A method and system for predicting air pump failures based on multi-source data

Abstract

This invention relates to the field of air pump failure prediction technology, and discloses a method and system for air pump failure prediction based on multi-source data. The method includes collecting vibration, temperature, pressure, gas flow, and environmental characteristic data, performing time-series alignment and standardization, extracting a polyvariant index, and enhancing and modeling its weak, continuous drift. Combined with anomaly confidence calculation and multi-dimensional data verification, it achieves reliable prediction of early cylinder seal failure and generates a real-time monitoring report containing abnormal time points and characteristic changes. This method can identify the continuous drift characteristics of the polyvariant index under noise and operating condition fluctuations, achieving reliable prediction of early cylinder seal failure.

Description

Technical Field

[0001] This invention relates to the field of air pump fault prediction technology, and in particular to an air pump fault prediction method and system based on multi-source data. Background Technology

[0002] As a key power component in industrial production, automotive repair, medical equipment, and various pneumatic systems, the operating status of air pumps directly affects the safety and continuity of the entire system. A failure in an air pump often leads to production interruptions, equipment performance degradation, and even safety hazards. Therefore, continuous monitoring of the air pump's operating status and effective early warning before a failure occurs are crucial for ensuring reliable equipment operation.

[0003] Currently, air pump condition monitoring largely relies on IoT chips to perform threshold judgments or simple trend analyses on single physical quantities such as vibration, temperature, or pressure. With technological advancements, several fault management solutions based on multi-source data monitoring and analysis have emerged. For example, deploying sensor networks to collect vibration, temperature, and pressure signals, combined with predictive models, allows for condition assessment and early warning (e.g., CN120974145B, US20230407863A1). However, these existing solutions primarily target already obvious anomalies. During long-term operation, early degradation processes such as cylinder seal wear and micro-leakage in air pumps manifest as slow and minimal changes. These anomalies are easily masked by environmental noise and operating condition fluctuations, leading existing monitoring methods to often trigger alarms only after the fault has progressed to a more severe stage, making true early warning difficult.

[0004] Existing technologies cannot reliably predict early failure of cylinder seals. Summary of the Invention

[0005] This invention provides a method and system for predicting air pump failures based on multi-source data, which can identify the continuous drift characteristics of polytropic indices under noise and operating condition fluctuations, and achieve reliable prediction of early cylinder seal failure.

[0006] Firstly, to address the aforementioned technical problems, this invention provides a method for predicting air pump failures based on multi-source data, comprising: collecting environmental feature data and simulated signals including vibration, temperature, pressure, and gas flow rate during air pump operation; performing time series alignment on the simulated signals to obtain an initial operating state dataset; performing wavelet decomposition on the initial operating state dataset to remove noise, obtaining a reconstructed signal sequence; performing local weighted regression scatter smoothing processing on the reconstructed signal sequence to obtain a standardized dataset; performing feature enhancement processing on the small amplitude changes of polytropic exponential drift based on the standardized dataset to obtain a feature enhancement sequence; performing feature regression mapping based on the feature enhancement sequence to obtain a feature vector set that reflects continuous trend characteristics; calculating the drift rate based on the feature vector set; and performing feature regression mapping based on the drift rate. A nonlinear fitting process is performed to update a preset evolution probability distribution, generating a polyvariable index reflecting early signs of cylinder seal failure. An evolution probability threshold range is constructed based on the polyvariable index, and a cumulative deviation is calculated based on this threshold range. When the cumulative deviation exceeds a preset anomaly tolerance limit, a potential anomaly signal is generated, and an anomaly confidence level is calculated to obtain a preliminary judgment result for the anomaly signal. Variational mode decomposition is performed based on the preliminary judgment result and the standardized dataset to obtain a multidimensional feature vector. Based on the multidimensional feature vector, it is determined whether the potential anomaly signal is related to cylinder seal failure, resulting in the final early fault warning information. The trigger timestamp of the anomaly is determined based on the early fault warning information. Based on the trigger timestamp, the environmental feature data is retrieved back and matched with the status description text, aggregating and generating a structured real-time monitoring report of the air pump status.

[0007] Secondly, this invention provides a fault prediction system for an air pump based on multi-source data, comprising: a data acquisition module for acquiring environmental characteristic data and simulated signals including vibration, temperature, pressure, and gas flow rate during air pump operation, and performing time series alignment on the simulated signals to obtain an initial operating state dataset; a standard processing module for performing wavelet decomposition to remove noise from the initial operating state dataset to obtain a reconstructed signal sequence, and performing local weighted regression scatter smoothing processing on the reconstructed signal sequence to obtain a standardized dataset; a feature vector module for performing feature enhancement processing on small amplitude changes of polytropic exponential drift based on the standardized dataset to obtain a feature enhancement sequence, and performing feature regression mapping based on the feature enhancement sequence to obtain a feature vector set that reflects continuous trend characteristics; and a polytropic exponential module for calculating the drift rate based on the feature vector set, and performing feature regression mapping based on the drift rate. The system performs nonlinear fitting and updates a preset evolution probability distribution to generate a polyvariable index reflecting early signs of cylinder seal failure. A preliminary judgment module is used to construct an evolution probability threshold range based on the polyvariable index, calculate a cumulative deviation based on the evolution probability threshold range, and generate a potential abnormal signal and calculate an abnormality confidence level when the cumulative deviation exceeds a preset anomaly tolerance limit, thus obtaining a preliminary judgment result for the abnormal signal. An early fault module is used to perform variational mode decomposition based on the preliminary judgment result and the standardized dataset to obtain a multidimensional feature vector, and determine whether the potential abnormal signal is related to cylinder seal failure based on the multidimensional feature vector, thus obtaining the final early fault warning information. A real-time monitoring module is used to determine the trigger timestamp of the abnormality based on the early fault warning information, backtrack and retrieve the environmental feature data based on the trigger timestamp, match the status description text, and aggregate to generate a structured real-time monitoring report of the air pump status.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention integrates multi-source data such as vibration, temperature, pressure and gas flow rate, and introduces a polyvariate index to comprehensively characterize the actual gas compression process. It transforms the originally dispersed and noise-sensitive physical quantity changes during the operation of the air pump into unified and comparable trend characteristics. By enhancing the characteristics of the small drift of the polyvariate index, the slow changes caused by early seal degradation become observable, thereby enabling early identification of the budding stage of cylinder seal failure and improving the sensitivity of early fault detection. (2) The present invention dynamically models the evolution process of the polyvariable index and combines the drift rate and evolution probability to determine the continuous deviation, avoiding false alarms or missed alarms caused by a single threshold judgment; by calculating the anomaly confidence, the short-term fluctuations are distinguished from the real degradation trend, improving the stability and reliability of the anomaly judgment, and making the fault prediction results more consistent with the actual characteristics of the long-term operation of the air pump. (3) Based on the preliminary anomaly judgment, the present invention introduces a multidimensional physical quantity cross-verification mechanism to confirm the correlation between the abnormality of the variable index and the failure of the cylinder seal, and generates a monitoring report containing the abnormal time point and characteristic changes, providing a clear basis for maintenance decisions; thereby reducing unplanned downtime and blind maintenance, and improving the safety of the air pump operation. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the air pump fault prediction method based on multi-source data provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the air pump fault prediction system based on multi-source data provided in the second embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] Reference Figure 1 The first embodiment of the present invention provides a method for predicting air pump failures based on multi-source data, including the following steps: S11, Collect environmental characteristic data and simulated signals including vibration, temperature, pressure and gas flow during the operation of the air pump, and perform time series alignment on the simulated signals to obtain the initial operating state dataset. S12, perform wavelet decomposition on the initial running state dataset to remove noise, obtain a reconstructed signal sequence, and perform local weighted regression scatter smoothing processing on the reconstructed signal sequence to obtain a standardized dataset; S13, Based on the standardized dataset, feature enhancement processing is performed on the small amplitude changes of the polytropic exponential drift to obtain a feature enhancement sequence, and feature regression mapping is performed based on the feature enhancement sequence to obtain a feature vector set that can reflect the continuous trend characteristics; S14. Based on the feature vector set, the drift rate is calculated, and a nonlinear fitting is performed based on the drift rate to update the preset evolution probability distribution, generating a variable index that reflects early signs of cylinder seal failure. S15, construct an evolution probability threshold range based on the variable index, calculate the cumulative deviation based on the evolution probability threshold range, and when the cumulative deviation exceeds the preset anomaly tolerance limit, generate a potential anomaly signal and calculate the anomaly confidence level to obtain a preliminary judgment result of the anomaly signal. S16. Based on the preliminary judgment result and the standardized dataset, variational mode decomposition is performed to obtain a multidimensional feature vector. Based on the multidimensional feature vector, it is determined whether the potential abnormal signal is related to cylinder seal failure, and the final early fault warning information is obtained. S17. Determine the trigger timestamp of the abnormality based on the early fault warning information, retrieve the environmental feature data back based on the trigger timestamp and match the status description text, and aggregate to generate a structured real-time monitoring report of the air pump status.

