A centrifugal fan fault trend prediction method based on multi-source information fusion

Abstract

This invention belongs to the field of centrifugal fan fault prediction technology. It provides a method for predicting centrifugal fan fault trends based on multi-source information fusion, comprising the following steps: During the operation of the centrifugal fan, by monitoring and statistically analyzing multi-source data, the changing trends of data distribution are captured to identify whether the centrifugal fan is experiencing non-stationary data changes due to equipment aging. Based on a time-space dual-dimensional comparative analysis framework, combined with physical mechanism verification, suspicious aging features are identified by calculating relative deviation and aging trend index quantification indicators. Then, the reliability of the features is verified by constructing an association consistency index using the average correlation coefficient and trend consistency. This provides the ability to locate specific aging components, optimizing the misjudgment problem caused by sensor failure or environmental interference in traditional fault diagnosis, improving the accuracy of aging component identification, and providing a clear target for targeted maintenance.

Description

Technical Field

[0001] This invention belongs to the field of centrifugal fan fault prediction technology, specifically a centrifugal fan fault trend prediction method based on multi-source information fusion. Background Technology

[0002] Centrifugal fans, as key equipment in industrial production, are widely used in ventilation, dust removal, and pneumatic conveying. Their long-term stable operation is crucial for production continuity and safety. During long-term operation, due to factors such as friction, corrosion, and fatigue, the core components of the equipment (such as bearings, impellers, and motors) inevitably age, manifesting as bearing wear, impeller corrosion, and winding aging. These aging processes are not sudden but are accompanied by slow changes in statistical characteristics. For example, bearing wear will cause the average vibration value to gradually increase, and impeller corrosion will cause the wind pressure baseline to gradually decrease. This gradual change in statistical characteristics over time is non-stationarity, which is the core characteristic of equipment deterioration.

[0003] Traditional centrifugal fan failure trend prediction methods are mostly based on the assumption of stationary data, that is, the statistical characteristics (such as mean and variance) of equipment operating data are assumed to remain stable, and model training relies on historical health data. However, in actual operation, the non-stationarity caused by equipment aging will lead to continuous changes in data distribution. For example, the mean of bearing vibration increases from the value when it is in a healthy state. This slow and continuous shift cannot be effectively captured by traditional static models. When the non-stationarity of multi-source data is not identified and processed, the model will continue to rely on outdated health data features for prediction, and will not be able to track the dynamic changes caused by aging. This will cause the model's cognition to lag behind the actual failure development, and thus the prediction of the failure time and severity will lag behind the actual equipment state, missing the best maintenance opportunity.

[0004] Therefore, this invention provides a method for predicting the failure trend of centrifugal fans based on multi-source information fusion. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0006] The technical solution adopted by this invention to solve its technical problem is: a method for predicting the failure trend of centrifugal fans based on multi-source information fusion, comprising the following steps:

[0007] During the operation of the centrifugal fan, by monitoring multi-source data and analyzing statistical characteristics, the changing trend of data distribution can be captured, and it can be identified whether the centrifugal fan causes non-stationary changes in data due to equipment aging.

[0008] If the non-stationary change is caused by equipment aging, then the initial health status data of the equipment is used as the baseline value, and a two-dimensional comparative analysis of time and space is performed.

[0009] The process of comparing and analyzing the time and space dimensions is as follows: First, for the features corresponding to non-stationary changes, the relative deviation of each data point is calculated within a preset time window. The aging trend index is output by analyzing the relative deviation. Suspicious aging features are extracted by the aging trend index. Then, for the suspicious aging components mapped by the suspicious aging features, the correlation consistency index of the sensor in the healthy state is calculated. The reliability of the suspicious aging features corresponding to the suspicious aging components is identified by the correlation consistency index.

[0010] If reliable, then combine physical mechanisms to verify the target component that has aged;

[0011] For target components that are aging, a fault trend prediction strategy is automatically switched to a dynamically updated reinforcement learning model. The reinforcement learning model includes setting a dynamic sliding data window and introducing a health forgetting factor.

