A method for detecting a tankway anomaly based on audio signal analysis
By setting up sensors around the tank passage to collect audio signals, preprocessing them, and combining incremental learning and long short-term memory networks to establish a dynamic feature change model, the problem that existing audio signal analysis methods cannot adapt to complex equipment changes is solved, and real-time and accurate anomaly detection and classification of the tank passage system is realized.
Patent Information
- Application Number
- CN202511071911.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-01
- Publication Date
- 2025-11-21
- Estimated Expiration
- 2045-08-01
AI Technical Summary
Existing audio signal analysis methods cannot adapt to the complex changes under different operating conditions of equipment, and are difficult to effectively handle noise interference and temporal characteristic changes, resulting in low anomaly detection accuracy, lack of automatic classification mechanism, and delay in handling equipment faults.
By setting up sensors around the tank passage to collect audio signals in real time, preprocessing them to remove noise, and combining incremental learning and long short-term memory networks to establish a dynamic feature change model, the temporal variation pattern of the audio signals is analyzed, and abnormal signal types are identified through a classification model.
It enables real-time and accurate anomaly detection of the tank conveyor system, can identify various types of abnormal signals, improves the accuracy and adaptability of detection, reduces human intervention, and enhances the stability and safety of the system.
Smart Images
Figure CN120600049B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of intelligent detection, and particularly relates to a tank abnormality detection method based on audio signal analysis. BACKGROUND
[0002] With the continuous advancement of industrialization, tank equipment plays a crucial role in many industrial productions, especially in the fields of chemical industry, energy, metallurgy, etc. In the long running process of tank equipment, abnormal situations may occur due to equipment aging, environmental changes or improper operation. These abnormalities are often difficult to be warned in advance through traditional monitoring means. Traditional fault detection methods usually rely on physical indicators or offline diagnostic means of equipment, and it is difficult to monitor the running state of the equipment in real time and comprehensively. In addition, the accuracy of many traditional detection methods is low, which cannot effectively distinguish different types of abnormalities, and can only identify fault signals, and it is difficult to provide more detailed abnormal classification and timely feedback.
[0003] In the prior art, audio signals, as an important reflection of the running state of equipment, have potential value for abnormality detection. By analyzing the audio signals of the equipment, the running characteristics of the equipment can be identified, and whether the equipment is in a normal working state can be monitored. However, the existing audio signal analysis method usually adopts a fixed signal processing model, which cannot adapt to the complex changes under different running states of the equipment. Moreover, the traditional method relies on simple signal comparison, which is difficult to effectively process noise interference and complex time sequence characteristic changes, resulting in low accuracy of abnormality detection. In addition, the prior art lacks an automatic classification mechanism for abnormal signals, which makes it difficult to accurately identify the type and cause of the abnormality, resulting in insufficient pertinence of subsequent processing measures and delay of the processing time of equipment failure. SUMMARY
[0004] Based on the above purpose, the present application provides a tank abnormality detection method based on audio signal analysis.
[0005] A tank abnormality detection method based on audio signal analysis, comprising the following steps:
[0006] S1: Real-time acquisition of audio signals of the equipment under different running states through a plurality of sensors arranged around the tank;
[0007] S2: Preprocessing the collected audio signals to remove noise and interference signals in the signals, and extracting audio features related to the running state of the tank through a high-pass filtering signal processing method;
[0008] S3: Based on the audio features extracted in S2, a dynamic feature change model is established, which adopts a combination of incremental learning and long short-term memory network to analyze the frequency and amplitude features of the audio signal changing over time, and establish the time sequence change rule of the audio signal;
[0009] S4: The dynamic feature change model established in S3 is applied to the analysis of real-time audio signals, and by comparing the features of real-time audio signals with the features of historical normal signals, it is determined whether an abnormality occurs;
[0010] S5: If S4 determines that the audio signal is abnormal, the pre-trained classification model is used to classify and process the abnormal signal.
[0011] Optionally, S1 specifically comprises:
[0012] S11: Audio acquisition sensors are arranged at the positions of the two ends and the middle of the tank channel, the audio acquisition sensors are microphones, which can capture sound signals generated during the operation of the equipment in real time, and convert the collected analog audio signals into digital signals through an analog-digital converter;
[0013] S12: The sampling frequency of the audio acquisition sensor is set to at least 1000 data points per second.
