Dry gas seal operating parameter overrun identification processing method, system, medium and product

Abstract

A kind of dry gas seal operating parameter over-limit identification processing method, system, medium and product, relate to parameter identification field.The method comprises: presetting safe operation interval for each monitoring parameter, and setting reasonable fluctuation interval with interval boundary narrower than safe operation interval inside safe operation interval;Real-time acquisition of the acquisition value of each monitoring parameter, when the acquisition value falls in the transition zone between safe operation interval and reasonable fluctuation interval, the current acquisition point and corresponding time sequence fragment are marked as suspected over-limit sample;The authenticity diagnosis of suspected over-limit sample is carried out by preset diagnostic model;According to the diagnosis result, corresponding operation is executed.The above-mentioned technical scheme can accurately identify dry gas seal operating parameter over-limit condition.

Description

Technical Field

[0001] This application relates to the field of parameter identification, and in particular to a method, system, medium, and product for identifying and processing excessive operating parameters of dry gas seals. Background Technology

[0002] In industrial production, the stable operation of equipment is crucial, especially for critical components such as dry gas seals. Dry gas seals are widely used in various rotating machinery, such as compressors and pumps, and their operating status directly affects the reliability and safety of the entire equipment and even the production system.

[0003] Previously, the monitoring and processing of dry gas seal operating parameters typically employed a simple threshold judgment method. This involved pre-setting fixed upper and lower limits; when the collected parameter values ​​exceeded these ranges, an equipment malfunction was identified and an alarm was triggered. This method was relatively simple to operate and easy to implement. Additionally, some methods performed simple statistical analyses on the collected data, such as calculating the average and standard deviation, to assess the stability and rationality of the data. Some companies also conducted regular manual inspections of the equipment, judging its operating status by observing its appearance and operating sounds.

[0004] However, these traditional methods have significant drawbacks. Simple threshold judgment methods are too rigid, unable to distinguish whether anomalies in collected values ​​are due to genuine changes in operating conditions or distortions during data transmission, easily leading to false alarms or missed alarms. Statistical analysis methods can only assess data macroscopically, making it difficult to accurately identify anomalies at individual data points. Manual inspections, on the other hand, suffer from strong subjectivity and low efficiency, failing to provide real-time and comprehensive understanding of equipment operating status. These shortcomings make traditional methods ill-suited to meeting the demands of modern industry for precise monitoring and effective processing of dry gas seal operating parameters. Summary of the Invention

[0005] This application provides a method, system, medium, and product for identifying and processing excessive operating parameters of dry gas seals. It can accurately identify situations where operating parameters of dry gas seals exceed limits, distinguish between changes in actual operating conditions and data transmission distortion, effectively process distorted signals, and ensure stable operation of equipment.

[0006] In a first aspect, this application provides a method for identifying and processing excessive operating parameters of a dry gas seal, the method comprising: For each monitoring parameter, a first upper limit value and a first lower limit value are preset. A safe operating range is determined based on the first upper limit value and the first lower limit value. Within the safe operating range, a reasonable fluctuation range with a range boundary narrower than the safe operating range is set. The second upper limit value of the reasonable fluctuation range is less than the first upper limit value, and the second lower limit value of the reasonable fluctuation range is greater than the first lower limit value. The system acquires the collected values ​​of each monitoring parameter in real time. When the collected value exceeds the safe operating range, it is determined to be an equipment malfunction and an alarm is triggered. When the collected value falls within the reasonable fluctuation range, it is determined to be a normal operating condition. When the collected value falls into the transition zone between the safe operating range and the reasonable fluctuation range, the current collection point and the corresponding time series segment are marked as suspected out-of-limit samples. The suspected out-of-limit samples are subjected to a real-world diagnostic test using a pre-defined diagnostic model to determine whether they are a precursor to changes in actual operating conditions or distortions in the data transmission link. If the diagnosis is a precursor to a change in actual operating conditions, the original data is retained and an early warning of the corresponding level is triggered. If the diagnosis is a distortion in the data transmission link, the distorted signal is repaired or removed, and the processed data is recorded in the database.

[0007] By adopting the above technical solution, a transition zone is set between the safe operating range and the reasonable fluctuation range, directly decoupling over-limit alarms and suspected fluctuation diagnosis in space. This avoids directly misjudging link distortion as equipment anomaly, effectively reducing the system's false alarm rate. The introduction of a genuine diagnostic mechanism for suspected over-limit samples enables the differentiation between precursors to actual operating conditions and transmission link distortion from numerical out-of-bounds phenomena, solving the problem of data trust crisis caused by the inability of traditional methods to identify the source of anomalies. Differential handling is implemented based on the diagnostic results: precursors to actual operating conditions are retained and warned against, while link distortion is repaired or removed. This achieves self-purification of monitoring data before storage, ensuring that early fault signals are not incorrectly filtered, thus balancing data quality and operational safety.

[0008] In some embodiments, the step of performing a real-world diagnosis on the suspected out-of-limit sample using a preset diagnostic model includes: In the multidimensional state space of dry gas seal operation, the real-time sequences of three parameters, namely leakage, pressure and temperature, are extracted to construct the operating attractor trajectory in high-dimensional phase space. Calculate the local manifold radius of curvature and the maximum Lyapunov exponent of the running attractor trajectory within a preset sliding time window; Compare the current local manifold radius of curvature and maximum Lyapunov exponent with the topological deviation of the attractor baseline model established under historical normal operating conditions. If the local topology of the running attractor trajectory undergoes an anomaly exceeding a preset sensitivity threshold, it is determined to be a precursor to a change in the actual operating conditions; if there are only point outliers and the attractor topology remains unchanged, it is determined to be a distortion in the data transmission link.

[0009] By employing the aforementioned technical solution, and mapping multiple parameters to a high-dimensional phase space and analyzing the local manifold distortion of the attractor topology, microscopic instability in system dynamics can be captured even before parameter values ​​significantly deviate from thresholds. This overcomes the detection blind spot where parameter changes are subtle and difficult to identify in the early stages of a fault. Using the local manifold curvature radius and Lyapunov exponent as metrics for topological singularity, the system distinguishes between multi-parameter coupled evolution and isolated outliers at the level of overall system dynamics, achieving essential identification of precursors to real-world operating conditions and data link distortions. Based on topological deviation comparison using the attractor baseline model, it is naturally immune to non-persistent disturbances such as sensor transient jitter and instantaneous electromagnetic interference, responding only to persistent evolutions that alter the system's dynamic structure, significantly reducing the probability of misdiagnosis caused by data link noise.

[0010] In some embodiments, calculating the local manifold radius of curvature and the maximum Lyapunov exponent of the running attractor trajectory within a preset sliding time window specifically includes: Multivariable phase space reconstruction is performed on the real-time sequences of the leakage amount, the pressure, and the temperature to determine the optimal embedding delay and embedding dimension for each dimension, thereby obtaining a multivariable delayed embedding vector sequence. Within each sliding time window, weighted principal component analysis is performed on the local embedded vector point set to extract the eigenvalue spectrum, and the local manifold radius of curvature is defined by the ratio of the largest eigenvalue to the second largest eigenvalue. Within the sliding time window, the evolution distance between each reference point and its nearest neighbor along the trajectory of the running attractor is tracked, and the logarithmic growth rate of the evolution distance is linearly fitted to obtain the maximum Lyapunov exponent.

[0011] By employing the aforementioned technical solution, through multivariate phase space reconstruction and optimal embedding parameter determination, the abstract attractor topological distortion is transformed into calculable local manifold curvature radius and maximum Lyapunov exponent. This provides accurate numerical criteria for subsequent topological deviation comparison, solving the problem of the difficulty in implementing chaotic dynamics analysis in engineering monitoring. The local manifold curvature radius is defined by the eigenvalue ratio of weighted principal component analysis, capturing the instantaneous bending degree of the attractor in phase space. The maximum Lyapunov exponent is extracted by the logarithmic growth rate of the nearest neighbor evolution distance, quantifying the system's divergence rate to small perturbations. These two methods, from the complementary dimensions of static local morphology and dynamic evolution trend, comprehensively characterize the dynamic behavior of the attractor trajectory. By performing weighted principal component analysis on the local embedded vector point set to suppress outlier noise, and by using linear fitting of the logarithmic growth rate within a sliding window to smooth instantaneous fluctuations, the calculation of the curvature radius and Lyapunov exponent possesses natural resistance to sensor noise and sporadic distortions in the data link, ensuring the reliability of the diagnostic indicators themselves.

[0012] In some embodiments, the step of comparing the local manifold radius of curvature and the maximum Lyapunov exponent at the current moment with the attractor baseline model established under historical normal operating conditions specifically includes: Using long-term operating data accumulated under the health status of the equipment, a multidimensional dynamic benchmark model under normal operating conditions is pre-constructed. The benchmark model includes a set of reference topological descriptors for normal attractors, and the topological descriptors include barcodes or persistent graphs generated by continuous cohomology. For the local manifold radius of curvature and the maximum Lyapunov exponent calculated within the current sliding window, a real-time topological descriptor for the current attractor segment is constructed, and topological features are extracted using a continuous homology algorithm. The topology deviation value between the real-time topology descriptor and the reference topology descriptor set in the baseline model is quantified using Wasserstein distance or Bottleneck distance, and the topology deviation value is compared with the preset sensitivity threshold.

