A gas bearing vacuum pump performance prediction system based on a specific computational model

Abstract

The application discloses a kind of based on specific computing model's air floatation vacuum pump performance prediction system, it is related to mechanical engineering technical field, including multi-parameter fluctuation perception module, weak feature sensitive extraction module, abnormal coupling graph construction module, instability precursor identification module, prediction model regulation and control module and model adaptive optimization module: multi-parameter fluctuation perception module, obtain the real-time data of temperature, flow and vibration signal, calculate the instantaneous change rate of each signal, and carry out fusion in preset time window, construct multi-parameter synchronous fluctuation perception model, output synchronous abnormal index.The application realizes the accurate identification and early warning to the early instability trend of air floatation vacuum pump by introducing multi-parameter synchronous fluctuation perception, weak feature extraction and abnormal coupling graph, and combines model regulation and control and adaptive optimization mechanism, constructs intelligent system with real-time perception and closed-loop prediction ability, improves operation safety and stability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical engineering, in particular to a gas bearing vacuum pump performance prediction system based on a specific calculation model. BACKGROUND

[0002] The gas bearing vacuum pump performance prediction system based on a specific calculation model refers to an intelligent system that dynamically simulates and trends the key performance indicators (such as pumping speed, ultimate vacuum, energy efficiency, axial floating stability, etc.) of a gas bearing vacuum pump under different operating conditions by constructing a mathematical model or a data-driven model that highly matches the operating mechanism of the gas bearing vacuum pump. The system usually integrates a sensor acquisition module, a historical data analysis module, a model reasoning engine, and a prediction decision unit. Through feature extraction of the multi-parameter input of the gas flow state inside the pump body, the rotor suspension state, and the external load conditions, combined with specific algorithms (such as physical mechanism models, neural network models, or hybrid prediction models), the system can output real-time prediction results of future performance fluctuations, assisting engineers in optimizing the operation scheduling and maintenance strategies of the gas bearing vacuum pump, thereby improving the safety, stability, and energy efficiency of the equipment operation.

[0003] The prior art has the following disadvantages:

[0004] During the long-term high-load operation of the gas bearing vacuum pump, multiple types of sensors (including but not limited to temperature sensors, flow sensors, and vibration sensors) inside the system may simultaneously experience micro-amplitude deviations from the normal state due to the coupling effects of operating condition boundary approximation, material fatigue, or micro-disturbances. Such disturbances usually do not constitute significant abnormalities in a single parameter, but when multiple parameter abnormalities coincide in time, they may form an early composite feature pattern that indicates potential instability trends of the equipment. However, since the calculation model in the existing performance prediction system does not design feature modeling and response mechanisms for this rare coupled abnormal state during the construction process, the system often misjudges it as a common short-term fluctuation or environmental disturbance in normal operation, and the adaptive prediction mechanism automatically filters or downweights such data, thereby outputting a "trend stable" or "no warning" prediction result. This problem not only prevents the equipment from providing risk warnings during the early instability stage, but also may lead to serious operating consequences such as gas film support imbalance, pumping speed drop, or rotor drift due to the lack of prediction intervention, severely restricting the accuracy and safety assurance capability of the gas bearing vacuum pump performance prediction system under critical operating conditions.

[0005] The above information disclosed in the background section is only used to enhance the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] The application aims to provide a gas bearing vacuum pump performance prediction system based on a specific calculation model to solve the problems in the background art.

[0007] To achieve the above-mentioned purpose, the application provides the following technical scheme: a gas bearing vacuum pump performance prediction system based on a specific calculation model, comprising a multi-parameter fluctuation perception module, a weak feature sensitive extraction module, an abnormal coupling graph construction module, a instability precursor identification module, a prediction model regulation module and a model adaptive optimization module:

[0008] The multi-parameter fluctuation perception module acquires real-time data of temperature, flow and vibration signals, calculates the instantaneous change rate of each signal, and fuses in a preset time window to construct a multi-parameter synchronous fluctuation perception model and output a synchronous abnormality index;

[0009] The weak feature sensitive extraction module constructs a sensitive mapping based on the synchronous abnormality index, identifies a high-confidence abnormal combination area through time consistency filtering and frequency domain reconstruction, eliminates inconsistent data segments, and generates a composite interference candidate set;

[0010] The abnormal coupling graph construction module inputs the composite interference candidate set into a dynamic correlation structure construction process, establishes an abnormal resonance coupling graph based on the time offset and trend coupling degree between signals, and uses the abnormal resonance coupling graph to represent potential instability paths;

[0011] The instability precursor identification module compares the abnormal resonance coupling graph with historical working condition evolution trajectories, extracts instability precursor factor clusters, and outputs risk perception intensity values through a critical proximity evaluation model;

[0012] The prediction model regulation module constructs a feedforward regulation criterion set based on the risk perception intensity values, dynamically adjusts the input weight and output credibility of the prediction model, and improves the prediction robustness and response sensitivity to composite interference states;

[0013] The model adaptive optimization module continuously verifies the regulated prediction output and real-time data for consistency, triggers model self-update when the deviation rate exceeds a threshold, and completes local structure optimization and feature reconstruction under sample augmentation conditions to form an adaptive prediction closed loop.

[0014] Preferably, the step of constructing the multi-parameter synchronous fluctuation perception model comprises:

[0015] The temperature sensor signals installed in the main bearing part, the inner wall of the pump body shell and the gas film cavity inlet, the thermal mass flowmeter signals installed in the air inlet pipe and the air outlet connection section, and the three-axis accelerometer signals installed in the rotor support area and the side wall of the suspension module are acquired, and the sampling frequency is uniformly set to 100 Hz;

[0016] First-order low-pass filtering and normalizing the signals to the range of [0, 1];

[0017] Calculating the instantaneous change rate of each signal in a 0.01-second time interval, and extracting the mean, maximum, and fluctuation amplitude of the change rate in a 1-second sliding window with a 0.1-second step;

[0018] In a 5-second synchronization window, analyze the consistency, amplitude difference, and standard deviation of the three types of change rate sequences to determine whether they simultaneously satisfy the preset conditions, and calculate a synchronization anomaly index in a weighted manner to represent the state of the micro-amplitude linkage disturbance.

[0019] Preferably, the step of generating a composite interference candidate set comprises:

[0020] Constructing a synchronization anomaly index mapping diagram and extracting time periods with an anomaly index greater than or equal to 0.7 as potential abnormal time periods;

[0021] Filtering the potential abnormal time periods for direction consistency and amplitude consistency, and only retaining time periods where the three types of signal change rates are simultaneously direction-consistent and the normalized average value difference is less than 0.05;

[0022] Performing a fast Fourier transform on the retained time periods to select time periods with a main peak frequency difference of no more than 1 Hz and a main peak energy intensity ratio of no more than 1.5 times;

[0023] Time periods that simultaneously satisfy the above three conditions are considered as members of the composite interference candidate set, and their time start and end points, change rate statistics, and frequency spectrum information are recorded.

