Defect recognition method and system for oil accessories mud pump casing based on deep learning

Abstract

The present application discloses a method and system for defect identification of a mud pump casing of a petroleum accessory based on deep learning, which belongs to the field of machine learning technology. The method includes: collecting the original acoustic signal of the inner cavity of the mud pump casing, and preprocessing the original signal to obtain the acoustic response characteristics of the mud pump casing under the target excitation; calculating the frequency domain energy distribution characteristics, time-frequency correlation characteristics and modal decomposition characteristics of the acoustic response characteristics, and performing discretization processing within a preset frequency range to obtain a feature vector; inputting the feature vector into a trained deep learning model based on a convolutional neural network and a long short-term memory network to obtain the probability distribution of various defect types of the mud pump casing. Through the scheme of the present application, the acoustic characteristics of the casing can be utilized, and combined with signal processing and machine learning technology, to achieve rapid, non-destructive and automated defect detection of the mud pump casing.

Description

Technical Field

[0001] The present application relates to the field of machine learning, and in particular to a method and system for identifying defects in a petroleum accessory mud pump casing based on deep learning. Background Art

[0002] Defect identification of mud pump casings has become a major quality control issue in the field of petroleum industry. It is closely related to the performance and service life of mud pumps, but not all defects will lead to failure or failure of mud pumps, which leads to the concepts of acceptable defects ("AD") and unacceptable defects ("UD"). The identification of AD and UD involves distinguishing the differences in performance and safety of mud pump casing defects. Although AD has certain defects, it will not significantly affect the performance and service life of mud pumps; while UD may cause early failure of mud pumps or serious safety hazards. Therefore, the identification of UD is particularly important for early detection, prevention and management of mud pump quality problems, which helps to reduce maintenance costs and improve equipment reliability. At present, the methods for identifying defects in mud pump casings are mainly based on non-destructive testing technologies such as visual inspection, penetration testing, magnetic particle testing and ultrasonic testing. Although these methods are effective to a certain extent, they have certain limitations, such as relying on the experience and skills of the inspectors, low detection efficiency, and difficulty in detecting internal defects.

[0003] In recent years, with the advancement of artificial intelligence technology, defect recognition methods based on machine learning and deep learning have gradually attracted attention. By collecting a large amount of mud pump housing image data and using complex algorithm models to learn the characteristic patterns of defects from them, UD can be automatically, efficiently and accurately identified, greatly improving the efficiency and reliability of defect detection. However, the existing machine learning methods mainly identify defects based on the final processing state of the mud pump housing, ignoring the different process stages that the mud pump housing undergoes during the manufacturing process. In fact, if the defects introduced in the early process stages such as welding, heat treatment, and rough processing can be identified and processed in time, the waste of materials and time caused by subsequent processing can be effectively avoided. However, due to the large differences in the surface state of the mud pump housing at different process stages, traditional machine learning methods are difficult to adapt to such differences, resulting in low recognition accuracy. In addition, the existing deep learning defect recognition methods are mainly aimed at surface defects, and the recognition effect of internal defects of the mud pump housing is limited.

[0004] Therefore, there is an urgent need for a technical solution that can utilize the acoustic properties of the shell and combine signal processing and machine learning technology to achieve rapid, non-destructive and automated defect detection of mud pump shells. Summary of the invention

[0005] In order to solve the deficiencies of the prior art, the embodiment of the present application provides a method and system for identifying defects of a mud pump casing of a petroleum accessory based on deep learning. The present application solves the technical problems that the prior art is difficult to detect internal defects at a low cost.

[0006] The embodiment of the present application provides a defect identification method for a mud pump casing of a petroleum accessory based on deep learning, comprising: collecting an acoustic original signal from an inner cavity of a mud pump casing, and preprocessing the original signal to obtain an acoustic response characteristic of the mud pump casing under a target excitation; calculating frequency domain energy distribution characteristics, time-frequency correlation characteristics, and modal decomposition characteristics of the acoustic response characteristics, and performing discretization processing within a preset frequency range to obtain a feature vector; inputting the feature vector into a trained deep learning model constructed based on a convolutional neural network and a long short-term memory network to obtain a probability distribution of various defect types of the mud pump casing; comparing the sizes of all probability values ​​in the probability distribution, selecting the category corresponding to the maximum probability value as a primary identification result, and calculating the confidence of the identification result to compare the confidence with a target threshold; calculating a consistency index and a stability index based on the results of repeated identification, and obtaining a credibility index of the defect identification result in combination with the confidence to serve as a reference for defect handling decisions.

[0007] In a possible implementation, the acoustic original signal of the inner cavity of the mud pump housing is collected, and the original signal is preprocessed to obtain the acoustic response characteristics of the mud pump housing under target excitation, including: a high-precision microphone array is evenly arranged in a ring at the inlet and outlet flanges of the mud pump housing, and 8 microphones are arranged at each flange to form a 16-channel acquisition system; an acoustic wave exciter is fixed at the center of the outer wall of the housing; a sinusoidal wave pulse with preset parameters is generated by the acoustic wave exciter, and the microphone array is started at the same time to collect acoustic signals at a preset sampling rate to obtain multiple complete excitation-response cycles; the collected original signal is band-pass filtered to remove low-frequency environmental noise and high-frequency electromagnetic interference; the signals of multiple cycles are superimposed and averaged to obtain a signal with a high signal-to-noise ratio; the averaged signal is short-time Fourier transformed to obtain a time-frequency spectrum diagram to obtain the acoustic response characteristics of the mud pump housing under target excitation, and the acoustic response characteristics include frequency, amplitude and phase information.