[0013] By employing the above technical solution, barcodes or persistent graphs generated through continuous coherence are used as topological descriptors. This elevates the comparison of single-index numerical thresholds in traditional methods to a comparison of the overall dynamic topology of the system. This allows the diagnostic model to perceive structural degradation hidden behind numerical changes, fundamentally improving the depth and accuracy of anomaly identification. A benchmark model, constructed based on long-term health data and incorporating a set of reference topological descriptors, can cover various dynamic manifestations under normal operating conditions, rather than relying solely on a single threshold. This gives the diagnostic model a more tolerant generalization ability to reasonable fluctuations during normal operation, reducing false triggers caused by normal switching of operating conditions. The Wasserstein distance or Bottleneck distance is used to measure the topological deviation between the real-time topological descriptor and the benchmark model. This transforms abstract topological spatial differences into single, continuous, and comparable deviation values, achieving engineering-level quantification of the degree of topological change. This provides a stable and interpretable numerical basis for setting and triggering preset sensitivity thresholds.

[0014] In some embodiments, the data repair or removal of distorted signals includes: A generative adversarial network model consisting of a generator and a discriminator is constructed and pre-trained. The generator learns the mapping from a masked multi-parameter incomplete signal to a complete and real sealed state signal. When the current signal segment is determined to be a distorted region, a binary masking operation is performed on the distorted data points in the current signal segment, and the multi-channel parameter matrix carrying the mask is input into the generator; The generator generates reconstructed data point by point under the continuous feature constraints of the running parameters and the three-dimensional spatiotemporal constraints of the associated parameters. The discriminator provides the generator with an adversarial gradient to enhance the reconstruction details. The original distorted fragments are replaced with the reconstructed data generated by the generator, and the confidence labels of the repaired data records are quantified.

[0015] By employing the aforementioned technical solution, the generator generates reconstructed data under multi-parameter continuous feature constraints and three-dimensional spatiotemporal constraints, rather than simple smoothing or interpolation. This enables the restoration of complex correlation features that should exist in distorted segments, ensuring that the repaired data matches the actual physical response of the device and avoiding the introduction of secondary distortions. Through the adversarial gradient provided by the discriminator, the generator can recover the subtle signal structure obscured by the mask, effectively overcoming the shortcomings of traditional repair methods in smoothing out abrupt changes and high-frequency details. This allows the reconstructed signal to approximate the actual sealed state signal in both statistical distribution and local morphology. A quantified confidence label is attached to the repaired data, enabling downstream analysis and decision-making modules to perceive that the data point was generated by the model rather than directly collected. This allows for differentiated weighting when using the repaired data, ensuring transparency and auditability throughout the entire lifecycle of the monitoring data.

[0016] In some embodiments, the step of generating reconstructed data point by point through the generator under the continuous feature constraints of the running parameters and the three-dimensional spatiotemporal constraints of the associated parameters specifically includes: The generator's dual-channel feature extraction encoder processes the multi-channel parameter matrix. The first channel of the dual-channel feature extraction encoder performs autocorrelation function morphology extraction and wavelet scattering transformation on each single-channel time series in the multi-channel parameter matrix, and outputs continuous feature constraint vectors for each channel. The second channel calculates the mutual information delay spectrum and dynamic conditional entropy transfer matrix between channels in the multi-channel parameter matrix, and outputs a cross-channel three-dimensional spatiotemporal constraint matrix. At each scale level of the generator, the temporal consistency loss between the hierarchical reconstructed feature map and the continuous feature constraint vector, and the linkage physical consistency loss between the cross-channel reconstructed feature map and the three-dimensional spatiotemporal constraint matrix are calculated, and the losses at each level are weighted and accumulated into the total loss function of the generator. Iterative adversarial training is performed based on the total loss function to generate reconstructed data point by point.

[0017] Using the above technical solution, the first channel captures the continuous feature constraints of each parameter through autocorrelation function morphology extraction and wavelet scattering transform. The second channel constructs a cross-channel three-dimensional spatiotemporal constraint matrix through mutual information delay spectrum and dynamic conditional entropy transfer matrix, realizing the structured separation and collaborative representation of the intrinsic dynamics of a single channel and the coupling relationship of multiple channels, providing a complete physical prior for reconstruction generation. At each scale level of the generator, the temporal consistency loss between the reconstructed feature map and the continuous feature constraint vector, as well as the linkage physical consistency loss between the cross-channel reconstructed feature map and the three-dimensional spatiotemporal constraint matrix, are calculated simultaneously. The physical constraints are permeated layer by layer to each resolution level of the reconstruction, ensuring that the physical mechanism of device operation is strictly met from macroscopic trends to microscopic details. The temporal consistency loss and the linkage physical consistency loss are weighted and accumulated into the total loss function, and iterative training is performed under the adversarial gradient feedback of the discriminator. This allows the generator to pursue realistic data distribution while forcibly obeying the temporal evolution law of the parameters themselves and the correlation constraints between parameters. The generated reconstructed data has both statistical realism and physical interpretability.

[0018] In some embodiments, before marking the current collection point and its corresponding time series segment as a suspected out-of-limit sample, the method further includes: The signal quality is evaluated on the real-time acquired values. The signal quality evaluation includes signal-to-noise ratio detection, data frame integrity verification, and reading of sensor health status indicators, and outputs a quantitative reliability score. When the quantitative confidence score is lower than the preset minimum confidence threshold, the current collection point is directly marked as a distortion in the data transmission link; When the quantitative confidence score is higher than or equal to the minimum confidence threshold and lower than the preset high confidence threshold, the current collection point is marked as a suspected out-of-limit sample, and the quantitative confidence score is added as a prior input weight for the diagnostic model. When the quantitative confidence score is lower than the minimum confidence threshold for a consecutive preset number of collection points, and the collection values ​​all fall within the transition zone, it is determined that the sensor or the collection link has a continuous failure, triggering a sensor maintenance warning.

[0019] By employing the above technical solution, a three-dimensional quality assessment is introduced before marking suspected out-of-limit samples: signal-to-noise ratio detection, data frame integrity verification, and sensor health status identification. Low-confidence signals are directly intercepted and marked as distortion, preventing the ineffective consumption of diagnostic model computational resources by low-quality data at the source and reducing the frequency of misdiagnosis. For acquisition points in the medium-confidence range, the quantified confidence score is used as a priori input weight for the diagnostic model, enabling the model to adaptively adjust the judgment confidence level based on the signal's inherent quality level. This avoids misjudgments caused by applying the same decision criteria to low-quality and high-quality signals, improving the reliability of diagnostic results. By monitoring the duration and spatial distribution of quantified confidence scores continuously falling below the minimum confidence threshold, maintenance warnings are triggered early in the progressive failure of sensors or acquisition links, shifting the maintenance mode from passive response to proactive predictive maintenance and preventing long-term contamination of monitoring data due to undetected chronic sensor drift.

[0020] In a second aspect, embodiments of this application provide a computer system including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the steps of the method described in any possible implementation of the first aspect.

[0021] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method described in any possible implementation of the first aspect.

[0022] Fourthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in any possible implementation of the first aspect.

[0023] It is understood that the computer system provided in the second aspect, the storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the method provided in this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. By simultaneously setting a safe operating range and a reasonable fluctuation range within each monitoring parameter, a transition zone is constructed. This architecture breaks down the traditional single threshold judgment into a three-level classification: absolute danger outside the safe operating range, absolute normality within the reasonable fluctuation range, and a suspected state in between. This hierarchical mechanism prevents the system from treating any data exceeding a single boundary as an alarm signal. Instead, it categorizes data that only exceeds the reasonable fluctuation range but does not reach the safety boundary into a gray area requiring further screening. This cuts off the path of occasional distortion in the data link directly triggering abnormal alarms at the source of the judgment logic, effectively reducing the false alarm rate introduced by the signal transmission link. 2. By using a pre-set diagnostic model to diagnose the authenticity of suspected out-of-limit samples, the system adds an intelligent identification step before triggering final action. This diagnostic model does not rely on fixed threshold rules for simple classification, but rather learns the deep differences between normal fluctuations and true anomalies in a multi-dimensional feature space, thereby making a probabilistic judgment on the source of suspected samples. This mechanism solves the trust dilemma in traditional monitoring systems where "data is questioned or accepted as soon as it crosses the line," enabling the system to provide well-founded causal conclusions for each out-of-line event, significantly improving the credibility and transparency of the monitoring data chain. 3. Data confirmed as precursors to actual changes in operating conditions is retained and triggers tiered early warnings to ensure that early fault signals are not overlooked, thus gaining response time for maintenance decisions. Data confirmed as link distortion is repaired or removed before being stored in the database, ensuring the cleanliness of the historical database and preventing inferior data from contaminating subsequent advanced applications such as trend analysis and lifespan prediction. This differentiated handling strategy achieves a balance between data quality assurance and equipment operational safety. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a method for identifying and processing excessive operating parameters of dry gas seals in an embodiment of this application. Figure 2 This is another flowchart illustrating the method for identifying and processing excessive operating parameters of dry gas seals in this application. Figure 3 This is a schematic diagram of the safe operating range and reasonable fluctuation range in the embodiments of this application; Figure 4 This is a schematic diagram of an exemplary hardware structure of a computer system in an embodiment of this application. Detailed Implementation

[0026] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0027] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0028] The following is combined Figure 1 The method of the embodiments of this application will be described below.

