Vibration spectrum classification and pipe blockage prediction method based on adaptive power spectral density function

Abstract

The invention relates to a vibration spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function, which comprises the following steps of: complementing a missing value of time domain data and replacing an abnormal value, deleting a duplicate value of the time domain data by taking the data appearing first as a criterion, and smoothing the time domain data to obtain a smooth time domain data; frequency domain feature extraction and dimension reduction are carried out on vibration acceleration data through discrete Fourier transform (DFT), state classification is carried out on time domain and frequency domain features after dimension reduction through a machine learning classification model, and on the basis that the classification result is the working state, the vibration acceleration data are classified; a Transformer model established by using historical accumulated data is combined with a sliding window mechanism to carry out adaptive prediction on power spectral density function characteristics obtained by carrying out discrete Fourier transform on actually measured data input in batches, and when characteristics represented by a prediction window meet blockage approaching characteristics represented by historical data, a system sends out pipe blockage early warning. Compared with the prior art, the states of the pipeline in the slurry conveying process can be effectively classified, and pipeline blockage early warning can be achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of building materials, and particularly to a vibration frequency spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function. Background Art

[0002] With the continuous advancement of urban construction work, the construction projects of urban rail transit have increased significantly. To meet the construction quality requirements of underground projects, subway shield tunneling machines are widely used in urban subway construction. Compared with traditional tunneling technologies, the shield tunneling technology using shield machines has the advantages of safety and reliability, good working environment, less excavated earthwork, high mechanization level, fast construction speed, and low construction cost. Especially when the construction conditions are complex and the tunnel burial depth is large, only shield machines can be relied on.

[0003] The shield machine will send bentonite or clay slurry prepared according to certain requirements to the excavation face from the upper part of the shield through a slurry pump and a slurry pipeline at a certain pressure. The problem of slurry blockage during shield construction caused by component wear has always been very difficult to solve, and in severe cases, it can lead to major risks such as ground heave and collapse in the tunneling section of the shield machine, resulting in the shield machine being buried. For a long time, in shield construction, in order to reduce the losses caused by the slurry blockage problem, more attention has been paid to the improvement of hardware equipment and the minimization of post-loss treatment, while ignoring the time dividend of early warning.

[0004] For example, CN117072182A discloses a safety system and method for subway shield construction. Through preset monitoring sensors, the tunnel environment data, operating status data, and construction process data of the shield machine are monitored in real time, and the tunnel environment data, operating status data, and construction process data of the shield machine are transmitted to the central control module. According to the preset construction requirement indicators, the tunnel environment data, operating status data, and construction process data are analyzed, and it is judged whether there is an abnormal situation in the shield machine according to the preset construction requirement indicators to obtain a judgment result. When the judgment result is an abnormal judgment result, the safety factor of the subway shield construction is evaluated, and based on the current safety factor of the subway shield construction, the regulation plan stored in the control terminal is retrieved. However, this safety system and method for subway shield construction focuses on judging whether there is an abnormal situation in the shield machine by comparing and analyzing the real-time monitored multiple sensor data (monitoring sensors, vibration sensors, temperature sensors, humidity sensors, and oxygen sensors) with the preset construction requirement indicators. However, this method is based on real-time data to judge the real-time state and is not applicable to the analysis of the working state of the pipeline and the trend of pipe blockage.

[0005] Therefore, there is an urgent need for a vibration frequency spectrum classification and pipe blockage prediction method to improve the quality of tunnel construction and ensure construction safety. Summary of the Invention

[0006] The object of the present invention is to provide a vibration spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function, which monitors the vibration spectrum of the easily blocked position of the pipeline, extracts its time-frequency domain characteristics, analyzes the operation of the slurry in the pipeline, predicts and warns of the slurry pipe blockage problem, and has important significance for improving the quality of tunnel construction and ensuring construction safety.

[0007] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] The object of the present invention is to provide a vibration spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function, including the following steps:

[0009] S1. Determine the time period of the acceleration data used to construct and train the model, collect relevant historical pipe blockage data in the engineering database, and intercept the amplitude data of the acceleration sensor at the pipeline bending section or near the bending section within a preset time range before and after the pipe blockage record time as the data set. Use positive and negative signs to represent the vibration direction. The collected vibration acceleration data in the data set includes the amplitude and the vibration direction.

[0010] S2. Perform data preprocessing on the collected vibration acceleration data.

[0011] S3. Extract and reduce the data time-frequency domain characteristics of the preprocessed vibration acceleration data. Use the principal component analysis method and compare with the time-frequency domain atlas to obtain the optimal feature set, and obtain the optimal feature set suitable for classification.

[0012] S4. Perform binary classification on the characteristics of the vibration acceleration data after dimensionality reduction to determine whether it is in the working state. If so, execute step S5 to enter the prediction and warning mode of the system. If not, do not enter the prediction and warning mode of the system.

[0013] S5. Enter the prediction and warning mode of the system, extract the vibration acceleration data in real time in batch form, and obtain the time-frequency domain characteristic function and characteristic critical value suitable for prediction and pipe blockage warning based on the historical data. Use the deep learning method to predict the power spectral density function characteristics of the next window based on the change of the power spectral density function in the form of a sliding window.

[0014] S6. Determine whether the predicted power spectral density function characteristics of the next window are outside the characteristic critical value determined by the historical data. If so, the system issues an alarm. If not, the system does not issue an alarm and returns to S5 to continue the prediction and warning mode of the system.

