state monitoring method based on voiceprint recognition to solve the above problems. SUMMARY

[0009] To solve the above problems, the present application provides an industrial equipment running state monitoring method based on voiceprint recognition, which performs wavelet transform and peak extraction on the equipment running state voiceprint signal, solves the problem that the effective acoustic signal of the equipment running state is seriously interfered by background noise in a complex noise environment, constructs a device normal running state voiceprint library, monitors the target voiceprint signal, solves the problem that the equipment fault acoustic sample is difficult to collect and cannot be directly based on the fault acoustic sample to realize equipment fault monitoring, and finally trains a fault recognition model based on the existing fault sample to recognize the fault category of the target fault voiceprint signal.

[0010] The present application is an industrial equipment running state monitoring method based on voiceprint recognition, comprising the following steps:

[0011] Acquiring the voiceprint signal containing background noise of the equipment running based on the acoustic sensor;

[0012] Performing wavelet transform on the collected voiceprint signal containing background noise to obtain the voiceprint signal after preliminary noise reduction processing, and then performing peak extraction to obtain the discrete time sequence corresponding to the peak signal, and further suppressing the background noise;

[0013] Continuously accumulating the peak signals under the normal running state of the equipment to construct a device normal running state voiceprint library, calculating the indexes of the peak signals under the normal running state of multiple equipment, combining the multi-index fusion equipment fault monitoring method, judging whether the target peak signal corresponding to the current running state of the equipment matches the constructed device normal running state voiceprint library, if matched, it means that the running state of the equipment is normal, otherwise, the target branch signal is considered as a fault state voiceprint signal, and the current equipment running state is faulty;

[0014] Obtaining a plurality of device fault operation voiceprint signal samples, constructing a device fault operation state voiceprint library, and labeling different labels for different fault categories, extracting features of peak signals of each device fault operation voiceprint signal sample, dividing into a plurality of fault sample data sets, and using a support vector machine to train and learn the plurality of sample data sets respectively to obtain a multi-classifier composed of a plurality of binary classifiers, thereby realizing device operation fault recognition.

[0015] Further, based on the acoustic sensor collecting the device operation voiceprint signal, the sampling frequency, sampling data duration and sampling interval time need to be set in advance.

[0016] Further, the wavelet transform of the collected background noise voiceprint signal comprises the following steps:

[0017] The selected basis function and the number of decomposition layers are used for wavelet decomposition of the collected voiceprint signal, wavelet coefficient processing is performed based on the 3sigma rule in Chebyshev inequality, the noise corresponding wavelet coefficient is removed, and then wavelet reconstruction is performed to obtain the voiceprint signal with preliminary noise reduction.

[0018]

[0019] σ=median({|w i |,i≤k}) / 0.6745

[0020] wherein w σ is the wavelet coefficient obtained after wavelet coefficient processing of the wavelet coefficient w, sgn(w) is the sign of the wavelet coefficient, and sigma is the noise variance estimate, i.e., the wavelet coefficient value within the interval [-3sigma, 3sigma] is set to 0, the wavelet coefficient greater than 3sigma is uniformly reduced by 3sigma, and the wavelet coefficient less than -3sigma is uniformly increased by 3sigma. i |,i≤k} is a set of wavelet coefficient modules, and median({|w i |,i≤k}) is the median of the wavelet coefficient module.

[0021] Further, the peak extraction comprises the following steps:

[0022] Based on the selected time interval, the peak value of the wavelet transformed voiceprint signal is extracted, the maximum and minimum values of the voiceprint signal in each time interval are obtained, and the discrete time sequence of the peak value signal corresponding to the voiceprint signal under the device operation state is obtained.

[0023] Further, the peak signal S i (n) obtained after wavelet transform and peak extraction combined noise suppression under the normal operation state of the device is continuously accumulated to construct a device normal operation state voiceprint library {S i(n), i≤s}, set the normal operating state of the device and store s peak signals in the voiceprint library;

[0024] Calculate the midline value mid(S) of each peak signal. i (n)), maximum value max(S) i (n)), minimum value min(S) i (n)), and the resulting vector [min(S i (n)), mid(S i (n)),max(S i [n] is associated with and stored in relation to the corresponding peak signal;

[0025] Peak signal S in the voiceprint library during normal operation of computing devices i (n) The average DWT distance d between the constructed equipment and the voiceprint database under normal operating conditions i1 Mean deviation of the centerline d i2 Maximum value offset from mean d i3 Minimum value offset from mean d i4 The aforementioned feature values ​​are then sorted in ascending order, fitted with a normal distribution, and the upper limit values ​​of each feature are obtained, forming a feature upper limit value vector D. max =[d1,d2,d3,d4], which is associated with the voiceprint library for normal device operation.