[0029] Please see Figure 1 This is a flowchart illustrating a method for identifying and processing excessive operating parameters of a dry gas seal in an embodiment of this application. Figure 1 As shown, the method includes the following steps: S101. For each monitoring parameter, a first upper limit value and a first lower limit value are preset. A safe operating range is determined based on the first upper limit value and the first lower limit value. Within the safe operating range, a reasonable fluctuation range with a range boundary narrower than the safe operating range is set. The second upper limit value of the reasonable fluctuation range is less than the first upper limit value, and the second lower limit value of the reasonable fluctuation range is greater than the first lower limit value. S102. Real-time acquisition of the collected value of each monitoring parameter. When the collected value exceeds the safe operating range, it is determined to be an equipment abnormality and an alarm is triggered. When the collected value falls within the reasonable fluctuation range, it is determined to be a normal operating condition. When the collected value falls into the transition zone between the safe operating range and the reasonable fluctuation range, the current collection point and the corresponding time series segment are marked as suspected over-limit samples. S103. Perform a realism diagnosis on the suspected out-of-limit sample using a preset diagnostic model to determine whether the suspected out-of-limit sample is a precursor to changes in actual operating conditions or a distortion in the data transmission link. S104. If the diagnosis is a precursor to a change in actual operating conditions, the original data is retained and a warning of the corresponding level is triggered. If the diagnosis is a distortion in the data transmission link, the distorted signal is repaired or removed, and the processed data is recorded in the database.

[0030] The safe operation interval is the parameter boundary within which the dry gas seal can operate continuously and reliably, defined by a first upper limit value and a first lower limit value. The setting of these two threshold values is not arbitrary, but is determined by comprehensively considering factors such as seal type and design parameters, operating pressure and working conditions, historical operation data, relevant standards and engineering specifications. Seal type and design parameters: Different types of dry gas seals (such as single-face, double-face, tandem, with intermediate labyrinth seal, etc.) have different tolerance capacities for leakage, pressure and temperature due to structural differences, and design documents usually contain the rated operating range. Operating pressure and working conditions: The operating conditions such as process medium pressure, rotation speed and temperature of the main unit where the seal is located (such as a turbocompressor) directly determine the safety boundary of the parameters. Historical operation data: Through statistical analysis of long-term operation data under the healthy state of the equipment, the parameter distribution range under normal working conditions is extracted as the empirical basis for interval setting. Relevant standards and engineering specifications: Recommended setting guidelines for alarm values of dry gas seal monitoring parameters in standards such as API 617 and API 692. The determined safe operation interval is denoted as [L, H], where L is the first lower limit value and H is the first upper limit value. When the collected value of the monitoring parameter exceeds this interval, it means that the dry gas seal is in or close to the critical state of failure, and an alarm needs to be triggered immediately. The reasonable fluctuation interval is a narrower operation window defined inside the safe operation interval, whose second upper limit value (H2) is less than the first upper limit value, and the second lower limit value (L2) is greater than the first lower limit value (i.e. L<L2<H2<H). The width of this interval usually accounts for 10% to 30% of the width of the safe operation interval, with a default value of 20%. The core logic for setting this interval lies in: during the normal operation of the equipment, the monitoring parameters are not constant, but present a certain range of normal fluctuations due to factors such as slight working condition adjustments, environmental changes, and measurement noise. This part of the fluctuation is numerically reflected as a small-amplitude oscillation around a certain steady-state center, which belongs to the normal dynamic behavior of the system and should not be regarded as an abnormality. The setting methods of the reasonable fluctuation interval may include: (1) Based on the median value of the safe operation interval: take C=(H+L) / 2 as the center, and extend to both sides with bw=(H-L)×r as the width. (2) Based on the historical steady-state mean value: through statistical analysis of normal operation historical data, take the steady-state mean value as the center, and take several multiples of the standard deviation as the half-width. (3) Adaptive adjustment: dynamically adjust according to different working condition stages to realize differentiated settings under different operation modes. The annular area inside the safe operation interval and outside the reasonable fluctuation interval is the transition zone. This area is the core innovation of the dual-interval architecture in the embodiments of the present application — it transforms the all-or-nothing judgment logic in the conventional solution into a ternary classification: absolutely normal (within the reasonable fluctuation interval), absolutely dangerous (outside the safe operation interval), and suspected state (within the transition zone). The data falling into the transition zone neither requires immediate shutdown like serious over-limit, nor can it be safely ignored like normal fluctuation, but needs to start a subsequent diagnosis mechanism for deep traceability.The system acquires the collected values of each monitoring parameter from the sensor at a fixed time interval T. The collected parameter types include but are not limited to: leakage amount, such as the primary leakage, secondary leakage or total leakage of the dry gas seal; pressure, such as seal cavity pressure, isolation gas pressure, buffer gas pressure, etc.; temperature, such as seal end face temperature, seal cavity temperature, leaked gas temperature, etc.; flow rate, such as supply gas flow rate, exhaust gas flow rate, etc. The collection frequency is set according to the seal type and application scenario. In high-frequency collection scenarios, T can be at the millisecond level, and in conventional scenarios, it is at the second level or minute level. The following judgments are performed on each collected value: First level - safety over-limit judgment: When the collected value vi exceeds the safe operation interval (vi>H or vi<L), it is determined that the equipment is abnormal. At this time, the parameter has crossed the safety boundary, the dry gas seal may have suffered a serious failure or be at risk of failure, and the system immediately triggers an alarm mechanism to notify operation and maintenance personnel to take emergency measures. Second level - normal judgment: When the collected value vi falls within the reasonable fluctuation interval (L2≤vi≤H2), it is determined that the working condition is normal. This collected value is regarded as credible normal operation data, and is directly recorded into the database for subsequent applications such as trend analysis and life prediction. Third level - suspected judgment: When the collected value vi falls into the transition zone (L<vi<L2 or H2<vi<H), the system does not immediately make a final judgment of normal or abnormal, but marks the current collection point and the corresponding time series segment as a suspected over-limit sample. The time series segment here does not only refer to an isolated collection point, but includes the point and the continuous sampling values within a period of time window before and after it, so as to provide sufficient time series context information for the subsequent diagnosis model. The diagnosis model does not start calculation for every collection point, but is triggered only when the collected value is marked as a suspected over-limit sample. This mechanism avoids consuming unnecessary computing resources during the vast majority of normal operation periods. The core task of the diagnosis model is to answer a key question: Is the currently observed value out of bounds a real deterioration of the physical state of the dry gas seal (a precursor of real working condition change), or a false signal caused by link factors such as sensor aging drift, data conversion quantization error, encoding and decoding error, and electromagnetic interference (distortion in the data transmission link)? The diagnosis model can realize the above judgment from multiple dimensions and by adopting multiple technical routes: (1) Data continuity verification: Perform trend analysis on multiple consecutive collection points to judge whether the over-limit is a continuous unidirectional change or an isolated jump. Continuous change tends to be real working condition evolution, and isolated jump tends to be data distortion. (2) Multi-parameter linkage verification: Check whether other parameters with physical correlation synchronously change in a manner consistent with the mechanism. For example, a real increase in leakage should be accompanied by a corresponding change in the pressure of the seal cavity; if there is no synchronous response of associated parameters, it is more likely to be a single-point data distortion. (3) High-dimensional phase space analysis: Map the multi-parameter time series to a high-dimensional phase space, and judge whether the dynamic behavior of the system has an essential change by analyzing the change of the attractor topological structure.The attractor under normal operating conditions has a stable topological structure. The real fault precursor will cause the local manifold of the attractor to be distorted, while the link noise only manifests as point outliers and will not change the overall topology of the attractor. (4) Statistical pattern recognition: Extract the multidimensional features of the suspected sample and use a trained classifier (such as support vector machine, random forest, etc.) to perform binary classification. (5) Deep learning discrimination: Use deep models such as convolutional neural network and recurrent neural network to learn the discrimination boundary between normal fluctuation and real anomaly from the original time series segment end to end. The diagnostic model outputs the classification conclusion of the suspected over-limit sample, usually with the confidence value of the truth to quantify the degree of certainty of the judgment. When the diagnostic model determines that the suspected over-limit sample is the precursor of the real operating condition change, it performs the following operations: (1) Original data retention: The collected value and the associated time series segment are completely retained without any modification or filtering to ensure that the data information generated by the real physical process is not lost; (2) Graded warning triggering: According to the confidence value of the truth or the degree of parameter deviation, the corresponding level of warning is triggered. The warning levels can be divided into: prompt level: the parameter deviates slightly and is in the early stage. The operation and maintenance personnel should be notified to pay attention and no immediate intervention is required; attention level: the parameter continues to deviate and shows a deteriorating trend. It is recommended to arrange an inspection plan; warning level: the parameter is close to the boundary of the safe operation range and control or shutdown measures should be taken immediately. (3) Data marking: the data is specially marked to indicate that it is diagnosed as a precursor of the real working condition, which is convenient for subsequent fault tracking and root cause analysis. When the diagnostic model determines that the suspected over-limit sample is a link distortion, the following operations are performed: (1) Data repair or removal: the distorted data points are processed. Repair strategies include: removal method: directly delete the distorted data points, which is suitable for isolated distortion and high sampling density scenarios. neighbor weighted interpolation method: fill with the weighted average of normal points before and after the distorted point, which is suitable for isolated distorted points. generative reconstruction method: use the physical constraints and temporal characteristics of the associated parameters to reconstruct the real signal of the distorted segment through the generative model, which is suitable for scenarios where multiple consecutive sampling points are distorted. smoothing method: such as moving average, exponential smoothing, etc., which is suitable for small random jitter. (2) Data Recording and Storage: The processed data (repaired data or null value markers after removal) is recorded and stored in the database. For the repaired data, a quantitative confidence label can be attached to indicate that the data point was reconstructed by the model rather than directly collected, for downstream applications to perform differentiated weighting. The choice of repair strategy can be adaptively determined according to the density and distribution pattern of the distorted data: for isolated distorted points, interpolation or removal methods with lower computational overhead are used; for continuous distorted segments, generative reconstruction methods with higher accuracy are used to achieve the optimal balance between repair quality and system resource consumption.