[0024] Preferably, the step of constructing an abnormal resonance coupling graph comprises:

[0025] Extracting the change rate sequences of temperature, flow, and vibration in the composite interference candidate time periods and unifying the time axis;

[0026] Calculating the direction consistency score of each signal pair within a ±0.5-second offset range to determine the dominant response time difference and response order;

[0027] Counting the number of point pairs that are direction-consistent and have a normalized amplitude difference value less than 0.05 at the position corresponding to the dominant response time difference, and calculating a coupling strength score;

[0028] Taking temperature, flow, and vibration as nodes in the graph, representing the response order, dominant time difference, and coupling strength as directed edges to form a graph structure, and forming an abnormal resonance coupling graph.

[0029] Preferably, the step of calculating the coupling strength score comprises:

[0030] At the position corresponding to the dominant response time difference, the difference of the normalized change rate of each data point with consistent direction in the two change rate sequences is calculated. If the difference is less than 0.05, it is recorded as an effective linkage point. The coupling strength score of the parameter pair is the ratio of the number of effective linkage points to the total number of data points, which is used to measure the linkage closeness between the two parameters in the time period.

[0031] Preferably, the output risk perception intensity value is calculated as follows:

[0032] A historical working condition trajectory library containing typical unstable evolution processes is constructed, and temperature change rate sequences, flow rate change rate sequences and vibration change rate sequences are extracted to generate standard graphs containing response time sequences, coupling strength and lag time;

[0033] The abnormal resonance coupling graph in the current operation cycle is compared with the historical standard graph to judge the similarity of the three standards of dominant response sequence, response time difference and coupling strength, and the closest historical graph is identified;

[0034] The unstable precursor factor cluster in the current graph is extracted, including the dominant disturbance path, the response time structure of the three types of signals and the coupling strength value;

[0035] The feature deviation of the current precursor factor and the reference graph is calculated, and the critical proximity value is obtained by normalization and weighted summation. The risk perception intensity value is obtained by subtracting the value from 1.

[0036] Preferably, the construction of the feedforward control criterion set includes the following steps:

[0037] The risk perception intensity value is divided into five level intervals, and a parameter set containing the input weight vector of the sensing signal, the prediction output confidence adjustment coefficient and the response sensitivity adjustment coefficient is configured for each level interval;

[0038] According to the level to which the current risk perception intensity value belongs, the corresponding input weight setting value is called to weight the temperature, flow rate and vibration three types of sensing signals, and is updated to the input end of the prediction model;

[0039] The output confidence correction parameter of the current level is called synchronously to adjust the upper and lower bounds of the confidence interval of the model prediction result, and prompt information is added;

[0040] All parameters in the risk level item are loaded into the model configuration structure to complete the feedforward control process of the prediction model.

[0041] Preferably, loading all parameters in the risk level item into the model configuration structure includes the following steps:

[0042] The sensing signal input weight vector corresponding to the current risk perception intensity value, the prediction output confidence adjustment coefficient and the model internal response sensitivity adjustment coefficient are written into the model calling interface in sequence, parameter synchronization injection is completed through unified configuration mapping relationship, it is ensured that all parameters are loaded before model prediction task initialization, and the input structure, output structure and intermediate response path of the model are affected, a risk-driven prediction feedforward control mechanism is formed.

[0043] Preferably, the continuous consistency verification of the regulated model prediction output and the real-time collected data and the triggering of the model self-updating process include the following steps:

[0044] After each prediction cycle ends, the relative deviation rate between the prediction value and the real-time collected value of the temperature, flow and vibration signals is calculated and recorded in the deviation buffer area;

[0045] The continuous deviation rate threshold of each type of signal is set, when the deviation rates of any two types of signals in the last three cycles are continuously out of limits and the directions are consistent, it is determined that the model is mismatched and the updating process is triggered;

[0046] After triggering the update, the historical sampling data containing the features before and after the mismatch are extracted, sample disturbance enhancement processing is performed to construct an extended training set, sample enhancement includes disturbance amplitude adjustment, time scale stretching and compression and historical micro-disturbance sequence embedding;

[0047] Based on the extended training set, partial optimization is performed on part of the structure and parameters of the model, the model is updated and the old model structure is covered, and the prediction and consistency verification process is re-entered.

[0048] Preferably, in the process of performing local optimization of the model, the input feature structure of the model is reconstructed, specifically including combining the change rate of each type of signal with its sliding average value for input, and introducing the first derivative of the change rate as a new feature variable, and only updating the parameters of the first two layers of feature extraction structure and the output mapping layer of the model, to reduce the risk of overfitting and reduce the calculation burden.

[0049] In the above technical solution, the technical effects and advantages provided by the present application are:

[0050] The scheme of the present application takes multi-parameter synchronous wave fluctuation perception as the starting point, first introduces weak feature sensitive extraction and abnormal resonance coupling graph construction means in the prediction mechanism, so that the system can accurately capture the hidden linkage disturbance structure and potential instability channel when multiple key operating parameters have not reached the traditional alarm threshold. By structurally comparing the identified instability precursor factor cluster with the historical evolution trajectory, the system can output a dynamic and quantifiable risk perception intensity, realizing the prediction paradigm change from "post-alarm" to "pre-warning". At the same time, through the introduction of the feedforward control mechanism and the model self-adaptive optimization mechanism, the dynamic response capability and robustness of the model under complex working conditions are significantly enhanced, avoiding the decline of prediction ability caused by external disturbance or equipment state drift, and finally building an intelligent prediction system with "continuous perception-dynamic prediction-real-time adaptation" closed loop capability, which comprehensively improves the safety, stability and operation efficiency of the gas floating vacuum pump under key working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0052] Figure 1 A module schematic diagram of the gas floating vacuum pump performance prediction system based on a specific calculation model. DETAILED DESCRIPTION

[0053] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the scope of example implementations to those skilled in the art.

[0054] The present application provides a gas floating vacuum pump performance prediction system based on a specific calculation model as shown in Figure 1 The present application provides a gas floating vacuum pump performance prediction system based on a specific calculation model as shown in

[0055] The multi-parameter wave fluctuation perception module acquires real-time data of temperature signals, flow signals and vibration signals, and calculates the instantaneous change rate of the signals. In a preset time synchronization window, the multi-parameter wave fluctuation perception module fuses the change rates of multiple signals to construct a multi-parameter synchronous wave fluctuation perception model, and outputs a synchronous abnormality index for representing the initial feature state of a micro-amplitude linkage disturbance.