[0012] The beneficial effects of this invention are as follows:

[0013] This invention performs real-time monitoring and processing of multi-source data during the operation of centrifugal fans. By using the ADF unit root test to identify non-stationary changes in data caused by equipment aging, it has the ability to capture the progressive aging characteristics of equipment. This optimizes the shortcomings of traditional methods that are difficult to distinguish between real aging and random interference, and lays an accurate preliminary judgment foundation for subsequent fault trend prediction.

[0014] This invention is based on a time and space dual-dimensional comparison and analysis framework, combined with physical mechanism verification. It locks down suspicious aging features by calculating relative deviation and aging trend index quantitative indicators, and then uses the average correlation coefficient and trend consistency to construct an association consistency index to verify the reliability of the features. Thus, it has the ability to locate specific aging components, optimizes the misjudgment problem caused by sensor failure or environmental interference in traditional fault diagnosis, improves the accuracy of aging component identification, and provides a clear target for targeted maintenance.

[0015] This invention is based on a dynamically updated reinforcement learning model strategy, which introduces a dynamic sliding data window and a health forgetting factor. This allows the model to incorporate new data in real time and weaken the influence of old health data, thereby enabling it to adapt to the aging and evolution trend of components. It optimizes the prediction lag problem of static models for non-stationary data, improves the prediction accuracy of failure occurrence time and severity, and provides more scientific decision support for the preventive maintenance of centrifugal fans. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of the steps of a centrifugal fan fault trend prediction method based on multi-source information fusion according to the present invention;

[0018] Figure 2 This is a flowchart of step S20 in the centrifugal fan fault trend prediction method based on multi-source information fusion of the present invention. Detailed Implementation

[0019] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0020] Example 1

[0021] Please see Figure 1 and Figure 2 As shown in the embodiment of the present invention, a method for predicting the fault trend of centrifugal fans based on multi-source information fusion includes the following steps:

[0022] Step S10: During the operation of the centrifugal fan, the changing trend of data distribution is captured through multi-source data monitoring and statistical characteristic analysis to identify whether the centrifugal fan causes non-stationary changes in data due to equipment aging.

[0023] During long-term operation of centrifuges, equipment aging (such as bearing wear and impeller corrosion) will cause slow changes in data statistical characteristics (such as a gradual increase in the average vibration value). The essence of this phenomenon is the "non-stationarity" of the equipment state, that is, the dynamic characteristics of the system change gradually over time. The core purpose of this step is to identify this non-stationarity phenomenon.

[0024] In some embodiments, during the operation of the centrifugal fan, sensors are installed on the core components of the centrifugal fan to collect multi-source data in real time. The core components include, but are not limited to, bearings, impellers, motors, and air inlets.

[0025] The multi-source data includes, but is not limited to: vibration data, temperature data, pressure and flow data, and operating condition data;

[0026] All data is transmitted to the data acquisition terminal via industrial buses (such as Modbus and Profinet), and stored after being aligned with timestamps to form a time-series database containing time-multi-source features.

[0027] The acquired multi-source data is preprocessed, including data cleaning and standardization to eliminate noise and interference;

[0028] For preprocessed multi-source data, a time window is set, and statistical characteristics are extracted within the time window to capture the trend changes of multi-source data over time.

[0029] Among them, statistical characteristics include: basic statistics, frequency domain characteristics, and trend indicators;

[0030] Specifically, the basic statistics include, but are not limited to: calculating the mean, variance, and maximum value of the vibration signal within each time window; the mean and heating rate of the temperature signal; and the baseline value and fluctuation amplitude of the pressure signal.

[0031] Frequency domain features include, but are not limited to: performing Fourier transform on vibration signals to extract the power spectral density of bearing characteristic frequencies and the energy proportion of impeller blade passing frequencies, reflecting the wear or imbalance state of mechanical structures.

