A wellbore casing collision monitoring method based on artificial intelligence acoustic identification

Abstract

The application discloses a kind of wellbore tankway collision monitoring methods based on artificial intelligence acoustic identification.The application, through acoustic acquisition device, real-time and continuously collects pipeline operation sound, and quickly carries out sound data cleaning, feature processing, model identification process, can carry out quick alarm to abnormal collision in short time, effectively improve the safety of wellbore operation.Through the multi-index data cleaning of original sound signal, the background noise and invalid signal in sound are effectively filtered, and the representation ability and distinguishing ability of feature vector to tankway collision sound are enhanced by calculating multiple acoustic characteristics, to improve the accuracy of the trained model in complex wellbore environment.The support vector data description method is used to distinguish normal and abnormal acoustic patterns, without a large number of abnormal samples training, suitable for the situation that abnormal collision samples are single and normal samples are complex and diverse in actual wellbore scene, with good environmental adaptability and expansibility.

Description

Technical Field

[0001] This invention belongs to the field of mine anomaly detection technology, specifically a method for monitoring collisions in mine shafts and tunnels based on artificial intelligence acoustic recognition. Background Technology

[0002] Track rollers typically consist of a metal core and a rubber outer layer. During operation, friction between the track roller and the track wears down the outer rubber layer. When the rubber wears down to a certain extent, collisions occur between the track and the track lugs. Track roller training in mine shafts is usually conducted through manual inspections or periodic maintenance, relying on human experience to determine the presence of abnormal collisions. These methods suffer from long delays, low efficiency, and poor accuracy, making it difficult to detect abnormal collisions caused by track roller rubber wear in a timely manner. During production operations, manual inspection and listening are insufficient for timely and accurate identification and location; in most cases, by the time abnormalities are discovered manually, they have already escalated, causing additional losses. Currently, some research utilizes visual inspection and vibration sensing for anomaly diagnosis, but these methods are limited by the complex environment of mines, influenced by factors such as dust, lighting, and vibration interference from other equipment. In contrast, sound signals, as a non-contact, real-time, and highly sensitive feature, can effectively characterize equipment operating status and represent an important research area in intelligent equipment operation and maintenance.

[0003] Currently, some studies have attempted to identify abnormal sounds in other industrial scenarios using template matching methods. However, such methods rely on pre-defined features and have poor robustness, making them unsuitable for complex acoustic environments such as mines where there is wind noise, noise from other equipment, and other noise.

[0004] Therefore, a novel collision monitoring method for well shafts and tank passages based on artificial intelligence acoustic recognition is proposed to accurately and timely detect abnormal collision sounds during the operation of well shafts and tank passages, thereby improving the safety of the system and the level of automation of inspections. Summary of the Invention

[0005] The purpose of this invention is to provide a method for monitoring collisions in well shafts and tank passages based on artificial intelligence acoustic recognition, in order to solve the problems mentioned above.

[0006] The technical solution adopted in this invention is as follows: A method for detecting collisions in well shaft tank passages based on artificial intelligence acoustic recognition, characterized in that: the method includes the following steps:

[0007] S1: Collect sound during operation;

[0008] S2: Perform data reading and preprocessing;

[0009] S3: Perform feature extraction;

[0010] S4: Construct a model for monitoring collisions in well shafts and tank passages;

[0011] S5: Train and deploy the model;

[0012] S6: Demonstrate the model's test performance;

[0013] In step S4, kernel principal component analysis is used to perform unsupervised feature compression on the normalized features. Compared with traditional principal component analysis, this method can extract linear components in high-dimensional implicit space and features with non-linear relationships. Let the sample... Each sample This indicates that there are d=10 features; first, a linear kernel function is calculated, and the kernel matrix is ​​constructed using the linear kernel function. :

[0014]

[0015] in, This indicates the calculation of the inner product of two vectors; then, the kernel matrix is ​​centered, as follows:

[0016]

[0017] in, for The whole Matrix; finally, the centered kernel matrix Perform eigenvalue decomposition and retain the first eigenvalue. One principal component, i.e., solving:

[0018]

[0019] in, and Representing the eigenvalues ​​and eigenvectors respectively, we finally obtain the dimensionality-reduced features. .

[0020] In a preferred embodiment, in step S1, a microphone is used as a sound acquisition device and installed near key points of the well shaft tank passage equipment to continuously collect the sound of the tank passage operation process, save the sound file according to a certain duration, and name the sound file by the device number and the time of the first second of the audio to indicate the time when the sound occurred.

