Pump valve failure detection method

Abstract

This invention provides a pump and valve fault detection method, which includes: constructing a pump and valve fault index system reflecting the characteristics of vibration signals indicative of the pump and valve's state; establishing an LSTM-AE model using an LSTM deep model and based on the autoencoder principle, and setting a fault discrimination threshold; inputting training data into the LSTM-AE model based on the pump and valve fault index system to train the LSTM-AE model; and inputting the detection data of the pump and valve to be detected into the trained LSTM-AE model based on the pump and valve fault index system, comparing the output value obtained by the LSTM-AE model with the set fault discrimination threshold to determine whether the detected pump and valve has a fault, and outputting a first judgment result. The method of this application improves the universality and accuracy of the model in practical applications.

Description

Technical Field

[0001] This invention relates to the field of fault detection, and more specifically, to a method for detecting pump and valve faults. Background Technology

[0002] Reciprocating pumps, such as plunger pumps, often have complex structures and numerous vibration signal sources, making detection and maintenance relatively difficult. The suction and discharge valves are typically the most maintenance-intensive components in pump equipment. Valve failure is considered a major cause of unexpected pump downtime. Current research on pump valve fault detection technologies involves building algorithmic models using pressure, temperature, and vibration signals. Due to the high-pressure environment inside the pump body, installing in-cylinder sensors is difficult. External sensors typically capture vibration signals to create algorithmic models for fault detection, resulting in higher accuracy. Existing research on fault detection models using vibration signals mainly includes: setting thresholds based on signals collected by vibration sensors, and using neural network structures such as Naive Bayes classification, backpropagation (BP), CNN, autoencoders, and LSTM for fault detection. However, most of this research is based on laboratory environments and has not been applied to actual production operations. Furthermore, the algorithmic model indicator systems in these studies are relatively incomplete, and the prediction methods are often simplistic, leading to low prediction accuracy. Summary of the Invention

[0003] Technical problems to be solved

[0004] However, in the existing technology for fault detection of pumps and valves such as plunger pumps as described above, the following problems exist:

[0005] 1. Detection methods based on neural network structures often do not consider both time-domain and frequency-domain features when constructing the model input indicator system, resulting in a relatively incomplete indicator system construction.

[0006] 2. Most detection methods do not consider the temporal fluctuations of characteristic indicators during the signal failure process and do not incorporate them into the signal characteristic indicator system.

[0007] 3. The neural network model used is only a single network structure, and the combination of multiple network structures is not considered to improve prediction accuracy.

[0008] 4. Many studies are based on laboratory environments and lack real-world equipment operation data, resulting in incomplete consideration of influencing factors and making it impossible to apply model predictions to actual industrial production operations.

[0009] 5. Lack of practical expert experience to assist in judgment, such as pump and valve replacement logs related to actual equipment operation on-site, and experience summaries from technical backbone experts.

[0010] To address the aforementioned issues, this invention provides a pump and valve fault detection method. This method establishes a more comprehensive and accurate signal feature index system, improving the model's prediction accuracy. Furthermore, this detection method considers the temporal fluctuation trend of signal features and combines LSTM and autoencoder network structures, thereby enhancing the model's versatility and accuracy in practical applications.

[0011] Technical solution

[0012] To achieve the above objectives, a pump valve fault detection method is provided. This method includes: constructing a pump valve fault index system that reflects the characteristics of vibration signals reflecting the pump valve's state; establishing an LSTM-AE model using an LSTM deep model and based on the autoencoder principle, and setting a fault discrimination threshold; inputting training data into the LSTM-AE model based on the pump valve fault index system to train the LSTM-AE model; and inputting the detection data of the pump valve to be detected into the trained LSTM-AE model based on the pump valve fault index system, comparing the output value obtained by the LSTM-AE model with the set fault discrimination threshold to determine whether the detected pump valve has a fault, and outputting a first judgment result.

[0013] Furthermore, the pump and valve fault index system that reflects the characteristics of vibration signals in the pump and valve state includes: constructing time-domain statistical indicators that reflect the time-domain characteristics of vibration signals based on vibration signals; and constructing frequency-domain indicators that reflect the spectral differences of vibration signals based on vibration signals.

[0014] Furthermore, the construction of a pump and valve fault index system that reflects the characteristics of vibration signals in the pump and valve state further includes: obtaining sensitive statistical indicators for pump and valve faults by comparing the sensitivity of time-domain statistical indicators; selecting frequency-domain characteristic indicators from the frequency-domain indicators by calculating the frequency-domain energy and the proportion of frequency-domain energy of the vibration signal; and constructing a pump and valve fault index system based on the sensitive statistical indicators for pump and valve faults and the selected frequency-domain characteristic indicators.

