Abnormal sound recognition and positioning method for pressure-bearing pipeline based on sound and vibration fusion

Abstract

The invention discloses a pressure-bearing pipeline abnormal sound recognition and positioning method based on sound-vibration fusion, and belongs to the technical field of pressure-bearing pipeline abnormal sound recognition. The method comprises the following steps: respectively training a plurality of abnormal sound recognition models; respectively training a plurality of abnormal sound positioning models; for each pressure-bearing pipeline in the to-be-detected pressure-bearing pipeline group, acquiring a sound signal and a vibration signal, performing abnormal sound recognition by using the plurality of trained abnormal sound recognition models, and performing decision-level fusion on an output initial result of abnormal sound recognition to output a final result of abnormal sound recognition of the pressure-bearing pipelines; and after the sound signals and the vibration signals of all the pressure-bearing pipelines with the abnormal sounds are fused and / or mixed, abnormal sound positioning is carried out by utilizing the trained abnormal sound positioning models, decision-level fusion is carried out on the output initial abnormal sound positioning results, and a final abnormal sound positioning result is output. The sound and vibration signals are deeply fused, and abnormal sound recognition and positioning can be effectively and accurately carried out on the pressure-bearing pipeline.

Description

Technical Field

[0001] The present invention belongs to the technical field of abnormal sound recognition in pressure pipes, and in particular relates to a method for identifying and locating abnormal sound in pressure pipes based on acoustic-vibration fusion. Background Art

[0002] As the core equipment of the medium transportation and storage system, pressure pipelines are widely used in industrial fields such as electricity and chemical industry. As a key transportation equipment, pressure pipelines need to operate under variable conditions for a long time in harsh environments such as high temperature, high pressure, and corrosion. In this process, they are easily damaged, which in turn causes abnormal vibration failures in the pipelines and radiates abnormal sounds outward, which will have a serious impact on the safe and stable operation of the pipeline system. Therefore, timely and effective identification of abnormal noise in pressure pipelines can reduce operation and maintenance costs and improve the safety and reliability of pipeline systems, which has important theoretical value and practical significance. In addition, in actual projects, the sound and vibration signals of pressure pipelines are often mixed with background noises such as phase regulators, steam turbines, fans, pumps, cooling towers, etc., resulting in unsatisfactory application effects of existing recognition and positioning methods, resulting in low accuracy of recognition and positioning and unreliable detection results. In addition, how to effectively fuse information and improve the accuracy of abnormal sound recognition and positioning deserves in-depth research.

[0003] At present, information fusion methods mainly include data-level fusion, feature-level fusion and decision-level fusion. Data-level fusion directly fuses the original data, which can retain the initial information of the data as much as possible, but has high requirements for data synchronization. Feature-level fusion first extracts features from the data and then fuses these features. This method can effectively reduce the amount of data while retaining key information. Decision-level fusion is based on the decisions made by each subsystem. It has high flexibility but may lose some information. However, in actual engineering, the sound and vibration signals of pressure pipelines are often background noise, resulting in a single fusion method with low accuracy, insufficient applicability and stability. Therefore, there is an urgent need for a method that can perform deep acoustic and vibration fusion based on the characteristics of sound and vibration signals, and realize reliable and accurate recognition and positioning of abnormal sounds in pressure pipelines. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a method for identifying and locating abnormal sounds in pressure pipes based on acoustic-vibration fusion, which can effectively and accurately identify and locate abnormal sounds in pressure pipes.

[0005] The present invention provides the following technical solutions:

[0006] The abnormal sound identification and positioning method of pressure pipe based on acoustic and vibration fusion includes:

[0007] Using the normal sound signal, abnormal sound signal, normal vibration signal and abnormal vibration signal of each pressure pipe in the training pressure pipe group, multiple abnormal sound recognition models are trained respectively;

[0008] Using abnormal sound signals and abnormal vibration signals of all pressure pipes with abnormal sounds in the training pressure pipe group, multiple abnormal sound location models are trained respectively;

[0009] For each pressure pipe m in the pressure pipe group to be detected, the acoustic signal and the vibration signal are collected, and the trained multiple noise recognition models are used to perform noise recognition respectively, and the preliminary noise recognition results output by all noise recognition models are fused at the decision level to output the final result of noise recognition of the pressure pipe m;

[0010] After fusing and / or mixing the acoustic signals and vibration signals of all pressure pipes with abnormal sounds, the abnormal sounds are located using multiple trained abnormal sound localization models, and multiple preliminary abnormal sound localization results are output one by one. The multiple preliminary abnormal sound localization results are fused at the decision level to output the final result of the abnormal sound localization.

[0011] Optionally, the training of multiple heterophone recognition models separately is specifically as follows:

[0012] For each pressure-bearing pipeline in the training pressure-bearing pipeline group, the first normal and abnormal sound signals A1.1 and B1.1 are collected by the first acoustic sensor, the second normal and abnormal sound signals A1.2 and B1.2 are collected by the second acoustic sensor, the first normal and abnormal vibration signals C1.1 and D1.1 are collected by the first vibration sensor, and the second normal and abnormal vibration signals C1.2 and D1.2 are collected by the second vibration sensor; the sampling frequencies of the two acoustic sensors are different, and the sampling frequencies of the two vibration sensors are also different;

[0013] The normal sound signals and the abnormal sound signals of the two sound sensors are respectively fused at the data level to obtain a third normal sound signal A1.3 and a third abnormal sound signal B1.3; the normal vibration signals and the abnormal vibration signals of the two vibration sensors are respectively fused at the data level to obtain a third normal vibration signal C1.3 and a third abnormal vibration signal D1.3;

[0014] For each pressure-bearing pipeline, the first normal and abnormal sound signals A1.1 and B1.1 are mixed, the second normal and abnormal sound signals A1.2 and B1.2 are mixed, the third normal and abnormal sound signals A1.3 and B1.3 are mixed, the first normal and abnormal vibration signals C1.1 and D1.1 are mixed, the second normal and abnormal vibration signals C1.2 and D1.2 are mixed, and the third normal and abnormal vibration signals C1.3 and D1.3 are mixed to obtain the first sound signal X1, the second sound signal X2, the third sound signal X3, the first vibration signal Y1, the second vibration signal Y2, and the third vibration signal Y3 for recognition training;

[0015] Six different heterosound recognition models are trained by respectively using the recognition training of the first acoustic signal X1, the second acoustic signal X2, the third acoustic signal X3, the first vibration signal Y1, the second vibration signal Y2 and the third vibration signal Y3.