[0021] In a preferred embodiment, in step S2, the audio file collected by the microphone is read at the original sampling rate to ensure that the accuracy on the time axis is not affected by resampling, and the frequency, phase and amplitude information are completely preserved, so as to obtain the continuous amplitude value that changes with time as the audio signal, which reflects the sound pressure change generated by the collision of the can ears in the well shaft.

[0022] The obtained audio signal is processed by frame segmentation, which divides the audio signal into frame segments according to the set frame length and frame shift. This process is implemented by sliding window. Each frame segment is a short audio signal that represents the local dynamic characteristics of the sound and does not disturb or lose temporal resolution.

[0023] By reading audio and processing it in frames, each frame segment is treated as an audio signal sample for data cleaning to improve the signal-to-noise ratio and stability of the data. This patent selects five data cleaning indicators: root mean square energy value, lower limit of dominant frequency, upper limit of dominant frequency, local energy value, and kurtosis value.

[0024] The root mean square energy value measures the total energy level of each sample audio signal by calculating the root mean square of the sample. When the root mean square of a sample is less than a set threshold, it is considered to have too low energy, and is a silent, no signal or background noise segment, and is therefore removed.

[0025] The lower limit of the main frequency indicates that samples with the highest frequency below the set threshold are considered inactive or invalid and should be removed. The frequency of the sample is calculated by combining the bandpass filter and the envelope harmonic rate. The upper limit of the main frequency indicates that samples with the highest frequency above the set threshold may contain high-frequency interference or abrupt changes and should be removed.

[0026] The local energy value is obtained by dividing the signal into several time segments and calculating the root mean square of each segment. If the root mean square of a time segment is greater than a set threshold, it is considered that there are local anomalies such as short-term strong pulses and sudden noise, which is helpful for cleaning up non-steady-state abnormal interference.

[0027] Kurtosis is a measure of the steepness of a signal's distribution, calculated by kurtosis, and is used to remove samples containing sharp abrupt changes or anomalous pulses.

[0028] Feature cleaning using the five indicators mentioned above can effectively improve data quality, which is beneficial for improving the accuracy and robustness of the training model.

[0029] In a preferred embodiment, in step S3, the samples retained after data cleaning are subjected to feature extraction to obtain features that can characterize and distinguish normal sound from abnormal collision sound; this patent calculates a feature vector of the sample composed of ten feature values, including roughness, mean square value, root mean square value, peak-to-peak value, impulse index, kurtosis index, envelope root mean square value, low-frequency root mean square value, mid-frequency root mean square value, and high-frequency root mean square value;

[0030] Roughness represents the ratio between the peak value and the root mean square energy value in a signal. The larger the value, the more obvious the instantaneous impact sound in the sound signal, which is suitable for the identification of abnormal collision sounds.

[0031] The mean square value represents the overall energy level of the signal, reflecting the overall intensity of the sound;

[0032] The root mean square value represents the overall amplitude of the signal in the energy dimension and can be used to determine whether the signal is in an effective working state.

[0033] Peak-to-peak value indicates the range of maximum amplitude of a sound signal, reflecting large instantaneous changes in sound.

[0034] The impulse index refers to the average of the ratios of the maximum absolute value of the amplitude to the absolute values ​​of all amplitudes in a signal. It can measure whether there are instantaneous impulse components in a sound signal and can reflect whether the sound comes from a "sudden event".

[0035] The kurtosis index is the ratio of the fourth power of the deviation of a signal point from the mean to the standard deviation. It characterizes the "sharpness" of the sound signal distribution and is very sensitive to shocks and outliers in the sound signal.

[0036] The root mean square value of the envelope refers to the root mean square value of the signal envelope after demodulation. It can extract the modulation information and has a significant effect in detecting structurally abnormal sounds.

[0037] The root mean square value of low frequency refers to the root mean square value of the low frequency band after the signal passes through the bandpass filter. It reflects the baseband or low frequency energy of the signal and can be used to determine background sound or low frequency signal.

[0038] The root mean square value of the intermediate frequency (RMS) refers to the root mean square value of the signal in the intermediate frequency band after passing through a bandpass filter. It is used to distinguish between normal and abnormal sounds.

[0039] The root mean square value of a high-frequency signal is the root mean square value of the high-frequency band after the signal passes through a bandpass filter. It is used to detect short-time high-frequency abnormal characteristics of high-frequency impact or collision sounds.

[0040] By calculating the feature vector composed of the above ten features, it can cover multiple dimensions such as energy, pulse, and frequency band energy distribution, and has high sensitivity to abnormal sounds, making it robust for recognition in environments with complex background noise.