[0014] Optionally, S2 specifically comprises:
[0015] S21: The collected audio signals are subjected to noise filtering, specifically through short-time Fourier transform to analyze the time-frequency of the audio signals, decompose the signals into several frequency bands, and identify and remove low-frequency noise and environmental background noise;
[0016] S22: A high-pass filter is used to attenuate the low-frequency part in the frequency domain, and the cutoff frequency of the high-pass filter is set to 300Hz to remove low-frequency noise and irrelevant signal components, and only signals in the frequency range related to the tank channel operation state are retained;
[0017] S23: The signal processed by the high-pass filter is subjected to time domain smoothing processing, and the amplitude of the audio signal is smoothed by using the sliding average method to reduce the fluctuation caused by instantaneous change;
[0018] S24: The frequency spectrum and time domain features of the preprocessed audio signal are calculated to extract the audio features related to the tank channel operation state, including the frequency spectrum peak, frequency response curve and time domain standard deviation.
[0019] Optionally, S24 specifically comprises:
[0020] S241: The frequency spectrum of the preprocessed audio signal is calculated, and the frequency spectrum of the audio signal is obtained by applying fast Fourier transform to the audio signal;
[0021] S242: Extract the peak value in the spectrum, first calculate the amplitude of the spectrum, then extract the maximum amplitude value and the corresponding frequency in the spectrum by the peak detection algorithm, and record the frequency and amplitude of the spectrum peak value;
[0022] S243: Calculate the frequency response curve, which represents the amplitude response of the audio signal at different frequencies; by drawing the curve of the amplitude of the spectrum varying with frequency, the frequency response curve is obtained;
[0023] S244: Calculate the standard deviation of the audio signal to evaluate the amplitude variation of the signal.
[0024] Optionally, the S3 specifically comprises:
[0025] S31: Based on the audio features extracted in S2, including the spectrum peak value, the frequency response curve and the time domain standard deviation, an audio feature dataset is constructed, wherein each data sample is composed of a timestamp and a corresponding audio feature vector;
[0026] S32: Use an incremental learning algorithm to train the audio feature dataset, and gradually update the model parameters to adapt to new audio feature data;
[0027] S33: Apply the long short-term memory network to the audio feature data to capture the timing changes of the audio signal and learn the rules of the audio signal changing over time;
[0028] S34: Combine incremental learning and long short-term memory network, first train the preliminary data with incremental learning, and then optimize the model using long short-term memory network to accurately capture the timing characteristics of the audio signal;
[0029] S35: Continuously update the long short-term memory network model through incremental learning, so that it can adapt to new audio feature data in real time and capture the dynamic changes of the audio signal;
[0030] S36: Through continuous updating and optimization, the establishment of the dynamic feature change model is completed, ensuring that the model can analyze the dynamic feature changes of the real-time audio signal.
[0031] Optionally, the S32 specifically comprises:
[0032] S321: Input the audio feature dataset constructed in S31 into the incremental learning model and initialize the model parameters, and the objective function used by the incremental learning model is: wherein, is the model parameter, is the loss function, is the actual label, is the model output, Input features, Sample quantity;
[0033] S322: According to the strategy of incremental learning, select the corresponding learning rate and the step of incremental update , and update the model parameters by the following formula: , wherein, is the updated model parameter, is the current model parameter, is the learning rate, is the gradient of the current model parameter;
[0034] S323: In each incremental training process, use the new audio feature data and its corresponding label , incremental update by the following formula: , wherein, represents the current parameter on the gradient of the new input data ;
[0035] S324: By iterating the above incremental learning process, gradually adjust the model parameters , so that the model can adapt to new audio feature data.
[0036] Optionally, the S33 specifically includes:
[0037] S331: The audio feature data updated by the incremental learning in S32 is input into the long short-term memory network to capture the timing changes of the audio signal;
[0038] S332: In the long short-term memory network, the control mechanism of the input gate, the forget gate, the candidate memory unit and the output gate is used to gradually calculate the hidden state and the memory unit state of each time step;
[0039] S333: Through the hidden state sequence of the long short-term memory network, learn the timing change rule of the audio signal over time;
[0040] S334: After training, the final hidden state of the long short-term memory network is taken as the timing feature of the audio signal;
[0041] S335: Through the back propagation algorithm to optimize the weight parameters in the long short-term memory network, so that the network can better fit the timing change rule of the audio signal.
