Sound-recognition-based method for monitoring operating condition abnormity of oil pump electric motor and outlet pipes thereof

Abstract

A sound-recognition-based method for monitoring an operating condition abnormity of an oil pump electric motor and outlet pipes thereof. Applying soundprint recognition technology to a speed-regulating hydraulic system of a hydropower station, and rational layout of on-site soundprint sensors realize real-time monitoring of operating conditions of an oil pump electric motor and outlet pipes thereof, thereby improving the intelligence level of device management, and achieving an industrial demonstration effect. By means of deep learning analysis of a soundprint fault sample library combined with multi-dimensional factors such as pressure, liquid level and temperature, a soundprint-recognition-based multi-dimensional-factor logical determination process for abnormal operating conditions of an oil pump electric motor and outlet pipes is designed, such that real-time monitoring and rapid locating of abnormal operating conditions such as base vibration, lubricating oil deficiency in an oil pump electric motor, pump body blade fracture, check valve malfunction, incomplete closure or internal leakage of a drain valve, frequent loading and unloading of an oil pump, and servomotor pipe pulsation are realized, so as to promptly use response measures such as automatic switching of main and standby oil pump electric motor units or automatic adjustment of unit load. Using soundprint sensors and soundprint recognition technology to monitor the state of on-site devices in real time eliminates uncertainties in manual inspection, and improves the fault diagnosis and analysis efficiency.

Description

Oil pump motor and outlet pipeline working condition abnormality monitoring method based on sound recognition TECHNICAL FIELD

[0001] The present application belongs to the technical field of abnormal working condition monitoring of hydropower stations, and particularly relates to an oil pump motor and an outlet pipeline working condition abnormality monitoring method based on sound recognition. BACKGROUND

[0002] The oil pump motor is an important driving device of the speed-regulating hydraulic system of a hydropower station, and also includes a base, a valve, a pipeline and other auxiliary equipment connected therewith. Failure of the oil pump motor and the outlet pipeline thereof can cause failure of the speed-regulating hydraulic system, affect normal standby of the unit, and even cause forced shutdown of the unit in serious cases. Therefore, real-time monitoring of the operating conditions of the oil pump motor and the outlet pipeline equipment is required to facilitate timely discovery and treatment.

[0003] The monitoring method for the oil pump motor and the outlet pipeline thereof at the present stage is as follows:

[0004] 1. Periodic on-site inspection of the oil pump motor by personnel to judge abnormal sound of the oil pump motor, motor reverse rotation to determine lack of lubricating oil, failure of the pipeline check valve and vibration of the oil pump motor base and other faults.

[0005] 2. Recording of the pressure oil tank pressure oil level, artificial analysis of the frequency of loading and unloading of the oil pump motor, comparison with the experience value, determination of whether the loading and unloading is frequent, and on-site inspection of whether the pipeline is twitching.

[0006] 3. On-site inspection or image monitoring to check whether the oil pump motor and the auxiliary equipment thereof leak oil.

[0007] The disadvantages of the prior art are as follows:

[0008] 1. Since the sensitivity of the human body to sound is different, the judgment of whether the sound of the oil pump motor is abnormal also varies from person to person, which is low in efficiency and lacks real-time performance, and the causes are many, such as lack of lubricating oil, pump body blade fracture, base vibration and other abnormal causes that cannot be accurately located.

[0009] 2. Whether the check valve function is failed needs to be indirectly judged by artificial inspection of whether the oil pump motor is reversed, and the rotating part of the oil pump motor has a protective cover, so it is difficult for artificial inspection to check the reverse condition, and if the rotating part is approached for careful checking, there is a safety hazard.

[0010] 3. Image monitoring can only check the leakage or spraying of oil of the peripheral part of the equipment, and cannot distinguish the internal leakage caused by the valve not being tightly closed and the internal sealing failure of the equipment, and artificial judgment also needs to consume a long time to analyze a large amount of data, which is low in efficiency and lacks real-time performance.

[0011] 4. Manually analyzing the frequency of oil pump loading and unloading has problems such as large workload and time lag. Furthermore, the determination of experience values ​​is not uniform, and frequent loading and unloading can cause pipeline pulling, which requires on-site inspection by personnel.

[0012] In summary, existing technologies mainly rely on manual periodic on-site inspections, which are inefficient, lack reliability and real-time performance, and pose certain safety hazards. Image monitoring cannot detect internal leaks caused by valves not being closed tightly or internal seal failures in equipment, making manual judgment difficult. The diagnosis of initial faults often lags behind the occurrence of secondary faults.

[0013] Summary of the Invention

[0014] The technical problem to be solved by the present invention is to provide a method for monitoring the abnormal operating conditions of oil pump motor and its outlet pipeline based on sound recognition. By using a voiceprint sensor and voiceprint recognition technology, the status of field equipment can be monitored in real time, eliminating the uncertainty of manual inspection and improving the efficiency of fault diagnosis and analysis.

[0015] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0016] A method for monitoring abnormal operating conditions of an oil pump motor and its outlet pipeline based on sound recognition, comprising the following steps:

[0017] Step 1: Collect voiceprint information.

[0018] Step 1.1: Install acoustic fingerprint sensors at the pump body, motor, base, and valve outlet pipeline of the oil pump motor; install temperature measuring resistors at the motor and oil pump bearings; install pressure sensors at the oil pump outlet, pressure tank, and pressure regulating valve inlet; and install level sensors at the oil collection tank and pressure tank. Use acoustic fingerprint sensors to collect acoustic fingerprint information, use temperature measuring resistors to collect temperature information, use pressure sensors to collect pressure information, and use level sensors to collect level information.

[0019] Step 1.2: After the voiceprint information is initially processed by the front-end box, it is transmitted to the edge computing gateway for preliminary screening, and then transmitted to the server for data analysis via the switch;

[0020] Temperature, pressure, and liquid level information are sent to the local LCU unit via the remote I / O module, and then transmitted to the server via the switch for comprehensive data analysis.