[0026] Furthermore, the upper limit value of the feature is the feature value at the specified upper limit probability of the corresponding feature.

[0027] Furthermore, DTW is an algorithm used to calculate the similarity between time series of unequal lengths. It uses dynamic programming to find the shortest distance between two time series. Midline offset, maximum value offset, and minimum value offset are used to measure the amplitude and fluctuation range of the two series. For two peak signals S... i (n) and S j For (n), the line offset Maximum offset Minimum offset The formulas are as follows:

[0028]

[0029] Furthermore, the specific steps for determining whether the target peak signal corresponding to the current operating state of the device matches the constructed voiceprint database of the device's normal operating state are as follows:

[0030] Real-time acquisition of the peak signal S of the target acoustic signature under the current device operating status c (n), and obtain the line value mid(S) within it. c (n)), maximum value max(S)c (n))、min(S c (n));

[0031] Calculate the feature vector D c (n) composed of the average of the DWT distance, the average of the median offset, the average of the maximum offset and the average of the minimum offset between the target peak signal S c (n) and the device normal operating state voiceprint library c1 , c2 , c3 , c4 ;

[0032] If each element in the feature vector of the target peak signal S c (n) is less than the corresponding element in the feature upper limit value vector D max , it is determined that the target peak signal S c (n) matches the device normal operating state voiceprint library, it is a normal peak signal, and the current device operating state is normal, otherwise it is determined to be a fault peak signal, and the current device operating state is abnormal.

[0033] Further, a plurality of device fault operating voiceprint signal samples are obtained, a device fault operating state voiceprint library is constructed, and different labels are marked for different fault categories, obtaining m fault categories;

[0034] The multi-scale entropy and power spectrum entropy of the peak signal of the device fault operating voiceprint signal sample are extracted to obtain a feature vector, and according to the fault categories, m category sample data {X1}, {X2}, …, {X m}, {X i}, 1≤i≤m represents a set composed of all feature vectors corresponding to the i-th fault category;

[0035] The fault sample data is divided into m positive and negative sample combinations in the form of OVR, that is, ({X1}, {X o1}), ({X2}, {X o2}), …, ({X m}, {X om}), and {X oi}, 1≤i≤m represents samples randomly and uniformly extracted from sample sets of other categories except the i-th fault category sample set;

[0036] Using support vector machines, m sets of sample sets are trained and learned, and m binary classifiers {f1(x), f2(x), …, f m (x)} are obtained, which can form a device fault recognition model;

[0037] When the target acoustic signature signal is detected as an equipment malfunction acoustic signature signal, the multi-scale entropy and power spectral entropy of the corresponding peak signal are extracted and substituted into the obtained equipment fault identification model to obtain the probability {f1,f2,…,f} corresponding to each fault category of the target acoustic signature signal. m The fault category with the highest probability is the fault category of the current target voiceprint signal.

[0038] Furthermore, as the equipment continues to operate, the continuously acquired voiceprint signals of the equipment's operating status are divided into normal voiceprint signals or fault voiceprint signals, and added to the corresponding voiceprint libraries respectively. The voiceprint libraries for normal equipment operation status and equipment fault operation status are updated regularly. For the updates of the voiceprint library for equipment fault operation status, the equipment fault identification model is retrained periodically or quantitatively to ensure the accuracy and robustness of equipment fault identification.

[0039] The beneficial effects of this invention are as follows:

[0040] 1. This invention is based on a joint noise suppression method of wavelet transform and peak extraction. It combines peak extraction and wavelet transform. First, wavelet transform is used to denoise the acquired device operation acoustic signal. Then, peak extraction is performed on the denoised signal based on a selected time interval to obtain the discrete time sequence of the peak signal. This reduces the impact of wavelet basis selection in wavelet transform on the final result, thereby avoiding the problem of wavelet basis interfering with all subsequent detection steps. Wavelet transform is used to replace the high-pass filter commonly used in peak detection, making the noise reduction step more thorough.

