Unsupervised plunger pump anomaly detection method and device based on time-dependent feature reconstruction

Abstract

The application provides a time correlation feature reconstruction unsupervised plunger pump anomaly detection method and device, and relates to the technical field of intelligent detection, which comprises the following steps: statistical feature extraction and feature screening are performed on the current steady-state operation vibration signal of a target plunger pump to obtain the current correlation statistical features of the target plunger pump; the current correlation statistical features are input into a detection model to obtain a fault prediction result of the target plunger pump; and fault early warning information of the target plunger pump is generated according to the fault prediction result; wherein the detection model is obtained by unsupervised training of an initial model based on the historical correlation statistical features of a first sample plunger pump; and the initial model is constructed based on a variational autoencoder and a long short-term memory network. The application automatically performs plunger pump anomaly detection based on the unsupervised detection model trained based on the fusion of the variational autoencoder and the long short-term memory network, thereby effectively improving the detection efficiency of the plunger pump and the accuracy of the detection result.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology, and in particular to an unsupervised method and apparatus for detecting anomalies in a plunger pump based on time-related feature reconstruction. Background Technology

[0002] In practical engineering applications, once a piston pump malfunctions, workers often stop the entire hydraulic system and replace the piston pump, rather than repairing it after determining the cause of the failure. This is because piston pumps are consumable parts, and replacing the entire pump is often more cost-effective and less efficient than disassembling and repairing it. Therefore, timely and accurate detection of abnormal piston pump conditions, allowing for early intervention and preventing damage to the entire hydraulic system due to a single piston pump malfunction, is of great significance in engineering practice.

[0003] Traditional methods for diagnosing plunger pump faults typically rely on expert experience and knowledge, which is not only inefficient but also limited in accuracy.

[0004] Therefore, there is an urgent need to provide an unsupervised method and apparatus for detecting anomalies in piston pumps based on time-related feature reconstruction to solve the above problems. Summary of the Invention

[0005] This invention provides an unsupervised method and apparatus for detecting anomalies in plunger pumps based on time-related feature reconstruction, which solves the defects of low detection efficiency and low accuracy in the prior art, and achieves efficient and accurate detection of plunger pumps.

[0006] This invention provides an unsupervised method for detecting anomalies in a plunger pump based on time-related feature reconstruction, comprising:

[0007] Statistical features are extracted from the current steady-state vibration signal of the target plunger pump to obtain the multidimensional current operating statistical features of the target plunger pump;

[0008] By performing feature filtering on the current operating statistical features described in the multidimensional analysis, the current associated statistical features of the target plunger pump are obtained.

[0009] The current correlation statistical features are input into the detection model to obtain the fault prediction result of the target plunger pump;

[0010] Based on the fault prediction results, fault warning information for the target plunger pump is generated;

[0011] The detection model is obtained by unsupervised training of the initial model based on the historical correlation statistical features of the first sample plunger pump; the historical correlation statistical features are obtained by extracting and filtering statistical features from the historical steady-state vibration signals of the first sample plunger pump; the initial model is constructed based on variational autoencoder (VAE) and long short-term memory network (LSTM) to mine reconstructed features related to feature time information; the first sample plunger pump is a sample plunger pump in normal operation.

[0012] According to the present invention, an unsupervised plunger pump anomaly detection method based on time-related feature reconstruction is provided, wherein the detection model is trained based on the following steps:

[0013] Statistical feature extraction and feature filtering were performed on the historical steady-state operation vibration signal of the first sample plunger pump to obtain the historical correlation statistical features of the first sample plunger pump.

[0014] The historical correlation statistical features are input into the initial model to obtain the first encoded features output by the encoding layer of the initial model and the first reconstructed features output by the decoding layer of the initial model;

[0015] With the goal of minimizing the reconstruction error between the historical correlation statistical features and the first reconstructed features, and maximizing the posterior probability of the first encoded features, the initial model is iteratively trained to obtain the detection model.

[0016] According to the present invention, an unsupervised plunger pump anomaly detection method based on time-related feature reconstruction is provided, wherein the encoding layer of the initial model is constructed based on the encoder of the VAE, the first feature embedding layer, and the LSTM; and the decoding layer of the initial model is constructed based on the second feature embedding layer and the decoder of the VAE.

[0017] The step of inputting the historical correlation statistical features into the initial model to obtain the first encoded feature output by the encoding layer of the initial model and the first reconstructed feature output by the decoding layer of the initial model includes:

[0018] The historical correlation statistical features are input into the encoder to obtain the initial encoded features of the historical correlation statistical features output by the encoder;

[0019] The initial encoded features are input into the first feature embedding layer to obtain the embedded features of the initial encoded features output by the first feature embedding layer;

[0020] The embedding features of the initial encoded features are input into the LSTM to obtain the first encoded features output by the LSTM;

[0021] The first encoded feature is input into the second feature embedding layer to obtain the embedded feature of the first encoded feature output by the second feature embedding layer;

[0022] The embedding features of the first encoded feature are input into the decoder to obtain the first reconstructed feature output by the decoder.

[0023] According to the unsupervised plunger pump anomaly detection method based on time-related feature reconstruction provided by the present invention, the step of acquiring the current steady-state operating vibration signal includes:

[0024] The motor speed, motor torque, and pressure of the fracturing truck to which the target plunger pump belongs are obtained at multiple moments within the current cycle.

[0025] Based on the information on the changes in motor speed, the information on the changes in motor torque, and the information on the changes in pressure at multiple times, the steady-state operating time of the target plunger pump is obtained within the current cycle.

[0026] The vibration signals of the target plunger pump at multiple moments during the steady-state operation time are extracted to obtain the current steady-state operation vibration signal.

[0027] According to the unsupervised plunger pump anomaly detection method based on time-related feature reconstruction provided by the present invention, the method involves extracting statistical features from the current steady-state vibration signal of the target plunger pump to obtain multidimensional current operational statistical features of the target plunger pump, including:

[0028] The current steady-state vibration signal is denoised based on the Savitzky-Golay digital filter.

[0029] The noise-reduced current steady-state vibration signal is processed by sliding window to obtain multiple window data.

[0030] Multi-dimensional statistical features are extracted from the data of each window to obtain the multi-dimensional current operating statistical features of the target plunger pump.

[0031] The multidimensional current operating statistical features include the maximum value, minimum value, mean, peak-to-peak value, absolute mean, root mean square value, root mean square amplitude, standard deviation, kurtosis, skewness, margin index, waveform index, impulse index, peak index, kurtosis index, frequency center, average frequency, frequency domain root mean square, frequency variance, and the multidimensional energy spectrum features of the last layer of wavelet packet decomposition.

[0032] According to the present invention, an unsupervised plunger pump anomaly detection method based on time-related feature reconstruction includes the following step: feature filtering of the multidimensional current operating statistical features to obtain the current associated statistical features of the target plunger pump.

[0033] Statistical features were extracted from the historical steady-state vibration signal of the second sample plunger pump to obtain the multidimensional historical operational statistical features of the second sample plunger pump; the second sample plunger pump is a sample plunger pump of the same type as the target plunger pump.

[0034] Calculate the monotonicity index of the historical operational statistical characteristics of each dimension;

[0035] Based on the monotonicity index, at least one-dimensional candidate statistical features are selected from the multidimensional historical operation statistical features.

[0036] Calculate the Pearson correlation coefficient, Spearman correlation coefficient, and Kendall coefficient among at least one of the candidate statistical features;

[0037] The Pearson correlation coefficient, the Spearman correlation coefficient, and the Kendall coefficient are fused to obtain at least one-dimensional correlation index among the candidate statistical features;

[0038] Based on the correlation index, the historical correlation statistical features of the second sample plunger pump are selected from at least one-dimensional candidate statistical features.