[0016] Optionally, the training of multiple heterophonic localization models separately is specifically as follows:

[0017] For each pressure-bearing pipeline with abnormal sound in all training pressure-bearing pipeline groups, two acoustic sensors with different acquisition frequencies are used to collect the first abnormal sound signal B2.1 and the second abnormal sound signal B2.2, and two vibration sensors with different acquisition frequencies are used to collect the first abnormal vibration signal D2.1 and the second abnormal vibration signal D2.2;

[0018] For each pressure-bearing pipeline with abnormal sound, the first abnormal sound signal B2.1 and the first abnormal sound signal B2.2 are data-level fused to obtain the third abnormal sound signal B2.3 of the pressure-bearing pipeline, and the first abnormal vibration signal D2.1 and the second abnormal vibration signal D2.2 are data-level fused to obtain the third abnormal vibration signal D2.3 of the pressure-bearing pipeline;

[0019] Mix the first abnormal sound signal B2.1, the second abnormal sound signal B2.2, the third abnormal sound signal B2.3, the first abnormal vibration signal D2.1, the second abnormal vibration signal D2.2 and the third abnormal vibration signal D2.3 of all the pressure pipes with abnormal sound to obtain the first sound signal P1, the second sound signal P2, the third sound signal P3, the first vibration signal Q1, the second vibration signal Q2 and the third vibration signal Q3 for positioning training;

[0020] Six different heterosound localization models are trained by using the first sound signal P1, the second sound signal P2, the third sound signal P3, the first vibration signal Q1, the second vibration signal Q2 and the third vibration signal Q3 for localization training respectively.

[0021] Optionally, the collecting of sound signals and vibration signals and using a plurality of trained different sound recognition models to perform different sound recognition respectively are specifically as follows:

[0022] For each pressure-bearing pipeline of the pressure-bearing pipeline group to be detected, two sound sensors with different acquisition frequencies are used to collect two sound signals, and two vibration sensors with different acquisition frequencies are used to collect two vibration signals; the two sound signals are fused at the data level to form a third sound signal; the two vibration signals are fused at the data level to form a third vibration signal;

[0023] The three sound signals and three vibration signals of each pressure pipe are respectively input into the corresponding six trained abnormal sound recognition models, and six preliminary abnormal sound recognition results of each pressure pipe are output.

[0024] Optionally, after fusing and / or mixing the acoustic signals and vibration signals of all the pressure pipes with abnormal sounds, the abnormal sounds are located using multiple trained abnormal sounds location models, specifically:

[0025] Mixing the first sound signal, the second sound signal, the third sound signal, the first vibration signal, the second vibration signal and the third vibration signal of all the pressure pipes with abnormal sounds;

[0026] The six mixed signals are input into the corresponding six heterophonic localization models for heterophonic localization.

[0027] Optionally, the data level fusion is specifically:

[0028] Obtain the sampling frequency, the highest sampling frequency, and the lowest sampling frequency of the sensors corresponding to the two signals to be fused at the data level, so as to obtain the weight coefficients of the two signals to be fused at the data level;

[0029]

[0030] Among them, ω 1 and ω 2 are the weight coefficients of the two sensors corresponding to the two signals, F S1 and F S2 are the sampling frequencies of the two sensors corresponding to the two signals, F H1 and F L1 are the highest and lowest sampling frequencies of the first sensor, F H2 and F L2 are the highest and lowest sampling frequencies of the second sensor, respectively;

[0031] Determine the segmentation frequency of each signal according to the weight coefficients of the two signals and the sampling frequencies of the sensors corresponding to the two signals, so as to divide each signal into a low frequency band and a high frequency band;

[0032] The first signal segment frequency F DS1 for:

[0033]

[0034] The second signal's segment frequency F DS2 for:

[0035]

[0036] The expressions for the low and high frequency bands of the first signal are:

[0037] F LS1 =xf 1 [0,F DS1 ],F HS1 =xf 1 (F DS1 ,F S1 ]

[0038] The expressions for the low and high frequency bands of the second signal are:

[0039] F LS2 =xf 2 [0,F DS2 ],F HS2 =xf 2 (F DS2 ,F S2 ]

[0040] Among them, F LS1 and F HS1 are the low frequency band and high frequency band of the first signal respectively; F LS2 and F HS2 are the low frequency and high frequency bands of the second signal, respectively, xf 1 and xf 2 are the frequency domains of the first and second signals respectively;

[0041] According to the sampling frequencies of the two sensors, the high frequency band of the signal corresponding to the sensor with high sampling frequency is fused with the low frequency band of the signal corresponding to the sensor with low sampling frequency:

[0042]

[0043] xf=[xf L ,xf H ]

[0044] Among them, xf is the signal after data-level fusion, xf L and xf H They are the low frequency band and high frequency band of the data-level fusion signal respectively.

[0045] Optionally, the plurality of foreign voice recognition models are all one-dimensional convolutional neural network structures, and the expression of the preliminary foreign voice recognition result output by the foreign voice recognition model is:

[0046]

[0047] Among them, Softmax(i) is the probability of label i, label i indicates normal or there is a foreign sound, Z(i) is the output of label i in the fully connected layer, It is the sum of the original scores of all labels after being processed by the exponential function;

[0048] The plurality of the heterophonic localization models are all one-dimensional convolutional neural network structures, and the expression of the preliminary result of heterophonic localization output by the heterophonic localization model is:

[0049]

[0050] Among them, Softmax(p) is the probability of label p, label p is any pressure-bearing pipe with abnormal sound; Z(p) is the output of p labels in the fully connected layer, is the sum of the original scores of all labels after being processed by the exponential function, and M is the number of pressure pipes with abnormal sounds.