[0056] To achieve early identification of the micro-amplitude linkage disturbance of the aerodynamic vacuum pump during high-load operation, the following specific steps are taken to construct a feature extraction process based on multi-parameter synchronous fluctuation perception to generate a synchronous anomaly index to represent the potential abnormal trend:

[0057] By arranging three types of sensors at key structural positions of the aerodynamic vacuum pump, real-time operation data of temperature, flow and vibration signals are obtained. Temperature data is collected by high-precision thermistor temperature sensors installed at the main bearing part, the inner wall of the pump body and the inlet of the gas film cavity. Flow data is obtained by a thermal mass flow meter installed at the connection section of the gas inlet pipe and the gas outlet. Vibration signals are collected by a three-axis accelerometer attached to the support area below the rotor and the side wall of the suspension module. All sensors sample at a frequency of 100 Hz and are synchronized by a clock synchronization chip with a timestamp to ensure consistent time reference for all collected data. The raw data collected is first preprocessed on the edge device using a first-order low-pass filter to suppress noise, with a filter cutoff frequency set to 30 Hz to retain the main low-frequency disturbance components while removing high-frequency interference from external mechanical vibrations. After filtering, linear normalization is performed on the three types of signals to unify their numerical range to the [0, 1] interval, providing standardized input for subsequent change rate calculation.

[0058] The instantaneous change rate of the preprocessed temperature signal, flow signal and vibration signal on the time axis is calculated. The change rate is calculated by dividing the difference between the current time and the previous time of each signal by the corresponding time interval, which is set to 0.01 seconds. To improve the stability of the change rate, a fixed-length sliding window is used for smoothing, with a window size of 1 second and an update step of 0.1 seconds. In each sliding window, the mean, maximum and fluctuation amplitude of all change rate data in the window are calculated and output as the characteristics of the window. The change rates of the three types of signals are saved as independent sequences and used together with the original normalized signals for subsequent joint feature analysis. This approach ensures that the original signal trend information is not lost while real-time tracking of the occurrence, intensity and duration of small disturbances is achieved.

[0059] The alignment analysis of the three types of signal change rate sequences is performed within a defined synchronization time window, which is fixed at 5 seconds, covering 500 data points each time. In this window, the three types of change rate sequences are one-to-one corresponding at each time point, and the consistency test method is used to determine whether they show consistent changes in direction at adjacent time points, i.e., whether the three signals show an increase or a decrease at a certain time. If more than two-thirds of the consecutive time points satisfy the consistency condition, and the maximum difference in numerical amplitude of each change rate is not more than 0.05 (after normalization), it is determined that there is a linkage fluctuation trend in this time window. In addition, the standard deviations of each signal in the window are compared, and if the standard deviation of the vibration signal is greater than 0.02 and the standard deviations of the temperature and flow signals are both greater than 0.015, it indicates that there is a synchronous fluctuation amplitude basis for the signals in this time period, further confirming the existence of linkage disturbance.

[0060] Based on the above analysis results, a synchronization anomaly index is constructed to quantify the intensity of linkage disturbance between signals. This index considers three aspects of indicators: first, the proportion of time points with consistent change rate direction of the three types of signals in the total time points of the window; second, the average of the normalized amplitude difference of the change rates of the three types of signals; third, the proportion of time points where the standard deviations of the three types of signals simultaneously exceed their respective preset baseline thresholds. The calculation formula of the synchronization anomaly index is the weighted sum of the three indicators, where the weight of the direction consistency is 0.4, the weight of the amplitude similarity is 0.3, and the weight of the standard deviation synchronization is 0.3, and the final result is limited in the interval [0, 1]. When the synchronization anomaly index in a certain window period exceeds 0.7, it is determined that this time period is an abnormal state with significant micro-linkage disturbance. This determination will be used to trigger further analysis and response mechanism design for the abnormal area.

[0061] The core role of this step is to realize the early identification and quantitative characterization of the multi-parameter micro-amplitude linkage disturbance of the gas floatation vacuum pump, to provide accurate and reliable input basis for subsequent performance prediction and abnormality judgment. In the high load operation process of the gas floatation vacuum pump, the system may not show obvious single parameter abnormality, but multiple key operating parameters such as temperature, flow and vibration often show slight but synchronous change trend due to factors such as material fatigue, gas film instability or load disturbance. Although this kind of linkage disturbance is not enough to trigger the traditional single variable threshold alarm mechanism in amplitude, it may be an early sign of the device gradually entering an unstable state. Therefore, through this step, real-time signal acquisition, change rate calculation and synchronous fusion are realized, which not only captures the cooperative change behavior between parameters from two dimensions of time and value, but also makes the implicit trend characteristics explicit and structured through the construction of a quantitative index, namely the synchronous abnormality index, thereby breaking through the technical bottleneck of the traditional model that "can't see, can't identify and can't quantify". In addition, the implementation of this step effectively improves the sensitivity and identification accuracy of the abnormal area in the subsequent data analysis process, ensuring that the whole prediction method has enough foresight and response capability in the early warning stage. This mechanism is especially suitable for weak disturbance detection of nonlinear dynamic systems, and has strong engineering applicability and popularization value.

[0062] The weak feature sensitive extraction module constructs a weak feature sensitive mapping based on the synchronous abnormality index, identifies the high confidence combination area of synchronous abnormality through time consistency filtering method and frequency domain reconstruction strategy, and eliminates the interference data segments with inconsistent change amplitudes and time inconsistencies, to generate a composite interference candidate set;

[0063] In order to use the synchronous abnormality index to extract the linkage micro-disturbance feature area in the operation process of the gas floatation vacuum pump, the following steps are adopted to construct a complete abnormal information extraction process to form a composite interference candidate set with high confidence, as the basis data source for subsequent identification and judgment. Specifically, the following steps are included:

[0064] According to the time sequence output of the synchronous abnormality index, a two-dimensional numerical mapping covering the entire running time is constructed. The horizontal axis of the mapping is the sampling time point, with a time interval of 0.01 seconds; the vertical axis is the synchronous abnormality index corresponding to the time point, and the numerical range is fixed between 0 and 1. Time points with abnormality index greater than or equal to 0.7 are marked as potential abnormal points. In the entire mapping, if the time interval between two consecutive potential abnormal points is less than 0.5 seconds, it is determined that they belong to the same abnormal time period, and the start and end time, corresponding abnormality index, time length and other information of the time period are recorded to generate an initial potential abnormal time period set. This processing method ensures that isolated high value points are merged into continuous event segments with more engineering significance, avoiding interference caused by accidental jitter.

[0065] The consistency filtering operation is performed on the above-mentioned potential abnormal time period to verify whether the three types of signals exhibit a sufficient degree of synchronous behavior within the time period. For each potential time period, the temperature change rate sequence, the flow rate change rate sequence, and the vibration change rate sequence are extracted respectively. For each time point, it is determined whether the three change rates are positive at the same time or negative at the same time, and if so, it is recorded as a direction consistent point. The number of direction consistent points in the time period is counted, and if the proportion exceeds 60% of all time points, it is determined that the time period meets the change direction consistency. At the same time, the average values of the three types of change rate sequences in the time period are calculated respectively, and the difference between any two average values is calculated. If the difference between any two average values is less than 0.05 (after normalization), it is determined that the amplitude change is consistent. Only when a time period meets the above conditions in both direction consistency and amplitude consistency, the time period is retained as an effective synchronous abnormal time period, otherwise it is excluded.