[0015] Furthermore, the pump and valve fault detection method also includes: constructing an expert experience base and processing the text data of the expert experience base into structured text, wherein the expert experience base includes daily maintenance logs of pump and valve structures, equipment maintenance specifications, after-sales maintenance records, and experience summaries of technical personnel; and judging whether the pump and valve have malfunctioned based on the detection data of the pump and valve to be tested according to the expert experience base, and outputting a second judgment result; and obtaining a final judgment result for judging whether the pump and valve have malfunctioned based on the first judgment result and the second judgment result.

[0016] Furthermore, if the first judgment result is the same as the second judgment result, then the first judgment result or the second judgment result is output as the final judgment result; if the first judgment result determines that the pump valve is normal and the second judgment result determines that the pump valve is faulty, then the first judgment result is output as the final judgment result; if the first judgment result determines that the pump valve is faulty and the second judgment result determines that the pump valve is normal, then the fault alarm is continuously identified backward and the trend of fault occurrence is judged forward to determine whether the pump valve has malfunctioned.

[0017] Furthermore, if the subsequent continuous identification of fault alarms is ongoing, and the forward judgment shows a trend of fault occurrence, then the pump valve is determined to be faulty and this is taken as the final judgment result.

[0018] Furthermore, the pump valve fault detection method also includes: acquiring vibration signals by using a vibration sensor installed on the pump valve to collect data of the pump valve under fault conditions and under normal conditions.

[0019] Furthermore, the statistical indicators sensitive to pump and valve failures include kurtosis, root mean square (RMS), peak value, and pulse value.

[0020] Furthermore, the statistical indicators sensitive to pump and valve failures include root mean square value, peak value, pulse value, and skewness value.

[0021] Further, setting the fault discrimination threshold includes: setting a threshold interval [err_train_min, err_train_max*1000], where err_train_min represents the minimum mean square error of the training data and err_train_max represents the maximum mean square error of the training data; dividing the above interval into equal intervals, and extracting the threshold one by one based on the interval, calculating the values ​​of the evaluation metrics AUC and F1 of the LSTM-AE model at the extracted threshold; and determining the value or range of the fault discrimination threshold based on the obtained evaluation metrics AUC and F1.

[0022] Furthermore, the construction of time-domain statistical indicators that reflect the time-domain characteristics of vibration signals based on vibration signals includes: constructing time-domain statistical indicators by calculating indicators such as mean, absolute mean, variance, standard deviation, root square amplitude, root mean square value, peak value, maximum value, minimum value, waveform index, peak value index, impulse index, margin index, skewness index, and kurtosis index based on the time-domain characteristics of vibration signals.

[0023] Furthermore, constructing a frequency domain index that reflects the spectral difference characteristics of the vibration signal based on the vibration signal includes: performing EEMD decomposition on the vibration signal to obtain multiple components; and performing FFT transformation on each component to obtain the frequency domain signal of each component.

[0024] Furthermore, the frequency domain index is screened by calculating the sum of the frequency domain energy and the proportion of the frequency domain energy of the vibration signal. This includes: calculating the sum of the frequency domain energy and the proportion of the frequency domain energy of each component signal; sorting the frequency domain signals of each component according to the frequency domain energy and proportion; and using the sum of the frequency domain energy of the top 80% of the sorted component signals as the frequency domain characteristic index.

[0025] According to another aspect of this application, a computer device is provided, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the steps of the above-described pump and valve fault detection method.

[0026] According to another aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above-described pump and valve fault detection method.

[0027] Beneficial effects

[0028] The pump and valve fault detection method of this application considers signal feature extraction from both the time domain and the frequency domain. Furthermore, the frequency domain features are decomposed into signals and fault features are maximized, which can establish a more complete and accurate signal feature index system and improve the prediction accuracy of the detection model.

[0029] Furthermore, the pump and valve fault detection method of this application considers the temporal fluctuation trend of signal characteristics and combines two network structures, LSTM and autoencoder, to improve the universality and accuracy of the model in practical applications.

[0030] Furthermore, the pump and valve fault detection method of this application, through the structured processing of historical work logs and experience summaries of existing technicians and experts, and the generation of an expert experience database, has accumulated rich knowledge on equipment maintenance and subsequent equipment upgrades and modifications.

[0031] Furthermore, the pump and valve fault detection method of this application is based on a dual-track verification detection model of LSTM-AE model and expert experience base, which makes the pump and valve fault prediction more accurate, the risk of misjudgment less, and the practical applicability stronger. Attached Figure Description

[0032] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0033] Figure 1 A flowchart of a pump valve fault detection method according to an embodiment of this application is shown;

[0034] Figure 2A comparative experimental diagram showing the kurtosis factor fluctuation when screening time-domain statistical indicators by sensitivity comparison according to an embodiment of this application is shown.

[0035] Figure 3 A comparative experimental diagram is shown, illustrating the fluctuation of the time margin factor of time-domain statistical indicators by sensitivity comparison according to an embodiment of this application;

[0036] Figure 4 The experimental fluctuation curves of AUC and F1 values ​​during the fault discrimination threshold setting process in the AE (autoencoder) algorithm according to an embodiment of this application are shown.

