An edge-computing-based Christmas tree fault prediction and health management system

Abstract

This invention provides an edge computing-based fault prediction and health management system for wellhead equipment, relating to the field of wellhead equipment monitoring technology. It includes an edge node layer, a cloud management layer, and a cross-site federated collaboration layer. The edge node layer is deployed at the wellhead site for on-site preprocessing, fault prediction, and health assessment of wellhead sensor data. The cloud management layer is deployed at the oilfield monitoring center. This invention utilizes a lightweight edge-end fusion preprocessing architecture with a built-in multi-protocol adaptation stack, enabling plug-and-play adaptation to mainstream oilfield industrial protocols such as Modbus, CANopen, LoRaWAN, and HART. This solves the compatibility issues of sensors from different manufacturers, eliminating the need to modify existing sensors and reducing deployment costs. An improved translation-invariant wavelet denoising algorithm is optimized for low-computing-power edge scenarios, effectively eliminating noise interference caused by wellhead operating condition fluctuations at the edge.

Description

Technical Field

[0001] This invention relates to the field of oil wellhead equipment monitoring technology, and in particular to an oil wellhead fault prediction and health management system based on edge computing. Background Technology

[0002] As a core and critical piece of equipment in oilfield development, the operational status of the wellhead directly determines the safety, stability, and economy of oilfield production. A malfunction can lead to serious consequences such as oil spills and production shutdowns, causing significant economic losses and environmental risks. Currently, the monitoring and fault management of wellhead equipment mainly rely on traditional manual inspections and simple online monitoring methods. Some scenarios have introduced cloud-based centralized processing monitoring systems, which collect sensor data such as pressure, temperature, and vibration of the wellhead by deploying various sensors, and use data processing and analysis to achieve fault detection. Existing technologies include some systems attempting to use wavelet denoising and machine learning models for fault identification, while simultaneously enabling on-site data processing through edge nodes to reduce cloud bandwidth pressure. Regarding cross-site data interaction, some solutions attempt to improve model generalization capabilities through data sharing, providing support for wellhead fault early warning and maintenance.

[0003] However, in existing technologies, wellhead sensors come from different manufacturers and employ various industrial protocols such as Modbus, CANopen, LoRaWAN, and HART. Existing systems lack a unified multi-protocol adaptation mechanism, resulting in poor access compatibility and difficulties in plug-and-play functionality, leading to ineffective integration of data from different sensors. Furthermore, wellhead operating conditions fluctuate significantly, and sensor data is subject to strong noise interference. Existing noise reduction algorithms largely rely on high-performance cloud computing, failing to achieve efficient noise removal under low-performance edge conditions. This can easily lead to the loss of fault characteristic signals, affecting the accuracy of fault identification. Finally, wellhead fault samples are scarce, and the computing power of industrial-grade chips at the edge is limited. Existing fault prediction models are large in size and have poor generalization ability, failing to achieve millisecond-level inference on-site at the edge, thus failing to meet the requirements. The need for fault prediction in small-sample scenarios; existing systems lack a comprehensive health measurement and assessment system, mostly only detecting single fault types, failing to achieve a full-dimensional assessment of equipment structural integrity, operational performance, and sealing reliability, and having low accuracy in remaining life prediction, leading to unreasonable maintenance strategies and problems of over-maintenance or under-maintenance; existing systems are mostly deployed in isolation, resulting in serious data silos, and model iteration relies on uploading raw data, which not only consumes a lot of cloud bandwidth but also poses a risk of data privacy leakage. At the same time, cross-site data cannot be securely shared, making it difficult to improve the model's generalization ability and achieve collaborative management of all oilfield equipment. Therefore, this invention proposes an edge computing-based oil well tree fault prediction and health management system to solve the problems existing in the prior art. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes an edge computing-based tree fault prediction and health management system. This system utilizes a lightweight edge-end fusion preprocessing architecture with a built-in multi-protocol adaptation stack to achieve plug-and-play adaptation of mainstream oilfield industrial protocols such as Modbus, CANopen, LoRaWAN, and HART. This solves the compatibility issues of sensors from different manufacturers, eliminating the need to modify existing sensors and reducing deployment costs. An improved translation-invariant wavelet denoising algorithm is optimized for low-computing-power edge scenarios, effectively eliminating noise interference caused by wellhead operating condition fluctuations while preserving fault characteristic signals. The signal-to-noise ratio of the denoised data is improved by ≥20dB. Simultaneously, the federated feature alignment operator achieves high-rate data compression, significantly reducing cloud bandwidth usage and providing high-quality feature input for subsequent fault prediction models.

[0005] To achieve the objectives of this invention, the invention is implemented through the following technical solution: an edge computing-based tree fault prediction and health management system, comprising an edge node layer, a cloud management layer, and a cross-site federated collaboration layer. The edge node layer is deployed at the tree wellhead site for on-site preprocessing, fault prediction, and health assessment of tree sensor data. The cloud management layer is deployed at the oilfield monitoring center for full lifecycle management of tree equipment and global model optimization across the entire oilfield. The cross-site federated collaboration layer is used to achieve model knowledge sharing and data privacy protection between different oilfield sites.

[0006] The edge node layer incorporates a lightweight fusion preprocessing architecture, a lightweight fault prediction model, and a three-dimensional health assessment module. The cloud management layer and the edge node layer achieve collaborative iteration of the model through federated learning, forming an edge-cloud collaborative closed-loop architecture. The cross-site federated collaboration layer adopts a hierarchical federated learning framework to achieve cross-site knowledge sharing and privacy protection. The entire system is based on edge computing technology to complete fault prediction and health management of the entire oil production tree process.

[0007] Further improvements are made in the following aspects: The lightweight fusion preprocessing architecture of the edge node layer includes a multi-protocol adaptation stack, an improved translation-invariant wavelet denoising module, and a federated feature alignment module, which are connected in series to realize the entire process of sensor data from access, denoising to standardization processing on-site. The multi-protocol adaptation stack supports plug-and-play adaptation to mainstream oilfield industrial protocols such as Modbus, CANopen, LoRaWAN, and HART, and has a built-in protocol parsing unit and data format conversion unit. The protocol parsing unit is used to parse sensor data from different protocols, and the data format conversion unit is used to convert the parsed heterogeneous data into a unified format, solving the access compatibility problem of sensors from different manufacturers. The improved translation-invariant wavelet denoising module is used to remove noise interference caused by wellhead operating condition fluctuations and retain fault feature signals. The federated feature alignment module is used to complete the time-series synchronization, standardization processing, and key feature extraction of sensor data of different dimensions and ranges, and to achieve high-rate data compression.