[0021] Step 2: Extract voiceprint features from the voiceprint information and process the voiceprint features:

[0022] Mel Frequency Cepstral Coefficients (MFCC) are used to extract the voiceprint information, and the Universal Background Model (UBM) and i-vector methods are used to convert the voiceprint features into model vectors for calculation and analysis. The Probability Linear Discriminant Analysis (PLDA) method is used to compare and score the voiceprint features.

[0023] Step 3: Comprehensive logical judgment is made on the four analysis results.

[0024] Step 4: The final fault result is output.

[0025] Preferably, the MFCC extraction method process in step 2 is as follows:

[0026] 1) The collected voiceprint information is pre-emphasized through a high-pass filter, then segmented into overlapping frames, and a Hamming window function is applied to each frame to offset the assumption of infinite data made by the Fast Fourier Transform and reduce spectral leakage.

[0027] 2) Fast Fourier Transform is performed on each frame and the power spectrum is calculated. The frequency spectrum is extracted through a Mel filter bank, the log energy is calculated, and the discrete cosine transform is applied for decorrelation.

[0028] 3) The MFCC feature vector composed of N-dimensional MFCC parameters + frame energy is finally obtained by extracting dynamic difference parameters. The dynamic difference parameters include first-order difference and second-order difference.

[0029] Preferably, the UBM calculation method process in step 2 is as follows:

[0030] First, collect a large amount of non-fault state voiceprint information, train a UBM, then use a small amount of fault state voiceprint information to adjust the parameters of the UBM through an adaptive algorithm to obtain the target model parameters.

[0031] Preferably, in step 2, the i-vector analysis method process is as follows:

[0032] i-vector defines a low-dimensional vector R x 1, w ~ N(0, I) to represent a certain audio segment.

[0033] M = m + T x w;

[0034] M is the ideal feature hyper-vector of a certain device corresponding to the i-vector multiplied by the T matrix plus the UBM mean hyper-vector;

[0035] m is the UBM mean hyper-vector. Assuming that UBM contains C Gaussian mixture components g, then m is the mean vector of all mixture components m ccombinations of c = 1,...,C, m has dimension C*F, T has dimension C*F x R, F represents the dimension of MFCC features;

[0036] For each mixing component c, parameter mixing weight w c , and covariance matrix∑c; further split T matrix into C combinations of Vc, Vc has dimension F*R,

[0037] μ c = m c + V c w;

[0038] Define a piece of audio feature data X, which has feature dimension F and time sequence T, i.e. X = X1,...,X T ; The subset of X belonging to the cth Gaussian component is X c , and the certain frame X t c , then:

[0039] v ~ N(0,∑ c );

[0040] Thus, the following formula can be obtained:

[0041] The calculation formula of i-vector is:

[0042] Wherein, N(u) is a diagonal matrix with dimension CF*CF, and the diagonal block is NcI, (c = 1,...,C), is a super vector with dimension CF x 1, which is composed of all first-order BW statistics .

[0043] Preferably, in step 2, the PLDA analysis method process is as follows:

[0044] Suppose the training data audio is composed of audio of I devices, wherein each device has J pieces of different audio; The i-vector of the jth piece of audio of the ith device is denoted as D ij , and then the PLDA is defined as: D ij = μ + Fh i + Gω ij + ε ij ;

[0045] Wherein, the device information part is μ + F ij , which is only related to device i and describes the difference between devices;

[0046] The noise part is Gω ij + εij , describes the difference between the inside of the device;

[0047] μ represents the mean of all training data;

[0048] F is regarded as the identity space, which contains information that can be used to represent various devices;

[0049] h i is regarded as the identity of a specific device or the position of the device in the identity space;

[0050] G is regarded as the error space, which contains information that can be used to represent different audio variations of the same device;

[0051] ω ij represents the position in the G space;

[0052] ε ij is the last residual noise term, which is used to represent the unexplained things;

[0053] After the PLDA model is trained, the device discrimination is performed by using the likelihood ratio strategy as follows:

[0054] In the above formula, x i and x p are i-vector vectors of two audios, the assumption that the two audios come from the same space is M1, and the assumption that the two audios come from different spaces is M0, wherein P(x i ,x p |M1) is the likelihood function of the two audios coming from the same space; P(x i |M0), P(x p |M0) are the likelihood functions of x i and x p coming from different spaces; by calculating the log-likelihood ratio, the similarity of the two audios is calculated and scored;

[0055] The higher the ratio, the higher the score, and the greater the possibility that the two audios belong to the same device;

[0056] The lower the ratio, the lower the score, and the smaller the possibility that the two audios belong to the same device.

[0057] Preferably, the sub-steps in step 3 are as follows:

[0058] Through PLDA model analysis, a voiceprint feature fault sample library corresponding to the equipment is established. If the voiceprint feature data collected on site is successfully matched with the fault sample library, the fault type is directly output. The voiceprint data that fails to match is scored by the PLDA model. The scorers are comprehensively determined based on multi-dimensional factors including pressure, liquid level and temperature, and fault determination is performed. Artificial confirmation is used to mark the fault samples, and the fault sample library is supplemented. The steps are as follows:

[0059] ① Motor lubricating oil loss:

[0060] Motor voiceprint data S 1-1 or S 1-2 The scorers are abnormal, the rising trend of the motor bearing temperature T1 is analyzed, the rising rate k of T1 is calculated with Δt=1min as a period and the current time t as a starting point, and the formula is as follows: t

[0061] If the relationship among the rising rates k of T1 in the last three periods k t-2Δt , k t-Δt and k t is k t-2Δt <k t-Δt <k t , it is determined that the motor bearing temperature rises abnormally, and the abnormal scorers of the voiceprint data are comprehensively determined, so that the motor lubricating oil loss is determined.

[0062] ② Oil pump lubricating oil loss:

[0063] Oil pump voiceprint data S 2-1 or S 2-2 The scorers are abnormal, the rising trend of the oil pump bearing temperature T2 is analyzed, and if the rising rates of T2 in the last three periods satisfy k t-2Δt <k t-Δt <k t , it is determined that the oil pump bearing temperature rises abnormally, and the abnormal scorers of the voiceprint data are comprehensively determined, so that the oil pump lubricating oil loss is determined.