[0015] Further, in step S2, in data preprocessing, the nearest neighbor interpolation is used to complete the missing values and replace the outliers in the time-domain data. The duplicate values in the time-domain data are deleted based on the earliest-occurring data, and the time-domain data is smoothed by the five-point cubic smoothing method.

[0016] Further, the method of data preprocessing in step S2 is specifically as follows:

[0017] (1) Clean the relevant plugging history data, and use the nearest neighbor interpolation to replace the outliers and missing values in the relevant plugging history data with the sample values of the original data that are closest to the sample points where the outliers and missing values are located;

[0018] (2) Smooth the time-domain data (the relevant plugging history data within the acceleration data time period) by the five-point cubic smoothing method:

[0019] f(x) = ax 3 + bx 2 + cx + d

[0020] where a, b, c, and d are the coefficients of the function, x represents the position in the data, and to ensure smoothness, the values of the four coefficients are obtained based on the adjacent five data points.

[0021] Further, in step S3, in the extraction and dimensionality reduction of data time-frequency domain features and the acquisition of the optimal feature set, the discrete Fourier transform is used to extract the frequency-domain features and perform dimensionality reduction on the collected vibration acceleration data.

[0022] Further, before performing the extraction and dimensionality reduction of data time-frequency domain features in step S3, the vibration acceleration data is normalized based on the following formula:

[0023]

[0024] where S i represents the sequence of the collected preprocessed vibration acceleration data, S min and S max represent the minimum value and the maximum value in the collected preprocessed vibration acceleration data respectively, and S' i represents the sequence of the preprocessed vibration acceleration data after normalization.

[0025] Further, in step S4, the features on which the classifier is based are multiple features in the time-frequency domain of the vibration acceleration data, and are input in the form of a multi-dimensional matrix.

[0026] Further, in step S4, in the binary classification of the states of the features of the vibration acceleration data after dimensionality reduction, machine learning classification models such as decision tree models and support vector machine models are used to classify the states of the features of the vibration acceleration data after dimensionality reduction.

[0027] Further, the historical vibration acceleration data and the extracted features are divided into a normal working state, a non-working state, and a state of approaching or blocking of the pipeline.

[0028] Optionally, in step S4, the support vector machine model is adopted as the machine learning classification model;

[0029] The steps for constructing the support vector machine model include:

[0030] (1) Data partitioning: The existing feature matrix data is partitioned into a training set and a test set (70% and 30%);

[0031] (2) Model construction: The training set is used to train the SVM model. The SVM realizes classification by finding an optimal hyperplane in the feature space. A linear kernel function or a non-linear kernel function, such as a Gaussian kernel function, can be selected according to the data characteristics to handle the case of linear inseparability. The model is trained by adjusting the parameters of the SVM model, such as the regularization parameter C, the kernel function parameter, etc.;

[0032] (3) Model evaluation and optimization: Confusion matrices are used to calculate metrics such as accuracy, precision, and recall to evaluate and optimize the classifier.

[0033] Further, in step S4, the optimal feature sets in the normal working state and the non-working state are taken, and a support vector machine trained by the maximum margin method strategy is used to obtain the separation hyperplane and the classification decision function;

[0034] w * ·x + b * = 0

[0035] G(x) = sign(w * ·x + b * )

[0036] where w * represents the optimal solution of the normal vector, b * represents the optimal solution of the intercept, and G(x) represents the classification decision function.

[0037] Optionally, in step S4, the decision tree classifier model is adopted as the machine learning classification model;

[0038] The steps for constructing the decision tree classifier model include:

[0039] (1) Data division: Divide the existing feature matrix data into a training set and a test set (70% and 30%);

[0040] (2) Model training: Use the training set to train a decision tree model. The decision tree realizes classification by constructing a series of judgment nodes in the feature space. When constructing the decision tree, the Gini index is used as the division criterion, and the stopping conditions (such as the threshold of the number of samples in the leaf node, the depth of the tree, etc.) are determined. According to the features and labels of the training set, a decision tree model is generated by recursively selecting the best division feature and division point;

[0041] (3) Use the confusion matrix to calculate metrics such as accuracy, precision, and recall to evaluate and optimize the classifier.

[0042] Furthermore, in step S4, a decision tree model based on the Gini coefficient (Gini index) is adopted to make the data as pure as possible in each divided subset, and the calculation formula of the Gini coefficient is obtained under the binary classification condition;

[0043]

[0044] D1 = {(x, y) ∈ D | A(x) = a)}; D2 = D - D1

[0045] where C k is the subset of samples in D that belong to the k-th class, and the sample set is divided into two parts, D1 and D2, according to whether the feature A takes a certain possible value. Gini(D, A) represents the Gini coefficient of the set D under the condition of the feature A, that is, the uncertainty of the set D after being divided by A = a; the larger the Gini coefficient value, the greater the uncertainty of the sample set.

[0046] Furthermore, a machine learning classification model can be reasonably selected and established based on the classification situations on the test set and validation set of the optimal features of historical vibration acceleration.

