Pressure switch malfunction detection method and system

Abstract

The present application provides a method and system for detecting malfunction of a pressure switch, which relates to the field of switch detection technology. The method includes: measuring the capacitance characteristics between a sensing electrode and a metal component of a pressure switch, and generating corrosion status information based on its change relative to a reference state; identifying and characterizing cracks based on the response of the metal component of the pressure switch to an external electromagnetic detection field and an acoustic detection wave, and generating crack information including position, size, depth and historical expansion trend; utilizing a coupled model of corrosion and crack expansion to generate predictive crack expansion information based on corrosion status information and crack information; obtaining real-time dynamic stress information through a stress sensor; finally, determining whether the pressure switch has a risk of malfunction or has already malfunctioned, and generating a detection result; the present application can integrate multi-source sensing and model prediction to achieve tracking and prediction of internal damage to the pressure switch, thereby improving the early warning capability, accuracy and adaptability of detection to working conditions.

Description

Technical Field

[0001] The present application relates to the field of switch detection technology, and in particular to a method and system for detecting malfunction of a pressure switch. Background Art

[0002] As a critical automatic control and safety protection component, pressure switches are widely used in industrial process control, hydraulic and pneumatic systems, HVAC, transportation, and many other fields. Their core function is to automatically open or close a circuit when fluid pressure reaches a preset threshold, thereby controlling, alarming, or providing safety interlocks for associated equipment. Common pressure switch types include mechanical and electronic. Regardless of the type, reliable operation of the pressure switch is crucial to the proper functioning of the system in which it operates. However, in practical applications, pressure switches may experience "malfunctions" for various reasons—failure to switch at the correct pressure point and in the correct manner as designed. These malfunctions can manifest in various forms, including set point drift, inaction, non-reset, false triggering, switch chatter, sticking, and leakage. These malfunctions can impact system efficiency and control accuracy at the very least, or even lead to equipment damage, production interruptions, or even safety incidents.

[0003] Existing technologies primarily focus on the external functional performance of pressure switches, but lack in-depth insight into and tracking of the evolving internal physical states that lead to malfunctions. In particular, critical metal components in pressure switches are inevitably subject to corrosion from the working medium and fatigue damage from cyclic pressure during service. This can lead to material degradation, the formation of microcracks that gradually expand, and ultimately component failure and switch malfunctions. Existing technologies are generally unable to effectively monitor these progressive internal damage processes or predict their progression under future operating conditions, making it difficult to achieve early warning and accurate assessment of malfunction risks.

[0004] Therefore, there is an urgent need to propose a new pressure switch malfunction detection method that can monitor the physical damage status of key components in real time, predict the damage development based on the actual working load and operation mode, and make accurate and timely judgments on the malfunction risk based on the comprehensive status information. Summary of the Invention

[0005] In response to the deficiencies of the prior art, the present application provides a pressure switch malfunction detection method and system.

[0006] In a first aspect, the present application provides a method for detecting malfunction of a pressure switch, comprising:

[0007] By using a sensing electrode provided on a surface of a metal component of the pressure switch, the capacitance characteristic between the sensing electrode and the metal component of the pressure switch is measured; and based on a change in the capacitance characteristic relative to a reference state, corrosion status information is generated;

[0008] Identifying and characterizing cracks in the metal component of the pressure switch based on responses of the metal component to an external electromagnetic detection field and an acoustic detection wave, and generating crack information; the crack information includes: location information, size information, depth information, and historical growth trend;

[0009] Using a coupled model of corrosion and crack growth, predictive crack growth information is generated based on the corrosion state information and the crack information; dynamic stress information of the metal component of the pressure switch during operation is obtained through a stress sensor;

[0010] Based on the dynamic stress information, the corrosion state information, the predictive crack growth information and the action mode required by the currently requested control logic, it is determined whether the pressure switch has a risk of malfunction or has malfunctioned, and a detection result is generated.

[0011] As an optional implementation manner, generating crack information includes:

[0012] At a first time point, applying a time-varying electromagnetic detection field to a predetermined monitoring area of ​​the pressure switch metal component to detect a corresponding electromagnetic response signal, and transmitting an acoustic energy detection wave to the pressure switch metal component to detect a corresponding acoustic echo signal;

[0013] Calculating a first set of geometric feature information of the crack based on the electromagnetic response signal and the acoustic echo signal acquired at the first time point; the first set of geometric feature information includes: crack position information, size information, and depth information at the first time point;

[0014] At a second time point, the process of acquiring the corresponding electromagnetic response signal and the acoustic echo signal and calculating the geometric feature information based thereon is again performed to obtain a second set of geometric feature information of the crack; the second set of geometric feature information comprising: crack location information, size information, and depth information at the second time point;

[0015] Comparing the second set of geometric feature information with the first set of geometric feature information, and calculating a change amount and a change direction of the position, size, and depth of the crack during a time interval between the first time point and the second time point;

[0016] Determining a historical crack growth trend based on the change amount and change direction;

[0017] The second set of geometric feature information is combined with the historical expansion trend to generate the crack information.

[0018] As an optional implementation manner, generating corrosion status information includes:

[0019] accessing a preset stress-capacitance response relationship, the relationship representing a capacitance characteristic of a metal component of the pressure switch in a known corrosion state versus an applied stress;

[0020] Calculating an expected corrosion-free capacitance characteristic reference value corresponding to the real-time dynamic stress information using the stress-capacitance response relationship and the dynamic stress information acquired in real time;

[0021] comparing the capacitance characteristic measured in real time with a calculated baseline value of the expected capacitance characteristic without corrosion to determine an amount of capacitance characteristic deviation caused by corrosion;

[0022] Based on the capacitance characteristic deviation, a preset correlation model or calibration data for correlating the capacitance deviation to the corrosion state is applied to generate the corrosion state information.

[0023] As an optional implementation manner, measuring the capacitance characteristic between the sensing electrode and the metal component of the pressure switch includes:

[0024] Applying an AC excitation signal with a predetermined frequency between the sensing electrode and the metal component of the pressure switch;

[0025] measuring an electrical response parameter generated between the sensing electrode and the metal component of the pressure switch in response to the AC excitation signal, wherein the electrical response parameter is selected from at least one of an AC response current, an AC response voltage, or a complex impedance;

[0026] The capacitance characteristic is calculated based on the measured electrical response parameter and the predetermined frequency.

[0027] As an optional implementation, generating predictive crack growth information using a coupled model of corrosion and crack growth includes:

[0028] mapping the corrosion state information and the crack information into damage states of multiple spatial regions on the metal component of the pressure switch;

[0029] In the coupling model, establishing and utilizing a mutual influence relationship between the damage states of the plurality of spatial regions;

[0030] Calculating predictive crack growth information for each spatial region based on the mutual influence relationship of the damage states and the current damage state of each region;

[0031] The mutual influence relationship of the damage states at least reflects one or more of the following effects: the change of the local stress distribution of other regions due to damage in one region, and the electrochemical coupling effect between different regions.

[0032] As an optional implementation manner, establishing and utilizing the mutual influence relationship of damage states among the multiple spatial regions includes:

[0033] Abstracting the multiple spatial regions into nodes of a graph neural network, and constructing a node feature vector for each node; wherein the node feature vector includes: the corrosion state information, the crack information, the dynamic stress information, and the capacitance characteristic;

[0034] Determine the edge connection and edge feature vector between the nodes based on whether the physical proximity relationship between the nodes, the physical field gradient between the nodes, and the state parameter difference between the nodes meet preset conditions;

[0035] Calling a message passing graph neural network pre-trained based on historical prototype data, inputting the node feature vector and the edge feature vector into the message passing graph neural network, and outputting a crack propagation driving force index of the corresponding node;

[0036] The crack growth driving force index is used in the coupling model to calculate the predictive crack growth information of each spatial region.

[0037] As an optional implementation manner, establishing and utilizing the mutual influence relationship of damage states among the multiple spatial regions further includes:

[0038] monitoring the predicted crack growth information generated by the coupling model, comparing it with the first set of geometric feature information and the second set of geometric feature information acquired at different subsequent time points, and determining a deviation between actual crack geometry changes;

[0039] In response to the deviation exceeding a preset threshold, triggering an incremental learning process;

[0040] The incremental learning process uses a monitoring data set to perform online adjustment on the weights of the graph neural network; the monitoring data set is associated with a monitoring period corresponding to the deviation, and the monitoring data set includes input information obtained during the monitoring period for generating the node feature vector and the edge feature vector, as well as the corresponding actual crack geometry changes.

[0041] As an optional implementation, determining whether the pressure switch has a risk of malfunction or has malfunctioned includes:

[0042] Extracting or aggregating the final state representations of the nodes generated after processing by the message passing graph neural network to generate a feature vector representing the overall damage state of the metal component of the pressure switch;

[0043] Identify the action mode required by the control logic of the current request;

[0044] Inputting the feature vector and the representation of the identified current action mode into a pre-trained risk discrimination model;

[0045] The detection result is generated using the output result of the risk discrimination model.