[0041] 2. This invention provides a multi-indicator fusion-based equipment fault monitoring method. It uses the average dynamic time curvature distance, average centerline offset, average maximum value offset, and average minimum value offset between the peak signal in normal state and the constructed normal state acoustic signature database as judgment indicators. The upper limit value vector of the corresponding feature in the normal state acoustic signature database is used as a benchmark to determine whether the target peak information matches the database. If they match, it is considered a normal state acoustic signature signal, indicating normal equipment operation; otherwise, it is considered a fault state acoustic signature signal, indicating abnormal equipment operation. This multi-indicator fusion not only effectively describes the overall trend of the signal but also effectively avoids the problem of unequal time series lengths in acoustic signature signals, thus ensuring the accuracy of equipment fault monitoring and exhibiting good robustness.

[0042] 3. This invention provides a device fault identification method based on joint entropy features and support vector machines. It selects the multi-scale entropy and power spectral entropy of the fault acoustic signature signal as fault feature parameters. Based on the fault category, the fault samples are divided into m positive and negative sample combinations using OVR (Optical Variation Responsibility). Support vector machines are then used to train and learn on each of the m sample sets, resulting in a device fault identification model composed of m binary classifiers. This invention selects multi-scale entropy and power spectral entropy, which can effectively characterize the operating state of the equipment, as feature parameters. It fully extracts the multi-timescale and energy distribution features of the acoustic signature signal and uses support vector machines to construct the device fault identification model, thereby obtaining a fault identification model with high accuracy and robustness. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort, wherein:

[0044] Figure 1 This is an overall flowchart of the method of the present invention;

[0045] Figure 2 This is a graph showing the peak signal and the signal midline.

[0046] Figure 3 This is a joint noise suppression process based on wavelet transform and peak extraction;

[0047] Figure 4 Peak extraction signal diagram of the voiceprint signal;

[0048] Figure 5 Select a graph for the upper limit of the feature. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0050] It should be noted that the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0051] This invention proposes a method for monitoring the operating status of industrial equipment based on voiceprint recognition, which specifically includes three parts:

[0052] I. Addressing the problem that effective acoustic signature signals are severely interfered with by background noise during equipment operation in complex noise environments.

[0053] Many critical devices operate in very complex environments. Friction between components and complex external interference noise can make it extremely difficult to extract truly useful acoustic signals, making it hard to detect weak signals in the early stages of the equipment. Therefore, it is essential to perform broadband background noise suppression on signals collected by acoustic sensors under complex operating conditions.

[0054] In complex noise environments, the acoustic signature of equipment operation exhibits non-stationarity under noise interference. Wavelet transform can extract and analyze non-stationary signals through time-frequency domain transformation. However, in the wavelet decomposition process, the selection of different wavelet basis functions will have a significant impact on the subsequent signal processing. Once the wavelet basis is determined, it cannot be changed in the subsequent analysis process, which will affect the detection and identification of fault acoustic signature signals.

[0055] This invention employs wavelet transform and peak extraction to suppress background noise in voiceprint signals. First, wavelet transform is performed on the acquired voiceprint signal. Based on selected basis functions and decomposition levels, the voiceprint signal is decomposed into wavelets, and the wavelet coefficients are processed according to the 3sigma criterion to remove noise-corresponding wavelet coefficients. Then, wavelet reconstruction is performed to obtain a preliminary denoised signal. Next, peak extraction is performed on the denoised signal based on selected time intervals to obtain a discrete-time sequence of the peak signals corresponding to the effective voiceprint signals under device operating conditions. This reduces the impact of wavelet basis selection in the wavelet transform on the final result, thus solving the problem of interference from a fixed wavelet basis in subsequent detection. Furthermore, wavelet transform replaces the high-pass filtering commonly used in peak detection techniques, making the denoising process more thorough and further suppressing the influence of background noise. Therefore, the joint noise suppression method based on wavelet transform and peak extraction has strong noise suppression and signal enhancement effects under complex background noise conditions.

[0056] Second, regarding the problem of limited acoustic samples for equipment failures, difficulty in collecting them, and inability to directly monitor equipment failures based on these samples.