[0039] Among the current operational statistical features described in the multidimensional analysis, statistical features that match the historical correlation statistical features are selected as the current correlation statistical features.

[0040] According to the present invention, an unsupervised plunger pump anomaly detection method based on time-related feature reconstruction includes inputting the current correlation statistical features into the detection model to obtain the fault prediction result of the target plunger pump, comprising:

[0041] The current correlation statistical features are input into the detection model to obtain the second reconstructed features output by the encoding layer of the detection model;

[0042] The reconstruction error between the second reconstructed feature and the current associated statistical feature is compared with the target error to obtain the fault prediction result of the target plunger pump.

[0043] The present invention also provides an unsupervised plunger pump anomaly detection device based on time-related feature reconstruction, comprising:

[0044] The feature extraction unit is used to extract statistical features from the current steady-state vibration signal of the target plunger pump to obtain the multidimensional current operating statistical features of the target plunger pump.

[0045] The feature filtering unit is used to filter the current operating statistical features in multiple dimensions to obtain the current associated statistical features of the target plunger pump.

[0046] The detection unit is used to input the current correlation statistical features into the detection model to obtain the fault prediction result of the target plunger pump;

[0047] The early warning unit is used to generate early warning information for the target plunger pump based on the fault prediction results.

[0048] The detection model is obtained by unsupervised training of the initial model based on the historical correlation statistical features of the first sample plunger pump; the historical correlation statistical features are obtained by extracting and filtering statistical features from the historical steady-state vibration signals of the first sample plunger pump; the initial model is constructed based on variational autoencoder (VAE) and long short-term memory network (LSTM) to mine reconstructed features related to feature time information; the first sample plunger pump is a sample plunger pump in normal operation.

[0049] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements an unsupervised plunger pump anomaly detection method as described in any of the above-described time-related feature reconstruction methods.

[0050] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an unsupervised plunger pump anomaly detection method for time-related feature reconstruction as described above.

[0051] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements an unsupervised plunger pump anomaly detection method for time-related feature reconstruction as described above.

[0052] This invention provides an unsupervised plunger pump anomaly detection method and apparatus based on time-related feature reconstruction. It extracts and filters statistical features from the current steady-state vibration signal of the target plunger pump online to obtain high-quality, lightweight current correlation statistical features. An unsupervised detection model trained based on a variational autoencoder and a long short-term memory network is then used to reconstruct these current correlation statistical features. The reconstruction loss of the detection model is used to obtain the target plunger pump's fault prediction result. Based on the fault prediction result, the plunger pump is automatically and accurately warned of faults. The entire detection process is completed online in real time, minimizing manual intervention and operation, effectively reducing the workload and cost of manual detection, and significantly improving detection efficiency and accuracy. Furthermore, it can provide timely warnings before the plunger pump malfunctions, promptly notifying maintenance personnel and effectively improving the real-time nature of anomaly maintenance. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0054] Figure 1 This is one of the flowcharts of the unsupervised plunger pump anomaly detection method based on time-related feature reconstruction provided by the present invention;

[0055] Figure 2 This is a schematic diagram of the plunger pump provided by the present invention;

[0056] Figure 3 This is the second flowchart of the unsupervised plunger pump anomaly detection method based on time-related feature reconstruction provided by the present invention;

[0057] Figure 4 This is a schematic diagram of the structure of the initial model provided by the present invention;

[0058] Figure 5 This is one of the schematic diagrams showing the distribution of steady-state vibration signals provided by the present invention;

[0059] Figure 6 This is the second schematic diagram of the distribution of steady-state vibration signals provided by the present invention;

[0060] Figure 7 This is one of the schematic diagrams illustrating the distribution of reconstruction errors provided by the present invention;

[0061] Figure 8 This is the second schematic diagram of the distribution of reconstruction error provided by the present invention;

[0062] Figure 9 This is the third schematic diagram of the distribution of reconstruction error provided by the present invention;

[0063] Figure 10 This is a schematic diagram of the unsupervised plunger pump anomaly detection device for time-related feature reconstruction provided by the present invention;

[0064] Figure 11 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0066] Fracturing trucks are mainly used for various fracturing operations in oil, gas, and water wells. They are specialized vehicles that inject high-pressure, high-volume fracturing fluid into the well to break up the formation and force proppant into the fractures. They can operate as stand-alone units or in multiple units. The main working equipment is a plunger pump. The pump unit controls the reciprocating motion of the plunger through the power end, changing the valve chamber volume and internal pressure, thereby indirectly controlling the opening and closing of the inlet and outlet pipes. The operating pressure is generally between 10MPa and 100MPa. The plunger pump power end assembly mainly consists of a drive motor, planetary gear mechanism, crankshaft, etc.; the hydraulic end assembly mainly consists of a valve box, discharge sealing combination valve, suction sealing combination valve, plunger, and its sealing components, etc. The core component of the suction and discharge sealing combination valves is the valve body. Currently, plunger pumps have short operating times, and the main failure modes are valve body damage, leading to seal failure, and poor bearing lubrication causing a rapid temperature rise and black carbonization on the plunger surface.

[0067] In practical engineering applications, once a piston pump malfunctions, workers often stop the entire hydraulic system and replace the piston pump, rather than repairing it after determining the cause of the failure. This is because piston pumps are consumable parts, and replacing the entire pump is often more cost-effective and less efficient than disassembling and repairing it. Therefore, timely and accurate detection of abnormal piston pump conditions, allowing for early intervention and preventing damage to the entire hydraulic system due to a single piston pump malfunction, is of great significance in engineering practice.

[0068] Traditional methods for diagnosing plunger pump faults typically rely on expert experience and knowledge, which is not only inefficient but also limited in accuracy.

[0069] To address the aforementioned issues, this application provides an unsupervised plunger pump anomaly detection method and apparatus based on time-related feature reconstruction for early warning of plunger pump failures. LSTM (Long Short-Term Memory) networks are widely used in time series modeling and have achieved outstanding results in time series prediction. VAE (Variational Autoencoder), as a deep learning generative model, can effectively learn the latent distribution of data, encode data into a latent feature space, and decode features to reconstruct the original data. Based on its powerful data reconstruction capabilities, VAE is widely used in anomaly detection. Therefore, this embodiment combines the advantages of VAE and LSTM, proposing a hybrid model of VAE and LSTM for plunger pump anomaly detection. This unsupervised detection model using a hybrid of VAE and LSTM automatically predicts plunger pump failures and provides early warning of plunger pump damage, thereby improving the efficiency and accuracy of plunger pump detection, promptly reminding maintenance personnel to replace the plunger pump, and significantly saving labor costs and downtime risks.

[0070] Figure 1 One of the flowcharts for an unsupervised plunger pump anomaly detection method based on time-related feature reconstruction provided in this application embodiment; as shown Figure 1 As shown, the method includes:

[0071] Step 110: Extract statistical features from the current steady-state vibration signal of the target plunger pump to obtain the multidimensional current operating statistical features of the target plunger pump;

[0072] The target plunger pump here refers to various types of plunger pumps for which fault detection is required; the target plunger pump can be a multi-cylinder plunger pump or a single-cylinder plunger pump, and this embodiment does not specifically limit it. Figure 2 This is a schematic diagram of the plunger pump provided in the embodiments of this application; as shown. Figure 2 As shown, the target plunger pump can be a 5-cylinder Jereh plunger pump; the following explanation will be based on the example of a multi-cylinder plunger pump as the target plunger pump.

[0073] Alternatively, vibration signals of the target plunger pump at each moment in the current cycle can be acquired by vibration sensors installed between the cylinders of the target plunger pump.