[0046] As an optional implementation, the inputting into a pre-trained risk discrimination model includes:

[0047] Obtaining the capacitance characteristic deviation caused by corrosion;

[0048] The capacitance characteristic deviation caused by corrosion is used as an additional input feature and is transmitted to the risk discrimination model together with the eigenvector and the identified representation of the current action mode;

[0049] The risk discrimination model is trained to jointly utilize the eigenvector, the representation of the current action mode, and the capacitance characteristic deviation to determine the malfunction risk level or state of the pressure switch.

[0050] In a second aspect, the present application provides a pressure switch malfunction detection system, comprising:

[0051] an acquisition module for measuring capacitance characteristics between a sensing electrode disposed on a surface of a metal component of the pressure switch and the metal component of the pressure switch, and generating corrosion status information based on a change in the capacitance characteristics relative to a reference state;

[0052] a first processing module for identifying and characterizing cracks in the metal component of the pressure switch based on responses of the metal component of the pressure switch to an external electromagnetic detection field and an acoustic detection wave, and generating crack information; the crack information includes: location information, size information, depth information, and historical growth trend;

[0053] The second processing module generates predictive crack growth information based on the corrosion state information and the crack information using a coupled model of corrosion and crack growth; and obtains dynamic stress information of the metal component of the pressure switch during operation through a stress sensor;

[0054] The detection module determines whether the pressure switch has a risk of malfunction or has malfunctioned based on the dynamic stress information, the corrosion state information, the predictive crack growth information, and the action mode required by the currently requested control logic, and generates a detection result.

[0055] Compared to existing technologies, this application, by monitoring the physical damage states of key components, such as corrosion and cracks, in real time and using coupled models to predict future development trends, can identify potential risks before functional malfunctions occur. This shift from passive fault detection to active state prediction significantly enhances early warning capabilities. By integrating information from multiple sensors (capacitive characteristics, electromagnetic / acoustic response, dynamic stress), a comprehensive understanding of equipment health is obtained, and a comprehensive assessment is made based on actual operating conditions. This avoids the limitations of a single information source or simple threshold judgments, more accurately distinguishing between real risks and normal fluctuations, and significantly reducing false alarm and missed alarm rates. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A flow chart of a method for detecting malfunction of a pressure switch provided in an embodiment of the present application;

[0057] Figure 2 A flowchart of a method for generating predictive crack growth information provided in an embodiment of the present application;

[0058] Figure 3 A schematic diagram of a metal component of a pressure switch provided in an embodiment of the present application;

[0059] Figure 4 Schematic diagram of a pressure switch malfunction detection system provided in an embodiment of the present application.

[0060] Figure numerals: 1. pressure switch body; 2a. first damage area; 2b. second damage area; 2c. third damage area; 3. threaded connection part; 10. acquisition module; 20. first processing module; 30. second processing module; 40. detection module. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments.

[0062] This application aims to obtain multi-dimensional information about the corrosion status, crack status, real-time stress status and operating conditions of key metal components of pressure switches by integrating multiple sensing technologies, and combine it with physics-based or data-driven models for analysis and prediction, so as to achieve early monitoring of the risk of malfunction of pressure switches.

[0063] See also Figure 1 FIG. 1 is a flow chart of a pressure switch malfunction detection method according to an embodiment of the present application. The pressure switch malfunction detection method includes steps S101 to S104, wherein:

[0064] S101: Measuring capacitance characteristics between a sensing electrode disposed on a surface of a metal component of a pressure switch and the metal component of the pressure switch by using the sensing electrode; and generating corrosion status information based on a change in the capacitance characteristics relative to a reference state;

[0065] S102: Identifying and characterizing cracks in the metal component of the pressure switch based on responses of the metal component of the pressure switch to an external electromagnetic detection field and an acoustic detection wave, and generating crack information; the crack information includes: location information, size information, depth information, and historical expansion trend;

[0066] S103: using a coupled model of corrosion and crack growth, based on the corrosion state information and the crack information, generating predictive crack growth information; obtaining dynamic stress information of the metal component of the pressure switch during operation through a stress sensor;

[0067] S104: Based on the dynamic stress information, the corrosion state information, the predictive crack growth information and the action mode required by the currently requested control logic, determine whether the pressure switch has a risk of malfunction or has malfunctioned, and generate a detection result.

[0068] Regarding S101 above:

[0069] The purpose of this step is to obtain real-time or periodic corrosion damage accumulation information on the surface of the key metal parts of the pressure switch. In a specific implementation, these sensing electrodes can be pre-set on a specific surface area of ​​the metal parts of the pressure switch, or placed close to the surface. These areas can be corrosion-prone areas determined based on experience or stress analysis, such as the surface of a diaphragm or bellows that is in direct contact with the corrosive medium, weld areas, gap areas, or stress concentration areas. The sensing electrode itself can be made of corrosion-resistant conductive materials, such as precious metals platinum, gold, or conductive polymers, and attached to a flexible or rigid insulating substrate, such as a polyimide film, ceramic sheet, etc., to ensure that there is a stable and known insulating layer or small gap between the electrode and the metal part being measured.

[0070] During measurement, a specific electrical signal is applied between the sensing electrode and the metal component of the pressure switch through an external circuit, and the capacitance characteristics between the two are measured.

[0071] As an optional embodiment, an AC excitation signal with a predetermined frequency, such as a small-amplitude sinusoidal voltage, can be applied, and the resulting electrical response parameters, such as the AC current and its phase, or the complex impedance, can be measured. Based on these measurements and the known frequency, capacitive properties that characterize the electrode-interface-component system, such as capacitance (C), dielectric loss (tan δ), or the imaginary part of the complex impedance, can be calculated.

[0072] When metal components corrode, their surface develops physical and chemical properties, particularly a dielectric constant, that differ from that of the base metal. This can also cause changes in surface roughness, or the absorption or filling of dielectrics into microcracks. These changes can alter the equivalent capacitance between the sensing electrode and the metal component. Therefore, by comparing the currently measured capacitance characteristics with the capacitance characteristics under a baseline state, a capacitance change can be determined. For example, the corrosion-free state of the component at initial installation, or the state during the previous measurement cycle, may require correction based on environmental factors such as temperature.

[0073] This variation mainly reflects the change in interface state caused by the corrosion process. In order to obtain more useful corrosion state information for subsequent steps, it is usually necessary to convert this capacitance characteristic variation.

[0074] As an optional implementation, this conversion process can consider and compensate for the effects of stress on capacitance measurement and utilize a preset correlation model or calibration data. This model or data can convert the corrected or uncorrected capacitance characteristic changes into a more specific description of the corrosion state, such as the equivalent thickness growth value of the corrosion layer, surface corrosion coverage, specific types of corrosion indicators such as the pitting tendency index, or a comprehensive corrosion damage grade. This processed and interpreted information constitutes the "corrosion state information" required by S103 and S104.

[0075] Regarding S102 above:

[0076] At least two detection methods based on different physical principles are combined to obtain more accurate and comprehensive information. Specifically, this is achieved by analyzing the response of the metal components of the pressure switch to externally applied electromagnetic detection fields and acoustic detection waves.

[0077] For example, eddy current testing (ECT) can be used. Using an eddy current probe, eddy currents are induced on the component surface. Cracks (particularly surface and subsurface cracks) impede the flow of eddy currents, causing a change in the probe coil's impedance. Analysis of this electromagnetic response, such as the impedance amplitude and phase, can effectively identify surface-opening cracks, determine their location, and estimate their subsurface depth and length.

[0078] For example, ultrasonic testing (UT) can be used. An ultrasonic transducer transmits acoustic pulses into a component. These pulses are reflected and scattered upon encountering crack interfaces. Analysis of the received acoustic echo signals, such as echo arrival time, amplitude, and frequency spectrum, allows the detection of surface and internal cracks, and the precise determination of their location, depth (e.g., penetration depth), and dimensions such as length, height, and orientation. Advanced UT techniques, such as phased array ultrasonic testing (PAUT), also enable rapid imaging and more accurate dimensional measurements.

[0079] In a preferred embodiment, the results of electromagnetic and acoustic detection can be integrated. For example, eddy currents can be used to determine the location and initial depth of surface cracks, and ultrasound can be used to accurately measure their depth and internal morphology. Alternatively, ultrasound can be used to detect internal defects, and eddy currents can be used to confirm whether they extend to the surface.

[0080] The crack information obtained in this step is a comprehensive dataset that includes not only the geometric state of the crack at the current moment, but also its historical expansion trend. This historical trend is obtained by repeating the above crack identification and characterization process at different time points, such as t1 and t2. Then, the crack geometric information at the two time points, such as the difference in position, size, depth, etc., is compared to calculate the change amount and direction of the crack within the time interval, and then determine the expansion rate and dominant expansion direction that characterize the historical expansion trend. The expansion rate can be the surface expansion rate and the depth expansion rate. The final "crack information" generated provides key input for the subsequent prediction model (S103) and risk assessment (S104) regarding the structural damage state and its dynamic evolution.

[0081] Regarding S103 above:

[0082] This step consists of two parallel subtasks: making predictions using the model and obtaining real-time stress.

[0083] Generate predictive crack growth information: This part uses a coupled corrosion and crack growth model. The core purpose of this model is to predict the crack growth trend in the future based on the current known state information.

[0084] The model inputs include at least the corrosion state information obtained in step S101 and the crack information obtained in step S102, including the current geometric state and historical trends. The historical trend information can be used for model calibration or verification.