[0008] Further improvements include: the multi-protocol adaptation stack adopts a modular design, with each industrial protocol corresponding to an independent adaptation module. The adaptation module can be dynamically loaded and unloaded according to the actual type of sensor accessed. The adaptation process includes four steps: protocol identification, parameter configuration, data parsing, and format conversion. Specifically, the protocol identification step automatically matches the corresponding adaptation module by detecting the protocol identification information sent by the sensor; the parameter configuration step automatically configures the parsing parameters of the adaptation module based on parameters such as the sensor's range and sampling frequency; the data parsing step parses the raw data sent by the sensor to extract valid sensing parameters; and the format conversion step converts the parsed valid parameters into a preset unified data format to ensure compatibility between different sensor data.

[0009] The further improvement lies in the following: the improved translation-invariant wavelet denoising module employs an improved translation-invariant wavelet denoising algorithm, which is optimized for low-computing-power scenarios at the edge. Its calculation formula is as follows:

[0010] ,

[0011] Where y(t) represents the acquired noisy sensor data, x(t) represents the original noise-free sensor data, h(t) represents the wavelet basis function, and k represents the translation coefficient. Noise signals caused by fluctuations in wellhead operating conditions. The sensor data after noise reduction. This is an improved translation-invariant wavelet transform operator. The adaptive noise reduction threshold is adaptively adjusted based on the variance of the sensor data and the noise intensity, ensuring efficient noise reduction under low computing power conditions.

[0012] Further improvements are made in that: the federated feature alignment module employs a federated feature alignment operator to achieve temporal synchronization, standardization, and key feature extraction of sensor data with different dimensions and ranges, achieving a data compression ratio of no less than 15:1. The key feature extraction calculation formula is as follows:

[0013] ;

[0014] Where F is the extracted key feature vector, Z is the multi-dimensional sensor data matrix after time synchronization, W is the feature weight matrix used to highlight fault-related features, and b is the bias term. The federated feature alignment operator eliminates the differences between data of different ranges through standardization, keeps the timestamps of data of different dimensions consistent through time synchronization algorithm, and extracts key features related to faults through feature screening to achieve high-rate data compression.

[0015] Further improvements are made in the following aspects: The lightweight fault prediction model is based on a Siamese network and integrates a lightweight multi-head attention mechanism. After optimization through three steps—pre-training, transfer tuning, and quantization pruning—the model size is compressed to less than 8MB, enabling it to run locally on industrial-grade low-power edge chips and achieve millisecond-level inference. The model achieves full-dimensional prediction of six core faults in the oil production tree, including valve internal leakage, flange seal failure, casing corrosion, actuator jamming, safety valve failure, and pressure anomalies. It also completes fault type localization, fault severity classification, and remaining service life (RUL) prediction. The pre-training step uses a large sample dataset of general rotating machinery faults to learn general fault feature extraction capabilities. The transfer tuning step uses a small amount of labeled fault data from the oil production tree field to fine-tune the parameters of the pre-trained model, improving the model's adaptability to oil production tree faults. The quantization pruning step converts the model parameters from 32-bit floating-point to 8-bit integer through quantization and removes redundant neurons and connections in the model, reducing model size and computational consumption.

[0016] A further improvement lies in the fact that the transfer fine-tuning process of the lightweight fault prediction model adopts a transfer learning algorithm, and its loss function is calculated as follows:

[0017] ;

[0018] Where L is the total loss function, α is the weight coefficient with a value ranging from 0.6 to 0.8, used to balance classification loss and regression loss, Lcls is the fault type classification loss function, which adopts the cross-entropy loss function to optimize the fault classification accuracy of the model, and Lreg is the fault severity and RUL regression loss function, which adopts the mean squared error loss function to optimize the regression accuracy of the model and ensure that the fault identification accuracy is ≥96% in small sample scenarios.

[0019] Further improvements are made in the following aspects: The three-dimensional health assessment module is used to establish a three-dimensional health assessment system covering structural integrity, operational performance, and sealing reliability, including a health index calculation unit and a remaining life prediction unit. The health index calculation unit calculates structural integrity, operational performance, and sealing reliability indices based on preprocessed sensor data, determines the weight of each index using the analytic hierarchy process (AHP), and quantifies the overall equipment health score. The remaining life prediction unit uses an improved particle filter algorithm to predict the remaining effective life of key components such as the main valve and wing valve online. The calculation formula is as follows:

[0020] ;

[0021] in, The remaining service life is the predicted value, where t is the current time and T is the time of equipment failure. For all sensor data up to and including the current moment, The posterior probability density function at the time of failure is updated iteratively using an improved particle filter algorithm to ensure the accuracy of RUL prediction, with an early fault warning lead time of ≥72h.

[0022] Further improvements are made in the following ways: The edge-cloud collaborative closed-loop architecture adopts an edge-to-ground inference + cloud-to-global iteration mode. The edge node layer and the cloud management layer are connected through an encrypted communication link, supporting data interruption resumption. The edge node layer is responsible for real-time data processing, local fault early warning, emergency control, and local data caching. When the network is interrupted, the edge node layer operates independently to complete fault prediction and health assessment. After the network is restored, it automatically uploads the cached data and operation logs to the cloud management layer. The cloud management layer is responsible for the full life cycle ledger management of all oilfield production tree equipment, fault root cause analysis, global model optimization, and maintenance decision management. The model iteration adopts a horizontal federated learning framework. Each edge node only uploads the model gradient parameters. The cloud management layer aggregates the gradient parameters through a federated aggregation algorithm to generate a globally optimized model, and then distributes the optimized model to each edge node for updating, without interrupting the edge-side inference business throughout the process.