[0041] In a preferred embodiment, in step S4, after obtaining the feature vectors of all samples, the mean and variance of each feature are calculated, and the features are normalized using the following formula:

[0042]

[0043] in, The feature matrix represents the feature vectors of all samples. and These represent the mean and variance of each column in the eigenvector, respectively.

[0044] In a preferred embodiment, in step S5, in order to collect sufficient training data, the gap between the pipe wheel and the tank channel is widened to collect the sound of the operation process. The abnormal collision sound and the normal operation sound are separated using an audio editing tool to form negative samples and positive samples respectively. In order to train a more accurate model, the positive samples should include as many normal operation sounds as possible, such as wind noise, sounds generated by other equipment, and human voices. After obtaining the data, the data is preprocessed and features are extracted according to the above process. Then, the model part is trained and saved to obtain the trained abnormal collision sound recognition model.

[0045] In a preferred embodiment, in step S5, the trained model is deployed to the target application scenario, and abnormal sounds are identified based on the data collected in real time by the microphone. If an abnormal collision is detected, the audio of 7 seconds, from 3 seconds before to 3 seconds after the collision, is saved, and the alarm information and audio file are fed back to the staff.

[0046] In a preferred embodiment, in step S6, data is collected by increasing the gap between the pipe wheel and the guide in a coal mine shaft scenario. The audio is edited to obtain 6138 seconds of normal sound samples, including wind noise, normal equipment operation sounds, human voices, and equipment maintenance sounds, and 15 seconds of abnormal collision sound samples. An abnormal collision sound recognition model is obtained through data preprocessing, feature extraction, and model training.

[0047] In a preferred embodiment, in step S6, in order to test the performance of the model, data was collected under normal operating conditions of the scenario, totaling 21 audio files containing both normal and abnormal sounds, including 3709 seconds of normal audio and 33 seconds of abnormal audio; the trained abnormal collision sound recognition model was used to identify the data.

[0048] In a preferred embodiment, in step S4, binary classification of normal and abnormal sound signals is achieved using a support vector data description method based on the dimensionality-reduced features. Let the set of dimensionality-reduced feature samples be... This method aims to find a minimum sphere in the feature map space, with its center at . , radius is The goal is to include as many normal sounds as possible; the objective function for optimization is:

[0049]

[0050]

[0051] in, This represents the feature map corresponding to the kernel function. The penalty coefficient for slack variables, As slack variables, some samples are allowed to be outside the sphere; this is achieved by introducing Lagrange multipliers and kernel functions. The optimization problem is transformed into a dual form:

[0052]

[0053]

[0054] The final discriminant function judges a new sample. Whether to perform binary classification on normal sound samples within a sphere:

[0055]

[0056] like ,but If it is a normal sound, then it is an abnormal sound.

[0057] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0058] 1. In this invention, a non-contact acoustic acquisition device is used to capture the operating sound of the shaft in real time. Combined with artificial intelligence technology, a recognition model is constructed to achieve continuous online analysis and real-time alarms. The acoustic signal is sensitive to instantaneous abnormal collisions, capturing the high-frequency impact sound at the moment of contact between the shaft lug and the shaft, effectively compensating for the shortcomings of manual inspection. Simultaneously, the AI ​​model, through self-learning optimization, gradually adapts to noise interference in different mine environments, significantly reducing the false alarm rate, decreasing the frequency of manual intervention, and promoting the intelligent transformation of shaft inspection.

[0059] 2. In this invention, a multi-index data cleaning process is designed to address complex interferences such as wind noise and equipment vibration in the wellbore environment. For example, low-energy background noise is filtered using the root mean square energy threshold, high-frequency abrupt changes are eliminated using the dominant frequency range, and abnormal pulse signals are identified by combining the kurtosis index. In the feature extraction stage, ten acoustic features, including roughness, pulse index, and frequency band energy distribution, are integrated, covering the time domain, frequency domain, and energy dimensions to enhance the model's sensitivity to abnormal collisions. This multi-dimensional feature fusion strategy effectively distinguishes between normal friction sound and abnormal collision sound, improves the model's robustness in complex acoustic environments, and ensures high accuracy of the detection results.

[0060] 3. In this invention, kernel principal component analysis (KPCA) is used to perform nonlinear dimensionality reduction on high-dimensional features, extracting implicit acoustic pattern features and avoiding information redundancy. Subsequently, a minimum hypersphere boundary is constructed based on support vector data description (SVDD) to enclose normal sound samples, while abnormal samples are distributed outside the boundary. This method requires only a small number of normal samples for training, without relying on a large amount of abnormal data, perfectly adapting to the challenge of scarce abnormal samples in real-world scenarios. Simultaneously, by adjusting the slack variable parameters, the model's tolerance to noise is flexibly controlled, further improving generalization ability.

