Valve remote operation and maintenance and fault diagnosis system based on industrial internet

Abstract

The invention relates to the technical field of industrial valve operation and maintenance, and discloses an industrial internet-based valve remote operation and maintenance and fault diagnosis system, which comprises a health baseline construction and tiny offset sensing module, a system-level collaborative diagnosis and root cause tracing module, a self-evolution intelligent optimization and decision module and a full-flow closed-loop operation and maintenance execution module. The system firstly collects all-working-condition parameters of a valve, constructs an exclusive health fingerprint, carries out dynamic calibration, and captures early-stage tiny offset through trend analysis and statistical control; linking upstream and downstream equipment by means of a digital twinborn model, distinguishing faults and process disturbance, and tracing root causes; through a digital shadow simulation evaluation scheme, combining a human-in-the-loop mechanism iteration model and an industrial knowledge graph; and finally, closed-loop operation and maintenance are realized through a layered hardware architecture. The invention aims to solve the problems of early warning lagging, high false alarm rate and poor adaptability in the prior art, realize active pre-judgment and accurate management and control, reduce the operation and maintenance cost, and adapt to complex industrial scene requirements.

Description

Technical Field

[0001] This invention relates to the field of industrial valve operation and maintenance technology, specifically to a valve remote operation and maintenance and fault diagnosis system based on the Industrial Internet. Background Technology

[0002] As core fluid control equipment in complex industrial scenarios such as chemical, power, and long-distance pipelines, valves directly determine the stability of process flows and the economics of operation and maintenance due to their operational reliability. With the popularization of industrial internet technology, remote valve operation and maintenance and fault diagnosis are gradually replacing the traditional on-site inspection mode, becoming an important direction for intelligent management and control of industrial equipment. Existing technologies mostly adopt a passive response architecture of "sensing-alarm-diagnosis," which collects operating parameters by deploying sensors at key parts of the valve, judges parameter anomalies based on fixed thresholds, and deduces the cause of the fault by combining single-point parameter analysis, thereby achieving basic remote alarm and preliminary fault identification.

[0003] Current mainstream solutions primarily analyze single equipment operating parameters, focusing on valve body parameter monitoring and fault diagnosis, without deeply integrating the operating status of upstream and downstream equipment and process flow logic. Their model construction largely relies on historical fault cases and fixed algorithms, and the health benchmark system is mostly statically set, failing to adapt to dynamic changes throughout the equipment's entire lifecycle. The overall architecture emphasizes basic alarm and preliminary diagnostic functions, which can meet the basic operation and maintenance needs of industrial scenarios under simple operating conditions.

[0004] However, as industrial scenarios evolve towards more complex media, more frequent fluctuations in operating conditions, and longer equipment service lives, the adaptability limitations of these traditional technologies are becoming increasingly apparent. Several technical limitations urgently need to be addressed. Specifically, existing technologies rely on fixed parameter thresholds to trigger alarms, only capable of capturing obvious faults where parameters significantly deviate from normal ranges. They cannot identify minute parameter deviations in the early stages of equipment degradation, resulting in warnings lagging behind the actual degradation process. Furthermore, the lack of unified degradation quantification indicators makes it difficult to accurately assess the degree of equipment degradation and remaining useful life, leading to a lack of data support for maintenance planning and often resulting in reactive repairs. The analytical dimensions of existing technologies are limited to single-point valve parameters, severing the collaborative relationship with upstream and downstream pumps, pipelines, flow meters, and other equipment, making it impossible to distinguish whether parameter anomalies are due to equipment malfunctions. The problem stems from process disturbances, making the system susceptible to false alarms due to fluctuations in operating conditions. This increases ineffective maintenance costs and reduces maintenance personnel's trust in alarm signals. Existing diagnostic and early warning models are built on fixed algorithms and initial parameters, lacking the ability to adapt to dynamic scenarios such as equipment aging, changes in media properties, and adjustments in process load. After long-term operation, the model accuracy significantly decreases with changes in equipment status, requiring frequent manual parameter adjustments, resulting in high maintenance costs and poor adaptability. Existing technologies mostly remain at the "diagnosis-alarm" stage, failing to form a closed-loop link of "diagnosis-decision-execution-feedback," and the effectiveness of maintenance and handling cannot feed back into model optimization. Furthermore, the lack of a structured accumulation and reuse mechanism for maintenance knowledge makes it difficult to quickly transform the experience of handling new fault cases into system capabilities, relying heavily on the inheritance of manual experience. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, this invention provides a valve remote operation and maintenance and fault diagnosis system based on the Industrial Internet, in order to solve the problems existing in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a valve remote operation and maintenance and fault diagnosis system based on the Industrial Internet, characterized in that it includes a health baseline construction and micro-offset perception module, a system-level collaborative diagnosis and root cause tracing module, a self-evolving intelligent optimization and decision-making module, and a full-process closed-loop operation and maintenance execution module connected in sequence.

[0007] The health baseline construction and minute offset sensing module is used to construct a health baseline with a health fingerprint matrix as the core based on the valve's full-condition operation data collected by a multi-dimensional sensor array. The health fingerprint matrix associates the operating condition parameters with a feature vector composed of the mean, standard deviation, peak factor, and waveform factor of the operating parameter sequence. It also senses minute offsets of the operating parameters based on sliding window trend analysis and statistical process control methods.

[0008] The system-level collaborative diagnosis and root cause tracing module is used to perform multi-parameter linkage verification based on the digital twin model of the covered valve and upstream and downstream equipment after receiving the offset warning, and to locate the root cause of the fault using the fault correlation matrix constructed based on historical fault cases and the Bayesian inference algorithm.