[0032] Trend indicators include: calculating the slope of features within a continuous window, such as the daily growth rate of the mean vibration, the weekly decay rate of the pressure baseline, and the rate of change of quantitative statistical characteristics;

[0033] The ADF unit root test method was used to analyze time series data from multiple sources to identify whether statistical characteristics changed significantly over time. The process was as follows:

[0034] The ADF unit root test is used to determine whether a time series has a "unit root": if the series has a unit root, it is a non-stationary series, and its statistical characteristics change over time, which is consistent with the non-stationarity characteristics caused by equipment aging; if there is no unit root, it is a stationary series, and its statistical characteristics are stable with no significant aging-driven trend changes.

[0035] A regression model based on the ADF test was established for the statistical characteristic sequence of multi-source data of centrifugal fans.

[0036] For example, the multi-source data statistical feature sequence of a centrifugal fan is: vibration mean - time series, bearing characteristic frequency PSD - time series;

[0037] The form of the regression model: ;in, Let be the statistical characteristic value at time t; The first difference of the sequence; For constant terms; For time trend items; The coefficient for the trend term. The coefficients to be estimated are... The characteristic value is lagged by one period. These are lagged difference terms; Here, k represents the random error term, and k is the lag order.

[0038] Set the test hypothesis: Null hypothesis: The sequence has a unit root (γ=0), that is, the sequence is non-stationary;

[0039] Alternative hypothesis: The sequence does not have a unit root (γ<0), that is, the sequence is stationary;

[0040] The parameters of the regression model are estimated using the least squares method, and the ADF statistic, i.e., the test index of the parameter to be estimated, is calculated.

[0041] Compare the ADF statistic with the critical value at the pre-set significance level, which is determined by the theoretical distribution of the ADF test;

[0042] If the ADF statistic is less than the critical value, the null hypothesis is rejected and the series is determined to be stationary.

[0043] If the ADF statistic is greater than or equal to the critical value, the null hypothesis cannot be rejected, and the series is determined to be non-stationary.

[0044] It should be noted that the ADF test method described above can only determine whether the series is stationary. It is necessary to combine this with the significance of the trend term in the regression model to determine whether the data non-stationarity is caused by equipment aging. The process is as follows:

[0045] If the sequence is determined to be non-stationary and the trend term is significant, i.e., the t-test shows a p-value < 0.05;

[0046] The trend term is greater than zero. Combined with the physical mechanism verification, it is determined to be a non-stationary change caused by the equipment.

[0047] For example, the physical mechanism is that if the trend term of the vibration mean sequence is positive, it conforms to the physical law that bearing wear will cause the vibration mean to gradually increase.

[0048] The trend term is less than zero. Combined with the physical mechanism verification, it is determined to be a non-stationary change caused by the equipment.

[0049] For example, the physical mechanism is that if the trend term of the pressure baseline sequence is negative, it matches the characteristic that impeller corrosion causes the wind pressure to gradually decrease.

[0050] If the sequence is determined to be non-stationary and the trend term is not significant, i.e., the t-test shows a P-value ≥ 0.05, it is not determined to be a non-stationary change caused by the device. Non-stationarity may be caused by random fluctuations, sensor noise, or non-aging sudden interference.

[0051] For example, the non-stationarity determination of the mean sequence of vibrations caused by bearing wear;

[0052] After three years of operation, the vibration acceleration signal in the X direction was collected by a vibration sensor and preprocessed to form a "vibration mean-time" sequence, in g, with a sampling interval of 1 hour, for a total of 30 days of data.

[0053] ADF unit root test: Construct a regression model for the "vibration mean-time" series and determine the lag order k=2 using the AIC criterion;

[0054] The calculated ADF statistic is -2.35, and the critical value at the 5% significance level is -3.43.

[0055] Since the ADF statistic (-2.35) is greater than or equal to the critical value (-3.43), the null hypothesis cannot be rejected, and the series is determined to be a non-stationary series.