[0074] Next, in order to select effective vibration signals and improve the detection efficiency and accuracy of the plunger pump, the steady-state operating time of the target plunger pump can be analyzed based on signals collected by at least one of the following sensors: strain sensors between cylinders of the target plunger pump, speed sensor, torque sensor, and pressure sensor of the fracturing truck to which the target plunger pump belongs. The steady-state vibration signals of the target plunger pump at each moment in the current cycle can be filtered based on the steady-state operating time to obtain the current steady-state operating vibration signal of the target plunger pump.

[0075] The so-called steady-state operating cycle analysis can be achieved by inputting signals collected by at least one of the strain sensor, the speed sensor, torque sensor, and pressure sensor of the fracturing truck to which the target plunger pump belongs into a pre-trained steady-state recognition model to identify the steady-state operating time from the current cycle; or by performing logical judgment between the signals collected by at least one of the strain sensor, the speed sensor, torque sensor, and pressure sensor of the fracturing truck to which the target plunger pump belongs and their respective steady-state signal thresholds to obtain the corresponding steady-state operating time based on the logical judgment.

[0076] Next, statistical features are extracted from the current steady-state vibration signal of the target plunger pump to obtain multidimensional current operating statistical features of the target plunger pump. The multidimensional current operating statistical features include multidimensional time-domain statistical features, multidimensional frequency-domain statistical features, and multidimensional time-frequency-domain statistical features.

[0077] The methods for obtaining the multidimensional current operational statistical features include directly extracting multidimensional statistical features from the current steady-state vibration signal in the time domain, frequency domain, and time-frequency domain to obtain the multidimensional current operational statistical features; or, preprocessing the current steady-state vibration signal to improve its data quality, and then extracting multidimensional statistical features from the preprocessed signal in the time domain, frequency domain, and time-frequency domain to obtain the multidimensional current operational statistical features; the preprocessing may include noise reduction, outlier removal, etc.

[0078] Step 120: Perform feature filtering on the current operating statistical features of the multidimensional pump to obtain the current associated statistical features of the target plunger pump;

[0079] Optionally, in order to further simplify the complexity of the anomaly detection model, shorten the training time, and improve the model's generalization ability, after obtaining the multidimensional current operating statistical features, feature filtering can be performed on the current operating statistical features to remove redundant features and retain effective features, thereby obtaining lightweight current correlation statistical features that can accurately characterize the operating performance of the plunger pump.

[0080] The feature selection mentioned can be implemented based on logical calculations such as correlation and / or monotonicity, or based on machine learning algorithms such as principal component analysis, linear discriminant analysis, and kernel principal component analysis. This embodiment does not specifically limit it in this way.

[0081] Step 130: Input the current correlation statistical features into the detection model to obtain the fault prediction result of the target plunger pump; wherein, the detection model is obtained by unsupervised training of the initial model based on the historical correlation statistical features of the first sample plunger pump; the historical correlation statistical features are obtained by statistical feature extraction and feature screening of the historical steady-state operating vibration signal of the first sample plunger pump; the initial model is constructed based on variational autoencoder (VAE) and long short-term memory network (LSTM) to mine reconstructed features related to feature time information; the first sample plunger pump is a sample plunger pump in normal operating condition.

[0082] The detection model here includes an encoding layer and a decoding layer; the encoding layer is used to extract encoded features of the associated statistical features, and the decoder is used to mine the temporal information of the associated statistical features and reconstruct features based on the encoded features; the detection model is obtained by unsupervised training of the initial model based on the historical associated statistical features of the first sample plunger pump in normal operation.

[0083] For VAEs: A VAE is a generative model used to learn a latent representation of data and generate new samples. The goal of a VAE is to encode the input data into a mean and variance vector in a latent space, and then sample from that distribution to generate reconstructed samples. The encoder part of a VAE maps the input sequence x to the mean and variance of the latent space: z ~ Q(z∣x), where Q is the encoder function. The decoder part generates the reconstructed sequence x^ from the sampled z in the latent space: x^ ~ P(x∣z), where P is the decoder function.

[0084] For LSTM: LSTM is a recurrent neural network suitable for processing temporal data. LSTM captures and remembers long-term dependencies through memory units and gating mechanisms. LSTM includes input gates, forget gates, output gates, and cell states, which can process the input sequence stepwise and update the memory units at each time step.

[0085] In summary, to learn the latent space representation and generate reconstructed features from temporal data, this embodiment constructs an initial model (hereinafter referred to as the VAE-LSTM model) based on a joint VAE and LSTM. This VAE-LSTM model combines the encoder-decoder framework of VAE with the temporal modeling capabilities of LSTM. For example, the encoding layer of the initial model can be constructed using an encoder based on a variational autoencoder and an LSTM, and the decoding layer can be constructed using a decoder based on a variational autoencoder. That is, an LSTM layer can be used to model the input sequence, and the hidden states of the LSTM can be used as the input to the encoder. The latent space is modeled by the encoder part of the VAE, while the decoder part is responsible for generating the reconstructed sequence from the latent space.

[0086] Optionally, before performing step 130, the detection model can be pre-trained. The specific training method is unsupervised training, and the training steps include: collecting a large number of historical steady-state vibration signals from a large number of normally operating first sample plunger pumps. The first sample plunger pump is a sample plunger pump in normal operating condition. The historical steady-state vibration signals include the steady-state vibration signals of the first sample plunger pump collected in each historical period prior to the current period.

[0087] After collecting the historical steady-state vibration signal of the first sample plunger pump, the historical correlation statistical features of the first sample plunger pump can be obtained by performing statistical feature extraction and feature filtering on the historical steady-state vibration signal, referring to steps 110-120.

[0088] Then, the historical correlation statistical features of the first sample plunger pump are input into VAE-LSTM, and unsupervised training is performed on VAE-LSTM so that the trained VAE-LSTM can accurately reconstruct the correlation statistical features of the normally operating plunger pump. This allows the detection model built based on the trained VAE-LSTM to accurately learn the distribution characteristics of the correlation statistical features of the plunger pump under normal operating conditions.

[0089] After obtaining a detection model with good reconstruction performance of the correlation statistical features of a normally operating plunger pump, the decoder of the detection model reconstructs the current correlation statistical features to obtain the reconstructed statistical features corresponding to the current correlation statistical features. Based on the reconstruction error between the reconstructed statistical features corresponding to the current correlation statistical features and the current correlation statistical features, the model predicts whether the target plunger pump has failed, thereby obtaining the failure prediction result of the target plunger pump.

[0090] The so-called prediction of whether the target plunger pump has failed based on the reconstruction error can be achieved by comparing the reconstruction error with an error threshold to predict whether the target plunger pump has failed; or it can be achieved by first performing a health score based on the reconstruction error to predict whether the target plunger pump has failed based on the health score. This embodiment does not specifically limit this method.

[0091] Step 140: Based on the fault prediction results, generate fault warning information for the target plunger pump;

[0092] To address this issue, a novel intelligent fault diagnosis method for plunger pumps based on deep learning theory has been developed. This method involves converting one-dimensional time-series signals into two-dimensional images to adapt to large image models with multiple parameters, thereby improving the fault diagnosis rate of plunger pumps. However, the aforementioned fault diagnosis method traces the fault source after component damage rather than providing early warning. In real-world scenarios, due to the short service life of plunger pumps, fault diagnosis and repair are time-consuming and labor-intensive. Failure to provide timely early warnings can severely impact the normal operation and safety of fracturing trucks. Therefore, this embodiment addresses this problem by using fault prediction results from an unsupervised detection model to provide early warnings of plunger pump damage, reminding maintenance personnel to replace the plunger pump, significantly saving labor costs and downtime risks. The specific warning steps include:

[0093] After obtaining the fault prediction results, it can be determined whether the target plunger pump has failed. If a failure occurs, the fault warning information generated based on the fault prediction results will be alarmed in the form of text, voice, light, etc., to give early warning of the plunger pump failure and notify the maintenance personnel in a timely manner so that the maintenance personnel can make maintenance decisions in a timely manner based on the fault warning information, thereby improving the operational stability and safety of the fracturing truck.