[0009] The self-evolving intelligent optimization and decision-making module is used to evaluate early warning and maintenance schemes based on digital shadow simulation, and to make optimization decisions by combining human-in-the-loop learning mechanism and industrial knowledge graph.

[0010] The full-process closed-loop operation and maintenance execution module is used to execute operation and maintenance decisions through a layered hardware architecture and feed back the handling effect data to the self-evolving intelligent optimization and decision-making module.

[0011] The technical effects and advantages of this invention are as follows:

[0012] 1. This invention constructs a unique health fingerprint matrix by collecting full-condition data during the early stages of valve service through a health baseline construction and micro-offset sensing module. Combined with sliding window trend analysis and statistical process control methods, it accurately captures early micro-parameter offsets. At the same time, it introduces the Performance Degradation Index (PDI) to quantify the degree of degradation and predicts the remaining useful life through linear degradation equations and Wiener process models. This upgrades the operation and maintenance mode from passive emergency repair to proactive prediction, providing accurate data support for operation and maintenance planning.

[0013] 2. This invention relies on a system-level collaborative diagnosis and root cause tracing module to construct a high-fidelity digital twin model covering valves and upstream and downstream equipment. Through multi-parameter linkage verification, it distinguishes between faults and process disturbances. For multi-equipment alarm scenarios, it combines a fault correlation matrix and a Bayesian inference algorithm to accurately locate the root cause and propagation path of the fault, effectively avoiding false alarms caused by operating condition fluctuations and improving the accuracy of fault identification.

[0014] 3. This invention introduces a human-in-the-loop learning mechanism and incremental learning algorithm through a self-evolving intelligent optimization and decision-making module. It collects on-site handling data to iteratively update model parameters and industrial knowledge graphs, adapting to equipment aging, process changes, and new failure modes. Simultaneously, a dynamic calibration mechanism periodically optimizes the health fingerprint to avoid baseline drift, ensuring stable long-term model accuracy and reducing manual maintenance costs.

[0015] 4. This invention uses digital shadow simulation to perform quantitative analysis of the cost, risk and benefit of candidate maintenance solutions, and selects the optimal solution through a comprehensive evaluation function; combined with industrial knowledge graphs, it realizes the structured accumulation and reasoning reuse of operation and maintenance experience, quickly derives the root cause of the fault and the optimal maintenance measures, shortens the diagnosis and handling time, avoids blind maintenance, and balances maintenance costs and system operation risks.

[0016] 5. This invention establishes a complete "perception-diagnosis-decision-execution-feedback" chain through a closed-loop operation and maintenance execution module, feeding back operation and maintenance data to the self-evolution module to drive continuous iteration of the model and knowledge graph; the layered hardware architecture adapts to different industrial scenarios, enabling collaborative remote parameter adjustment, work order dispatch, and emergency management, reducing unplanned downtime losses, and balancing operation and maintenance reliability and economy. Attached Figure Description

[0017] Figure 1 This is a block diagram of the overall structure of the present invention;

[0018] Figure 2 This is a flowchart illustrating the health baseline construction and offset sensing process of the present invention.

[0019] Figure 3 This is a flowchart of the system-level collaborative diagnosis and root cause tracing of the present invention;

[0020] Figure 4 This is a flowchart of the self-evolving intelligent optimization and decision-making process of the present invention;

[0021] Figure 5 This is a diagram illustrating the closed-loop operation and maintenance execution and feedback mechanism of the present invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Reference Figure 1-5 This invention provides a remote valve operation and maintenance and fault diagnosis system based on the Industrial Internet, including a health baseline construction and micro-offset perception module, a system-level collaborative diagnosis and root cause tracing module, a self-evolving intelligent optimization and decision-making module, and a full-process closed-loop operation and maintenance execution module connected in sequence.

[0024] The health baseline, characterized by a valve-specific "health fingerprint," must be established within the initial service life of the valve (within 72 hours of successful installation and commissioning) or within 48 hours of successful overhaul acceptance, ensuring comprehensive coverage of all operating conditions. The primary step in establishing the health fingerprint is sensor deployment. In this embodiment, data is collected using a multi-dimensional sensor array deployed at key valve locations. Specific selection and installation methods are as follows: a three-dimensional vibration sensor is magnetically mounted 10±2cm from the valve stem end (model PCB 356A16, sampling accuracy ±0.01m / s², frequency response range 0.5Hz~10kHz); a torque sensor is integrated into the actuator output (model HBM T40B, accuracy ±0.1N·m, measurement range 0~500N·m); a pressure transmitter (model Rosemount 3051, accuracy ±0.01MPa) and a temperature sensor (model PT100, accuracy ±0.1℃) are deployed on the valve's inlet and outlet pipes and valve body, respectively. It should be noted that the above sensor model is only a preferred option for this embodiment. In actual applications, it can be replaced with other sensor models that meet the same measurement accuracy and industrial scenario adaptability.