[0031] Figure 2 This is another flowchart illustrating the method for identifying and processing excessive operating parameters of dry gas seals in this application, as shown in the embodiments. Figure 2As shown, a safe operating range and a reasonable fluctuation range are defined. Real-time data acquisition and over-limit judgment are performed. If the data exceeds the safe operating range, an alarm is triggered. If the data is within the reasonable fluctuation range, the data is stored in the database. If the data is within the safe operating range but exceeds the reasonable fluctuation range, over-limit identification processing is performed. Finally, the model start-up conditions are determined, and the handling results are output.

[0032] Figure 3 This is a schematic diagram of the safe operating range and reasonable fluctuation range in the embodiments of this application, as shown below. Figure 3 As shown, the positional relationship between the high alarm value H, the low alarm value L, and the baseline value C is displayed. The safe operating range a = [L, H], the width of the safe operating range w = HL, the ratio of the reasonable fluctuation range width r (default 20%), and the reasonable fluctuation range width bw = w × r are all shown.

[0033] In some embodiments, the step of performing a real-world diagnosis on the suspected out-of-limit sample using a preset diagnostic model includes: In the multidimensional state space of dry gas seal operation, the real-time sequences of three parameters, namely leakage, pressure and temperature, are extracted to construct the operating attractor trajectory in high-dimensional phase space. Calculate the local manifold radius of curvature and the maximum Lyapunov exponent of the running attractor trajectory within a preset sliding time window; Compare the current local manifold radius of curvature and maximum Lyapunov exponent with the topological deviation of the attractor baseline model established under historical normal operating conditions. If the local topology of the running attractor trajectory undergoes an anomaly exceeding a preset sensitivity threshold, it is determined to be a precursor to a change in the actual operating conditions; if there are only point outliers and the attractor topology remains unchanged, it is determined to be a distortion in the data transmission link.

[0034] The operating state of a dry gas seal cannot be fully described by a single parameter. The magnitude of leakage is influenced by a combination of factors, including the gap between the sealing end faces, the sealing gas pressure, and the temperature field distribution; conversely, changes in leakage will also cause dynamic responses in system pressure and temperature. These three parameters are physically coupled, and together they constitute the set of fundamental variables describing the operating state of the dry gas seal. This application selects three core parameters—leakage, pressure, and temperature—and maps their respective real-time acquisition sequences to a three-dimensional state space (which can be extended to higher dimensions if necessary using phase space reconstruction technology). In this state space, the three measurements at each sampling moment constitute a state point, and connecting consecutive sampling points forms a trajectory—this is the operating attractor trajectory of the dry gas seal system. An attractor is a stable set that a dynamic system tends to after long-term evolution in phase space. Under the health condition of the equipment, the operating state of the dry gas seal fluctuates around a certain steady-state region in phase space, forming an attractor with stable geometric structure and topological characteristics. This attractor characterizes the inherent dynamic characteristics of the system, including damping characteristics, stiffness characteristics, time delay relationships between parameters, and energy dissipation paths. When a device exhibits early signs of a real malfunction (such as end-face wear, film instability, or sealing surface deformation), the system's dynamic structure changes, and the local geometry of the attractor becomes distorted (e.g., increased local bending, altered trajectory winding, or changes in convergence rate). However, when only isolated noise points exist in the data transmission link, the corresponding state point may jump away from the attractor, forming a sudden point-like outlier. These isolated noise points do not have a lasting impact on the overall structure of the attractor. The local manifold radius of curvature is a geometric quantity that measures the degree of local bending of the attractor in phase space. Under normal operating conditions, the attractor's manifold is relatively smooth with a large radius of curvature. When the system's dynamic characteristics change, the attractor's local manifold may experience increased bending, manifested as a decrease in the radius of curvature. By calculating the spatial distribution of the locally embedded vector point set within a sliding time window, extracting its principal component analysis eigenvalue spectrum, and quantifying the curvature characteristics of the local region by the ratio of the largest to the second largest eigenvalue, the local manifold radius of curvature at the current moment can be obtained. This index, from the perspective of static geometry, characterizes the local structural features of the attractor near the current moment, capturing minute deformations of the system's dynamic structure over a short period. The maximum Lyapunov exponent is a quantitative indicator measuring the sensitivity of a dynamic system to small initial disturbances, representing the average exponential rate of divergence or convergence of adjacent orbits in phase space. For a dry gas-sealed system: when the system is healthy, the trajectory exhibits a certain degree of convergence or a stable divergence rate, and the maximum Lyapunov exponent remains within a stable range. When the system shows signs of a real fault, the system's dynamic behavior changes, the evolution rate of adjacent orbits changes accordingly, and the maximum Lyapunov exponent shows a significant deviation.When only data link noise exists, since the noise lacks physical causal continuity, it does not make a continuous and stable contribution to the logarithmic growth rate of the nearest neighbor's evolution distance, and therefore does not lead to a significant change in the maximum Lyapunov exponent. By tracking the evolution distance of each reference point and its nearest neighbor along the attractor trajectory within a sliding time window and linearly fitting the logarithmic growth rate of the evolution distance, the maximum Lyapunov exponent for the current window can be obtained. This index characterizes the stability changes of the system's dynamic behavior from the perspective of dynamic evolution trends. The local manifold curvature radius and the maximum Lyapunov exponent characterize the attractor from two dimensions: "static geometry" and "dynamic evolution," respectively. The former is sensitive to short-term, local structural changes and is suitable for detecting instantaneous deformation of the attractor; the latter is sensitive to medium-term, systemic evolution trends and is suitable for assessing the overall stability changes of the system. Combining the two allows for a more comprehensive capture of multi-dimensional dynamic changes caused by real fault precursors, while avoiding misjudgments based on fluctuations in a single index. In the initial stage of equipment healthy operation or during normal operating conditions after thorough maintenance and confirmation, a baseline model of the dry gas seal in a multi-dimensional state space is pre-constructed using long-term operating data. This baseline model not only records the macroscopic geometry of the attractor under normal conditions but also extracts its topological descriptors—such as barcodes or persistent graphs generated by a continuous cohomology algorithm—to characterize the attractor's connectivity, ring structure, void distribution, and other topological features, forming a complete formal description of the dynamic structure under normal operating conditions. The construction of the baseline model is not a one-time event but rather a continuous accumulation of historical normal operating condition data, forming a set of topological descriptors covering various normal operating condition manifestations, enabling the diagnostic model to tolerate reasonable fluctuations in operating conditions during operation. The real-time topological descriptor of the current attractor segment is constructed based on the local manifold radius of curvature and the maximum Lyapunov exponent calculated within the current sliding time window. Similarly, a continuous cohomology algorithm can be used to extract topological features (such as generating the current barcode or persistent graph) from the phase space point set corresponding to the current window data. The topological deviation between the real-time topological descriptor and the reference topological descriptor set in the baseline model is quantified using Wasserstein distance or Bottleneck distance. These distance metrics effectively measure the difference between two topological representations and have good mathematical properties and computational feasibility. If the deviation exceeds the sensitivity threshold, it indicates a significant difference in topological structure between the current attractor segment and the attractor under normal operating conditions, and the dynamic nature of the system has changed. If the deviation does not exceed the sensitivity threshold, it indicates that the current dynamic structure of the system is still within the normal range, and the observed numerical deviation is only a temporary perturbation. When the local topological structure of the running attractor trajectory undergoes an anomaly exceeding the preset sensitivity threshold, it is judged as a precursor to a change in the actual operating conditions.The fact that the acquired value falls into the transition zone at this point is not accidental, but rather indicates that the system dynamics of the dry gas seal have indeed changed—perhaps the end-face gap has begun to change gradually, microscopic wear on the sealing surface has begun to accumulate, or the gas film stiffness is declining. These real physical processes manifest as deformation of the attractor structure in phase space, and this deformation is continuous. This diagnostic conclusion requires retaining the original data and triggering the corresponding level of warning. When only point-like outliers exist but do not change the attractor topology, it is determined to be a distortion in the data transmission link. At this time, although the acquired value numerically exceeds the reasonable fluctuation boundary, the overall dynamics of the system have not undergone a detectable structural change, and isolated data points cannot cause attractor topology distortion at the phase space level. This diagnostic conclusion requires data repair or removal of the distorted signal, and the processed data should be recorded in the database.