[0042] Optionally, the S36 specifically includes:
[0043] S361: According to the training results of incremental learning and long short-term memory network, continuously update each parameter in the dynamic feature change model;
[0044] S362: By calculating the prediction error during each training process, the model weights are adjusted using an optimization algorithm to minimize the model's prediction error;
[0045] S363: The structure of the dynamic feature change model is gradually improved based on the incremental learning method and the iterative updates of the long short-term memory network;
[0046] S364: When the model error is lower than the preset threshold, it indicates that the dynamic feature change model has been optimized.
[0047] Optionally, S4 specifically includes:
[0048] S4 specifically includes:
[0049] S41: Input the audio features of the real-time acquired audio signal after processing in step S2 into the dynamic feature change model to obtain the dynamic features of the real-time audio signal.
[0050] S42: Calculate the difference between the dynamic characteristics of the real-time audio signal and the dynamic characteristics of the historical normal signal, wherein the difference is measured by calculating the Euclidean distance, using the following formula:
[0051] ,in, Indicates a measure of difference. The first part representing the real-time audio signal 1 eigenvalue, The first normal audio signal in history 1 eigenvalue, Dimensions representing audio features;
[0052] S43: Based on the calculated difference measure Compare with the set threshold, if If the value is greater than the threshold, it is considered an abnormal signal.
[0053] Optionally, S5 specifically includes:
[0054] S51: If S4 determines that the audio signal is abnormal, then the feature vector of the abnormal signal is input into the pre-trained classification model for classification.
[0055] S52: The classification model receives the abnormal signal features from the input and outputs a category label based on the patterns learned during training. This label indicates the type of abnormal signal; specifically when This indicates a device malfunction. Indicates external interference. This indicates an operational error.
[0056] S53: Different processing strategies are adopted according to the category label obtained in S52, for the abnormal signal of the equipment failure type, a maintenance notice will be sent; for the abnormal signal of the external interference type, signal enhancement or shielding will be performed; for the abnormal signal of the operation abnormality type, real-time prompts or warnings will be sent to the operator, and the existing operation errors are prompted.
[0057] Advantages of the present application:
[0058] The present application can capture the timing changes of audio signals in real time, recognize and classify various types of abnormal signals by combining dynamic feature change models and incremental learning techniques. Compared with traditional abnormal detection methods, the present application has higher accuracy and adaptability, can effectively eliminate the interference of environmental noise, and accurately reflects the real running state of the tank channel system.
[0059] The present application uses an incremental learning algorithm to continuously update the model and adapt to new operating states and audio features. This adaptive feature enables the system to maintain efficient abnormal detection capabilities in complex and variable environments, reduces the need for human intervention, and improves the stability and safety of the tank channel system. BRIEF DESCRIPTION OF DRAWINGS
[0060] In order to more clearly illustrate the technical solutions in the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only a part of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.
[0061] Fig. 1 The figure is a schematic diagram of the tank abnormality detection method of the embodiment of the present application.
[0062] Fig. 2 The figure is a schematic diagram of the process of establishing a dynamic feature change model of the embodiment of the present application. DETAILED DESCRIPTION
[0063] The present application will be described in detail below in combination with the drawings and specific embodiments. It should be noted that in order to make the embodiments more detailed, the following embodiments are the best, preferred embodiments, and other alternative ways can also be used by those skilled in the art to implement them; and the drawings are only used to more specifically describe the embodiments, and are not intended to specifically limit the present application.