[0023] Further improvements are made in the following aspects: The cross-site federated collaboration layer adopts a hierarchical federated learning framework, including an edge node layer, a regional aggregation layer, and a global aggregation layer. It combines contrastive learning and differential privacy protection techniques to achieve cross-site knowledge sharing and data privacy protection. Contrastive learning is used to enhance the commonality of cross-site fault characteristics, and its loss function is calculated as follows:

[0024] ;

[0025] Where Lcontrast is the contrastive learning loss, N is the number of samples, and xi is the anchor sample. For positive samples of the same type as anchor samples, For negative samples that are of a different class than the anchor sample, For feature extraction function, The cosine similarity function is used. The temperature coefficient is used; differential privacy protection technology adds Gaussian noise to the uploaded model parameters to protect parameter privacy, ensuring that cross-site data is not leaked, while improving the model's generalization ability.

[0026] The beneficial effects of this invention are as follows:

[0027] 1. This invention utilizes a lightweight edge-end fusion preprocessing architecture with a built-in multi-protocol adaptation stack to achieve plug-and-play adaptation of mainstream oilfield industrial protocols such as Modbus, CANopen, LoRaWAN, and HART. This solves the compatibility issues of sensors from different manufacturers, eliminating the need to modify existing sensors and reducing deployment costs. The improved translation-invariant wavelet denoising algorithm is optimized for low-computing-power scenarios at the edge, effectively eliminating noise interference caused by wellhead operating condition fluctuations while retaining fault feature signals. The signal-to-noise ratio of the denoised data is improved by ≥20dB. Simultaneously, the federated feature alignment operator achieves high-rate data compression, significantly reducing cloud bandwidth usage and providing high-quality feature input for subsequent fault prediction models. Compared to existing technologies, data processing efficiency is improved by more than 50%, and data quality is significantly improved.

[0028] 2. The lightweight fault prediction model designed in this invention uses a pre-training-transfer fine-tuning mode. It is pre-trained using a large sample dataset of general rotating machinery faults, and then fine-tuned using a small amount of labeled fault data from the oil well site. This effectively solves the problem of poor model generalization ability in small sample scenarios. After quantization and pruning optimization, the model size is compressed to less than 8MB, enabling millisecond-level inference on industrial-grade low-power edge chips. Compared with existing models, the size is reduced by more than 70%, and the inference speed is improved by more than 60%. This model can achieve full-dimensional prediction of 6 major categories of core faults. In small sample scenarios, the fault identification accuracy is ≥96%, the early fault warning lead time is ≥72h, and the fault type positioning accuracy is ≤10cm. It can effectively achieve early warning and accurate positioning of oil well faults, and prevent the fault from escalating.

[0029] 3. The three-dimensional health assessment system constructed in this invention, encompassing structural integrity, operational performance, and sealing reliability, can comprehensively quantify the health status of the wellhead. The weights of each health indicator are determined through the analytic hierarchy process (AHP), ensuring the scientific rigor and accuracy of the health assessment. The improved particle filter algorithm enables accurate online prediction of the remaining lifespan of key components, with a prediction error ≤10%. The system automatically adjusts its inspection strategy based on the health score, transforming traditional "scheduled maintenance" into "condition-based maintenance," effectively avoiding over-maintenance or under-maintenance. Compared to the existing scheduled maintenance model, maintenance costs are reduced by more than 30%, and equipment lifespan is extended by more than 15%, providing a reliable basis for oilfield production planning.

[0030] 4. This invention constructs an edge-cloud collaborative closed-loop architecture of edge-to-ground inference and cloud-based global iteration. The edge supports offline operation during network outages and automatically resumes transmission after network recovery, ensuring system availability in harsh network environments. The cloud enables full lifecycle management and global model optimization for all oilfield production tree equipment. Model iteration adopts a horizontal federated learning framework, where each edge node only uploads model gradient parameters, not raw data, solving the data silo problem and avoiding the privacy leakage risk caused by uploading raw data. Data security compliance meets relevant national standards. The edge-based inference business is not interrupted during model iteration, enabling continuous model iteration and optimization. Compared with existing systems, model iteration efficiency is improved by more than 40%, and the level of collaborative management of all oilfield equipment is significantly improved.

[0031] 5. The hierarchical federated learning framework adopted in this invention combines contrastive learning and differential privacy protection technology. Edge nodes only upload model parameters and do not transmit raw data. Contrastive learning enhances the commonality of fault features across sites, improving the model's generalization ability. Differential privacy protection technology ensures parameter privacy and avoids data leakage. Compared with existing cross-site data sharing schemes, this invention not only achieves knowledge sharing between different oilfield sites, improving the fault identification accuracy of each site's model by ≥1.5%, but also ensures data privacy and security, reaching the national level three privacy protection standard. This provides strong support for the optimization of the oilfield-wide tree fault prediction model. Attached Figure Description

[0032] Figure 1 This is a diagram illustrating the composition of the present invention. Detailed Implementation

[0033] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0034] Example 1

[0035] according to Figure 1As shown, this embodiment proposes an edge computing-based oil well tree fault prediction and health management system, including a specific implementation of a lightweight fusion preprocessing architecture at the edge. This architecture is deployed at the edge node of the oil wellhead and uses an industrial-grade low-power edge chip (such as STM32H743) with a computing power of 500MIPS and a power consumption of ≤5W, meeting the deployment requirements of low power consumption and small size at the wellhead.

[0036] The multi-protocol adapter stack adopts a modular design, with four independent adapter modules built-in: Modbus, CANopen, LoRaWAN, and HART. Each module occupies ≤512KB of storage space and can be dynamically loaded according to the actual type of sensor connected. When sensors from different manufacturers are deployed at the wellhead of the wellhead, edge nodes detect the protocol identification information sent by the sensors through interfaces such as serial ports and LoRa modules, and automatically match the corresponding adapter modules: For Modbus protocol sensors, the adapter module adopts RTU mode, with a baud rate set to 9600bps, 8 data bits, 1 stop bit, and no parity bit, and parses the register data sent by the sensor to extract effective parameters such as pressure and temperature; for CANopen protocol sensors, the adapter module connects through the CAN bus interface, parses PDO (Process Data Object) and SDO (Service Data Object) data, and extracts parameters such as vibration and displacement; for LoRaWAN protocol sensors, the adapter module connects through a LoRa gateway, adopts Class A mode, sets the spreading factor to 12, and parses the wirelessly transmitted sensor data; for HART protocol sensors, the adapter module connects through a HART modem, parses the digital data in the 4-20mA current signal, and extracts parameters such as flow rate and liquid level. All parsed parameters are converted into a unified JSON format through the data format conversion unit. The data sampling frequency can be adjusted according to the working conditions, ranging from 1Hz to 10Hz, to ensure compatibility and integration of sensor data from different manufacturers.