[0056] A t-test was performed on the trend term coefficient, and the p-value was calculated to be 0.02 (< 0.05), indicating that the trend term was significant.

[0057] The estimated trend term coefficient is 0.015 (>0), which means that the mean vibration value shows a significant upward trend over time (an average daily increase of 0.015g).

[0058] The physical mechanism is that bearing wear leads to an increase in the clearance between the rolling elements and the inner and outer rings, which intensifies friction and impact. This is directly manifested as a gradual increase in the mean vibration over time, consistent with the statistical result of "trend term > 0".

[0059] The non-stationarity of the vibration mean-time series was determined to be caused by the aging (wear) of bearing #1.

[0060] Step S20: If the non-stationary change is caused by equipment aging, then the initial health status data of the equipment is used as a benchmark to perform a two-dimensional comparative analysis of time and space, and the target component that has aged is verified by combining the physical mechanism.

[0061] In some embodiments, the non-stationary feature sequence determined in step S10 above to be caused by equipment aging is obtained, along with the corresponding statistical feature type, associated components, and trend of change.

[0062] For example, vibration mean-time series (#1 bearing X direction): non-stationary, trend term β=0.015 (>0), P=0.02;

[0063] Pressure baseline-time series (outlet): non-stationary, trend term β=-5.2 (<0), P=0.03;

[0064] The initial health status data of the centrifugal fan was extracted and used as the comparison benchmark data. The comparison benchmark data was then preprocessed.

[0065] The benchmark data includes statistical features corresponding to non-stationary features;

[0066] For example: the baseline of the average vibration of bearing #1 is 2.5g, the baseline of the air outlet pressure is 1200Pa, and the baseline of the bearing housing temperature is 45℃;

[0067] The process of comparative analysis based on the time dimension is as follows:

[0068] For each non-stationary feature, calculate the relative deviation between the current value and the benchmark value;

[0069] The relative deviation is calculated as follows: within the time window, the absolute value of the difference between the current value and the benchmark value corresponding to each data point is calculated, and the ratio of the absolute value of the difference to the benchmark value is obtained.

[0070] The difference between the relative deviation and the preset relative deviation threshold is calculated, and the ratio of the difference to the preset relative deviation threshold is calculated to obtain the severity of deviation from the benchmark. The proportion of data points with a positive deviation from the benchmark severity to the total number of data points is calculated as the percentage of data points exceeding the benchmark.

[0071] Within the time window, the relative deviations of adjacent data points are calculated, data points with positive differences are extracted, and the proportion of data points with positive differences in the total number of data points is calculated as the proportion of trend growth.

[0072] The aging trend index is calculated by multiplying the percentage of deviations exceeding the standard with the percentage of trend increases.

[0073] It should be noted that the percentage of deviations exceeding the standard reflects the range of features that deviate from the benchmark value and exceed the preset threshold within the time window. The higher the percentage, the more times the feature is in a significantly abnormal state and the wider the range of deviation from the health state of the equipment. The percentage of trend-increasing numbers reflects the direction of the relative deviation over time. The higher the percentage, the greater the proportion of deviations at adjacent times showing an increasing trend, and the more the abnormality of the feature is continuously deteriorating, which is consistent with the physical law of gradual performance degradation during equipment aging.

[0074] The ratio of the number of deviations exceeding the standard and the ratio of the number of deviations showing an increasing trend are used to characterize the coupling strength between the range of abnormalities and the continuous deterioration trend from both static and dynamic dimensions.

[0075] If the aging trend index is greater than the aging trend index limit, the corresponding non-stationary feature is determined to be a suspicious aging feature. The components corresponding to the suspicious aging features are listed as a set of suspicious aging components through the mapping relationship between features and components.

[0076] For example, the average vibration value in the X direction corresponds to the drive end bearing, and the air outlet pressure baseline corresponds to the impeller;

[0077] If the aging trend index is less than or equal to the aging trend index limit, the corresponding non-stationary feature is determined to be a non-suspicious aging feature.