[0025] After the sensors are deployed, the valves are controlled to operate stably step by step according to a preset operating condition sequence, and core parameters are collected synchronously. The operating condition parameters cover the medium pressure P (20%~100% of the design pressure, in 10% increments, with each pressure level stabilizing for 3 minutes), the medium temperature T (-20%~100% of the design temperature, in 5°C increments, with each temperature level stabilizing for 5 minutes), and the valve opening α (0°~90°, in 5° increments, with each opening level stabilizing for 2 minutes). The collected parameters include the three-dimensional vibration acceleration of the valve stem a, the actuator output torque M, the action response time tr (the time from the issuance of the command to the stabilization of the opening), the leakage of the sealing surface q (measured using an ultrasonic leak detector with an accuracy of 0.01mL / min), and the actuator operating current I, etc. The sampling frequency is set to 20Hz, and each operating condition is run stably for 10 minutes to obtain a complete time-series data sequence, ensuring the integrity and representativeness of the data.

[0026] After the full-condition time series data acquisition is completed, standardization preprocessing is performed to eliminate the influence of dimensions. The Z-score standardization algorithm is used, and the formula is as follows: Where x is the original data, Let σ be the mean of the dataset and σ be the standard deviation of the dataset. After preprocessing, a health fingerprint matrix is ​​constructed. Where m is the number of operating condition groups (17 pressure groups × 25 temperature groups × 19 opening groups = 7775 operating conditions), n is the number of parameter types (n=6, corresponding to the above 6 key parameters), and k is the number of sampling data points under a single operating condition (k=1200, calculated from a sampling frequency of 20Hz × 60 seconds × 10 minutes). For the parameter sequence of each operating condition, the mean is calculated one by one. Peak factor (Peak to RMS ratio, RMS value calculated using the root mean square algorithm), waveform factor (Ratio of effective value to average value) forms a four-dimensional feature vector. Finally, a dedicated health fingerprint database linking operating condition parameters and feature vectors is established and stored in a local database on the edge node. This embodiment preferably uses an SQLite database to support offline access, but other lightweight industrial databases can also be used in practice.

[0027] To maintain the long-term stability of the healthy baseline and avoid monitoring deviations caused by baseline drift, the valve automatically triggers a dynamic fingerprint calibration process every 6 months of continuous operation or when it undergoes a major adjustment in operating conditions. The criteria for determining a major adjustment in operating conditions are clearly defined as: change of media type (complete replacement with a media of different chemical properties, or a change in the same media component by ≥30% for a duration of ≥20 minutes), operating pressure fluctuation exceeding the design value by 30% for a duration of ≥10 minutes, or media temperature fluctuation exceeding the design value by 20% for a duration of ≥15 minutes. The calibration process only supplements the data collected during the changed operating condition interval to update the feature vector. A weighted average method is used to fuse the new and old data, with a weight of 0.7 for new data and 0.3 for old data. This weighting is determined based on 100 sets of calibration experiments, balancing data timeliness and baseline stability. Historical fingerprint versions are also archived quarterly to provide data for comparing degradation trends. The above calibration cycle, weight allocation, and criteria can be flexibly adjusted according to the operational and maintenance needs of the actual application scenario.

[0028] Based on a stable health baseline, a combination of trend analysis and statistical process control (SPC) is used to accurately capture minute parameter shifts. Specifically, this is applied to parameter sequences acquired in real time. The sliding window method was used to calculate the local trend slope. The window size was set to 100 sampling points (corresponding to 5 seconds of data), and the step size was 10 sampling points (corresponding to 0.5 seconds). The slope calculation formula was as follows: ,in This represents the average time within the window. This represents the mean of the parameters within the window. When the slope... Five consecutive windows maintain the same sign, and the absolute value exceeds the trend slope threshold corresponding to the healthy fingerprint condition. When the standard deviation of the trend slope of this parameter in the healthy fingerprint is taken as twice the value obtained by statistical analysis of 100 sets of historical data under the same working conditions, it is determined that the parameter has a slight decay shift trend.

[0029] To further improve the accuracy of offset warnings, a single-value-movement range (X-MR) control chart is simultaneously constructed for core parameters such as actuator torque and valve stem vibration amplitude. The single-value control limits are set as follows: The moving range control limit is set to This corresponds to the control chart coefficients for a sample size of n=2. When seven consecutive parameter points fall within the same side of the control limits, or when two out of three consecutive points fall within the 2σ~3σ range, an offset warning is triggered immediately without waiting for the parameter to exceed the fixed threshold.

[0030] To achieve quantitative early warning of degradation, a performance degradation index (PDI) is calculated and the remaining useful life (RUL) is predicted by combining multi-parameter offsets and a mechanistic model. First, valve stem friction is screened. Wear of sealing surface Actuator efficiency For key parameters exhibiting offsets, the weights of each parameter are determined using the Analytic Hierarchy Process (AHP). ,satisfy A three-layer judgment matrix was constructed and passed through a consistency test (test criterion CR < 0.1), and the final weight allocation was determined. =0.4、 =0.35、 =0.25, this weight allocation can be flexibly adjusted according to valve type and maintenance focus.

[0031] Based on the Archard wear model, real-time parameter offsets are converted into wear offset rates. ( (Wear amount under the corresponding working conditions for healthy fingerprints), the model formula is: For PTFE sealing surfaces, the wear coefficient k is taken as follows under normal temperature and pressure (25℃, 0.1MPa). m² / N, and dynamically adjusted to adapt to operating conditions: for every 50°C increase in medium temperature, the k value increases by 1.2 times; for every 1 MPa increase in pressure, the k value increases by 1.05 times. Normal pressure It is calculated from the actuator torque and valve stem radius, using the following formula: Where r is the actual radius of the valve stem (taken from the valve design drawings); sliding distance N represents the cumulative number of switching operations (counted by the PLC system), and L represents the total valve stem stroke (value determined according to valve model parameters). Material hardness H adapts to temperature; PTFE has a hardness of 20 HV at room temperature, decreasing to 15 HV above 100℃. If other sealing surface materials are used, the wear coefficient k and hardness H values ​​can be adjusted accordingly, such as for stainless steel sealing surfaces. m² / N, H=200HV.