[0008] In a possible implementation, the frequency domain energy distribution characteristics, time-frequency correlation characteristics and modal decomposition characteristics of the acoustic response characteristics are calculated, and discretization processing is performed within a preset frequency range to obtain a feature vector, including: using the complex spectrum and microphone weight coefficients of the acoustic response characteristics within the preset frequency range to calculate the frequency domain energy distribution characteristics to obtain a frequency domain energy feature vector; based on time delay and frequency parameters, calculating the time-frequency correlation characteristics between different microphone pairs to generate a time-frequency correlation matrix; extracting the inherent vibration modal characteristics by linearly combining multiple microphone signals, and constructing a modal feature matrix; discretizing and flattening the vectors of the frequency domain energy distribution characteristics, time-frequency correlation characteristics and modal decomposition characteristics to obtain a feature vector.

[0009] In a possible implementation, all probability values ​​in the probability distribution are compared, the category corresponding to the maximum probability value is selected as the primary recognition result, and the confidence of the recognition result is calculated to compare the confidence with the target threshold, including: obtaining the output probability distribution of the defect recognition model, comparing all probability values, and selecting the category corresponding to the maximum probability value as the primary recognition result; calculating the confidence of the recognition result, including: determining the maximum probability value and the second largest probability value; counting the total number of defect categories; multiplying each probability value by its own logarithm and summing them, and then taking the negative value to obtain the probability distribution entropy; based on the maximum probability value, the second largest probability value, the total number of defect categories and the probability distribution entropy, calculating the difference between the maximum probability and the second largest probability and the degree of discreteness of the overall distribution to obtain the confidence; numerically comparing the calculated confidence with the threshold, outputting the recognition result when the confidence is greater than the threshold, and marking the result as a state to be inspected when the confidence is less than or equal to the threshold.

[0010] In a possible implementation, a consistency index and a stability index are calculated based on the results of repeated identification, and a credibility index of the defect identification result is obtained in combination with the confidence level to serve as a reference for defect handling decisions, including: obtaining multiple groups of detection data of the same mud pump casing, each group of detection data including input features and output probability distribution, and the output probability distribution is generated by a defect recognition model; based on the output probability distribution of the multiple groups of detection data, the consistency index of the defect identification result is calculated by a consistency evaluation formula, and the consistency evaluation formula measures the degree of discreteness of multiple detection results; the original input features of the same mud pump casing are evaluated. Add random noise to generate multiple groups of disturbance input features, input the multiple groups of disturbance input features into the defect recognition model, and obtain corresponding multiple groups of disturbance output probability distributions; based on the original input features, the original output probability distribution, the multiple groups of disturbance input features and the multiple groups of disturbance output probability distributions, calculate the stability index of the defect recognition result through a stability evaluation formula, and the stability evaluation formula measures the sensitivity of the model to the input disturbance; according to the confidence, consistency index and stability index output by the defect recognition model, calculate the credibility index of the defect recognition result through a weighted sum formula, and use the credibility index as a reference for defect handling decisions.

[0011] In a possible implementation, the frequency domain energy distribution characteristics include:

[0012] ,in, Indicates the frequency The energy distribution characteristics at Indicates frequency, Indicates the number of the microphone. Indicates the total number of microphones, Indicates Microphones at frequency The complex spectrum at Indicates The weighting factor for each microphone.

[0013] In a possible implementation, the time-frequency correlation feature includes:

[0014] ,in, Indicates time delay and frequency The correlation characteristics of Indicates time delay, and Indicates the number of the microphone. Indicates the total number of microphones, Indicates the total duration of the signal, Indicates microphones at a time and frequency The time spectrum at Indicates microphones at a time and frequency The conjugate of the time-frequency spectrum at .

[0015] In a possible implementation, the modal decomposition feature includes:

[0016] ,in, Indicates The order mode at the frequency The amplitude at represents the order of the mode, Indicates the number of the microphone. Indicates the total number of microphones, Indicates Microphone pair The order mode at the frequency The contribution coefficient at Indicates Microphones at frequency The complex spectrum at .

[0017] The embodiment of the present application also provides a defect recognition system for a mud pump housing of a petroleum accessory based on deep learning, comprising: an acquisition unit, a deep learning unit and a decision unit; wherein the acquisition unit is used to acquire the acoustic original signal of the inner cavity of the mud pump housing, and pre-process the original signal to obtain the acoustic response characteristics of the mud pump housing under the target excitation; the frequency domain energy distribution characteristics, time-frequency correlation characteristics and modal decomposition characteristics of the acoustic response characteristics are calculated, and discretization processing is performed within a preset frequency range to obtain a feature vector; the deep learning unit is used to input the feature vector into a trained deep learning model based on a convolutional neural network and a long short-term memory network to obtain the probability distribution of various defect types of the mud pump housing; the decision unit is used to compare the sizes of all probability values ​​in the probability distribution, select the category corresponding to the maximum probability value as the primary recognition result, and calculate the confidence of the recognition result to compare the confidence with the target threshold; based on the results of repeated multiple recognitions, the consistency index and the stability index are calculated, and the credibility index of the defect recognition result is obtained in combination with the confidence to serve as a reference for defect disposal decisions.

[0018] In the deep learning-based defect identification method and system for the mud pump casing of a petroleum accessory provided above, the embodiment of the present application utilizes the acoustic characteristics of the casing and combines signal processing and machine learning techniques to achieve rapid, non-destructive and automated defect detection of the mud pump casing. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 A schematic flow chart of a method for identifying defects of a mud pump casing of a petroleum accessory based on deep learning provided in an embodiment of the present application; Figure 2 A schematic diagram of acoustic signal acquisition provided in an embodiment of the present application; Figure 3 A schematic block diagram of a deep learning-based defect recognition system for a petroleum accessory mud pump casing provided in an embodiment of the present application. DETAILED DESCRIPTION

[0021] Various exemplary embodiments of the present application will now be described in detail with reference to the accompanying drawings. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present application.