[0012] In step S11, it is necessary to collect environmental characteristic data and simulated signals including vibration, temperature, pressure, and gas flow rate during the operation of the air pump, and perform time series alignment on the simulated signals to obtain an initial operating state dataset. This includes: converting the simulated signals into discrete digital sequences; assigning timestamps to the digital sequences and performing time series alignment to generate multidimensional raw data synchronized with the time axis; if the fluctuation amplitude of the multidimensional raw data exceeds the fluctuation amplitude threshold, removing environmental noise interference from the multidimensional raw data to obtain a smoothed multi-source physical quantity data stream; collecting environmental characteristic data and simulated signals including vibration, temperature, pressure, and gas flow rate during the operation of the air pump, and performing time series alignment on the simulated signals to obtain an initial operating state dataset.

[0013] Environmental characteristic data includes at least one of ambient temperature, ambient humidity, and ambient air pressure, which are collected synchronously by the corresponding environmental sensors; vibration, temperature, pressure, and gas flow signals related to the operation status of the air pump are collected by sensors installed on the air pump body or pipeline, and each sensor outputs continuously changing analog voltage or current signals.

[0014] The environmental characteristic data and operational status analog signals are respectively input into a multi-channel analog-to-digital converter (ADC) module for digital processing. The resolution of the ADC module is set to 12 bits or 16 bits. Taking an ambient temperature sensor as an example, its range is −20℃ to 80℃, corresponding to a 0–5V output. When using a 12-bit ADC, the temperature resolution is approximately 0.025℃. The pressure sensor has a range of 0–10 bar, and when using a 16-bit ADC, the pressure resolution is approximately 0.00015 bar. The vibration signal sampling frequency is set according to the frequency band of interest, usually not lower than 2kHz, preferably 2kHz to 5kHz, to capture the mid-to-high frequency characteristics of components such as bearings and gears. The sampling frequency for temperature, pressure, and environmental characteristic data is set to 10–100Hz, and the sampling frequency for gas flow signals is set to 10–50Hz, forming corresponding digital quantity sequences.

[0015] Each digital sequence is time-stamped by a unified system clock during sampling, with timestamp accuracy set to milliseconds. For digital sequences with different sampling frequencies, a common time reference sampling frequency of 100Hz is set, and each sequence undergoes resampling. For sequences with sampling frequencies lower than the common sampling frequency, linear interpolation is used to fill in the time points; for sequences with sampling frequencies higher than the common sampling frequency, a sliding window averaging method is used for downsampling. This ensures that environmental characteristic data, vibration, temperature, pressure, and gas flow data form a multi-dimensional original data matrix on the same time axis.

[0016] For the multidimensional raw data, a sliding time window is used to calculate the local fluctuation amplitude of each physical quantity, with the window length set to 0.5–2 seconds. The fluctuation amplitude threshold is statistically determined based on historical data collected by the air pump during a preset stable operation phase, where the stable operation phase refers to the time period during which the air pump runs continuously for no less than 10 minutes under rated speed and rated load conditions without any fault alarm records. For each physical quantity data within this time period, its mean and standard deviation within the sliding time window are calculated, with the time window length set to 0.5–2 seconds. The fluctuation amplitude threshold for the corresponding physical quantity is set to mean ± k times the standard deviation, where k takes a preset range of 2.5–3.5. When the change amplitude of a certain physical quantity within the corresponding time window exceeds the fluctuation amplitude threshold, it is determined that there is an abnormal disturbance in the data segment, and the data segment is smoothed using median filtering or moving average filtering.

[0017] The smoothed environmental feature data and air pump operating status data are stored in timestamp order. Each data record includes a timestamp field and corresponding values ​​for ambient temperature, humidity, vibration, pressure, and gas flow rate. An initial operating status dataset is constructed based on these data records for subsequent data preprocessing, feature enhancement, polyvariate index calculation, and fault prediction analysis.

[0018] In step S12, the initial running state dataset needs to be decomposed by wavelet to remove noise, resulting in a reconstructed signal sequence. Then, based on the reconstructed signal sequence, local weighted regression scatter smoothing is performed to obtain a standardized dataset. This includes: performing wavelet decomposition on the initial running state dataset and eliminating noise interference based on a preset noise threshold processing coefficient to obtain a reconstructed signal sequence; removing outliers from the reconstructed signal sequence to obtain a corrected signal sequence, and performing Z-score standardization on the corrected signal sequence to obtain a dimensionless numerical sequence; and performing local weighted regression scatter smoothing on the dimensionless numerical sequence to fit the trend component of the dimensionless numerical sequence, generating a smoothed standardized dataset.