[0047] Furthermore, in step S5, predicting the power spectral density function features of the next window by a deep learning method based on the change of the power spectral density function in the form of a sliding window includes the following process:

[0048] S5.1. Construct data units and solve the power spectral density function features: Determine the size of a single batch of data stream (such as 200 acceleration data) according to the performance of the acceleration sensor and the requirements of the early warning lead time. A single batch of data stream is called a data unit. Use the two-dimensional discrete Fourier transform (DFT) to obtain the two-dimensional spectral image and the power spectral density function features of the data unit of the historical data respectively;

[0049] S5.2. Analyze the characteristics of the power spectral density functions of data units in three states (normal working state, non-working state, and approaching congestion or pipe blockage state) in historical data through comparison, and obtain the critical values applicable to prediction and early warning. The division of the three states is as follows: Divide the historical data and the extracted features into the normal working state, non-working state, and approaching congestion or pipe blockage state;

[0050] S5.3. Establish a sliding window mechanism. Each window contains a number of data units. Extract the characteristics of the power spectral density function within the window, construct an adaptive Transformer model for predicting the power spectral density function, and use the trained adaptive Transformer model to predict the characteristics of the power spectral density function of the next window based on the changes in the power spectral density functions of several windows during the actual measurement process.

[0051] Furthermore, in step S5.1, the construction of data units and the solution of the power spectral density function characteristics specifically include the following process:

[0052] S5.1.1. Select historical data containing three states, perform data preprocessing, and divide the continuous historical data into a large number of data units with equal data volumes;

[0053] S5.1.2. Use the two-dimensional discrete Fourier transform (DFT) on the data units to obtain the power spectral density function, and certain characteristics can be used to characterize the characteristics of the power spectral density function image. The formula for obtaining the power spectral density function / image through the discrete Fourier transform is as follows:

[0054]

[0055] In the formula, f(x, y) represents an N×N matrix. The coordinate system where f(x, y) is located is called the spatial domain, F(u, v) represents the Fourier transform of f(x, y), and the coordinate system where F(u, v) is located is called the frequency domain.

[0056] Furthermore, in step S5.3, the construction of the adaptive Transformer model for predicting the power spectral density function is as follows:

[0057] S5.3.1. Use historical data for training, and divide the processed data units into a training set and a validation set;

[0058] S5.3.2. According to needs, features can be extracted from each power spectral density function in the training set and the validation set. Common features include peak frequency, main frequency bandwidth, energy distribution, etc.;

[0059] S5.3.3. Adjust the structures of the self-attention layer and the feed-forward neural network layer to construct a Transformer model, predict the new power spectral density function features of the next window based on the power spectral density function features of several windows of historical data, and use the mean square error as the loss function;

[0060] S5.3.4. After completing the training and validation of the model, use the power spectral density function features of several measured windows as the input matrix to predict the new power spectral density function features of the next window.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] (1) A vibration spectrum classification and plugging prediction method based on an adaptive power spectral density function proposed by the present invention uses a new index of pipeline vibration acceleration to analyze the working state of the pipeline and the trend of plugging.

[0063] (2) The present invention adopts a sliding window mechanism and a discrete Fourier transform (DFT) method, which can track the changes of the time-frequency spectrum features of acceleration data in real time, is also conducive to summarizing the historical data of plugging rules, and obtains a more reliable and representative working state classification and plugging prediction model with high classification and prediction accuracy; data is processed in data units, and the processing speed is fast.

[0064] (3) A vibration spectrum classification and plugging prediction method based on an adaptive power spectral density function proposed by the present invention utilizes machine learning models (such as support vector machines, decision trees, etc.) and deep learning models (such as Transformer, etc.), has low requirements for the hardware facilities of the system, and has a high degree of automation, thus gaining a time dividend for the control of engineering accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 is a test result image of the fluctuation of the average frequency of the power spectral density function in different working states of a vibration spectrum classification and plugging prediction method based on an adaptive power spectral density function according to Embodiment 1 of the present invention.

[0066] Figure 2 is a test result image of the fluctuation of the low-frequency band energy ratio of the power spectral density function in different working states of a vibration spectrum classification and plugging prediction method based on an adaptive power spectral density function according to Embodiment 1 of the present invention.

[0067] Figure 3 is a test result image of the fluctuation of the average frequency of the power spectral density function in different working states of a vibration spectrum classification and plugging prediction method based on an adaptive power spectral density function according to Embodiment 2 of the present invention.

[0068] Figure 4 It is a test result image of the fluctuation of the low-frequency band energy ratio of the power spectral density function in different working states of a vibration spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function in Embodiment 2 of the present invention.

[0069] Figure 5 It is a schematic flow chart of a vibration spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function of the present invention. Detailed implementation manners

[0070] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and gives detailed implementation manners and specific operation processes, but the protection scope of the present invention is not limited to the following embodiments. Features such as component models, material names, connection structures, control methods, etc. that are not clearly described in this technical solution are regarded as common technical features disclosed in the prior art.

[0071] The present invention discloses a vibration spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function. Through the above method, it is possible to effectively classify the state during the pipeline transportation of slurry and realize the early warning of pipeline blockage.

[0072] The present invention discloses a vibration spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function. The missing values and abnormal values of the time-domain data are filled and replaced by nearest neighbor interpolation. The duplicate values of the time-domain data are deleted based on the earliest data, and the time-domain data is smoothed by the five-point cubic smoothing method. The frequency-domain features of the vibration acceleration data are extracted and dimension-reduced by the discrete Fourier transform (DFT). Machine learning classification models such as decision tree models and support vector machine models are used to classify the dimension-reduced time-domain and frequency-domain features. Based on the classification result of the working state, an adaptive prediction is made on the power spectral density function features obtained by the discrete Fourier transform of the measured data input in batches by a Transformer model established with historical cumulative data combined with a sliding window mechanism. When the features shown in the prediction window meet the trend-blocking features reflected by the historical data, the system issues a pipeline blockage warning.