[0032] The degree of decline is quantified using the following formula: ,in The real-time offset of the i-th parameter. The maximum allowable deviation for this parameter is determined according to GB / T 12224-2019 "General Requirements for Steel Valves". Maximum deviation 15%, The maximum deviation is 20%, Maximum deviation 10%), These are the parameter values ​​corresponding to healthy fingerprints. The PDI value ranges from 0% to 100%, with 0%–30% representing mild degradation, 30%–60% representing moderate degradation, and 60%–100% representing severe degradation. Based on PDI time-series data, a linear degradation equation is used. Predicting RUL Let k be the current recession index, and k be the recession rate (obtained by least squares fitting of the PDI data over the past 30 days, with a goodness of fit R² ≥ 0.85). For the nonlinear recession scenario, a Wiener process model is introduced to supplement the stochastic fluctuation term; the model expression is as follows: Where μ is the drift coefficient, σ is the diffusion coefficient, and B(t) is the standard Brownian motion, the model parameters are solved using the maximum likelihood estimation method. Pilot testing has verified that this prediction method can effectively improve the accuracy of RUL prediction and provide a reliable basis for operation and maintenance planning.

[0033] The system-level collaborative diagnosis and root cause tracing module is used to receive the offset warning signal and performance degradation index (PDI) output by the health baseline module, synchronously retrieve the real-time operating parameters of upstream and downstream equipment based on the full-process digital twin model, distinguish faults from process disturbances through multi-parameter linkage verification, and for scenarios where multiple devices alarm simultaneously, use the fault correlation matrix and Bayesian inference algorithm to locate the root cause of the fault and the propagation path, and output accurate fault identification results and tracing information.

[0034] This embodiment preferably uses Simcenter 3D software to build a full-process digital twin model, covering core equipment such as valves, pumps, compressors, pipelines, and flow meters. The model scaling ratio is 1:1, the mesh generation accuracy is 5mm, and the simulation step size is 0.1 seconds. Other industrial twin software with equivalent high-fidelity simulation capabilities can also be used. This model integrates geometric, physical, and control models to achieve high-fidelity replication of the process flow: the geometric model is reconstructed based on equipment CAD drawings (STEP 214 format) with millimeter-level accuracy, completely restoring the equipment installation location, connection relationships, and structural features, including key details such as valve sealing surface roughness and valve stem fit clearance; the physical model uses mechanistic equations to characterize the equipment's operating characteristics, covering pump head equations. ( For the rated head, (Flow loss coefficient, obtained by fitting pump nameplate parameters) and pipeline resistance equation. Where λ is the friction coefficient (calculated using the Colebrook formula), ε is the pipe inner wall roughness, D is the pipe inner diameter, Re is the Reynolds number, ρ is the medium density, and v is the medium flow velocity, accurately mapping the intrinsic relationship between the parameters; the control model replicates the field PLC control logic. In this embodiment, Siemens S7-1500 PLC is preferred, and the programming language is Structured Control Language. It can be replaced by other mainstream industrial PLCs and corresponding programming languages. The control model covers PID regulation algorithms and interlocking protection rules, such as the triggering conditions and execution actions of overpressure interlocking and overtemperature interlocking.

[0035] The model synchronizes the operating parameters of the field equipment at a frequency of 1Hz. It integrates measured data with model predictions using an Extended Kalman Filter (EKF) algorithm to form a dynamic parameter correction link, focusing on correcting key parameters such as the drag coefficient λ and wear coefficient k. The correction process is as follows: the deviation between the model's predicted value and the measured value is calculated every 10 seconds. When the deviation exceeds 5%, correction is initiated, and the parameters are iteratively updated using the EKF algorithm, with ≤20 iterations until the deviation is ≤5%. If the deviation is ≤5%, the current parameters are maintained to avoid over-correction leading to model oscillation. It should be noted that if the parameter deviation is caused by fluctuations in operating conditions (such as sudden changes in medium flow), correction is not initiated until the operating conditions stabilize (parameter fluctuation ≤2% / min) before deviation determination, ensuring consistency between the model and the physical system. The prediction error is controlled within ±5%, and this accuracy index is derived from the optimized parameters and algorithm of this embodiment; adjustments to parameters or algorithms can be made accordingly.

[0036] After the preceding module triggers an offset warning, this module immediately initiates a context-aware verification process, without immediately determining the fault type. Simultaneously, it retrieves real-time operating data from upstream and downstream equipment via the OPC UA protocol (other common industrial communication protocols can also be used). The data transmission delay is ≤100ms. The parameters collected include the upstream pump outlet pressure. ,flow Downstream flow meter reading Pipeline pressure and process controller output signals A multi-parameter linkage verification logic is constructed based on a digital twin model. Taking a suspected valve jamming fault as an example, a clear verification threshold is set: the rate of change of the upstream pump outlet pressure. If the downstream flow rate decreases by more than 15% of the health fingerprint threshold, and the pump motor current fluctuates by more than 5%, it is determined to be a parameter anomaly caused by pump operating condition fluctuations. This is considered a false fault, and only the operating condition fluctuation is recorded (stored in the cloud database for one year), without triggering a fault alarm. If the parameters of upstream and downstream equipment are stable within the health fingerprint ±1σ range, and the valve torque continues to increase accompanied by abnormal amplitude of valve stem vibration frequency in the 200~500Hz range (exceeding the health fingerprint threshold by 20%), combined with the PDI level (moderate or above degradation), it can be determined to be a valve jamming fault, triggering the corresponding level of warning: minor faults are only notified locally, moderate faults are pushed to the industrial-grade mobile APP, and severe faults trigger on-site audible and visual alarms. This warning classification method can be flexibly adjusted according to the enterprise's operation and maintenance process.