[0064] ③ Abnormal base vibration:

[0065] The scorers of the base upper surface voiceprint data S3 are abnormal, and the motor and oil pump lubricating oil loss features are not matched, so that the base vibration is determined to be abnormal.

[0066] ④ Pump body blade fracture:

[0067] Oil pump voiceprint data S 2-1 or S 2-2 The scorers are abnormal, the rising trend of the oil pump bearing temperature T2 is normal, the oil pump outlet pressure P1 is lower than the normal loading and unloading pressure under the conditions of loading and unloading, i.e. P 1加 <P 加 and P​1卸 <P 卸 If the pump efficiency is abnormal, and the motor sound signature data is normal, it can be determined that the pump blades are broken.

[0068] ⑤ Check valve malfunction:

[0069] When the oil pump motor stops, if the S5 score of the oil pump outlet pipeline acoustic fingerprint data is abnormal and P1>0, and the acoustic fingerprint data score of the oil pump motor is also abnormal, it can be determined that the oil pump outlet check valve function has failed, causing the oil pump motor to reverse.

[0070] ⑥ Frequent loading and unloading of the oil pump:

[0071] The acoustic signature data S6 of the loading / unloading valve assembly is abnormal. Analyze the time interval t between each loading / unloading operation of the oil pump. jx When the pressure in the oil tank P3 = 6.1 MPa, the oil pump starts to load, and this is recorded as time t. j The next time P3 = 6.1 MPa is recorded as time t. j1 , then t jx =t j1 -t j If t jx <t z , where t z If the loading / unloading interval is the set value, it is determined that the oil pump is loading / unloading frequently;

[0072] ⑦ The drain valve is not closed tightly or there is internal leakage:

[0073] The acoustic signature data S4 of the oil tank drain valve is abnormal, and the time t for one oil pump loading is also abnormal. jz Too long, i.e., t jz >t z1 , t z1 The loading time is set to a value that matches the frequent loading and unloading characteristics. The total oil volume V of the hydraulic system is calculated by considering the liquid levels H1 and H2 of the pressure tank and the oil collection tank. If V = V0, where V0 is the initial total oil volume of the hydraulic system, it is determined that the drain valve is not closed tightly or there is internal leakage.

[0074] ⑧ Relay line pulling:

[0075] If the acoustic data S7 of the pressure regulating valve body and the acoustic data S8 or S9 of the pressure regulating valve outlet pipeline are abnormal, the frequent loading and unloading characteristics of the oil pump have been matched, and the total oil volume V = V0, then it can be determined that the relay pipeline is being pumped.

[0076] Preferably, after the above fault matching or confirmation, the alarm signal and fault message are output through the host computer human-machine interaction. After any of the faults ①-⑤ are confirmed to occur, the main standby automatic switching function of the oil pump motor unit should be activated, the standby oil pump motor unit should be activated, the faulty group should be deactivated and marked as unavailable for maintenance.

[0077] After confirming the occurrence of fault 6, only an alarm signal is outputted;

[0078] Faults 6 and 7 occur simultaneously, and are considered as the unsealed or internal leakage of the oil discharge valve, an alarm signal is outputted, and the start-up condition is locked;

[0079] Faults 6 and 7 occur simultaneously, and are confirmed as the pumping of the relay pipe, an alarm signal is outputted, the AGC regulating unit is started, the unit output is increased or decreased by 10 MW according to the actual output of the unit, and the unit working condition is adjusted, which needs to be monitored manually, and manual intervention and on-site confirmation are needed when necessary.

[0080] Preferably, in step 1, the voiceprint sensor adopts a patch type bone conduction microphone, which is arranged on the two sides of the motor shell, the two sides of the pump body shell, the upper surface of the base, the oil pump outlet pipeline, the loading and unloading valve group, the pressure tank oil discharge valve, the pressure regulating valve body and the pressure regulating valve outlet pipeline.

[0081] The temperature measuring resistor is arranged on the motor bearing and the oil pump bearing.

[0082] The pressure sensor is arranged on the oil pump outlet, the pressure tank and the pressure regulating valve inlet.

[0083] The liquid level sensor is arranged on the oil collecting tank and the pressure tank.

[0084] The present application can achieve the following beneficial effects:

[0085] 1. The voiceprint sensor and voiceprint recognition technology are used to monitor the on-site equipment state in real time, eliminate the uncertain factors of manual inspection, and improve the fault diagnosis and analysis efficiency.

[0086] 2. The voiceprint fault sample library of the oil pump motor and the outlet pipeline is established, which has industry universality and great reference value.

[0087] 3. The voiceprint recognition technology is creatively applied to the speed regulating hydraulic system of the hydropower station, the on-site voiceprint sensor is reasonably arranged, the running condition of the oil pump motor and the outlet pipeline is monitored in real time, the intelligent level of equipment management is improved, and the industry demonstration effect is achieved.

[0088] 4. Through the deep learning analysis of the voiceprint fault sample library combined with pressure, liquid level, temperature and other multi-dimensional factors, a multi-dimensional factor logic judgment process of the abnormal working condition of the oil pump motor and the outlet pipeline based on voiceprint recognition is designed, the real-time monitoring and rapid positioning of the abnormal working conditions such as base vibration, oil pump motor lubricating oil loss, pump body blade fracture, check valve function failure, unsealed or internal leakage of the oil discharge valve, frequent loading and unloading of the oil pump, pumping of the relay pipe and the like are realized, and the automatic switching of the standby oil pump motor or the automatic adjustment of the unit load adjustment and other measures are taken in time.