[0094] The method provided in this embodiment extracts and filters statistical features from the current steady-state vibration signal of the target plunger pump online to obtain high-quality, lightweight current associated statistical features. An unsupervised detection model trained based on a variational autoencoder and a long short-term memory network is then constructed to reconstruct these associated statistical features. The fault prediction result of the target plunger pump is obtained based on the reconstruction loss of the detection model. Based on the fault prediction result, the plunger pump is automatically and accurately warned of faults. The entire detection process is completed online in real time, minimizing human intervention and operation, effectively reducing the workload and cost of manual detection, and significantly improving detection efficiency and accuracy. Furthermore, timely warnings can be issued before the plunger pump malfunctions, allowing maintenance personnel to be notified promptly and improving the real-time nature of abnormal maintenance.

[0095] In some embodiments, the detection model is trained based on the following steps:

[0096] Statistical feature extraction and feature filtering were performed on the historical steady-state operation vibration signal of the first sample plunger pump to obtain the historical correlation statistical features of the first sample plunger pump.

[0097] The historical correlation statistical features are input into the initial model to obtain the first encoded features output by the encoding layer of the initial model and the first reconstructed features output by the decoding layer of the initial model;

[0098] With the goal of minimizing the reconstruction error between the historical correlation statistical features and the first reconstructed features, and maximizing the posterior probability of the first encoded features, the initial model is iteratively trained to obtain the detection model.

[0099] Figure 3 This is a second schematic flowchart of the unsupervised plunger pump anomaly detection method based on time-related feature reconstruction provided in this application embodiment; as shown... Figure 3 As shown, the specific training steps for the detection model include:

[0100] First, based on the three indicator signals of the fracturing truck to which the first sample plunger pump belongs—motor speed, motor torque, and fracturing truck pressure—the continuous steady-state vibration signal of the fracturing truck plunger pump is extracted to obtain the historical steady-state operating vibration signal of the first sample plunger pump.

[0101] Next, referring to steps 110-120, statistical features are extracted from the historical steady-state vibration signal of the first sample plunger pump to obtain the multidimensional historical operation statistical features of the first sample plunger pump, and feature filtering is performed on the multidimensional historical operation statistical features of the first sample plunger pump to obtain the historical correlation statistical features of the first sample plunger pump.

[0102] Next, the historical correlation statistical features are input into the encoding layer of the initial model. The encoding layer encodes the historical correlation statistical features to obtain the first encoded features. The first encoded features are input into the decoding layer of the initial model. The decoding layer reconstructs the historical correlation statistical features according to the first encoded features to obtain the reconstructed features of the historical correlation statistical features, which is the first reconstructed features.

[0103] Subsequently, the distribution of the VAE part of the learning data is forced by minimizing the reconstruction error between the first reconstructed feature and the historical correlation statistical feature, and maximizing the posterior probability of the latent variable (i.e., the first encoded feature). That is, a loss function is constructed by minimizing the reconstruction error between the first reconstructed feature and the historical correlation statistical feature, and maximizing the KL divergence between the distribution of the first encoded feature and the prior distribution (e.g., the normal distribution). The VAE-LSTM model is trained based on this loss function, thereby obtaining a detection model that can be used to accurately and efficiently detect anomalies in the time-series signal of the plunger pump, i.e., the steady-state vibration signal.

[0104] Subsequently, after the model training is completed, the current associated statistical features of the input plunger pump under test can be accurately and efficiently reconstructed to obtain the reconstruction error; based on the reconstruction error, fault detection can be performed to achieve early warning of plunger pump faults.

[0105] Traditional machine learning-based anomaly detection models, such as Isolation Forest, the "3Q" rule, and One-Class SVM (One-Class Support Vector Machine), have achieved good results in anomaly detection. However, for time series signals, effectively capturing the temporal information in the data becomes particularly important. These anomaly detection models ignore temporal order information, only considering the distance between data points or the distribution of features, and cannot fully utilize temporal dependencies. Secondly, these methods usually use fixed window sizes to process data, making them unable to adapt to anomaly patterns at different time scales, potentially missing anomalies or misclassifying normal data. Furthermore, traditional methods are based on linear models or assume that the data follows a specific distribution, making it difficult to handle complex nonlinear relationships, while anomalies often exhibit nonlinear patterns or trends.

[0106] In contrast, the embodiments of this application use the historical correlation statistical features of a normally operating plunger pump to perform unsupervised training on the VAE-LSTM model, thereby obtaining a detection model that combines the autonomous learning ability of variational autoencoders with the temporal learning ability of long short-term memory networks, and can handle complex nonlinear relationships. This detection model can model normal patterns by learning the distribution of time-series data and detect anomalies by using reconstruction errors. This effectively solves the defects of the above-mentioned technologies that cannot fully utilize temporal dependencies and cannot learn nonlinear relationships, and effectively improves the accuracy of anomaly detection.

[0107] In some embodiments, the encoding layer of the initial model is constructed based on the encoder of the VAE, the first feature embedding layer, and the LSTM; the decoding layer of the initial model is constructed based on the second feature embedding layer and the decoder of the VAE.

[0108] The step of inputting the historical correlation statistical features into the initial model to obtain the first encoded feature output by the encoding layer of the initial model and the first reconstructed feature output by the decoding layer of the initial model includes:

[0109] The historical correlation statistical features are input into the encoder to obtain the initial encoded features of the historical correlation statistical features output by the encoder;

[0110] The initial encoded features are input into the first feature embedding layer to obtain the embedded features of the initial encoded features output by the first feature embedding layer;

[0111] The embedding features of the initial encoded features are input into the LSTM to obtain the first encoded features output by the LSTM;

[0112] The first encoded feature is input into the second feature embedding layer to obtain the embedded feature of the first encoded feature output by the second feature embedding layer;

[0113] The embedding features of the first encoded feature are input into the decoder to obtain the first reconstructed feature output by the decoder.

[0114] Figure 4 This is a schematic diagram of the structure of the initial model provided in the embodiments of this application; as shown below. Figure 4 As shown, the encoding layer of the initial model is formed by sequentially stacking a VAE-based encoder, a first feature embedding layer, and an LSTM; the decoding layer of the initial model is formed by sequentially stacking a second feature embedding layer and a VAE-based decoder. The encoder and decoder can be constructed based on a multi-layer feedforward neural network.

[0115] Alternatively, the initial model can be trained using the following steps:

[0116] First, the sliding time window method is adopted to extract window sequences from the time series dataset of historical correlation statistical features to form a subsequence window dataset. The sliding time window method refers to the method of obtaining data by sliding a data window (referred to as a time window) of fixed time length on the time series (i.e., historical correlation statistical features arranged in time series). Assuming that the length of the time window is l and the length of the time series is t, then t-l+1 windows containing subsequences (referred to as subsequence window sets) can be extracted.

[0117] Next, the LSTM-VAE model is trained in an unsupervised manner using the healthy state subsequence window dataset. The specific training steps are as follows:

[0118] Perform the following steps for each subsequence window:

[0119] The subsequence window data in the subsequence window dataset is input into the LSTM-VAE model to perform encoding and reconstruction learning based on each subsequence window data. The encoder outputs the initial encoded features of each subsequence window data, the first feature embedding layer outputs the embedded features of the initial encoded features, the LSTM outputs the first encoded features, the second feature embedding layer outputs the embedded features of the first encoded features, and the decoder outputs the first reconstructed features of each subsequence window data. Then, the average value of the reconstruction error between the first reconstructed features of each subsequence window data and the reconstruction error between each subsequence window data is calculated, and the first encoded features of each subsequence window data are used to train the LSTM-VAE model to obtain a detection model that can accurately detect anomalies in plunger pumps.