[0064] It is noted that the recitations "one embodiment," "an embodiment,” “one example embodiment,” “some embodiments,” etc. indicate the described features, structures, or characteristics can be included in one or more embodiments, but each as such an embodiment can not necessarily include that particular feature, structure, or characteristic. Additionally, means-plus-function or step-plus-function clauses can be used in describing features, structures, or characteristics of one or more embodiments. The
[0065] Generally, the terminology can be understood at least in part from usage in context. For example, the term “one or more” as used herein, depending at least in part upon context, can be used to describe any feature, structure, or characteristic in a singular sense or can be used to describe combinations of features, structures or characteristics in a plural sense. Additionally, the term, “based on,” can be understood as not necessarily of a
[0066] As shown in Figs. 1-2 FIG. 1, a tank way anomaly detection method based on audio signal analysis includes the following steps:
[0067] S1: Real-time acquisition of audio signals of the device under different operating states through multiple sensors arranged around the tank way;
[0068] S2: Preprocessing of the collected audio signals to remove noise and interference signals in the signals, and extracting audio features related to the tank way operating state through a high-pass filtering signal processing method;
[0069] S3: Based on the audio features extracted in S2, a dynamic feature change model is established, which uses a combination of incremental learning and long short-term memory (LSTM) network to analyze the frequency and amplitude characteristics of the audio signal changes over time, and establish the time sequence change rule of the audio signal;
[0070] S4: Applying the dynamic feature change model established in S3 to the analysis of real-time audio signals, and comparing the features of the real-time audio signals with the features of historical normal signals to determine whether an anomaly has occurred;
[0071] S5: If S4 determines that the audio signal is abnormal, a pre-trained classification model is used to classify and process the abnormal signal.
[0072] S1 specifically includes:
[0073] S11: Set up audio acquisition sensors at the positions of both ends and the middle of the tank channel, the audio acquisition sensors are microphones, which can capture the sound signals generated during the operation of the equipment in real time, and convert the collected analog audio signals into digital signals through an analog-digital converter;
[0074] S12: Set the sampling frequency of the audio acquisition sensor to at least 1000 data points per second to ensure sufficient resolution and accuracy for subsequent analysis; provide accurate data support for subsequent anomaly detection.
[0075] S2 specifically includes:
[0076] S21: Noise filtering of the collected audio signal, specifically through short-time Fourier transform (STFT) for time-frequency analysis of the audio signal, decomposing the signal into several frequency bands, identifying and removing low-frequency noise and environmental background noise; the formula of short-time Fourier transform is: Where, is the time of the audio signal, is the window function, is the frequency, is the time variable, is the complex spectrum at time and frequency ;
[0077] S22: Use a high-pass filter to attenuate the low-frequency part in the frequency domain, set the cutoff frequency of the high-pass filter to 300Hz to remove low-frequency noise and irrelevant signal components, and only keep the signals in the frequency range related to the tank channel operation state;
[0078] S23: Time domain smoothing processing of the signal after high-pass filtering, using the sliding average method to smooth the amplitude of the audio signal to reduce the fluctuations caused by instantaneous changes and ensure the stability and accuracy of the signal;
[0079] S24: Extract audio features related to the tank channel operation state by calculating the frequency spectrum and time domain features of the preprocessed audio signal, including spectral peak, frequency response curve and time domain standard deviation, as input features for subsequent modeling; by comprehensively applying time-frequency analysis, noise filtering and high-pass filtering, low-frequency noise and environmental interference signals in the audio signal can be effectively removed, ensuring that the signal retains effective features closely related to the tank channel operation state; This process provides clean and stable signal data for subsequent feature extraction and anomaly detection.
[0080] S24 specifically includes:
[0081] S241: Calculate the frequency spectrum of the pre-processed audio signal, obtain the frequency spectrum by applying the fast Fourier transform to the audio signal , the formula is: , wherein, is the audio signal at time , is the total number of sampling points of the signal, is the frequency, is the amplitude of the signal at frequency ;
[0082] S242: Extract the peak value in the frequency spectrum, first calculate the amplitude of the frequency spectrum, then extract the maximum amplitude value and the corresponding frequency in the frequency spectrum by the peak detection algorithm, and record the frequency and amplitude of the frequency spectrum peak; the formula is: ; , wherein, is the frequency of the frequency spectrum peak, is the amplitude of the frequency spectrum peak, is the amplitude of the frequency spectrum;
[0083] S243: Calculate the frequency response curve, which represents the amplitude response of the audio signal at different frequencies; by drawing the curve of the amplitude of the frequency spectrum with the frequency, the frequency response curve is obtained, the formula is as follows: , wherein, is the value of the frequency response curve at frequency , unit: decibel ;
[0084] S244: Calculate the standard deviation of the audio signal to evaluate the amplitude variation of the signal, the formula is:
[0085] , wherein, is the mean of the audio signal, is the total number of sampling points of the signal, is the standard deviation of the audio signal, reflecting the amplitude variation of the signal; by fast Fourier transform (FFT) to calculate the frequency spectrum of the audio signal, the frequency spectrum peak and the frequency response curve can be effectively extracted, so as to reflect the frequency characteristics of the audio signal.