[0057] Considering the low frequency of equipment failures in actual operation, resulting in limited fault sample data, the equipment operating status is divided into fault state and normal state. A fault detection method based on a voiceprint library of normal operating states is adopted. First, peak signals corresponding to voiceprint signals under normal operating conditions are continuously accumulated to construct a voiceprint library for normal operating conditions. Then, combined with a multi-index fusion method for equipment fault monitoring, the centerline value, maximum value, and minimum value of the target peak signal are obtained. Next, the average DTW distance, average centerline offset, average maximum value offset, and average minimum value offset between the target peak signal and the constructed voiceprint library for normal operating conditions are calculated. These are compared with the corresponding upper limit values ​​of the features in the voiceprint library for normal operating conditions to determine whether the target peak signal is a normal peak signal. If it is determined to be a normal peak signal, the equipment operating state is normal; otherwise, the equipment operating state is abnormal.

[0058] Third, regarding the problem of numerous equipment fault types and the difficulty in identifying fault types.

[0059] To address the challenges of limited equipment fault samples, numerous fault categories, and relatively high difficulty in fault identification in practical work, we selected the multiscale sample entropy (MSE) and power spectral entropy (PSE) of fault acoustic signals as feature parameters, and combined them with support vector machines (SVM) to achieve the identification of multiple types of equipment faults.

[0060] The fault acoustic signature signal in the operating state of the equipment contains rich fault information. Entropy is a measure of time series complexity and can be used to describe the randomness of the equipment. When the operating state of the equipment changes, the entropy value of the acoustic signature signal will also change with the increase of complexity. The complexity of the time series is positively correlated with the multi-scale entropy value corresponding to the scale factor, and the power spectrum entropy can reflect the complexity of the energy distribution of the time series signal in its frequency domain. Therefore, the multi-scale entropy and power spectrum entropy of the fault acoustic signature signal can be used as fault feature parameters.

[0061] By labeling fault acoustic signal samples with different tags according to different fault categories, the multi-scale entropy and power spectral entropy of the peak signals of equipment fault operation acoustic signal samples are extracted. The fault acoustic signal samples are divided in the form of OVR (One vs Rest), and a support vector machine is used for training and learning to obtain a multi-class fault recognition model composed of binary classifiers, thereby realizing equipment fault recognition.

[0062] Example:

[0063] This invention is a method for monitoring the operating status of industrial equipment based on voiceprint recognition. The overall process is as follows: Figure 1 As shown.

[0064] 1. Voiceprint signal acquisition

[0065] After pre-setting the sampling frequency (1kHz), sampling data duration (2s), and sampling interval (3s), the acoustic sensor is used to collect the acoustic signature signal containing background noise from the device.

[0066] 2. Joint noise suppression based on wavelet transform and peak extraction

[0067] Wavelet transform and peak extraction operations are performed on the acquired voiceprint signal containing background noise to obtain the discrete-time sequence of the peak signal corresponding to the voiceprint signal. The specific operation process is as follows: Figure 3 As shown.

[0068] Considering the significant and rapid attenuation characteristic of acoustic fingerprint signals from industrial equipment collected by acoustic sensors, this embodiment selects the Daubechies wavelet as the basis function for wavelet transform and chooses a decomposition level of 3-5 layers to perform wavelet decomposition on the acoustic fingerprint signal containing background noise. Then, based on the 3sigma criterion in Chebyshev's inequality, wavelet coefficient processing is performed to remove wavelet coefficients corresponding to noise, followed by wavelet reconstruction to obtain the preliminary denoised acoustic fingerprint signal. The wavelet coefficient processing formula is as follows:

[0069]

[0070] σ=median({|w i |,i≤k}) / 0.6745

[0071] Among them, w σ The wavelet coefficients are obtained after wavelet coefficient processing, where sgn(w) represents the sign of the wavelet coefficients, and σ is the noise variance estimate. Specifically, wavelet coefficients within the interval [-3σ, 3σ] are set to 0, wavelet coefficients greater than 3σ are uniformly subtracted by 3σ, and wavelet coefficients less than -3σ are uniformly added by 3σ. i |,i≤k} is the set of wavelet coefficient moduli obtained from wavelet decomposition, median({|w i |,i≤k}) represents the median of the wavelet coefficient modulus.