[0037] When normal changes occur in the process conditions (such as production load adjustments, with a load change rate ≤10% / min), the model automatically updates the condition matching interval of the health fingerprint, using a linear interpolation algorithm to supplement the feature vector of the newly added condition, avoiding misjudgments caused by changes in conditions. For complex scenarios where multiple devices simultaneously alarm, fault root cause tracing is achieved through a fault correlation matrix and Bayesian inference. Specifically, the analysis is based on the process flow logic and over 1000 historical fault cases. Cases are selected from fault data of similar devices under the same conditions over the past 5 years, eliminating faults caused by human error or natural disasters, retaining only faults caused by the equipment itself and process disturbances, covering common fault modes of valves, pumps, pipelines, and other equipment. An equipment fault correlation matrix is ​​established. 'p' represents the number of system devices; in this embodiment, p=8, including valves, feed pumps, circulating pumps, compressors, inlet and outlet pipes, flow meters, pressure gauges, and temperature sensors. (Matrix elements) This represents the probability that a malfunction in device i will trigger an alarm in device j. The value ranges from 0 to 1, and the initial value is generated through historical case statistics, such as the probability that a valve jamming will cause a downstream flow meter to alarm. This value is determined based on the statistical results of 672 out of 820 valve jamming cases triggering flow meter alarms. It is updated quarterly using a weighted average method based on newly added cases, with a weight of 0.6 for each new case, balancing the timeliness of new cases with the stability of historical data. The matrix dimensions and probability values ​​mentioned above can be adjusted according to the system equipment composition.

[0038] Define device alarm events (1 indicates an alarm, 0 indicates normal) and fault events (1 represents a malfunction, 0 represents normal operation) is a random variable, with a set prior probability. (Based on equipment failure rate statistics, such as valve failure rate) Based on Bayes' theorem The system calculates the posterior probability of each device as the root cause of the fault, selecting the device with the highest probability as the initial fault point. The posterior probability threshold is ≥0.7; if it is below the threshold, the range of associated devices is expanded and the calculation is repeated. Simultaneously, a fault propagation path diagram is drawn based on a digital twin model, marking the fault transmission probability and time delay at each stage. For example, the transmission time from pump cavitation to pipeline pressure fluctuation is 2-3 seconds, with a probability of 0.91. This provides clear fault handling guidance for maintenance personnel, avoiding blind repairs. Pilot applications have shown that this root cause tracing mechanism can significantly improve the accuracy and efficiency of fault location.

[0039] When collaborative diagnostics confirms that the valve is in good condition and that abnormal parameters are caused by process fluctuations, the valve control parameters are adaptively optimized based on the system's digital twin model to improve system regulation performance. For the valve PID controller, the optimization objective is established as "shortest system settling time and minimum overshoot," and an objective function is defined accordingly. ,in This refers to the system settling time (time when steady-state error is ≤2%). The overshoot is a percentage. The PID parameters (proportional coefficient) are iteratively optimized using a digital twin model through simulation. Integral Time Differential time The iteration step size is dynamically adjusted based on process sensitivity: for high-frequency fluctuation conditions (pressure fluctuation frequency ≥ 0.5Hz), a small step size is used. Step size 0.01, Step length 0.1s, The step size is 0.01s, and a large step size is used in stable operating conditions. Step size 0.05, Step length 0.5s, With a step size of 0.05s, the iteration terminates when the objective function J converges (the difference between two adjacent iterations of J is ≤0.01), which can significantly shorten the adjustment time and improve the operating efficiency of the entire process.

[0040] The self-evolving intelligent optimization and decision-making module, based on the fault determination results and root cause analysis conclusions output by collaborative diagnosis, uses digital shadow simulation to deduce the PDI change trend to assess the early warning level, performs cost, risk and benefit quantitative analysis on candidate maintenance solutions, and collects on-site handling data in conjunction with the human-in-the-loop learning mechanism to iteratively update model parameters and industrial knowledge graphs, forming an adaptive decision-making capability that adapts to equipment aging and process changes.

[0041] Based on a high-fidelity digital twin model, a digital shadow is constructed. This embodiment uses Python to build the simulation script and calls the Simcenter 3D API interface to achieve data interaction with the twin model. In practice, other programming languages ​​and corresponding API interfaces can also be used. The core is used to quickly simulate and evaluate the early warning conclusions and candidate maintenance solutions output by the preceding modules. The simulation time is ≤30 seconds, providing data support for operation and maintenance decisions. In the early warning rationality assessment stage, the digital shadow simulates the PDI change trajectory in the next 24 hours based on the current operating parameters and health fingerprint, and outputs the simulation results with 1-hour time nodes. If the daily average increase of PDI is <1% and does not affect the stability of the process flow (system pressure and flow fluctuation ≤3%), it is judged as a low-risk warning. It is recommended to adopt the observation operation mode, review the parameter changes every 2 hours, and automatically generate a review report. If the daily average increase of PDI is >3%, or the simulation predicts that PDI will exceed 30% within 12 hours, it is judged as a high-risk warning, and it is recommended to immediately start the disposal process.