[0086] S3 specifically includes:
[0087] S31: Based on the audio features extracted in S2, including the frequency spectrum peak, the frequency response curve and the time domain standard deviation, an audio feature dataset is constructed, wherein each data sample is composed of a timestamp and a corresponding audio feature vector, which is used as the input of subsequent modeling;
[0088] S32: Train the audio feature dataset using the incremental learning algorithm, gradually update the model parameters to adapt to new audio feature data;
[0089] S33: Apply long short-term memory network (LSTM) to the audio feature data to capture the timing changes of the audio signal and learn the rules of the audio signal changes over time;
[0090] S34: Combine incremental learning and long short-term memory network, first train the preliminary data using incremental learning, then optimize the model using long short-term memory network to accurately capture the timing characteristics of the audio signal;
[0091] S35: Continuously update the long short-term memory network model through incremental learning to enable it to adapt to new audio feature data in real time and capture dynamic changes in the audio signal;
[0092] S36: Through continuous updating and optimization, the dynamic feature change model is established to ensure that the model can analyze the dynamic feature changes of the real-time audio signal to determine whether there are abnormal changes in the audio signal.
[0093] S32 specifically includes:
[0094] S321: Input the audio feature dataset constructed in S31 into the incremental learning model and initialize the model parameters. The objective function used by the incremental learning model is: where, is the model parameter, is the loss function, is the actual label, is the model output, is the input feature, is the number of samples;
[0095] S322: According to the strategy of incremental learning, select the corresponding learning rate and the step size of incremental update , and update the model parameters by the following formula: where, is the updated model parameter, is the current model parameter, is the learning rate, is the gradient of the current model parameter;
[0096] S323: In each incremental training process, use new audio feature data and its corresponding label to perform incremental update by the following formula: where, indicates the current parameter in the new input data gradient on;
[0097] S324: By iterating through the above incremental learning process, the model parameters are gradually adjusted. This enables the model to adapt to new audio feature data and continuously optimizes its accuracy and generalization ability. By gradually updating the model parameters, it can adapt to new audio data in real time, ensuring the model's adaptability in dynamic environments. By optimizing the model update mechanism, the model's accuracy and generalization ability can be effectively improved, enhancing the system's adaptability to audio signals under different operating conditions, thereby improving the accuracy and stability of tank passage anomaly detection.
[0098] S33 specifically includes:
[0099] S331: Input the audio feature data updated by incremental learning in S32 into the Long Short-Term Memory network to capture the temporal changes of the audio signal;
[0100] S332: In a long short-term memory network, the hidden state and memory unit state at each time step are calculated step by step using the control mechanism of input gate, forget gate, candidate memory unit and output gate;
[0101] S333: Learn the temporal variation of audio signals over time by using the hidden state sequence of the Long Short-Term Memory network;
[0102] S334: After training, the final hidden state of the Long Short-Term Memory network is used as the temporal feature of the audio signal;
[0103] S335: The weight parameters in the Long Short-Term Memory network are optimized through the backpropagation algorithm, enabling the network to better fit the temporal variation of audio signals.