[0072] Peak extraction is performed on the wavelet-transformed acoustic signature signal based on selected time intervals to obtain the maximum and minimum values ​​of the acoustic signature signal within each time interval, resulting in a discrete-time sequence of the peak signal corresponding to the acoustic signature signal under device operating conditions. Peak extraction is as follows: Figure 4 As shown.

[0073] 3. Construction of voiceprint database for normal equipment operation

[0074] The peak signal S obtained after continuous accumulation of noise suppression based on wavelet transform and peak extraction under normal equipment operating conditions. i (n), construct the voiceprint database of the normal operating status of the equipment {S} i (n), i≤s}, set the normal operating state of the device and store s peak information in the voiceprint library;

[0075] Calculate the midline value mid(S) of each peak signal. i (n)), maximum value max(S) i (n)), minimum value min(S) i (n)), such as Figure 2 As shown, the resulting vector [min(S)] is... i (n)), mid(S i (n)),max(S i [n] is stored in association with the peak signal;

[0076] Calculate the peak signal S of each normal state in the voiceprint library. i (n) Mean d of the dynamic time warping distance (DWT distance) between the constructed normal-state voiceprint database and the database. i1 Mean deviation of the centerline d i2 Maximum value offset from mean d i3 Minimum value offset from mean d i4 Then, all the above features (mean DWT distance, mean median offset, mean maximum offset, and mean minimum offset) are sorted in ascending order, and a normal distribution is fitted. The feature values ​​at which each feature value falls within a specified upper limit probability (90%, which can be adjusted based on the actual scenario) are taken as feature upper limit values, and a feature upper limit value vector D is formed. max =[d1,d2,d3,d4], associated with the voiceprint database of equipment production and operation status, with the upper limit of the feature values ​​selected as follows. Figure 5 As shown, where d i This is the upper limit value of the feature when the upper probability is 90%.

[0077] DTW is a method for calculating the similarity between time series of unequal lengths. This algorithm uses dynamic programming to find the shortest distance between two time series. Midline offset, maximum offset, and minimum offset are used to measure the amplitude and fluctuation range of the two series. For two peak signals S... i (n) and S i For (n), the line offset Maximum offset Minimum offset As shown in the following formula.

[0078]

[0079] Among the above features, centerline offset, maximum value offset, and minimum value offset are used to measure the amplitude and fluctuation range of the two sequences. DWT can effectively avoid the problem of unequal lengths of acoustic signal time series. By integrating these indicators to construct the equipment fault monitoring model, the peak signal features can be better integrated, which can effectively improve the accuracy of equipment fault monitoring and have good robustness.

[0080] 4. Equipment fault monitoring based on multi-indicator fusion

[0081] Real-time acquisition of the peak signal S of the target acoustic signature during device operation. c (n), and obtain the line value mid(S) within it. c (n)), maximum value max(S) c (n)), minimum value min(S) c (n)); Calculate the target peak signal S c (n) The feature vector D is composed of the mean DWT distance, mean centerline offset, mean maximum offset, and mean minimum offset between the DWT distance and the voiceprint database under normal equipment operation. c =[d c1 ,d c2 ,d c3 ,d c4 ];

[0082] If the target peak signal S c (n) Each element in the eigenvector is less than the upper limit value of the eigenvector D. max The corresponding element in the value is then used to determine the target peak signal S. c (n) If it matches the voiceprint library of the normal operating status of the equipment, it is a normal peak signal and the current operating status of the equipment is normal. Otherwise, it is identified as a fault peak signal and the current operating status of the equipment is abnormal.

[0083] Furthermore, as the equipment continues to operate, the voiceprint database of the normal operating status can be updated regularly, thereby ensuring that the voiceprint database of the normal operating status is more consistent with the actual status of the equipment, and ensuring the accuracy and robustness of equipment fault monitoring based on multi-indicator fusion.