[0035] In some embodiments, calculating the local manifold radius of curvature and the maximum Lyapunov exponent of the running attractor trajectory within a preset sliding time window specifically includes: Multivariable phase space reconstruction is performed on the real-time sequences of the leakage amount, the pressure, and the temperature to determine the optimal embedding delay and embedding dimension for each dimension, thereby obtaining a multivariable delayed embedding vector sequence. Within each sliding time window, weighted principal component analysis is performed on the local embedded vector point set to extract the eigenvalue spectrum, and the local manifold radius of curvature is defined by the ratio of the largest eigenvalue to the second largest eigenvalue. Within the sliding time window, the evolution distance between each reference point and its nearest neighbor along the trajectory of the running attractor is tracked, and the logarithmic growth rate of the evolution distance is linearly fitted to obtain the maximum Lyapunov exponent.

[0036] According to the embedding theorem, a time series of a single variable contains information about all dynamic variables of the system. Using time-delayed embedding, a high-dimensional phase space equivalent to the original system dynamics can be reconstructed from the scalar sequence. This application employs a multivariate phase space reconstruction strategy, jointly considering the real-time sequences of three parameters to construct an embedding vector sequence that more completely reflects the system state. The quality of the phase space reconstruction depends on two key parameters: embedding delay and embedding dimension. The embedding delay determines how far apart points in the time series are considered independent variables; the embedding dimension determines the size of the reconstructed phase space. Too small an embedding delay leads to high correlation between adjacent coordinates and the reconstructed attractor being compressed near the diagonal; too large an embedding delay makes the causal relationships between coordinates too discrete. Too low an embedding dimension prevents the attractor structure from being fully unfolded, resulting in spurious trajectory intersections; too high an embedding dimension amplifies noise and increases computational burden. The optimal embedding delay and embedding dimension are determined for each dimension. The embedding delay can be determined by the first minimum point of the mutual information method, where the time-delayed sequence and the current sequence share a balanced amount of information, neither excessively redundant nor excessively independent. The embedding dimension can be determined using the spurious nearest neighbor method. The core idea is to gradually increase the dimension and observe whether neighboring points in the high-dimensional space maintain their proximity after the dimension increase—when the proportion of spurious neighbors drops sufficiently, the corresponding dimension is the optimal embedding dimension. After determining the optimal parameters for each dimension, the time series of the three monitoring parameters are arranged with time delay embedding according to their respective embedding delays and embedding dimensions, generating a high-dimensional embedding vector at each sampling time. All embedding vectors within the entire sampling interval are arranged in chronological order, forming a multivariate time-delayed embedding vector sequence, which fully characterizes the trajectory of the dry gas sealing system in the reconstructed phase space. The local manifold radius of curvature of the attractor measures the degree of curvature of the attractor in a local region of phase space. Geometrically, if the attractor is relatively flat at a certain point, the set of embedding vector points at that point is roughly distributed on a low-dimensional hyperplane with a large radius of curvature; if the attractor is curved at a certain point, the point set will be dispersed in multiple directions with a small radius of curvature. When a dry gas sealing system exhibits early signs of a real fault, changes in the dynamic structure may initially manifest as increased local bending of the attractor and a more tortuous trajectory in phase space. Therefore, a decrease in the radius of curvature of the local manifold is an early geometric signal of a fault precursor. Within a sliding time window, a set of embedding vectors near the current moment is selected as the local point set. Weighted principal component analysis is then performed on this local point set—points closer to the current moment are assigned greater weights, while points farther away are assigned smaller weights. This aims to ensure that the analysis results better reflect the latest state at the current moment, while not completely ignoring information provided by neighboring historical points, achieving a balance between time-varying sensitivity and estimation stability. Principal component analysis then performs eigenvalue decomposition on the weighted point set, yielding a set of eigenvalues ​​and corresponding eigenvectors.The magnitude of eigenvalues ​​reflects the degree of dispersion of a local point set along the corresponding eigenvector direction—the largest eigenvalue corresponds to the principal direction where the point set is most fully expanded, and the second largest eigenvalue corresponds to the secondary principal direction. The curvature characteristic of the local region is defined by the ratio of the largest to the second largest eigenvalue. The maximum Lyapunov exponent measures the rate at which a dynamic system amplifies (or diminishes) small initial differences. Intuitively, if two very close initial points are selected in phase space, after evolving along their respective trajectories for a period of time, the distance between them will change—for chaotic or unstable systems, the distance increases exponentially; for stable systems, the distance remains constant or decreases. The logarithm of this divergence (or convergence) rate is the Lyapunov exponent. Under dry gas seal health conditions, the system has certain damping characteristics, the trajectory remains within a stable region, and the evolution distance between adjacent points will not continuously increase; when real fault precursors appear, the system stability decreases, the sensitivity to small disturbances increases, and the Lyapunov exponent may increase. Therefore, this exponent can sensitively reflect changes in the dynamic stability of the system. Within a sliding time window, for each reference point, its nearest neighbor is found in phase space. These two points form a pair with similar initial states. Then, along their respective time evolution directions, the positions of these two points after the same number of evolution steps are traced, and the distance between them—the evolutionary distance—is calculated. The growth rate of the evolutionary distance reflects the system's sensitivity to small initial differences. If the system tends to be unstable, the distance between adjacent orbits will increase rapidly over time; if the system is stable, the distance will tend to remain constant or decrease slowly. A linear fit is performed on the relationship between the logarithmic distance and the evolution time. The average slope of the logarithmic distance over time is the estimate of the maximum Lyapunov exponent. Calculating the exponent separately using multiple reference points within the sliding window and then combining the results effectively smooths out the interference of individual noise points on the calculation results, improving the robustness of the estimation. The Lyapunov exponent only shows significant deviations when the actual dynamic behavior of the system changes; isolated noise points in the link, lacking causal continuity in physical evolution, do not have a systematic impact on the overall fit.

[0037] In some embodiments, the step of comparing the local manifold radius of curvature and the maximum Lyapunov exponent at the current moment with the attractor baseline model established under historical normal operating conditions specifically includes: Using long-term operating data accumulated under the health status of the equipment, a multidimensional dynamic benchmark model under normal operating conditions is pre-constructed. The benchmark model includes a set of reference topological descriptors for normal attractors, and the topological descriptors include barcodes or persistent graphs generated by continuous cohomology. For the local manifold radius of curvature and the maximum Lyapunov exponent calculated within the current sliding window, a real-time topological descriptor for the current attractor segment is constructed, and topological features are extracted using a continuous homology algorithm. The topology deviation value between the real-time topology descriptor and the reference topology descriptor set in the baseline model is quantified using Wasserstein distance or Bottleneck distance, and the topology deviation value is compared with the preset sensitivity threshold.

[0038] The benchmark model is built upon long-term operational data accumulated under healthy operating conditions. The core of the benchmark model is a "set of reference topological descriptors for normal attractors." It is not a single descriptor, but a set, where each reference topological descriptor corresponds to a normal operating condition manifestation, collectively forming a complete characterization of the dynamic structure under normal operating conditions. The topological descriptors are generated using barcodes or persistent graphs based on continuous cohomology. Continuous cohomology is a method in computational topology that can extract topologically invariant features from point cloud data. For each typical data segment under normal operating conditions, the corresponding reference topological descriptor is generated according to the following sub-steps: mapping the embedding vector sequence corresponding to the data segment to the phase space to form a normal attractor segment; performing continuous cohomology analysis on the point set in the segment to generate a barcode or persistent graph; and storing this barcode or persistent graph as a reference topological descriptor in the benchmark model set. After processing massive amounts of normal operating condition data, the benchmark model ultimately contains a set of reference topological descriptors covering multiple normal operating manifestations, providing a comprehensive reference basis for subsequent comparisons. When the system reaches a certain point in time, and the collected values ​​fall into the transition zone, triggering the diagnostic model, the local manifold radius of curvature and the maximum Lyapunov exponent are calculated for the data within the current sliding time window. The sequence of these two indices constitutes the dynamic characteristic representation of the attractor segment within the current window. The data point set of the calculated local manifold radius of curvature and maximum Lyapunov exponent within the current window is considered as discrete sampling of the attractor segment in the feature space. The following operations are performed on this point set: the point set is mapped to an appropriate metric space as input for continuous cohomology analysis; a continuous cohomology algorithm is used to gradually change the scale parameter, tracking the appearance and disappearance of each topological feature in the geometric structure formed by the point set at different scales; the birth and disappearance scales of each topological feature are recorded, generating a barcode or persistent map corresponding to the current attractor segment, which is the real-time topological descriptor. To quantify the degree of difference between the real-time topological descriptor and the reference topological descriptor set in the baseline model, a mathematically rigorous distance metric is required. This application uses Wasserstein distance or Bottleneck distance in its embodiments. Both of these distance metrics are widely recognized and applied in the field of persistent cohomology, specifically used to measure the similarity between two barcodes or persistent graphs. They share the following characteristics: they have a rigorous mathematical foundation, satisfying the axiomatic requirements of distance metrics; they exhibit good continuity under small deformations, without drastic jumps due to minor perturbations; and their computational cost is within an acceptable range for engineering applications. The distance between the real-time topology descriptor and each reference topology descriptor in the baseline model set is calculated one by one, and the minimum value is taken as the final topology deviation value. Taking the minimum value means determining whether the current attractor segment is topologically consistent with at least one known normal operating condition in the baseline model.If the current segment's topological structure is sufficiently close to a normal form, the system is considered to be within the normal dynamic range. Conversely, if the current segment's real-time topological descriptor is significantly different from all reference topological descriptors in the baseline model set, the minimum value will also be large, indicating that the current system's dynamic topological structure has deviated from all normal forms and undergone a fundamental change. The calculated topological deviation value is compared with a preset sensitivity threshold: if the topological deviation value exceeds the preset sensitivity threshold, it is determined that the local topological structure of the attractor trajectory has undergone an anomaly, and this suspected out-of-limit sample is a precursor to changes in actual operating conditions; if the topological deviation value does not exceed the preset sensitivity threshold, it is determined that the attractor topological structure has not undergone a fundamental change, and this suspected out-of-limit sample is a distortion in the data transmission link. The sensitivity threshold can be determined by cross-validation using historical normal operating condition data, calculating the statistical distribution of topological deviation values ​​between segments under normal operating conditions, and selecting a value that covers the vast majority of normal fluctuation ranges as the sensitivity threshold while retaining sensitivity to real anomalies.