[0022] Those skilled in the art will appreciate that the terms "first", "second" and the like in the embodiments of the present application are only used to distinguish different steps, devices or modules, etc., and do not represent any specific technical meaning, nor do they represent the necessary logical order between them. It should also be understood that in the embodiments of the present application, "multiple" may refer to two or more, and "at least one" may refer to one, two or more. It should also be understood that for any component, data or structure mentioned in the embodiments of the present application, in the absence of explicit limitation or contrary revelation given in the context, it can generally be understood as one or more. In addition, the term "and / or" in the present application is only a kind of association relationship describing the associated objects, indicating that there can be three kinds of relationships, for example, A and / or B, which can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in the present application generally indicates that the associated objects before and after are an "or" relationship. It should also be understood that the description of each embodiment in the present application emphasizes the differences between the embodiments, and the same or similar parts can refer to each other. For the sake of brevity, they will not be repeated one by one.

[0023] At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present application and its application or use. The techniques, methods and devices known to ordinary technicians in the relevant fields may not be discussed in detail, but where appropriate, the techniques, methods and devices should be considered as part of the specification. It should be noted that similar numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0024] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0025] Figure 1 A flow chart of a defect identification method for a petroleum accessory mud pump housing based on deep learning provided in an embodiment of the present application. The petroleum accessory mud pump housing refers to an important equipment component used in the oil extraction process, and its main function is to accommodate and support the core components of the mud pump to ensure the normal operation of the mud pump. The manufacturing process of the mud pump housing generally includes but is not limited to steps such as welding, heat treatment, rough machining and fine machining. Among them, welding refers to the process of connecting the various parts of the housing by molten metal; heat treatment refers to changing the internal structure of the metal material by heating and cooling to obtain the desired mechanical properties; rough machining refers to preliminary mechanical machining of the housing to remove excess material; fine machining is the final fine machining step to achieve the desired dimensional accuracy and surface quality.

[0026] The core idea of ​​the embodiment of the present application is to identify defects after heat treatment and before subsequent processing. The advantage of this approach is that defects can be discovered in a timely manner, avoiding unnecessary subsequent processing of defective shells, thereby reducing waste of materials and energy and improving production efficiency. The embodiment of the present application makes full use of the cavity structure characteristics of the mud pump shell to identify internal defects through acoustic analysis. This non-destructive detection method can maintain the integrity of the shell while detecting deep internal defects. It has a fast detection speed and is suitable for mass production.

[0027] Specifically, the embodiment of the present application includes the following steps: First, a high-precision microphone array is installed at the inlet and outlet flanges of the mud pump housing. The high-precision microphone array refers to a sound collection device composed of multiple high-sensitivity microphones arranged in a specific manner, which is used to accurately collect acoustic signals. These microphones can be of different types such as capacitive, dynamic or piezoelectric, and are selected according to specific application scenarios. The arrangement of the microphone array can be linear, planar or stereo, etc., to ensure that the acoustic information inside the housing can be fully collected.

[0028] Next, a sound wave exciter of a specific frequency is used to generate standing waves inside the shell. The sound wave exciter is a device that can generate sound waves of a specific frequency, which can be a piezoelectric transducer, an electromagnetic transducer, or a pneumatic transducer. Standing waves refer to waves formed by the superposition of incident waves and reflected waves in a closed space, and are characterized by certain points of the waves (called nodes) always remaining stationary. By generating standing waves inside the shell, the sound waves can form a stable sound field distribution in the cavity, which is beneficial for subsequent defect identification.

[0029] Then, the acoustic response signal inside the shell is collected, including frequency, amplitude, phase and other characteristics. Frequency refers to the speed of sound wave vibration, usually in Hertz (Hz); amplitude refers to the intensity of sound wave vibration, which can be expressed in decibels (dB); phase describes the relative position of the sound wave in time or space. These characteristics together constitute a complete description of the acoustic response inside the shell.

[0030] The core of the embodiment of the present application is to use machine learning algorithms, especially convolutional neural networks (CNNs), to analyze these acoustic features to identify internal defects. Convolutional neural networks are a deep learning model that is particularly suitable for processing data with grid topology, such as time series data or image data. Through structures such as convolutional layers, pooling layers, and fully connected layers, it can automatically learn the hierarchical features of data, thereby achieving efficient pattern recognition.

[0031] The cavity structure inside the mud pump housing produces a specific response pattern to sound waves. This response pattern can be regarded as the "acoustic fingerprint" of the housing, reflecting the internal structure and material properties of the housing. When there are defects inside the housing, such as cracks, pores, etc., these defects will change the propagation path and reflection characteristics of the sound waves, resulting in changes in the response pattern. Through machine learning algorithms, the differences in the response patterns of normal and defective housings can be learned to identify these anomalies.

[0032] The specific implementation steps of the embodiment of the present application are as follows: Figure 1As shown, at step S101, the acoustic original signal of the inner cavity of the mud pump housing is collected, and the original signal is preprocessed to obtain the acoustic response characteristics of the mud pump housing under the target excitation. The preprocessing may include but is not limited to signal denoising, normalization, time-frequency transformation and other operations. Denoising may use methods such as wavelet transform or Fourier transform; normalization may use methods such as maximum and minimum value normalization or Z-score normalization; time-frequency transformation may use methods such as short-time Fourier transform (STFT) or wavelet transform.