[0066] The frequency domain feature reconstruction is performed on the retained synchronous abnormal time period to further exclude atypical disturbances with obvious differences in frequency characteristics. Specifically, the fast Fourier transform is performed on the three types of change rate sequences in each time period to extract the frequency spectrum distribution in the frequency range of 1 Hz to 10 Hz. In each frequency spectrum, the position of the main peak with the maximum frequency energy is found, and the main peak frequency value and energy intensity are recorded. Then, the main peak frequencies of the three groups of signals are compared with each other, and if the difference between any two main peak frequencies is not more than 1 Hz and the corresponding energy intensity ratio is not more than 1.5 times, it is considered that the time period has consistency in frequency structure. Otherwise, the time period is excluded and does not enter the next processing flow. The purpose of this frequency domain reconstruction step is to ensure that the linkage disturbance not only has synchronization in time, but also exhibits coordination in energy release and periodic behavior, so as to enhance the discrimination accuracy of the disturbance.

[0067] The time periods that pass through the time consistency filtering and frequency domain feature reconstruction double verification are recorded as members of the composite disturbance candidate set. Each member includes the start time, the end time, the time period duration, the average change rate values of the three types of signals, the standard deviation, the direction consistency ratio, the main peak frequency and its intensity information. All data are stored in the cache structure of the processing unit in a unified format for subsequent feature modeling and correlation analysis module calling. The set not only provides numerical support at the data level, but also retains the structure characteristics and frequency behavior of the time sequence, facilitating reliable instability trend identification, factor extraction and risk reasoning in subsequent operations.

[0068] The core role of this step is to convert the synchronization anomaly index identified in the previous stage into a structured, operable and high-confidence abnormal data region, thereby providing basic data support for subsequent anomaly evolution identification and dynamic trend inference. During the operation of the gas floatation vacuum pump, multiple parameters such as temperature, flow and vibration may exhibit coordinated changes in a short period of time due to system perturbations, which can be initially perceived through the synchronization anomaly index. However, a single index cannot directly provide a data structure that can be used for trend modeling or instability channel construction. Therefore, this step further verifies whether the changes among multiple parameters occur simultaneously, are in the same direction, and have similar amplitudes by constructing a two-dimensional mapping, performing time consistency filtering and frequency domain feature reconstruction, thereby eliminating false positives caused by isolated fluctuations or noise interference and retaining time periods that truly reflect the potential abnormal trends of the system. The final composite interference candidate set not only has multiple dimensional features such as time, value, and frequency, but also has clear boundaries and stable expression forms, which can provide precise input for subsequent parameter correlation structure construction, instability factor identification, and model regulation mechanism. This process realizes the effective conversion from the original abnormal perception results to data blocks that can be used for modeling, is a key bridge link for intelligent prediction and accurate intervention, and significantly enhances the adaptive ability and engineering practicality of the prediction system to weak abnormal states.

[0069] The abnormal coupling graph construction module inputs the composite interference candidate set into the dynamic correlation structure construction process, establishes a cross-parameter abnormal resonance coupling graph based on the relative time offset pattern and coupling degree of change trends between different signals, and uses it to represent the correlation structure of the potential instability path;

[0070] In order to extract the dynamic linkage relationship between the three key parameters (temperature, flow and vibration) from the identified composite interference candidate set, and establish an abnormal propagation structure with time sequence, response direction and coupling strength characteristics for revealing the potential instability path of the gas floatation vacuum pump during the disturbance evolution process, an abnormal resonance coupling graph based on actual time response offset and fluctuation trend consistency is proposed. This process includes the following steps:

[0071] For each composite interference candidate time period, extract the temperature rate of change sequence, flow rate of change sequence and vibration rate of change sequence within that time period, with a time resolution of 0.01 seconds and a total length of 500 sampling points, corresponding to a 5-second operation process. In order to facilitate alignment and comparison, all data are synchronized and calibrated with the starting point of the time period as the unified reference zero point, i.e. the time axes of all signals are unified to start at 0 seconds and end at 5 seconds, while retaining their original sampling frequency and amplitude information, without time scaling or interpolation processing.

[0072] The relative response order between the three types of change rate sequences is determined. With the temperature change rate sequence as the fixed reference, the flow change rate sequence and the vibration change rate sequence are sequentially translated within a time offset range of 0 to ±0.5 seconds, and the direction consistency ratio at each offset point is calculated. The direction consistency ratio is calculated by counting the number of data points with the same change direction at the alignment time point, and dividing the total number of data points to obtain the consistency score. The consistency score of each pair of signals at all offset positions is recorded, and the position with the highest score is determined as the dominant response time difference of the signal pair. For example, if the vibration change rate has the highest direction consistency score relative to the temperature change rate at a delay of 0.3 seconds, it is recorded that the temperature responds first and the vibration responds later, and the dominant response time difference is 0.3 seconds.

[0073] After determining the dominant response time difference between each pair of parameters, the change trend coupling relationship is further analyzed. The number of point pairs with the same direction change in the two change rate sequences is re-extracted at the corresponding time offset position, and the change amplitude difference is analyzed. Specifically, at each direction consistent point, the change rate difference of the two signals is calculated, and if the difference is less than 0.05 (after normalization), it is considered that the two have trend amplitude consistency at that point. The number of data points with trend direction consistency and amplitude difference less than 0.05 is counted, and the ratio with the total number of points is taken as the coupling strength score of the signal pair. The score reflects the degree of linkage between the two parameters during the disturbance process.

[0074] The response direction, dominant time difference and coupling strength between each pair of parameters obtained above are represented in a graph structure. The graph includes three nodes corresponding to the temperature, flow and vibration parameters. The directed line between any two nodes represents the response order between the parameters, with the arrow pointing from the first responding parameter to the second responding parameter. The thickness of the line represents the level of coupling strength, with thicker lines indicating tighter linkage. The value marked next to the line represents the dominant response time difference in seconds. During a disturbance period, if the temperature parameter responds first to both the flow and vibration parameters and has a high coupling strength, it can be inferred that the temperature is a possible trigger source for the disturbance period. If the vibration parameter is the end of the response multiple times and has a high coupling strength, it indicates that the vibration has high sensitivity to system disturbance. By collecting and analyzing the graphs of multiple disturbance periods, the main abnormal propagation path, influence priority order and potential instability link of the device under different operating states can be obtained.