[0039] In some embodiments, the data repair or removal of distorted signals includes: A generative adversarial network model consisting of a generator and a discriminator is constructed and pre-trained. The generator learns the mapping from a masked multi-parameter incomplete signal to a complete and real sealed state signal. When the current signal segment is determined to be a distorted region, a binary masking operation is performed on the distorted data points in the current signal segment, and the multi-channel parameter matrix carrying the mask is input into the generator; The generator generates reconstructed data point by point under the continuous feature constraints of the running parameters and the three-dimensional spatiotemporal constraints of the associated parameters. The discriminator provides the generator with an adversarial gradient to enhance the reconstruction details. The original distorted fragments are replaced with the reconstructed data generated by the generator, and the confidence labels of the repaired data records are quantified.

[0040] This application employs a generative adversarial network (GAN) model consisting of a generator and a discriminator. The generator's task is to learn the mapping relationship from the masked multi-parameter incomplete signal to the complete and true sealed state signal; the discriminator's task is to distinguish the reconstructed signal output by the generator from the true complete signal, providing feedback gradients to the generator. The generator's core learning objective is: given a multi-parameter channel matrix where some data positions are masked, the generator needs to infer and fill in the true signal values ​​at the masked positions based on the unmasked context information and the physical relationships between the multi-parameters. After sufficient training, the generator can internalize the physical constraints and signal characteristics of the dry gas sealing system, forming an accurate mapping capability from incomplete signals to complete signals. The pre-training data comes from high-quality historical data accumulated during the healthy operation of the equipment. By performing random masking operations on the complete data, training sample pairs of "incomplete signal-complete signal" are constructed, enabling the generator to learn the signal reconstruction capability under different masking positions and different masking ratios. When the diagnostic model completes the authenticity diagnosis of the current signal segment and outputs the conclusion of "distortion in the data transmission link," the diagnostic result also indicates the specific location of the distorted data points within the segment. These distorted data points are the targets that need to be repaired. A binary masking operation is performed on the distorted data points in the current signal segment, that is, the value at the distorted position is replaced with a special mask marker, and the information on whether each position is covered is recorded in the mask matrix. For normal data points that are not determined to be distorted, their original acquired values ​​are retained unchanged. The multi-channel parameter matrix carrying the mask marker—containing the time series of various monitoring parameters such as leakage, pressure, and temperature, where the distorted positions have been masked—is used as the input to the generator. The multi-channel input design ensures that when the generator repairs the distortion of a certain parameter, it can simultaneously refer to the normal acquired values ​​of other parameters at the same time, making full use of the physical correlation information between parameters. Continuity feature constraints refer to the continuity characteristics of each monitoring parameter itself on the time axis—for example, the temperature signal will not undergo instantaneous step jumps, and the change in leakage under normal operating conditions is smooth, etc. When generating reconstructed data, the generator is constrained by these single-parameter temporal continuity characteristics, ensuring that the reconstructed signal conforms to physical laws in the time dimension and avoiding unreasonable abrupt changes or high-frequency oscillations. Three-dimensional spatiotemporal constraints refer to the correlations between parameters and time delays—for example, there is a specific time-delay response relationship between changes in leakage and sealing cavity pressure, and temperature fluctuations and pressure changes satisfy thermodynamic laws. When generating reconstructed data, the generator ensures that the data from the repaired channel satisfies these preset physical correlations with the data from other channels at the same or adjacent times, ensuring the physical consistency of the linkage between multi-parameter reconstructed signals. During the generator's data generation process, the discriminator simultaneously performs authenticity checks on the generated results.The discriminator identifies unrealistic detail regions in the reconstructed data by comparing the statistical distribution and local morphology of the reconstructed data fragments output by the generator with those of real, normal data, and propagates an adversarial gradient back to the generator. Based on this gradient feedback, the generator further optimizes the detail representation of the reconstructed signal, making it more closely approximate the real sealed signal in terms of statistical features and local morphology. The generator reconstructs each mask position in the distorted fragment point-by-point, outputting a reconstructed data sequence with the same sampling rate as the original signal. This point-by-point generation strategy ensures the temporal resolution and detail granularity of the reconstructed signal, avoiding smooth transition artifacts caused by block filling. After the generator completes the reconstruction of the mask positions, it seamlessly replaces the distorted fragments in the original signal with the generated reconstructed data. Normal data points remain unchanged, while distorted data points are replaced by the reconstructed values ​​output by the generator, maintaining the continuity and integrity of the entire signal fragment on the time axis. For data points repaired by the generator, the system records a quantized confidence label. This label reflects the confidence level of the repair process, and its value can be determined by considering the following factors: the generator's confidence in the output of the mask location, the density of distorted points in the distorted fragment, the quantity and quality of the normal data points in the context upon which the repair depends, and the performance of this type of distortion pattern in historical repairs. The purpose of the confidence label is to enable downstream advanced applications such as trend analysis, fault diagnosis, and lifetime prediction to distinguish between "directly collected real data" and "reconstructed data repaired by the model," thereby allowing for differentiated weighting when using repaired data for decision-making. For example, when the confidence label is low, downstream applications can reduce the contribution weight of that data point in trend fitting; when the confidence label is high, the repaired data can serve as an effective supplement to normal data.

[0041] In some embodiments, the step of generating reconstructed data point by point through the generator under the continuous feature constraints of the running parameters and the three-dimensional spatiotemporal constraints of the associated parameters specifically includes: The generator's dual-channel feature extraction encoder processes the multi-channel parameter matrix. The first channel of the dual-channel feature extraction encoder performs autocorrelation function morphology extraction and wavelet scattering transformation on each single-channel time series in the multi-channel parameter matrix, and outputs continuous feature constraint vectors for each channel. The second channel calculates the mutual information delay spectrum and dynamic conditional entropy transfer matrix between channels in the multi-channel parameter matrix, and outputs a cross-channel three-dimensional spatiotemporal constraint matrix. At each scale level of the generator, the temporal consistency loss between the hierarchical reconstructed feature map and the continuous feature constraint vector, and the linkage physical consistency loss between the cross-channel reconstructed feature map and the three-dimensional spatiotemporal constraint matrix are calculated, and the losses at each level are weighted and accumulated into the total loss function of the generator. Iterative adversarial training is performed based on the total loss function to generate reconstructed data point by point.