[0033] Specifically, the method includes: arranging a high-precision microphone array in a circular and uniform manner at the inlet and outlet flanges of the mud pump housing, with 8 microphones arranged at each flange to form a 16-channel acquisition system; fixing an acoustic wave exciter at the center of the outer wall of the housing; using the acoustic wave exciter to generate a sinusoidal wave pulse with preset parameters, and simultaneously starting the microphone array to collect acoustic signals at a preset sampling rate to obtain multiple complete excitation-response cycles; performing bandpass filtering on the collected original signal to remove low-frequency environmental noise and high-frequency electromagnetic interference; superimposing and averaging signals of multiple cycles to obtain a signal with a high signal-to-noise ratio; performing short-time Fourier transform on the averaged signal to obtain a time-frequency spectrum diagram to obtain the acoustic response characteristics of the mud pump housing under target excitation, wherein the acoustic response characteristics include frequency, amplitude and phase information.

[0034] In one embodiment (such as Figure 2 As shown in the figure, during the acquisition process, the exciter is first used to generate a sine wave pulse with a duration of 1 ms and a peak amplitude of 100 V, with a repetition frequency of 100 Hz. At the same time, the microphone array is started, the sampling rate is set to 100 kHz, and the acquisition time is 1 second. In this way, 100 complete stimulus-response cycles can be obtained.

[0035] The collected raw signal was preprocessed as follows: 1. A bandpass filter (passband 10 kHz - 50 kHz) was applied to remove low-frequency environmental noise and high-frequency electromagnetic interference. 2. The 100-cycle signal was averaged to improve the signal-to-noise ratio. 3. The averaged signal was subjected to short-time Fourier transform (STFT) with a window length of 1 ms and an overlap rate of 50% to obtain a time-frequency spectrum.

[0036] Through these steps, the acoustic response characteristics of the mud pump casing under specific excitation can be obtained, including frequency, amplitude and phase information. These characteristics will serve as input data for subsequent defect identification.

[0037] At step S102, the frequency domain energy distribution feature, time-frequency correlation feature and modal decomposition feature of the acoustic response feature are calculated, and discretization processing is performed within a preset frequency range to obtain a feature vector. The frequency domain energy distribution feature can be obtained by power spectral density (PSD) analysis; the time-frequency correlation feature can be obtained by calculating the frequency correlation coefficient of different time windows; the modal decomposition feature can be obtained by methods such as empirical mode decomposition (EMD) or variational mode decomposition (VMD). Discretization processing can use equal frequency binning or equal width binning methods to convert continuous feature values ​​into discrete categories.

[0038] Specifically, it includes: using the complex spectrum of the acoustic response characteristics within a preset frequency range and the microphone weight coefficient to calculate the frequency domain energy distribution characteristics to obtain the frequency domain energy feature vector; based on the time delay and frequency parameters, calculating the time-frequency correlation characteristics between different microphone pairs to generate a time-frequency correlation matrix; extracting the inherent vibration modal characteristics by linearly combining multiple microphone signals, and constructing a modal feature matrix; discretizing and flattening the vectors of the frequency domain energy distribution characteristics, the time-frequency correlation characteristics and the modal decomposition characteristics to obtain a feature vector.

[0039] In one embodiment, in order to extract effective features from the collected acoustic signals, the present application discloses the following mathematical formulas for preliminary calculation and preprocessing: wherein the frequency domain energy distribution features include:

[0040] ,in, Indicates the frequency The energy distribution characteristics at Indicates frequency, Indicates the number of the microphone. Indicates the total number of microphones, Indicates Microphones at frequency The complex spectrum at Indicates The formula calculates the weighted energy sum of all microphones at each frequency point, reflecting the energy distribution characteristics of the sound wave inside the shell.

[0041] Among them, the time-frequency correlation features include:

[0042] ,in, Indicates time delay and frequency The correlation characteristics of Indicates time delay, and Indicates the number of the microphone. Indicates the total number of microphones, Indicates the total duration of the signal, Indicates microphones at a time and frequency The time spectrum at Indicates microphones at a time and frequency This formula calculates the time-frequency correlation between different microphone pairs and can reflect the propagation characteristics of sound waves inside the shell.

[0043] Among them, the modal decomposition features include:

[0044] ,in, Indicates The order mode at the frequency The amplitude at represents the order of the mode, Indicates the number of the microphone. Indicates the total number of microphones, Indicates Microphone pair The order mode at the frequency The contribution coefficient at Indicates Microphones at frequency This formula extracts the natural vibration mode characteristics of the shell by linearly combining multiple microphone signals.

[0045] Using the above three formulas, a series of intermediate features can be extracted from the original acoustic signal. The specific steps are as follows: 1. Calculate E(f) for each frequency point (10 kHz - 50 kHz, step size 100 Hz) to obtain a 401-dimensional frequency domain energy feature vector. 2. Calculate C(τ, f) for time delay τ from 0 to 10 ms (step size 0.1 ms) and frequency f from 10 kHz to 50 kHz (step size 1 kHz) to obtain a 101x41-dimensional time-frequency correlation matrix. 3. Calculate Φk(f) (k=1, 2, ..., 10) of the first 10 modes. Each mode takes 41 frequency points in the range of 10 kHz - 50 kHz to obtain a 10x41-dimensional modal feature matrix.

[0046] These feature vectors and matrices are flattened and concatenated to finally obtain a 6551-dimensional feature vector as the input of the subsequent machine learning model.

[0047] At step S103, the feature vector is input into a trained deep learning model based on a convolutional neural network and a long short-term memory network to obtain the probability distribution of various defect types of the mud pump casing. The deep learning model is a hybrid model that combines the advantages of a convolutional neural network (CNN) and a long short-term memory network (LSTM). CNN is suitable for processing spatial features, while LSTM is good at processing temporal features. The combination of the two can more comprehensively analyze the spatiotemporal features of acoustic signals.