[0042] During the maintenance decision-making and evaluation phase, for three candidate solutions—immediate downtime maintenance, planned downtime maintenance (scheduled during off-peak hours within 24 hours), and online parameter adjustment—implementation costs (including labor costs, downtime losses, and consumable costs), downtime, system operational risks (probability of fault escalation), and overall benefits are simulated and calculated using a comprehensive evaluation function. The optimal solution is selected, with the one having the smallest F-value as the final execution plan. In the function, α=0.4, β=0.3, and γ=0.3, set according to the enterprise's operation and maintenance strategy. C represents the implementation cost (in ten thousand yuan), R represents the operational risk (value 0~1), and B represents the comprehensive benefit (in ten thousand yuan). These weighting coefficients can be adjusted according to the enterprise's cost priority and risk tolerance. After maintenance is completed, the actual operating results are compared with the simulation evaluation data to revise the weighting coefficients in the evaluation function and the cost and risk models. The revision cycle is after each maintenance, based on the deviation rate between the actual and simulated values, continuously improving the accuracy of decision-making and evaluation.

[0043] A human-in-the-loop learning mechanism is introduced to transform the on-site handling results of maintenance personnel into core feedback for model optimization. Maintenance personnel input relevant information through an industrial-grade mobile app, which supports Android 10.0 and above, and iOS 14.0 and above. An encrypted transmission protocol ensures data security. Input includes the authenticity of the fault (real / false), the actual fault type (classified according to GB / T 7221-2021 "Valve Terminology"), specific repair measures (such as sealing surface replacement, valve stem lubrication, actuator calibration, etc.), and the handling effect (qualified / unqualified, with the qualified standard being a PDI regression within 30% and stable operation for 4 hours). The system automatically adds tags to the case, forming a standardized labeled dataset in JSON format, containing timestamps, equipment numbers, fault information, handling information, and effect data. The app system configuration can be adjusted and adapted to the enterprise's existing maintenance terminals.

[0044] An incremental learning algorithm based on stochastic gradient descent is adopted, iteratively updating the parameters of each module using newly labeled data to ensure that the model adapts to newly emerging fault modes and equipment aging characteristics. The algorithm hyperparameters are optimized using a grid search method, with a learning rate of 0.001, 1000 iterations, and a batch size of 32, balancing training efficiency and convergence accuracy. When correcting the feature vectors corresponding to the working conditions of the health fingerprint, momentum gradient descent is used with a momentum coefficient of 0.9 to accelerate convergence and avoid local optima, while simultaneously adjusting the elements of the fault correlation matrix. The probability value is used to update the degradation rate k-fit coefficient of the RUL prediction model. For every 100 labeled cases (statistics for single valve maintenance systems; multiple parallel systems are counted independently), model performance is evaluated using 5-fold cross-validation. Validation metrics include accuracy, recall, and RUL prediction precision. If the improvement in any of these three metrics is less than 5%, model structure optimization is automatically triggered, including adjusting parameter weight allocation, adding feature dimensions (such as adding a new medium viscosity parameter), and optimizing algorithm hyperparameters. This ensures the effectiveness of the learning mechanism, allowing the system to continuously accumulate maintenance experience and gradually improve the intelligence level of diagnostic decision-making.

[0045] Building an industrial knowledge graph enables structured storage and reasoning reuse of operational experience, improving diagnostic and decision-making efficiency. This embodiment preferably uses the Neo4j graph database, but other industrial-grade graph databases can also be used. Graph node types include equipment, components, fault types, media, maintenance tools, and maintenance measures. Relationship types include "fault-component association," "fault-media association," and "maintenance measure-fault association," etc. Attributes cover fault characteristic parameters, component materials, media physicochemical properties, maintenance tool models, and maintenance steps. An automated knowledge extraction and manual review mechanism is adopted to extract knowledge from historical fault cases, equipment operation and maintenance manuals, and industry standards and specifications. The BERT model is fine-tuned based on an industrial fault text pre-training model. The pre-training dataset contains 100,000 industrial equipment fault description texts, and the fine-tuning dataset contains 20,000 labeled valve operation and maintenance texts. The model extracts entities and relationships from the text with an accuracy of ≥92%. The extraction results are reviewed and confirmed by operation and maintenance experts and then entered into the knowledge graph with a review pass rate of ≥95%. The review criteria are that the entities and relationships conform to industry terminology standards and are consistent with actual operation and maintenance scenarios, thus constructing an initial knowledge graph. After a new fault case is handled, it is automatically added to the knowledge graph after review and confirmation by operation and maintenance personnel, realizing dynamic updates of knowledge.

[0046] When a new fault occurs, the system first uses a cosine similarity algorithm to locate similar historical cases, with a similarity threshold of ≥85%. Then, relying on a knowledge graph, it uses a combination of forward and backward reasoning to deduce the root cause of the fault and the optimal maintenance measures. For example, when a valve stem seal leak is detected, the system identifies media crystallization as a common cause through knowledge graph reasoning, automatically prompting maintenance personnel to check the media purity (pass standard ≥99.5%) and operating temperature (controlled above the media crystallization temperature by 5°C), significantly shortening the fault diagnosis time.

[0047] The full-process closed-loop operation and maintenance execution module is used to receive the optimal decision scheme output by the self-evolution module. Through a layered hardware architecture, it realizes remote parameter adjustment, on-site work order dispatch and emergency control. It simultaneously collects full-process handling data and effect verification results, and feeds them back to the self-evolution module to complete the iterative update of the model and industrial knowledge graph, thus constructing a closed-loop operation and maintenance link of "decision-execution-feedback".