[0104] The steps for using long short-term memory networks to learn patterns of change are as follows:
[0105] First, the incrementally updated audio feature data from S32 is input into the Long Short-Term Memory (LSTM) network to capture the temporal changes of the audio signal. The input to the LSTM model is the time series of the audio feature dataset. ,in, Indicates the first Audio feature vectors at each time step, Indicates the length of the time series;
[0106] Then, in the LSTM network, the hidden state at each time step is calculated based on the following formula. and memory cell state :
[0107] ;
[0108] ;
[0109] ;
[0110] ;
[0111] ;
[0112] ; wherein, is the activation value of the input gate; is the activation value of the forget gate; is the candidate memory cell; is the memory cell state at the current time step; is the activation value of the output gate; is the hidden state at the current time step; is the LSTM weight matrix, is the bias term; denotes the sigmoid function, anhdenotes the hyperbolic tangent activation function;
[0113] Next, the hidden state of the LSTM network is updated as The LSTM network learns the temporal variation of the audio signal over time, and the long-term and short-term dependency information of the audio signal is updated at each time step in the network to capture the potential temporal patterns in the signal; and after the training is completed, the final hidden state of the LSTM network is used as the temporal feature vector of the audio signal, as the input for subsequent model analysis and prediction;
[0114] Finally, the weight parameters of the LSTM network are optimized using the backpropagation algorithm, and the learning ability of the model for the temporal variation of the audio signal is gradually improved by minimizing the loss function; the loss function usually uses mean square error (MSE) or cross-entropy loss function, and the formula is: wherein, is the actual label, is the predicted value of the model, is the number of samples.
[0115] S36 specifically comprises:
[0116] S361: According to the training results of the incremental learning and the long short-term memory network, the parameters in the dynamic feature change model are continuously updated; specifically, when new audio signal feature data is input each time, the model will learn based on the current audio feature vector and the historical audio feature, update the weights and biases in the model, and thus optimize the adaptability of the model to the temporal feature change;
[0117] S362: By calculating the prediction error in each training process, the weights of the model are adjusted using an optimization algorithm (such as gradient descent) to minimize the prediction error of the model; specifically, through the backpropagation algorithm, the model adjusts all network parameters according to the error between the actual observation value and the predicted value, so that the model can more accurately fit the new audio feature data;
[0118] S363: According to the incremental learning method and the iterative update of the long short-term memory network, the structure of the dynamic feature change model is gradually improved, so that it can more stably capture the long-term and short-term changes of the audio signal; after each update, the dynamic feature change model optimizes the model according to the new audio signal, thereby enhancing its dynamic response capability to the running state of the tankway;
[0119] S364: When the error of the model is lower than the preset threshold, it indicates that the dynamic feature change model has completed optimization and can be further applied; the final expression of the dynamic feature change model is: wherein, is the input audio feature vector, representing the audio signal feature at time ; is the updated model parameter at time , which contains all parameters (such as weight matrix, bias, etc.) of the LSTM network; is the model output at time , i.e. the predicted value; the above steps can make the model have stronger adaptability and accuracy by continuously updating and optimizing the parameters of the dynamic feature change model, so that it can cope with audio signal changes under different running states; the model can gradually adapt to new audio feature data through incremental learning, and accurately capture the time sequence features of the audio signal using the powerful time sequence learning ability of the LSTM network; finally, the optimized dynamic feature change model can provide more accurate anomaly detection in practical applications, and improve the robustness and response capability of the system.
[0120] S4 specifically includes:
[0121] S41: input the audio features processed by step S2 from the real-time collected audio signal into the dynamic feature change model to obtain the dynamic features of the real-time audio signal;
[0122] S42: calculate the difference between the dynamic features of the real-time audio signal and the dynamic features of the historical normal signal, which is measured by calculating the Euclidean distance, and the formula is:
[0123] wherein, represents the difference measure, represents the first 1 eigenvalue, The first normal audio signal in history 1 eigenvalue, Dimensions representing audio features;
[0124] S43: Based on the calculated difference measure Compare with the set threshold, if If the signal exceeds the threshold, it is considered an abnormal signal. By applying the dynamic feature change model to real-time audio signal analysis, it is possible to determine whether an anomaly has occurred in real time and accurately. Furthermore, by using Euclidean distance measurement, the difference between the real-time signal and the historical normal signal can be precisely measured, thereby effectively improving the accuracy and response speed of anomaly detection.
[0125] S5 specifically includes:
[0126] S51: If S4 determines that the audio signal is abnormal, then the feature vector of the abnormal signal is input into the pre-trained classification model for classification; the classification model is trained based on a large amount of labeled abnormal signal data and can automatically determine the type of abnormal signal based on the audio features of the abnormal signal (including spectral features, time domain features, etc.).
[0127] S52: The classification model receives the abnormal signal features from the input and outputs a category label based on the patterns learned during training. This label indicates the type of abnormal signal; specifically when This indicates a device malfunction. Indicates external interference. This indicates an operational error.