[0048] Specifically, we can combine the advantages of convolutional neural network (CNN) and long short-term memory network (LSTM). The mathematical expression of the model is as follows:

[0049] H1 = ReLU(Conv1D(X, W1) + b1), H2 = MaxPool(H1), H3 = ReLU(Conv1D(H2, W2) + b2), H4 = MaxPool(H3), H5 = LSTM(H4, Wf, Wi, Wo, Wc), Y = Softmax(Dense(H5, Wd) + bd), where X is the input 6551-dimensional feature vector and Y is the output defect type probability distribution. The specific implementation steps are as follows: Reshape the input feature X into a 101x65 two-dimensional matrix to retain the local structure of time-frequency correlation. The first convolution layer uses 32 convolution kernels of size 3x3, with a step size of 1 and a padding mode of "same". The activation function uses ReLU. The first maximum pooling layer uses a pooling window of 2x2 with a step size of 2. The second convolution layer uses 64 convolution kernels of size 3x3, and the other parameters are the same as the first convolution layer. The parameters of the second max pooling layer are the same as the first max pooling layer. The LSTM layer uses 128 hidden units and the input is a 25x64 sequence (25 time steps, 64-dimensional features per step). The fully connected layer maps the output of the LSTM to the number of defect categories (assuming 5 categories: no defects, cracks, pores, inclusions, and deformation). The Softmax function converts the output of the fully connected layer into a probability distribution. The model training uses the cross entropy loss function and the Adam optimizer, with a learning rate of 0.001, a batch size of 32, and 100 training rounds. To prevent overfitting, a dropout layer is added before the fully connected layer with a dropout rate of 0.5.

[0050] At step S104, all probability values ​​in the probability distribution are compared, the category corresponding to the maximum probability value is selected as the primary recognition result, and the confidence of the recognition result is calculated to compare the confidence with the target threshold. The confidence can be calculated by the output value of the softmax function or other statistical methods. The setting of the target threshold needs to be adjusted according to the specific application scenario and requirements to balance the accuracy and recall rate of recognition.

[0051] Specifically, it includes: obtaining the output probability distribution of the defect recognition model, comparing the sizes of all probability values, and selecting the category corresponding to the maximum probability value as the primary recognition result; calculating the confidence of the recognition result, including: determining the maximum probability value and the second largest probability value; counting the total number of defect categories; multiplying each probability value by its own logarithm and summing them up, and then taking the negative value to obtain the probability distribution entropy; based on the maximum probability value, the second largest probability value, the total number of defect categories and the probability distribution entropy, calculating the difference between the maximum probability and the second largest probability and the degree of discreteness of the overall distribution to obtain the confidence; numerically comparing the calculated confidence with the threshold, outputting the recognition result when the confidence is greater than the threshold, and marking the result as a pending inspection state when the confidence is less than or equal to the threshold.

[0052] The model outputs the probability distribution of various defect types. In one embodiment, in order to obtain the final defect recognition result, the following processing is required: take the category with the highest probability as the preliminary recognition result. The confidence of the recognition result is calculated using the following formula: Confidence = (Pmax - Psecond) / (1 - 1 / N) * (1 - H / log2(N)), where Pmax is the maximum probability value, Psecond is the second highest probability value, N is the total number of defect categories, and H is the entropy of the output probability distribution: H = -∑[i=1 to N] Pi * log2(Pi). This formula takes into account the gap between the maximum probability and the second highest probability, as well as the uncertainty of the overall probability distribution. When the maximum probability is much greater than other probabilities and the overall distribution is concentrated, the confidence is high. Set a confidence threshold (such as 0.8), and only recognition results that exceed the threshold are considered reliable. Results below the threshold are marked as "uncertain" and require further manual inspection.

[0053] In addition, for the results identified as defects, the location and size of the defects need to be estimated. The time delay information in the time-frequency correlation feature C(τ, f) can be used in combination with the propagation speed of the sound wave in the material to roughly estimate the depth position of the defect. The size of the defect can be indirectly evaluated by the degree of abnormality of the modal decomposition feature Φk(f).

[0054] At step S105, the consistency index and stability index are calculated based on the results of repeated identification, and the credibility index of the defect identification result is obtained in combination with the confidence level as a reference for defect handling decisions. The consistency index can be obtained by calculating the consistency ratio of multiple identification results; the stability index can be obtained by calculating the variance or standard deviation of multiple identification results. The credibility index can be a weighted combination of these indicators, and the setting of the weight needs to be optimized according to actual application requirements.

[0055] Specifically, it includes: obtaining multiple groups of detection data of the same mud pump casing, each group of detection data includes input features and output probability distribution, and the output probability distribution is generated by a defect recognition model; based on the output probability distribution of the multiple groups of detection data, calculating the consistency index of the defect recognition result through a consistency evaluation formula, and the consistency evaluation formula measures the discreteness of multiple detection results; adding random noise to the original input features of the same mud pump casing to generate multiple groups of disturbance input features, and inputting the multiple groups of disturbance input features into the defect recognition model to obtain corresponding multiple groups of disturbance output probability distributions; based on the original input features, the original output probability distribution, the multiple groups of disturbance input features and the multiple groups of disturbance output probability distributions, calculating the stability index of the defect recognition result through a stability evaluation formula, and the stability evaluation formula measures the sensitivity of the model to input disturbances; according to the confidence, consistency index and stability index output by the defect recognition model, calculating the credibility index of the defect recognition result through a weighted sum formula, and using the credibility index as a reference for defect handling decisions.

[0056] In one implementation scenario, in order to further evaluate the reliability of the defect identification result, the following mathematical formula is disclosed: Reliability = α * Confidence + β * Consistency + γ * Stability, where Confidence is the confidence calculated in the fourth step, Consistency is the consistency of multiple test results, and Stability is the stability to input disturbances. α, β, and γ are weight coefficients, and α + β + γ = 1.