[0019] For the time series of each physical quantity in the initial running state dataset, discrete wavelet decomposition is performed. The Daubechies wavelet family is selected as the wavelet basis function. The number of decomposition levels is determined based on the signal length and the lowest frequency component to be analyzed, typically set to 3 to 6 levels to ensure that the lowest frequency approximation coefficients effectively characterize the signal's trend. The specific number of levels is determined based on the sampling frequency. Wavelet decomposition decomposes the original signal into approximation coefficients and several levels of detail coefficients, where the detail coefficients correspond to high-frequency noise components. The noise threshold is set based on the statistical characteristics of the detail coefficients, using a soft thresholding method to suppress detail coefficients with absolute values ​​less than the noise threshold. The noise threshold is determined by estimating the noise standard deviation based on the median absolute deviation, and the threshold coefficient is set in the range of 2.5–3.5. After thresholding, wavelet reconstruction is performed on the processed coefficients to obtain a reconstructed signal sequence free of high-frequency noise.

[0020] For the reconstructed signal sequence, a sliding time window is used to detect outliers in the data points, with the window length set to 1–3 seconds. Within the window, the median and absolute deviation of the median are calculated. When a data point deviates from the median by more than a preset multiple, it is identified as an outlier, with the deviation multiple set to 3–4. The identified outliers are replaced using linear interpolation of adjacent normal data to obtain a corrected signal sequence. Subsequently, Z-score standardization is performed on the corrected signal sequence, where the standardization parameters are taken from historical data collected by the air pump during a preset stable operating phase. The stable operating phase refers to the operating interval during which the air pump runs continuously at rated speed and rated load for no less than 10 minutes, and no fault alarm is triggered within the corresponding time period. For the time series of each physical quantity within this interval, its arithmetic mean and standard deviation are calculated. These mean and standard deviation are then used as fixed standardization parameters to perform dimensionless processing on the corrected signal sequence, making the mean of each physical quantity sequence zero and the standard deviation one.

[0021] For dimensionless numerical sequences, a locally weighted regression scatter smoothing method is used to fit the trend of the time series. Specifically, using timestamps as independent variables and dimensionless numerical values ​​as dependent variables, a local regression window is constructed at each time point. This window contains data points centered at that time point, representing 5%–15% of the total sample length before and after it, with the number of data points within the window not less than a preset minimum sample size. During local regression, a weighting function based on time distance is used to assign weights to the samples within the window. Samples closer in time have greater weights, while the weights of samples farther in time decrease monotonically according to a preset kernel function. The kernel function is either a Gaussian kernel or a cubic kernel, and the regression model uses a first-order polynomial form. By performing least-squares fitting on the weighted samples within the window, the local regression estimate corresponding to the current time point is obtained. This local regression calculation is performed sequentially on each time point in the dimensionless numerical sequence to form a corresponding trend term sequence. This trend term sequence is used as a smoothed, standardized dataset for subsequent feature enhancement and polyvariate modeling analysis.

[0022] In step S13, feature enhancement processing is performed on the small amplitude changes of the polytropic index drift based on the standardized dataset to obtain a feature enhancement sequence. Then, feature regression mapping is performed on the feature enhancement sequence to obtain a feature vector set that reflects the continuous trend characteristics. This includes: performing gas thermodynamic processing on the standardized dataset to obtain an instantaneous polytropic index sequence; performing differential and nonlinear gain transformation on the instantaneous polytropic index sequence to generate a feature enhancement sequence; calculating the local regression slope of the feature enhancement sequence; and mapping the local regression slope to the instantaneous polytropic index sequence to a feature vector set that reflects the continuous trend characteristics.

[0023] Pressure and temperature data at the same timestamp are selected from a standardized dataset and arranged chronologically to form a continuous data sequence. The pressure and temperature data are segmented using a fixed-length sliding time window, containing 20–50 consecutive sampling points with a corresponding time span of 0.5–2 seconds. Within each time window, the trends of pressure and temperature data are fitted based on their correspondence on the time axis. The polytropic index of the gas compression or expansion process within the corresponding time window is determined based on the degree of pressure response to temperature changes in the fitting results, reflecting the actual state of the gas deviating from an ideal adiabatic or isothermal process during that time period.

[0024] By moving the sliding time window point by point along the time axis, the entire operation process is repeatedly calculated to form a transient polytropic index sequence that corresponds one-to-one with the time axis. The transient polytropic index sequence is used to describe the local change characteristics of the gas state during the operation of the air pump and serves as input data for subsequent feature enhancement processing.

[0025] It should be noted that the purpose of the aforementioned gas thermodynamic treatment is to calculate a physical quantity that comprehensively reflects the relationship between pressure and temperature changes, namely the instantaneous polyvariance index, based on the thermodynamic characteristics of the actual compression or expansion process of the gas in the cylinder of the air pump. This index is used in thermodynamics to describe the quantitative relationship between pressure and volume (or temperature) during the process of gas state change. When the cylinder is well sealed and the process is ideal, the index tends to a stable theoretical value. However, when the seal deteriorates slightly, gas leakage or frictional heat generation will cause a slight but continuous shift in the relationship between the actual pressure change and temperature change, and the index value will drift slowly accordingly.

[0026] Specifically, based on the pressure and temperature signals aligned at the same time in the standardized dataset, their continuous changes on the time axis are first paired. For a selected short-term analysis window, such as 0.5 to 2 seconds, the window contains a series of pressure and temperature data pairs arranged in chronological order. A linear trend analysis method, such as the least squares method, is applied to quantify the overall proportional relationship between the magnitude of pressure change and the magnitude of temperature change within this window, i.e., to determine the average response slope of pressure with temperature change. This response slope is directly related to the polytropic index in thermodynamics; in polytropic gas processes, this response slope and the polytropic index are correlated. The one-to-one monotonic function relationship, whose specific transformation relationship is determined by the fundamental laws of thermodynamics, is common knowledge in this field. In the calculation, the obtained response slope value is converted into the corresponding instantaneous polytropic index value through a certain transformation relationship based on thermodynamic principles. This transformation relationship is a well-known relationship in the field of gas thermodynamics, indicating that the polytropic index is a single-valued function of the response slope. By sliding the analysis window along the time axis and repeating the above steps of pairing, trend analysis and transformation for each window, an instantaneous polytropic index sequence corresponding to the time axis can be generated to continuously characterize the subtle evolution of the gas state change characteristics during the operation of the gas pump.