[0089] 5. By the MFCC algorithm, UBM method and i-vector method, the voiceprint information of the oil pump motor and the outlet pipeline equipment is converted into a model vector for calculation and analysis, so that the recognition and analysis of the voiceprint information are more accurate and reasonable, and the determination of equipment failure is more reliable. BRIEF DESCRIPTION OF DRAWINGS

[0090] The application will be further described below in combination with the drawings and examples:

[0091] Fig. 1 is a flowchart of the present application;

[0092] Fig. 2 is a layout of the voiceprint sensor, temperature measuring resistor, pressure sensor and liquid level sensor of the present application;

[0093] Fig. 3 is a data transmission diagram of the present application;

[0094] Fig. 4 is a voiceprint feature extraction processing method diagram of the present application;

[0095] Fig. 5 is a 3D diagram of 13-order MFCC feature extraction of 500ms voiceprint data of the present application;

[0096] Fig. 6 is a motor fault voiceprint comprehensive evaluation flowchart of the present application;

[0097] Fig. 7 is an oil pump fault voiceprint comprehensive evaluation flowchart of the present application;

[0098] Fig. 8 is a check valve fault voiceprint comprehensive evaluation flowchart of the present application;

[0099] Fig. 9 is an oil pump frequent loading and unloading voiceprint comprehensive evaluation flowchart of the present application;

[0100] Fig. 10 is a discharge valve internal leakage voiceprint comprehensive evaluation flowchart of the present application;

[0101] Fig. 11 is a servomotor pipeline twitching voiceprint comprehensive evaluation flowchart of the present application;

[0102] Fig. 12 is a fault result output and pre-control measure logic diagram of the present application.

[0103] In Fig. 2: S 1-1 / S 1-2 , S 2-1 / S 2-2 , S3-S9 are voiceprint sensors (S 1-1 , S 1-2 are symmetrically arranged, S 2-1 , S 2-2 are symmetrically arranged); P1-P3 are pressure sensors; H1 / H2 are liquid level sensors; T1 is a motor bearing temperature measuring resistor; T2 is an oil pump bearing temperature measuring resistor. DETAILED DESCRIPTION

[0104] The present application utilizes voiceprint sensors and voiceprint recognition technology to monitor the real-time state of the oil pump motor and outlet pipeline of the speed regulating hydraulic system of a hydropower station, replaces manual regular on-site inspection, eliminates the lagging nature of manual inspection and unreliable factors in judging equipment faults, processes the collected voiceprint information using MFCC, PLDA and other voiceprint recognition technologies, establishes a corresponding equipment voiceprint fault sample library through repeated training, and can be universally applied to the judgment of voiceprint fault characteristics of related equipment in the industry; the collected voiceprint feature data is matched with the fault sample library to accurately and quickly determine; the unmatched voiceprint data is scored by the PLDA model, and the abnormally scored data is comprehensively judged based on multiple factors such as pressure, liquid level and temperature, and the fault sample is marked by manual confirmation to supplement the fault sample library; a front-end storage device is arranged in the voiceprint collection section, and the original sound information can be accessed at the front end or offline, greatly improving the storage capacity of the terminal device; the real-time fault information matched or marked is sent by the terminal man-machine interaction to issue an alarm information, and is viewed in the form of fault message, achieving real-time monitoring of on-site equipment, timely and accurate alarm prompt for abnormal conditions, and automatically taking countermeasures to eliminate equipment abnormalities in the early stage and avoid serious accidents.

[0105] The present application creatively applies voiceprint recognition technology to the speed regulating hydraulic system of a hydropower station, realizes real-time monitoring of the operating conditions of the oil pump motor and its outlet pipeline through reasonable layout of on-site voiceprint sensors, improves the intelligent level of equipment management, and plays a role in industry demonstration.

[0106] Through deep learning analysis of the voiceprint fault sample library combined with multiple factors such as pressure, liquid level and temperature, a multi-dimensional factor logic judgment process for abnormal operating conditions of the oil pump motor and its outlet pipeline based on voiceprint recognition is designed, realizing real-time monitoring and rapid positioning of abnormal operating conditions such as base vibration, oil pump motor lubricating oil loss, pump body blade fracture, check valve function failure, oil pump unsealed or internal leakage, frequent oil pump loading and unloading, and servomotor pipeline twitching, and timely taking countermeasures such as automatic switching of standby oil pump motor or automatic adjustment of unit load. The specific scheme is as follows:

[0107] The preferred scheme is shown in FIGS. 1 to 12, an oil pump motor and its outlet pipeline operating condition abnormality monitoring method based on sound recognition, mainly including information collection and transmission, voiceprint feature processing, comprehensive logic judgment, fault result output and the like, as shown in FIG. 1. The on-site equipment information is collected and transmitted to the upper computer monitoring system in real time, and the voiceprint feature processing, comprehensive logic judgment and fault result output are automatically completed on the upper computer. The specific steps are as follows:

[0108] Step 1, information collection and transmission:

[0109] Information collection by sensor device: acoustic fingerprint sensor, temperature resistance, pressure sensor, liquid level sensor, etc. are arranged at the pump body, motor, base and valve outlet pipeline of the oil pump motor, among which the acoustic fingerprint sensor adopts a patch type bone conduction sound pickup, which is arranged at: one on each side of the motor shell (symmetrically arranged), one on each side of the pump body shell (symmetrically arranged), one on the upper surface of the base, one near the check valve at the oil pump outlet pipeline, one at the outlet pipeline of the loading and unloading valve group, one at the outlet pipeline of the oil tank drain valve, one at the body of the pressure regulating valve, one at the outlet pipeline of the pressure regulating valve; temperature resistance arrangement: one at the motor bearing, one at the oil pump bearing; pressure sensor arrangement: one at the oil pump outlet, one at the oil tank, one at the inlet of the pressure regulating valve; liquid level sensor arrangement: one at the oil collecting tank, one at the oil tank, specific position as shown in Figure 2.

[0110] Data transmission by computer: the collected acoustic fingerprint information is first processed by the front-end box, then screened by the edge computing gateway, and then transmitted to the server by the switch for data analysis. Man-machine interaction can read and write data through the server. The front-end box has the functions of storage and acoustic fingerprint initial processing, can store the original acoustic information in the front-end box, is convenient for reading historical data, and saves storage space for terminal devices; temperature, pressure, liquid level and other information are sent to the present LCU unit through the remote I / O module, transmitted to the server through the switch for data comprehensive analysis, as shown in Figure 3.

[0111] Step 2, acoustic fingerprint feature processing by computer:

[0112] The acoustic fingerprint feature processing algorithm includes: acoustic fingerprint MFCC extraction, UBM calculation, i-vector calculation and PLDA analysis, as shown in Figure 4.