[0051] Optionally, the decision-level fusion is specifically:

[0052] The six preliminary results were combined into a preliminary results matrix PR;

[0053]

[0054] Among them, PS 1j ,PS 2j ,PS 3j , PZ 1j , PZ 2j and PZ 3j are the probabilities of label j in the preliminary results of the six signals, respectively, and T is the number of labels;

[0055] Set the weight matrix ω of the six preliminary results;

[0056]

[0057] Multiply the preliminary result matrix PR and the weight matrix ω of the preliminary result to obtain the preliminary result fusion probability FPR;

[0058]

[0059] The label corresponding to the maximum probability in the preliminary result fusion probability FPR is the final label, and the final label is used as the decision-level fusion result, and the expression is:

[0060] DFR=n,maxP Fn

[0061] Among them, DFR is the decision-level fusion result, and n is the final label.

[0062] Optionally, in the process of training the six different noise recognition models, the first sound signal X1, the second sound signal X2, the third sound signal X3, the first vibration signal Y1, the second vibration signal Y2 and the third vibration signal Y3 are respectively identified and trained, and divided into a number of first identification training sound samples, second identification training sound samples, third identification training sound samples, first identification training vibration samples, second identification training vibration samples and third identification training vibration samples according to a set length, and the six different noise recognition models are trained using the corresponding identification training samples.

[0063] Optionally, in the process of training the six different sound localization models, the first sound signal P1, the second sound signal P2, the third sound signal P3, the first vibration signal Q1, the second vibration signal Q2 and the third vibration signal Q3 are respectively positioned and trained, and divided into a number of first positioning training sound samples, second positioning training sound samples, third positioning training sound samples, first positioning training vibration samples, second positioning training vibration samples and third positioning training vibration samples according to a set length, and the corresponding positioning training samples are used to train the six different sound localization models.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] (1) The present invention realizes data-level fusion of the signals of two acoustic sensors and two vibration sensors respectively, obtains a number of preliminary results by using the abnormal sound recognition and positioning model of the pressurized pipeline, and realizes decision-level fusion of the preliminary results by allocating weights of the preliminary results, thereby realizing deep fusion of the two stages of sound and vibration, improving the fusion quality of the sound and vibration signals, thereby providing support for realizing abnormal sound recognition and positioning of the pressurized pipeline, and ensuring the accuracy of abnormal sound recognition and positioning of the pressurized pipeline.

[0066] (2) The present invention utilizes two acoustic sensors and two vibration sensors to collect normal and abnormal sound signals of a pressure pipeline in different frequency ranges, and respectively fuses the signals of the acoustic sensor and the acoustic sensor, and the signals of the vibration sensor and the vibration sensor at data level to obtain data-level fused acoustic signals and vibration signals, thereby training a pressure pipeline abnormal sound recognition model, and inputting the test acoustic and vibration signals into the trained abnormal sound recognition model after data-level fusion to obtain several preliminary results, and fuses the several preliminary results at decision level to obtain the final pressure pipeline abnormal sound recognition result, thereby improving the accuracy of pressure pipeline abnormal sound recognition.

[0067] (3) The present invention uses two acoustic sensors and two vibration sensors to collect abnormal sound signals of different pressure pipes in different frequency ranges, and respectively fuses the signals of the acoustic sensor and the acoustic sensor, and the signals of the vibration sensor and the vibration sensor at the data level to obtain data-level fused acoustic signals and vibration signals, thereby training the pressure pipe abnormal sound positioning model. The acoustic vibration signal whose abnormal sound recognition result of the pressure pipe is abnormal sound is input into the trained pressure pipe abnormal sound positioning model to obtain several preliminary results, and the several preliminary results are fused at the decision level to obtain the final pressure pipe abnormal sound positioning result, thereby improving the accuracy of the pressure pipe abnormal sound positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a flowchart of the method for identifying and locating abnormal sound in a pressure pipeline based on acoustic-vibration fusion of the present invention;

[0069] Figure 2 It is a flow chart of the training of the abnormal sound recognition model of the pressure pipe of the present invention;

[0070] Figure 3 It is a flow chart of the pressure pipe abnormal sound location model training of the present invention;

[0071] Figure 4 It is a data level fusion flow chart in the process of heterophonic recognition of the present invention;

[0072] Figure 5 It is a decision-level fusion flow chart in the process of heterophonic recognition of the present invention;

[0073] Figure 6 It is a structural diagram of a typical pipeline system given in Example 2 of the present invention. DETAILED DESCRIPTION

[0074] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0075] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the term "comprising" and any variation thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0076] Example 1

[0077] A method for identifying and locating abnormal sound in a pressure pipe based on acoustic-vibration fusion is provided, comprising the following steps:

[0078] S1: using the normal sound signal, abnormal sound signal, normal vibration signal and abnormal vibration signal of each pressure pipe in the training pressure pipe group, respectively training a plurality of abnormal sound recognition models.

[0079] Specifically, Figure 2 As shown, step S1 includes the following sub-steps:

[0080] Step S11: For each pressure pipeline in the training pressure pipeline group, use the first acoustic sensor to collect the first normal and abnormal sound signals A1.1 and B1.1, use the second acoustic sensor to collect the second normal and abnormal sound signals A1.2 and B1.2, use the first vibration sensor to collect the first normal and abnormal vibration signals C1.1 and D1.1, and use the second vibration sensor to collect the second normal and abnormal vibration signals C1.2 and D1.2; the sampling frequencies of the two acoustic sensors are different, and the sampling frequencies of the two vibration sensors are also different.

[0081] The two acoustic sensors can be installed at either end of the pipeline, or at a set distance. The two vibration sensors can be installed at one horizontal direction and the other vertical direction, so as to facilitate the measurement of the vibration of the pipeline in two directions.

[0082] As an option, the sampling frequencies of the two acoustic sensors are 2560 Hz and 25600 Hz respectively, the sampling frequencies of the two vibration sensors are 2560 Hz and 25600 Hz respectively, and the sampling times are preferably 100 s and 10 s respectively.