[0078] The aging trend index serves several purposes: First, by multiplying the percentage of deviations exceeding the standard by the percentage of trend-increasing deviations, it couples the abnormal range and deterioration trend of non-stationary features into a single quantitative indicator, intuitively reflecting the severity of the impact of aging on features. Second, by comparing with preset aging trend index limits, it accurately determines which non-stationary features are suspicious aging features, providing a basis for subsequently identifying related components. If the index exceeds the limit, the corresponding feature is marked as suspicious, and then a set of suspicious aging components is listed through the mapping relationship between features and components, reducing the omission of real aging signals or misjudgment of irrelevant features. Third, this index integrates information from both static and dynamic dimensions, ensuring the identification of features that significantly deviate from the healthy state while capturing the trend of abnormal and continuous deterioration, which conforms to the physical law of gradual performance degradation during equipment aging. It provides a reliable quantitative judgment standard for time-dimensional comparative analysis, improving the scientificity and accuracy of aging feature identification.

[0079] The process of comparative analysis based on spatial dimensions is as follows:

[0080] For any one of the suspected aging components in the set of suspected aging components;

[0081] By using the Pearson correlation coefficient, the correlation strength of different sensor data in the same component of the aging equipment is quantified to determine whether the non-stationary characteristics are caused by the actual aging of the component, rather than by sensor failure or other reasons.

[0082] For any combination of two different sensors in the same component, calculate the Pearson correlation coefficient under healthy conditions, and average the Pearson correlation coefficients of all combinations to obtain the average correlation coefficient.

[0083] The number of sensors in a component containing suspicious aging characteristics is counted and compared with the total number of sensors in that component to obtain the trend consistency.

[0084] The correlation consistency index is obtained by multiplying the average correlation coefficient by the trend consistency.

[0085] The role of the correlation consistency index is as follows: First, by integrating the average correlation coefficient under healthy conditions with the consistency of the current trend, it quantifies whether non-stationary characteristics are caused by actual component aging rather than sensor failure, interference, or other factors, providing a core basis for distinguishing between real aging and false anomalies. Second, the higher the value, the higher the physical correlation between the current abnormal characteristics and the healthy state of the component, and the simultaneous abnormal trend of multiple sensors indicates stronger feature credibility, which can screen out reliable suspicious aging characteristics, eliminate misjudgments caused by single sensor failures, and improve the accuracy of spatial dimension analysis. Third, only components with a correlation consistency index exceeding the limit are considered reliable for their corresponding suspicious aging characteristics. This is then combined with physical mechanisms to verify and output the final aging target component, ensuring the rigor of the derivation logic from suspicious components to identified aging components, providing important support for accurately locating aging components and reducing invalid analysis.

[0086] It should be noted that the correlation consistency index is the coupling of the baseline correlation under healthy conditions and the current abnormal trend. The core is whether the current non-stationary characteristics are a true reflection of the abnormal state of the component itself. The higher the value, the higher the credibility of the non-stationary characteristics, providing a key basis for outputting aging components.

[0087] If the correlation consistency index is less than or equal to the correlation consistency index limit, the suspected aging characteristics corresponding to the component are determined to be unreliable.

[0088] If the correlation consistency index is greater than the correlation consistency index limit, then the suspected aging characteristics corresponding to the component are determined to be reliable.

[0089] If the suspected aging characteristics of the component are reliable, then the suspected aging component that conforms to the physical mechanism is output as the target component that has aged, in conjunction with the physical mechanism verification.

[0090] It should be noted that the physical mechanism is a well-known physical manifestation familiar to those skilled in the art. For example, bearing aging: the physical law is "wear leads to an increase in the gap between the rolling element and the inner and outer rings → an increase in the average vibration value and an increase in the energy of high-frequency vibration (bearing characteristic frequency) → increased friction leading to an increase in temperature". If the suspicious feature is "the average vibration value and temperature increase simultaneously, and the proportion of high-frequency vibration energy increases", then it is consistent with the bearing aging law.