[0027] For a transient polytropic index sequence, the amplitude of change between adjacent time points is calculated sequentially to characterize the trend of the polytropic index over a short period. This amplitude reflects the rate of increase or decrease of the polytropic index, but in the early stages of cylinder seal degradation, the amplitude is small and easily masked by noise. To enhance sensitivity to such small but continuous changes, nonlinear gain processing is applied to the amplitude sequence. This nonlinear gain processing, according to preset rules, amplifies smaller amplitude changes by a relatively higher proportion while suppressing larger abrupt changes, ensuring that short-term drastic fluctuations do not dominate the analysis results. The nonlinear gain parameters are set based on the amplitude distribution during normal operation of the equipment, resulting in a feature enhancement sequence that highlights early degradation characteristics.

[0028] For feature-enhanced sequences, timestamps are used as the sorting criterion, and the sequences are processed point by point in chronological order. At each time point, a fixed number of adjacent data points are selected forward and backward from that time point to form a local analysis window. The number of data points contained in the window is preset to 5%–15% of the total sample size based on the total sample length, and the number of data points in the window is not less than the preset minimum sample size to ensure the stability of trend analysis.

[0029] Within each local analysis window, the numerical changes of the feature enhancement sequence within the window are trend-fitted according to the chronological order of the data points. By comparing the direction and magnitude of the changes in the values ​​at the beginning and end of the window, the trend of the feature enhancement sequence within that time window is determined, and the trend slope value at the corresponding time point is obtained accordingly. This slope is used to characterize whether the polymorphism index shows a continuous upward or downward trend during that time period.

[0030] Subsequently, the instantaneous variability index value corresponding to the time point is combined with the trend slope value in a preset order to form a feature vector describing the operating state at that time point, wherein the instantaneous variability index is used to reflect the current gas state level, and the trend slope is used to reflect the intensity of its changing trend.

[0031] Repeat the above local window analysis and feature combination process for each time point in the feature enhancement sequence to construct a data set containing multiple sets of feature vectors. The feature vector set is used for subsequent multivariable index trend modeling and anomaly detection analysis.

[0032] In step S14, the drift rate needs to be calculated based on the feature vector set, and a nonlinear fitting is performed based on the drift rate to update the preset evolution probability distribution, generating a polytropic index reflecting early signs of cylinder seal failure. This includes: performing state filtering and smoothing on the feature vector set to obtain a state estimation sequence; calculating the drift rate of the state estimation sequence and determining the portion of the drift rate greater than a preset rate threshold as a drift feature interval set; performing nonlinear fitting on the drift feature interval set to obtain a dynamic feature change descriptor; updating the preset evolution probability distribution based on the dynamic feature change descriptor to generate a polytropic index reflecting early signs of cylinder seal failure.

[0033] The feature vector set is sorted according to timestamp order and used sequentially as data input for discrete time steps. A single state variable represents the evolution of the polyvariable index during operation, and the state variable has a unique value at each time step. The instantaneous polyvariable index value and trend strength value contained in the feature vector are used as reference data for correcting the state variable at the current time step.

[0034] During the initialization phase, the instantaneous polymorphic index value corresponding to the first time step is directly set as the initial state value. At the same time, based on the historical data of the feature vector during the stable operation phase, the normal range of change of the polymorphic index value between adjacent time steps is statistically analyzed, and this range of change is used as the allowable adjustment boundary in the subsequent state update process.

[0035] In each subsequent time step, state update processing is performed in a fixed order. First, the state value of the previous time step is used as the prediction baseline value for the current time step. Then, the feature vector corresponding to the current time step is read, and the deviation between the prediction baseline value and the current instantaneous variability index value is compared. When the deviation is within the normal range of change, the prediction baseline value is kept unchanged and used as the state value for the current time step. When the deviation exceeds the normal range of change, the prediction baseline value is corrected according to the direction of change indicated by the trend intensity value in the feature vector. The correction magnitude is limited to within a preset maximum adjustment ratio, thereby obtaining the state value corresponding to the current time step.

[0036] By repeatedly performing the above prediction, comparison and correction processes on each time step of the feature vector set, a continuously updated state estimation sequence is formed on the time axis. The state estimation sequence is used to describe the actual evolution of the polytropic index over time.

[0037] The state estimation sequence is processed chronologically, using a fixed-length time window as the analysis unit, with the time window length set to 30–120 seconds. Within each time window, the change in the state estimation value corresponding to the start and end times of the window is calculated, and combined with the window duration, the rate of change corresponding to that window is obtained, which is used to characterize the degree of drift of the polymorphism index within that time period.

[0038] The drift rate threshold is determined statistically based on the state estimation sequence during the stable operation phase. By statistically analyzing the rate of change distribution corresponding to each time window within this phase, the rate of change not exceeding 95% is selected as the normal upper limit. When the rate of change within a certain consecutive time window exceeds the normal upper limit, and this over-limit state continues to occur within at least three adjacent time windows, the corresponding time period is determined as the drift characteristic interval.

[0039] By scanning the entire state estimation sequence window by window, all time periods that meet the above conditions are extracted, forming a set of drift feature intervals.

[0040] For each drift characteristic interval, a sequence of state estimates within that interval is extracted, and the change pattern of this sequence is fitted and analyzed according to time sequence. The fitting process is constrained by the overall trend of the state estimates within the interval, allowing only monotonic changes to avoid interference from short-term reverse fluctuations on the fitting results.

[0041] After fitting, a set of parameters describing the drift behavior is extracted from the fitting results. This set of parameters includes at least: the average rate of change within the interval, the maximum magnitude of change, and the increasing or decreasing trend of the rate of change over time. These parameters together constitute a change descriptor characterizing the dynamic properties of the drift interval.

[0042] In the initial stage of the system, based on the historical data of the air pump's normal operation, an evolution probability distribution of the polytropic index under fault-free conditions is established to characterize the statistical characteristics of the polytropic index's changes under normal circumstances. This evolution probability distribution is stored as a baseline distribution.

[0043] After acquiring new dynamic characteristic change descriptors, the descriptor parameters are compared with the corresponding parameters of the baseline distribution. Based on the magnitude, rate, and duration of change reflected by the descriptors, the evolution probability distribution is updated step by step. When the probability of the polymorphism index entering an abnormal evolution state in the updated evolution probability distribution exceeds a preset risk threshold, and this probability maintains an upward trend over multiple consecutive time windows, the polymorphism index for the corresponding time period is determined as the polymorphism index reflecting early signs of cylinder seal failure.