[0120] In some embodiments, the step of acquiring the current steady-state vibration signal includes:

[0121] The motor speed, motor torque, and pressure of the fracturing truck to which the target plunger pump belongs are obtained at multiple moments within the current cycle.

[0122] Based on the information on the changes in motor speed, the information on the changes in motor torque, and the information on the changes in pressure at multiple times, the steady-state operating time of the target plunger pump is obtained within the current cycle.

[0123] The vibration signals of the target plunger pump at multiple moments during the steady-state operation time are extracted to obtain the current steady-state operation vibration signal.

[0124] It should be noted that during the non-steady-state operation of the equipment, the abnormal fluctuations generated are significant, making it difficult to distinguish these abnormal fluctuations from noise and degradation signals during data analysis, thus hindering effective equipment analysis. Therefore, to achieve accurate fault detection of the plunger pump, it is necessary to remove the vibration signals generated during non-steady-state operation and retain only the valid vibration signals from steady-state operation.

[0125] Optionally, based on the speed sensor, torque sensor, and pressure sensor of the fracturing truck to which the target plunger pump belongs, the motor speed, motor torque, and pressure of the fracturing truck can be collected at multiple moments within the current cycle. Based on the changes in motor speed, motor torque, and pressure at multiple moments, the steady-state operating time of the plunger pump with stable working pressure can be selected within the current cycle. The data of the time period corresponding to the vibration data can then be identified as the current steady-state operating vibration signal, so that subsequent plunger pump anomaly detection can be performed efficiently and accurately based on the steady-state operating vibration signal.

[0126] Figure 5 One of the schematic diagrams showing the distribution of steady-state vibration signals provided in the embodiments of this application; Figure 6 This is the second schematic diagram of the distribution of steady-state vibration signals provided in the embodiments of this application; as shown... Figure 5 and Figure 6 As shown, by extracting the continuous steady-state vibration signal of the plunger pump of the fracturing truck based on three indicator signals—motor speed, motor torque, and fracturing truck pressure—the interference of unstable signals can be effectively eliminated, thereby obtaining a clearer and more reliable steady-state vibration signal. This helps to improve the accuracy and usability of the data, and thus improves the efficiency and accuracy of plunger pump anomaly detection.

[0127] In some embodiments, step 110 involves extracting statistical features from the current steady-state vibration signal of the target plunger pump to obtain multidimensional current operational statistical features of the target plunger pump, including:

[0128] The current steady-state vibration signal is denoised based on the Savitzky-Golay digital filter.

[0129] The noise-reduced current steady-state vibration signal is processed by sliding window to obtain multiple window data.

[0130] Multi-dimensional statistical features are extracted from the data of each window to obtain the multi-dimensional current operating statistical features of the target plunger pump.

[0131] The multidimensional current operating statistical features include the maximum value, minimum value, mean, peak-to-peak value, absolute mean, root mean square value, root mean square amplitude, standard deviation, kurtosis, skewness, margin index, waveform index, impulse index, peak index, kurtosis index, frequency center, average frequency, frequency domain root mean square, frequency variance, and the multidimensional energy spectrum features of the last layer of wavelet packet decomposition.

[0132] like Figure 5 and Figure 6 As shown, since the raw data is based on the current steady-state vibration signal, it contains a large amount of noise, which affects the feature selection effect and the detection effect of the detection model. Therefore, before feature selection, it is necessary to use a Savitzky-Golay filter to denoise the current steady-state vibration signal to eliminate noise interference.

[0133] The Savitzky-Golay filter is widely used for data stream smoothing and noise reduction. It is a filtering method based on local polynomial least squares fitting in the time domain. The most significant feature of this filter is that it can remove noise while ensuring that the shape and width of the signal remain unchanged.

[0134] The Savitzky-Golay filter is a local polynomial fitting filtering method used for smoothing and denoising signals. It can reduce the impact of noise to a certain extent while preserving the overall shape and characteristics of the signal. The basic idea of ​​the Savitzky-Golay filter is to perform a polynomial fit within a given window and use points on the fitted curve to approximate the original signal, thus obtaining a smoothed result. This method can be adapted to different types of signals by adjusting the polynomial order and the window size.

[0135] The advantages of the Savitzky-Golay filter include: preserving the overall shape and characteristics of the signal and avoiding information loss due to over-smoothing; reducing the impact of noise and, in some cases, outperforming conventional smoothing methods; and offering some resistance to outliers.

[0136] Suppose the current steady-state vibration signal (hereinafter referred to as the original signal) is y, it can be smoothed to obtain a denoised current steady-state vibration signal (hereinafter referred to as the new signal) y′. This process can be achieved using polynomial fitting. Specifically, a polynomial function f(x) is used to approximate the original signal within each window, and then points on the fitted curve are used as the smoothed result. Assuming f(x) is an m-th order polynomial function, it can be expressed as:

[0137]

[0138] Where a0, a1, ..., a m These are the coefficients of the polynomial. Specifically, the least squares method can be used to determine these coefficients, minimizing the squared error between the data points within a given window and the fitted curve. Specifically, an error function E can be defined, representing the squared difference between all data points within a given window and the fitted curve:

[0139]

[0140] Where y i The original signal at position x i The value at point n is half the window size. During the fitting optimization process, a set of coefficients a0, a1, ..., a... needs to be found. m This minimizes the error function E.

[0141]

[0142] The specific optimization steps include:

[0143] Taking the derivative of E and setting it equal to zero, we obtain the following matrix equation:

[0144]

[0145] This equation can be solved using matrix inversion in linear algebra to obtain the polynomial coefficients a0, a1, ..., a m Then, the points on the fitted curve can be used as the smoothed result.

[0146] Subsequently, after obtaining the noise-reduced current steady-state vibration signal, the biggest problem with the data is its excessive length. To more effectively characterize the operating features of the plunger pump, a sliding window process can be applied to the noise-reduced current steady-state vibration signal to obtain multiple window data points. Multidimensional statistical features of the current operation can then be extracted from these window data points. The window size of the sliding window can be configured according to actual needs, such as 2048 time-series data points.

[0147] Here, sliding window processing helps analyze local features of time series data within each window, unaffected by overall data trends. This helps identify patterns and trends in data across different time periods. Through sliding window processing, various statistics, frequency domain features, and time-frequency domain features can be extracted within each window. Simultaneously, sliding window processing helps reduce the dimensionality and complexity of the data. This simplifies data processing and makes subsequent modeling and analysis more efficient.

[0148] The so-called multidimensional current operating statistical features can include 28 statistical features, namely 16 time-domain features, 4 frequency-domain features, and 8 time-frequency-domain features. Specifically, these include maximum value, minimum value, mean, peak-to-peak value, absolute mean, root mean square value, root mean square amplitude, standard deviation, kurtosis, skewness, margin index, waveform index, impulse index, peak index, kurtosis index, frequency center, average frequency, frequency domain root mean square, frequency variance, and the capability ratio of the last sub-band of wavelet packet decomposition, totaling 8 time-frequency-domain features. The specific extracted features are shown in Table 1.

[0149] In most sensor data feature extraction processes, only time-domain features are extracted. However, extracting features solely from the time-domain signal misses the signal differences caused by frequency variations. Therefore, this embodiment adds frequency domain and time-frequency domain feature extraction to the extracted sensor data.