[0061] 4. In this invention, the acoustic sensor can collect sound signals without contacting the mine shaft equipment, avoiding physical modifications to the shaft structure. The sensor can be flexibly deployed near key locations in the mine shaft, adapting to the spatial constraints of different mines. The system adopts an edge computing architecture, with the model deployed on local devices for real-time processing, reducing data transmission pressure and cloud dependency. Furthermore, the modular design supports rapid integration into existing monitoring platforms, is compatible with various hardware devices, and significantly reduces deployment and maintenance costs. For example, a coal mine completed system deployment within two weeks by simply installing microphones and an edge server. Within six months of operation, the accuracy rate of abnormal collision recognition remained stable at over 99%, verifying the engineering practicality and scalability of the solution. Attached Figure Description

[0062] Figure 1 This is an overall flowchart of the present invention;

[0063] Figure 2 This is a diagram showing the effect of training the model in this invention;

[0064] Figure 3 This is a mixture matrix diagram of the test results in this invention. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0066] Example:

[0067] Reference Figure 1-3 This invention discloses a collision monitoring method for well shaft guideways based on artificial intelligence acoustic recognition. The invention addresses the problem of collisions between the guideway and the guideway lugs caused by friction between the external rubber of the guideway impeller during operation. It utilizes sound acquisition equipment to capture the sounds of the operation process and constructs an abnormal collision sound recognition model to detect abnormal collisions between the guideway and the guideway lugs. Figure 1As shown, the collision sound recognition model first installs sound acquisition equipment at key locations in the well shaft and tank passage to collect sound during operation in real time; it reads the sound file according to the original sampling rate to obtain the audio signal, and performs frame processing on the audio signal according to the set frame length and frame shift, dividing the continuous time domain signal into multiple overlapping short time frames, which facilitates the subsequent extraction and analysis of local time domain feature information; the frame data is cleaned according to multi-dimensional feature indicators, including frame energy, frequency characteristics, local frame energy, frequency domain energy ratio, and kurtosis, etc. Through comprehensive judgment of multi-dimensional feature indicators, abnormal frames with insufficient energy, abnormal spectrum, or abrupt noise are screened out to improve the signal-to-noise ratio and quality of the audio signal and ensure the stability and accuracy of subsequent feature extraction and modeling;

[0068] Key features in audio files are calculated based on audio continuity and psychoacoustic indices to comprehensively characterize the acoustic patterns generated during well collision, such as loudness, sharpness, roughness, and wave intensity, forming feature vectors. The feature vectors are then normalized, and the original high-dimensional features are mapped to a low-dimensional feature space using kernel principal component analysis to preserve sample structure information while improving the interpretability of the model.

[0069] After feature dimensionality reduction, an abnormal collision sound recognition model is constructed using a kernel-based support vector data description method. By constructing a closed hypersphere with a minimum radius in the feature space to include the data samples of normal sound, binary classification is achieved by dividing the boundary between normal and abnormal sound samples.

[0070] The model is deployed in the target application scenario. First, the wheel spacing of the can guide is increased to cause collision between the can guide and the can ear. Sufficient audio data is collected using a sound acquisition device. The normal operation sound and abnormal collision sound in the collected sound are separated using an audio editing tool and divided into positive and negative samples for binary classification. The data is processed according to the above process and the abnormal collision sound recognition model is trained. The model is then deployed in the target application scenario and recognizes abnormal collisions based on the audio data continuously collected by the sound acquisition device. This enables real-time alarms for abnormal collisions.

[0071] Specifically, it includes the following steps:

[0072] (a) Sound recording during operation

[0073] A microphone is installed near key points of the wellbore tank conveyor equipment to continuously collect the sound of the tank conveyor operation. The sound files are saved according to a certain duration, and the sound files are named by the equipment number and the time of the first second of the audio to indicate the time when the sound occurred.

[0074] (ii) Data reading and preprocessing

[0075] The audio file collected by the microphone is read at the original sampling rate to ensure that the accuracy on the time axis is not affected by resampling, and the frequency, phase and amplitude information are completely preserved. The continuous amplitude value that changes over time is obtained as the audio signal to reflect the sound pressure change caused by the collision of the can ears in the well shaft.

[0076] The obtained audio signal is processed by frame segmentation, dividing the audio signal into frame segments according to the set frame length and frame shift. This process is implemented using a sliding window method. Each frame segment is a short audio signal that represents the local dynamic characteristics of the sound without disturbing or losing temporal resolution.