[0037] Figure 5 The internal structure and training process of an LSTM-AE model according to an embodiment of this application are illustrated schematically.

[0038] Figure 6 The internal structure of an LSTM-AE model according to an embodiment of this application and its testing process are illustrated schematically.

[0039] Figure 7 This illustration schematically depicts the process of constructing a regular expression library and a keyword mining process according to embodiments of this application; and

[0040] Figure 8 A flowchart of a pump valve fault detection method combining an LSTM-AE model and an expert experience base, according to another embodiment of this application, is shown. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0042] like Figure 1 As shown, according to a preferred embodiment of this application, a pump valve fault detection method is provided. The method includes: constructing a pump valve fault index system that reflects the characteristics of vibration signals reflecting the pump valve state (S101); establishing an LSTM-AE model using an LSTM deep model and based on the autoencoder principle, and setting a fault discrimination threshold (S102); inputting training data into the LSTM-AE model based on the pump valve fault index system to train the LSTM-AE model (S103); and inputting the detection data of the pump valve to be detected into the trained LSTM-AE model based on the pump valve fault index system, comparing the output value obtained by the LSTM-AE model with the set fault discrimination threshold, thereby determining whether the detected pump valve has a fault, and outputting a first judgment result (S104).

[0043] Furthermore, the pump and valve fault detection method according to this application improves the universality and accuracy of the model application by combining LSTM and autoencoder networks.

[0044] The following section will describe in detail the construction of the pump and valve fault index system, the construction method of the LSTM-AE model, and the method of using the LSTM-AE model according to this application to detect pump and valve faults.

[0045] According to a preferred embodiment of this application, a vibration sensor is installed on a pump valve (e.g., a plunger pump valve), and a signal acquisition device is used to collect data of the vibration signal of the pump valve under fault conditions and under normal conditions as raw data, such as the amplitude of vibration.

[0046] The following example, a five-cylinder plunger pump, illustrates how to acquire vibration signals using a vibration sensor.

[0047] According to a preferred embodiment of this application, a unidirectional acceleration vibration sensor is used, with a sampling frequency set to 10 kHz. One vibration sensor is installed on each of the five cylinder upper and lower valves on the hydraulic end of the plunger pump to collect vibration data from each valve. Specifically, ten unidirectional acceleration vibration sensors are installed on the ten check valves on the hydraulic end, with the specific installation locations shown in Table 1.

[0048]

[0049]

[0050] Table 1

[0051] Of course, other vibration sensors, different from the unidirectional acceleration vibration sensors mentioned above, can be used to collect vibration signals at the required frequency.

[0052] With the above settings, we can obtain enough data to truly reflect the state of the pump and valve, and thus, we can use this data to better train the LSTM-AE model.

[0053] According to a preferred embodiment of this application, the vibration signal obtained above is preprocessed to obtain a time-domain statistical index that reflects the time-domain characteristics of the vibration signal data.

[0054] According to a preferred embodiment of this application, time-domain statistical indicators are constructed by calculating, but not limited to, 15 indicators, including, but not limited to, mean, absolute mean, variance, standard deviation, root mean square amplitude, root mean square value, peak value, maximum value, minimum value, waveform index, peak index, impulse index, margin index, skewness index, and kurtosis index, based on the time-domain characteristics of the vibration signal. These time-domain statistical indicators are then normalized. The data normalization standard employs the MIN-MAX normalization method.

[0055] Furthermore, according to a preferred embodiment of this application, the acquired vibration signal is subjected to EEMD (Ensemble Empirical Mode Decomposition), that is, white noise is added to the original spectral signal x(t) to uniformly adjust the original signal distribution, and then EMD (Empirical Mode Decomposition) is performed to generate n components imf, i.e., x(t) = {imf1(t), imf2(t), ..., imf n (t)}. Furthermore, FFT (Fast Fourier Transform) is performed on each component to obtain the frequency domain signal of each component.

[0056] Therefore, a frequency domain index reflecting the spectral differences of vibration signals is constructed based on the vibration signals.

[0057] Preferably, the sensitive statistical indicators for pump and valve failure are obtained by comparing the sensitivity of the above-mentioned time-domain statistical indicators.

[0058] Specifically, a sensitivity comparison was conducted on the aforementioned 15 time-domain statistical indicators. This involved dividing the collected vibration signal data into segments with a given time window, calculating the values ​​of each time-domain statistical indicator within each time window, and then comparing the temporal fluctuation sensitivity of the same time-domain statistical indicator under the same coordinate system in the normal and fault states of the pump and valve. Time-domain statistical indicators with clear difference boundaries and value ranges were retained as pump and valve fault sensitive statistical indicators, while other statistical indicators were not considered as pump and valve fault sensitive statistical indicators for this model.