[0150] The following describes the specific steps for extracting multidimensional current operational statistical features:

[0151] First, based on the feature extraction method in Table 1, the current steady-state vibration signal after noise reduction is subjected to time-domain sliding window processing to obtain multiple time-domain statistical features: maximum value, minimum value, mean value, peak-to-peak value, absolute mean value, root mean square value, root mean square amplitude, standard deviation, kurtosis, skewness, margin index, waveform index, impulse index, peak index, and kurtosis index.

[0152] Next, because the signals generated during the reciprocating motion of a plunger pump are complex and variable, time-domain analysis alone is insufficient to fully explain their components and cannot describe the speed of signal fluctuations. Therefore, it is necessary to transform the time-domain signal into the frequency domain to analyze the frequency changes of the signal.

[0153] Specifically, time-frequency domain conversion is performed using Fourier transform:

[0154]

[0155] Where X(ω) is the spectrum result after Fourier transform, and x(t) is the time-domain signal. The Fourier transform can truly reflect the phase changes of signals at different frequencies.

[0156] Since the acquired signals are discrete in practical applications, a Discrete Fourier Transform is then performed:

[0157]

[0158] Where x(kt) is the time-domain sampled signal (i.e., the vibration signal at each moment in the current steady-state vibration signal), N is the number of sampling points, t is the sampling interval, and n and k are both the serial numbers of discrete points.

[0159] After converting the time-domain signal to the frequency domain, we can obtain the frequency components of the signal and their distribution. To further characterize the characteristics of the plunger pump in different states, four frequency domain statistical features were extracted from the spectrum: average frequency, root mean square frequency, frequency center, and frequency variance.

[0160] Furthermore, time-domain and frequency-domain features can only be analyzed from the perspective of signal amplitude changes over time or frequency components, which cannot comprehensively reflect the signal characteristics. Therefore, it is necessary to extract time-frequency domain features to represent the signal from the perspective of time-frequency-amplitude. Wavelet Packet Decomposition (WPD) decomposes the high-frequency and low-frequency components of the signal simultaneously, which has high computational efficiency and resolution accuracy. Therefore, this embodiment uses WPD to extract time-frequency domain features.

[0161] Optionally, the noise-reduced current steady-state vibration signal can be decomposed using the Mallat algorithm:

[0162]

[0163]

[0164] Where d is the wavelet packet coefficient, h n-2k and g n-2k It is a dual operator.

[0165] The extracted time-frequency feature is the average energy E of the wavelet packet coefficients. X-L-M :

[0166]

[0167] Where X is the current steady-state vibration signal after noise reduction, M is the sub-band marker, and L is the layer number.

[0168] Analysis of the collected signals revealed that the best results were achieved by performing a three-level decomposition. Therefore, in this embodiment, the noise-reduced current steady-state vibration signal was decomposed into three levels, and eight energy features were extracted from each channel as time-frequency statistical features.

[0169] Then, based on the extracted frequency domain statistical features, time domain statistical features, and time-frequency statistical features, multidimensional current operating statistical features are obtained, so as to accurately predict the failure prediction results of the plunger pump based on the multidimensional current operating statistical features.

[0170] This embodiment extracts frequency domain statistical features, time domain statistical features, and time-frequency statistical features after denoising the current steady-state vibration signal. This not only maps the vibration signal from a high-dimensional feature space to a low-dimensional feature space, but also ensures that the mapped feature data still has good separability, thereby improving fault detection efficiency and fault detection accuracy.

[0171] In some embodiments, step 120 involves feature filtering of the multidimensional current operating statistical features to obtain the current associated statistical features of the target plunger pump, including:

[0172] Statistical features were extracted from the historical steady-state vibration signal of the second sample plunger pump to obtain the multidimensional historical operational statistical features of the second sample plunger pump; the second sample plunger pump is a sample plunger pump of the same type as the target plunger pump.

[0173] Calculate the monotonicity index of the historical operational statistical characteristics of each dimension;

[0174] Based on the monotonicity index, at least one-dimensional candidate statistical features are selected from the multidimensional historical operation statistical features.

[0175] Calculate the Pearson correlation coefficient, Spearman correlation coefficient, and Kendall coefficient among at least one of the candidate statistical features;

[0176] The Pearson correlation coefficient, the Spearman correlation coefficient, and the Kendall coefficient are fused to obtain at least one-dimensional correlation index among the candidate statistical features;

[0177] Based on the correlation index, the historical correlation statistical features of the second sample plunger pump are selected from at least one-dimensional candidate statistical features.

[0178] Among the current operational statistical features described in the multidimensional analysis, statistical features that match the historical correlation statistical features are selected as the current correlation statistical features.

[0179] Optionally, after obtaining the multidimensional current operational statistical features of the denoised current steady-state vibration signal, the study of high-dimensional data is extremely challenging. While ensuring the prediction accuracy of the learning algorithm, the sample requirement during training increases exponentially with the increase in feature dimensionality. Therefore, this embodiment performs feature filtering on the multidimensional current operational statistical features to remove some irrelevant and redundant features, thereby reducing dimensionality, the number of features, running time, and improving the algorithm's running time. The specific filtering steps are as follows:

[0180] First, statistical features were extracted from the historical steady-state vibration signal of the second sample plunger pump, which is the same type as the target plunger pump, according to the feature extraction method shown in Table 1, so as to obtain the multidimensional historical operation statistical features of the second sample plunger pump.

[0181] Next, since a good feature often has the ability to characterize the overall trend of a set of data, if a feature has low monotonicity, meaning that the monotonic relationship between the feature's value and the target variable is not obvious, it can be considered for removal. Therefore, by calculating the monotonicity index of each dimension of historical statistical features, features with monotonicity below the monotonicity threshold can be removed to obtain at least one-dimensional candidate statistical features, thereby reducing the dimensionality of the feature space and improving the simplification and interpretability of the model. The formula for calculating the monotonicity index is as follows:

[0182]

[0183] Where Mon is the monotonicity coefficient, R is any historical operational statistical feature, L is time, T is the total sampling time, and δ(·) is the unit step function. The larger this value, the stronger the characterization ability of the historical operational statistical feature.

[0184] Next, based on monotonicity screening, the correlation index between features in at least one-dimensional candidate statistical features is further calculated using Pearson correlation coefficient, Spearman correlation coefficient, and Kendall coefficient for unsupervised feature screening. This allows for the removal of any candidate statistical feature whose correlation index is greater than a correlation threshold from the at least one-dimensional candidate statistical features, thereby eliminating redundant features and obtaining the historical correlation statistical features of the second sample plunger pump. The fusion can be achieved through weighted summation or direct summation.

[0185] The Pearson correlation coefficient measures the degree of linearity between two variables, ranging from -1 to 1, where -1 indicates a perfect negative correlation, 1 indicates a perfect positive correlation, and 0 indicates no linear relationship. The formula for the Pearson correlation coefficient r is as follows:

[0186]

[0187] Where X and Y represent the values ​​of the two variables (i.e., candidate statistical features) involved in the correlation calculation. and These represent the means of the two variables, respectively.

[0188] The Spearman correlation coefficient measures the degree of monotonicity between two variables, even if the relationship is not linear. The steps to calculate the Spearman correlation coefficient are to convert candidate statistical features into ordinal (ranked) data, and then calculate the Pearson correlation coefficient for the ordinal data. The formula is as follows:

[0189]

[0190] Where ρ represents the Spearman correlation coefficient, ∑d 2 Let represent the sum of squares of the differences between the ranks corresponding to any two candidate statistical features, and n represent the number of sample points in the candidate statistical features.