[0044] It should be noted that the preset evolution probability distribution is established based on a large amount of historical data of normal operation of the air pump during system initialization. Specifically, the parameters contained in the dynamic feature change descriptors in each time period of the historical data are calculated, such as the statistical distribution of the average rate of change. For example, it is assumed that it follows a Gaussian distribution, and the initial mean and variance of the distribution are determined by maximum likelihood estimation, which is the initial evolution probability distribution. During online monitoring, the distribution parameters are updated using Bayesian update or sliding window statistics based on the newly calculated dynamic feature change descriptors. For example, recursive Bayesian estimation is used to treat the new descriptor parameters as observations and to update the mean and variance of the distribution posteriorly, thereby obtaining the latest evolution probability distribution reflecting the current equipment status.

[0045] In step S15, an evolution probability threshold range needs to be constructed based on the polyvariable index, and a cumulative deviation is calculated based on the evolution probability threshold range. When the cumulative deviation exceeds a preset anomaly tolerance limit, a potential anomaly signal is generated and an anomaly confidence level is calculated to obtain a preliminary judgment result of the anomaly signal. This includes: constructing an evolution probability threshold range containing dynamic upper and lower limits based on the polyvariable index and the associated evolution probability distribution; calculating the deviation magnitude between the polyvariable index and the evolution probability threshold range to obtain a trend feature deviation sequence; identifying data segments in the trend feature deviation sequence that satisfy a preset minimum persistence constraint, and calculating the cumulative deviation of the data segments; when the cumulative deviation exceeds a preset anomaly tolerance limit, a potential anomaly signal is generated, and an anomaly confidence level is calculated in conjunction with the evolution probability distribution to obtain a preliminary judgment result of the anomaly signal.

[0046] Based on historical data collected during the stable operation phase of the polyvariable index, its value distribution characteristics under normal operating conditions are statistically analyzed. Specifically, during a stable operation phase in which the air pump operates continuously at rated speed and rated load for no less than 10 minutes, the average level of the polyvariable index and its natural fluctuation range are calculated, and this statistical result is used as the benchmark range for normal operation.

[0047] In a preferred embodiment, the baseline interval is extended upwards and downwards by 2 to 3 times the typical fluctuation range, respectively, to construct the dynamic upper limit and dynamic lower limit of the polymorphic index. At the same time, the upper and lower limits are slightly adaptively adjusted in combination with the evolution probability change trend of the polymorphic index in the current time period, so that the threshold range can be dynamically updated with the change of the operating state, thereby forming an evolution probability threshold range for subsequent judgment.

[0048] After obtaining the evolution probability threshold range, the polyvariance index corresponding to each time point is compared and processed. When the polyvariance index is within the threshold range, the deviation value of that time point is recorded as zero; when the polyvariance index exceeds the threshold range, the deviation magnitude of its exceeding the upper or lower limit is recorded, and a continuous trend feature deviation sequence is formed in chronological order.

[0049] In practical engineering applications, the sampling period of the polyvariable index can be set to 1 to 5 seconds. Therefore, the trend feature deviation sequence can reflect the continuous deviation of the polyvariable index from the normal evolution range in a relatively precise manner.

[0050] To address the deviation sequence of the trend characteristics, a minimum duration constraint is introduced to avoid misjudgments caused by instantaneous fluctuations. In a preferred embodiment, time periods in which the deviation values ​​of at least 5 to 10 consecutive sampling points are all greater than zero are identified as candidate abnormal segments, with a corresponding duration of approximately 5 to 50 seconds.

[0051] For each candidate anomaly segment, the deviation magnitude at each time point within that time period is accumulated to obtain the cumulative deviation, which is used to comprehensively characterize the duration and intensity of the anomaly. When the cumulative deviation exceeds a preset anomaly tolerance limit, the anomaly is considered no longer a random disturbance but has practical engineering significance.

[0052] When the cumulative deviation exceeds the anomaly tolerance limit, a potential anomaly signal for the corresponding time period is generated. The anomaly tolerance limit is preferably determined based on the deviation statistics during the stable operation phase, for example, set to 2 to 3 times the maximum cumulative deviation during the stable operation phase, to ensure that anomaly detection is triggered only when there is a significant deviation from normal evolutionary behavior.

[0053] After generating potential anomalous signals, the anomaly confidence level is further calculated by combining the changing trend of the polytropic index evolution probability within that time period. When the probability of the polytropic index entering an anomalous evolution state increases for multiple consecutive sampling periods during the anomaly occurrence, and the duration exceeds 10 seconds, the anomaly confidence level is determined to be high; when the probability increase is discontinuous or the duration is short, the anomaly confidence level is determined to be medium or low. The final output includes the anomaly occurrence time period, anomaly level, and confidence level as a preliminary judgment result.

[0054] In step S16, variational mode decomposition is performed based on the preliminary judgment result and the standardized dataset to obtain a multidimensional feature vector. Based on the multidimensional feature vector, it is determined whether the potential abnormal signal is related to cylinder seal failure, thus obtaining the final early fault warning information. This includes: performing variational mode decomposition based on the preliminary judgment result and the standardized dataset to obtain a multidimensional feature vector; calculating the dynamic time warping distance between the multidimensional feature vector and the polytropic feature deviation sequence associated with the potential abnormal signal to obtain a synchronization feature value; determining the degree of correlation between the potential abnormal signal and cylinder seal failure based on the synchronization feature value to obtain a failure correlation probability value; and generating early fault warning information containing failure type confirmation if the failure correlation probability value is within a preset high-risk range.

[0055] Based on the preliminary judgment results, standardized data within a certain time range before and after the anomaly occurrence time period are selected as the analysis object. The preferred time range is 30 to 120 seconds before and after the anomaly trigger time point. For the multi-source physical quantity time series data within this time range, signal decomposition processing is performed to split the original time series into several components with different time scale characteristics.

[0056] In a preferred embodiment, each physical quantity signal is decomposed into 3 to 6 feature components, where the low-time-scale component is used to characterize long-term slow changes, and the high-time-scale component is used to characterize transient fluctuations. For each component, its average amplitude change, energy proportion change, and duration of change within the stated time range are extracted, and these indicators are combined in chronological order to construct a multidimensional feature vector set for subsequent analysis.

[0057] After obtaining the set of multidimensional feature vectors, a time alignment analysis is performed between them and the polytropic exponential feature deviation sequence corresponding to the potential anomalous signal. Using the time point of the first trigger of the polytropic exponential anomaly as the alignment benchmark, local time alignment processing is performed on the multidimensional feature vectors along the time axis to compensate for the lag in the response of different physical quantities to the anomaly.