[0128] S53: Based on the category labels obtained in S52, different processing strategies are adopted. For abnormal signals of equipment failure type, a maintenance notification will be sent; for abnormal signals of external interference type, signal enhancement or shielding will be performed; for abnormal signals of operation abnormality type, real-time prompts or warnings will be sent to the operator, indicating the existing operational errors. The above steps, by using a pre-trained classification model, can quickly and accurately classify and process abnormal signals, effectively distinguish different types of abnormal signals, and thus take targeted countermeasures, improving the system's detection accuracy and response speed, and reducing the need for manual intervention.
[0129] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.
[0130] The above merely describes the preferred embodiments of the present application, and it should be pointed out that those skilled in the art can make several improvements and refinements without departing from the principles of the present application, and these improvements and refinements should also be considered as falling within the protection scope of the present application.
Claims
1. A method for detecting tank tunnel anomalies based on audio signal analysis, characterized in that, Includes the following steps: S1: Audio signals of the equipment under different operating conditions are collected in real time by multiple sensors set around the tank passage; S2: Preprocess the acquired audio signal to remove noise and interference signals, and extract audio features related to the operating status of the tank passage through a high-pass filtering signal processing method; S2 specifically includes: S21: Noise filtering is performed on the acquired audio signal. Specifically, the audio signal is analyzed in time and frequency by short-time Fourier transform, the signal is decomposed into several frequency bands, and low-frequency noise and environmental background noise are identified and removed. S22: Use a high-pass filter to attenuate the low-frequency part in the frequency domain. Set the cutoff frequency of the high-pass filter to 300Hz to remove low-frequency noise and irrelevant signal components, and retain only the signal in the frequency range related to the operating status of the tank passage. S23: Perform time-domain smoothing on the signal after high-pass filtering, and use the moving average method to smooth the amplitude of the audio signal to reduce fluctuations caused by instantaneous changes; S24: By calculating the frequency spectrum and time domain characteristics of the preprocessed audio signal, extract audio features related to the operating status of the tank passage, including spectral peak value, frequency response curve and time domain standard deviation; S3: Based on the audio features extracted in S2, a dynamic feature change model is established. The dynamic feature change model adopts a combination of incremental learning and long short-term memory network to analyze the frequency and amplitude features of the audio signal as time changes and establish the temporal change law of the audio signal. S3 specifically includes: S31: Based on the audio features extracted in S2, including spectral peaks, frequency response curves, and time-domain standard deviation, construct an audio feature dataset, where each data sample consists of a timestamp and a corresponding audio feature vector; S32: Use an incremental learning algorithm to train the audio feature dataset and gradually update the model parameters to adapt it to new audio feature data; S33: Applying Long Short-Term Memory (LSTM) networks to audio feature data to capture temporal changes in audio signals and learn the patterns of audio signal changes over time; S34: Combining incremental learning and long short-term memory networks, the model is first trained on the initial data using incremental learning, and then optimized using long short-term memory networks to accurately capture the temporal characteristics of audio signals. S35: The long short-term memory network model is continuously updated through incremental learning, enabling it to adapt to new audio feature data in real time and capture the dynamic changes of audio signals. S36: Through continuous updates and optimizations, the dynamic feature change model is established to ensure that the model can analyze its dynamic feature changes based on real-time audio signals; S4: Apply the dynamic feature change model established in S3 to the analysis of real-time audio signals. By comparing the characteristics of real-time audio signals with the characteristics of historical normal signals, determine whether any abnormalities have occurred. S4 specifically includes: S41: Input the audio features of the real-time acquired audio signal after processing in step S2 into the dynamic feature change model to obtain the dynamic features of the real-time audio signal. S42: Calculate the difference between the dynamic characteristics of the real-time audio signal and the dynamic characteristics of the historical normal signal, wherein the difference is measured by calculating the Euclidean distance, using the following formula: ,in, Indicates a measure of difference. The first part representing the real-time audio signal 1 eigenvalue, The first normal audio signal in history 1 eigenvalue, Dimensions representing audio features; S43: Based on the calculated difference measure Compare with the set threshold, if If the value exceeds the threshold, it is considered an abnormal signal; S5: If S4 determines that the audio signal is abnormal, then the abnormal signal is classified and processed using a pre-trained classification model.