[0037] The improved translation-invariant wavelet denoising module uses the db4 wavelet as the basis function and sets the decomposition level to 3. The adaptive denoising threshold λ is adjusted in real time according to the variance of the sensor data and the noise intensity. Specifically, it is calculated as λ = σ × √(2 × lnN), where σ is the noise standard deviation and N is the data length. After acquiring noisy sensor data, the module first normalizes the data, mapping the data range to [0,1]. Then, it performs 3-level wavelet decomposition to obtain cA3 (3-level low-frequency components), cD1, cD2, and cD3 (1-3-level high-frequency components). The high-frequency components cD1, cD2, and cD3 are thresholded using an adaptive threshold to remove the high-frequency components corresponding to noise and retain the high-frequency components corresponding to fault characteristics. Finally, the processed low-frequency components and high-frequency components are subjected to inverse wavelet transform to obtain the denoised sensor data. Tests showed that the module can process 1,000 data points at the edge in ≤50ms, and the signal-to-noise ratio of the data after noise reduction is improved by ≥20dB. It can effectively eliminate noise interference caused by wellhead operating condition fluctuations and retain fault characteristic signals.

[0038] The federated feature alignment module first performs time-series synchronization processing on the multi-dimensional sensor data, using linear interpolation to fill in missing timestamps, ensuring consistency in timestamps across different dimensions with a synchronization accuracy ≤100ms. Then, it uses Z-score normalization to standardize the synchronized data, eliminating differences between data from different measurement ranges. The normalization formula is as follows: The algorithm uses a federated feature alignment operator to extract key features. The feature weight matrix W is trained using a stochastic gradient descent algorithm, highlighting fault-related features such as pressure anomalies and vibration peaks. The bias term b is set to 0.01. The extracted key features include time-domain features (mean, variance, peak value, kurtosis) and frequency-domain features (fundamental frequency, harmonic amplitude), totaling 12-dimensional feature vectors. Testing shows that this module achieves a data compression ratio of 18:1, significantly reducing cloud bandwidth usage. The compressed feature data volume is ≤1KB / s, providing high-quality feature input for subsequent fault prediction models.

[0039] Example 2

[0040] according to Figure 1 As shown, this embodiment proposes an edge computing-based oil well tree fault prediction and health management system, including a specific implementation of a lightweight fault prediction model. This model is deployed on an industrial-grade low-power edge chip at the edge node, with the model size compressed to 7.2MB and the inference time ≤50ms, meeting the millisecond-level inference requirements of the edge.

[0041] The model is based on a Siamese network, which consists of two identical feature extraction branches with shared parameters. Each branch comprises four convolutional layers, two pooling layers, and one fully connected layer. The convolutional layers use 3×3 kernels with 32, 64, 128, and 128 kernels respectively, and employ ReLU activation to avoid gradient vanishing. The pooling layers use 2×2 max pooling with a stride of 2 to reduce feature dimensionality. The fully connected layer has 256 neurons, and the output feature vector has a dimension of 128. A lightweight multi-head attention mechanism employs four attention heads, each with a dimension of 32. By assigning attention weights to the feature vectors output by the feature extraction branches, fault-related features are highlighted, irrelevant features are suppressed, and the model's feature extraction capability is improved.

[0042] The model was pre-trained using a large-sample dataset of general rotating machinery faults (such as the CWRU bearing fault dataset and the XJTU-SY gear fault dataset). This dataset includes various fault types, such as normal faults, inner race faults, outer race faults, and rolling element faults, with over 100,000 samples, a sampling frequency of 10kHz, and a data length of 1024. The pre-training process used the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 100 iterations. The cross-entropy loss function was used. After pre-training, the model achieved a general fault feature extraction accuracy of ≥98%.

[0043] The transfer learning fine-tuning employed a small dataset of labeled fault data from the oilfield. This dataset included six major fault categories: valve internal leakage, flange seal failure, casing corrosion, actuator jamming, safety valve failure, and abnormal pressure. Each fault category had 50-100 labeled samples, with 500 normal samples. The dataset length was 1024 bytes, and the sampling frequency was 5kHz. During transfer learning fine-tuning, the first three convolutional layers of the Siamese network's feature extraction branch were frozen, and only the parameters of the fourth convolutional layer, the fully connected layer, and the multi-head attention mechanism were fine-tuned. The Adam optimizer was used, with a learning rate of 0.0001, a batch size of 16, and 50 iterations. The total loss function was used, with an α value of 0.7. After fine-tuning, the model achieved a fault identification accuracy of 97.2%, meeting the fault prediction requirements in small-sample scenarios.

[0044] Model quantization and pruning optimization utilize TensorRT to convert model parameters from 32-bit floating-point to 8-bit integer, compressing the model size from 32MB to 7.2MB. Structured pruning is employed, removing redundant connections and neurons with weights less than 0.01, achieving a pruning ratio of 60%. This pruning reduces inference time from 120ms to 45ms and computational power consumption by 50%, enabling stable operation on industrial-grade low-power edge chips. Field testing demonstrates that the model accurately identifies six core fault categories in the oil wellhead, with fault type location accuracy ≤10cm. Fault severity is categorized into mild, moderate, and severe levels with a classification accuracy ≥95%. Remaining lifetime prediction error is ≤10%, and early fault warning leads to 80 hours, effectively achieving early warning and precise location of oil wellhead faults.

[0045] Example 3

[0046] according to Figure 1 As shown, this embodiment proposes an edge computing-based tree failure prediction and health management system, including a specific implementation of a three-dimensional health assessment system. This system is deployed on edge nodes and works in conjunction with a lightweight fusion preprocessing architecture and a lightweight failure prediction model to achieve quantitative assessment of tree health and prediction of remaining life.