[0085] Coupled mechanisms: The model needs to reflect the interaction between corrosion and stress / cracking.

[0086] For example, corrosion can accelerate crack propagation, and the presence of cracks can also change the local corrosion environment. Models can be constructed based on modified fracture mechanics formulas, such as the Paris law that considers environmental effects, damage mechanics theory, or data-driven methods.

[0087] The output of the model is “predictive crack growth information”, which typically includes a prediction of the future crack growth rate and an estimate of the future crack size / depth or remaining service life based on this prediction.

[0088] Acquiring dynamic stress information: Simultaneously, stress sensors, such as strain gauges and fiber Bragg gratings (FBGs), are deployed at key locations on the pressure switch's metal components to monitor the dynamic stresses of the components in real time or periodically when subjected to actual operating pressures and environmental loads. This information reflects the true state of the mechanical loads driving crack growth, such as amplitude, frequency, and cyclic characteristics. This information serves as input for the coupled model to calculate parameters such as the stress intensity factor and is the basis for the final misoperation determination (S104).

[0089] Regarding S104 above:

[0090] This is the final decision-making step of the entire method. The real-time and predicted information obtained and generated in all previous steps is comprehensively analyzed and combined with the current operating instructions to determine the functional status of the pressure switch.

[0091] The core of this step is to integrate the following four aspects of information:

[0092] Real-time dynamic stress information (from S103): reflects the current mechanical load level.

[0093] Corrosion status information (from S101, can be compensated by S103): reflects the current degree of material chemical damage.

[0094] Predictive crack growth information (from S103): reflects the expected future structural damage state and risk.

[0095] The action mode required by the currently requested control logic provides context for evaluating whether the current state is “abnormal” or “risky.” For example, a small predicted crack growth might be acceptable in a low-pressure hold mode, but could pose a high risk in an impending high-pressure rapid switchover mode.

[0096] Regarding judgment logic and implementation, the judgment logic aims to evaluate whether the current comprehensive status, such as physical damage, predicted risk, real-time load, etc., meets the safety and functional requirements of the current action mode.

[0097] For example, thresholds / rules can be used to set safety thresholds for different action modes that take into account corrosion and crack conditions, such as maximum allowable stress or critical crack size. When the real-time stress or predicted crack condition exceeds the threshold for that mode, it is determined to be a risk or a false action.

[0098] For another example, a comprehensive risk index can be calculated based on the risk score according to various input information, and an alarm will be issued if the risk index exceeds a threshold.

[0099] For example, based on state comparison, the current stress and sensor response can be compared with the expected response of the "healthy" state in the action mode. Significant deviations may indicate functional abnormalities such as internal jamming and stiffness changes.

[0100] For another example, based on machine learning, a trained model, such as a classifier, can be used to directly determine the probability or category of a misoperation based on the four types of input information.

[0101] This ultimately generates a test result that reports the status of the pressure switch to the user or upper-level control system. The result can range from a simple normal / abnormal / high-risk indication to more detailed diagnostic information, such as a crack approaching criticality, suggesting a shutdown for inspection and maintenance.

[0102] In this way, the embodiment of the present invention systematically integrates in-situ corrosion monitoring, advanced non-destructive testing, physics-based predictive models, real-time stress monitoring and operating condition information to construct a comprehensive and intelligent pressure switch malfunction detection method, which can significantly improve the accuracy, early warning and reliability of detection.

[0103] As an optional implementation manner, generating crack information includes:

[0104] At a first time point, applying a time-varying electromagnetic detection field to a predetermined monitoring area of ​​the pressure switch metal component to detect a corresponding electromagnetic response signal, and transmitting an acoustic energy detection wave to the pressure switch metal component to detect a corresponding acoustic echo signal;

[0105] Calculating a first set of geometric feature information of the crack based on the electromagnetic response signal and the acoustic echo signal acquired at the first time point; the first set of geometric feature information includes: crack position information, size information, and depth information at the first time point;

[0106] At a second time point, the process of acquiring the corresponding electromagnetic response signal and the acoustic echo signal and calculating the geometric feature information based thereon is again performed to obtain a second set of geometric feature information of the crack; the second set of geometric feature information comprising: crack location information, size information, and depth information at the second time point;

[0107] Comparing the second set of geometric feature information with the first set of geometric feature information, and calculating a change amount and a change direction of the position, size, and depth of the crack during a time interval between the first time point and the second time point;

[0108] Determining a historical crack growth trend based on the change amount and change direction;

[0109] The second set of geometric feature information is combined with the historical expansion trend to generate the crack information.

[0110] The specific implementation includes the following main stages:

[0111] Stage 1: Determination of initial crack state:

[0112] First, the metal components of the pressure switch need to be inspected at an initial or baseline time point (t1) to determine the crack status at that moment. Inspections are typically focused on predetermined monitoring areas, which are areas identified as prone to cracking based on stress analysis, historical failure data, or engineering experience. These areas include the center of the pressure switch diaphragm, edge transition areas, weld-affected areas, or other high-stress concentration areas.

[0113] At this first time point t1, detection and calculation are performed:

[0114] For applying a probe field / wave and detecting a response signal:

[0115] As described above with respect to S102 , a time-varying electromagnetic detection field is applied to the predetermined monitoring area, for example, by an eddy current probe, and a corresponding electromagnetic response signal caused by the crack is detected, for example, a change in impedance or induced voltage.

[0116] The two methods also emit acoustic energy probe waves to the same area or related volume. For example, an ultrasonic transducer transmits ultrasonic pulses and detects the corresponding acoustic echo signals reflected or scattered back by the crack. The raw signals obtained by these two detection methods are collected and recorded separately or simultaneously.

[0117] To calculate the first set of geometric feature information:

[0118] The electromagnetic response signal and acoustic echo signal collected at time t1 are subjected to necessary signal processing, such as filtering, noise reduction, amplification, and extraction of key features, such as the phase and amplitude of the eddy current signal; the time of flight ToF, amplitude, and frequency component of the ultrasonic echo.

[0119] Based on these processed signal characteristics, an appropriate non-destructive testing inversion algorithm or physical model is applied to calculate the geometric parameters of the cracks existing in the monitoring area at the first time point t1.

[0120] For example: estimating depth based on eddy current phase, positioning and depth measurement based on ToF, determining crack length and width based on echo amplitude or imaging results, etc.

[0121] In addition, a data fusion strategy can be used to combine electromagnetic and acoustic detection results to improve the accuracy and completeness of geometric feature information. The calculation results constitute the first set of geometric feature information, which describes the crack state at time t1. Specifically, it includes crack location information, such as coordinates within the monitoring area, size information, such as length, width, or area, and depth information, such as maximum penetration depth or depth profile.

[0122] Phase 2: Subsequent crack state tracking and change quantification:

[0123] In order to determine whether and how the crack is growing, repeated detection and analysis is required at a predetermined second time point (t2) after the first time point.

[0124] In practice, at time t2 (t2 > t1), the same process as at time t1 is repeated to acquire the corresponding electromagnetic response signal and acoustic echo signal, and the geometric feature information is calculated based on these newly acquired signals. To ensure comparability, the detection at time t2 should generally use the same or equivalent detection equipment, setting parameters, and scanning strategy as at time t1, and cover the same predetermined monitoring area.

[0125] The result of the calculation is a second set of geometric feature information, which describes the crack state at time t2, including, for example, position, size, and depth.

[0126] The second set of geometric feature information is then compared with the first set of previously stored geometric feature information. During this comparison, it is important to ensure that the damage evolution of the same crack or region is being compared. Image registration or feature matching algorithms can be used to accurately match the crack features at times t1 and t2.

[0127] By comparison, the amount and direction of change in each dimension of the crack within the time interval Δt (Δt = t2 - t1) can be calculated. For example, the change vectors at the two end points of the crack can be calculated to determine the surface propagation direction and length increment, the increase in maximum depth, and the change in crack area.

[0128] Phase 3: Identify historical trends and generate final information:

[0129] Based on the change amount and direction calculated in the previous step, and combined with the known time interval Δt (or the number of working cycles experienced during the period ΔN), the historical expansion trend of the crack in the most recent monitoring period can be quantitatively determined.

[0130] This trend information can be expressed in the form of an expansion rate, for example:

[0131] The surface expansion rate (in mm / hour, mm / day, or mm / cycle) and the corresponding main expansion direction, i.e., the angle relative to a coordinate axis.

[0132] Depth expansion rate (same units as above).

[0133] It may also include a description of the expansion mode, for example, whether it is primarily length expansion, primarily depth expansion, or both.

[0134] Finally, the second set of geometric feature information, such as the current state of the representative crack at the latest monitoring time point t2, is combined with the calculated historical expansion trend of the recent development dynamics of the representative crack.

[0135] This combination means that the resulting crack information includes not only where the crack is, how big it is, and how deep it is, but also dynamic information about how fast it is developing and in which direction.

[0136] This comprehensive “crack information” package, which includes the current geometric state and historical expansion trends, is the key input data provided to the coupling model in step S103 for future predictions and to step S104 for final risk judgment.

[0137] In this way, the present application can achieve a leap from static characterization to dynamic tracking of internal crack damage in pressure switches, laying the foundation for state-based and predictive malfunction detection.