[0083] Step S12: Perform data-level fusion on the first and second normal sound signals A1.1 and A1.2 to obtain a third normal sound signal A1.3; perform data-level fusion on the first and second abnormal sound signals B1.1 and B1.2 to obtain a third abnormal sound signal B1.3; perform data-level fusion on the first and second normal vibration signals C1.1 and C1.2 to obtain a third normal vibration signal C1.3; perform data-level fusion on the first and second abnormal vibration signals D1.1 and D1.2 to obtain a third abnormal vibration signal D1.3.

[0084] Step S13: Mix the first normal and abnormal sound signals A1.1 and B1.1 to obtain the first sound signal X1 for recognition training.

[0085] Step S14: Mix the second normal and abnormal sound signals A1.2 and B1.2 to obtain a recognition training second sound signal X2;

[0086] Step S15: Mix the third normal and abnormal sound signals A1.3 and B1.3 to obtain a third recognition training sound signal X3;

[0087] Step S16: Mix the first normal and abnormal vibration signals C1.1 and D1.1 to obtain the first vibration signal Y1 for identification training;

[0088] Step S17: Mix the second normal and abnormal vibration signals C1.2 and D1.2 to obtain the second vibration signal Y2 for identification training;

[0089] Step S18: Mix the third normal and abnormal vibration signals C1.3 and D1.3 to obtain a third vibration signal Y3 for recognition training;

[0090] Step S19: Six different heterosound recognition models are trained respectively by using the first acoustic signal X1 for recognition training, the second acoustic signal X2 for recognition training, the third acoustic signal X3 for recognition training, the first vibration signal Y1 for recognition training, the second vibration signal Y2 for recognition training and the third vibration signal Y3 for recognition training.

[0091] Furthermore, in the process of training six different abnormal sound recognition models, the recognition training first sound signal X1, the recognition training second sound signal X2, the recognition training third sound signal X3, the recognition training first vibration signal Y1, the recognition training second vibration signal Y2 and the recognition training third vibration signal Y3 are respectively divided into a number of first recognition training sound samples, second recognition training sound samples, third recognition training sound samples, first recognition training vibration samples, second recognition training vibration samples and third recognition training vibration samples according to the set length, and the corresponding recognition training samples are used to train six different abnormal sound recognition models.

[0092] As an option, each recognition training sound sample and vibration sample has 100 samples.

[0093] Each foreign voice recognition model is a one-dimensional convolutional neural network structure. The expression of the preliminary foreign voice recognition result output by the foreign voice recognition model is:

[0094]

[0095] Among them, Softmax(i) is the label i The probability of label i indicates normal or abnormal sound, Z(i) is the output of label i in the fully connected layer, It is the sum of the original scores of all labels after being processed by the exponential function.

[0096] Specifically, the sample types of the heterophonic recognition model are 0 and 1, 0 indicates normal, and 1 indicates the presence of heterophonic sound.

[0097] S2: using the abnormal sound signals and abnormal vibration signals of all the pressure pipes with abnormal sound in the training pressure pipe group, respectively training a plurality of abnormal sound location models.

[0098] Specifically, Figure 3 As shown, step S2 includes the following sub-steps:

[0099] Step S21: For each pressure-bearing pipeline with abnormal sound in all training pressure-bearing pipeline groups, use two acoustic sensors with different collection frequencies to collect the first abnormal sound signal B2.1 and the second abnormal sound signal B2.2, and use two vibration sensors with different collection frequencies to collect the first abnormal vibration signal D2.1 and the second abnormal vibration signal D2.2.

[0100] The sampling frequency and sampling time of the acoustic sensor used in the training of the foreign sound recognition model and the foreign sound localization model are consistent. Similarly, the sampling frequency and sampling time of the vibration sensor used in the training of the foreign sound recognition model and the foreign sound localization model are also consistent.

[0101] Step S22: For each pressure pipeline with abnormal sound, the first abnormal sound signal B2.1 and the first abnormal sound signal B2.2 are data-level fused to obtain the third abnormal sound signal B2.3 of the pressure pipeline, and the first abnormal vibration signal D2.1 and the second abnormal vibration signal D2.2 are data-level fused to obtain the third abnormal vibration signal D2.3 of the pressure pipeline.

[0102] Step S23: Mix the first abnormal sound signals B2.1 of all the pressure pipes with abnormal sound to obtain the first sound signal P1 for positioning training;

[0103] Step S24: Mix the second abnormal sound signals B2.2 of all the pressure pipes with abnormal sound to obtain the second sound signal P2 for positioning training;

[0104] Step S25: Mix the third abnormal sound signals B2.3 of all the pressure pipes with abnormal sound to obtain the third sound signal P3 for positioning training;

[0105] Step S26: Mix the first abnormal vibration signals D2.1 of all the pressure pipes with abnormal sounds to obtain the first vibration signal Q1 for positioning training;

[0106] Step S27: Mix the second abnormal vibration signals D2.2 of all the pressure pipes with abnormal sounds to obtain the second vibration signal Q2 for positioning training;

[0107] Step S28: Mix the third abnormal vibration signals D2.3 of all the pressure pipes with abnormal sounds to obtain the third vibration signal Q3 for positioning training;

[0108] Step S29: Six different noise localization models are trained using the first sound signal P1 for localization training, the second sound signal P2 for localization training, the third sound signal P3 for localization training, the first vibration signal Q1 for localization training, the second vibration signal Q2 for localization training and the third vibration signal Q3 for localization training.

[0109] More specifically, in the process of training six different heterosound localization models, the first sound signal P1, the second sound signal P2, the third sound signal P3, the first vibration signal Q1, the second vibration signal Q2 and the third vibration signal Q3 are respectively positioned and trained, and a number of first positioning training sound samples, second positioning training sound samples, third positioning training sound samples, first positioning training vibration samples, second positioning training vibration samples and third positioning training vibration samples are divided according to the set length, and the corresponding positioning training samples are used to train six different heterosound localization models.