2. The method for detecting tank tunnel anomalies based on audio signal analysis according to claim 1, characterized in that, S1 specifically includes: S11: Audio acquisition sensors are installed at both ends and the middle of the tank passage. The audio acquisition sensors are microphones, which are used to capture the sound signals generated during the operation of the equipment in real time, and convert the acquired analog audio signals into digital signals through an analog-to-digital converter. S12: Set the sampling frequency of the audio acquisition sensor to collect at least 1000 data points per second.
3. The method for detecting tank tunnel anomalies based on audio signal analysis according to claim 2, characterized in that, S24 specifically includes: S241: Calculate the frequency spectrum of the preprocessed audio signal by applying a fast Fourier transform to the audio signal to obtain its spectrum. S242: Extract the peak values in the spectrum. First, calculate the amplitude of the spectrum. Then, use a peak detection algorithm to extract the maximum amplitude value and the corresponding frequency in the spectrum, and record the frequency and amplitude of the peak values. S243: Calculate the frequency response curve, which represents the amplitude response of an audio signal at different frequencies; obtain the frequency response curve by plotting the amplitude of the spectrum as a function of frequency. S244: Calculate the standard deviation of the audio signal to assess the amplitude variation of the signal.
4. The method for detecting tank tunnel anomalies based on audio signal analysis according to claim 1, characterized in that, Specifically, S32 includes: S321: Input the audio feature dataset constructed in S31 into the incremental learning model and initialize the model parameters. The objective function used by the incremental learning model is: ,in, For model parameters, For loss function, For actual labels, For model output, For input features, The number of samples; S322: Select the appropriate learning rate based on the incremental learning strategy. and the step size of incremental updates The model parameters are updated using the following formula: ,in, For the updated model parameters, For the current model parameters, For learning rate, The gradient of the current model parameters; S323: Use new audio feature data in each incremental training process. and its corresponding tags Incremental updates are performed using the following formula: ,in, Indicates the current parameter New input data gradient on; S324: By iterating through the above incremental learning process, the model parameters are gradually adjusted. This enables the model to adapt to new audio feature data.
5. The method for detecting tank tunnel anomalies based on audio signal analysis according to claim 4, characterized in that, Specifically, S33 includes: S331: Input the audio feature data updated by incremental learning in S32 into the Long Short-Term Memory network to capture the temporal changes of the audio signal; S332: In a long short-term memory network, the hidden state and memory unit state at each time step are calculated step by step using the control mechanism of input gate, forget gate, candidate memory unit and output gate; S333: Learn the temporal variation of audio signals over time by using the hidden state sequence of the Long Short-Term Memory network; S334: After training, the final hidden state of the Long Short-Term Memory network is used as the temporal feature of the audio signal; S335: The weight parameters in the Long Short-Term Memory network are optimized through the backpropagation algorithm, enabling the network to better fit the temporal variation of audio signals.
6. The method for detecting tank tunnel anomalies based on audio signal analysis according to claim 5, characterized in that, Specifically, S36 includes: S361: Based on the training results of incremental learning and long short-term memory networks, continuously update the parameters in the dynamic feature change model; S362: By calculating the prediction error during each training process, the model weights are adjusted using an optimization algorithm to minimize the model's prediction error; S363: The structure of the dynamic feature change model is gradually improved based on the incremental learning method and the iterative updates of the long short-term memory network; S364: When the model error is lower than the preset threshold, it indicates that the dynamic feature change model has been optimized.
7. The method for detecting tank tunnel anomalies based on audio signal analysis according to claim 1, characterized in that, S5 specifically includes: S51: If S4 determines that the audio signal is abnormal, then the feature vector of the abnormal signal is input into the pre-trained classification model for classification. S52: The classification model receives the abnormal signal features from the input and outputs a category label based on the patterns learned during training. This label indicates the type of abnormal signal; specifically when This indicates a device malfunction. Indicates external interference. This indicates an operational error. S53: Based on the category labels obtained in S52, different processing strategies will be adopted. For abnormal signals of equipment failure type, a maintenance notification will be sent; for abnormal signals of external interference type, signal enhancement or shielding will be performed; for abnormal signals of operation abnormality type, real-time prompts or warnings will be sent to the operator, indicating the existing operational error.