[0077] By reading audio and processing it in frames, each frame segment is treated as a sample audio signal for data cleaning to improve the signal-to-noise ratio and stability of the data. This patent selects five data cleaning metrics: root mean square energy value, lower limit of dominant frequency, upper limit of dominant frequency, local energy value, and kurtosis value.

[0078] The root mean square energy value measures the total energy level of each sample audio signal by calculating the root mean square. When the root mean square of a sample is less than a set threshold, it is considered to have too low energy, and is a silent, no signal, or background noise segment, and is therefore discarded.

[0079] The lower limit of the main frequency indicates that samples with the highest frequency below the set threshold are considered inactive or invalid and should be removed. The frequency of the sample is calculated by combining the bandpass filter and the envelope harmonic rate. The upper limit of the main frequency indicates that samples with the highest frequency above the set threshold may contain high-frequency interference or abrupt changes and should be removed.

[0080] The local energy value is obtained by dividing the signal into several time segments and calculating the root mean square of each segment. If the root mean square of a time segment is greater than a set threshold, it is considered that there are local anomalies such as short-term strong pulses and sudden noise, which is beneficial for cleaning up non-steady-state abnormal interference.

[0081] Kurtosis is a measure of the steepness of a signal's distribution, calculated by kurtosis, and is used to remove samples containing sharp abrupt changes or anomalous pulses.

[0082] Feature cleaning using the five indicators mentioned above can effectively improve data quality, which is beneficial for improving the accuracy and robustness of the training model.

[0083] (III) Feature Extraction

[0084] Feature extraction is performed on the samples retained after data cleaning to obtain features that can characterize and distinguish normal sounds from abnormal collision sounds. This patent calculates a feature vector of the sample composed of ten feature values, including roughness, mean square value, root mean square value, peak-to-peak value, impulse index, kurtosis index, envelope root mean square value, low-frequency root mean square value, mid-frequency root mean square value, and high-frequency root mean square value.

[0085] Roughness represents the ratio between the peak value and the root mean square energy value in a signal. The larger the value, the more obvious the instantaneous impact sound in the sound signal, which is suitable for the identification of abnormal collision sounds.

[0086] The mean square value represents the overall energy level of the signal, reflecting the overall intensity of the sound.

[0087] The root mean square (RMS) value represents the overall amplitude of a signal in the energy dimension and can be used to determine whether a signal is in a valid operating state.

[0088] Peak-to-peak value indicates the range of maximum amplitude of a sound signal, reflecting large instantaneous changes in sound.

[0089] The impulse index refers to the average of the ratios of the maximum absolute value of the amplitude to the absolute values ​​of all amplitudes in a signal. It can measure whether there are instantaneous impulse components in a sound signal and can reflect whether the sound comes from a "sudden event".

[0090] The kurtosis index is the ratio of the fourth power of the deviation of a signal point from its mean to its standard deviation. It characterizes the "sharpness" of the sound signal distribution and is very sensitive to shocks and outliers in the sound signal.

[0091] The root mean square value of the envelope refers to the root mean square value of the signal envelope after demodulation. It can extract the modulation information and has a significant effect in detecting structurally abnormal sounds.

[0092] The low-frequency root mean square value refers to the root mean square value of the low-frequency band after the signal passes through the bandpass filter. It reflects the baseband or low-frequency energy of the signal and can be used to determine background sound or low-frequency signals.

[0093] The root mean square value of the intermediate frequency (RMS) refers to the root mean square value of the signal in the intermediate frequency band after passing through a bandpass filter. It is used to distinguish between normal and abnormal sounds.

[0094] The root mean square value of a high-frequency signal is the root mean square value of the high-frequency band after the signal passes through a bandpass filter. It is used to detect short-term high-frequency abnormal characteristics of high-frequency impact or collision sounds.

[0095] By calculating the feature vector composed of the above ten features, it can cover multiple dimensions such as energy, pulse, and frequency band energy distribution, and has high sensitivity to abnormal sounds, making it robust for recognition in environments with complex background noise.

[0096] (iv) Model Section

[0097] After obtaining the feature vectors of all samples, calculate the mean and variance of each feature, and then normalize the features using the following formula:

[0098]

[0099] in, The feature matrix represents the feature vectors of all samples. and These represent the mean and variance of each column in the eigenvector, respectively.