[0047] The three-dimensional health assessment system includes three dimensions: structural integrity, operational performance, and sealing reliability. Each dimension contains several health indicators (HI): the structural integrity dimension includes three indicators: sleeve wall thickness, valve wear, and flange deformation; the operational performance dimension includes three indicators: valve opening and closing response time, actuator output torque, and pressure regulation accuracy; and the sealing reliability dimension includes three indicators: sealing surface leakage, sealing pressure holding time, and sealing ring aging degree, for a total of nine health indicators.

[0048] The health indicator calculation unit calculates the specific values ​​of each health indicator based on the standardized sensor data output by the lightweight fusion preprocessing architecture: the sleeve wall thickness is calculated using data collected by the ultrasonic sensor, and the calculation formula is as follows: Where v is the ultrasonic wave propagation speed and t is the ultrasonic wave round-trip time; valve wear is calculated from the vibration peak value collected by the vibration sensor, and the wear is positively correlated with the vibration peak value, and the wear value is obtained through linear fitting; flange deformation is directly obtained from the data collected by the displacement sensor; valve opening and closing response time is calculated from the valve opening and closing time collected by the stroke sensor; actuator output torque is directly obtained from the data collected by the torque sensor; pressure regulation accuracy is calculated from the pressure fluctuation value collected by the pressure sensor, and the smaller the fluctuation value, the higher the regulation accuracy; sealing surface leakage is calculated from the leakage flow rate collected by the flow sensor; sealing pressure holding time is directly obtained from the pressure holding time collected by the pressure sensor; sealing ring aging degree is calculated by combining the sealing ring temperature collected by the temperature sensor and the service time, and the aging degree is positively correlated with temperature and service time.

[0049] The weights of each health indicator were determined using the analytic hierarchy process (AHP): Structural integrity had a weight of 0.4, with sleeve wall thickness at 0.15, valve wear at 0.15, and flange deformation at 0.1; operational performance had a weight of 0.3, with valve on / off response time at 0.1, actuator output torque at 0.1, and pressure regulation accuracy at 0.1; and sealing reliability had a weight of 0.3, with sealing surface leakage at 0.1, sealing pressure holding time at 0.1, and sealing ring aging at 0.1. The health score was calculated on a 100-point scale using the following formula: Where HITtotal is the overall health score, wi is the weight of the i-th health indicator, and HII is the score of the i-th health indicator (0-100 points). Based on the health score, the equipment health status is divided into four levels: Excellent (90-100 points), Good (80-89 points), Acceptable (60-79 points), and Unacceptable (<60 points). The system automatically adjusts the inspection strategy according to the health status: Excellent status is inspected once every 7 days, Good status once every 3 days, Acceptable status once a day, and Unacceptable status immediately triggers inspection and maintenance warnings.

[0050] The remaining lifetime prediction unit employs an improved particle filter algorithm, with the main valve as the prediction target, a particle count of 1000, and a linear state equation. The observation equation adopts Where xk is the state vector of the main valve at time k (including wear, fatigue level, etc.), xk+1 is the state vector of the main valve at time k+1, A is the state transition matrix, B is the input matrix, uk is the input vector at time k (including working pressure, working temperature, etc.), wk is the process noise, following a Gaussian distribution N(0,Q), Q is the process noise covariance matrix, zk is the observation vector at time k (i.e., preprocessed sensor data), C is the observation matrix, vk is the observation noise, following a Gaussian distribution N(0,R), and R is the observation noise covariance matrix. The improved particle filtering algorithm enhances particle diversity and avoids particle degradation by adaptively adjusting the particle weight update strategy. The predicted remaining service life error of the main valve is ≤8%, providing a reliable decision-making basis for oilfield production planning and maintenance strategies.

[0051] Example 4

[0052] according to Figure 1 As shown, this embodiment proposes an edge computing-based oil well tree fault prediction and health management system, including a specific implementation of an edge-cloud collaborative closed-loop architecture. This architecture consists of an edge node layer (100 wellhead edge nodes) and a cloud management layer (1 oilfield monitoring center cloud platform). The edge nodes and the cloud are connected via a 5G / fiber optic communication link with a communication latency of ≤50ms. Data transmission uses the AES-256 encryption algorithm to ensure data transmission security.

[0053] Each edge node in the edge node layer deploys one edge controller, several sensors (pressure, temperature, vibration, displacement, etc.), one local cache module (16GB capacity), and one emergency control module. The edge controller uses an industrial-grade edge gateway, supporting multi-protocol access and low-computing-power data processing. The local cache module is used to cache pre-processed sensor data, fault prediction results, and operation logs. The cache time can be set to 1-7 days. When the network is interrupted, the edge node automatically switches to offline operation mode, while the edge controller continues to complete data preprocessing, fault prediction, and health assessment. Based on the fault prediction results, the emergency control module automatically triggers an emergency valve shut-off command when a serious fault occurs (such as safety valve failure or crude oil leakage) to prevent the fault from escalating. When the network is restored, the edge node automatically resumes the transmission of cached data and operation logs to the cloud management layer, ensuring that no data is lost.

[0054] The cloud-based management layer is deployed at the oilfield monitoring center, using a cloud server cluster (4 cloud servers, each with 16 CPU cores, 64GB of memory, and 1TB of hard drive) to build a cloud-based management platform. This platform enables full lifecycle ledger management, root cause analysis of failures, global model optimization, and maintenance decision management for all oilfield production tree equipment. The full lifecycle ledger management module records equipment information (model, installation time, maintenance records, etc.), operational data, and failure records for each production tree, supporting searches by keywords such as equipment number and failure type. The root cause analysis module uses failure prediction results and sensor data uploaded from edge nodes, combined with a failure knowledge base, to analyze the root causes of failures using decision tree algorithms (e.g., valve internal leakage may be caused by aging seals, media corrosion, etc.) and provides corresponding maintenance suggestions. The maintenance decision management module automatically generates maintenance plans based on equipment health scores and remaining life prediction results, allocates maintenance personnel and materials, and tracks maintenance progress.