[0057] The calculation formula for Consistency is:

[0058] ,in, is the output probability distribution of the i-th detection, is the average distribution of M tests, and Var(Y) is the theoretical maximum variance. This formula measures the degree of dispersion of multiple test results. The more concentrated, the higher the consistency.

[0059] The calculation formula of Stability is: Stability = exp(-λ * ∑[j=1 to K] ||Yj - Y0||² / ||Xj - X0||²), where X0 and Y0 are the original input and output, Xj and Yj are the input and output after adding small disturbances, K is the number of disturbances, and λ is the scaling factor. This formula measures the sensitivity of the model to input disturbances. The less sensitive it is, the higher the stability.

[0060] The specific implementation steps are as follows: Repeat the inspection of the same mud pump casing M times (such as M=5) to obtain multiple groups of output probability distributions. Calculate the Consistency of these M inspection results. Add K groups (such as K=10) of random Gaussian noise to the original input feature X, and the standard deviation is set to 1% of the original feature standard deviation. Send these K groups of perturbed inputs into the model to obtain K groups of outputs. Calculate the Stability of the original output and the K groups of perturbed outputs. Set α=0.5, β=0.3, γ=0.2, and calculate the final Reliability. Reliability is used as a credibility indicator of defect recognition results, which can be used to guide subsequent manual re-inspection or rework decisions.

[0061] Figure 2 A schematic diagram of acoustic signal acquisition provided in an embodiment of the present application. Figure 2 As shown, the mud pump housing (the housing body is represented by cyan), the inlet flange (the blue ring on the lower side), the outlet flange (the blue ring on the upper side), the microphone array (the red dots at the inlet and outlet flanges), and the ultrasonic transducer (the green triangle in the middle of the housing).

[0062] An embodiment of the present application provides an acoustic signal acquisition device and method, wherein the acoustic signal acquisition device includes a mud pump housing, an inlet flange, an outlet flange, a microphone array, and an ultrasonic transducer. It should be understood that an acoustic signal refers to a mechanical wave propagating in a medium, which can be propagated through air, liquid, or solid, and can be sensed by the human ear or a specialized device. Acoustic signal acquisition refers to the use of specific equipment or devices to capture, record, and analyze these sound waves. In industrial applications, acoustic signal acquisition is often used in fields such as equipment monitoring, fault diagnosis, and performance evaluation.

[0063] The mud pump housing refers to the external structure of the mud pump, which is used to accommodate and protect the internal components of the mud pump. In the present embodiment, the mud pump housing is simplified into a cylindrical shape for ease of display and explanation. However, it should be emphasized that the scope of application of the present application is not limited to this specific shape. In fact, the present application can be applied to any mud pump housing with openings at both ends, including but not limited to elliptical, polygonal or irregularly shaped housings. This flexibility enables the present application to adapt to mud pumps of various models and designs, thereby having wide practicality.

[0064] The inlet flange and outlet flange are important structural components on the mud pump housing, located on the left and right sides of the housing respectively. A flange is a mechanical part used to connect pipes, valves, pumps and other equipment, usually in the shape of a disc with multiple bolt holes. In this embodiment, the inlet flange and outlet flange not only play the role of connection and sealing, but also serve as an ideal location for installing a microphone array. This design makes full use of the structural characteristics of the flange, which does not affect the normal operation of the mud pump and can effectively collect acoustic signals.

[0065] The microphone array is one of the core components in the embodiment of the present application, which is used to collect acoustic signals generated during the operation of the mud pump. The microphone array consists of multiple high-precision microphones, which are arranged in a specific way to achieve comprehensive and accurate acoustic signal collection. In this embodiment, the microphone array is installed at the inlet and outlet flanges to form an annular distribution. This arrangement has several significant advantages: first, it can capture the acoustic characteristics inside the mud pump in all directions; second, by collecting signals at both ends of the inlet and outlet at the same time, the propagation characteristics of sound waves inside the mud pump can be analyzed; finally, the annular distribution design can minimize the blind spots of signal collection.

[0066] Specifically, this embodiment uses a 1 / 2-inch free-field microphone with a wide frequency response range of 6.3 Hz to 20 kHz, which can cover the acoustic signal frequencies in most industrial applications. Its sensitivity of 50 mV / Pa ensures the accurate capture of weak sounds. At each flange, 8 microphones are evenly arranged in a ring to form a 16-channel acquisition system. This configuration not only improves the spatial resolution of the acquisition, but also provides a rich data source for subsequent signal processing and analysis.

[0067] The ultrasonic transducer is another key component in the embodiment of the present application, which is used to generate sound wave excitation. This embodiment uses an ultrasonic transducer with a center frequency of 1 MHz. This high-frequency ultrasonic transducer can generate narrow-band sound wave excitation, which is of great significance for accurately analyzing the internal structure and fluid properties of the mud pump. The ultrasonic transducer is fixed at the center of the outer wall of the shell. This arrangement ensures that the sound waves can be evenly propagated throughout the entire inner cavity of the mud pump. By controlling the excitation parameters of the ultrasonic transducer, researchers can simulate different working conditions to comprehensively evaluate the performance and status of the mud pump.

[0068] The acoustic signal acquisition method of the embodiment of the present application may include the following steps: First, install the microphone array at the inlet and outlet flanges of the mud pump, ensuring that each microphone is firmly fixed at a predetermined position. Then, install the ultrasonic transducer at the center of the outer wall of the mud pump housing. Next, start the mud pump and make it reach normal working state. During the operation of the mud pump, activate the ultrasonic transducer to generate sound wave excitation of a preset frequency. At the same time, the acoustic signals inside the mud pump are collected through the microphone array. These signals are recorded and stored through the data acquisition system for subsequent analysis.