[0100] To further extract effective low-dimensional recognition features and achieve separation of normal and abnormal signals in sound signals, kernel principal component analysis (KPCA) is used to perform unsupervised feature compression on the normalized features. Compared with traditional KPCA, this method can extract linear components in high-dimensional implicit space and features with non-linear relationships. Let the sample... Each sample This indicates that there are d=10 features. First, a linear kernel function is calculated, and the kernel matrix is ​​constructed using the linear kernel function. :

[0101]

[0102] in, This indicates the calculation of the inner product of two vectors; then, the kernel matrix is ​​centered, as follows:

[0103]

[0104] in, for The whole Matrix; finally, the centered kernel matrix Perform eigenvalue decomposition and retain the first eigenvalue. One principal component, i.e., solving:

[0105]

[0106] in, and Representing the eigenvalues ​​and eigenvectors respectively, we finally obtain the dimensionality-reduced features. .

[0107] Based on the dimensionality-reduced features, a support vector data description method is used to achieve binary classification of normal and abnormal sound signals. Let the set of dimensionality-reduced feature samples be . This method aims to find a minimum sphere in the feature map space, with its center at . , radius is The goal is to include as many normal sounds as possible. The objective function for optimization is:

[0108]

[0109]

[0110] in, This represents the feature map corresponding to the kernel function. The penalty coefficient for slack variables, As slack variables, some samples are allowed to be outside the sphere. This is achieved by introducing Lagrange multipliers and a kernel function. The optimization problem is transformed into a dual form:

[0111]

[0112]

[0113] The final discriminant function judges a new sample. Whether to perform binary classification on normal sound samples within a sphere:

[0114]

[0115] like ,but If it is a normal sound, then it is an abnormal sound.

[0116] (v) Model training and deployment

[0117] To collect sufficient training data, the gap between the pipe wheel and the tank guide was widened to collect the sounds during operation. Audio editing tools were used to separate abnormal collision sounds from normal operating sounds, creating negative and positive samples respectively. To train a more accurate model, the positive samples should include as many normal operating sounds as possible, such as wind noise, sounds from other equipment, and human voices. After obtaining the data, preprocessing and feature extraction were performed according to the above process. The model was then trained and saved, resulting in a trained abnormal collision sound recognition model.

[0118] The trained model is deployed to the target application scenario. It identifies abnormal sounds based on the data collected in real time by the microphone. If an abnormal collision is detected, it saves 7 seconds of audio from 3 seconds before to 3 seconds after the collision and sends the alarm information and audio file back to the staff.

[0119] (vi) Demonstration of the model's test performance

[0120] In a coal mine shaft scenario, data was collected by increasing the gap between the pipe wheel and the guide rail. Audio was edited to obtain 6138 seconds of normal sound samples, including wind noise, normal equipment operation sounds, human voices, and equipment maintenance sounds, as well as 15 seconds of abnormal collision sound samples. An abnormal collision sound recognition model was obtained through data preprocessing, feature extraction, and model training. Figure 2As shown, the trained model separates positive and negative samples, obtaining a boundary for positive samples. Samples inside the boundary are normal sound samples, while samples outside the boundary are abnormal sound samples.

[0121] To test the model's performance, data was collected under normal operating conditions in this scenario, totaling 21 audio files containing both normal and abnormal sounds, comprising 3709 seconds of normal audio and 33 seconds of abnormal audio. The trained abnormal collision sound recognition model was then used to identify these sounds. Figure 3 As shown in the mixed-sound matrix, there was only one second of false alarm where a normal sound was predicted as an abnormal sound, while the other 3741 seconds were predicted correctly, achieving an accuracy rate of 99.97%, which is sufficient for use in this scenario.

[0122] This invention provides a collision monitoring method for wellbore tank passages based on artificial intelligence acoustic recognition. First, the operating sounds of the tank passage devices in the wellbore are collected in real time using acoustic acquisition devices. Samples are obtained through reading and frame segmentation, and data processing is performed based on five indicators, including root mean square energy value and lower limit of dominant frequency. Second, ten features, including roughness, mean square value, and root mean square value, are calculated. Then, kernel principal component analysis is used to reduce the dimensionality of the features, and binary classification based on support vector data description method is used to identify normal and abnormal sounds. Finally, the model is trained and deployed for application in real-world scenarios.

[0123] Based on the specific production process of a certain mine, this invention has the following advantages:

[0124] 1) High real-time performance and rapid response. By continuously collecting pipeline operation sounds in real time through acoustic acquisition devices, and quickly performing sound data cleaning, feature processing, and model recognition, it can quickly issue alarms for abnormal collisions in a short period of time, effectively improving the safety of well operation and reducing the number and cost of manual inspections.

[0125] 2) High recognition accuracy and strong robustness. By cleaning the original sound signal with multiple indicators, background noise and invalid signals are effectively filtered out. Furthermore, by calculating multiple acoustic features, the feature vector is enhanced to improve the ability of the feature vector to represent and distinguish the sound of the tank collision, thereby improving the accuracy of the trained model in complex well environments.