[0055] The model's global iteration employs a horizontal federated learning framework with an iteration cycle of 7 days. During each iteration, each edge node trains the model based on local data, calculates the model's gradient parameters (gradient parameter volume ≤ 100KB), and uploads them to the cloud management layer via an encrypted communication link. The cloud management layer then aggregates the gradient parameters uploaded by all edge nodes using a federated averaging algorithm to generate a globally optimized model. The aggregation formula is as follows: Here, Wglobal represents the global model parameters, N is the number of edge nodes, and Wi is the model gradient parameters of the i-th edge node. After aggregation, the cloud management layer distributes the globally optimized model to each edge node, and the edge nodes automatically update their local models without interrupting the on-site inference service at the edge. Testing showed that after three federated iterations, the model's fault identification accuracy improved from 97.2% to 98.5%, achieving continuous iterative optimization of the model while avoiding bandwidth consumption and data security risks associated with uploading raw data.

[0056] Example 5

[0057] according to Figure 1 As shown, this embodiment proposes an edge computing-based oilfield tree fault prediction and health management system, including a specific implementation of a cross-site federated learning mechanism. This mechanism is used to connect three different oilfield sites (site A, site B, and site C), with 30-40 edge nodes deployed at each site, to achieve cross-site knowledge sharing and data privacy protection.

[0058] The cross-site hierarchical federated learning framework includes an edge node layer, a regional aggregation layer, and a global aggregation layer: the edge node layer consists of edge nodes for each oilfield site, responsible for local model training and uploading gradient parameters; the regional aggregation layer consists of regional aggregation nodes for each oilfield site, responsible for aggregating the gradient parameters of all edge nodes at this site to generate regional model parameters; the global aggregation layer consists of global aggregation nodes at the cloud management layer, responsible for aggregating the regional model parameters of all regional sites, generating global model parameters, and distributing them to each regional aggregation node and edge node.

[0059] Contrastive learning is used to enhance the commonality of cross-site fault features. During local training, each edge node randomly selects anchor point samples xi and positive samples from the local dataset. (Same type of fault as anchor point sample) and negative sample (For faults of a different class than those of the anchor sample), the feature vector of the sample is extracted using the feature extraction function f(·), and the cosine similarity is calculated. and The loss value was calculated based on the contrastive learning loss function to optimize the model's feature extraction capability, resulting in higher similarity for similar fault features across different sites and lower similarity for dissimilar fault features. The temperature coefficient τ for contrastive learning was set to 0.1, and the sample size N was set to 64. After contrastive learning optimization, the similarity of similar fault features across sites increased by ≥30%.

[0060] Differential privacy protection is used to protect the privacy of model parameters. Before each edge node uploads its gradient parameters, Gaussian noise is added to the gradient parameters. The noise intensity is set according to the privacy budget ε, with ε set to 1.0 (the smaller the privacy budget, the higher the level of privacy protection). The Gaussian noise follows a set order. Distribution, among which , This represents the sensitivity of the gradient parameter (with a value of 0.1). The probability of privacy leakage (with a value of 10) -5 The gradient parameters with added noise are uploaded to the regional aggregation node. The regional aggregation node aggregates the gradient parameters of its local site, adds Gaussian noise again, and then uploads them to the global aggregation node to ensure that the gradient parameters do not leak the original data information during transmission.

[0061] Tests showed that after the cross-site federated learning mechanism was implemented, the fault identification accuracy of the models at each site improved by ≥1.5%. Specifically, the accuracy of the model at site A increased from 97.8% to 99.4%, that at site B from 97.1% to 98.7%, and that at site C from 96.9% to 98.5%. At the same time, through differential privacy protection technology, even if the gradient parameters are stolen, the original sensor data cannot be restored. The data privacy protection level reaches the national level three privacy protection standard, realizing the synergy between data privacy and knowledge sharing.

[0062] Validation data:

[0063] This invention has undergone on-site testing and verification, and all performance indicators have met the design requirements. Specific data summary is as follows:

[0064] Multi-protocol adaptation performance: Supports more than 4 mainstream oilfield industrial protocols such as Modbus, CANopen, LoRaWAN, and HART, with plug-and-play adaptation time ≤30s and 100% compatibility with sensors from different manufacturers; Data processing performance: Improved translation-invariant wavelet denoising increases the signal-to-noise ratio by ≥20dB, data compression ratio ≥15:1 (actual test reaches 18:1), edge data processing latency ≤50ms, and data processing efficiency is improved by more than 50%; Fault prediction performance: Model size ≤8MB (actual test 7.2MB), edge inference time ≤50ms (actual test 45ms), fault identification accuracy ≥96% in small sample scenarios (actual test 98.5%), early fault warning lead time ≥72h (actual test 80h), full coverage of 6 major categories of core faults, and fault type positioning accuracy ≤10cm. The system boasts the following performance characteristics: Fault severity classification accuracy ≥95%; Health assessment and lifespan prediction performance: 3D health assessment accuracy ≥98%, critical component remaining lifespan prediction error ≤10% (actual test error 8%), optimized inspection strategy reduces maintenance costs by over 30%, and extends equipment lifespan by over 15%; Edge-cloud collaboration and cross-site performance: Edge-cloud communication latency ≤50ms, offline operation time ≥7 days, 100% success rate of resume transmission after network recovery, improved model federation iteration efficiency by over 40%, improved cross-site model fault identification accuracy by ≥1.5%, and data privacy protection level meets the national Level 3 privacy protection standard; Overall operational performance: System operational stability ≥99.5%, edge node power consumption ≤5W, deployment cost reduced by over 25%, and stable operation in harsh wellhead environments ranging from -40℃ to 85℃, meeting the actual application needs of oilfields.