[0073] The present invention provides a vibration spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function, which includes: (1) selecting training set and validation set data segments in the database; (2) data preprocessing; (3) time-frequency domain feature extraction and obtaining the optimal feature set, obtaining the power spectral density function and its features; (4) the preliminary model building processes such as classifier training, validation, and Transformer model training, validation, etc., and (1) batch input of real-time data; (2) data preprocessing; (3) time-frequency domain feature extraction, obtaining the optimal feature set, and classifier judgment; (4) obtaining the power spectral density function and its features, and the actual measurement process such as Transformer model prediction and warning.

[0074] Further, in data preprocessing, the nearest neighbor interpolation is used to fill in the missing values of the time-domain data and replace the abnormal values, the duplicate values of the time-domain data are deleted based on the earliest occurring data, and the time-domain data is smoothed by the five-point cubic smoothing method.

[0075] Further, in time-frequency domain feature extraction and obtaining the optimal feature set, the discrete Fourier transform (DFT) is used to extract the frequency-domain features and reduce the dimension of the vibration acceleration data.

[0076] Further, in classifier classification, machine learning classification models such as decision tree models and support vector machine models are used to classify the states of the dimension-reduced time-domain and frequency-domain features.

[0077] Further, based on the classification result of the working state, the Transformer model established with historical cumulative data combines with the sliding window mechanism to adaptively predict the power spectral density function features obtained by discrete Fourier transform of the batch input measured data. When the features shown in the prediction window satisfy the pipe-blocking tendency features reflected by the historical data, the system issues a pipe blockage warning.

[0078] The following will further specifically illustrate with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the protection scope of the present invention.

[0079] The following illustrates the implementation manners of the present invention through specific specific embodiments. Those skilled in the art can understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the principles and spirits of the present invention.

[0080] Embodiment 1

[0081] As Figure 5 shown, this embodiment provides a vibration spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function, including the following steps:

[0082] S1. Determine the time period of acceleration data for constructing and training the model. Collect relevant pipe blockage history data (related pipe blockage historical data) in the engineering database. Intercept the amplitude data of the acceleration sensors at the pipe bending section or near the bending section for 1 hour before and after the initial pipe blockage time (22:28:32 on May 25, 2022) on a certain day as the verification data set. Use the acceleration sensor data in the remaining working and pipe blockage time periods in May 2022 as the training data set. Use positive and negative signs to represent the vibration direction. The data set includes the collected vibration acceleration data, and the collected vibration acceleration data includes amplitude and vibration direction.

[0083] S2. Perform data preprocessing on the collected vibration acceleration data (including amplitude and vibration direction). The method of data preprocessing in step S2 is specifically as follows:

[0084] (1) Clean the relevant pipe blockage historical data. Use the nearest neighbor interpolation to replace the outliers and missing values in the relevant pipe blockage historical data with the sample values of the original data closest to the sample points where the outliers and missing values are located.

[0085] (2) Smooth the time-domain data (relevant pipe blockage historical data within the acceleration data time period) by the five-point cubic smoothing method:

[0086] f(x) = ax 3 + bx 2 + cx + d

[0087] where a, b, c, and d are the coefficients of the function, x represents the position in the data, and to ensure smoothness, the values of the four coefficients are obtained according to the adjacent five data points.

[0088] S3. Respectively perform data time-frequency domain feature extraction on the vibration acceleration data of different acceleration sensors after preprocessing. Use the principal component analysis method and compare with the time-frequency domain atlas to obtain the optimal feature set suitable for classification, including two features: the standardized standard deviation in the time domain and the energy spectral density (area integral) in the frequency domain. In data time-frequency domain feature extraction, dimensionality reduction, and obtaining the optimal feature set, use the discrete Fourier transform (DFT) to perform frequency domain feature extraction and dimensionality reduction on the collected vibration acceleration data.

[0089] Table 1 shows an example of time-frequency domain features extracted by a vibration spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function.

[0090] Table 1 Example of time-frequency domain features.

[0091]

[0092] In step S3, before extracting the time-frequency domain features and reducing the dimension of the data, the vibration acceleration data is normalized based on the following formula:

[0093]

[0094] where S i represents the sequence of preprocessed vibration acceleration data collected, S min and S max represent the minimum and maximum values respectively in the preprocessed vibration acceleration data collected, and S' i represents the sequence of preprocessed vibration acceleration data after normalization;

[0095] S4. Classify the features of the vibration acceleration data after dimensionality reduction into two categories of states using a support vector machine or a decision tree model to determine whether it is in operation. If it is in the operating state, the next step is executed. If it is not in the operating state, it does not enter the prediction and early warning mode of the system. In step S4, the machine learning classification model uses a support vector machine model;

[0096] The steps for constructing the support vector machine model include:

[0097] (1) Data division: Divide the existing feature matrix data into a training set and a test set (70% and 30%);