[0126] 3) The model exhibits strong generalization ability and good adaptability. It distinguishes between normal and abnormal acoustic patterns using support vector data description, eliminating the need for extensive training with a large number of abnormal samples. This makes it suitable for real-world wellbore scenarios where abnormal collision samples are simple while normal samples are complex and diverse. It is applicable to collision monitoring in different wellbore structures, operating equipment, and geological environments, demonstrating excellent environmental adaptability and scalability.

[0127] 4) Easy deployment and low maintenance cost. Acoustic sensors are easy to deploy, the system structure is simple, and it can be deployed in existing tank line operation systems without large-scale hardware modifications. After training, the model can run on edge devices to achieve local real-time monitoring and early warning, reducing data transmission pressure and subsequent operation and maintenance costs.

[0128] In summary, this invention improves the safety of wellbore tank passage operation while taking into account identification accuracy, model adaptability, and engineering practicality, and has broad application value and promotion prospects.

[0129] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting collisions in well shafts and tank passages based on artificial intelligence acoustic recognition, characterized in that: The method includes the following steps: S1: Collect sound during operation; S2: Perform data reading and preprocessing; S3: Perform feature extraction; S4: Construct a model for monitoring collisions in well shafts and tank passages; S5: Train and deploy the model; S6: Demonstrate the model's test performance; In step S4, kernel principal component analysis is used to perform unsupervised feature compression on the normalized features, which can extract linear components in the high-dimensional implicit space and features with non-linear relationships; let the sample Each sample This indicates that there are d=10 features; first, a linear kernel function is calculated, and the kernel matrix is ​​constructed using the linear kernel function. : ; in, This indicates the calculation of the inner product of two vectors; then, the kernel matrix is ​​centered, as follows: , in, for The whole Matrix; finally, the centered kernel matrix Perform eigenvalue decomposition and retain the first eigenvalue. One principal component, i.e., solving: , in, and Representing the eigenvalues ​​and eigenvectors respectively, we finally obtain the dimensionality-reduced features. ; In step S4, after obtaining the feature vectors of all samples, the mean and variance of each feature are calculated, and the features are normalized using the following formula: , in, The feature matrix represents the feature vectors of all samples. and These represent the calculation of the mean and variance of each column in the eigenvector, respectively. In step S5, in order to collect sufficient training data, the gap between the pipe wheel and the tank is widened to collect the sound of the operation process. The abnormal collision sound and the normal operation sound are separated using an audio editing tool to form negative samples and positive samples respectively. In order to train a more accurate model, the positive samples should include as many normal operation sounds as possible. After obtaining the data, the data is preprocessed and features are extracted according to the above process. Then the model part is trained and saved to obtain the trained abnormal collision sound recognition model. In step S5, the trained model is deployed to the target application scenario, and abnormal sounds are identified based on the data collected in real time by the microphone. If an abnormal collision is detected, the audio of 7 seconds, from 3 seconds before the collision to 3 seconds after the collision, is saved, and the alarm information and audio file are fed back to the staff. In step S6, data is collected by increasing the gap between the pipe wheel and the guide in a coal mine shaft scenario. The audio is edited to obtain 6138 seconds of normal sound samples including wind noise, normal equipment operation sound, human voice, and equipment maintenance sound, and 15 seconds of abnormal collision sound samples. An abnormal collision sound recognition model is obtained through data preprocessing, feature extraction, and model training. In step S6, in order to test the performance of the model, data was collected under normal operating conditions in the actual operation scenario of the well shaft and tank passage. A total of 21 audio files containing both normal and abnormal sounds were collected, including 3709 seconds of normal audio and 33 seconds of abnormal audio. The trained abnormal collision sound recognition model was used to identify the data. In step S4, based on the dimensionality-reduced features, a support vector data description method is used to achieve binary classification of normal and abnormal sound signals. Let the set of dimensionality-reduced feature samples be... This method aims to find a minimum sphere in the feature map space, with its center at . , radius is The goal is to include as many normal sounds as possible; the objective function for optimization is: , in, This represents the feature map corresponding to the kernel function. The penalty coefficient for slack variables, As slack variables, some samples are allowed to be outside the sphere; this is achieved by introducing Lagrange multipliers and kernel functions. The optimization problem is transformed into a dual form: ; The final discriminant function judges a new sample. Whether to perform binary classification on normal sound samples within a sphere: ; like ,but If it is a normal sound, then it is an abnormal sound. In step S2, the audio file collected by the microphone is read at the original sampling rate to ensure that the accuracy on the time axis is not affected by resampling, and the frequency, phase and amplitude information are completely preserved. The continuous amplitude value that changes with time is obtained as the audio signal to reflect the sound pressure change generated by the collision of the can ears in the well shaft. The obtained audio signal is processed by frame segmentation, which divides the audio signal into frame segments according to the set frame length and frame shift. This process is implemented by sliding window. Each frame segment is a short audio signal that represents the local dynamic characteristics of the sound and does not disturb or lose temporal resolution. By reading audio and processing it in frames, each frame segment is treated as a sample audio signal for data cleaning to improve the signal-to-noise ratio and stability of the data. Five data cleaning metrics are selected: root mean square energy value, lower limit of dominant frequency, upper limit of dominant frequency, local energy value, and kurtosis value. The root mean square energy value measures the total energy level of each sample audio signal by calculating the root mean square. When the root mean square of a sample is less than a set threshold, it is considered to have too low energy, and is a silent, no signal, or background noise segment, and is therefore removed. The lower limit of the main frequency indicates that samples with the highest frequency below the set threshold are considered inactive or invalid and should be removed. The frequency of the sample is calculated by combining the bandpass filter and the envelope harmonic rate. The upper limit of the main frequency indicates that samples with the highest frequency above the set threshold may contain high-frequency interference or abrupt changes and should be removed. The local energy value is obtained by dividing the signal into several time segments and calculating the root mean square of each segment. If the root mean square of a time segment is greater than a set threshold, it is considered that there is a short-term strong pulse or sudden noise local anomaly, which is beneficial for cleaning up non-steady-state abnormal interference. Kurtosis is a measure of the steepness of a signal's distribution, calculated by kurtosis, and is used to remove samples containing sharp abrupt changes or anomalous pulses. Feature cleaning using the above five indicators effectively improves data quality, which is beneficial for improving the accuracy and robustness of the training model.