[0138] In order to further improve the accuracy of the corrosion status information generated in step S101, especially in actual working conditions where the pressure switch is usually subjected to dynamic stress loads, an embodiment of the present invention provides a preferred technical solution, which includes a process for compensating for stress effects when generating corrosion status information.

[0139] As an optional implementation manner, generating corrosion status information includes:

[0140] accessing a preset stress-capacitance response relationship, the relationship representing a capacitance characteristic of a metal component of the pressure switch in a known corrosion state versus an applied stress;

[0141] Calculating an expected corrosion-free capacitance characteristic reference value corresponding to the real-time dynamic stress information using the stress-capacitance response relationship and the dynamic stress information acquired in real time;

[0142] comparing the capacitance characteristic measured in real time with a calculated baseline value of the expected capacitance characteristic without corrosion to determine an amount of capacitance characteristic deviation caused by corrosion;

[0143] Based on the capacitance characteristic deviation, a preset correlation model or calibration data for correlating the capacitance deviation to the corrosion state is applied to generate the corrosion state information.

[0144] Among them, the mechanical stress applied to the metal components of the pressure switch is not only a key factor in driving structural damage such as cracks, but it may also directly cause changes in the capacitance characteristics measured by the sensing electrodes. This change may be due to minor geometric deformation of the components caused by stress, such as changing the effective distance or relative posture between the sensing electrode and the surface of the metal component, or the stress directly acting on the passivation film or the formed corrosion product layer on the metal surface, changing its microstructure or dielectric properties, which is called the pressure-induced dielectric effect. This change in capacitance characteristics directly induced by stress will be superimposed on the change in capacitance characteristics caused by corrosion processes, such as the accumulation of corrosion products and the evolution of surface morphology. If the changes from these two sources cannot be effectively distinguished, and the overall change in capacitance characteristics is directly attributed to corrosion, especially in the case of large stress fluctuations or weak initial corrosion signals, it may lead to significant deviations in the assessment of the actual corrosion state, thereby affecting the accuracy of subsequent predictions and judgments, and even resulting in false positives or missed reports.

[0145] Therefore, the core of this optional embodiment is to separate the capacitance characteristic changes mainly contributed by the corrosion process. To achieve this goal, the method utilizes the dynamic stress information that can be obtained in real time in the overall solution of the present invention.

[0146] Specifically, the process of generating the corrosion state information, in one embodiment, first requires access to a preset stress-capacitance response relationship. This relationship is pre-established and stored, and it characterizes how the capacitance characteristics of the metal component of the pressure switch measured by a specific sensing electrode configuration change with the stress applied thereto when the component is in a known baseline corrosion state, for example, the initial uncorroded state of the component, or the state where the corrosion has reached a certain stable stage. This relationship is equivalent to establishing a "pure" influence map of stress on capacitance measurement in the absence of additional corrosion. This relationship can be established by applying precisely known stress loads of different levels to representative samples under controlled conditions, and synchronously recording the corresponding capacitance characteristics using the same sensing electrodes and measurement methods as those used in actual monitoring, and finally obtaining it through data fitting, such as obtaining a functional relationship, or constructing a lookup table.

[0147] During real-time monitoring, the newly accessed stress-capacitance response relationship is then used in conjunction with the dynamic stress information acquired in real time from the stress sensor, such as the current stress measurement, to calculate a baseline value for the expected corrosion-free capacitance characteristic. This calculated baseline value represents the expected capacitance characteristic reading under the current real-time stress, assuming the corrosion state remains at the known baseline state from when the response relationship was established, and no further corrosion occurs. Because the real-time stress changes dynamically, this expected baseline value is also dynamically adjusted.

[0148] Next, the capacitance characteristic measured in real time by the sensing electrodes (the direct measurement result from S101) is compared with the expected baseline capacitance characteristic value for a corrosion-free state calculated in the previous step. This comparison typically involves calculating the difference between the two. This difference represents the capacitance characteristic deviation due to corrosion. Because the expected baseline value already includes the direct impact of the current stress state, this deviation theoretically primarily reflects the cumulative capacitance characteristic change caused by the corrosion process itself (relative to the baseline corrosion state when the relationship was established).

[0149] Finally, based on this capacitance characteristic deviation that has been corrected for stress effects, another preset correlation model or calibration data is needed to associate the capacitance deviation with the corrosion state to generate the final corrosion state information that can be used in subsequent steps. The purpose of this correlation model or data is to "translate" the electrical deviation into a corrosion state description with physical or engineering significance. Its establishment can be based on physical principles, such as an equivalent circuit model or analytical model that simulates the effect of corrosion layer growth on capacitance, or obtained through a large number of controlled corrosion experiments, such as measuring the capacitance deviation of samples with different corrosion levels and calibrating them, or using machine learning technology to learn this mapping relationship from data. The corrosion state information finally generated can be a quantitative parameter, such as an estimate of the equivalent thickness of the corrosion layer, an indirect indicator of the corrosion rate, or a qualitative description such as a corrosion type indicator, corrosion damage level, or a combination of these information.

[0150] In this way, the present application can significantly improve the accuracy and reliability of corrosion monitoring in a dynamic stress environment, and provide more realistic and reliable input for the coupling model prediction in the subsequent step S103 and the malfunction risk judgment in step S104.

[0151] In order to obtain the corrosion status information by measuring the capacitance characteristics mentioned in step S101, as an optional implementation, measuring the capacitance characteristics between the sensing electrode and the metal component of the pressure switch includes:

[0152] Applying an AC excitation signal with a predetermined frequency between the sensing electrode and the metal component of the pressure switch;

[0153] measuring an electrical response parameter generated between the sensing electrode and the metal component of the pressure switch in response to the AC excitation signal, wherein the electrical response parameter is selected from at least one of an AC response current, an AC response voltage, or a complex impedance;

[0154] The capacitance characteristic is calculated based on the measured electrical response parameter and the predetermined frequency.

[0155] It is important to understand that the “capacitance characteristics” obtained here are raw measurements that are affected by multiple factors, including corrosion and stress states.

[0156] In a specific implementation, an AC excitation signal with a predetermined frequency is applied between the sensing electrode and the metal component of the pressure switch. The "sensing electrode" and "metal component" here refer to the electrodes arranged on or near the surface of the metal component of the pressure switch and the metal component itself as mentioned above. The applied AC excitation signal can be a small-amplitude sinusoidal wave signal to avoid excessive disturbance to the measured system, especially the electrochemical interface that may exist, to ensure the effectiveness and linearity of the measurement. The signal can be an AC voltage source signal generated by a signal generator or a microcontroller, applied between the electrode and the component; or it can be an AC current source signal, forcing a known AC current to flow between the two.

[0157] The AC excitation signal has a predetermined frequency (f). This frequency is a key parameter that is pre-selected or set based on application requirements. Its selection may require consideration of multiple factors.

[0158] For example, maximizing sensitivity to expected changes in the corrosion product layer, as its dielectric properties are often frequency-dependent; effectively distinguishing capacitive effects from resistive effects or interfacial electrochemical processes such as double-layer charging, the latter of which is more significant at low frequencies; avoiding known ambient electromagnetic interference frequency bands; and the optimal operating frequency range of the measurement circuitry and sensors used.

[0159] Although the use of a predetermined frequency is described herein, in other embodiments, measurements may be performed using multiple different frequencies to obtain richer information, or the frequency itself may be dynamically adjusted.

[0160] Next, after applying the AC excitation signal, it is necessary to measure the electrical response parameters between the sensing electrode and the metal component in response to the AC excitation signal. The specific parameters measured depend on the type of excitation applied and the design of the measurement system.

[0161] For example, if an AC voltage signal is applied, it is necessary to measure the AC response current flowing through the "electrode-interface-component" equivalent capacitor. The key is to accurately measure its amplitude and phase difference relative to the excitation voltage.

[0162] For example, if an AC current signal is applied, it is necessary to measure the AC response voltage generated between the sensing electrode and the component, and also to obtain its amplitude and phase difference relative to the excitation current.

[0163] Alternatively, the complex impedance (Z) between the two can be measured directly using impedance measurement equipment. Complex impedance is a complex number that contains information about its magnitude (|Z|) and phase (φ), or equivalently, resistance (R, real part) and reactance (X, imaginary part).

[0164] Therefore, the concept of electrical response parameters is at least one of the above-mentioned AC response current, AC response voltage, or complex impedance, or a combination thereof, such as amplitude and phase. The equipment performing the measurement needs to have sufficient accuracy and bandwidth to process the signal at the selected frequency.

[0165] Finally, the capacitance characteristic is calculated based on the measured electrical response parameter and the predetermined frequency. This step is to extract or calculate the capacitance characteristic from the original electrical measurement result.

[0166] The specific calculation method depends on the measured parameters and the equivalent circuit model used to describe the "electrode-interface-component" system. The simplest model is the series RC or parallel RC model.

[0167] For example, if the real part R and imaginary part X of the complex impedance Z are measured and the angular frequency ω = 2πf is known, for the series model, the capacitive reactance Xc = X, then the capacitance characteristic C can be calculated as C = -1 / (ωX).

[0168] If the voltage amplitude |V|, current amplitude |I|, and phase difference φ are measured, the impedance amplitude |Z| = |V| / |I|, the real part R = |Z|cosφ, and the imaginary part X = |Z|sinφ can be calculated first, and then the capacitance characteristics can be calculated.