[0098] (2) Model construction: Use the training set to train the SVM model. The SVM realizes classification by finding an optimal hyperplane in the feature space. A linear kernel function or a non-linear kernel function, such as a Gaussian kernel function, can be selected according to the data characteristics to handle the case of linear inseparability. The model is trained by adjusting the parameters of the SVM model, such as the regularization parameter C, the kernel function parameter, etc.;

[0099] (3) Model evaluation and optimization: Use the confusion matrix to calculate indicators such as accuracy, precision, recall, etc. to evaluate and optimize the classifier;

[0100] The steps for constructing the support vector machine model specifically include: In step S4, take the optimal feature sets in the normal working state and the non-working state, and use the support vector machine trained by the maximum margin method strategy to obtain the separation hyperplane and the classification decision function;

[0101] w * ·x + b * = 0

[0102] G(x) = sign(w * ·x + b * )

[0103] where w *Denote the optimal solution of the normal vector as b * Denote the optimal solution of the intercept, and G(x) represents the classification decision function;

[0104] S5. Enter the predictive warning mode of the system, extract vibration acceleration data in real time in batch form (200 data / 0.2 s), and obtain the time-frequency domain characteristic function and characteristic critical value applicable to prediction and pipe blockage warning based on historical data (acceleration data in May 2022). Based on the change of the characteristic function (power spectral density function) in the form of a sliding window (window size is 10 data units, window moving step is 5 data units), use deep learning method to predict the characteristics of the next window. In step S5, the process of predicting the power spectral density function characteristics of the next window based on the change of the power spectral density function in the form of a sliding window using deep learning method includes the following processes:

[0105] S5.1. Construct data units and solve the power spectral density function characteristics: Determine the size of a single batch of data stream (such as 200 acceleration data) according to the performance of the acceleration sensor and the requirements of warning lead time. A single batch of data stream is called a data unit. Use two-dimensional discrete Fourier transform (DFT) to obtain the two-dimensional spectrum image and power spectral density function characteristics of the data unit of historical data respectively;

[0106] S5.2. Compare and analyze the power spectral density function characteristics of data units in three states (normal working state, non-working state, and state of approaching or blocked pipe) in historical data to obtain the critical value applicable to predictive warning. The division of the three states is as follows: Divide the historical data and the extracted characteristics into normal working state, non-working state, and state of approaching or blocked pipe;

[0107] S5.3. Establish a sliding window mechanism. Each window contains several data units. Extract the power spectral density function characteristics within the window, construct an adaptive Transformer model for power spectral density function prediction, and use the trained adaptive Transformer model to predict the characteristics of the power spectral density function of the next window based on the change of the power spectral density function of several windows in the actual measurement process;

[0108] In step S5.1, the process of constructing data units and solving the power spectral density function characteristics specifically includes the following processes:

[0109] S5.1.1. Select historical data containing three states, perform data preprocessing, and divide the continuous historical data into a large number of data units with equal data volume;

[0110] S5.1.2. Apply the two-dimensional discrete Fourier transform (DFT) to the data unit to obtain the power spectral density function, and the characteristics of the power spectral density function image can be characterized by certain features. The formula for obtaining the power spectral density function / image through the discrete Fourier transform is as follows:

[0111]

[0112] In the formula, f(x, y) represents an N×N matrix. The coordinate system where f(x, y) is located is called the spatial domain, F(u, v) represents the Fourier transform of f(x, y), and the coordinate system where F(u, v) is located is called the frequency domain;

[0113] In step S5.3, construct an adaptive Transformer model for power spectral density function prediction. The steps are as follows:

[0114] S5.3.1. Use historical data for training, and divide the processed data units into a training set and a validation set;

[0115] S5.3.2. According to needs, features can be extracted from each power spectral density function in the training set and the validation set. Common features include peak frequency, main frequency bandwidth, energy distribution, etc.;

[0116] S5.3.3. Adjust the structures of the self-attention layer and the feed-forward neural network layer to construct a Transformer model, predict the new power spectral density function features of the next window based on the power spectral density function features of several windows of historical data, and use the mean square error as the loss function;

[0117] S5.3.4. After completing the training and validation of the model, use the power spectral density function features of several measured windows as the input matrix to predict the new power spectral density function features of the next window;

[0118] S6. Compare the power spectral density function features of the predicted next window with the critical value determined by the historical law (historical data). The rules for critical value determination are as follows:

[0119] (1) The mutation of the average frequency shown by the power spectral density functions of the two adjacent sliding windows is greater than 18 Hz, and the mutation of the low-frequency band energy ratio (0 - 30 Hz) is greater than 0.15;

[0120] (2) The average frequency shown by the power spectral density function of the latter sliding window is less than the critical value of 35 Hz, and the low-frequency band energy ratio (0 - 30 Hz) is higher than the critical value of 0.8.

[0121] If it is outside the critical value, the system issues an alarm; if it is within the critical value, the system does not issue an alarm.

[0122] Figure 1 , Figure 2 It is a test result image obtained by using a vibration spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function. Figure 1 It shows the fluctuation of the average frequency of the power spectral density function in different working states, Figure 2 It shows the fluctuation of the low-frequency band energy ratio of the power spectral density function in different working states.