[0110] Each heterophonic localization model is a one-dimensional convolutional neural network structure. The expression of the preliminary result of heterophonic localization output by the heterophonic localization model is:

[0111]

[0112] Among them, Softmax(p) is the label p The probability of label p is any pressure pipe with abnormal sound; Z(p) is p The output of the fully connected layer is is the sum of the original scores of all labels after being processed by the exponential function, and M is the number of pressure pipes with abnormal sounds.

[0113] The sample type of the abnormal sound location model is the label of the pressure pipe, such as sample type 1 represents pressure pipe 1, and the sample type 2 represents pressure pipe 2; the one-dimensional convolutional neural network structure refers to the prior art.

[0114] S3: For each pressure pipe m in the pressure pipe group to be detected, the acoustic signal and the vibration signal are collected, and the abnormal sound recognition is performed respectively using multiple trained abnormal sound recognition models, and the preliminary results of the abnormal sound recognition output by all the abnormal sound recognition models are fused at the decision level to output the final result of the abnormal sound recognition of the pressure pipe m.

[0115] Specifically, Figure 1 As shown, for each pressure-bearing pipeline of the pressure-bearing pipeline group to be detected, two sound sensors with different acquisition frequencies are used to collect two sound signals, and two vibration sensors with different acquisition frequencies are used to collect two vibration signals; the two sound signals are fused at the data level to form a third sound signal; the two vibration signals are fused at the data level to form a third vibration signal;

[0116] After the three sound signals and three vibration signals of each pressure pipeline are divided into samples, they are respectively input into the corresponding six trained abnormal sound recognition models, and the six preliminary abnormal sound recognition results of each pressure pipeline are output.

[0117] S4: After fusing and / or mixing the acoustic signals and vibration signals of all the pressure pipes with abnormal sounds, use the trained multiple abnormal sound localization models to perform abnormal sound localization, output multiple preliminary abnormal sound localization results one by one, and perform decision-making-level fusion on the multiple preliminary abnormal sound localization results to output the final result of abnormal sound localization.

[0118] Specifically, the first sound signal, the second sound signal, the third sound signal, the first vibration signal, the second vibration signal and the third vibration signal of all the pressure pipes with abnormal sound are mixed;

[0119] The six mixed signals are divided into samples and input into the corresponding six heterophonic localization models for heterophonic localization.

[0120] In this embodiment, if Figure 4 As shown in the figure, data-level fusion is specifically as follows:

[0121] Obtain the sampling frequency, the highest sampling frequency, and the lowest sampling frequency of the sensors corresponding to the two signals to be fused at the data level, so as to obtain the weight coefficients of the two signals to be fused at the data level;

[0122]

[0123] Among them, ω 1 and ω 2 are the weight coefficients of the two sensors (which can be two acoustic sensors or two vibration sensors) corresponding to the two signals, respectively, S1 and F S2 are the sampling frequencies of the two sensors corresponding to the two signals, F H1 and F L1 are the highest and lowest sampling frequencies of the first sensor, F H2 and F L2 are the highest and lowest sampling frequencies of the second sensor, respectively;

[0124] Determine the segmentation frequency of each signal according to the weight coefficients of the two signals and the sampling frequencies of the sensors corresponding to the two signals, so as to divide each signal into a low frequency band and a high frequency band;

[0125] The first signal segment frequency F DS1 for:

[0126]

[0127] The second signal's segment frequency F DS2 for:

[0128]

[0129] The expressions for the low and high frequency bands of the first signal are:

[0130] F LS1 =xf 1 [0,F DS1 ],F HS1 =xf 1 (F DS1 ,F S1 ]

[0131] The expressions for the low and high frequency bands of the second signal are:

[0132] F LS2 =xf 2 [0,F DS2 ],F HS2 =xf 2 (F DS2 ,F S2 ]

[0133] Among them, F LS1 and F HS1 are the low frequency band and high frequency band of the first signal respectively; F LS2 and F HS2 are the low frequency and high frequency bands of the second signal, respectively, xf 1 and xf 2 are the frequency domains of the first and second signals respectively;

[0134] According to the sampling frequencies of the two sensors, the high frequency band of the signal corresponding to the sensor with high sampling frequency is fused with the low frequency band of the signal corresponding to the sensor with low sampling frequency:

[0135]

[0136] xf=[xf L ,xf H ]

[0137] Among them, xf is the signal after data-level fusion, xf L and xf H They are the low frequency band and high frequency band of the data-level fusion signal respectively.

[0138] That is, if the sampling frequency of the first sound sensor (or the first vibration sensor) is greater than or equal to the sampling frequency of the second sound sensor (or the second vibration sensor), the low frequency band of the second sound signal (or the second vibration signal) and the high frequency band of the first sound signal (or the first vibration signal) are combined to obtain a data-level fusion signal; if the sampling frequency of the first sound sensor (or the first vibration sensor) is less than the sampling frequency of the second sound sensor (or the second vibration sensor), the low frequency band of the first sound signal (or the first vibration signal) and the high frequency band of the second sound signal (or the second vibration signal) are combined to obtain a data-level fusion signal.

[0139] In this embodiment, if Figure 5 As shown in Figure 2, the decision-level fusion is specifically as follows:

[0140] Step a: merge the six preliminary results into a preliminary result matrix PR;

[0141]

[0142] Among them, PS 1j ,PS 2j ,PS 3j , PZ 1j , PZ 2j and PZ 3j are the probabilities of label j in the preliminary results of the six signals, respectively, and T is the number of labels;

[0143] Step b: Set the weight matrix ω of the six preliminary results;

[0144]

[0145] Step c: Multiply the preliminary result matrix PR and the weight matrix ω of the preliminary results, perform statistical calculations on the label probabilities of all preliminary results, and obtain the preliminary result fusion probability FPR;

[0146]

[0147] Step d: The label corresponding to the maximum probability in the preliminary result fusion probability FPR is the final label, and the final label is used as the decision-level fusion result, expressed as:

[0148] DFR=n,maxP Fn

[0149] Among them, DFR is the decision-level fusion result, and n is the final label.