[0169] The final calculated capacitance characteristic primarily refers to the capacitance value C (in pF or nF), but depending on the specific application scenario and subsequent processing requirements, it may also refer to the capacitive reactance Xc, the imaginary part of the complex impedance X, or the complex capacitance containing capacitance and loss information, etc., which can reflect the capacitive behavior of the system. This calculated value is the raw measurement result used to generate corrosion status information directly when performing stress compensation calculations or ignoring stress effects.

[0170] In this way, the present application can reliably complete the basic electrical measurement of the capacitance characteristics between the sensing electrode and the metal parts of the pressure switch, providing one of the key raw sensing data for the entire malfunction detection process.

[0171] See Figure 2 , Figure 2 A flowchart of a method for generating predictive crack growth information provided in an embodiment of the present application includes steps S201 to S203, wherein:

[0172] S201: Mapping the corrosion state information and the crack information into damage states of multiple spatial regions on the metal component of the pressure switch;

[0173] S202: establishing and utilizing, in the coupling model, a mutual influence relationship between the damage states of the plurality of spatial regions;

[0174] S203: Calculating the predicted crack growth information for each spatial region based on the mutual influence relationship of the damage states and the current damage state of each region; wherein the mutual influence relationship of the damage states reflects at least one or more of the following effects: changes in the local stress distribution of other regions due to damage in one region, and electrochemical coupling effects between different regions.

[0175] See Figure 3 , Figure 3 A schematic diagram of a metal component of a pressure switch provided in an embodiment of the present application, which schematically shows the distribution of multiple possible damage areas.

[0176] In a preferred embodiment, Figure 3 As shown, the metal component of the pressure switch includes a pressure switch body 1, a threaded connection 3, and multiple damaged areas distributed on the metal component's surface. The pressure switch body 1 is a metal package that carries the primary structure and sensing functions. Its surface is typically prone to damage such as corrosion and cracking, and is typically covered with sensing electrodes, stress sensors, or non-destructive testing interfaces. The threaded connection 3 is located at the lower end of the pressure switch body 1 and is used to reliably install the entire pressure switch into the interface structure of the target system, forming an airtight or liquid-tight seal to ensure pressure measurement accuracy and system stability.

[0177] like Figure 3 As shown, multiple independent damage areas exist at different locations on the surface of the pressure switch body 1, including a first damage area 2a, a second damage area 2b, and a third damage area 2c. These multiple damage areas can be formed by various mechanisms such as local stress concentration under long-term service conditions, electrochemical corrosion, mechanical impact, or residual welding defects. They usually manifest as surface cracks, pits, localized material erosion, or depressions.

[0178] The spatial distribution between different damage areas is non-uniform, and there may be mechanical and electrochemical interactions, that is, the damage state of one area may affect the damage evolution process of the adjacent area through stress field coupling or corrosion galvanic mechanism. In order to improve the accuracy and physical authenticity of the predictive crack extension information generated in step S103, especially when there may be multiple damage areas or uneven damage distribution on the metal component of the pressure switch, a preferred embodiment of the present application adopts the above-mentioned coupling model that considers spatial interactions. Traditional coupling models may only be calculated based on a single crack or local information of a certain monitoring point, or the entire component may be treated as a whole for averaging, which may ignore the important physical interactions between different damage areas. The damage state of a region, whether it is corrosion or cracks, does not exist in isolation. It can affect the damage evolution process of adjacent or even distant regions through a variety of physical mechanisms.

[0179] For example, the presence of a major crack can significantly change the stress field distribution around it and even throughout the entire component. Corrosion severity or material differences in different regions can lead to different electrochemical potentials, forming corrosion couples in the presence of electrolytes, thereby accelerating or slowing down corrosion in specific areas. Ignoring these spatial coupling effects can lead to significant deviations in predictions of damage growth rates and remaining life.

[0180] In a specific implementation, the corrosion status information and crack information obtained from steps S101 and S102 are mapped from point or local measurement results to the damage status of multiple spatial areas on the key pressure-bearing or pressure-sensing areas of the pressure switch metal parts; the key areas may be diaphragms, bellows, high-stress connections or other areas prone to cracks.

[0181] The critical areas of the component must first be spatially discretized. For example, this can be divided into a two-dimensional or three-dimensional grid, which can be regular or irregular, or using the cell grid used in finite element analysis, or defining monitoring sub-areas based on the physical arrangement of sensors, such as multiple capacitive electrodes.

[0182] Then, the acquired sensor data with spatial attributes, such as the corrosion status information corresponding to each capacitor electrode and the location, size, and depth of cracks obtained by non-destructive testing (NDT), are assigned or interpolated to corresponding spatial regions, such as grid cells, finite element cells, or sub-regions.

[0183] In this way, a digital representation is obtained that can describe the spatial distribution of damage states in the entire monitoring range at the current moment, where the damage states can include corrosion and crack information, etc. For example, a damage field or a list of state vectors for each area / unit is obtained.

[0184] Furthermore, within the coupled model, the mutual influence of damage states between the multiple spatial regions is established and utilized. This means that when calculating the future damage evolution of any specific region, the model not only considers the region's current damage state and external loads, but also the influence of the states of other regions on it.

[0185] The establishment and utilization of this mutual influence relationship can be achieved based on different model frameworks. For example, an "influence matrix" or a set of "influence functions" can be pre-calculated or estimated in real time to quantify the degree to which damage in one area affects the key physical quantities in another area. During the prediction calculation, the driving force of the target area is corrected by superimposing the influence of all other areas. Among them, damage can be a crack per unit length, a corrosion layer per unit thickness, etc.; key physical quantities can be stress intensity factors, local corrosion potential / current density, etc.

[0186] For example, multiple physical field equations can be coupled and solved directly on the finite element mesh. This approach can include solving solid mechanics equations to calculate the stress and strain field distributions that are updated as the crack state changes, thereby reflecting the mechanical interaction; at the same time, the method can also include solving electrochemical equations to calculate the potential and current field distributions that reflect the galvanic effect between regions, thereby reflecting electrochemical coupling. Within this finite element framework based on multi-physics coupling, the impact of damage in one region on other regions is naturally and comprehensively reflected by solving the global system of equations.

[0187] Furthermore, based on the above-mentioned establishment and utilization of the mutual influence relationship between damage states and the current damage state of each region (derived from the mapping results), the coupled model performs damage evolution calculations on each spatial region separately or as a whole coupled model to generate predictive crack growth information for the region.

[0188] Because the model accounts for these interactions, its predictions better reflect real-world conditions. For example, the model can predict that a small defect located near an existing larger crack will grow significantly faster than if it existed in isolation due to the transfer of stress concentration effects. Alternatively, the corrosion rate or stress corrosion crack growth rate of metal located in the anodic region (relative to adjacent regions) will be accelerated due to electrochemical coupling, even if the local stress is not high.

[0189] The mutual influence relationship of the damage states at least reflects one or more of the following physical effects:

[0190] For example, damage in one area alters the local stress distribution in other areas, a major coupling mechanism in mechanics. Crack formation and propagation, or localized material removal due to corrosion, alter the stiffness and geometry of a component, leading to a redistribution of the stress field throughout the component. This directly impacts stress levels and concentrations in other areas, regardless of pre-existing damage, and thus affects the fatigue life or crack growth drivers in those areas.

[0191] Another example is the electrochemical coupling effect between different regions: This is the main coupling mechanism in electrochemistry. When different spatial regions exhibit different corrosion potentials due to differences in material composition, such as welds and base materials, different surface conditions such as differences in passivation film integrity, differences in corrosion product coverage, or even different local stress levels, if there are ion conductive pathways between them, microscopic or macroscopic corrosion couples will be formed, such as through surface condensation water films, internal leakage media, or conductive corrosion products. The area with lower potential acts as the anode, and corrosion is accelerated; the area with higher potential acts as the cathode, and corrosion is suppressed, but may be accompanied by other reactions such as hydrogen evolution. This electrochemical coupling can significantly change the actual local corrosion rate and corrosion morphology.

[0192] It is understandable that more complex models can also consider other interactions, such as the coupling of temperature field and stress field, thermal effects of corrosion process, etc., which will not be discussed here.

[0193] By considering the coupled model of spatial interactions, the method of the present invention can more accurately and realistically predict the complex, non-uniform, and interconnected damage evolution processes that may occur on pressure switch components. This provides high-quality input information for reliable malfunction risk assessment and remaining life prediction in step S104.

[0194] As an optional implementation manner, establishing and utilizing the mutual influence relationship of damage states among the multiple spatial regions includes:

[0195] Abstracting the multiple spatial regions into nodes of a graph neural network, and constructing a node feature vector for each node; wherein the node feature vector includes: the corrosion state information, the crack information, the dynamic stress information, and the capacitance characteristic;

[0196] Determine the edge connection and edge feature vector between the nodes based on whether the physical proximity relationship between the nodes, the physical field gradient between the nodes, and the state parameter difference between the nodes meet preset conditions;

[0197] Calling a message passing graph neural network pre-trained based on historical prototype data, inputting the node feature vector and the edge feature vector into the message passing graph neural network, and outputting a crack propagation driving force index of the corresponding node;

[0198] The crack growth driving force index is used in the coupling model to calculate the predictive crack growth information of each spatial region.