[0191] Kendall's coefficient is also used to measure the degree of monotonicity between two variables. It is calculated based on the number of comparisons made between pairs of ordinal data corresponding to any two candidate statistical features. The formula for Kendall's coefficient is as follows:

[0192]

[0193] Where τ represents the Kendall coefficient, P represents the number of pairs with the same order, Q represents the number of pairs with the opposite order, T represents the number of pairs with the same value but not the same order, and U represents the number of pairs with the same order but not the same value.

[0194] Next, in practical applications, statistical features that match historical correlation statistical features can be filtered from the multidimensional current running statistical features to obtain the current correlation statistical features.

[0195] The method provided in this embodiment calculates the monotonicity of the features and the correlation between the features after obtaining the statistical characteristics of the noise-reduced signal, and then filters the most representative features to help achieve real-time and effective fault detection of the target plunger pump, reduce detection time overhead, and increase detection effect.

[0196] In some embodiments, step 130 involves inputting the current correlation statistical features into the detection model to obtain the fault prediction result of the target plunger pump, including:

[0197] The current correlation statistical features are input into the detection model to obtain the second reconstructed features output by the encoding layer of the detection model;

[0198] The reconstruction error between the second reconstructed feature and the current associated statistical feature is compared with the target error to obtain the fault prediction result of the target plunger pump.

[0199] Optionally, the fault prediction results can be obtained through the following steps:

[0200] The current correlation statistical feature is extracted into subsequence window data, and the subsequence window data of the current correlation statistical feature is input into the encoding layer and decoding layer in the detection model. The encoder, first feature embedding layer and LSTM of the encoding layer encode the features of the subsequence window data of the current correlation statistical feature to obtain the second encoded features of the subsequence window data of the current correlation statistical feature. The second feature embedding layer and decoder of the decoding layer reconstruct the features of the second encoded features to obtain the second reconstructed features of the subsequence window data of the current correlation statistical feature.

[0201] Next, the reconstruction error between the second reconstruction feature of each sub-sequence window data of the current associated statistical feature and the reconstruction error of each sub-sequence window data of the current associated statistical feature is calculated to obtain the reconstruction error corresponding to each sub-sequence window data of the current associated statistical feature. The reconstruction error corresponding to each sub-sequence window data of the current associated statistical feature is then compared with the target error. If the reconstruction error corresponding to any sub-sequence window data of the current associated statistical feature is greater than or equal to the target error, the fault prediction result is used to characterize the target plunger pump malfunctioning at the time point corresponding to that sub-sequence window data within the current cycle. Based on this, a fault warning message is sent to alert maintenance personnel that the target plunger pump has malfunctioned at the time point corresponding to that sub-sequence window data within the current cycle, enabling timely maintenance and improving the stable operation of the fracturing truck and the user experience. If the reconstruction error corresponding to all sub-sequence window data of the current associated statistical feature is less than the target error, the fault prediction result is used to characterize the target plunger pump as being in normal condition within the current cycle, thereby achieving efficient and accurate anomaly detection of the plunger pump.

[0202] The target error here can be obtained by analyzing the distribution characteristics of the reconstruction error between the historical steady-state vibration signal of each first sample plunger pump and the reconstruction characteristics corresponding to the historical steady-state vibration signal, and obtaining the error threshold.

[0203] To verify the effectiveness of the method provided in this embodiment, experimental verification will be conducted using data from a fracturing truck plunger pump measured in a real-world scenario.

[0204] By using data from fracturing truck plunger pumps collected in real-world scenarios, the data is first denoised, then features are extracted and filtered, and finally the obtained effective representation features are used as input to a detection model based on VAE-LSTM unsupervised training for modeling. Figure 7 One of the schematic diagrams illustrating the distribution of reconstruction errors provided in the embodiments of this application; Figure 8 A second schematic diagram illustrating the distribution of reconstruction errors provided in an embodiment of this application; Figure 9 This is the third schematic diagram illustrating the distribution of reconstruction errors provided in the embodiments of this application; as shown below. Figure 7 and Figure 8 As shown, by inputting the selected effective representation features into a detection model that combines the advantages of LSTM and VAE, the data reconstruction error can be accurately obtained. Figure 9 As shown, by comparing the target error configured based on the reconstruction error corresponding to the normally operating sample plunger pump with the reconstruction error corresponding to the target plunger pump, the failure prediction result of the target plunger pump can be predicted efficiently and accurately in advance, thus verifying the effectiveness of the algorithm.

[0205] The unsupervised plunger pump anomaly detection device based on time-related feature reconstruction provided by the present invention will be described below. The unsupervised plunger pump anomaly detection device based on time-related feature reconstruction described below can be referred to in correspondence with the unsupervised plunger pump anomaly detection method based on time-related feature reconstruction described above.

[0206] Figure 10 A schematic diagram of the unsupervised plunger pump anomaly detection device for time-related feature reconstruction provided in this application embodiment; as shown. Figure 10As shown, the device includes: a feature extraction unit 1010 for extracting statistical features from the current steady-state vibration signal of the target plunger pump to obtain multidimensional current operating statistical features of the target plunger pump; a feature filtering unit 1020 for filtering the multidimensional current operating statistical features to obtain current associated statistical features of the target plunger pump; a detection unit 1030 for inputting the current associated statistical features into a detection model to obtain a fault prediction result of the target plunger pump; and an early warning unit 1040 for generating a fault early warning information for the target plunger pump based on the fault prediction result. The detection model is obtained by unsupervised training of an initial model based on the historical associated statistical features of a first sample plunger pump. The historical associated statistical features are obtained by extracting and filtering statistical features from the historical steady-state vibration signal of the first sample plunger pump. The initial model is constructed based on a variational autoencoder (VAE) and a long short-term memory (LSTM) network. The first sample plunger pump is a sample plunger pump in normal operating condition.

[0207] The device provided in this embodiment extracts and filters statistical features from the current steady-state vibration signal of the target plunger pump online to obtain high-quality, lightweight current associated statistical features. An unsupervised detection model trained based on a variational autoencoder and a long short-term memory network is then used to reconstruct these associated statistical features. The reconstruction loss of the detection model is used to obtain the fault prediction result of the target plunger pump. Based on the fault prediction result, the device automatically and accurately provides fault warnings for the plunger pump. The entire detection process is completed online in real time, minimizing human intervention and operation, effectively reducing the workload and cost of manual detection, and significantly improving detection efficiency and accuracy. Furthermore, it can provide timely warnings before the plunger pump malfunctions, promptly notifying maintenance personnel and effectively improving the real-time nature of abnormal maintenance.

[0208] Figure 11 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 11 As shown, the electronic device may include a processor 1110, a communications interface 1120, a memory 1130, and a communication bus 1140. The processor 1110, communications interface 1120, and memory 1130 communicate with each other via the communication bus 1140. The processor 1110 can call logical instructions in the memory 1130 to execute the unsupervised plunger pump anomaly detection method based on time-related feature reconstruction provided by the aforementioned methods.

[0209] Furthermore, the logical instructions in the aforementioned memory 1130 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0210] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the unsupervised plunger pump anomaly detection method for time-related feature reconstruction provided by the above methods.

[0211] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an unsupervised plunger pump anomaly detection method for reconstructing time-related features provided by the methods described above.