[0059] Based on the above method, draw the time-series fluctuation sensitivity diagrams of each of the 15 time-domain statistical indicators under the normal and fault states of the pump and valve in the same coordinate system.

[0060] Taking kurtosis factor and margin factor as examples, Figure 2 and Figure 3 This figure shows the time-series fluctuation sensitivity diagrams of kurtosis factor and margin factor under normal and fault conditions of pumps and valves, respectively, on the same coordinate system. As can be seen from the figure, the x-axis represents the number of consecutive tests, with each test sequentially extracted in time order, for example, 10 seconds of vibration signal data, to calculate the kurtosis factor and margin factor, and their values ​​are used as the y-axis values. As shown in the figure, a comparison diagram of the fluctuation trends and magnitudes of kurtosis factor and margin factor under faulty and normal equipment conditions is constructed in the time domain.

[0061] like Figure 2 and Figure 3As shown, the overall value range of the margin index and kurtosis index under fault conditions is much larger than that under normal conditions, and there is a clear dividing line between the two. Therefore, the margin index and / or kurtosis index can be selected as statistical indicators sensitive to pump and valve faults.

[0062] By comparing fault sensitivity using the above method, time-domain statistical indicators that are more sensitive to pump and valve faults are obtained as pump and valve fault sensitivity statistical indicators. Preferably, these pump and valve fault sensitivity statistical indicators may include kurtosis indicators. Root mean square value Peak indicators and pulse index Alternatively, these pump and valve failure sensitivity statistics could include the root mean square value. Peak indicators Pulse Indicators and skewness index It should be understood that those skilled in the art, upon reading this article, can obviously select any suitable combination of index data from the above 15 time-domain statistical indicators as sensitive statistical indicators for pump and valve failures, as needed.

[0063] According to a preferred embodiment of this application, the frequency domain signals of each component obtained above... The formulas for calculating the sum of frequency domain energies and their proportions are as follows:

[0064]

[0065]

[0066] Among them, |imf' ji | represents the magnitude of the discrete frequency corresponding to the j-th imf component, E j E represents the sum of the energies of the j-th imf component. all E' is the total energy of the original vibration signal after FFT transformation. j This represents the energy of the j-th imf component and its proportion to the total energy of the original vibration signal.

[0067] Then, based on the frequency domain energy and proportion, each IMF component is ranked in descending order of energy and proportion. IMF components with frequency domain energy and ranking in the bottom 20% are removed. The sum of the frequency domain energy of the remaining IMF components is used as the frequency domain characteristic index of the original signal, i.e., {E1, E2, ..., E...}. k That is, the sum of the frequency domain energy of the top 80% of the IMF component frequency domain signals is used as the frequency domain characteristic index of the original signal, where k is the number of retained IMF components. The sum of the frequency domain energy of the IMF component frequency domain signals is then normalized.

[0068] As described above, frequency domain characteristic indicators are selected from the frequency domain indicators by calculating the frequency domain energy and the proportion of the frequency domain energy of the vibration signal.

[0069] Based on the pump and valve failure sensitivity statistical indicators obtained through the above methods and the selected frequency domain characteristic indicators, a pump and valve failure indicator system is constructed:

[0070] The following describes, based on a preferred embodiment of the present application, the use of an LSTM deep model and the establishment of an LSTM-AE model according to the principle of an autoencoder.

[0071] The LSTM-AE model structure according to this application comprises two parts: an encoder and a decoder. Both the encoder and decoder are network structures generated through model tuning based on sample data training. According to a preferred embodiment of this application, the encoder and decoder each comprise two LSTM recurrent network layers and one fully connected layer.

[0072] Preferably, the LSTM layer in the Encoder contains 15-60 neurons, and the fully connected layer contains 10-20 neurons.

[0073] Preferably, the LSTM layer in the Decoder contains 15-60 neurons, and the fully connected layer contains neurons (number of input sample features * feedforward time step).

[0074] The specific number of neurons in each layer can be determined through iterative tuning within the range mentioned above, based on the accuracy predicted from the sample data. Additionally, dropout is set to >0 for each LSTM layer to randomly select neurons, preventing overfitting during model training.

[0075] Furthermore, preferably, the loss function of the LSTM-AE model uses the mean squared error function to calculate the error between the input data and the prediction result.

[0076] Furthermore, a pump and valve fault discrimination threshold is set for the LSTM-AE model to identify pump and valve fault anomalies.

[0077] The threshold is dynamically selected based on the model performance evaluation metrics (such as AUC and F1 scores). First, the threshold range [err_train_min, err_train_max*1000] is set, where err_train_min represents the minimum mean square error of the training data and err_train_max represents the maximum mean square error of the training data. The training data here is based on the pump and valve failure index system mentioned above.

[0078] The upper limit of the above interval can be any number greater than the maximum mean square error, and the lower limit can be any number less than the minimum mean square error. However, considering the validity of the interval and based on experimental results, it is preferable that the interval is taken between the maximum mean square error of the training data and the minimum mean square error of the training data, which are 1000 times the maximum mean square error. The upper and lower limits can be adjusted within a small range.