2. The method for detecting collisions in well shafts and tank passages based on artificial intelligence acoustic recognition as described in claim 1, characterized in that: In step S1, a microphone is used as a sound acquisition device and installed near key points of the well shaft tank passage equipment to continuously collect the sound of the tank passage operation process. The sound file is saved according to a certain duration, and the sound file is named by the equipment number and the time of the first second of the audio to indicate the time when the sound occurs.

3. The method for detecting collisions in well shafts and tank passages based on artificial intelligence acoustic recognition as described in claim 1, characterized in that: In step S3, after data cleaning, the remaining samples are subjected to feature extraction to obtain features that can characterize and distinguish normal sounds from abnormal collision sounds. Calculate the feature vector of the sample composed of ten feature values, including roughness, mean square value, root mean square value, peak-to-peak value, impulse index, kurtosis index, envelope root mean square value, low-frequency root mean square value, mid-frequency root mean square value, and high-frequency root mean square value. Roughness represents the ratio between the peak value and the root mean square energy value in a signal. The larger the value, the more obvious the instantaneous impact sound in the sound signal, which is suitable for the identification of abnormal collision sounds. The mean square value represents the overall energy level of the signal, reflecting the overall intensity of the sound; The root mean square value represents the overall amplitude of the signal in the energy dimension and can be used to determine whether the signal is in an effective working state. Peak-to-peak value represents the range of the maximum amplitude of a sound signal, reflecting the instantaneous and significant changes in sound. The impulse index refers to the average of the ratios of the maximum absolute value of the amplitude to the absolute values ​​of all amplitudes in a signal. It measures whether there are instantaneous impulse components in a sound signal and can reflect whether the sound comes from a "sudden event". The kurtosis index is the ratio of the fourth power of the deviation of a signal point from the mean to the standard deviation. It characterizes the "sharpness" of the sound signal distribution and is very sensitive to shocks and outliers in the sound signal. The root mean square value of the envelope refers to the root mean square value of the signal envelope after demodulation. The root mean square value of low frequency refers to the root mean square value of the low frequency band after the signal passes through the bandpass filter. It reflects the baseband or low frequency energy of the signal and can be used to determine background sound or low frequency signals. The root mean square value of the intermediate frequency (RMS) refers to the root mean square value of the signal in the intermediate frequency band after passing through a bandpass filter. It is used to distinguish between normal and abnormal sounds. The high-frequency root mean square value refers to the root mean square value of the high-frequency band after the signal passes through the bandpass filter. It is used to detect short-time high-frequency abnormal characteristics of high-frequency impact or collision sounds. By calculating the feature vector composed of the above ten features, it can cover multiple dimensions such as energy, pulse, and frequency band energy distribution, and has high sensitivity to abnormal sounds, making it robust for recognition in environments with complex background noise.

Images