[0058] After time alignment, the similarity between the changing trend of the multidimensional feature vector and the changing trend of the polyvariable exponential feature deviation sequence is evaluated, and a synchronicity feature value is generated. The synchronicity feature value is used to characterize the degree of synchronization between the two within the abnormal time period, and its value range is preferably normalized to the interval of 0–1.

[0059] It should be noted that the synchronization characteristic value is used to characterize the degree of synchronization between the two during the abnormal time period. The smaller the value, the higher the synchronization. The higher the failure association probability value obtained based on the synchronization characteristic value, the greater the possibility that the current abnormality is caused by cylinder seal failure. If the failure association probability value exceeds the preset high-risk threshold, an early fault warning information containing failure type confirmation is generated.

[0060] For example, when the synchronicity characteristic value is less than 0.3, it indicates that the changes in multidimensional physical quantity characteristics and the abnormal changes in polyvariance index are highly synchronized; when the synchronicity characteristic value is between 0.3 and 0.6, it indicates that the two have a certain correlation but the degree of synchronization is generally low; when the synchronicity characteristic value is greater than 0.6, it indicates that the synchronicity between the changes in multidimensional physical quantity and the abnormal polyvariance index is weak and the correlation is low.

[0061] For example, in the case of slight leakage in the cylinder seal, the increased fluctuation of exhaust temperature and the continuous deviation of the polyvariance index occur almost simultaneously, and the corresponding synchronicity characteristic value can fall in the range of 0.15–0.25; while in the case of abnormal conditions caused by short-term fluctuations in external load, the changes in multidimensional physical quantities and the deviation of the polyvariance index are not synchronized, and the synchronicity characteristic value is greater than 0.7.

[0062] After obtaining the synchronicity feature value, the correlation between the anomaly and cylinder seal failure is determined by combining the multidimensional feature vector within the abnormal time period. The failure correlation probability is used to characterize the possibility that the current anomaly is caused by cylinder seal failure, and its value range is preferably normalized to the interval 0–1.

[0063] In a preferred embodiment, when the synchronicity feature value is less than 0.3 and at least two types of features related to sealing failure appear simultaneously in the multidimensional feature vector, the failure association probability is determined to be in the high-risk range, and the corresponding value is preferably greater than 0.7; when the synchronicity feature value is between 0.3 and 0.6 and only some related features appear, the failure association probability is determined to be in the medium-risk range, and the corresponding value is preferably between 0.4 and 0.7; when the synchronicity feature value is greater than 0.6, or the multidimensional features do not show consistent abnormal changes, the failure association probability is determined to be in the low-risk range, and the corresponding value is preferably less than 0.4.

[0064] For example, in a certain monitoring scenario, the synchronicity characteristic value is 0.22, and the amplitude of exhaust temperature fluctuation and crankcase pressure pulsation both increase significantly, so the failure correlation probability can be determined to be above 0.8; while in another scenario, the synchronicity characteristic value is 0.68, and only the vibration characteristics are abnormal for a short time, so the failure correlation probability can be determined to be around 0.3.

[0065] The failure correlation probability value is compared with a preset risk range. When the failure correlation probability value is in the high-risk range, an early fault warning is generated. To avoid false alarms caused by occasional fluctuations, in a preferred embodiment, the final warning decision is triggered only when the failure correlation probability remains in the high-risk range for at least three consecutive time windows.

[0066] In engineering implementation, the time window length is preferably set to 3–5 seconds, so the minimum duration required to trigger the warning is approximately 10–15 seconds. For example, when the failure correlation probability remains above 0.75 for 12 consecutive seconds, an early fault warning is generated; if the probability only briefly exceeds 0.7 within a single time window and then quickly falls back, no warning is triggered.

[0067] When generating early fault warning information, the failure type is confirmed by combining the dominant abnormal feature types in the multi-dimensional feature vector. For example, when increased pressure pulsation and abnormal exhaust temperature occur simultaneously, the failure type is preferably confirmed as minor leakage of cylinder seal; when the abnormality is mainly manifested as sudden vibration and weak synchronicity, it is not determined as seal failure.

[0068] In step S17, it is necessary to determine the trigger timestamp of the abnormality based on the early fault warning information, backtrack and retrieve the environmental feature data based on the trigger timestamp and match the status description text, and aggregate to generate a structured real-time monitoring report of the air pump status. This includes: parsing the failure types contained in the early fault warning information and locking the trigger timestamp of the failure type; backtracking and retrieving the environmental feature data based on the trigger timestamp to construct a feature change set containing feature amplitude fluctuations and trend directions; mapping the feature change set to a preset fault semantic description library, matching the status description text, and aggregating to generate a structured real-time monitoring report of the air pump status.

[0069] The early fault warning information is parsed to read the failure type identifier and its corresponding trigger timestamp. The trigger timestamp is the time when the system first confirms that the failure association probability has entered the high-risk range and continues to meet preset conditions, and is used as a reference time for the occurrence of the anomaly.

[0070] In a preferred embodiment, if the failure correlation probability remains above 0.7 for at least 10–15 seconds, the start time of this sustained interval is determined as the abnormal trigger time point. For example, if the failure correlation probability is above 0.75 for 12 consecutive seconds starting from 14:32:10, then 14:32:10 is recorded as the abnormal trigger timestamp, and the failure type is analyzed as a minor cylinder seal leak or a seal wear failure.

[0071] Centered on the anomaly trigger timestamp, data segments within a certain time range before and after the anomaly are retrieved from the environmental feature data. The time range is preferably set to 5 minutes before and 2 minutes after the trigger time point to comprehensively reflect the feature evolution process before and after the anomaly occurs.

[0072] For the environmental characteristic data obtained through retrospection, the changes in physical quantities related to the anomalies are extracted, including but not limited to vibration amplitude, exhaust temperature, pressure fluctuations, and gas flow rate trends. In a preferred embodiment, the maximum amplitude, average rate of change, and direction of change of each physical quantity within the retrospective time range are calculated to describe the characteristic amplitude fluctuations and trends.

[0073] For example, in a scenario involving minor cylinder seal leakage, the exhaust temperature gradually increases by 4–6 degrees Celsius from a stable value within 3 minutes before the abnormality is triggered, and the crankcase pressure pulsation amplitude increases by 20%–35%. These changes are recorded as characteristic change details of the corresponding physical quantities. A feature change set containing information on the amplitude and trend direction of each physical quantity is then constructed for subsequent semantic description.

[0074] The set of feature changes is matched against a pre-defined fault semantic description library, which stores typical feature combinations and their textual description rules corresponding to various fault types. The state description text with the highest matching degree is selected by comparing the current set of feature changes with the matching conditions in the semantic description library.