[0084] 5. Equipment Fault Identification Based on Joint Entropy Features and Support Vector Machines

[0085] As the equipment continues to operate, the number of equipment failures increases. Multiple equipment failure operation acoustic signal samples are acquired to construct an equipment failure operation status acoustic database. Different labels are assigned to different failure categories, resulting in m failure categories. Once the number of failures in each category reaches a certain requirement, the equipment failure identification model can be trained based on the obtained failure samples. The specific training process is as follows:

[0086] By extracting the multiscale sample entropy (MSE) (based on the characteristics of peak extraction, the scale factor must be an integer multiple of 2; in this embodiment, the scale factor is set to 2) and the power spectral entropy (PSE) of the peak signal from the acoustic signature of equipment malfunction operation samples as sample features, m categories of sample data {X1}, {X2}, ..., {X...} can be obtained. m}, {X i}, 1≤i≤m represents the set of all feature vectors corresponding to the i-th fault category.

[0087] The fault sample data is divided into m positive and negative sample combinations ({X1}, {X2}) in the form of OVR (One vs. Rest). o1}), ({X2}, {X o2}),…,({X m},{X om}), while {X oi}, 1≤i≤m represents a sample randomly and uniformly drawn from the sample sets of all categories except the i-th fault category sample set;

[0088] By training m sets of samples using support vector machines, we can obtain m binary classifiers {f1(x), f2(x), ..., f...} m The m binary classifiers (x) can form the fault identification model of the equipment.

[0089] When the target acoustic signature signal is detected as an equipment malfunction acoustic signature signal, the multi-scale entropy and power spectral entropy of the corresponding peak signal are extracted and substituted into the obtained equipment fault identification model (m binary classifiers) to obtain the probability {f1, f2, ..., f} corresponding to each fault category. m The fault category with the highest probability is the most likely fault category at present.

[0090] Furthermore, as equipment continues to generate faults during operation, the equipment fault operation status voiceprint database also expands. This allows for periodic or quantitative retraining of the equipment fault identification model, thereby ensuring the accuracy and robustness of equipment fault identification based on joint entropy features and support vector machines.

[0091] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A method for monitoring the operating status of industrial equipment based on voiceprint recognition, characterized in that, Includes the following steps: Acoustic sensors are used to collect acoustic signature signals from devices that contain background noise. Wavelet transform is performed on the collected voiceprint signal containing background noise to obtain the voiceprint signal after preliminary noise reduction. Then, peak extraction is performed to obtain the discrete time series of the corresponding peak signal to further suppress background noise. By continuously accumulating peak signals under normal operating conditions, a soundprint library for normal operating conditions is constructed. Indicators of multiple peak signals under normal operating conditions are calculated. Combined with a multi-indicator fusion method for equipment fault monitoring, it is determined whether the target peak signal corresponding to the current operating state of the equipment matches the constructed soundprint library for normal operating conditions. If they match, it indicates that the equipment is operating normally. Otherwise, the target branch signal is considered a fault soundprint signal, and the current operating state of the equipment is faulty. Acquire several equipment fault operation acoustic signal samples, construct an equipment fault operation status acoustic database, and label different fault categories. Extract features from the peak signals of each equipment fault operation acoustic signal sample, divide it into multiple fault sample datasets, and use support vector machines to train and learn on multiple sample datasets respectively, to obtain a multi-class fault recognition model composed of multiple binary classifiers, so as to realize the fault recognition of equipment operation. The peak signal obtained after wavelet transform and peak extraction combined with noise suppression is continuously accumulated under normal equipment operating conditions. Build a voiceprint database of normal equipment operation status Set the device to normal operating status and store s peak signals in the voiceprint library; Calculate the midline value of each peak signal. Maximum value Minimum value and the resulting vector Associate and store the data with the corresponding peak signal; Peak signal in the voiceprint library during normal operation of computing devices Average DWT distance between the constructed equipment and the voiceprint database under normal operating conditions Mean deviation of the midline Maximum value offset from mean Minimum value offset from mean The aforementioned feature values ​​are then sorted in ascending order, fitted with a normal distribution, and the upper limit values ​​of each feature are obtained, forming a feature upper limit value vector. It is stored in association with the voiceprint database in relation to the normal operating status of the equipment; The upper limit value of the feature is the feature value at the specified upper limit probability of the corresponding feature; DTW is an algorithm used to calculate the similarity between time series of unequal lengths. It uses dynamic programming to find the shortest distance between two time series. Midline offset, maximum offset, and minimum offset are used to measure the amplitude and fluctuation range of the two series. For two peak signals... and In other words, the line offset Maximum value offset Minimum value offset The formulas are as follows: ； The specific steps for determining whether the target peak signal corresponding to the current operating state of the device matches the constructed voiceprint database of the device's normal operating state are as follows: Real-time acquisition of the peak signal of the target acoustic signature under the current device operating status. And obtain its line value. Maximum value Minimum value ; Calculate the target peak signal The feature vector composed of the mean DWT distance, mean centerline offset, mean maximum offset, and mean minimum offset between the voiceprint library and the device's normal operating status. ; If the target peak signal Each element in the feature vector is less than the upper limit of the feature vector. The corresponding element in the data is then used to identify the target peak signal. If the signal is matched with the voiceprint database of the normal operating status of the equipment, it is considered a normal peak signal and the current operating status of the equipment is normal; otherwise, it is considered a fault peak signal and the current operating status of the equipment is abnormal.