[0113] 1) MFCC feature extraction (this process needs to use computer program):

[0114] MFCC feature extraction includes preprocessing (pre-emphasis, framing, windowing), fast Fourier transform FFT, Mel filter bank, logarithmic operation, discrete cosine transform, dynamic feature extraction. The collected voiceprint information is pre-emphasized through a high-pass filter, then segmented into (overlapping) frames, and a Hamming window function is applied to each frame to offset the assumption of infinite data made by FFT and reduce spectral leakage; then, fast Fourier transform (FFT) is performed on each frame and the power spectrum is calculated, the spectrum is extracted through the Mel filter bank, the logarithmic energy of the spectrum is calculated, and then the discrete cosine transform (DCT) is applied to correlation; finally, by extracting dynamic difference parameters (including first-order difference and second-order difference), the MFCC feature vector composed of N-dimensional MFCC parameters (N / 3 MFCC coefficients + N / 3 first-order difference parameters + N / 3 second-order difference parameters) + frame energy is obtained.

[0115] Taking 500 milliseconds of voiceprint data as an example, using 0.97 as the pre-emphasis coefficient, the voiceprint data is pre-emphasized, then segmented into multiple frames, each frame is 32 milliseconds, and the frame shift is 16 milliseconds (the frame data length is 32 ms, the sample number is the sampling frequency multiplied by the time: 16000*0.032=512; the frame shift length is 16 ms, the sample number is the sampling frequency multiplied by the time: 16000*0.016=256), the number of Mel filters is set to 24, the MFCC features of 500 milliseconds of voiceprint data are extracted (frame length 512; frame shift 256; MFCC coefficient number 24), and a three-dimensional graph of 13-order MFCC is drawn, as shown in FIG. 5.

[0116] 2) Calculate UBM (this process needs to use a computer program):

[0117] Universal Background Model (UBM, Universal Background Model), first collect a large amount of voiceprint information in a non-fault state, train a UBM, and the training of UBM is actually the training of Gaussian Mixture Model (GMM). The specific method is: the voiceprint feature MFCC parameters in the non-fault state are iteratively optimized through the expectation maximization algorithm EM to obtain the UBM model parameters, and then a small amount of voiceprint data in a fault state is used to adjust the parameters of the UBM through the adaptive algorithm MAP to obtain the target model parameters. Generally, only the mean m is adjusted by the adaptive algorithm. This method can reduce the number of fault samples and the amount of parameters required for training, and facilitate fast training convergence and decoding calculation in mobile terminals.

[0118] 3) Calculate i-vector:

[0119] i-vector defines a low-dimensional vector R x 1, w ~ N (0, I) to represent a certain audio segment.

[0120] M = m + T x w;

[0121] M is the ideal feature super-vector of a certain device being modeled, multiplied by the T matrix and added to the UBM mean super-vector; m is the UBM mean super-vector, assuming the UBM contains C Gaussian mixture components g, then m is the mean vector of all the mixture components c , c = 1,..., C, with dimension C*F, and T has dimension C*F x R, F represents the dimension of MFCC features.

[0122] For each mixture component c, there are also parameters mixture weight w c , and covariance matrix∑c. Further split the T matrix into C combinations of Vc, Vc has dimension F*R, μ c = m c + V c w;

[0123] Define a piece of audio feature data X, with feature dimension F, and time sequence T, i.e. X = X1,..., X T . The subset of X belonging to the c-th Gaussian component is X c , and a certain Frame (frame) in the subset is X

[0124] v ~ N(0,∑ c );

[0125] Thus, the following formula can be obtained

[0126] The calculation formula of i-vector

[0127] Where N(u) is a diagonal matrix with dimension CF*CF, and the diagonal block is NcI, (c = 1,..., C), is a super-vector with dimension CF x 1, composed of all first-order BW statistics .

[0128] 4. PLDA analysis (this process needs to use computer programs)

[0129] Probabilistic Linear Discriminant Analysis (PLDA) is also a channel compensation algorithm, also known as the probabilistic form of LDA algorithm. PLDA is also usually based on i-vector features, providing channel compensation for it. We assume that the training data audio is composed of audio of I devices, each of which has J different audio of its own. The i-vector of the j-th audio of the i-th device is denoted as D​ij Then PLDA defines D ij = μ + Fh i + Gω ij + ε ij ;

[0130] Where the device information part is μ + F ij , which is only related to device i and describes the difference between devices;

[0131] The noise part is Gω ij + ε ij , which describes the difference between the internal devices.

[0132] μ represents the mean of all training data;

[0133] F can be regarded as an identity space, which contains information that can be used to represent various devices;

[0134] h i can be regarded as the identity of a specific device (or the position of the device in the identity space);

[0135] G can be regarded as an error space, which contains information that can be used to represent different audio changes of the same device;

[0136] ω ij represents the position in the G space;

[0137] ε ij is the last residual noise term, which is used to represent things that have not been explained.

[0138] After the PLDA model is trained, the device discrimination is performed using the likelihood ratio strategy as follows:

[0139] In the above formula, x i and x p are i-vector vectors of two audios, the assumption that the two audios come from the same space is M1, and the assumption that the two audios come from different spaces is M0, where P(x i , x p |M1) is the likelihood function of the two audios coming from the same space;

[0140] P(x i |M0), P(x p |M0) are the likelihood functions of x i and x p coming from different spaces. By calculating the log likelihood ratio, the similarity of the two audios is calculated and scored.

[0141] The higher the ratio, the higher the score, and the greater the possibility that the two audios belong to the same device;

[0142] The lower the ratio, the lower the score, and the less likely the two audio belong to the same device.