[0069] It should be noted that the acoustic signal acquisition device and method of the embodiment of the present application can be flexibly adjusted according to actual needs. For example, the number of microphones can be increased or decreased, the arrangement of the microphones can be changed, or ultrasonic transducers with different frequency ranges can be used. In addition, additional sensors, such as acceleration sensors or pressure sensors, can be added to other locations of the mud pump housing to obtain more comprehensive equipment operation data.

[0070] Figure 3 A schematic block diagram of a defect identification system for a mud pump casing of an oil accessory based on deep learning provided in an embodiment of the present application. It should be understood that the system shown in the figure is exemplary and not restrictive. This means that the system architecture involved is not limited to a specific form or design, but is presented as an example. In other words, the architecture shown in the figure can be regarded as a way of expression to clearly describe related concepts and relationships, and does not exclude other forms of architecture. Therefore, when interpreting the architecture in the picture, it should be understood that the model is flexible and diverse, and its purpose is to provide an exemplary description rather than a restrictive provision on a specific form.

[0071] The embodiment of the present application discloses a defect recognition system for a mud pump housing of a petroleum accessory based on deep learning, comprising: an acquisition unit 301, a deep learning unit 302 and a decision unit 303; wherein the acquisition unit 301 is used to acquire the original acoustic signal of the inner cavity of the mud pump housing, and pre-process the original signal to obtain the acoustic response characteristics of the mud pump housing under the target excitation; calculate the frequency domain energy distribution characteristics, time-frequency correlation characteristics and modal decomposition characteristics of the acoustic response characteristics, and perform discretization processing within a preset frequency range to obtain a feature vector; the deep learning unit 302 is used to collect the original acoustic signal of the inner cavity of the mud pump housing, and pre-process the original signal to obtain the acoustic response characteristics of the mud pump housing under the target excitation; calculate the frequency domain energy distribution characteristics, time-frequency correlation characteristics and modal decomposition characteristics of the acoustic response characteristics, and perform discretization processing within a preset frequency range to obtain a feature vector; the deep learning unit 30 2 is used to input the feature vector into the trained deep learning model based on the convolutional neural network and the long short-term memory network to obtain the probability distribution of various defect types of the mud pump casing; the decision unit 303 is used to compare the sizes of all probability values ​​in the probability distribution, select the category corresponding to the maximum probability value as the primary recognition result, and calculate the confidence of the recognition result to compare the confidence with the target threshold; calculate the consistency index and the stability index based on the results of repeated multiple recognitions, and obtain the credibility index of the defect recognition result in combination with the confidence to serve as a reference for defect disposal decisions.

[0072] In summary, the advantages of the embodiments of the present application are: first, it is a non-destructive detection method that does not require cutting or destructive testing, and can maintain the integrity of the shell. This is very important for protecting the high-value mud pump shell, and it also saves detection time and cost. Secondly, this method can detect deep internal defects. Traditional surface detection methods are difficult to find internal defects, and the embodiments of the present application can "see" the situation inside the shell through acoustic analysis. Thirdly, the detection speed is fast and suitable for mass production. Compared with traditional X-ray or ultrasonic detection methods, the detection process of the embodiments of the present application is faster and can meet the needs of large-scale production. Finally, through the application of deep learning models, the embodiments of the present application have strong adaptability and learning ability, and can continuously optimize recognition performance.

[0073] Furthermore, an embodiment of the present application also provides a defect identification device for a mud pump casing of an oil accessory based on deep learning, comprising: a processor, a memory, and a system bus; the processor and the memory are connected via the system bus; the memory is used to store one or more programs, and the one or more programs include instructions, which, when executed by the processor, enable the processor to execute any of the above methods.

[0074] Furthermore, an embodiment of the present application also provides a computer program product, which, when executed on a terminal device, enables the terminal device to execute any of the above methods.

[0075] It can be known from the description of the above implementation mode that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., including several instructions for enabling a computer device (which can be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0076] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0077] It should also be noted that, in the embodiments of the present application, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements includes not only those elements, but also includes other elements that are not clearly listed, or also includes elements inherent to such process, method, article or equipment. In the absence of more restrictions, the elements limited by the statement "comprise one..." do not exclude the presence of other identical elements in the process, method, article or equipment including the elements.

[0078] The above description of the disclosed embodiments enables professionals and technicians in the field to implement or use the present application. Various modifications to these embodiments will be apparent to professionals and technicians in the field, and the general principles defined in the embodiments of the present application can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown in the embodiments of the present application, but will conform to the widest range consistent with the principles and novel features disclosed in the embodiments of the present application.

Claims

1. A method for identifying defects in a mud pump casing of a petroleum accessory based on deep learning, characterized in that: include: Acquiring an original acoustic signal from the inner cavity of a mud pump housing, and preprocessing the original signal to obtain an acoustic response characteristic of the mud pump housing under a target excitation; Calculating frequency domain energy distribution characteristics, time-frequency correlation characteristics, and modal decomposition characteristics of the acoustic response characteristics, and performing discretization processing within a preset frequency range to obtain a feature vector; Among them, it includes: using the complex spectrum and microphone weight coefficient of the acoustic response characteristics within a preset frequency range to calculate the frequency domain energy distribution characteristics to obtain the frequency domain energy feature vector; based on the time delay and frequency parameters, calculating the time-frequency correlation characteristics between different microphone pairs to generate a time-frequency correlation matrix; extracting the inherent vibration modal characteristics by linearly combining multiple microphone signals, and constructing a modal feature matrix; discretizing and flattening the vectors of frequency domain energy distribution characteristics, time-frequency correlation characteristics and modal decomposition characteristics to obtain a feature vector; inputting the feature vector into a trained deep learning model based on a convolutional neural network and a long short-term memory network to obtain the probability distribution of various defect types of the mud pump casing; comparing the sizes of all probability values ​​in the probability distribution, selecting the category corresponding to the maximum probability value as the primary recognition result, and calculating the confidence of the recognition result to compare the confidence with the target The method comprises: obtaining the output probability distribution of the defect recognition model, comparing the sizes of all probability values, and selecting the category corresponding to the maximum probability value as the primary recognition result; calculating the confidence of the recognition result, including: determining the maximum probability value and the second largest probability value; counting the total number of defect categories; multiplying each probability value by its own logarithm and summing them up, and then taking the negative value to obtain the probability distribution entropy; calculating the difference between the maximum probability and the second largest probability and the degree of dispersion of the overall distribution based on the maximum probability value, the second largest probability value, the total number of defect categories and the probability distribution entropy to obtain the confidence; comparing the calculated confidence with the threshold, outputting the recognition result when the confidence is greater than the threshold, and marking the result as a state to be inspected when the confidence is less than or equal to the threshold; calculating the consistency index and the stability index based on the results of repeated multiple recognitions, and obtaining the credibility index of the defect recognition result in combination with the confidence to serve as a reference for defect disposal decisions.