[0075] The core role of this step is to structure the timing response relationship and trend linkage between the three key operating parameters in the identified complex interference period, thereby establishing an abnormal resonance coupling graph that reflects the potential instability mechanism of the gas floatation vacuum pump, providing an interpretable and high-value basic data structure for subsequent early warning factor extraction, trend evolution modeling, and risk judgment. During the high-load or long-time operation of the gas floatation vacuum pump, multiple physical variables (such as temperature, flow, and vibration) inside the device often do not abnormally change alone, but interact with each other in the form of micro changes and gradually accumulate risks. If the dynamic relationship between these parameters cannot be revealed from the time and signal linkage dimension, it is difficult to accurately grasp the early signs of the system's evolution from stability to instability. Therefore, this step identifies which parameter changes first, whether other parameters respond, and how strong the response is by synchronously calibrating the change rate of each parameter in the perturbation period, identifying the response time difference, and analyzing the trend coupling degree. It also displays the entire disturbance propagation path in a clear directed relationship and intensity label through a graph structure. This structured graph enables engineering and technical personnel to intuitively identify which parameter is the main driving factor and which parameter is the most sensitive response item under a specific disturbance state, thereby clearly identifying the potential instability "trigger path" and "disturbed path" in the system. Compared with traditional single-parameter threshold alarm mechanisms or static feature recognition methods, this graph construction method has dynamic, directional, and causal structure expression capabilities, which can significantly improve the prediction system's understanding of complex coupled disturbances and forward-looking judgment capabilities. It is a key step to achieve multi-parameter early warning and risk intervention.

[0076] The instability precursor identification module compares the abnormal resonance coupling graph with the historical working condition critical evolution trajectory, identifies and extracts a multi-parameter instability precursor factor cluster, calculates the relative distance between the system state and the instability boundary through a critical proximity evaluation model, and outputs a real-time risk perception intensity value.

[0077] To realize real-time perception of the instability risk of the gas floatation vacuum pump under complex working conditions, the abnormal resonance coupling graph constructed in the previous step is compared with the established historical working condition critical evolution trajectory to identify the precursor signal characteristics, and the proximity between the current state and the instability state is calculated based on the preset judgment standard, finally outputting the risk perception intensity value. This process includes the following steps:

[0078] A historical trajectory library containing typical unstable evolution processes is constructed. The trajectory library is constructed based on actual engineering operation records, and the data is obtained from long-term monitoring of the entire process of the same type of gas floatation vacuum pump, covering various operating states, including normal operation state, mild fluctuation state, near instability state, and actual instability state. For each operating cycle, the change rate sequence of three key parameters, temperature, flow rate, and vibration, is extracted, and the entire process from the beginning of the anomaly to complete instability is clearly marked. With every 5 seconds as an analysis window, the change trend, response time sequence, coupling strength, and lag time of each parameter in these data windows are encapsulated into standard maps, and a one-to-one mapping relationship with the corresponding working condition category is established as a historical reference template, which is saved in the trajectory library for real-time comparison.

[0079] The abnormal resonance coupling map generated in the current operating stage is extracted and compared with all standard maps in the historical trajectory library. During the comparison process, three fixed standards are used for comparison and judgment: first, whether the dominant response order among the temperature change rate, flow rate change rate, and vibration change rate in the current map is the same as that in a certain historical map, for example, whether the temperature is the first changing parameter, the flow rate is the intermediate response, and the vibration is the end response; second, the response time difference between each parameter in the current map and the historical map is calculated, and if the difference is less than 0.2 seconds, it is considered that the response time sequence has high consistency; third, the coupling strength is compared, and if the coupling strength difference between the same pair of parameters in the two maps is not more than 0.15, it is considered that the connection relationship is stable. If the current map and a certain historical map meet at least two of the above three standards, it is considered that they have high similarity, and the instability stage corresponding to the historical map is taken as the reference stage of the current state.

[0080] According to the matching result with the highest similarity to the historical instability map, a set of instability precursor factors is extracted from the current map. This factor set includes three key elements: first, the dominant disturbance path, i.e., the direction relationship of disturbance propagation among temperature, flow rate, and vibration in the current time period; second, the response time structure, specifically the order and duration of response delay between parameters, with the maximum lag time between the three as the index; third, the numerical representation of coupling strength, recording the linkage strength in each parameter connection, ranging from 0 to 1, representing the disturbance transmission ability. All the precursor factors are packaged into structured data packets, containing parameter pairs, response direction, response time difference, and coupling strength, and saved in the precursor factor record area of the current operating cycle for further risk level assessment.

[0081] The extracted precursor factor cluster is used for feature deviation calculation with historical instability reference factors to obtain the relative distance between the current operating state and the instability boundary, and output the risk perception intensity value. The specific implementation manner is as follows: first, the deviation amount of three feature dimensions is calculated, that is, the current path order and the reference path consistency result (taking 0 or 1), the difference between the current response time difference and the reference response time difference, and the difference between the current coupling strength and the reference coupling strength; then, the three deviation values are normalized to the interval of 0 to 1, and the normalization standard is set to be that the response time difference deviation of 0.5 seconds is the upper limit, and the coupling strength difference of 0.2 is the upper limit; finally, the three normalized indexes are weighted and summed according to the weight to obtain the critical proximity value, and the risk perception intensity is obtained by subtracting the proximity value from 1. The higher the risk perception intensity, the closer the system is to the instability state. The value is updated every 5 seconds during operation, which can reflect the deviation degree of the system operating state compared with the historical instability experience model in real time, and can be used as a direct input for the prediction model regulation and the early warning triggering.

[0082] The main role of this step is to deeply compare the multi-parameter linkage structure revealed by the abnormal resonance coupling graph with the evolution trajectory of the instability that has occurred in the historical working condition, so as to identify the instability trend characteristics with high similarity, and evaluate the proximity between the current system operating state and the potential instability boundary, and finally output a quantifiable and operable risk perception intensity value. In the gas bearing vacuum pump and other complex devices that rely on high-speed rotation, non-contact suspension and high-precision gas film control, the system performance is often affected by multiple physical parameters, and the instability process is usually not caused by the mutation of a single parameter, but by the coupling evolution of multiple key variables in time and response path. Therefore, it is difficult to identify the progressive degradation or hidden failure trend of the system in advance by simply relying on instantaneous value or single variable abnormality. This step builds a historical trajectory library containing typical pre-instability evolution characteristics, and compares the current graph with the historical instability graph in response order, linkage strength, time delay and other dimensions, and then extracts a group of precursor factor clusters with the same structure mode. These factor clusters can be regarded as “warning signals” before the system gradually loses stability. On this basis, the multi-dimensional deviation between the current state and the historical instability state is calculated to form a quantitative “critical proximity”, which is inversely mapped to a “risk perception intensity value”. This value has high real-time and sensitivity, and can reflect whether the system state is approaching the risk boundary, and provide scientific decision basis for the dynamic regulation of the prediction model, the level adjustment of the early warning system and the immediate intervention of the operation and maintenance strategy. Compared with traditional rule judgment or static model output, this method significantly improves the identification depth and response speed of the performance prediction system to multi-parameter weak coupling disturbance, and is a key component of building a high-robustness and high-precision prediction mechanism.