[0079] Then, the interval is divided into multiple parts at equal intervals, such as 1000 parts, and threshold value tests are conducted for each interval. The AUC value and F1 value of the model prediction results are calculated for each value. The rationality of the threshold value or range of values ​​for pump and valve fault discrimination is determined according to these two evaluation index values ​​(if the AUC and F1 values ​​in a certain interval are continuously close to 1, then the threshold value is the average value of that interval).

[0080] Figure 4 An experimental fluctuation curve of AUC and F1 values ​​during the fault discrimination threshold setting process in the AE (autoencoder) algorithm according to an embodiment of this application is shown. As shown in the figure, the horizontal axis represents the pump valve fault discrimination threshold, and the vertical axis represents the evaluation index value. As shown in the figure, when the AUC value and F1 value approach 1, the average value of this interval can be taken as the pump valve fault discrimination threshold.

[0081] Therefore, according to this application, by setting a threshold interval [err_train_min, err_train_max*1000], dividing the interval into equal intervals, and extracting thresholds one by one based on the interval, the evaluation metrics AUC and F1 values ​​of the LSTM-AE model at the extracted thresholds are calculated, thereby determining a reasonable value or range of fault discrimination thresholds based on the obtained evaluation metrics AUC and F1.

[0082] Therefore, the obtained fault discrimination threshold value or its range is used to determine whether the pump or valve is abnormal. When the test value is greater than the pump or valve fault discrimination threshold, the pump or valve is judged to be faulty; when the test value is less than the pump or valve fault discrimination threshold, the pump or valve is judged to be normal.

[0083] like Figure 5 As shown, it illustrates the internal structure and training process of an LSTM-AE model according to an embodiment of this application.

[0084] The relevant parameters involved in model training are shown in Table 2 below:

[0085]

[0086] Table 2

[0087] As shown in the figure, training data is input into the model in epochs each time.

[0088] For example, the model training data consisted of operational data from a fracturing pump at a well site in Xinjiang. The equipment was a 5000-type electrically driven fracturing skid, with a power end speed of 1900 rpm, a hydraulic end with 5 cylinders, and 10 sets of upper and lower valve assemblies. The experimental data came from vibration signal acquisition of the upper valve assembly near the high-pressure discharge end of cylinder 1 (corresponding to the vibration sensor located at the 'plunger direction inside the discharge cap of cylinder 1' on the hydraulic end), at a frequency of 10 kHz. A total of 60 hours of normal operation data and 2 hours of fault operation data were collected from different well sections within the same well site.

[0089] Vibration signal data of 600,000 points (i.e., 1 minute) were extracted sequentially over time. Based on the aforementioned pump and valve fault index system, the values ​​of each index were calculated, resulting in 3600 rows of normal characteristic index data and 120 rows of fault characteristic index data. Both the normal and fault characteristic index data are based on the aforementioned pump and valve fault index system. Using, for example, a 3-minute time window, samples were slid forward to create a three-dimensional array with temporal trend fluctuation characteristics, i.e., (1, batch_size, (k+4)*3). Here, batch_size represents the length of the three-dimensional array sample, and k is the number of retained IMF components.

[0090] For example, 99% of the normal feature index data obtained above can be used as training data to train the LSTM-AE model.

[0091] The above training data is input, and the training data is processed by the neurons of each layer of the Encoder. The results are output to the Decoder for further processing by the neurons of each layer, and the Decoder outputs the calculation results with the same structure as the training data.

[0092] The loss value is calculated using the mean squared error loss function based on the output calculation results and the input training data, and the gradient information is backpropagated. The network parameters of each layer are updated using Adam (Adaptive Moment Estimation).

[0093] Furthermore, repeat the above training process until the model has been trained for all epochs, and the loss value no longer shows a decreasing trend. Calculate the AUC value after each epoch. When the AUC value approaches 1 and shows no increasing trend, the training is complete, and the trained LSTM-AE model is obtained.

[0094] like Figure 6 As shown, it schematically illustrates the LSTM-AE model testing process according to an embodiment of this application:

[0095] Input test data based on the above model pump and valve failure index system into the model. The test data can use 1% of the normal characteristic index data and failure characteristic index data obtained from the fracturing pump operation at a well site in Xinjiang. The above experimental data are labeled with positive and negative samples, with failure characteristic index data labeled as 1 and normal characteristic index data labeled as 0.

[0096] The test data is processed by the neurons in each layer of the Encoder, and the results are output to the Decoder for further processing by the neurons in each layer. The Decoder then outputs the results with the same structure as the training data.

[0097] The mean square error of the test data is calculated using the mean square error loss function, and the equipment status is determined based on the pump and valve fault discrimination threshold, thus ending the test process.