[0091] Impeller aging: The physical law is that "corrosion / dust accumulation leads to impeller imbalance → wind pressure (pressure baseline) decreases and flow fluctuations increase → the proportion of vibration energy at the blade passing frequency decreases". If the suspicious feature is "pressure baseline decreases + flow fluctuations increase", it is consistent with the impeller aging law.

[0092] Motor aging: The physical law is "winding aging → increased current fluctuations and increased temperature → decreased output efficiency (weakened correlation with load)". If the suspicious characteristic is "current fluctuations + simultaneous increase in temperature", it is consistent with the motor aging law.

[0093] The common goal of steps S20 and S10 is to capture the aging state of the equipment and provide a basis for predicting failure trends. Step S10 focuses on identifying whether the non-stationarity of data is caused by equipment aging. Step S20, on the other hand, further realizes the precise location of specific aging components and determines whether the equipment has entered the aging stage. This is a deeper analysis of the equipment status from the presence of aging phenomena to the clear location and severity of aging.

[0094] Step S30: For the target component that has aged, automatically switch to a dynamically updated reinforcement learning model strategy to predict the failure trend;

[0095] Among them, the dynamically updated reinforcement learning model strategy includes the following core mechanisms: dynamic sliding data window and healthy forgetting factor setting mechanism;

[0096] By incorporating newly collected multi-source data in real time and dynamically updating model parameters, the prediction lag problem caused by static models trained on historical health data can be effectively reduced, and the prediction accuracy of equipment failure evolution trends in the accelerated aging stage can be improved.

[0097] For the target component that has aged as output in step S20 above, a reinforcement learning model framework is constructed, which includes: state space, action space and reward function.

[0098] State space: The target component's multi-source features are used as input, including: real-time statistical features: current state parameters such as vibration mean, temperature mean, and pressure baseline within the time window;

[0099] Trend characteristics: aging trend index, relative deviation growth rate, and correlation consistency index;

[0100] Physical constraint characteristics: correlation parameters based on physical mechanisms;

[0101] Action space: Defines the dynamic adjustment strategy for model parameters, including: dynamic adjustment of the sliding window size, for example, expanding from a 2-hour window to 24 hours to adapt to data changes during the accelerated aging phase;

[0102] Weighting of multi-source features, for example, increasing the weight of high-frequency vibration features and weakening noise interference features;

[0103] Adjusting the forecast step size, for example, shortening the forecast from 7 days to 3 days, improves the accuracy of short-term forecasts;

[0104] Reward function: aims to minimize prediction error;

[0105] The formula is: ;

[0106] The relative prediction error is the ratio of the deviation between the predicted value and the actual value; the prediction lag time is the difference between the predicted fault occurrence time and the actual occurrence time; a and b are weighting coefficients (calibrated based on historical fault data, such as a=0.6 and b=0.4), which reward prediction results with "high accuracy and low lag".

[0107] The first specific mechanism for the real-time update of the dynamic sliding data window is as follows:

[0108] The data window is dynamically adjusted according to the aging stage of the target component to ensure that the model captures the latest state;

[0109] Set the initial data window length based on the historical aging data of the target component;

[0110] Optionally, bearing aging data is taken over 30 days, and impeller corrosion data is taken over 60 days, to match the aging evolution cycle of different components;

[0111] When newly collected multi-source data is incorporated into the time series database, if the average growth rate of the relative deviation of the target component within the initial data window length exceeds the preset growth rate threshold (2%), the data window will automatically slide forward by 1 time unit (e.g., 1 hour), removing the oldest data and retaining the latest data.

[0112] Among them, the core function of the average growth rate of relative deviation is to quantify the overall deterioration rate of the non-stationary characteristics of the target component within the initial data window;

[0113] Secondly, the specific mechanism for setting the healthy forgetting factor is as follows:

[0114] The formula for calculating the healthy forgetting factor is: ;in, The healthy forgetting factor is denoted by k, which is the decay coefficient, typically taken as 0.5-1.0, and m is the aging trend index.