[0199] In order to more accurately consider the prediction model of the mutual influence of damage states between spatial regions, a particularly preferred embodiment of the present application uses a graph neural network (GNN) technology. GNN is suitable for processing non-Euclidean data with complex connections and interactions. It can learn and quantify such mutual influence between regions from the data, such as discretized spatial regions on a component and their relationships.

[0200] In practice, the multiple spatial regions mentioned above are abstracted as nodes in a graph structure. Each node in the graph represents a specific spatial location or region on the metal component of the pressure switch. For example, it can be a cell defined by a finite element mesh or a monitoring area defined by the sensor location.

[0201] Then, for each node, a node feature vector needs to be constructed. This vector is a multi-dimensional digital representation of the current state of the node (region), and its information comes from the various upstream sensing and calculation results defined above.

[0202] In one embodiment, the node feature vector may include or encode features derived from the following information:

[0203] The corrosion state information: for example, the equivalent thickness of the corrosion layer in the area corresponding to the node, a corrosion type indicator, or a characteristic value converted from the capacitance characteristic deviation itself.

[0204] The crack information includes, for example, a Boolean flag indicating whether a crack exists in the node, and if a crack exists, its geometric characteristics in the region, such as length, maximum depth, and quantitative representation of area ratio, as well as historical crack expansion trend information in the region.

[0205] The dynamic stress information: for example, the real-time or periodic statistical stress state characteristics of the position corresponding to the node, such as the maximum principal stress, von Mises stress, stress amplitude, stress ratio, etc.

[0206] The measured capacitance characteristics, for example, the capacitance value and dielectric loss measured by the sensing electrode at the node, serve as additional information reflecting the surface state.

[0207] It should be noted that features from different sources and dimensions need to be normalized before being input into GNN.

[0208] In practice, a graph structure requires not only nodes but also connections between nodes, known as edges, and the attributes of these connections, known as edge feature vectors. Edges are constructed to represent the paths and strengths of possible interactions between regions.

[0209] Initial edge connections are established based on the physical proximity of nodes. For example, if the nodes represent FEA grid cells, connections are established between adjacent cells; or, if the nodes represent monitoring areas, connections are established between pairs of nodes whose spatial distance is less than a preset threshold, such as 5 mm. This constitutes the basic topological structure of the graph, primarily representing nearest-neighbor interactions.

[0210] As an optional implementation, in order to better capture possible long-range coupling or interactions driven by specific physical mechanisms, the edge connection relationship may not be fixed. The edge connection can be dynamically determined or adjusted based on whether the physical field gradient or state parameter difference between the nodes meets the preset conditions. For example, when the stress gradient or potential difference between two nodes exceeds a certain threshold, it can be considered that there is a significant mechanical or electrochemical interaction between them, thereby adding an edge between them or enhancing the weight of the existing edge. This adaptive edge construction method enables the graph structure to better reflect the key interaction paths in actual physical processes.

[0211] The physical field gradient is, for example, the stress gradient between nodes calculated based on stress sensor data or FEA; the state parameter difference is, for example, the difference between nodes calculated based on capacitance characteristics or corrosion potential inferred from corrosion state information.

[0212] For the construction of edge feature vectors:

[0213] For each determined graph edge, an edge feature vector can be constructed to describe the properties of this connection, which helps GNN distinguish the importance or type of different connections during message passing.

[0214] For example, edge feature vectors can include: geometric information between nodes, such as three-dimensional distances and relative direction vectors; gradient or difference information of the physical fields between nodes, such as stress gradient values, capacitance characteristic differences, and corrosion potential differences used to determine edge connections; and categorical features indicating the connection type, such as whether it is a purely mechanical edge or an electrochemical edge. These features also typically require normalization.

[0215] Furthermore, after constructing the graph structure containing node features and edges, a graph neural network is called to process it.

[0216] Among them, the message passing neural network (MPNN) architecture is preferred for network types. The core mechanism of this type of network is that the calculation of each layer includes two stages:

[0217] In the message aggregation phase, each node collects information ("messages") from its neighbor nodes. These messages are usually the result of some transformation of the features of the neighbor nodes and the features of the connecting edges, such as a neural network layer;

[0218] In the node update phase, each node combines the aggregated neighbor information with its current representation and updates its representation through another transformation, such as a neural network layer. The current representation can be a feature vector or a hidden state.

[0219] The MPNN can be pre-trained based on historical prototype data or simulation data. This means the network has already learned general laws regarding corrosion, cracks, stress, and the spatial interactions between them. The training goal is typically to enable the network to accurately predict a key indicator related to damage evolution.

[0220] In real-time monitoring, the currently constructed graph, including the latest node and edge features, is input into the pre-trained MPNN. After the network's multi-layer message passing and node update calculations, the MPNN outputs one or a set of values ​​for each node.

[0221] Furthermore, the MPNN output is designed to be an index of the crack growth driving force at the corresponding node. This index is a comprehensive quantity that not only reflects the damage state and stress experienced by the node itself, but also implicitly incorporates information about the mutual influences aggregated from neighboring and related regions via the MPNN's message passing mechanism. It can be a physically well-defined quantity, such as the equivalent stress intensity factor K_eff, the J-integral, or the energy release rate G, or a learned, abstract driving force index that is highly correlated with these physical quantities.

[0222] Finally, the coupled model uses the node-level crack growth driving force index output by the MPNN to calculate the final predictive crack growth information for each spatial region.

[0223] This means that MPNN in this scheme does not directly predict the final crack length or life, but provides a more accurate “driving force” input that takes into account spatial interactions.

[0224] It is understandable that subsequent calculations may be to substitute this driving force index into a physical crack growth law to calculate the crack growth rate of the node, such as the Paris law.

[0225] Then, based on this calculated rate, the crack growth amount in the future period can be predicted, the crack information can be updated, and finally the predictive crack growth information including the rate and future range can be obtained.

[0226] By adopting this MPNN-based implementation, the present application can efficiently and accurately capture and quantify complex spatial interactions of damage using a data-driven and physical-information-based approach, generating more reliable predictive crack growth information than traditional methods. This approach is particularly suitable for dealing with practical engineering problems with non-uniform damage distribution and complex coupling effects.

[0227] As an optional implementation manner, establishing and utilizing the mutual influence relationship of damage states among the multiple spatial regions further includes:

[0228] monitoring the predicted crack growth information generated by the coupling model, comparing it with the first set of geometric feature information and the second set of geometric feature information acquired at different subsequent time points, and determining a deviation between actual crack geometry changes;

[0229] In response to the deviation exceeding a preset threshold, triggering an incremental learning process;

[0230] The incremental learning process performs online fine-tuning on the weights of the graph neural network using a monitoring dataset; the monitoring dataset is associated with a monitoring period corresponding to the deviation, and the monitoring dataset includes input information obtained during the monitoring period for generating the node feature vector and the edge feature vector, as well as the corresponding actual crack geometry changes.

[0231] When using graph neural network-based methods to establish and utilize spatial damage interactions, although the pre-trained model may perform well initially, as the actual service life of the pressure switch increases, changes in operating conditions, slight drifts in material properties, and the cumulative effects of complex physical and chemical processes that the model itself fails to fully capture may cause the pre-trained model's prediction accuracy to gradually decrease over time. To maintain or improve the long-term accuracy and reliability of the model's predictions, this application further introduces an online fine-tuning or incremental learning mechanism.

[0232] The core idea of ​​this mechanism is to use the actual monitored crack extension conditions to provide feedback and adjust the parameters of the GNN model.

[0233] In a specific implementation, the deviation between the predicted crack growth information generated by the coupled model and the actual observed crack geometry changes is continuously monitored, for example, the predicted value of the crack growth amount or rate in each spatial region within a future monitoring period Δt.

[0234] The actual crack geometry change is determined by repeating the crack information acquisition process described above after the monitoring cycle ends, for example, at time t3. This involves performing a new electromagnetic / acoustic survey, calculating geometric signatures, and comparing these signatures with the geometric signatures at the beginning of the cycle, for example, at time t2. This comparison yields the actual expansion and direction of each monitoring region or critical crack over time Δt. A measure of the deviation between the predicted and observed values ​​is then calculated, such as the root mean square error (RMS) or maximum absolute error (MBE) between the predicted and actual expansions for all monitoring regions, or the prediction error for a specific critical crack.

[0235] Next, in response to the deviation exceeding the preset threshold, the incremental learning process is triggered. It is necessary to set one or more preset thresholds for prediction deviation. This threshold represents the acceptable lower limit for the model's prediction accuracy. When the monitored deviation continuously or significantly exceeds this threshold, the system determines that the current GNN model may no longer accurately reflect the actual damage evolution law and needs to be adjusted. At this point, the system automatically triggers the incremental learning process.

[0236] The core of the incremental learning process involves online fine-tuning of the graph neural network's weights using a monitoring dataset. This "monitoring dataset" for fine-tuning is crucial. It is associated with one or more monitoring cycles that triggered the incremental learning and includes real-world information from those cycles.

[0237] Exemplarily, its content includes at least: the original input information obtained at the beginning of the monitoring cycle and used to generate the node feature vectors and edge feature vectors of the GNN input at that time, such as the corrosion state, crack geometry, stress information, capacitance characteristics, etc. at that time.