[0098] The results of the mean square error calculation for the test data are shown in Table 3 below:

[0099] 1.3656744e-10 6.0024617e-11 6.6082535e-11 4.9380777e-11 4.9380777e-11 1.4042054e-10 6.2024413e-11 1.1417222e-10 … … … … 1.5851367e-06 1.4575928e-06 1.4876275e-06 1.4053102e-06 1.2826864e-06 1.3016524e-06 1.4301578e-06 2.2050192e-06 … … … …

[0100] Table 3

[0101] As shown in the table above, the mean square error of the prediction for the first half of the test dataset, which contains data on normal operation of the equipment, is on the order of e⁻¹⁰, while the mean square error of the prediction for the second half, which contains data on faulty operation, is on the order of e⁻⁶. The test results are shown in Table 4 below.

[0102] Data types Test set data Prediction results Normal operating data of equipment 36 groups 31 groups Equipment failure operation data 118 groups Group 123

[0103] Table 4

[0104] As shown in the table above, when identifying anomalies in the test data, 5 sets of data were judged incorrectly, accounting for 3.2% of the total sets. Relatively speaking, the LSTM-AE model has a higher recognition accuracy.

[0105] As can be seen from the above description, this application considers signal feature extraction from both the time domain and the frequency domain, and performs signal decomposition and fault feature maximization in the frequency domain, which can establish a more complete and accurate signal feature index system and improve the model prediction accuracy.

[0106] Furthermore, this application considers the temporal fluctuation trend of signal characteristics and combines two network structures, LSTM and autoencoder, to improve the universality and accuracy of the model in practical applications.

[0107] According to another preferred embodiment of this application, such as Figure 8As shown, the left side illustrates the pump and valve fault detection process based on LSTM-AE as described above, while the right side shows the process combined with an expert experience base. This process combines the LSTM-AE model with the personal operational experience of industry technical backbones and domain experts, as well as a large amount of expert experience base such as equipment technical specifications and equipment operation logs accumulated over the years to identify anomalies in equipment pumps and valves, thereby further improving the accuracy of prediction.

[0108] According to this application, keyword mining and structuring are performed on the personal operational experience of technical backbones and domain experts, as well as a large amount of unstructured text data accumulated over the years, such as equipment technical specifications, daily maintenance logs of pump and valve structures of various pump and valve equipment, equipment maintenance specifications, after-sales maintenance records, experience summary logs of technical personnel, and equipment operation logs, to generate a structured expert experience database.

[0109] Taking the daily maintenance log of the hydraulic end valve of a plunger pump at a well site as an example, this illustrates how to process unstructured text data into standardized structured text data.

[0110] Table 5 below shows the daily maintenance log of the hydraulic end valve of a plunger pump at a well site, in unstructured text format. The specific format is as follows:

[0111]

[0112] Table 5

[0113] For the aforementioned 'replacement record' text, a keyword mining regular expression formula library was constructed, including formulas for 'job duration,' 'job layer number,' and 'replacement quantity.' Using Python programming, keyword text mining was performed on the source text to extract and filter the keywords, and the extracted keywords were then stored in a structured format, such as... Figure 7 As shown in Table 6 below:

[0114]

[0115]

[0116] Table 6

[0117] Based on the keyword mining results mentioned above, an expert experience base for pump and valve fault detection rules was created, thereby obtaining standardized expert experience rules. For example, the specific format is shown in Table 7 below:

[0118]

[0119] Table 7

[0120] The rules in Table 7 are constructed based on the fields in Table 6. Specifically, the rules in Table 7 are constructed based on the fields such as 'working hours' and 'reason for replacement' in Table 6, and are customized according to expert experience.

[0121] As described above, a standardized expert experience database has been constructed. This application has structured and compiled the historical work logs and experience summaries of existing technical personnel and experts to generate an expert experience database, accumulating rich knowledge on equipment maintenance, equipment upgrades and modifications, etc.

[0122] Figure 8 The diagram shows a flowchart of a method for pump and valve fault detection by combining an LSTM-AE model and an expert experience base according to another embodiment of this application. The left side shows the construction process of the above-mentioned pump and valve fault index system and the LSTM-AE model, while the right side shows the construction process of the expert experience base.

[0123] To better understand this preferred embodiment of the application, the fracturing pump operation data obtained above from a well site in Xinjiang will be used as an example for illustration. Figure 8 The pump and valve fault detection process is shown. The construction of the pump and valve fault index system, the LSTM-AE model, and the expert experience base have already been described in detail above, and will not be repeated here.

[0124] As described above, the temporal characteristics of vibration signal data, which is obtained as raw data from the fracturing pump operation data of a well site in Xinjiang, are calculated to obtain temporal statistical indicators that reflect the temporal characteristics of vibration signal data. The obtained temporal statistical indicators are then compared for sensitivity to select sensitive statistical indicators for pump and valve failure.