[0083] The prediction model regulation module constructs a set of feedforward regulation criteria according to the risk perception intensity value, dynamically adjusts the input weight of various types of sensing signals in the performance prediction model and the confidence level of the model output, and improves the prediction response sensitivity and anti-interference robustness of the model under complex interference state.

[0084] To realize the dynamic adaptation and stable output of the performance prediction model of the air floating vacuum pump under the complex interference state, an implementation based on the risk perception intensity value to construct a set of feedforward regulation criteria is proposed to improve the response sensitivity and anti-interference ability of the prediction model in the early stage of instability trend. The process includes the following steps:

[0085] The risk perception intensity value output by the instability precursor identification process is received in real time. The risk perception intensity value is a floating point number, with a value range of 0 to 1, which is used to reflect the closeness of the current air floating vacuum pump operating state to the historical instability boundary. The risk perception intensity value is divided into five fixed intervals: when the risk perception intensity value is in [0, 0.2), it is defined as "very low risk"; when it is in [0.2, 0.4), it is defined as "low risk"; when it is in [0.4, 0.6), it is defined as "medium risk"; when it is in [0.6, 0.8), it is defined as "high risk"; when it is in [0.8, 1.0], it is defined as "high risk". Each risk level corresponds to a set of pre-set regulation parameters, which includes three contents: one is the input feature weight value of temperature, flow and vibration three types of sensing signals; the second is the model output confidence adjustment coefficient; the third is the model internal response sensitivity adjustment coefficient, which corresponds one by one and cannot be omitted.

[0086] According to the current risk level, the corresponding sensing signal input weight setting value is selected and called to dynamically update the input feature weighting structure of the performance prediction model. For example, when the current risk perception intensity value is 0.82, which belongs to the "high risk" level, the system will automatically call the input weight setting scheme corresponding to the "high risk level". Under this scheme, the vibration signal input weight is set to 0.40, the flow signal input weight is set to 0.35, and the temperature signal input weight is set to 0.25. The input weight is normalized in vector form, and the sum is kept as 1 to prevent the deviation of the model input caused by the proportional overrun of various types of signals. The updated weight vector directly acts on the feature input channel in the forward propagation structure of the model, ensuring that the current data is adjusted by weight before entering the model calculation, and strengthening the signal features highly related to the current risk level.

[0087] The synchronous call matches the model output confidence correction strategy with the current risk level. The correction strategy is based on the confidence correction coefficient corresponding to the risk level, and adjusts the prediction confidence interval attached to the model output result. For example, in the "high risk" state, the system reduces the lower bound of the confidence interval by 5%, expands the upper bound by 3%, and adds the "possible instability" label to remind the system operation and maintenance personnel to pay attention to the volatility of the prediction result. In the "medium risk" state, the system uses the standard confidence output, and the confidence interval remains the original range and does not need to be corrected. If it is in the "low risk" or "very low risk" state, the system will append the "running stable" prompt to the output result, but still retain the confidence interval for subsequent consistency verification. Through this dynamic confidence adjustment mechanism, the expression form and response mechanism of the prediction model output information under different risk levels are different, which enhances the actual reference value of the prediction output.

[0088] A set of feedforward control criteria indexed by risk perception intensity values is constructed and maintained. The criterion set consists of five fixed risk level entries, each containing specific sensor signal input weight vectors, confidence correction coefficients, and internal response sensitivity parameters, and is saved in the configuration database of the system control center in a structured format. The criterion set is automatically updated every 30 days, and during the update process, the system automatically corrects the control parameters corresponding to each risk level by analyzing the prediction error distribution, actual instability event occurrence rate, and prediction accuracy indicators under different parameter combinations in the past 30 days. Each time the model starts a prediction process, the system will first read the current risk perception intensity value and query the corresponding control entry, and then load all parameters in the entry into the model configuration structure. This feedforward control mechanism realizes a complete closed-loop adjustment process from risk identification to prediction response, significantly improving the sensitive response ability and prediction stability of the model under multiple disturbances and variable working conditions.

[0089] The core role of this step is to realize the dynamic response regulation and prediction stability enhancement of the gas float vacuum pump performance prediction model in the face of complex disturbance conditions. Specifically, when the gas float vacuum pump is running for a long time or approaching the limit load state, the temperature, flow and vibration signals inside the system may appear short-term linkage disturbance. This disturbance is often outside the traditional threshold and cannot be directly identified as a loss of stability risk. In this context, by building a feedforward control criterion set based on the "risk perception intensity value", the input feature weight of three key sensing signals can be dynamically adjusted according to different risk levels, so that the model pays more attention to the operating parameters that have the greatest impact at the current risk level. For example, at a high risk level, appropriately increasing the proportion of vibration signals and flow signals can more sensitively capture structural disturbances and gas channel changes. At the same time, the model output confidence interval can also be adjusted to amplify abnormal trends early, thereby improving the model's alertness to boundary instability trends. In addition, this step also has adaptive ability, which can adjust the model's internal structure and response mechanism by calling the corresponding parameter items of the risk level in real time, ensuring that the prediction results remain highly reliable under sudden working conditions. This mechanism not only breaks the limitations of traditional prediction models "static input, passive response", but also establishes a "risk-driven-model reconstruction" feedforward logic path, which enhances the robustness, adaptability and practicality of the model from the source. It is the key technical basis for realizing the intelligent and high-sensitivity early warning function of the prediction system.

[0090] The model self-adaptive optimization module continuously verifies the regulated model prediction output and real-time collected data for consistency. When the continuous deviation rate exceeds the preset adaptive threshold, the model self-update process is triggered, and the local structure optimization and feature parameter reconstruction of the model are performed under the condition of sample augmentation containing disturbance characteristics, completing the prediction closed loop with adaptive ability.

[0091] To ensure the long-term stability and accuracy of the gas float vacuum pump performance prediction, the prediction model after risk regulation needs to be continuously verified for consistency. When the model prediction results and the real-time collected data of the device continuously produce significant deviations within a certain time range, the system should automatically start the model update process and perform local structure reconstruction and parameter optimization after adding data samples containing disturbance characteristics, to realize adaptive adjustment to complex working condition changes. This process mainly includes the following steps:

[0092] At the end of each 5-second prediction cycle, the predicted values of the model output, including the temperature rate of change, the flow rate of change, and the vibration rate of change, are recorded. At the same time, the actual rate of change values of temperature, flow, and vibration are obtained by real-time acquisition of sensor data at the current time of the device. Based on a unified timestamp, the predicted values are compared one by one with the corresponding actual acquisition values. The calculation method uses the standard difference method, that is, the predicted value is subtracted from the actual value and then divided by the actual value to obtain the relative deviation rate of each parameter. The absolute value of the deviation rate value is used for subsequent consistency judgment. In each 5-second prediction cycle, the average deviation rate of the three parameters is calculated and recorded in the continuous deviation buffer area of the system as the basis data for continuous judgment.