[0150] Example 2

[0151] A specific example of a method for identifying and locating abnormal sound in a pressure pipe based on acoustic-vibration fusion is provided. Figure 6As shown in the figure, a pressure pipe system is built, including a motor, a water pump, a pressure reducing valve, a pressure pipe, a water tank, a hoisting device, a collection card, an acoustic sensor, etc. The water in the water tank is pumped out by the water pump, and is regulated by the pressure reducing valve to move along the pressure pipe, and finally flows back to the water tank to form a cycle. In order to simulate the generation of abnormal sound in the pipeline, a typical pipeline system is set up according to the actual structural characteristics of the power boiler, which consists of three pipes with three degrees of freedom.

[0152] In order to simulate the actual installation method of pipeline hoisting in the boiler system, the experimental pipeline is suspended in the air by a hoisting spring fixed at the outlet. The fixture is loosened to simulate the abnormal vibration of the pressure pipeline, thereby generating an abnormal sound signal. Three pressures of 5PSI, 10PSI and 15PSI are set through the water pump and the pressure reducing valve to simulate three working conditions, and three pipes are set to simulate three positions.

[0153] In order to collect signals of different frequencies, the sampling frequencies of the two acoustic sensors and the two vibration sensors are set to 2560 Hz and 25600 Hz respectively, and the sampling time is 100 s and 10 s respectively. The samples are divided according to the length of 2560, and there are 100 samples in each case. Finally, there are 2×2×2×3×3×100 (2 sensors: acoustic sensor and vibration sensor, 2 cases: normal and abnormal sound, 2 frequencies: high frequency and low frequency, 3 pressures: 5, 10 and 15 PSI, 3 pipes: No. 1, No. 2 and No. 3) signals in each sample.

[0154] Firstly, the normal and abnormal sound signals of high-frequency and low-frequency sound and vibration of a certain pressure and a certain pipeline are fused at the data level to obtain three types of 3×2×2×100 (3 types of signals: sound signal, vibration signal and data-level fusion signal, 2 situations: normal and abnormal sound, 2 frequencies: high frequency and low frequency) samples, which are input into the abnormal sound recognition model of pressure pipeline for training, and 6 (3 types of signals: sound signal, vibration signal and data-level fusion signal, 2 frequencies: high frequency and low frequency) abnormal sound recognition models of pressure pipeline after training are obtained.

[0155] The abnormal sound signals of high-frequency and low-frequency sound and vibration of a certain pressure and three pipelines were fused at the data level to obtain three types of 3×3×2×100 (3 types of signals: sound signal, vibration signal and data-level fusion signal, 3 pipelines: No. 1, No. 2 and No. 3, 2 frequencies: high frequency and low frequency) samples, which were input into the pressure pipeline abnormal sound localization model for training, and 6 (3 types of signals: sound signal, vibration signal and data-level fusion signal, 2 frequencies: high frequency and low frequency) trained pressure pipeline abnormal sound localization models were obtained.

[0156] 70% of the sound and vibration samples were randomly selected from 2×2×2×3×3×100 samples as test samples, with a total of 2×2×2×3×3×70 test samples. The high- and low-frequency sound and vibration samples were respectively fused at the data level to obtain data-level fusion test samples, and finally there were 3×2×2×3×3×70 (3 types of signals: sound signal, vibration signal, data-level fusion signal, 2 situations: normal and abnormal sound, 2 frequencies: high frequency and low frequency, 3 pressures: 5, 10 and 15PSI, 3 pipes: No. 1, No. 2 and No. 3) test samples. The test samples were respectively input into the trained pressure pipeline abnormal sound recognition model to obtain 6 (3 types of signals: sound signal, vibration signal and data-level fusion signal, 2 frequencies: high frequency and low frequency) abnormal sound recognition preliminary results, and the 6 abnormal sound recognition preliminary results were fused at the decision level to obtain the pressure pipeline abnormal sound recognition result. For samples with abnormal sound results, the test samples were further input into the trained pressure pipeline abnormal sound location model to obtain 6 preliminary results of abnormal sound location (3 types of signals: acoustic signal, vibration signal and data-level fusion signal, 2 frequencies: high frequency and low frequency). The 6 preliminary results of abnormal sound location were fused at the decision level to obtain the pressure pipeline abnormal sound location result. The three pressures were used as three working conditions, and the accuracy of the proposed method in pressure pipeline abnormal sound recognition and location was tested under 5PSI, 10PSI and 15PSI working conditions, respectively. The results are shown in Table 1.

[0157] Table 1 Results of abnormal sound identification and location in pressure pipes

[0158]

[0159] It can be seen that by fusing sound and vibration signals at the data level and decision level, the accuracy of abnormal sound recognition of pressure pipes under different pressures exceeds 92.86%, and the average value reaches 93.57%, which verifies the accuracy and reliability of the method of the present application. Since the location of abnormal sound of three pipelines is involved, the accuracy of pipeline abnormal sound location is lower than that of pipeline abnormal sound recognition, but the average accuracy reaches 91.75%, which verifies the accuracy and stability of the method of the present application to a certain extent.

[0160] In addition, in order to analyze the effect of the proposed decision-level fusion, taking 5PSI pressure as an example, the accuracy of the six preliminary results in pipeline abnormal sound recognition and positioning was calculated respectively, and compared and analyzed with the decision-level fusion results, and the results are shown in Table 2. It can be seen that due to the use of original sound and vibration signals, the accuracy of preliminary results 1, preliminary results 2, preliminary results 4 and preliminary results 5 is the lowest, and the accuracy of abnormal sound positioning is also lower than that of abnormal sound recognition. After using data-level fusion, the accuracy of preliminary results 3 and preliminary results 6 has improved. Compared with the original sound and vibration signals, the accuracy of abnormal sound recognition and positioning of the data-level fusion results has increased by about 6%, which further verifies the effectiveness and reliability of data-level fusion in the method of this application. After the six preliminary results are further fused at the decision level, the accuracy of the final decision-level fusion results in abnormal sound recognition and positioning exceeds 90%, and the accuracy is further improved by about 5% compared with the results of data-level fusion, which further verifies the effectiveness and reliability of decision-level fusion in the method of this application.