[0123] In the obtained image, green (stuck) is a section of test data with pipe blockage intercepted from the validation dataset, orange (not in function) is data not in the working state intercepted from the validation dataset, blue (function) is data in the normal working state intercepted from the validation dataset, and the abscissa is the time axis in seconds. The pipe blockage data shows a decrease in the average frequency and an increase in the low-frequency band energy ratio during the stage of approaching pipe blockage. The prediction result using the Transformer model and the threshold rule is that pipe blockage will occur at 22:28:30 on May 25th, with a certain early warning amount compared to the actual initial pipe blockage time.

[0124] Embodiment 2

[0125] As Figure 5 shown, this embodiment provides a vibration spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function, including the following steps:

[0126] S1. Determine the time period of acceleration data used to construct and train the model, collect relevant pipe blockage historical materials (relevant pipe blockage historical data) in the engineering database, intercept the amplitude data of the acceleration sensor at the pipe bending section or near the bending section for 1 hour before and after the initial pipe blockage time (03:19:22 on May 30, 2022) on a certain day as the validation dataset, use the acceleration sensor data in the remaining working and pipe blockage time periods in May 2022 as the training dataset, use positive and negative signs to represent the vibration direction, the dataset includes the collected vibration acceleration data, and the collected vibration acceleration data includes the amplitude and the vibration direction;

[0127] S2. Perform data preprocessing on the collected vibration acceleration data (including amplitude and vibration direction). The specific method of data preprocessing in step S2 is as follows:

[0128] (1) Clean the relevant pipe blockage historical data, and use the nearest neighbor interpolation to replace the outliers and missing values in the relevant pipe blockage historical data with the sample values of the original data closest to the sample points where the outliers and missing values are located;

[0129] (2) Smooth the time-domain data (relevant pipe blockage historical data within the acceleration data time period) by the five-point cubic smoothing method;

[0130] f(x) = ax 3 + bx 2 + cx + d

[0131] Among them, a, b, c, and d are the coefficients of the function, x represents the position in the data. To ensure smoothness, the values of the four coefficients are obtained based on five adjacent data points;

[0132] S3. Respectively perform data time-frequency domain feature extraction on the vibration acceleration data of different acceleration sensors after preprocessing. Use the principal component analysis method and compare with the time-frequency domain atlas to obtain the optimal feature set suitable for classification, including two features: the standardized standard deviation in the time domain and the energy spectral density (area integral) in the frequency domain. In data time-frequency domain feature extraction, dimensionality reduction, and obtaining the optimal feature set, the discrete Fourier transform (DFT) is used to perform frequency domain feature extraction and dimensionality reduction on the collected vibration acceleration data;

[0133] Table 1 is an example of time-frequency domain features extracted in a vibration spectrum classification and plugging prediction method based on an adaptive power spectral density function;

[0134] In step S3, before performing data time-frequency domain feature extraction and dimensionality reduction, the vibration acceleration data is normalized based on the following formula:

[0135]

[0136] Among them, S i represents the sequence of vibration acceleration data collected after preprocessing, S min and S max respectively represent the minimum and maximum values in the vibration acceleration data collected after preprocessing, and S' i represents the sequence of vibration acceleration data collected after preprocessing and normalized;

[0137] S4. Perform binary classification of the state of the features of the vibration acceleration data after dimensionality reduction using a support vector machine or a decision tree model to determine whether it is in operation. If it is in the working state, the next step is executed; if it is not in the working state, it does not enter the prediction and warning mode of the system. In step S4, the machine learning classification model uses a decision tree classifier model;

[0138] The construction steps of the decision tree classifier model include:

[0139] (1) Data division: Divide the existing feature matrix data into a training set and a test set (70% and 30%);

[0140] (2) Model training: Use the training set to train a decision tree model. The decision tree realizes classification by constructing a series of judgment nodes in the feature space. When constructing the decision tree, the Gini index is used as the splitting criterion, and the stopping conditions (such as the threshold of the number of samples in the leaf node, the depth of the tree, etc.) are determined. According to the features and labels of the training set, by recursively selecting the best splitting feature and splitting point, a decision tree model is generated;

[0141] (3) Use the confusion matrix to calculate metrics such as accuracy, precision, and recall to evaluate and optimize the classifier;

[0142] Adopt a decision tree model based on the Gini coefficient (Gini index) to make the data as pure as possible in each divided subset, and obtain the calculation formula of the Gini coefficient under the binary classification condition;

[0143]

[0144] D1 = {(x, y) ∈ D | A(x) = a}; D2 = D - D1

[0145] where C k is the subset of samples in D that belong to the k-th class. The sample set is divided into two parts, D1 and D2, according to whether the feature A takes a certain possible value. Gini(D, A) represents the Gini coefficient of the set D under the condition of the feature A, that is, the uncertainty of the set D after being split by A = a; the larger the Gini coefficient value, the greater the uncertainty of the sample set;

[0146] S5. Real-time extract vibration acceleration data in batch form (200 data / 0.2 seconds), and obtain the time-frequency domain feature function and feature critical value applicable to prediction and plugging warning based on historical data (acceleration data in May 2022). Based on the change of the feature function (power spectral density function) in the form of a sliding window (window size is 10 data units, window moving step size is 5 data units), use deep learning methods to predict the features of the next window. In step S5, predicting the power spectral density function features of the next window based on the change of the power spectral density function in the form of a sliding window using deep learning methods includes the following process:

[0147] S5.1. Construct data units and solve the power spectral density function features: Determine the size of a single batch of data stream (such as 200 acceleration data) according to the performance of the acceleration sensor and the requirements of the warning lead time. A single batch of data stream is called a data unit. Use the two-dimensional discrete Fourier transform (DFT) to obtain the two-dimensional spectrum image and power spectral density function features of the data unit of the historical data respectively;