[0143] Step 3, comprehensive logical judgment (this process needs to use computer program):

[0144] Through PLDA model analysis, establish the corresponding device voiceprint feature fault sample library, if the voiceprint feature data collected on site matches the fault sample library successfully, directly output the fault type; the voiceprint data that fails to match is scored by PLDA model, the abnormally scored one is comprehensively judged for fault according to multi-dimensional factors such as pressure, liquid level and temperature, and the fault sample is marked manually for confirmation, and the fault sample library is supplemented, as follows:

[0145] ① Motor lubricating oil loss:

[0146] Motor voiceprint data S 1-1 or S 1-2 Abnormal score, analyze the rising trend of motor bearing temperature T1, take Δt=1min as a period, and take the current time t as the starting point to calculate the rising rate k t of T1, as follows:

[0147] If the rising rates k t-2Δt , k t-Δt , and k t of T1 in the last 3 periods satisfy the relationship: k t-2Δt <k t-Δt <k t , it can be judged that the motor bearing temperature rises abnormally, and the abnormal score of the voiceprint data can be determined as motor lubricating oil loss, as shown in FIG. 6.

[0148] ② Oil pump lubricating oil loss:

[0149] Oil pump voiceprint data S 2-1 or S 2-2 Abnormal score, analyze the rising trend of oil pump bearing temperature T2 (calculation method is the same as above), if the rising rate of T2 in the last 3 periods satisfies k t-2Δt <k t-Δt <k t , it can be judged that the oil pump bearing temperature rises abnormally, and the abnormal score of the voiceprint data can be determined as oil pump lubricating oil loss, as shown in FIG. 7.

[0150] ③ Abnormal vibration of base:

[0151] Abnormal score of base upper surface voiceprint data S3, and no matching of motor and oil pump lubricating oil loss features, then determine that the base vibrates abnormally.

[0152] ④ Pump body blade fracture:

[0153] Oil pump voiceprint data S 2-1 Or S 2-2 Score anomaly, rising trend of oil pump bearing temperature T2 is normal, oil pump outlet pressure P1 is lower than normal loading and unloading pressure under loading and unloading conditions, i.e. P 1加 <P 加 And P 1卸 <P 卸 Oil pump efficiency is abnormal, combined with normal motor voiceprint data, it can be determined that the oil pump blade is broken, as shown in Fig. 7.

[0154] ⑤Check valve function failure:

[0155] When the oil pump motor is stopped, the oil pump outlet pipeline (near the check valve) voiceprint data S5 is abnormal, P1>0, and the oil pump motor voiceprint data score is abnormal, which can be determined as oil pump outlet check valve function failure, causing the oil pump motor to reverse, as shown in Fig. 8.

[0156] ⑥Oil pump loading and unloading frequently:

[0157] Loading and unloading valve group (outlet pipeline) voiceprint data S6 is abnormal, and the time interval t jx of oil pump loading and unloading once is analyzed j When the oil tank pressure P3=6.1MPa, the oil pump starts to load, and the time t j1 is recorded as t jx The next time P3=6.1MPa is recorded as t j1 , t j , if t jx <t z (t z is the loading and unloading interval setting value), it is determined that the oil pump loading and unloading is frequent, as shown in Fig. 9.

[0158] ⑦Oil discharge valve is not tightly closed or internal leakage:

[0159] The oil tank discharge valve (outlet pipeline) voiceprint data S4 is abnormal, and the time t jz of oil pump loading once (the time used for P3 to rise from 6.1MPa to 6.3MPa) is too long, i.e. t jz >t z1 (t z1 is the loading time setting value), and the loading and unloading frequency feature has been matched, the total oil volume V of the hydraulic system is calculated by combining the liquid levels H1 and H2 of the oil tank and the oil collecting tank, if V=V0 (V0 is the initial total oil volume of the hydraulic system), it is determined that the oil discharge valve is not tightly closed or internal leakage, as shown in Fig. 10.

[0160] ⑧Servomotor pipeline twitch:

[0161] The pressure regulating valve body voiceprint data S7 and the pressure regulating valve outlet pipeline voiceprint data S8 or S9 score abnormally, the oil pump loading and unloading frequency feature has been matched, and the total oil volume V=V0, so that the relay pipeline twitching can be determined, as shown in FIG. 11.

[0162] Step 4, fault result output (this process needs to use a computer program):

[0163] After the above fault matching or confirmation, an alarm signal and a fault message are output through human-computer interaction of the upper computer, and a corresponding pre-control measure is started, as shown in FIG. 12.

[0164] After any one of faults ①-⑤ is confirmed to occur, the main and standby automatic switching function of the oil pump motor set should be started, the standby oil pump motor set is started, the fault group is disabled and marked as unavailable at the same time, and maintenance is waited;

[0165] After fault ⑥ is confirmed to occur, only an alarm signal is output;

[0166] Faults ⑥ and ⑦ occur at the same time, considering that the oil discharge valve is not tightly closed or has internal leakage, an alarm signal is output, and the start-up condition is locked;

[0167] Faults ⑥ and ⑧ occur at the same time, confirming that the relay pipeline is twitching, an alarm signal is output, and the AGC regulating unit output is started, according to the actual output of the unit, the output is increased or decreased by 10 MW as a step, and the unit working condition is adjusted, which needs to be monitored manually, and manual intervention and on-site confirmation are needed.

[0168] The above embodiments are only preferred technical solutions of the present application, and should not be regarded as a limitation of the present application, and the protection scope of the present application should be based on the technical solutions claimed in the claims, including the equivalent replacement solutions of the technical features claimed in the claims. That is, within this range, equivalent replacement improvements are also within the protection scope of the present application.

Claims

1. A sound recognition-based oil pump motor and its outlet pipeline working condition abnormality monitoring method, characterized in that The method comprises the following steps: Step 1, collecting voiceprint information: Step 1.1, arranging voiceprint sensors at the pump body, motor, base and valve outlet pipeline of the oil pump motor, arranging temperature resistance at the motor and oil pump bearing, arranging pressure sensors at the oil pump outlet, oil tank and pressure regulating valve inlet, and arranging liquid level sensors at the oil collecting tank and oil tank; collecting voiceprint information by using voiceprint sensors, collecting temperature information by using temperature resistance, collecting pressure information by using pressure sensors, and collecting liquid level information by using liquid level sensors; Step 1.2, transmitting the voiceprint information to the edge computing gateway after preliminary processing by the front-end box for preliminary screening, and then transmitting to the server through the switch for data analysis; Transmit the temperature information, pressure information and liquid level information to the local control unit LCU through the remote I / O module, and transmit to the server through the switch for data comprehensive analysis; Step 2, extracting voiceprint features from the voiceprint information and processing the voiceprint features: The Mel frequency cepstral coefficient MFCC is used to extract the voiceprint information, the universal background model UBM method and the identity vector feature i-vector method are used to convert the voiceprint features into model vectors for calculation and analysis, and the probability linear discriminant analysis PLDA method is used to compare and score the voiceprint features; Step 3, for the voiceprint data with abnormal score, combining the multi-dimensional factors including pressure, liquid level and temperature, comprehensive logical judgment is carried out; Step 4, output the final fault result.