2. The defect identification method according to claim 1, characterized in that: in, The acoustic original signal of the inner cavity of the mud pump housing is collected, and the original signal is preprocessed to obtain the acoustic response characteristics of the mud pump housing under target excitation, including: a high-precision microphone array is evenly arranged in a ring at the inlet and outlet flanges of the mud pump housing, and 8 microphones are arranged at each flange to form a 16-channel acquisition system; an acoustic wave exciter is fixed at the center of the outer wall of the housing; a sine wave pulse with preset parameters is generated by the acoustic wave exciter, and the microphone array is started at the same time to collect acoustic signals at a preset sampling rate to obtain multiple complete excitation-response cycles; the collected original signal is band-pass filtered to remove low-frequency environmental noise and high-frequency electromagnetic interference; the signals of multiple cycles are superimposed and averaged to obtain a signal with a high signal-to-noise ratio; the averaged signal is short-time Fourier transformed to obtain a time-frequency spectrum diagram to obtain the acoustic response characteristics of the mud pump housing under target excitation, and the acoustic response characteristics include frequency, amplitude and phase information.

3. The defect identification method according to claim 1, characterized in that: in, Based on the results of repeated multiple identifications, the consistency index and stability index are calculated, and the credibility index of the defect identification result is obtained in combination with the confidence level as a reference for defect handling decisions, including: obtaining multiple groups of detection data of the same mud pump casing, each group of detection data includes input features and output probability distribution, and the output probability distribution is generated by the defect recognition model; based on the output probability distribution of the multiple groups of detection data, the consistency index of the defect identification result is calculated by a consistency evaluation formula, and the consistency evaluation formula measures the discrete degree of multiple detection results; adding random noise to the original input features of the same mud pump casing to generate A plurality of groups of disturbance input features are input into the defect recognition model to obtain a plurality of corresponding disturbance output probability distributions; based on the original input features, the original output probability distribution, the plurality of groups of disturbance input features and the plurality of groups of disturbance output probability distributions, a stability index of the defect recognition result is calculated by a stability evaluation formula, and the stability evaluation formula measures the sensitivity of the model to the input disturbance; according to the confidence, consistency index and stability index output by the defect recognition model, a credibility index of the defect recognition result is calculated by a weighted sum formula, and the credibility index is used as a reference for defect handling decisions.

4. The defect identification method according to claim 1, characterized in that: in, Frequency domain energy distribution characteristics, including: ,in, Indicates the frequency The energy distribution characteristics at Indicates frequency, Indicates the number of the microphone. Indicates the total number of microphones, Indicates Microphones at frequency The complex spectrum at Indicates The weighting factor for each microphone.

5. The defect identification method according to claim 1, characterized in that: in, Time-frequency correlation features, including: ,in, Indicates time delay and frequency The correlation characteristics of Indicates time delay, and Indicates the number of the microphone. Indicates the total number of microphones, Indicates the total duration of the signal, Indicates microphones at a time and frequency The time spectrum at Indicates microphones at a time and frequency The conjugate of the time-frequency spectrum at .

6. The defect identification method according to claim 1, characterized in that: in, Modal decomposition features, including: ,in, Indicates The order mode at the frequency The amplitude at represents the order of the mode, Indicates the number of the microphone. Indicates the total number of microphones, Indicates Microphone pair The order mode at the frequency The contribution coefficient at Indicates Microphones at frequency The complex spectrum at .

7. A deep learning-based defect recognition system for a petroleum accessory mud pump housing, applied to the defect recognition method according to any one of claims 1 to 6, characterized in that: include: An acquisition unit, a deep learning unit and a decision unit; wherein the acquisition unit is used to acquire the acoustic original signal of the inner cavity of the mud pump housing, and pre-process the original signal to obtain the acoustic response characteristics of the mud pump housing under the target excitation; the frequency domain energy distribution characteristics, time-frequency correlation characteristics and modal decomposition characteristics of the acoustic response characteristics are calculated, and discretization processing is performed within a preset frequency range to obtain a feature vector; the deep learning unit is used to input the feature vector into a trained deep learning model based on a convolutional neural network and a long short-term memory network to obtain the probability distribution of various defect types of the mud pump housing; the decision unit is used to compare the sizes of all probability values ​​in the probability distribution, select the category corresponding to the maximum probability value as the primary recognition result, and calculate the confidence of the recognition result to compare the confidence with the target threshold; based on the results of repeated multiple recognitions, the consistency index and stability index are calculated, and the credibility index of the defect recognition result is obtained in combination with the confidence to serve as a reference for defect disposal decisions.

Images