[0148] S5.2. Analyze and compare the characteristics of the power spectral density functions of data units in three states (normal working state, non-working state, and approaching congestion or pipe blockage state) in historical data to obtain critical values applicable to prediction and early warning. The division of the three states is as follows: Divide the historical data and the extracted features into a normal working state, a non-working state, and an approaching congestion or pipe blockage state;

[0149] S5.3. Establish a sliding window mechanism. Each window contains several data units. Extract the characteristics of the power spectral density function within the window, construct an adaptive Transformer model for power spectral density function prediction, and use the trained adaptive Transformer model to predict the characteristics of the power spectral density function of the next window based on the changes in the power spectral density functions of several windows during the actual measurement process;

[0150] In step S5.1, constructing data units and solving the characteristics of the power spectral density function specifically include the following process:

[0151] S5.1.1. Select historical data containing three states, perform data preprocessing, and divide the continuous historical data into a large number of data units with equal data volumes;

[0152] S5.1.2. Use the two-dimensional discrete Fourier transform (DFT) on the data units to obtain the power spectral density function, and certain characteristics can be used to represent the characteristics of the power spectral density function image. The formula for obtaining the power spectral density function / image through the discrete Fourier transform is as follows:

[0153]

[0154] In the formula, f(x, y) represents an N×N matrix. The coordinate system where f(x, y) is located is called the spatial domain, F(u, v) represents the Fourier transform of f(x, y), and the coordinate system where F(u, v) is located is called the frequency domain;

[0155] In step S5.3, to construct an adaptive Transformer model for power spectral density function prediction, the steps are as follows:

[0156] S5.3.1. Use historical data for training, and divide the processed data units into a training set and a validation set;

[0157] S5.3.2. According to needs, extract features from each power spectral density function in the training set and the validation set. Common features include peak frequency, main frequency bandwidth, energy distribution, etc.;

[0158] S5.3.3. Adjust the structures of the self-attention layer and the feed-forward neural network layer to construct a Transformer model, predict the new power spectral density function features of the next window using the power spectral density function features of several windows of historical data, and use the mean squared error as the loss function;

[0159] S5.3.4. After completing the training and verification of the model, use the power spectral density function features of several measured windows as the input matrix to predict the new power spectral density function features of the next window;

[0160] S6. Compare the predicted power spectral density function features of the next window with the critical value determined by the historical pattern (historical data). The rule for critical value determination is:

[0161] (1) The mutation of the average frequency represented by the power spectral density functions of the two adjacent sliding windows is greater than 18 Hz, and the mutation of the low-frequency band energy ratio (0 - 30 Hz) is greater than 0.15;

[0162] (2) The average frequency represented by the power spectral density function of the latter sliding window is less than the critical value of 35 Hz, and the low-frequency band energy ratio (0 - 30 Hz) is higher than the critical value of 0.8.

[0163] If it is outside the critical value, the system issues an alarm; if it is within the critical value, the system does not issue an alarm.

[0164] Figure 3 、 Figure 4 is the test result image obtained by using a vibration spectrum classification and pipe blockage prediction method based on the adaptive power spectral density function. Figure 3 It shows the fluctuations of the average frequency represented by the power spectral density function in different working states, Figure 4 It shows the fluctuations of the low-frequency band energy ratio represented by the power spectral density function in different working states.

[0165] In the obtained image, green (stuck) is a section of test data with pipe blockage intercepted from the validation dataset, orange (not in function) is the data not in the working state intercepted from the validation dataset, blue (function) is the data in the normal working state intercepted from the validation dataset, and the abscissa is the time axis in seconds. The pipe blockage data shows a decrease in the average frequency and an increase in the low-frequency band energy ratio during the stage of approaching pipe blockage. The prediction result using the Transformer model and the critical value rule is that pipe blockage will occur at 03:19:20 on May 30th, with a certain early warning amount compared to the actual initial pipe blockage time.

[0166] By Figure 1 、 Figure 2 、 Figure 3 、 Figure 4, referring to the pipe blockage data and code running time of the original research at the time node, it can be considered that the present invention can basically classify the working state of the pipeline more accurately according to the time-frequency domain characteristics of the pipeline, and predict the occurrence of mud pipe blockage as early as possible through the change of the power spectral density function characteristics in the frequency domain, reducing engineering losses.

[0167] The above are only partial embodiments of the present invention, and the scope of the rights of the present invention cannot be limited thereby. Therefore, equivalent changes made according to the scope of the patent application of the present invention still fall within the protection scope of the present invention.

[0168] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations according to the concept of the present invention without creative labor. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art should fall within the protection scope determined by the claims.