[0042] The dual-channel feature extraction encoder is the pre-module of the generator. Its function is to extract two types of physical constraint information of different properties from the input masked multi-channel parameter matrix. The two channels work in parallel and output continuous feature constraint vectors and three-dimensional spatiotemporal constraint matrices respectively. The first channel focuses on the time series characteristics of each monitoring parameter itself, processes each single-channel time series in the multi-channel parameter matrix separately, and extracts the continuous feature constraint vectors of each channel. Its core operation includes two parts: (1) Autocorrelation function morphology extraction: The autocorrelation function describes the degree of correlation between a time series and itself at different time delays, and can characterize the inherent dynamic characteristics of the signal such as periodicity, memory and smoothness. By calculating the autocorrelation function of each channel time series and extracting its morphological features - such as autocorrelation decay rate, principal period position, envelope shape, etc. - the inherent constraint law of the parameter's evolution over time can be captured. For example, the autocorrelation function of the temperature series usually shows a slow decay, reflecting its large thermal inertia; the pressure series may show periodic features related to the compressor speed. These morphological information provide the generator with the continuity prior that the parameter should follow on the time axis. (2) Wavelet Scatter Transform: Wavelet scatter transform is a nonlinear signal transformation method based on wavelet cascade. It extracts stable representations of signals at different time and frequency scales by performing iterative operations of multi-scale wavelet decomposition, modulus taking, and local averaging on the signal. Scatter transform has translation invariance and stability against small-amplitude deformation, and can effectively separate trend components, periodic components, and noise components in the signal. After performing scatter transform on the time series of each channel, the scattering coefficients of the signal at multiple time-frequency scales of the channel are obtained. These coefficients, together with the morphological characteristics of the autocorrelation function, constitute the continuous feature constraint vector of the channel. The continuous feature constraint vector finally output by the first channel encodes the inherent evolution law of each monitoring parameter in the time dimension—including the degree of smoothness, periodicity, multi-scale fluctuation mode, etc., which provide a specification for single-parameter temporal consistency for the generator during point-by-point reconstruction. The second channel focuses on the correlation between different monitoring parameters, calculates the coupling characteristics between channels in the multi-channel parameter matrix, and outputs a cross-channel three-dimensional spatiotemporal constraint matrix. Its core operation includes two parts: (1) Mutual information delay spectrum calculation: Mutual information measures the amount of information shared between two time series and is better at capturing nonlinear correlations than linear correlation coefficients. When calculating mutual information for two time series with different parameters, different time delays are introduced to obtain a spectrum of mutual information changing with time delay. The peak position in the mutual information delay spectrum reveals the time delay direction and time scale of the coupling relationship between the two parameters—for example, how many sampling periods after a change in leakage will cause a pressure response, this delay is the peak position of the mutual information delay spectrum at the corresponding delay. By calculating the mutual information delay spectrum between each channel pair, a time delay structure diagram of the correlation between parameters can be constructed.(2) Calculation of dynamic conditional entropy transfer matrix: Conditional entropy measures the uncertainty of one variable when another variable is known. The dynamic conditional entropy transfer matrix characterizes the transfer path and efficiency of information between parameters by calculating the change of conditional entropy between different time points. The basic approach is to calculate the conditional entropy of a parameter within a sliding time window under the condition that the current and historical values ​​of another parameter are known, observe the dynamic change of conditional entropy over time, and construct a transfer matrix describing the information flow and coupling strength between parameters. This quantitative description describes the dynamic influence relationship across parameters, such as "how the degree of elimination of pressure uncertainty by leakage changes evolves over time". The three-dimensional spatiotemporal constraint matrix finally output by the second channel encodes the cross-channel relationships such as the coupling strength, time delay direction, and information transfer path between parameters, providing a specification for the physical consistency of multi-parameter linkage when the generator is reconstructed point by point. The so-called "three-dimensional" means that the constraint covers both the parameter dimension (between different parameter channels) and the time dimension (relationship under different time delays); the so-called "spatiotemporal" means that it describes both the relationship between parameters in space and the law of the evolution of these relationships over time. This embodiment calculates two types of physical consistency loss at each scale level of the generator. At each scale level, the reconstructed feature map of the generator at that level is compared with the continuous feature constraint vector output by the first channel to calculate the degree of consistency between the two in terms of temporal characteristics. The temporal consistency loss measures how much the signal currently reconstructed by the generator deviates from the continuous feature constraints that a normal signal should have in terms of autocorrelation morphology and multi-scale scattering characteristics. If the signal reconstructed by the generator exhibits unreasonable abrupt changes, high-frequency noise, or fluctuation patterns inconsistent with normal temporal patterns at some point, its temporal consistency loss will increase. This loss term constrains the generator to maintain the natural evolution characteristics of the single-parameter signal in the time dimension at each scale. At each scale level, the cross-channel reconstructed feature map is compared with the three-dimensional spatiotemporal constraint matrix output by the second channel to calculate the degree of consistency between the two in terms of cross-parameter coupling relationship. The linkage physical consistency loss measures how much the signals currently reconstructed by the generator deviate from the coupling constraints that a normal signal should have in terms of mutual information delay structure and conditional entropy transfer relationship. If the mutual information delay spectrum between the leak and pressure sequences reconstructed by the generator is inconsistent with the normal coupling relationship—for example, due to time delay misalignment or excessive coupling strength deviation—the resulting physical consistency loss will increase. This loss term constrains the generator to maintain the correct correlation structure and information flow between the parameter signals at each scale. The temporal consistency loss and physical consistency loss calculated at each scale level are weighted and summed according to preset hierarchical weights, and accumulated into the generator's total loss function. The multi-level design allows physical constraints to penetrate layer by layer from macroscopic trends to microscopic details, ensuring that the final reconstructed data satisfies physical laws in both global structure and local details.The hierarchical weighting can assign greater weight to lower scales to prioritize the accuracy of macroscopic trends, while assigning appropriate weights to higher scales to ensure physical consistency in details. The generator's total loss function consists of three parts: adversarial loss (feedback from the discriminator, pushing the generated signal to statistically approximate the real signal), temporal consistency loss (from the first channel constraint, pushing the generated signal to satisfy single-parameter temporal patterns), and linkage physical consistency loss (from the second channel constraint, pushing the generated signal to satisfy multi-parameter coupling relationships). The three losses are weighted and combined according to preset coefficients. During the training phase, the generator and discriminator alternately update their parameters. In each iteration: the generator receives a masked multi-channel parameter matrix and outputs reconstructed data; the discriminator receives the generator's output and the real complete signal, performs authenticity discrimination, and calculates the adversarial loss; the dual-channel feature extraction encoder synchronously calculates the continuous feature constraint vector and the three-dimensional spatiotemporal constraint matrix, and calculates the temporal consistency loss and linkage physical consistency loss; the three losses are weighted and accumulated into a total loss function, which is then backpropagated to update the generator's parameters. During the inference phase (actual online repair), the trained generator receives a multi-channel parameter matrix with a mask, propagates forward once, and then outputs reconstructed data for each mask position.

[0043] In some embodiments, before marking the current collection point and its corresponding time series segment as a suspected out-of-limit sample, the method further includes: The signal quality is evaluated on the real-time acquired values. The signal quality evaluation includes signal-to-noise ratio detection, data frame integrity verification, and reading of sensor health status indicators, and outputs a quantitative reliability score. When the quantitative confidence score is lower than the preset minimum confidence threshold, the current collection point is directly marked as a distortion in the data transmission link; When the quantitative confidence score is higher than or equal to the minimum confidence threshold and lower than the preset high confidence threshold, the current collection point is marked as a suspected out-of-limit sample, and the quantitative confidence score is added as a prior input weight for the diagnostic model. When the quantitative confidence score is lower than the minimum confidence threshold for a consecutive preset number of collection points, and the collection values ​​all fall within the transition zone, it is determined that the sensor or the collection link has a continuous failure, triggering a sensor maintenance warning.

[0044] Signal-to-noise ratio (SNR) testing is used to evaluate the ratio of effective information components to noise components in a acquired signal. In dry gas seal monitoring, normal operating parameters change relatively smoothly and continuously, while noise exhibits high-frequency, small-amplitude random fluctuations. By performing spectral analysis or statistical fluctuation analysis on a time window near the current acquisition point, the local SNR of the current signal can be estimated. If the SNR is too low, it indicates that the acquired value may have been subjected to significant electromagnetic interference or instantaneous sensor disturbances, casting doubt on the physical authenticity of the data. The SNR test result is quantified as a component of the credibility score. During the transmission of data from the sensor through the acquisition unit to the server, it is typically encapsulated and transmitted in the form of data frames. Each data frame contains integrity verification fields such as a checksum. By checking the checksum of the received data frame, it can be determined whether bit errors or packet loss occurred during transmission. If the data frame integrity check fails, it indicates that a clear encoding / decoding error has occurred in the transmission link, and its value has deviated from the original sensor measurement. The integrity check result is incorporated into the credibility scoring system in the form of "pass / fail". When receiving collected values, the system simultaneously reads the health status identifier attached to the sensor. If the sensor reports itself to be in a sub-healthy or faulty state, the reliability of the collected value will be significantly reduced. The sensor health status identifier reflects the reliability of the signal generation source and is the most fundamental part of the reliability assessment. Combining the evaluation results of the signal-to-noise ratio detection, data frame integrity verification, and sensor health status identifier, the system outputs a quantitative reliability score through a preset fusion rule. This score is usually normalized to a certain range (such as between 0 and 1), with higher values ​​indicating higher signal reliability and a greater likelihood of being a true physical measurement value. When the quantitative reliability score is below the minimum reliability threshold, it means that there is a clear quality defect at the signal level in this collected value—it could be an extremely low signal-to-noise ratio, a failed data frame verification, or the sensor itself reporting a faulty state. Such out-of-bounds values ​​are almost certainly caused by signal link problems rather than actual physical processes. For such signals, the system skips all subsequent diagnostic processes and directly marks the current collection point as a distortion in the data transmission link, entering the data repair or removal process. When the quantization confidence score is higher than or equal to the minimum confidence threshold but lower than the preset high confidence threshold, it indicates that the signal of the collected value may have been subject to a certain degree of noise interference or slight transmission impairment. The out-of-bounds value could reflect real physical changes or be a false out-of-bounds error caused by signal quality degradation. For signals in this gray area, the system marks them as suspected out-of-bounds samples and initiates a diagnostic model for deep authenticity diagnosis. Simultaneously, the system adds the quantization confidence score as a priori input weight to this sample in the diagnostic model.In subsequent analysis, the diagnostic model will refer to this prior weight to adjust the confidence level of the judgment. For samples with low confidence scores, the diagnostic model needs to collect stronger evidence to determine that they represent real changes in operating conditions; for samples with high confidence scores, the diagnostic model can make judgments based on normal standards. The system continuously tracks the quantitative confidence score output by each acquisition point. When the following situation occurs, a maintenance warning for the sensor or acquisition link is triggered: a preset number of acquisition points have quantitative confidence scores that are all below the minimum confidence threshold, and the acquired values ​​of these acquisition points all fall within the transition zone (i.e., excluding the case where a safety over-limit alarm has been triggered). The theoretical basis for this judgment logic is that: low confidence of a single acquisition point may be due to occasional electromagnetic interference or instantaneous packet loss, which is an unavoidable transient disturbance in normal engineering sites; however, when multiple acquisition points continuously show low confidence, the probability of occasional transient disturbances is extremely low. A more reasonable explanation is that the sensor or acquisition link has experienced a persistent organic failure—such as aging and drifting of sensor sensing elements, unstable power supply modules, poor contact of signal cables, damage to the acquisition unit interface circuit, etc. More importantly, this persistent low-reliability signal is concentrated precisely in the transition zone (rather than the safe over-limit zone), indicating that the measured physical quantity itself has not undergone drastic changes, but rather that the measurement system itself is chronically deteriorating, continuously outputting distorted acquired values. When the above conditions are triggered, the system issues a sensor maintenance warning, notifying maintenance personnel to inspect and maintain the signal acquisition link, and to perform preventative replacement or repair before the sensor completely fails.