[0093] The maximum continuous deviation rate threshold within the system allowable range is set to determine whether the model deviates from the actual operating state. In this embodiment, the maximum continuous deviation rate allowed for the temperature rate of change is ±10%, the flow rate of change is ±12%, and the vibration rate of change is ±15%. If the deviation rate of at least two parameters continuously exceeds the corresponding threshold for three consecutive prediction cycles (total duration of 15 seconds) and the deviation direction is consistent (for example, both positive or both negative), the system determines that the current model is no longer suitable for the existing working condition, triggering the model update process. This judgment method avoids frequent updates caused by occasional errors and enhances stability.

[0094] After entering the model update process, the system automatically extracts update samples from the historical acquisition data of the last 60 seconds, ensuring that the selected data contains feature change information of the three stages before, during, and after the model deviation occurs. Each update sample includes the original value of the rate of change of the three parameters, the previous predicted value, the actual value deviation rate, the system operating state label, and the disturbance time marker. In order to enhance the learning ability of the model to weak disturbances and atypical states, sample enhancement processing is performed on the original samples. The specific processing methods include: adding a disturbance amplitude of ±5% to the existing rate of change sequence and re-normalizing; extending the time sequence to 1.5 times its length or compressing it to 0.7 times its length while maintaining the relative structure; adding a filtered historical disturbance sequence to some data segments to form a new training scenario. After the sample enhancement process is executed, 8 variant versions of the original sample are generated, which are combined into an expanded training set and provided to the model update.

[0095] Based on the new training set constructed by the augmented samples, the local structure optimization and parameter update of the model are performed. The local structure optimization includes redefining the feature combination mode of each input parameter in the model, such as replacing the original input structure with a single value of the change rate with a combination input of the change rate and its moving average, or introducing the change rate of the change rate as a new input variable; the output layer structure remains unchanged. In the parameter update stage, only the first two layers and the final output mapping layer of the model used for feature extraction are retrained to avoid the overfitting risk and computational overhead brought by full model reconstruction. After the model optimization is completed, the old model structure is automatically covered, and it serves as the running model for the next round of prediction process, and at the same time the system enters a new round of continuous consistency verification cycle, completing the closed-loop process of prediction, feedback, correction and re-prediction.

[0096] The core role of this step is to realize the closed-loop construction of the model adaptive ability in the performance prediction process of the gas floatation vacuum pump, ensuring that the prediction model can continuously maintain high precision, strong robustness and response sensitivity under complex operating conditions and dynamic changes of working conditions. With the passage of time of the gas floatation vacuum pump, the device state may change slightly but continuously due to factors such as temperature drift, structural micro-change, gas film change or external disturbance, resulting in a deviation between the input distribution applicable to the initial training of the model and the current actual operating data. If this deviation cannot be identified and corrected in time, it will directly affect the prediction accuracy and early warning ability of the model, and even output misleading results at critical moments. Therefore, this step continuously verifies the consistency of the model prediction output and the real-time collected data within a time window, monitors the deviation degree of the prediction performance in real time, and sets a fault tolerance threshold. When the continuous deviation rate exceeds the set range, the model self-update process is triggered immediately to ensure that the prediction results are consistent with the actual behavior of the device. Further, by introducing extended samples containing perturbation features for sample augmentation, and optimizing the feature extraction mode and weight parameters of the model within the local structure range, the adaptability of the model to the current working condition is improved, and its recognition ability to new disturbance patterns is also enhanced. Finally, through the closed-loop process of "prediction-verification-update-optimization", the model has the ability of self-evolution, significantly improves the reliability and long-term availability of the entire prediction system in actual engineering scenarios, and is particularly suitable for complex devices such as gas floatation vacuum pumps that rely on multi-parameter dynamic cooperative operation, providing a solid guarantee for system-level intelligent prediction.

[0097] The above-mentioned gas float vacuum pump performance prediction system based on a specific calculation model can effectively solve the problems in the prior art that multiple parameter micro-amplitude cooperative disturbance cannot be identified, early instability trend perception ability is insufficient, and model misjudgment and missed judgment frequently occur. The present application scheme takes multi-parameter synchronous fluctuation perception as the starting point, first introduces weak feature sensitive extraction and abnormal resonance coupling graph construction means in the prediction mechanism, so that the system can accurately capture the hidden linkage disturbance structure and potential instability channel when multiple key operating parameters have not reached the traditional alarm threshold. By structurally comparing the identified instability precursor factor cluster with the historical evolution trajectory, the system can output a dynamic and quantifiable risk perception intensity, realizing the prediction paradigm change from "post-alarm" to "pre-warning". At the same time, through the introduction of the feedforward control mechanism and the model adaptive optimization mechanism, the dynamic response ability and robustness of the model under complex working conditions are significantly enhanced, avoiding the decline of the prediction ability caused by external disturbance or equipment state drift, and finally building an intelligent prediction system with "continuous perception-dynamic prediction-real-time adaptation" closed loop capability, which fully improves the safety, stability and operation efficiency of the gas float vacuum pump under key working conditions.

[0098] The above only describes certain exemplary embodiments of the present application by way of illustration, and it is needless to say that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present application. Therefore, the above drawings and descriptions are illustrative in nature and should not be understood as limiting the scope of protection of the claims of the present application.

Claims

1. A gas bearing vacuum pump performance prediction system based on a specific computational model, characterized by, The method comprises a multi-parameter fluctuation perception module, a weak feature sensitive extraction module, an abnormal coupling graph construction module, an instability precursor identification module, a prediction model regulation module, and a model adaptive optimization module. The multi-parameter fluctuation perception module acquires real-time data of temperature, flow rate and vibration signals, calculates the instantaneous change rate of each signal, and fuses them within a preset time window to construct a multi-parameter synchronous fluctuation perception model and output a synchronous abnormality index. The weak feature sensitive extraction module constructs a sensitive mapping based on the synchronous abnormality index, identifies high-confidence abnormal combination areas through time consistency filtering and frequency domain reconstruction, eliminates inconsistent data segments, and generates a composite interference candidate set. The abnormal coupling graph construction module inputs the composite interference candidate set into a dynamic correlation structure construction process, establishes an abnormal resonance coupling graph based on the time offset and trend coupling degree between signals, and represents the potential instability path. The instability precursor identification module compares the abnormal resonance coupling graph with the historical working condition evolution trajectory, extracts instability precursor factor clusters, and outputs risk perception intensity values through a critical proximity evaluation model. The prediction model regulation module constructs a feedforward regulation criterion set based on the risk perception intensity values, dynamically adjusts the input weight and output credibility of the prediction model, and improves the prediction robustness and response sensitivity to composite interference states. The model adaptive optimization module continuously verifies the regulated prediction output and real-time data for consistency, triggers model self-update when the deviation rate exceeds the threshold, and completes local structure optimization and feature reconstruction under sample augmentation conditions to form an adaptive prediction closed loop.