[0125] Furthermore, EEMD is performed on the vibration signal data to calculate the IMF components, and FFT transformation is performed on each IMF component to obtain the frequency domain signal of each component; and frequency domain characteristic indicators are selected from the frequency domain indicators by calculating the frequency domain energy and the proportion of the frequency domain energy of the vibration signal.

[0126] Data standardization, such as normalization, is performed on the statistical indicators and frequency domain characteristic indicators sensitive to pump and valve failures to construct a pump and valve failure indicator system.

[0127] Normal characteristic index data based on the pump and valve failure index system were obtained from the fracturing pump operation data of a well site in Xinjiang. For example, 99% of the obtained normal characteristic index data was used as training data and input into the LSTM-AE model constructed above. The LSTM-AE model was trained and optimized through the Encoder, Decoder, prediction error calculation module and optimizer module to finally obtain the trained model.

[0128] As described above, by preprocessing unstructured text data, including the personal operational experience of technical backbones and domain experts, as well as a large amount of equipment technical specifications, daily maintenance logs of pump and valve structures, equipment maintenance specifications, after-sales maintenance records, experience summary logs of technical personnel, and equipment operation logs accumulated over many years, a keyword regular expression formula library is constructed as described above. Keywords are extracted from the unstructured text data, and the unstructured text data is transformed into structured text for storage and classification, thereby creating a structured expert experience rule library.

[0129] Using 1% of the normal characteristic index data (36 groups in Table 4) and fault characteristic index data (118 groups in Table 4) obtained from fracturing pump operations at a well site in Xinjiang as test data, the data were input into the trained LSTM-AE model. Based on the test results from the LSTM-AE model (see the prediction results in Table 4), it can be seen that 5 groups of data were judged incorrectly.

[0130] Furthermore, for the above five sets of false negative experimental data, the constructed expert experience base is used to determine whether the pump valve has malfunctioned based on the test data of the pump valve to be tested (i.e., the above five sets of false negative experimental data), and the judgment result is output. Specifically, the relevant field data in the fault index system of the five sets of prediction error data are matched with all rules in the expert experience base, and a judgment is made based on the rule matching results of the expert experience base.

[0131] Furthermore, based on the judgment results from the LSTM-AE model (i.e., the first judgment result) and the judgment results based on the expert experience base (i.e., the second judgment result), a final judgment result is obtained to determine whether the pump valve has malfunctioned.

[0132] If the judgment result based on the LSTM-AE model is the same as the judgment result based on the expert experience base, then the judgment result based on the LSTM-AE model or the judgment result based on the expert experience base shall be the final judgment result.

[0133] If the judgment result based on the LSTM-AE model determines that the pump valve is normal, while the judgment result based on the expert experience base determines that the pump valve is faulty, then the judgment result based on the LSTM-AE model shall be the final judgment result.

[0134] If the LSTM-AE model determines the pump / valve is faulty, while the expert experience base determines it is normal, the system continuously identifies whether fault alarms are ongoing and checks for fault trends to determine if a pump / valve has malfunctioned. Specifically, starting with the moment the LSTM-AE model identifies an anomaly, K time intervals are continuously taken forward or backward at a fixed time step T. The LSTM-AE model predicts the equipment status for each of the K time intervals. Based on the fault prediction results for K+1 consecutive time intervals, the system determines whether the fault is a true fault based on whether the fault conditions are continuous.

[0135] According to one embodiment of this application, the dual-track verification results using LSTM-AE and expert experience base follow the following rules (as shown in Table 8):

[0136]

[0137]

[0138] Table 8

[0139] Therefore, regarding the five sets of false negative experimental data tested based on the LSTM-AE model (i.e., experimental data that the LSTM-AE model judged as faulty), two sets were judged as normal by the expert experience base. Therefore, the LSTM-AE model further continuously identifies whether the fault alarm continues and judges whether there is a fault trend. Combining the verification results of the LSTM-AE model and the expert experience base, these two sets were ultimately judged as normal. The other three sets of false negative experimental data were judged as abnormal by the expert experience base, indicating that these three sets of data could not be accurately identified as faults.

[0140] The results above show that the accuracy of the prediction results has been further improved with the assistance of the expert experience base. Therefore, the dual-track verification method using LSTM-AE and the expert experience base has high predictive accuracy and practicality for fault identification of hydraulic end pumps and valves in fracturing equipment.

[0141] Therefore, the dual-track verification detection model based on the LSTM-AE model and expert experience base, as described in this application, achieves higher prediction accuracy, lower risk of misjudgment, and stronger practical applicability.