2. The sound recognition-based oil pump motor and its outlet pipeline working condition abnormality monitoring method according to claim 1, characterized in that: The MFCC extraction method in step 2 is as follows: 1) The collected voiceprint information is pre-emphasized by a high-pass filter, then segmented into overlapping frames, and a Hamming window function is applied to each frame to offset the assumption of infinite data made by the fast Fourier transform and reduce spectral leakage; 2) Perform fast Fourier transform on each frame and calculate the power spectrum, extract the frequency spectrum through the Mel filter bank, take the logarithmic energy of the frequency spectrum, and then apply the discrete cosine transform to decorrelation; 3) By extracting dynamic difference parameters, the MFCC feature vector composed of N-dimensional MFCC parameters and frame energy is finally obtained; the dynamic difference parameters include first-order difference and second-order difference.

3. The sound recognition based oil pump motor and its outlet pipeline working condition abnormality monitoring method according to claim 1, characterized in that: The UBM calculation method in step 2 is as follows: First, collect a large amount of voiceprint information in a non-fault state, train a universal background model UBM, and then use a small amount of fault state voiceprint information to adjust the parameters of the UBM through an adaptive algorithm to obtain target model parameters.

4. The sound recognition based oil pump motor and its outlet pipeline working condition abnormality monitoring method according to claim 1, characterized in that: In step 2, the i-vector analysis method process is as follows: i-vector defines a low-dimensional vector R x 1, w ~ N(0, I) to represent a certain audio segment; M = m + T x w; M is the ideal feature super vector corresponding to a certain device modeled by T matrix multiplied by i-vector, plus UBM mean super vector. m as the UBM mean super-vector, assuming the UBM contains C Gaussian mixture components g, then m is the combination of all the mixture component mean vectors m c , c = 1,..., C, the dimension of m is C*F, and the dimension of T is C*F x R, F represents the dimension of the MFCC feature. for each mixing component c, a parameter mixing weight w c and a covariance matrix Sc; further splitting the T matrix into C combinations of Vc, with Vc having a dimension of F * R, μ c = m c + V c w; Define a piece of audio feature data X, whose feature dimension is F, and time sequence is T, i.e. X = X1,...,X T The subset of X belonging to the cth Gaussian component is X c The frame X t c Then: From this the following equation is derived: The formula for calculating i-vector is: where N(u) is a diagonal matrix of dimension CF*CF with diagonal blocks NcI (c = 1,..., C), is the super vector of CF x 1 by all the first order BW statistics Step 2, the PLDA analysis method process is as follows:

5. The sound recognition based oil pump motor and its outlet pipeline working condition abnormality monitoring method according to claim 1, characterized in that: μ represents the mean of all training data; Suppose the training data audio is composed of audio from I devices, where each device has J segments of its own audio; the i-vector of the jth segment of the ith device is denoted as D ij , and then the PLDA is defined as: D ij = μ + Fh i + Gω ij + ε ij ; where the device information part is μ + F ij and is only related to device i, describing the difference between devices. Noise part is Gω ij +ε ij , describes the difference between the device internals; F is regarded as an identity space, which contains information that can be used to represent various devices; G is regarded as an error space, which contains information that can be used to represent different audio changes of the same device. h i to be considered as the identity of a particular device or the location of a device in identity space; ​ ω ij represents a position in G-space; ε ij is the last residual noise term, which represents the unexplained things; After the PLDA model is trained, the device discrimination is performed using a likelihood ratio strategy as follows: As in the above formula, x i and x p are i-vector vectors of two audio, the assumption that the two audio come from the same space is M1, and the assumption that the two audio come from different spaces is M2. the assumption of the space of M0, wherein P(x i ,x p | M1) is the likelihood function of two audio from the same space; P(x i | M0), P(x p | M0) are the likelihood functions of x i and x p from different spaces respectively; by calculating the log-likelihood ratio, the similarity of two audio is calculated and scored; The higher the ratio, the higher the score, and the greater the likelihood that the two audio belong to the same device; The lower the ratio, the lower the score, and the less likely that the two audio belong to the same device.

6. The sound recognition based oil pump motor and its outlet pipeline working condition abnormality monitoring method according to claim 1, characterized in that: The determination steps of motor and oil pump lubricating oil absence or base vibration in step 3 are as follows: Through PLDA model analysis, a voiceprint feature fault sample library corresponding to the device is established. If the voiceprint feature data collected on site is successfully matched with the fault sample library, the fault type is directly output. The voiceprint data that fails to match is scored by the PLDA model. The abnormally scored data is comprehensively determined for fault according to related temperature data. The fault sample is marked by manual confirmation, and the fault sample library is supplemented. The steps are as follows: ① Motor lubricating oil absence: Motor voiceprint data S 1-1 Or S 1-2 If the score is abnormal, the rising trend of the motor bearing temperature T1 is analyzed, with Δt=1 min as a period, the current time t as a starting point, and the rising rate k of T1 is calculated t , as follows: If the rising rate k of T1 in the last 3 cycles t-2Δt , k t-Δt , k t The relationship among the three is: k t-2Δt <k t-Δt <k t , the motor bearing temperature rises abnormally, and the comprehensive voiceprint data score is abnormal, which can be determined as motor lubricating oil loss; ② Oil pump lubricating oil absence: Oil pump voiceprint data S 2-1 Or S 2-2 If the score is abnormal, the rising trend of the oil pump bearing temperature T2 is analyzed, and if the rising rate of T2 in the last 3 cycles satisfies k t-2Δt <k t-Δt <k t Then it can be judged that the oil pump bearing temperature rises abnormally, and the abnormal score of the voiceprint data can be determined as the lack of oil pump lubricating oil. ③ Abnormal base vibration: If the voiceprint data S3 on the upper surface of the base is abnormally scored, and the motor and oil pump lubricating oil absence features are not matched, it is determined that the base vibration is abnormal.