2. The gas bearing vacuum pump performance prediction system based on a specific computational model according to claim 1, characterized in that, The steps of constructing a multi-parameter synchronous fluctuation perception model include: Acquiring temperature sensor signals installed on the main bearing part, the inner wall of the pump body shell, and the inlet of the gas film cavity, thermal mass flowmeter signals installed on the inlet pipe and the connection section of the outlet, and three-axis accelerometer signals installed on the rotor support area and the side wall of the suspension module, and unifying the sampling frequency to 100 Hz; First-order low-pass filtering and normalizing the signals of various sensors to the range [0, 1]; Calculating the instantaneous change rate of each signal at a 0.01-second time interval, and extracting the mean, maximum, and fluctuation amplitude of the change rate with a 1-second sliding window and a 0.1-second step; Within a 5-second synchronization window, analyze the consistency, amplitude difference, and standard deviation of the three types of change rate sequences to determine whether they simultaneously satisfy the preset conditions, and calculate the synchronous abnormality index in a weighted manner to represent the micro-amplitude linkage disturbance state.

3. The system for predicting the performance of an aerodynamic vacuum pump based on a specific computational model according to claim 1, characterized in that, The steps of generating a composite interference candidate set include: Constructing a synchronous abnormality index mapping and extracting time periods with abnormality indexes greater than or equal to 0.7 as potential abnormal time periods; Filtering the potential abnormal time periods for direction consistency and amplitude consistency, and only retaining time periods where the three types of signal change rates are directionally consistent and the normalized average value difference is less than 0.05; Performing a fast Fourier transform on the retained time periods, and selecting time periods with a main peak frequency difference of no more than 1 Hz and a main peak energy intensity ratio of no more than 1.5 times; Recording the time start and end points, change rate statistics, and frequency spectrum information of time periods that simultaneously satisfy the above three conditions as members of the composite interference candidate set.

4. The system for predicting the performance of an aerodynamic vacuum pump based on a specific computational model according to claim 1, characterized in that, The steps of constructing the abnormal resonance coupling graph include: Extracting the temperature, flow rate and vibration rate of change sequences in the composite interference candidate time period and unifying the time axis; Calculating the direction consistency score of each signal pair within a ±0.5 second offset range to determine the dominant response time difference and response order; Counting the number of points with consistent direction and normalized amplitude difference less than 0.05 at the position corresponding to the dominant response time difference, and calculating the coupling strength score; Taking temperature, flow rate and vibration as nodes in the graph, and taking response order, dominant time difference and coupling strength as directed edges to form a graph structure, and forming an abnormal resonance coupling graph.

5. The gas foil vacuum pump performance prediction system based on a specific computational model according to claim 4, characterized in that, The steps of calculating the coupling strength score include: At the position corresponding to the dominant response time difference, for each data point in the two rate of change sequences that are consistent in direction, calculate the difference in normalized rate of change. If the difference is less than 0.05, it is recorded as an effective linkage point. The ratio of the number of effective linkage points to the total number of data points is taken as the coupling strength score of the corresponding signal combination, which is used to measure the linkage closeness of the corresponding signal combination in the current time period.

6. The gas bearing vacuum pump performance prediction system based on a specific computational model according to claim 1, characterized in that, The steps of outputting the risk perception intensity value are as follows: Constructing a historical working condition trajectory library containing typical instability evolution processes, extracting temperature rate of change sequences, flow rate of change sequences and vibration rate of change sequences, and generating standard graphs containing response time order, coupling strength and lag time; Compare the abnormal resonance coupling graph in the current running period with the historical standard graph to judge the similarity of the three standards of dominant response order, response time difference and coupling strength, and identify the closest historical graph; Extracting the instability precursor factor cluster in the current graph, including the dominant disturbance path, the response time structure of the three types of signals and the coupling strength value; Calculate the feature deviation of the current precursor factor and the reference graph, and get the critical proximity value by normalization and weighted summation. The risk perception intensity value is obtained by subtracting this value from 1.

7. The system for predicting the performance of an aerodynamic vacuum pump based on a specific computational model according to claim 1, characterized in that, The steps of constructing the feedforward control criterion set include: Divide the risk perception intensity value into five level intervals, and configure a parameter set containing the input weight vector of the sensing signal, the confidence adjustment coefficient of the prediction output and the response sensitivity adjustment coefficient for each level interval; According to the level to which the current risk perception intensity value belongs, call the corresponding input weight setting value, weight the temperature, flow rate and vibration three types of sensing signals, and update to the input end of the prediction model; Synchronously call the output confidence correction parameter of the current level to adjust the upper and lower bounds of the confidence interval of the model prediction result, and add prompt information; Load all parameters in the risk level item into the model configuration structure to complete the prediction model feedforward control process.

8. The gas foil bearing vacuum pump performance prediction system based on a specific computational model according to claim 7, characterized in that, The steps of loading all parameters in the risk level item into the model configuration structure include: Write the sensing signal input weight vector, the prediction output confidence adjustment coefficient and the model internal response sensitivity adjustment coefficient corresponding to the current risk perception intensity value into the model calling interface in turn, complete the parameter synchronous injection through the unified configuration mapping relationship, ensure that all parameters are loaded before the model prediction task is initialized, and act on the input structure, output structure and intermediate response path of the model, forming a risk-driven prediction feedforward control mechanism.

9. The system for predicting the performance of an aerodynamic vacuum pump based on a specific computational model according to claim 1, characterized in that, The model prediction output after regulation is continuously verified with real-time collected data for consistency and triggers a model self-updating process, including the following steps: After each prediction cycle, the relative deviation rate between the predicted and real-time collected values of temperature, flow rate, and vibration signals is calculated and recorded in the deviation buffer; The continuous deviation rate threshold of each type of signal is set. When the deviation rates of any two types of signals in the last three cycles are continuously out of limits and consistent in direction, the model is determined to be mismatched and the updating process is triggered; After the updating is triggered, the historical sampling data containing the features before and after the mismatch are extracted, and sample perturbation enhancement processing is performed to build an extended training set. Sample enhancement includes perturbation amplitude adjustment, time scale stretching and compression, and historical micro-perturbation sequence embedding; Based on the extended training set, the partial structure and parameters of the model are locally optimized, the model is updated, and the old model structure is overwritten, and the prediction and consistency verification process is re-entered.

10. The gas foil vacuum pump performance prediction system based on a specific computational model according to claim 9, characterized in that, During the model local optimization process, the model input feature structure is reconstructed, specifically including combining the change rate of each type of signal with its sliding average value for input, and introducing the first derivative of the change rate as a new feature variable. At the same time, only the first two layers of feature extraction structure and the output mapping layer of the model are updated to reduce the risk of overfitting and reduce the computational burden.

Images