[0048] The hardware deployment adopts a layered design to adapt to the application needs of different industrial scenarios: The front-end sensing layer deploys multi-dimensional sensors at key locations such as valve stems, valve bodies, and actuators. The sensor selection and installation methods are consistent with the health baseline module. The sensor output signals are 4~20mA analog signals, which are converted into digital signals by a signal converter (accuracy ±0.001mA) before being transmitted to the edge nodes. The edge nodes use Intel Core i7-12700E processors, are equipped with 16GB of RAM and 512GB of SSD storage, support wide voltage input (12~24V DC), have an anti-interference rating of IP67, and can adapt to high temperature (-40℃~85℃), high humidity (relative humidity ≤95%, non-condensing), and highly corrosive conditions. They have built-in firewalls and encryption modules to achieve local data preprocessing, cache storage, and encrypted transmission. The above edge node hardware configuration can be adjusted according to data processing requirements, including processor model, memory, and storage capacity.

[0049] The transmission layer employs differentiated solutions to adapt to different scenarios: Industrial park scenarios utilize a dual-link transmission of 5G private network + fiber optic, with 5G private network bandwidth ≥100Mbps, transmission latency ≤20ms, and fiber optic transmission rate ≥1Gbps. The dual links serve as backups for each other, with a switching time ≤1 second, ensuring data transmission rate and stability. Remote scenarios employ LoRa + NB-IoT dual-mode transmission, with LoRa communication distance ≥3km and NB-IoT supporting full network compatibility. It features a low-power design (standby current ≤10mA) suitable for long-term outdoor operation. Fault alarm signals are set to the highest priority, using dedicated bandwidth transmission and equipped with a breakpoint resume mechanism (data loss rate ≤0.1%). Data transmission is encrypted using TLS / SSL 1.3 protocol to prevent data leakage and loss. The above transmission solutions can be replaced with other industrial communication methods depending on the scenario's communication conditions.

[0050] A server cluster is deployed in the cloud layer. This embodiment uses Huawei FusionServer Pro, which can be replaced with other industrial-grade servers. The configuration includes a CPU with at least 24 cores (Intel Xeon Gold 6348), at least 128GB of memory, and at least 10TB of storage (RAID 5 array). It is equipped with a time-series database (InfluxDB 2.7), a relational database (PostgreSQL 14), and a graph database (Neo4j 5.12) to store time-series runtime data (retained for 1 year), structured operation and maintenance data (permanently retained), and knowledge graph data (permanently retained), respectively. A GPU accelerator card (NVIDIA A100) is configured to support digital twin simulation and model training. The server cluster is deployed in the enterprise's private cloud, employing a redundant design with an availability of at least 99.9%. The terminal layer equips maintenance personnel with an industrial-grade mobile APP (consistent with the previous one), which supports offline data entry and synchronization; the central control room deploys a visualization platform. In this embodiment, WinCC configuration software is used, but it can be replaced with other industrial configuration software. It displays the digital twin model, fault propagation path, maintenance work order progress and equipment operating parameters in real time (refresh frequency 1Hz), and supports the issuance of remote control commands (such as valve opening and closing, parameter adjustment) to achieve global management and control.

[0051] The software execution process strictly follows a closed loop of "initialization-monitoring-diagnosis-decision-execution-iteration," with seamless integration and strict time control at each stage: During system initialization, health fingerprint construction, digital twin model building, industrial knowledge graph initialization, and parameter threshold configuration are completed within 2 hours. After initialization, self-verification is performed, including sensor communication, data transmission, and model accuracy, with a pass rate ≥99%. If verification fails, an audible and visual alarm is triggered, and the fault location is displayed (e.g., sensor communication failure, abnormal model parameters), meeting the requirements for rapid on-site deployment. In the real-time monitoring stage, data collected by front-end sensors is preprocessed by edge nodes (using a 5th-order Butterworth low-pass filter with a cutoff frequency of 10Hz to remove high-frequency interference; outlier removal uses the 3σ criterion, exceeding...). After the data within the range is marked as abnormal and replaced with valid data from the previous moment, it is transmitted to the cloud at a frequency of 20Hz. The health baseline module continuously senses minor parameter shifts, and abnormal data is marked in real time. In the collaborative diagnosis phase, the digital twin model is linked to complete context verification and root cause localization. The diagnosis time is controlled within 10 seconds, and the diagnosis results are pushed to the cloud and industrial-grade mobile APP in real time. In the decision optimization phase, the optimal solution is generated through digital shadow simulation and industrial knowledge graph reasoning. The evaluation time of a single solution does not exceed 30 seconds, and the solution automatically generates an operation and maintenance work order (including fault information, handling steps, required tools, and responsible person). In the operation and maintenance execution phase, mild degradation faults are handled through remote parameter adjustment. The instruction issuance delay is ≤500ms, and stability is continuously monitored for 10 minutes after adjustment. Moderate degradation faults are handled through remote parameter adjustment. Preventive maintenance work orders are dispatched to nearby maintenance personnel via industrial-grade mobile app push notifications and SMS reminders, with a 24-hour validity period. In cases of severe degradation, valves are immediately shut off remotely (with a response time of ≤1 second for the shut-off command when remote control is available) and emergency repairs are initiated. During on-site handling, maintenance personnel upload operation data and videos in real time (video resolution 1080P, frame rate 25fps, H.265 encoding compression storage) to achieve synchronous monitoring in the cloud. In the effect verification phase, parameter changes are continuously monitored for 4 hours. Once it is confirmed that the PDI has returned to the mild degradation range of less than 30% without rebound (fluctuation ≤2%), the entire chain of data (collected parameters, diagnostic process, treatment plan, and effect verification) is fed back to the self-evolution module to complete the iterative update of the model and industrial knowledge graph, forming a complete closed loop.