[0065] This edge computing-based oil well tree fault prediction and health management system achieves plug-and-play adaptation of mainstream oilfield industrial protocols such as Modbus, CANopen, LoRaWAN, and HART through a built-in multi-protocol adaptation stack in a lightweight fusion preprocessing architecture at the edge. This solves the access compatibility problem of sensors from different manufacturers, eliminating the need to modify existing sensors and reducing deployment costs. The improved translation-invariant wavelet denoising algorithm is optimized for low-computing-power scenarios at the edge, effectively eliminating noise interference caused by wellhead operating condition fluctuations at the edge while retaining fault feature signals. The signal-to-noise ratio of the denoised data is improved by ≥20dB. At the same time, the federated feature alignment operator achieves high-ratio data compression (compression ratio ≥15:1), significantly reducing cloud bandwidth usage and providing high-quality feature input for subsequent fault prediction models. Compared with existing technologies, data processing efficiency is improved by more than 50%, and data quality is significantly improved. The lightweight fault prediction model designed in this invention utilizes a pre-training-transfer fine-tuning approach. Pre-training is performed using a large-sample dataset of general rotating machinery faults, followed by fine-tuning using a small amount of labeled fault data from the oil well site. This effectively solves the problem of poor model generalization ability in small-sample scenarios. After quantization and pruning optimization, the model size is compressed to less than 8MB, enabling millisecond-level inference (inference time ≤50ms) on an industrial-grade low-power edge chip. Compared to existing models, this represents a reduction in size of over 70% and an improvement in inference speed of over 60%. This model can achieve full-dimensional prediction of six major categories of core faults, with a fault identification accuracy ≥96% in small-sample scenarios, an early fault warning lead time ≥72 hours, and a fault type location accuracy ≤10cm. It effectively enables early warning and precise location of oil well faults, preventing fault escalation. This invention constructs a three-dimensional health assessment system covering structural integrity, operational performance, and sealing reliability. This system can comprehensively quantify the health status of the wellhead. The weight of each health indicator is determined by the analytic hierarchy process (AHP) to ensure the scientific nature and accuracy of the health assessment. The improved particle filter algorithm enables online and accurate prediction of the remaining lifespan of key components with a prediction error of ≤10%. The system automatically adjusts the inspection strategy based on the health score, transforming the traditional "scheduled maintenance" into "condition-based maintenance," effectively avoiding over-maintenance or under-maintenance. Compared with the existing scheduled maintenance model, the operation and maintenance cost is reduced by more than 30%, and the equipment service life is extended by more than 15%, providing a reliable decision-making basis for oilfield production planning.The "edge-cloud collaborative closed-loop architecture" constructed in this invention, which combines "edge-side in-situ inference with cloud-side global iteration," supports offline operation during network outages at the edge, with automatic resume capability after network recovery, ensuring system availability in harsh network environments. The cloud enables full lifecycle management and global model optimization for all oilfield production tree equipment. Model iteration employs a horizontal federated learning framework, where each edge node uploads only model gradient parameters, not raw data, completely resolving the data silo problem and avoiding privacy risks associated with uploading raw data. Data security compliance meets relevant national standards. The edge-side inference process is uninterrupted during model iteration, enabling continuous model iteration and optimization. Compared to existing systems, model iteration efficiency is improved by over 40%, significantly enhancing the collaborative management level of all oilfield equipment. This invention employs a hierarchical federated learning framework that combines contrastive learning with differential privacy protection technology. Edge nodes only upload model parameters and do not transmit raw data. Contrastive learning enhances the commonality of fault characteristics across sites, improving the model's generalization ability, while differential privacy protection technology ensures parameter privacy and prevents data leakage. Compared with existing cross-site data sharing schemes, this invention achieves knowledge sharing between different oilfield sites, improving the fault identification accuracy of models at each site by ≥1.5%, while also ensuring data privacy and security. The privacy protection level reaches the national level three privacy protection standard, providing strong support for the optimization of oilfield-wide tree fault prediction models.

[0066] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An edge computing-based oil well tree fault prediction and health management system, comprising an edge node layer, a cloud management layer, and a cross-site federated collaboration layer, characterized in that: The edge node layer is deployed at the wellhead of the wellhead for on-site preprocessing, fault prediction, and health assessment of wellhead sensor data. The cloud management layer is deployed at the oilfield monitoring center for full lifecycle management of wellhead equipment and global model optimization. The cross-site federated collaboration layer is used to achieve model knowledge sharing and data privacy protection between different oilfield sites. The edge node layer incorporates a lightweight fusion preprocessing architecture, a lightweight fault prediction model, and a three-dimensional health assessment module. The cloud management layer and the edge node layer achieve collaborative iteration of the model through federated learning, forming an edge-cloud collaborative closed-loop architecture. The cross-site federated collaboration layer adopts a hierarchical federated learning framework to achieve cross-site knowledge sharing and privacy protection. The entire system is based on edge computing technology to complete fault prediction and health management of the entire oil production tree process.

2. The edge computing-based oil well tree fault prediction and health management system according to claim 1, characterized in that: The lightweight fusion preprocessing architecture of the edge node layer includes a multi-protocol adaptation stack, an improved translation-invariant wavelet denoising module, and a federated feature alignment module, which are connected in series to realize the entire process of sensor data from access, denoising to standardization processing on-site. The multi-protocol adaptation stack supports plug-and-play adaptation of mainstream oilfield industrial protocols such as Modbus, CANopen, LoRaWAN, and HART, and has built-in protocol parsing and data format conversion units. The protocol parsing unit is used to parse sensor data from different protocols, and the data format conversion unit is used to convert the parsed heterogeneous data into a unified format, solving the access compatibility problem of sensors from different manufacturers. The improved translation-invariant wavelet denoising module is used to remove noise interference caused by wellhead operating condition fluctuations and retain fault feature signals. The federated feature alignment module is used to complete the time-series synchronization, standardization processing and key feature extraction of sensor data of different dimensions and ranges, and to achieve high-rate compression of the data.

3. The edge computing-based oil well tree fault prediction and health management system according to claim 2, characterized in that: The multi-protocol adaptation stack adopts a modular design, with each industrial protocol corresponding to an independent adaptation module. The adaptation module can be dynamically loaded and unloaded according to the actual type of sensor connected. The adaptation process includes four steps: protocol identification, parameter configuration, data parsing, and format conversion. Specifically, the protocol identification step automatically matches the corresponding adaptation module by detecting the protocol identification information sent by the sensor; the parameter configuration step automatically configures the parsing parameters of the adaptation module according to the sensor's range, sampling frequency, and other parameters; the data parsing step parses the raw data sent by the sensor and extracts the effective sensing parameters; and the format conversion step converts the parsed effective parameters into a preset unified data format to ensure compatibility between different sensor data.