7. The sound recognition based oil pump motor and its outlet pipeline working condition abnormality monitoring method according to claim 1, characterized in that: The determination steps of pump body blade fracture in step 3 are as follows: Through PLDA model analysis, a voiceprint feature fault sample library corresponding to the device is established. If the voiceprint feature data collected on site is successfully matched with the fault sample library, the fault type is directly output. The voiceprint data that fails to match is scored by the PLDA model. The abnormally scored data is comprehensively determined for fault according to multiple factors including temperature and pressure. The fault sample is marked by manual confirmation, and the fault sample library is supplemented. The steps are as follows: ④ Pump body blade fracture: Oil pump voiceprint data S 2-1 Or S 2-2 Score anomaly, the rising trend of oil pump bearing temperature T2 is normal, and the oil pump outlet pressure P1 is lower than the normal loading and unloading pressure under loading and unloading conditions, that is, P 1加 <P 加 And P 1卸 <P 卸 Then it can be judged that the oil pump efficiency is abnormal, combined with the normal motor voiceprint data, it can be determined that the oil pump blade is broken.

8. The sound recognition based oil pump motor and its outlet pipeline working condition abnormality monitoring method according to claim 1, characterized in that: The determination steps of check valve function failure in step 3 are as follows: Through PLDA model analysis, a voiceprint feature fault sample library corresponding to the device is established. If the voiceprint feature data collected on site is successfully matched with the fault sample library, the fault type is directly output. The voiceprint data that fails to match is scored by the PLDA model. The abnormally scored data is comprehensively determined for fault according to related pressure data. The fault sample is marked by manual confirmation, and the fault sample library is supplemented. The steps are as follows: ⑤ Check valve function failure: When the oil pump motor is out of service, the voiceprint data S5 of the oil pump outlet pipeline is abnormally scored, P1>0, and the voiceprint data of the oil pump motor is abnormally scored. It can be determined that the oil pump outlet check valve function fails, causing the oil pump motor to reverse.

9. The sound recognition based oil pump motor and its outlet pipeline working condition abnormality monitoring method according to claim 1, characterized in that: The determination steps of oil pump frequent loading and unloading, oil discharge valve leakage, and servomotor pipeline twitching in step 3 are as follows: Through PLDA model analysis, a voiceprint feature fault sample library corresponding to the device is established. If the voiceprint feature data collected on site is successfully matched with the fault sample library, the fault type is directly output. The voiceprint data that fails to match is scored by the PLDA model. The abnormally scored data is comprehensively determined for fault according to multiple factors including pressure and liquid level. The fault sample is marked by manual confirmation, and the fault sample library is supplemented. The steps are as follows: ⑥ Oil pump frequent loading and unloading: The score of the sound print data S6 of the loading and unloading valve group is abnormal, and the time interval t of one loading and unloading of the oil pump is analyzed jx When the pressure of the oil tank P3=6.1MPa, the oil pump starts to load, and the time t is recorded j The next time P3=6.1MPa is recorded as time t j1 Then t jx =t j1 -t j If t jx <t z , wherein t z is the setting value of the loading and unloading interval, it is determined that the loading and unloading of the oil pump is frequent; ⑦ Oil discharge valve not tightly closed or internal leakage: The sound print data S4 of the oil tank drain valve is abnormal, and the loading time t of the oil pump is one jz Too long, that is, t jz >t z1 , t z1 is the loading time setting value, the loading and unloading frequency feature has been matched, the total oil volume V of the hydraulic system is calculated by combining the liquid levels H1 and H2 of the oil tank and the oil collecting tank, if V=V0, V0 is the initial total oil volume of the hydraulic system, it is determined that the drain valve is not tightly closed or there is internal leakage; ⑧ Servomotor pipeline twitching: If the voiceprint data S7 of the pressure regulating valve body and the voiceprint data S8 or S9 of the pressure regulating valve outlet pipeline are abnormally scored, the oil pump frequent loading and unloading feature is matched, and the total oil volume V=V0, it can be determined that the servomotor pipeline is twitching.

10. The sound recognition based oil pump motor and its outlet pipeline working condition abnormality monitoring method according to claims 6-9, characterized in that: After the above fault matching or confirmation, the alarm signal and fault message are output through the man-machine interaction of the upper computer, and the corresponding pre-control measures are started; after the confirmation of any one of faults ①-⑤, the main and standby automatic switching function of the oil pump motor set should be started, the standby oil pump motor set should be started, the fault group should be disabled and marked as unavailable at the same time, and the maintenance should be waited; After the confirmation of fault ⑥, only the alarm signal is output; Faults ⑥ and ⑦ occur at the same time, considering that the oil discharge valve is not tightly closed or has internal leakage, the alarm signal is output, and the start-up condition is locked; Faults ⑥ and ⑧ occur at the same time, confirming that the servomotor pipeline is pumping, the alarm signal is output, the AGC regulating unit output is started, and the actual output of the unit is increased or decreased by 10 MW as a step to adjust the unit operating condition. This process needs to be monitored manually, and manual intervention and on-site confirmation are necessary.

11. The sound recognition based oil pump motor and its outlet pipeline working condition abnormality monitoring method according to claim 1, characterized in that: In step 1, the voiceprint sensor uses a patch type bone conduction microphone, which is arranged on: the two sides of the motor shell, the two sides of the pump body shell, the upper surface of the base, the oil pump outlet pipeline, the loading and unloading valve group, the oil tank discharge valve, the pressure regulating valve body and the pressure regulating valve outlet pipeline; Temperature measuring resistor arrangement: motor bearing and oil pump bearing; Pressure sensor arrangement: oil pump outlet, oil tank and pressure regulating valve inlet; Liquid level sensor arrangement: oil collecting tank and oil tank.

Images