[0238] And the corresponding actual crack geometry changes determined by actual measurement and comparison at the end of the monitoring period, such as the actual growth of crack length / depth of each node / region, serve as the "true label" or target output for fine-tuning.

[0239] The fine-tuning dataset can be formed by collecting data from the most recent one or more deviation exceeding standard cycles.

[0240] Furthermore, online fine-tuning can be performed during normal equipment operation or during specific maintenance windows. It uses the aforementioned monitoring dataset to make small, targeted adjustments to the GNN model's weights. Fine-tuning can employ standard gradient descent optimization algorithms, or specialized strategies tailored to online, small-batch, and forget-prevention requirements, such as:

[0241] Use a low learning rate to avoid making large changes to the model weights that could destroy the general knowledge it has learned from pre-training.

[0242] You can fine-tune the weights of only some layers, such as the layers close to the output or the layers related to the attention mechanism, and freeze the weights of other layers.

[0243] Regularization techniques are used to limit the magnitude of weight updates or to constrain old weights to alleviate the "catastrophic forgetting" problem.

[0244] By executing this incremental learning and online fine-tuning process, the GNN model’s weights are adjusted and optimized based on the latest actual observation data, making it better adapted to the specific state and evolution of the current device. This improves the accuracy and timeliness of the misoperation risk assessment (S104) based on this information.

[0245] As an optional implementation, when a graph neural network (GNN)-based approach is used to model the mutual influence of damage states as described above, there is a particularly preferred decision-making implementation method that can fully utilize the rich information learned by the GNN. This method aims to achieve smarter and more accurate risk identification through the deep feature representation generated by the GNN combined with the operating conditions.

[0246] In a specific implementation, a feature vector that can represent the overall damage state of the metal parts of the pressure switch is extracted or aggregated from the final state representation of each node generated by the message passing graph neural network.

[0247] Among them, after GNN processes node and edge features through message passing, the hidden state of each node includes the characteristics of the node itself as well as information about the structure and characteristics of its neighborhood, comprehensively reflecting the complex state of the local area and its context. In order to obtain a vector representing the state of the entire monitored component, a "graph readout" or "graph pooling" operation needs to be performed. This can be to aggregate the final state representation vectors of all nodes or selected key nodes, such as by summing, averaging, taking the maximum value, etc., or through a specially learned readout function / network layer to map the node representation set into a fixed-dimensional graph feature vector. This graph feature vector is a highly concentrated and information-rich representation of the current damage state of the entire component by GNN.

[0248] Next, the action mode required by the currently requested control logic is identified. This information comes directly from the control system or its operating logic, indicating the current or upcoming operating mode, such as high-pressure hold, rapid cycling, or low-pressure standby. This action mode information is crucial for determining current risk, as different modes correspond to different safety requirements and failure consequences. This mode information is a categorical variable and requires appropriate representation to facilitate subsequent model processing, such as one-hot encoding or other categorical embedding methods.

[0249] Then, the feature vector representing the intrinsic damage state of the component generated in the previous step is input into a pre-trained risk discrimination model together with the representation of the identified current action mode (representing the external operating condition).

[0250] This risk discrimination model is an independent model or a model integrated with GNN. Its task is to directly output the judgment about the risk of misoperation based on the input component state representation and operation mode.

[0251] For example, the model can be of various types, such as:

[0252] Traditional machine learning classifiers: such as support vector machines, random forests, gradient boosting trees, etc., can directly process the concatenated feature vectors and action pattern representations.

[0253] Neural network models: such as a simple multi-layer perceptron, which receives concatenated input and outputs risk probability or category after passing through several hidden layers.

[0254] It can also be the last few layers of the GNN model, which are specially designed for graph-level classification tasks.

[0255] A pre-trained risk discrimination model requires prior training on a dataset that includes historical data to learn the mapping relationship from input to output. For example, this historical data can include feature vectors for different damage states, corresponding motion patterns, and labels indicating whether a misoperation occurred or the risk level.

[0256] Finally, the detection result is generated using the output result of the risk discrimination model.

[0257] The output of the risk identification model can be directly: probability of false operation, that is, a value between 0 and 1; risk level, for example, classification results of three levels: low, medium, and high; status category, for example, normal, need attention, alarm / false operation.

[0258] Based on this output, the system generates a final detection result, which can then be used to trigger an alarm, displayed to an operator, or input into a maintenance decision system.

[0259] In this way, compared with the single physical indicator that only relies on the output of GNN or the traditional simple fusion method of multi-sensor data, it has the following advantages: the feature vector can capture the overall, spatially distributed, and interacting complex damage state of the component more comprehensively and deeply, and improve the intelligence level and performance of pressure switch malfunction detection.

[0260] Furthermore, while the eigenvector, as a potential representation generated by a deep learning model, can well summarize the overall and complex structural damage state of a component, it may not be sensitive enough to certain real-time, localized, rapid physical and chemical changes, especially those occurring at the metal / electrolyte interface, or its response may be delayed. For example, if a crack suddenly fills completely with a conductive medium, or if the protective corrosion product layer on the surface rapidly dissolves or peels off locally, these events may first and significantly manifest themselves in real-time changes in capacitance characteristics, and the GNN's eigenvector may not yet be fully updated to reflect this sudden change.

[0261] To further improve the sensitivity and robustness of the final judgment, especially for potential functional anomalies directly related to the interface state, such as the risk of short circuit due to electrolyte bridging or the sudden acceleration of corrosion due to the destruction of the protective layer, as an optional implementation, the input to a pre-trained risk discrimination model includes:

[0262] Obtaining the capacitance characteristic deviation caused by corrosion;

[0263] The capacitance characteristic deviation caused by corrosion is used as an additional input feature and is transmitted to the risk discrimination model together with the eigenvector and the identified representation of the current action mode;

[0264] The risk discrimination model is trained to jointly utilize the eigenvector, the representation of the current action mode, and the capacitance characteristic deviation to determine the malfunction risk level or state of the pressure switch.

[0265] In a specific implementation, the corrosion-induced capacitance characteristic deviation (ΔC_corr) determined above is first obtained. This deviation is the result of the real-time capacitance characteristic measurement after compensating for stress effects. Therefore, it more purely reflects the degree to which the capacitance characteristic deviates from its non-corrosion baseline due to the current corrosion process (including changes in the corrosion layer and interface dielectric).

[0266] Alternatively, in addition to using the deviation itself, its time rate of change can also be calculated and used. This rate of change better reflects the dynamics or activity of corrosion or interface state changes and may be more sensitive to sudden events. Therefore, the additional input feature here can be the capacitance characteristic deviation itself, its time rate of change, or a combination of the two.

[0267] Then, when inputting the information into the risk discrimination model, this acquired capacitance characteristic deviation or its rate of change is provided as an additional input feature, along with the already used eigenvector representing the deep structural state and the representation of the current action mode representing the operating condition. This means that the input feature vector of the risk discrimination model now includes at least three pieces of information: the overall damage state representation learned by the GNN, the current operating mode, and the real-time interface capacitance state indicator.

[0268] Accordingly, regardless of whether the risk discrimination model is an SVM, neural network, or other type of classifier or regressor, it needs to be trained to be able to jointly utilize these three types of input features to jointly determine the risk level or state of the pressure switch's misoperation. During the training phase, the model learns how to judge risk based on different combinations of these three types of information. For example, the model can learn that even if the structural damage indicated by the GNN feature vector is not serious, if the capacitance characteristic deviation shows an abnormal and rapid increase, and at this time is in a demanding operation mode, then it should also be judged as high risk.

[0269] This approach combines potential signatures reflecting long-term structural damage with direct physical quantities reflecting short-term interfacial changes, resulting in a more comprehensive assessment. This approach also provides greater sensitivity for detecting certain failure modes that initially manifest as abrupt changes in interfacial conditions. Even if the GNN model is temporarily insensitive to a particular change, direct capacitance deviation measurements may still capture the anomaly, and vice versa, thus improving the reliability of the overall assessment.

[0270] For example, at a certain monitoring moment, the eigenvectors generated by the graph neural network processing comprehensively reflect the presence of "medium" levels of accumulated structural damage in the metal components of the pressure switch. At the same time, the system identifies the currently requested action mode as "high-voltage pulse," a condition that places high demands on the equipment's state. At this moment, the calculated, stress-compensated capacitance characteristic deviation shows an abnormally rapid increase, which typically indicates a sudden change in the physical and chemical state of the metal component's surface or interface, such as a microcrack causing a sharp change in conductivity due to dielectric infiltration.

[0271] The input features fed into the risk identification model now include: an eigenvector representing "moderate" structural damage, an action pattern representing a "high-voltage pulse" operating condition, and an abnormally increased capacitance characteristic deviation reflecting a "mutated interface." Because the risk identification model is trained to understand the high potential risk inherent in this combination, it outputs a detection result that indicates a high risk or a direct indication of potential misoperation.

[0272] In contrast, without the capacitance characteristic deviation input, which directly reflects the real-time state of the interface, the risk discrimination model may judge the risk level to be low based primarily on the "medium" damage state given by the eigenvector, potentially missing early warnings of potential functional abnormalities caused by rapid changes in the interface. Therefore, incorporating the stress-compensated capacitance characteristic deviation as an additional feature into the final judgment can significantly improve the detection sensitivity for specific types of failure modes, especially those related to rapid changes in surface / interface states, making the final risk assessment more comprehensive, timely, and reliable.