[0161] Table 2 Preliminary results and decision-level fusion results of abnormal noise recognition and location of pressure pipes under 5PSI conditions

[0162]

[0163]

[0164] Compared with the prior art, the method for identifying and locating abnormal sounds of pressure pipes based on acoustic and vibration fusion of the present invention uses two acoustic sensors and two vibration sensors to collect acoustic and vibration signals of pressure pipes in different frequency ranges, and fuses the two acoustic and vibration signals at the data level to train an abnormal sound identification and positioning model for pressure pipes. After training, the abnormal sound identification and positioning model for pressure pipes can further fuse the acoustic, vibration and data-level fusion results at the decision level to realize abnormal sound identification and positioning of pressure pipes, and improves the accuracy and reliability of abnormal sound identification and positioning of pressure pipes by deeply fusing acoustic and vibration information.

[0165] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. Those skilled in the art can clearly understand that the technology in the embodiments of the present invention can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the technical solutions in the embodiments of the present invention are essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present invention or certain parts of the embodiments.

[0166] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A method for identifying and locating abnormal sound in a pressure pipe based on acoustic-vibration fusion, characterized in that: include: Using the normal sound signal, abnormal sound signal, normal vibration signal and abnormal vibration signal of each pressure pipe in the training pressure pipe group, multiple abnormal sound recognition models are trained respectively; Using abnormal sound signals and abnormal vibration signals of all pressure pipes with abnormal sounds in the training pressure pipe group, multiple abnormal sound location models are trained respectively; For each pressure pipe m in the pressure pipe group to be detected, the acoustic signal and the vibration signal are collected, and the trained multiple noise recognition models are used to perform noise recognition respectively, and the preliminary noise recognition results output by all noise recognition models are fused at the decision level to output the final result of noise recognition of the pressure pipe m; After fusing and / or mixing the acoustic signals and vibration signals of all pressure pipes with abnormal sounds, the abnormal sounds are located using multiple trained abnormal sound localization models, and multiple preliminary abnormal sound localization results are output one by one. The multiple preliminary abnormal sound localization results are fused at the decision level to output the final result of the abnormal sound localization.

2. The method for identifying and locating abnormal sound in a pressure pipe based on acoustic-vibration fusion according to claim 1 is characterized in that: The training of multiple heterophone recognition models respectively includes: For each pressure-bearing pipeline in the training pressure-bearing pipeline group, the first normal and abnormal sound signals A1.1 and B1.1 are collected by the first acoustic sensor, the second normal and abnormal sound signals A1.2 and B1.2 are collected by the second acoustic sensor, the first normal and abnormal vibration signals C1.1 and D1.1 are collected by the first vibration sensor, and the second normal and abnormal vibration signals C1.2 and D1.2 are collected by the second vibration sensor; the sampling frequencies of the two acoustic sensors are different, and the sampling frequencies of the two vibration sensors are also different; The normal sound signals and the abnormal sound signals of the two sound sensors are respectively fused at the data level to obtain a third normal sound signal A1.3 and a third abnormal sound signal B1.3; the normal vibration signals and the abnormal vibration signals of the two vibration sensors are respectively fused at the data level to obtain a third normal vibration signal C1.3 and a third abnormal vibration signal D1.3; For each pressure-bearing pipeline, the first normal and abnormal sound signals A1.1 and B1.1 are mixed, the second normal and abnormal sound signals A1.2 and B1.2 are mixed, the third normal and abnormal sound signals A1.3 and B1.3 are mixed, the first normal and abnormal vibration signals C1.1 and D1.1 are mixed, the second normal and abnormal vibration signals C1.2 and D1.2 are mixed, and the third normal and abnormal vibration signals C1.3 and D1.3 are mixed to obtain the first sound signal X1, the second sound signal X2, the third sound signal X3, the first vibration signal Y1, the second vibration signal Y2, and the third vibration signal Y3 for recognition training; Six different heterosound recognition models are trained by respectively using the recognition training of the first acoustic signal X1, the second acoustic signal X2, the third acoustic signal X3, the first vibration signal Y1, the second vibration signal Y2 and the third vibration signal Y3.

3. The method for identifying and locating abnormal sound in a pressure pipe based on acoustic-vibration fusion according to claim 1 is characterized in that: The training of multiple heterophonic localization models is specifically as follows: For each pressure-bearing pipeline with abnormal sound in all training pressure-bearing pipeline groups, two acoustic sensors with different acquisition frequencies are used to collect the first abnormal sound signal B2.1 and the second abnormal sound signal B2.2, and two vibration sensors with different acquisition frequencies are used to collect the first abnormal vibration signal D2.1 and the second abnormal vibration signal D2.2; For each pressure-bearing pipeline with abnormal sound, the first abnormal sound signal B2.1 and the first abnormal sound signal B2.2 are data-level fused to obtain the third abnormal sound signal B2.3 of the pressure-bearing pipeline, and the first abnormal vibration signal D2.1 and the second abnormal vibration signal D2.2 are data-level fused to obtain the third abnormal vibration signal D2.3 of the pressure-bearing pipeline; Mix the first abnormal sound signal B2.1, the second abnormal sound signal B2.2, the third abnormal sound signal B2.3, the first abnormal vibration signal D2.1, the second abnormal vibration signal D2.2 and the third abnormal vibration signal D2.3 of all the pressure pipes with abnormal sound to obtain the first sound signal P1, the second sound signal P2, the third sound signal P3, the first vibration signal Q1, the second vibration signal Q2 and the third vibration signal Q3 for positioning training; Six different heterosound localization models are trained by using the first sound signal P1, the second sound signal P2, the third sound signal P3, the first vibration signal Q1, the second vibration signal Q2 and the third vibration signal Q3 for localization training respectively.