[0212] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0213] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0214] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An unsupervised method for detecting anomalies in a plunger pump based on time-related feature reconstruction, characterized in that, include: Statistical features are extracted from the current steady-state vibration signal of the target plunger pump to obtain the multidimensional current operating statistical features of the target plunger pump; By performing feature filtering on the current operating statistical features described in the multidimensional analysis, the current associated statistical features of the target plunger pump are obtained. The current correlation statistical features are input into the detection model to obtain the fault prediction result of the target plunger pump; Based on the fault prediction results, fault warning information for the target plunger pump is generated; The detection model is obtained by unsupervised training of an initial model based on the historical correlation statistical features of the first sample plunger pump. The historical correlation statistical features are obtained by extracting and filtering statistical features from the historical steady-state vibration signals of the first sample plunger pump. The initial model is constructed based on a variational autoencoder (VAE) and a long short-term memory network (LSTM) to mine reconstructed features related to feature time information. The first sample plunger pump is a sample plunger pump in normal operating condition. The steps for obtaining the current steady-state vibration signal include: The motor speed, motor torque, and pressure of the fracturing truck to which the target plunger pump belongs are obtained at multiple moments within the current cycle. Based on the information on the changes in motor speed, the information on the changes in motor torque, and the information on the changes in pressure at multiple times, the steady-state operating time of the target plunger pump is obtained within the current cycle. The vibration signals of the target plunger pump at multiple moments during the steady-state operation time are extracted to obtain the current steady-state operation vibration signal; The step of performing feature filtering on the multidimensional current operating statistical features to obtain the current associated statistical features of the target plunger pump includes: Statistical features were extracted from the historical steady-state vibration signal of the second sample plunger pump to obtain the multidimensional historical operational statistical features of the second sample plunger pump; the second sample plunger pump is a sample plunger pump of the same type as the target plunger pump. Calculate the monotonicity index of the historical operational statistical characteristics of each dimension; Based on the monotonicity index, at least one-dimensional candidate statistical features are selected from the multidimensional historical operation statistical features. Calculate the Pearson correlation coefficient, Spearman correlation coefficient, and Kendall coefficient among at least one of the candidate statistical features; The Pearson correlation coefficient, the Spearman correlation coefficient, and the Kendall coefficient are fused to obtain at least one-dimensional correlation index among the candidate statistical features; Based on the correlation index, the historical correlation statistical features of the second sample plunger pump are selected from at least one-dimensional candidate statistical features. Among the current operational statistical features described in the multidimensional analysis, statistical features that match the historical correlation statistical features are selected as the current correlation statistical features.

2. The unsupervised plunger pump anomaly detection method based on time-related feature reconstruction according to claim 1, characterized in that, The detection model was trained based on the following steps: Statistical feature extraction and feature filtering were performed on the historical steady-state operation vibration signal of the first sample plunger pump to obtain the historical correlation statistical features of the first sample plunger pump. The historical correlation statistical features are input into the initial model to obtain the first encoded features output by the encoding layer of the initial model and the first reconstructed features output by the decoding layer of the initial model; With the goal of minimizing the reconstruction error between the historical correlation statistical features and the first reconstructed features, and maximizing the posterior probability of the first encoded features, the initial model is iteratively trained to obtain the detection model.

3. The unsupervised plunger pump anomaly detection method based on time-related feature reconstruction according to claim 2, characterized in that, The encoding layer of the initial model is constructed based on the encoder of the VAE, the first feature embedding layer, and the LSTM; the decoding layer of the initial model is constructed based on the second feature embedding layer and the decoder of the VAE. The step of inputting the historical correlation statistical features into the initial model to obtain the first encoded feature output by the encoding layer of the initial model and the first reconstructed feature output by the decoding layer of the initial model includes: The historical correlation statistical features are input into the encoder to obtain the initial encoded features of the historical correlation statistical features output by the encoder; The initial encoded features are input into the first feature embedding layer to obtain the embedded features of the initial encoded features output by the first feature embedding layer; The embedding features of the initial encoded features are input into the LSTM to obtain the first encoded features output by the LSTM; The first encoded feature is input into the second feature embedding layer to obtain the embedded feature of the first encoded feature output by the second feature embedding layer; The embedding features of the first encoded feature are input into the decoder to obtain the first reconstructed feature output by the decoder.

4. The unsupervised plunger pump anomaly detection method based on time-related feature reconstruction according to any one of claims 1-3, characterized in that, The statistical feature extraction of the current steady-state vibration signal of the target plunger pump yields multidimensional current operational statistical features of the target plunger pump, including: The current steady-state vibration signal is denoised based on the Savitzky-Golay digital filter. The noise-reduced current steady-state vibration signal is processed by sliding window to obtain multiple window data. Multi-dimensional statistical features are extracted from the data of each window to obtain the multi-dimensional current operating statistical features of the target plunger pump. The multidimensional current operating statistical features include the maximum value, minimum value, mean, peak-to-peak value, absolute mean, root mean square value, root mean square amplitude, standard deviation, kurtosis, skewness, margin index, waveform index, impulse index, peak index, kurtosis index, frequency center, average frequency, frequency domain root mean square, frequency variance, and the multidimensional energy spectrum features of the last layer of wavelet packet decomposition.

5. The unsupervised plunger pump anomaly detection method based on time-related feature reconstruction according to any one of claims 1-3, characterized in that, The step of inputting the current correlation statistical features into the detection model to obtain the fault prediction result of the target plunger pump includes: The current correlation statistical features are input into the detection model to obtain the second reconstructed features output by the encoding layer of the detection model; The reconstruction error between the second reconstructed feature and the current associated statistical feature is compared with the target error to obtain the fault prediction result of the target plunger pump.

6. An unsupervised plunger pump anomaly detection device based on time-related feature reconstruction, characterized in that, include: The feature extraction unit is used to extract statistical features from the current steady-state vibration signal of the target plunger pump to obtain the multidimensional current operating statistical features of the target plunger pump. The feature filtering unit is used to filter the current operating statistical features in multiple dimensions to obtain the current associated statistical features of the target plunger pump. The detection unit is used to input the current correlation statistical features into the detection model to obtain the fault prediction result of the target plunger pump; The early warning unit is used to generate early warning information for the target plunger pump based on the fault prediction results. The detection model is obtained by unsupervised training of an initial model based on the historical correlation statistical features of the first sample plunger pump. The historical correlation statistical features are obtained by extracting and filtering statistical features from the historical steady-state vibration signals of the first sample plunger pump. The initial model is constructed based on a variational autoencoder (VAE) and a long short-term memory network (LSTM) to mine reconstructed features related to feature time information. The first sample plunger pump is a sample plunger pump in normal operating condition. The device is also used for: The motor speed, motor torque, and pressure of the fracturing truck to which the target plunger pump belongs are obtained at multiple moments within the current cycle. Based on the information on the changes in motor speed, the information on the changes in motor torque, and the information on the changes in pressure at multiple times, the steady-state operating time of the target plunger pump is obtained within the current cycle. The vibration signals of the target plunger pump at multiple moments during the steady-state operation time are extracted to obtain the current steady-state operation vibration signal; The feature filtering unit is specifically used for: Statistical features were extracted from the historical steady-state vibration signal of the second sample plunger pump to obtain the multidimensional historical operational statistical features of the second sample plunger pump; the second sample plunger pump is a sample plunger pump of the same type as the target plunger pump. Calculate the monotonicity index of the historical operational statistical characteristics of each dimension; Based on the monotonicity index, at least one-dimensional candidate statistical features are selected from the multidimensional historical operation statistical features. Calculate the Pearson correlation coefficient, Spearman correlation coefficient, and Kendall coefficient among at least one of the candidate statistical features; The Pearson correlation coefficient, the Spearman correlation coefficient, and the Kendall coefficient are fused to obtain at least one-dimensional correlation index among the candidate statistical features; Based on the correlation index, the historical correlation statistical features of the second sample plunger pump are selected from at least one-dimensional candidate statistical features. Among the current operational statistical features described in the multidimensional analysis, statistical features that match the historical correlation statistical features are selected as the current correlation statistical features.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the unsupervised plunger pump anomaly detection method based on time-related feature reconstruction as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the unsupervised plunger pump anomaly detection method for time-related feature reconstruction as described in any one of claims 1 to 5.

Images