4. The edge computing-based oil well tree fault prediction and health management system according to claim 3, characterized in that: The improved translation-invariant wavelet denoising module employs an improved translation-invariant wavelet denoising algorithm, optimized for low-computing-power scenarios at the edge. Its calculation formula is as follows: ， Where y(t) represents the acquired noisy sensor data, x(t) represents the original noise-free sensor data, h(t) represents the wavelet basis function, and k represents the translation coefficient. Noise signals caused by fluctuations in wellhead operating conditions. The sensor data after noise reduction. This is an improved translation-invariant wavelet transform operator. The adaptive noise reduction threshold is adaptively adjusted based on the variance of the sensor data and the noise intensity, ensuring efficient noise reduction under low computing power conditions.

5. The edge computing-based oil well tree fault prediction and health management system according to claim 4, characterized in that: The federated feature alignment module employs a federated feature alignment operator to achieve temporal synchronization, standardization, and key feature extraction of sensor data with different dimensions and ranges. The data compression ratio is no less than 15:1, and the key feature extraction calculation formula is as follows: ； Where F is the extracted key feature vector, Z is the multi-dimensional sensor data matrix after time synchronization, W is the feature weight matrix used to highlight fault-related features, and b is the bias term. The federated feature alignment operator eliminates the differences between data of different ranges through standardization, keeps the timestamps of data of different dimensions consistent through time synchronization algorithm, and extracts key features related to faults through feature screening to achieve high-rate data compression.

6. The edge computing-based oil well tree fault prediction and health management system according to claim 1, characterized in that: The lightweight fault prediction model is based on a Siamese network and integrates a lightweight multi-head attention mechanism. After optimization through three steps—pre-training, transfer learning, and quantization pruning—the model size is compressed to less than 8MB, enabling it to run locally on industrial-grade low-power edge chips and achieve millisecond-level inference. The model achieves full-dimensional prediction of six core faults in the oil production tree, including valve internal leakage, flange seal failure, casing corrosion, actuator jamming, safety valve failure, and pressure anomalies. It also completes fault type localization, fault severity classification, and remaining service life (RUL) prediction. The pre-training step uses a large sample dataset of general rotating machinery faults to learn general fault feature extraction capabilities. The migration fine-tuning step uses a small amount of labeled fault data from the oilfield to fine-tune the parameters of the pre-trained model, improving the model's adaptability to oilfield faults. The quantization pruning step converts the model parameters from 32-bit floating-point to 8-bit integer through quantization, and removes redundant neurons and connections in the model through pruning, reducing the model size and computational consumption.

7. The edge computing-based oil well tree fault prediction and health management system according to claim 6, characterized in that: The transfer learning algorithm is used for the fine-tuning process of the lightweight fault prediction model, and its loss function is calculated as follows: ； Where L is the total loss function, α is the weight coefficient with a value ranging from 0.6 to 0.8, used to balance classification loss and regression loss, Lcls is the fault type classification loss function, which adopts the cross-entropy loss function to optimize the fault classification accuracy of the model, and Lreg is the fault severity and RUL regression loss function, which adopts the mean squared error loss function to optimize the regression accuracy of the model and ensure that the fault identification accuracy is ≥96% in small sample scenarios.

8. The edge computing-based oil well tree fault prediction and health management system according to claim 1, characterized in that: The three-dimensional health assessment module is used to establish a three-dimensional health assessment system covering structural integrity, operational performance, and sealing reliability. It includes a health index calculation unit and a remaining life prediction unit. The health index calculation unit calculates structural integrity, operational performance, and sealing reliability indices based on preprocessed sensor data, determines the weights of each index using the analytic hierarchy process (AHP), and quantifies the overall equipment health score. The remaining life prediction unit uses an improved particle filter algorithm to predict the remaining effective life of key components such as the main valve and wing valve online. The calculation formula is as follows: ； in, The remaining service life is the predicted value, where t is the current time and T is the time of equipment failure. For all sensor data up to and including the current moment, The posterior probability density function at the time of failure is updated iteratively using an improved particle filter algorithm to ensure the accuracy of RUL prediction, with an early fault warning lead time of ≥72h.

9. The edge computing-based oil well tree fault prediction and health management system according to claim 1, characterized in that: The edge-cloud collaborative closed-loop architecture adopts an edge-to-ground inference + cloud-to-global iteration mode. The edge node layer and the cloud management layer are connected through an encrypted communication link, supporting data interruption resumption. The edge node layer is responsible for real-time data processing, local fault early warning, emergency control, and local data caching. When the network is interrupted, the edge node layer operates independently to complete fault prediction and health assessment. After the network is restored, it automatically uploads the cached data and operation logs to the cloud management layer. The cloud management layer is responsible for the full life cycle ledger management of all oilfield production tree equipment, fault root cause analysis, global model optimization, and maintenance decision management. The model iteration adopts a horizontal federated learning framework. Each edge node only uploads the model gradient parameters. The cloud management layer aggregates the gradient parameters through a federated aggregation algorithm to generate a globally optimized model, and then distributes the optimized model to each edge node for updating, without interrupting the edge-side inference business throughout the process.

10. The edge computing-based oil well tree fault prediction and health management system according to claim 9, characterized in that: The cross-site federated collaboration layer adopts a hierarchical federated learning framework, including an edge node layer, a regional aggregation layer, and a global aggregation layer. It combines contrastive learning and differential privacy protection techniques to achieve cross-site knowledge sharing and data privacy protection. Contrastive learning is used to enhance the commonality of cross-site fault characteristics, and its loss function is calculated as follows: ； Where Lcontrast is the contrastive learning loss, N is the number of samples, and xi is the anchor sample. For positive samples of the same type as anchor samples, For negative samples that are of a different class than the anchor sample, For feature extraction function, The cosine similarity function is used. The temperature coefficient is used; differential privacy protection technology adds Gaussian noise to the uploaded model parameters to protect parameter privacy, ensuring that cross-site data is not leaked, while improving the model's generalization ability.

Images