[0052] The foregoing describes exemplary embodiments of this application. It should be understood that the above exemplary embodiments are not restrictive but illustrative, and the scope of protection of this application is not limited thereto. It should be understood that those skilled in the art can make modifications and variations to the embodiments of this application without departing from the spirit and scope of this application, and such modifications and variations should be within the scope of protection of this application.

Claims

1. A remote valve operation and maintenance and fault diagnosis system based on the Industrial Internet, characterized in that, It includes a health baseline construction and micro-deviation perception module, a system-level collaborative diagnosis and root cause tracing module, a self-evolving intelligent optimization and decision-making module, and a full-process closed-loop operation and maintenance execution module, which are connected in sequence. The health baseline construction and micro-offset sensing module is used to construct a health baseline with a health fingerprint matrix as the core based on the valve full-condition operation data collected by the multi-dimensional sensor array. The health fingerprint matrix is ​​associated with the operating condition parameters and the feature vector composed of the mean, standard deviation, peak factor and waveform factor of the operating parameter sequence. and The method of sensing minute deviations in operating parameters is based on sliding window trend analysis and statistical process control. The system-level collaborative diagnosis and root cause tracing module is used to perform multi-parameter linkage verification based on the digital twin model of the covered valve and upstream and downstream equipment after receiving the offset warning, and to locate the root cause of the fault using the fault correlation matrix constructed based on historical fault cases and the Bayesian inference algorithm. The self-evolving intelligent optimization and decision-making module is used to evaluate early warning and maintenance schemes based on digital shadow simulation, and to make optimization decisions by combining human-in-the-loop learning mechanism and industrial knowledge graph. The full-process closed-loop operation and maintenance execution module is used to execute operation and maintenance decisions through a layered hardware architecture and feed back the handling effect data to the self-evolving intelligent optimization and decision-making module.

2. The valve remote operation and maintenance and fault diagnosis system based on the Industrial Internet as described in claim 1, characterized in that, The health baseline construction and micro-offset sensing module is also used to calculate the performance degradation index and, based on the time-series data of the performance degradation index, predict the remaining useful life of the valve through a linear degradation equation or Wiener process model.

3. The valve remote operation and maintenance and fault diagnosis system based on the Industrial Internet as described in claim 1, characterized in that, The digital twin model integrates a geometric model reconstructed from equipment CAD drawings, a physical model constructed from equipment operating mechanism equations, and a control model that replicates the on-site PLC control logic. It also dynamically fuses measured data with model predictions using an extended Kalman filter algorithm to correct key parameters.

4. The valve remote operation and maintenance and fault diagnosis system based on the Industrial Internet as described in claim 1, characterized in that, The method of locating the root cause of a fault using a fault correlation matrix and a Bayesian inference algorithm includes: calculating the posterior probability of each device as the root cause of a fault based on the fault correlation probability between devices in the fault correlation matrix, combined with the device alarm status and prior fault rate, and selecting devices whose posterior probability exceeds a preset threshold as the initial fault point.

5. The valve remote operation and maintenance and fault diagnosis system based on the Industrial Internet as described in claim 1, characterized in that, The digital shadow simulation-based evaluation of maintenance solutions includes: calculating the implementation cost, system operation risk, and overall benefits of different candidate solutions through simulation, and selecting the optimal solution through a comprehensive evaluation function.

6. The valve remote operation and maintenance and fault diagnosis system based on the Industrial Internet according to claim 1, characterized in that, The industrial knowledge graph stores devices, components, fault types, maintenance measures entities, and their interrelationships. When a new fault occurs, the system infers the root cause of the fault and maintenance measures by matching the similarity of historical cases and combining the industrial knowledge graph.

7. The valve remote operation and maintenance and fault diagnosis system based on the Industrial Internet according to claim 1, characterized in that, The layered hardware architecture includes: The front-end sensing layer is equipped with multi-dimensional sensors that collect valve operating parameters; Edge nodes are used for local preprocessing and caching of sensor data; The transport layer employs at least one of the following communication solutions: 5G private network, optical fiber, LoRa, or NB-IoT. The cloud layer contains a server cluster deployed for data storage, simulation, and model training. The terminal layer includes an industrial-grade mobile app for operation and maintenance personnel and a central control room visualization platform.

8. The valve remote operation and maintenance and fault diagnosis system based on the Industrial Internet according to claim 1, characterized in that, The closed-loop operation and maintenance execution module adaptively performs remote parameter adjustments, generates and dispatches preventive operation and maintenance work orders, or initiates emergency repair processes based on the level of fault or degradation.

9. The valve remote operation and maintenance and fault diagnosis system based on the Industrial Internet according to claim 1, characterized in that, The human-in-the-loop learning mechanism is used to collect on-site handling results entered by maintenance personnel through terminals, and based on this, uses an incremental learning algorithm to iteratively update the model parameters in the health baseline construction and micro-offset perception module, the system-level collaborative diagnosis and root cause tracing module, and the self-evolving intelligent optimization and decision-making module.

Images