[0045] The above describes the method for identifying and processing excessive operating parameters of dry gas seals in the embodiments of this application. The computer system in the embodiments of this application will be described in detail below in conjunction with the above method for identifying and processing excessive operating parameters of dry gas seals.

[0046] Please see Figure 4 This is a schematic diagram of an exemplary hardware structure of a computer system in an embodiment of this application.

[0047] In some embodiments, the computer system 400 includes a computer device, which may be a terminal device. The computer device includes a processor 401, a memory 402, a sensor module 403, a communication module 404, an input device 405, and an output device 406 connected via a system bus. The processor 401 of the computer device provides computing and control capabilities. The memory 402 of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database is used to store data.

[0048] Those skilled in the art will understand that Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0049] In some embodiments of this application, a computer-readable storage medium is provided, including instructions that, when executed on the computer system 400, cause the computer system 400 to perform the dry gas seal operating parameter over-limit identification and processing method of the embodiments of this application.

[0050] In some embodiments of this application, a computer program product is also provided, which, when run on a computer system 400, causes the computer system 400 to execute the dry gas seal operating parameter over-limit identification and processing method in the embodiments of this application.

[0051] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for identifying and processing excessive operating parameters of a dry gas seal, characterized in that, include: For each monitoring parameter, a first upper limit value and a first lower limit value are preset. A safe operating range is determined based on the first upper limit value and the first lower limit value. Within the safe operating range, a reasonable fluctuation range with a range boundary narrower than the safe operating range is set. The second upper limit value of the reasonable fluctuation range is less than the first upper limit value, and the second lower limit value of the reasonable fluctuation range is greater than the first lower limit value. The system acquires the collected values ​​of each monitoring parameter in real time. When the collected value exceeds the safe operating range, it is determined to be an equipment malfunction and an alarm is triggered. When the collected value falls within the reasonable fluctuation range, it is determined to be a normal operating condition. When the collected value falls into the transition zone between the safe operating range and the reasonable fluctuation range, the current collection point and the corresponding time series segment are marked as suspected out-of-limit samples. The suspected out-of-limit samples are subjected to a real-world diagnostic test using a pre-defined diagnostic model to determine whether they are a precursor to changes in actual operating conditions or distortions in the data transmission link. If the diagnosis is a precursor to a change in actual operating conditions, the original data is retained and an early warning of the corresponding level is triggered. If the diagnosis is a distortion in the data transmission link, the distorted signal is repaired or removed, and the processed data is recorded in the database.

2. The method according to claim 1, characterized in that, The step of performing a verifiable diagnosis on the suspected out-of-limit sample using a preset diagnostic model includes: In the multidimensional state space of dry gas seal operation, the real-time sequences of three parameters, namely leakage, pressure and temperature, are extracted to construct the operating attractor trajectory in high-dimensional phase space. Calculate the local manifold radius of curvature and the maximum Lyapunov exponent of the running attractor trajectory within a preset sliding time window; Compare the current local manifold radius of curvature and maximum Lyapunov exponent with the topological deviation of the attractor baseline model established under historical normal operating conditions. If the local topology of the running attractor trajectory undergoes an anomaly exceeding a preset sensitivity threshold, it is determined to be a precursor to a change in the actual operating conditions; if there are only point outliers and the attractor topology remains unchanged, it is determined to be a distortion in the data transmission link.

3. The method according to claim 2, characterized in that, The calculation of the local manifold radius of curvature and the maximum Lyapunov exponent of the running attractor trajectory within the preset sliding time window specifically includes: Multivariable phase space reconstruction is performed on the real-time sequences of the leakage amount, the pressure, and the temperature to determine the optimal embedding delay and embedding dimension for each dimension, thereby obtaining a multivariable delayed embedding vector sequence. Within each sliding time window, weighted principal component analysis is performed on the local embedded vector point set to extract the eigenvalue spectrum, and the local manifold radius of curvature is defined by the ratio of the largest eigenvalue to the second largest eigenvalue. Within the sliding time window, the evolution distance between each reference point and its nearest neighbor along the trajectory of the running attractor is tracked, and the logarithmic growth rate of the evolution distance is linearly fitted to obtain the maximum Lyapunov exponent.

4. The method according to claim 3, characterized in that, The step of comparing the current local manifold radius of curvature and maximum Lyapunov exponent with the attractor baseline model established under historical normal operating conditions for topological deviation specifically includes: Using long-term operating data accumulated under the health status of the equipment, a multidimensional dynamic benchmark model under normal operating conditions is pre-constructed. The benchmark model includes a set of reference topological descriptors for normal attractors, and the topological descriptors include barcodes or persistent graphs generated by continuous cohomology. For the local manifold radius of curvature and the maximum Lyapunov exponent calculated within the current sliding window, a real-time topological descriptor for the current attractor segment is constructed, and topological features are extracted using a continuous homology algorithm. The topology deviation value between the real-time topology descriptor and the reference topology descriptor set in the baseline model is quantified using Wasserstein distance or Bottleneck distance, and the topology deviation value is compared with the preset sensitivity threshold.

5. The method according to claim 1, characterized in that, The data repair or removal of distorted signals includes: A generative adversarial network model consisting of a generator and a discriminator is constructed and pre-trained. The generator learns the mapping from a masked multi-parameter incomplete signal to a complete and real sealed state signal. When the current signal segment is determined to be a distorted region, a binary masking operation is performed on the distorted data points in the current signal segment, and the multi-channel parameter matrix carrying the mask is input into the generator; The generator generates reconstructed data point by point under the continuous feature constraints of the running parameters and the three-dimensional spatiotemporal constraints of the associated parameters. The discriminator provides the generator with an adversarial gradient to enhance the reconstruction details. The original distorted fragments are replaced with the reconstructed data generated by the generator, and the confidence labels of the repaired data records are quantified.

6. The method according to claim 5, characterized in that, The process of generating reconstructed data point by point through the generator under the continuous feature constraints of the operating parameters and the three-dimensional spatiotemporal constraints of the associated parameters specifically includes: The generator's dual-channel feature extraction encoder processes the multi-channel parameter matrix. The first channel of the dual-channel feature extraction encoder performs autocorrelation function morphology extraction and wavelet scattering transformation on each single-channel time series in the multi-channel parameter matrix, and outputs continuous feature constraint vectors for each channel. The second channel calculates the mutual information delay spectrum and dynamic conditional entropy transfer matrix between channels in the multi-channel parameter matrix, and outputs a cross-channel three-dimensional spatiotemporal constraint matrix. At each scale level of the generator, the temporal consistency loss between the hierarchical reconstructed feature map and the continuous feature constraint vector, and the linkage physical consistency loss between the cross-channel reconstructed feature map and the three-dimensional spatiotemporal constraint matrix are calculated, and the losses at each level are weighted and accumulated into the total loss function of the generator. Iterative adversarial training is performed based on the total loss function to generate reconstructed data point by point.

7. The method according to claim 1, characterized in that, Before marking the current collection point and its corresponding time series segment as suspected out-of-limit samples, the following steps are also included: The signal quality is evaluated on the real-time acquired values. The signal quality evaluation includes signal-to-noise ratio detection, data frame integrity verification, and reading of sensor health status indicators, and outputs a quantitative reliability score. When the quantitative confidence score is lower than the preset minimum confidence threshold, the current collection point is directly marked as a distortion in the data transmission link; When the quantitative confidence score is higher than or equal to the minimum confidence threshold and lower than the preset high confidence threshold, the current collection point is marked as a suspected out-of-limit sample, and the quantitative confidence score is added as a prior input weight for the diagnostic model. When the quantitative confidence score is lower than the minimum confidence threshold for a consecutive preset number of collection points, and the collection values ​​all fall within the transition zone, it is determined that the sensor or the collection link has a continuous failure, triggering a sensor maintenance warning.

8. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

Images