Claims

1. A vibration spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function, characterized in that It includes the following steps: S1. Determine the time period of acceleration data for constructing and training the model. Collect relevant plugging history data in the engineering database, and intercept the amplitude data of the acceleration sensors at the pipeline bending section or near-bending section within a preset time range before and after the plugging record time as the dataset. Use positive and negative signs to represent the vibration direction. The collected vibration acceleration data in the dataset includes amplitude and vibration direction; S2. Perform data preprocessing on the collected vibration acceleration data; S3. Extract and reduce the time-frequency domain features of the preprocessed vibration acceleration data. Use the principal component analysis method and compare with the time-frequency domain spectrum to obtain the optimal feature set, and get the optimal feature set suitable for classification; S4. Perform binary classification on the features of the vibration acceleration data after dimensionality reduction to determine whether it is in the working state. If so, execute step S5 to enter the prediction and warning mode of the system. If not, do not enter the prediction and warning mode of the system; S5. Enter the prediction and warning mode of the system, extract the vibration acceleration data in real time in batches, and obtain the time-frequency domain feature function and feature critical value suitable for prediction and plugging warning based on historical data. Predict the power spectrum density function feature of the next window by deep learning method based on the change of the power spectrum density function in the form of a sliding window; S6. Judge whether the predicted power spectrum density function feature of the next window is outside the feature critical value determined by historical data. If so, the system issues an alarm. If not, the system does not issue an alarm.

2. The vibration frequency spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function according to claim 1, wherein In step S2, in data preprocessing, use the nearest neighbor interpolation to fill in the missing values and replace the abnormal values of the time domain data, delete the duplicate values of the time domain data based on the earliest data, and smooth the time domain data by the five-point cubic smoothing method.

3. A vibration frequency spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function according to claim 1, characterized in that, In step S3, in the extraction and reduction of time-frequency domain features of data and the acquisition of the optimal feature set, use the discrete Fourier transform to extract and reduce the frequency domain features of the collected vibration acceleration data.

4. A vibration frequency spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function according to claim 1, characterized in that, In step S3, perform normalization processing on the vibration acceleration data before extracting and reducing the time-frequency domain features of the data.

5. A vibration frequency spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function according to claim 1, characterized in that In step S4, in the binary classification of the features of the vibration acceleration data after dimensionality reduction, use a machine learning classification model to classify the features of the vibration acceleration data after dimensionality reduction.

6. A vibration frequency spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function according to claim 5, characterized in that, In step S4, the machine learning classification model adopts a support vector machine model; The construction steps of the support vector machine model include: (1) Divide the existing feature matrix data into a training set and a test set; (2) Use the training set to train the SVM model. The SVM realizes classification by finding an optimal hyperplane in the feature space. Adjust the parameters of the SVM model to train the model; (3) Use the confusion matrix to calculate the accuracy, precision, and recall rate to evaluate and optimize the classifier.

7. A vibration frequency spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function according to claim 5, characterized in that, In step S4, the machine learning classification model adopts a decision tree classifier model; The construction steps of the decision tree classifier model include: (1) Divide the existing feature matrix data into a training set and a test set; (2) Use the training set to train a decision tree model. The decision tree realizes classification by constructing a series of judgment nodes in the feature space. When constructing the decision tree, the Gini index is used as the splitting criterion, and the stopping condition is determined. According to the features and labels of the training set, a decision tree model is generated by recursively selecting the best splitting feature and splitting point. (3) Use the confusion matrix to calculate the accuracy, precision, and recall to evaluate and optimize the classifier.

8. A vibration frequency spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function according to claim 1, characterized in that, In step S5, predicting the power spectral density function features of the next window based on the change of the power spectral density function in the form of a sliding window by a deep learning method includes the following process: S5.

1. Construct a data unit and solve the power spectral density function features: Determine the size of a single batch of data stream according to the performance of the acceleration sensor and the requirement of early warning lead time. A single batch of data stream is called a data unit. Use the two-dimensional discrete Fourier transform to obtain the two-dimensional spectrum image and the power spectral density function features of the historical data of the data unit respectively. S5.

2. Compare and analyze the power spectral density function features of the data units in three states in the historical data to obtain the critical values applicable to prediction and early warning. The division of the three states is as follows: The historical data and the extracted features are divided into a normal working state, a non-working state, and a state of approaching or blocked pipeline. S5.

3. Establish a sliding window mechanism. Each window contains several data units. Extract the power spectral density function features within the window, construct an adaptive Transformer model for power spectral density function prediction, and use the trained adaptive Transformer model to predict the features of the power spectral density function of the next window based on the change of the power spectral density function of several windows in the actual measurement process.

9. A vibration frequency spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function according to claim 8, characterized in that In step S5.1, constructing a data unit and solving the power spectral density function features specifically includes the following process: S5.1.

1. Select historical data containing three states, perform data preprocessing, and divide the continuous historical data into multiple data units with equal data volume. S5.1.

2. Use the two-dimensional discrete Fourier transform on the data unit to obtain the power spectral density function, and characterize the characteristics of the power spectral density function image with features. Obtain the power spectral density function through the discrete Fourier transform.

10. A vibration frequency spectrum classification and pipe blockage prediction method based on an adaptive power spectral density function according to claim 8, characterized in that, In step S5.3, constructing an adaptive Transformer model for power spectral density function prediction, the steps are as follows: S5.3.

1. Use the historical data for training, and divide the processed data units into a training set and a validation set. S5.3.

2. Extract features from each power spectral density function in the training set and the validation set. S5.3.

3. Adjust the structures of the self-attention layer and the feed-forward neural network layer to construct a Transformer model to predict the new power spectral density function features of the next window based on the power spectral density function features of several windows of historical data, and use the mean square error as the loss function. S5.3.

4. After completing the training and validation of the model, use the power spectral density function features of several windows in the actual measurement as the input matrix to predict the new power spectral density function features of the next window.

Images