[0142] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0143] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A pump valve failure detection method, characterized by, The method includes: Construct a pump and valve fault index system that reflects the characteristics of vibration signals in the pump and valve state; An LSTM-AE model is established using an LSTM deep model and based on the principle of autoencoders, and a fault discrimination threshold is set. Based on the aforementioned pump and valve fault index system, training data is input into the LSTM-AE model to train the LSTM-AE model; and Based on the aforementioned pump and valve fault index system, the detection data of the pump and valve to be detected is input into the trained LSTM-AE model, and the output value obtained by the LSTM-AE model is compared with the set fault discrimination threshold to determine whether the detected pump and valve has a fault, and outputs a first judgment result for determining whether the pump and valve has a fault; wherein, the construction of the pump and valve fault index system reflecting the characteristics of the vibration signal of the pump and valve state includes: Based on the vibration signal, a time-domain statistical index reflecting the time-domain characteristics of the vibration signal is constructed, and a pump and valve fault sensitivity statistical index is obtained by comparing the sensitivity of the time-domain statistical index. Based on the vibration signal, a frequency domain index reflecting the spectral difference characteristics of the vibration signal is constructed. By calculating the frequency domain energy and the proportion of the frequency domain index of the vibration signal, the frequency domain signals of each component are sorted according to the frequency domain energy and proportion, and frequency domain feature indexes are selected from the frequency domain index. The pump and valve fault index system is constructed based on the aforementioned pump and valve fault sensitivity statistical indexes and the selected frequency domain feature indexes.

2. The pump and valve fault detection method according to claim 1, characterized in that, The method further includes: An expert experience database is constructed, and the text data in the database is processed into structured text. This database includes daily maintenance logs for pumps and valves, equipment maintenance specifications, after-sales maintenance records, and summaries of technical personnel's experience. Based on the detection data of the pump valve to be tested, the expert experience base is used to determine whether the pump valve has malfunctioned, and a second judgment result is output. Based on the first judgment result and the second judgment result, a final judgment result is obtained to determine whether the pump valve has malfunctioned.

3. The pump and valve fault detection method according to claim 2, characterized in that, If the first judgment result is the same as the second judgment result, then output the first judgment result or the second judgment result as the final judgment result; If the first judgment result determines that the pump valve is normal, and the second judgment result determines that the pump valve is faulty, then the first judgment result is output as the final judgment result. If the first judgment result determines that the pump valve is faulty, and the second judgment result determines that the pump valve is normal, then the system continuously identifies whether the fault alarm continues and determines whether there is a fault trend in the future, in order to determine whether the pump valve has malfunctioned.

4. The pump and valve fault detection method according to claim 3, characterized in that, If the fault alarm is continuously detected backwards and there is a trend of the fault occurring forwards, then the pump valve is determined to be faulty and this is taken as the final judgment result.

5. The pump and valve fault detection method according to claim 1, characterized in that, The method further includes: The vibration signal is obtained by collecting data of the pump valve under fault conditions and normal conditions using a vibration sensor installed on the pump valve.

6. The pump and valve fault detection method according to claim 1, characterized in that, The statistical indicators sensitive to pump and valve failures include kurtosis, root mean square, peak value, and pulse.

7. The pump and valve fault detection method according to claim 1, characterized in that, The statistical indicators sensitive to pump and valve failures include root mean square value, peak value, pulse value, and skewness value.

8. The pump and valve fault detection method according to claim 1, characterized in that, Setting fault detection thresholds includes: Set the threshold interval [err_train_min, err_train_max*1000], where err_train_min represents the minimum mean square error of the training data, and err_train_max represents the maximum mean square error of the training data; The intervals are divided into equal parts, and thresholds are extracted for each interval. The AUC and F1 values ​​of the LSTM-AE model at each extracted threshold are then calculated. The value or range of the fault discrimination threshold is determined based on the obtained evaluation indices AUC and F1.

9. The pump and valve fault detection method according to claim 1, characterized in that, The time-domain statistical indicators that reflect the time-domain characteristics of the vibration signal, based on the vibration signal, include: The time-domain statistical index is constructed by calculating the time-domain characteristics of the vibration signal, including the mean, absolute mean, variance, standard deviation, root square amplitude, root mean square value, peak value, maximum value, minimum value, waveform index, peak index, impulse index, margin index, skewness index, and kurtosis index.

10. The pump and valve fault detection method according to claim 1, characterized in that, The frequency domain indices that reflect the spectral differences of the vibration signals, constructed based on the vibration signals, include: The vibration signal was decomposed using EEMD to obtain multiple components; and Perform FFT transformation on each component to obtain the frequency domain signal of each component.

11. The pump and valve fault detection method according to claim 10, characterized in that, By calculating the frequency domain energy and proportion of the frequency domain indices of the vibration signal, and sorting the frequency domain signals of each component according to the frequency domain energy and proportion, frequency domain characteristic indices are selected from the frequency domain indices, including: Calculate the sum of the frequency domain energy and the proportion of the frequency domain energy of each component signal; The frequency domain signals of each component are sorted according to their frequency domain energy and proportion, and the sum of the frequency domain energy of the top 80% of the sorted component frequency domain signals is used as the frequency domain characteristic index.

12. A computer device, comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 11.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 11.

Images