2. The method for monitoring the operating status of industrial equipment based on voiceprint recognition according to claim 1, characterized in that, When collecting acoustic fingerprint signals from devices using acoustic sensors, the sampling frequency, sampling duration, and sampling interval must be set in advance.

3. The method for monitoring the operating status of industrial equipment based on voiceprint recognition according to claim 1, characterized in that, The specific steps for performing wavelet transform on the acquired speaker signal containing background noise are as follows: The selected basis functions and decomposition level are used to perform wavelet decomposition on the acquired voiceprint signal. Wavelet coefficient processing is performed based on the 3sigma criterion in Chebyshev's inequality to remove the wavelet coefficients corresponding to noise. Then, wavelet reconstruction is performed to obtain the preliminary denoised voiceprint signal. The wavelet coefficient processing formula is as follows: in, Wavelet coefficients The wavelet coefficients obtained after wavelet coefficient processing, The sign of the wavelet coefficients is indicated by the sign of the coefficients. For noise variance estimation, that is, the wavelet coefficients are located at... The wavelet coefficient values ​​within the interval are set to 0, and those greater than 0 are set to 0. Wavelet coefficients uniformly subtract less than Wavelet coefficients uniformly added , This is the set of wavelet coefficient moduli obtained from wavelet decomposition. This represents the median of the wavelet coefficient modulus.

4. The method for monitoring the operating status of industrial equipment based on voiceprint recognition according to claim 3, characterized in that, The peak extraction specifically involves: Peak extraction is performed on the wavelet-transformed acoustic signature signal based on the selected time interval to obtain the maximum and minimum values ​​of the acoustic signature signal within each time interval, thus obtaining the discrete time sequence of the peak signal corresponding to the acoustic signature signal under the device operating state.

5. The method for monitoring the operating status of industrial equipment based on voiceprint recognition according to claim 1, characterized in that, Acquire multiple equipment fault operation acoustic signal samples, construct an equipment fault operation status acoustic library, and label different fault categories to obtain m fault categories; The multi-scale entropy and power spectral entropy of the peak signals from the acoustic signature samples during equipment malfunction are extracted to obtain feature vectors. Based on the fault category, sample data for m categories can be obtained. , Indicates the first Each fault category corresponds to a set of all feature vectors; The fault sample data is divided into m positive and negative sample combinations in the form of OVR, i.e. ,and Indicates random uniformity from the excluding the Samples drawn from sample sets of other categories outside of the sample set of each fault category; By training m sets of samples using support vector machines, m binary classifiers can be obtained. These m binary classifiers can form a device fault identification model; When the target acoustic signature signal is detected as an equipment malfunction acoustic signature signal, the multi-scale entropy and power spectral entropy of the corresponding peak signal are extracted and substituted into the obtained equipment fault identification model to obtain the probability corresponding to each fault category of the target acoustic signature signal. The fault category with the highest probability is the fault category of the current target voiceprint signal.

6. The method for monitoring the operating status of industrial equipment based on voiceprint recognition according to claim 1, characterized in that, As the equipment continues to operate, the continuously acquired voiceprint signals of the equipment's operating status are divided into normal voiceprint signals or fault voiceprint signals, and added to the corresponding voiceprint libraries respectively. The voiceprint libraries for normal equipment operation status and equipment fault operation status are updated regularly. For the update of the voiceprint library for equipment fault operation status, the equipment fault identification model is retrained periodically or quantitatively to ensure the accuracy and robustness of equipment fault identification.

Images