[0273] Based on the same inventive concept, the embodiment of the present application also provides a pressure switch malfunction detection system corresponding to the pressure switch malfunction detection method. Since the principle of solving the problem by the system in the embodiment of the present application is similar to the above-mentioned pressure switch malfunction detection method in the embodiment of the present application, the implementation of the system can refer to the implementation of the method, and the repeated parts will not be repeated.

[0274] Reference Figure 4 FIG. 1 is a schematic diagram of a pressure switch malfunction detection system according to an embodiment of the present application, the system comprising:

[0275] The acquisition module 10 measures the capacitance characteristics between the sensing electrodes and the metal component of the pressure switch by using sensing electrodes provided on the surface of the metal component of the pressure switch; and generates corrosion status information based on the change of the capacitance characteristics relative to the reference state;

[0276] A first processing module 20 identifies and characterizes cracks in the metal component of the pressure switch based on the response of the metal component to the external electromagnetic detection field and the acoustic detection wave, and generates crack information; the crack information includes: location information, size information, depth information, and historical growth trend;

[0277] The second processing module 30 generates predictive crack growth information based on the corrosion state information and the crack information using a coupled model of corrosion and crack growth; and obtains dynamic stress information of the metal component of the pressure switch during operation through a stress sensor;

[0278] The detection module 40 determines whether the pressure switch has a risk of malfunction or has malfunctioned based on the dynamic stress information, the corrosion state information, the predictive crack growth information and the action mode required by the currently requested control logic, and generates a detection result.

[0279] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

Claims

1. A method for detecting malfunction of a pressure switch, characterized in that: include: By utilizing a sensing electrode disposed on a surface of a metal component of the pressure switch, measuring the capacitance characteristics between the sensing electrode and the metal component of the pressure switch; and generating corrosion state information based on a change in the capacitance characteristic relative to a baseline state; Identifying and characterizing cracks in the metal component of the pressure switch based on responses of the metal component of the pressure switch to an external electromagnetic detection field and an acoustic detection wave, and generating crack information; The crack information includes: location information, size information, depth information and historical expansion trend; Using a coupled model of corrosion and crack growth, predictive crack growth information is generated based on the corrosion state information and the crack information; dynamic stress information of the metal component of the pressure switch during operation is obtained through a stress sensor; Based on the dynamic stress information, the corrosion state information, the predictive crack growth information and the action mode required by the currently requested control logic, it is determined whether the pressure switch has a risk of malfunction or has malfunctioned, and a detection result is generated.

2. The pressure switch malfunction detection method according to claim 1, characterized in that: The generated crack information includes: At a first time point, applying a time-varying electromagnetic detection field to a predetermined monitoring area of ​​the pressure switch metal component to detect a corresponding electromagnetic response signal, and transmitting an acoustic energy detection wave to the pressure switch metal component to detect a corresponding acoustic echo signal; Calculating a first set of geometric feature information of the crack based on the electromagnetic response signal and the acoustic echo signal acquired at the first time point; the first set of geometric feature information includes: crack position information, size information, and depth information at the first time point; At a second time point, the process of acquiring the corresponding electromagnetic response signal and the acoustic echo signal and calculating the geometric feature information based thereon is again performed to obtain a second set of geometric feature information of the crack; the second set of geometric feature information comprising: crack location information, size information, and depth information at the second time point; Comparing the second set of geometric feature information with the first set of geometric feature information, and calculating a change amount and a change direction of the position, size, and depth of the crack during a time interval between the first time point and the second time point; Determining a historical crack growth trend based on the change amount and change direction; The second set of geometric feature information is combined with the historical expansion trend to generate the crack information.

3. The pressure switch malfunction detection method according to claim 1, characterized in that: Generating corrosion status information includes: accessing a preset stress-capacitance response relationship, the relationship representing a capacitance characteristic of a metal component of the pressure switch in a known corrosion state versus an applied stress; Calculating an expected corrosion-free capacitance characteristic reference value corresponding to the real-time dynamic stress information using the stress-capacitance response relationship and the dynamic stress information acquired in real time; comparing the capacitance characteristic measured in real time with a calculated baseline value of the expected capacitance characteristic without corrosion to determine an amount of capacitance characteristic deviation caused by corrosion; Based on the capacitance characteristic deviation, a preset correlation model or calibration data for correlating the capacitance deviation to the corrosion state is applied to generate the corrosion state information.

4. The pressure switch malfunction detection method according to claim 1, characterized in that: Measuring the capacitance characteristic between the sensing electrode and the metal component of the pressure switch includes: Applying an AC excitation signal with a predetermined frequency between the sensing electrode and the metal component of the pressure switch; measuring an electrical response parameter generated between the sensing electrode and the metal component of the pressure switch in response to the AC excitation signal, wherein the electrical response parameter is selected from at least one of an AC response current, an AC response voltage, or a complex impedance; The capacitance characteristic is calculated based on the measured electrical response parameter and the predetermined frequency.

5. The pressure switch malfunction detection method according to claim 2, characterized in that: Using the coupled corrosion and crack growth model, predictive crack growth information is generated, including: mapping the corrosion state information and the crack information into damage states of multiple spatial regions on the metal component of the pressure switch; In the coupling model, establishing and utilizing a mutual influence relationship between the damage states of the plurality of spatial regions; Calculating predictive crack growth information for each spatial region based on the mutual influence relationship of the damage states and the current damage state of each region; The mutual influence relationship of the damage states at least reflects one or more of the following effects: the change of the local stress distribution of other regions due to damage in one region, and the electrochemical coupling effect between different regions.

6. The pressure switch malfunction detection method according to claim 5, characterized in that: The establishing and utilizing the mutual influence relationship of damage states among the multiple spatial regions includes: Abstracting the multiple spatial regions into nodes of a graph neural network, and constructing a node feature vector for each node; wherein the node feature vector includes: the corrosion state information, the crack information, the dynamic stress information, and the capacitance characteristic; Determine the edge connection and edge feature vector between the nodes based on whether the physical proximity relationship between the nodes, the physical field gradient between the nodes, and the state parameter difference between the nodes meet preset conditions; Calling a message passing graph neural network pre-trained based on historical prototype data, inputting the node feature vector and the edge feature vector into the message passing graph neural network, and outputting a crack propagation driving force index of the corresponding node; The crack growth driving force index is used in the coupling model to calculate the predictive crack growth information of each spatial region.

7. The pressure switch malfunction detection method according to claim 6, characterized in that: The establishing and utilizing the mutual influence relationship of damage states among the multiple spatial regions further includes: monitoring the predicted crack growth information generated by the coupling model, comparing it with the first set of geometric feature information and the second set of geometric feature information acquired at different subsequent time points, and determining a deviation between actual crack geometry changes; In response to the deviation exceeding a preset threshold, triggering an incremental learning process; The incremental learning process uses a monitoring data set to perform online adjustment on the weights of the graph neural network; the monitoring data set is associated with a monitoring period corresponding to the deviation, and the monitoring data set includes input information obtained during the monitoring period for generating the node feature vector and the edge feature vector, as well as the corresponding actual crack geometry changes.

8. The pressure switch malfunction detection method according to claim 7, characterized in that: Determining whether the pressure switch has a risk of malfunction or has malfunctioned includes: Extracting or aggregating the final state representations of the nodes generated after processing by the message passing graph neural network to generate a feature vector representing the overall damage state of the metal component of the pressure switch; Identify the action mode required by the control logic of the current request; Inputting the feature vector and the representation of the identified current action mode into a pre-trained risk discrimination model; The detection result is generated using the output result of the risk discrimination model.

9. The pressure switch malfunction detection method according to claim 8, characterized in that: The input to a pre-trained risk discrimination model includes: Obtaining the amount of capacitance characteristic deviation caused by corrosion; The capacitance characteristic deviation caused by corrosion is used as an additional input feature and is transmitted to the risk discrimination model together with the eigenvector and the identified representation of the current action mode; The risk discrimination model is trained to jointly utilize the eigenvector, the representation of the current action mode, and the capacitance characteristic deviation to determine the malfunction risk level or state of the pressure switch.

10. Pressure switch malfunction detection system, characterized in that: include: an acquisition module, which measures the capacitance characteristics between the sensing electrode and the metal component of the pressure switch by utilizing the sensing electrode provided on the surface of the metal component of the pressure switch; and generating corrosion state information based on a change in the capacitance characteristic relative to a baseline state; a first processing module, for identifying and characterizing cracks in the metal component of the pressure switch according to responses of the metal component of the pressure switch to an external electromagnetic detection field and an acoustic detection wave, and generating crack information; The crack information includes: location information, size information, depth information and historical expansion trend; The second processing module generates predictive crack growth information based on the corrosion state information and the crack information using a coupled model of corrosion and crack growth; and obtains dynamic stress information of the metal component of the pressure switch during operation through a stress sensor; The detection module determines whether the pressure switch has a risk of malfunction or has malfunctioned based on the dynamic stress information, the corrosion state information, the predictive crack growth information, and the action mode required by the currently requested control logic, and generates a detection result.

Images