4. The method for identifying and locating abnormal sound in a pressure pipe based on acoustic-vibration fusion according to claim 1 is characterized in that: The collecting of sound signals and vibration signals and using the trained multiple noise recognition models to perform noise recognition respectively are specifically as follows: For each pressure-bearing pipeline of the pressure-bearing pipeline group to be detected, two sound sensors with different acquisition frequencies are used to collect two sound signals, and two vibration sensors with different acquisition frequencies are used to collect two vibration signals; the two sound signals are fused at the data level to form a third sound signal; the two vibration signals are fused at the data level to form a third vibration signal; The three sound signals and three vibration signals of each pressure pipe are respectively input into the corresponding six trained abnormal sound recognition models, and six preliminary abnormal sound recognition results of each pressure pipe are output.

5. The method for identifying and locating abnormal sound in a pressure pipe based on acoustic-vibration fusion according to claim 4 is characterized in that: After fusing and / or mixing the acoustic signals and vibration signals of all the pressure pipes with abnormal sounds, the abnormal sounds are located using multiple trained abnormal sound location models, specifically: Mixing the first sound signal, the second sound signal, the third sound signal, the first vibration signal, the second vibration signal and the third vibration signal of all the pressure pipes with abnormal sounds; The six mixed signals are input into the corresponding six heterophonic localization models for heterophonic localization.

6. The method for identifying and locating abnormal sound in a pressure pipe based on acoustic-vibration fusion according to any one of claims 2, 3 and 4, characterized in that: The data level fusion is specifically: Obtain the sampling frequency, the highest sampling frequency, and the lowest sampling frequency of the sensors corresponding to the two signals to be fused at the data level, so as to obtain the weight coefficients of the two signals to be fused at the data level; Among them, ω1 and ω2 are the weight coefficients of the two sensors corresponding to the two signals, F S1 and F S2 are the sampling frequencies of the two sensors corresponding to the two signals, F H1 and F L1 are the highest and lowest sampling frequencies of the first sensor, F H2 and F L2 are the highest and lowest sampling frequencies of the second sensor, respectively; Determine the segmentation frequency of each signal according to the weight coefficients of the two signals and the sampling frequencies of the sensors corresponding to the two signals, so as to divide each signal into a low frequency band and a high frequency band; The first signal segment frequency F DS1 for: The second signal's segment frequency F DS2 for: The expressions for the low and high frequency bands of the first signal are: F LS1 =xf1[0,F DS1 ],F HS1 =xf1(F DS1 ,F S1 ] The expressions for the low and high frequency bands of the second signal are: F LS2 =xf2[0,F DS2 ],F HS2 =xf2(F DS2 ,F S2 ] Among them, F LS1 and F HS1 are the low frequency band and high frequency band of the first signal respectively; F LS2 and F HS2 are the low frequency band and high frequency band of the second signal respectively, and xf1 and xf2 are the frequency domains of the first signal and the second signal respectively; According to the sampling frequencies of the two sensors, the high frequency band of the signal corresponding to the sensor with high sampling frequency is fused with the low frequency band of the signal corresponding to the sensor with low sampling frequency: xf=[xf L ,xf H ] Among them, xf is the signal after data-level fusion, xf L and xf H They are the low frequency band and high frequency band of the data-level fusion signal respectively.

7. The method for identifying and locating abnormal sound in a pressure pipe based on acoustic-vibration fusion according to claim 1 is characterized in that: The plurality of heterophonic recognition models are all one-dimensional convolutional neural network structures, and the expression of the preliminary heterophonic recognition result output by the heterophonic recognition model is: Among them, Softmax(i) is the probability of label i, label i indicates normal or there is a foreign sound, Z(i) is the output of label i in the fully connected layer, It is the sum of the original scores of all labels after being processed by the exponential function; The plurality of the heterophonic localization models are all one-dimensional convolutional neural network structures, and the expression of the preliminary result of heterophonic localization output by the heterophonic localization model is: Among them, Softmax(p) is the probability of label p, label p is any pressure-bearing pipe with abnormal sound; Z(p) is the output of p labels in the fully connected layer, is the sum of the original scores of all labels after being processed by the exponential function, and M is the number of pressure pipes with abnormal sounds.

8. The method for identifying and locating abnormal sound in a pressure pipe based on acoustic-vibration fusion according to claim 1 is characterized in that: The decision-level fusion is specifically: The six preliminary results were combined into a preliminary results matrix PR; Among them, PS 1j ,PS 2j ,PS 3j , PZ 1j , PZ 2j and PZ 3j are the probabilities of label j in the preliminary results of the six signals, respectively, and T is the number of labels; Set the weight matrix ω of the six preliminary results; Multiply the preliminary result matrix PR and the weight matrix ω of the preliminary result to obtain the preliminary result fusion probability FPR; The label corresponding to the maximum probability in the preliminary result fusion probability FPR is the final label, and the final label is used as the decision-level fusion result, and the expression is: DFR=n,maxP Fn Among them, DFR is the decision-level fusion result, and n is the final label.

9. The method for identifying and locating abnormal sound in a pressure pipe based on acoustic-vibration fusion according to claim 2 is characterized in that: In the process of training the six different heterophonic sound recognition models, the first acoustic signal X1, the second acoustic signal X2, the third acoustic signal X3, the first vibration signal Y1, the second vibration signal Y2 and the third vibration signal Y3 are respectively identified and trained, and a number of first identification training acoustic samples, second identification training acoustic samples, third identification training acoustic samples, first identification training vibration samples, second identification training vibration samples and third identification training vibration samples are divided according to the set length, and the six different heterophonic sound recognition models are trained using the corresponding identification training samples.

10. The method for identifying and locating abnormal sound in a pressure pipe based on acoustic-vibration fusion according to claim 3 is characterized in that: In the process of training the six different heterosound localization models, the first sound signal P1, the second sound signal P2, the third sound signal P3, the first vibration signal Q1, the second vibration signal Q2 and the third vibration signal Q3 are respectively positioned and trained, and a number of first positioning training sound samples, second positioning training sound samples, third positioning training sound samples, first positioning training vibration samples, second positioning training vibration samples and third positioning training vibration samples are divided according to the set length, and the corresponding positioning training samples are used to train the six different heterosound localization models.

Images