[0075] It is worth noting that the preset fault semantic description library is built based on domain knowledge and a historical fault case library. First, it summarizes the typical change patterns of various physical quantity characteristics under different fault modes, such as cylinder seal leakage and bearing wear, including vibration spectrum, temperature trend, and pressure fluctuation, forming a feature-fault mapping rule library. For example, the thresholds (X, Y) in the rules such as "IF (exhaust temperature continues to rise > X℃ / min) AND (pressure pulsation frequency amplitude increases > Y%) AND (polyvariance index drifts negatively) THEN Fault type: Cylinder micro-leakage; Description text: Cylinder sealing performance shows signs of degradation, and there may be a slight gas leak." are determined by statistical analysis of historical fault case data. The library is essentially a structured set of rules used to match the numerical feature change set to the most suitable, predefined fault semantic description.

[0076] In a preferred embodiment, when the set of characteristic changes simultaneously satisfies the three conditions of "continuous increase in exhaust temperature", "increased pressure pulsation", and "continuous deviation of polymorphism index", the matched state description text can be "decreased cylinder sealing performance, with a risk of minor leakage"; when only abnormal vibration occurs and other physical quantities do not change synchronously, the matching is "short-term abnormality caused by external operating condition disturbance".

[0077] Finally, the anomaly trigger timestamp, failure type confirmation result, details of characteristic changes in each physical quantity, and corresponding status description text are aggregated to generate a structured real-time monitoring report of the air pump status. The monitoring report preferably displays the anomaly evolution process in chronological order and is used for subsequent operation and maintenance decisions and maintenance plan formulation.

[0078] It should be noted that the specific values ​​of the preset thresholds involved in this invention, such as noise threshold coefficient, rate threshold, anomaly tolerance limit, and risk interval boundary, can be calculated based on sufficient historical data collected during the stable operation phase of the air pump. This involves calculating the distribution of relevant characteristic quantities, such as mean, standard deviation, and percentiles, and setting the threshold as the mean plus or minus a certain multiple of the standard deviation, or taking a specific percentile, such as the 95th percentile. Alternatively, initial values ​​can be given based on industry standards, equipment manuals, or expert experience. Alternatively, labeled historical datasets can be used to optimize the threshold parameters with the goal of maximizing the fault detection rate and minimizing the false alarm rate. Those skilled in the art can select appropriate methods to determine these thresholds based on specific application scenarios and equipment characteristics.

[0079] In summary, this invention relates to the field of air pump failure prediction technology, and discloses an air pump failure prediction method based on multi-source data. The method can identify the continuous drift characteristics of polytropic index under noise and operating condition fluctuations, and realize reliable prediction of early cylinder seal failure.

[0080] Reference Figure 2 The second embodiment of the present invention provides a fault prediction system for an air pump based on multi-source data, comprising: a data acquisition module, used to acquire environmental characteristic data and simulated signals including vibration, temperature, pressure, and gas flow rate during the operation of the air pump, and to perform time series alignment on the simulated signals to obtain an initial operating state dataset; a standard processing module, used to perform wavelet decomposition to remove noise from the initial operating state dataset to obtain a reconstructed signal sequence, and to perform local weighted regression scatter smoothing processing on the reconstructed signal sequence to obtain a standardized dataset; a feature vector module, used to perform feature enhancement processing on the small amplitude changes of polytropic exponential drift based on the standardized dataset to obtain a feature enhancement sequence, and to perform feature regression mapping based on the feature enhancement sequence to obtain a feature vector set that reflects continuous trend characteristics; and a polytropic exponential module, used to calculate the drift rate based on the feature vector set, and to perform feature regression mapping based on the drift rate. The system performs nonlinear fitting and updates a preset evolution probability distribution to generate a polyvariable index reflecting early signs of cylinder seal failure. A preliminary judgment module is used to construct an evolution probability threshold range based on the polyvariable index, calculate a cumulative deviation based on the evolution probability threshold range, and generate a potential abnormal signal and calculate an abnormality confidence level when the cumulative deviation exceeds a preset anomaly tolerance limit, thus obtaining a preliminary judgment result for the abnormal signal. An early fault module is used to perform variational mode decomposition based on the preliminary judgment result and the standardized dataset to obtain a multidimensional feature vector, and determine whether the potential abnormal signal is related to cylinder seal failure based on the multidimensional feature vector, thus obtaining the final early fault warning information. A real-time monitoring module is used to determine the trigger timestamp of the abnormality based on the early fault warning information, backtrack and retrieve the environmental feature data based on the trigger timestamp, match the status description text, and aggregate to generate a structured real-time monitoring report of the air pump status.

[0081] It should be noted that the air pump fault prediction system based on multi-source data provided in this embodiment of the invention is used to execute all the process steps of the air pump fault prediction method based on multi-source data in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0082] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0083] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for predicting air pump failures based on multi-source data, characterized in that, include: Collect environmental characteristic data and simulated signals including vibration, temperature, pressure and gas flow during the operation of the air pump, and perform time series alignment on the simulated signals to obtain the initial operating state dataset. The initial running state dataset is subjected to wavelet decomposition to remove noise, resulting in a reconstructed signal sequence. Based on this reconstructed signal sequence, local weighted regression scatter smoothing is performed to obtain a standardized dataset. According to the standardized dataset, feature enhancement processing is applied to small amplitude changes in the polytropic index drift, resulting in a feature enhancement sequence. Feature regression mapping is then performed on this feature enhancement sequence to obtain a feature vector set reflecting continuous trend characteristics. Based on this feature vector set, the drift rate is calculated, and nonlinear fitting is performed based on the drift rate to update a preset evolution probability distribution, generating a polytropic index reflecting early signs of cylinder seal failure. An evolution probability threshold range is constructed based on the polytropic index, and the cumulative deviation is calculated based on this threshold range. When the cumulative deviation exceeds a preset anomaly tolerance limit, a potential anomaly signal is generated, and the anomaly confidence level is calculated to obtain a preliminary judgment result for the anomaly signal. Variational mode decomposition is performed based on the preliminary judgment results and the standardized dataset to obtain multidimensional feature vectors. Based on the multidimensional feature vectors, it is determined whether the potential abnormal signal is related to cylinder seal failure, thus obtaining the final early fault warning information. The trigger timestamp of the abnormality is determined based on the early fault warning information. Based on the trigger timestamp, the environmental feature data is retrieved back and matched with the status description text to generate a structured real-time monitoring report of the air pump status.

2. The air pump fault prediction method based on multi-source data according to claim 1, characterized in that, The step of aligning the analog signal to a time series to obtain an initial operating state dataset includes: converting the analog signal into a discrete digital sequence; assigning a timestamp to the digital sequence and aligning it to time series to generate multidimensional raw data synchronized with the time axis; storing the multidimensional raw data to construct an initial operating state dataset containing information on various physical quantities.

3. The air pump fault prediction method based on multi-source data according to claim 1, characterized in that, The step of performing wavelet decomposition on the initial running state dataset to remove noise, obtaining a reconstructed signal sequence, and then performing local weighted regression scatter smoothing on the reconstructed signal sequence to obtain a standardized dataset includes: performing wavelet decomposition on the initial running state dataset, eliminating noise interference based on a preset noise threshold processing coefficient to obtain a reconstructed signal sequence; removing outliers from the reconstructed signal sequence to obtain a corrected signal sequence, and performing Z-score standardization on the corrected signal sequence to obtain a dimensionless numerical sequence; and performing local weighted regression scatter smoothing on the dimensionless numerical sequence to fit the trend component of the dimensionless numerical sequence and generate a smoothed standardized dataset.

4. The air pump fault prediction method based on multi-source data according to claim 1, characterized in that, The step involves performing feature enhancement processing on the small amplitude changes of the polytropic index drift based on the standardized dataset to obtain a feature enhancement sequence, and performing feature regression mapping based on the feature enhancement sequence to obtain a feature vector set that reflects the continuous trend characteristics. This includes: performing gas thermodynamic processing on the standardized dataset to obtain an instantaneous polytropic index sequence; performing differential and nonlinear gain transformation on the instantaneous polytropic index sequence to generate a feature enhancement sequence; calculating the local regression slope of the feature enhancement sequence, and mapping the local regression slope to the instantaneous polytropic index sequence to obtain a feature vector set that reflects the continuous trend characteristics.

5. The air pump fault prediction method based on multi-source data according to claim 1, characterized in that, The step of calculating the drift rate based on the feature vector set, performing nonlinear fitting based on the drift rate, updating the preset evolution probability distribution, and generating a polyvariant index reflecting early signs of cylinder seal failure includes: performing state filtering and smoothing on the feature vector set to obtain a state estimation sequence; calculating the drift rate of the state estimation sequence and determining the portion of the drift rate greater than a preset rate threshold as a drift feature interval set; performing nonlinear fitting on the drift feature interval set to obtain a dynamic feature change descriptor; and updating the preset evolution probability distribution based on the dynamic feature change descriptor to generate a polyvariant index reflecting early signs of cylinder seal failure.

6. The air pump fault prediction method based on multi-source data according to claim 1, characterized in that, The process of constructing an evolution probability threshold range based on the polyvariable index, calculating a cumulative deviation based on the evolution probability threshold range, and generating a potential abnormal signal and calculating an abnormal confidence level when the cumulative deviation exceeds a preset anomaly tolerance limit, to obtain a preliminary judgment result of the abnormal signal, includes: constructing an evolution probability threshold range with dynamic upper and lower limits based on the polyvariable index and the associated evolution probability distribution; calculating the deviation magnitude between the polyvariable index and the evolution probability threshold range to obtain a trend feature deviation sequence; identifying data segments in the trend feature deviation sequence that satisfy a preset minimum persistence constraint, and calculating the cumulative deviation of the data segments; generating a potential abnormal signal when the cumulative deviation exceeds a preset anomaly tolerance limit, and calculating an abnormal confidence level in conjunction with the evolution probability distribution to obtain a preliminary judgment result of the abnormal signal.

7. The air pump fault prediction method based on multi-source data according to claim 1, characterized in that, The step of performing variational mode decomposition based on the preliminary judgment result and the standardized dataset to obtain a multidimensional feature vector, and determining whether the potential abnormal signal is related to cylinder seal failure based on the multidimensional feature vector to obtain the final early fault warning information, includes: performing variational mode decomposition based on the preliminary judgment result and the standardized dataset to obtain a multidimensional feature vector; calculating the dynamic time warping distance between the multidimensional feature vector and the polytropic feature deviation sequence associated with the potential abnormal signal to obtain a synchronization feature value; determining the degree of correlation between the potential abnormal signal and cylinder seal failure based on the synchronization feature value to obtain a failure correlation probability value; and generating early fault warning information containing failure type confirmation if the failure correlation probability value is in a preset high-risk range.

8. The air pump fault prediction method based on multi-source data according to claim 1, characterized in that, The process of determining the trigger timestamp of the anomaly based on the early fault warning information, backtracking and retrieving the environmental feature data based on the trigger timestamp and matching the status description text, and aggregating to generate a structured real-time monitoring report of the air pump status includes: parsing the failure types contained in the early fault warning information and locking the trigger timestamp of the failure type; backtracking and retrieving the environmental feature data based on the trigger timestamp to construct a feature change set containing feature amplitude fluctuations and trend directions; mapping the feature change set to a preset fault semantic description library, matching the status description text, and aggregating to generate a structured real-time monitoring report of the air pump status.

9. A fault prediction system for an air pump based on multi-source data, characterized in that, include: The data acquisition module is used to collect environmental characteristic data and simulated signals including vibration, temperature, pressure and gas flow during the operation of the air pump, and to perform time series alignment on the simulated signals to obtain an initial operating state dataset. The standard processing module is used to perform wavelet decomposition to remove noise from the initial running state dataset, obtain a reconstructed signal sequence, and perform local weighted regression scatter smoothing processing based on the reconstructed signal sequence to obtain a standardized dataset. The feature vector module is used to perform feature enhancement processing on the small amplitude changes of the polytropic exponential drift based on the standardized dataset to obtain a feature enhancement sequence, and to perform feature regression mapping based on the feature enhancement sequence to obtain a feature vector set that can reflect the continuous trend characteristics. The polyvariable index module is used to calculate the drift rate based on the feature vector set, perform nonlinear fitting based on the drift rate and update the preset evolution probability distribution to generate a polyvariable index that reflects early signs of cylinder seal failure. The preliminary judgment module is used to construct an evolution probability threshold range based on the variable index, calculate the cumulative deviation based on the evolution probability threshold range, and when the cumulative deviation exceeds the preset anomaly tolerance limit, generate a potential anomaly signal and calculate the anomaly confidence level to obtain the preliminary judgment result of the anomaly signal. The early fault module is used to perform variational mode decomposition based on the preliminary judgment result and the standardized dataset to obtain a multidimensional feature vector, and to determine whether the potential abnormal signal is related to cylinder seal failure based on the multidimensional feature vector, so as to obtain the final early fault warning information. The real-time monitoring module is used to determine the trigger timestamp of the abnormality based on the early fault warning information, retrieve the environmental feature data back based on the trigger timestamp and match the status description text, and aggregate to generate a structured real-time monitoring report of the air pump status.

Images