[0115] For example, when the aging trend index is 0.8, the decay coefficient is 0.6, and the health forgetting factor is 0.61, then the weight of health data is only 61% of the current data.

[0116] The weight of the data at time n within the sliding data window is N represents the current time, thus making the weight of newly collected multi-source data higher than that of previously collected multi-source data, thereby enhancing the fault trend prediction model's ability to capture aging trends.

[0117] For example, if the current time is the 10th hour (N=10), and the window contains data from the 6th to the 10th hour (a total of 5 times), then n can take the values ​​6, 7, 8, 9, and 10.

[0118] The 10th hour (n=10) is the current data, with a weight of 1;

[0119] The data weight for the 9th hour (n=9) is 0.61;

[0120] The data weight for the 6th hour (n=6) is 0.13 (lowest weight).

[0121] Thirdly, specifically, the dynamic training and parameter update process of the reinforcement learning model is as follows:

[0122] Multi-source data within a dynamically sliding data window is mapped to state-space parameters and input into the reinforcement learning agent.

[0123] The agent selects the optimal strategy from the action space based on the current state and outputs the fault trend prediction result, including:

[0124] The time window for the failure to occur (e.g., "bearing jamming may occur within the next 15 days").

[0125] The severity level of the fault includes mild aging, moderate aging, and severe aging;

[0126] Key feature evolution curves, such as the predicted fitting curve of the vibration mean over time;

[0127] The actual new data collected is compared with the prediction results to calculate the reward value;

[0128] If the prediction error is less than 5% and there is no lag, a positive reward will be given (>).

[0129] If the prediction error is greater than 10% or the lag is greater than 2 days, a negative reward (<) will be given.

[0130] Based on the reward value, the agent's policy network parameters are updated using the Temporal Difference Algorithm (TD algorithm) to make the model prioritize retaining action policies corresponding to high rewards, such as optimized window size and feature weights.

[0131] The prediction results are compared with the physical mechanism of the target component. If they match the physical mechanism, the prediction is confirmed to be effective. If they do not match the physical mechanism, the model is retrained and the weight of the physical constraint features is increased.

[0132] Based on the predicted severity level of the fault, output the corresponding warning:

[0133] Mild aging: A trend warning will be sent, and it is recommended to increase the monitoring frequency;

[0134] Moderate aging: Push maintenance alerts and prompts you to prepare spare parts;

[0135] Severe aging: An emergency warning will be sent, and it is recommended to shut down the machine for maintenance to prevent the fault from escalating.

[0136] After each round of prediction, the newly collected actual data is incorporated into the dynamic window. Through continuous iteration, the model is made to always adapt to the aging evolution trend of the target component, thereby optimizing the lag problem of the static model for "non-stationary" data and ultimately achieving accurate prediction of the time and severity of the failure.

[0137] This embodiment has at least the following effects: by monitoring and analyzing multi-source data and statistical characteristics, combined with the ADF unit root test, it identifies the non-stationary changes in data caused by equipment aging, distinguishes between non-stationarity driven by real aging and random interference and sensor noise, reduces the interference of invalid data in subsequent analysis, lays a reliable preliminary judgment foundation for the entire fault trend prediction process, and ensures that in-depth analysis is carried out only on the characteristics related to equipment aging.

[0138] Based on time and space dual-dimensional comparative analysis, suspicious aging features are screened by calculating relative deviation and aging trend index, and the reliability of features is verified by combining correlation consistency index. Finally, the target components that have aged are identified through physical mechanism verification, effectively eliminating non-aging factors such as sensor failure and environmental interference, improving the accuracy of aging component identification, clarifying the specific objects of fault prediction, reducing the waste of resources caused by blind maintenance, and providing accurate basis for targeted maintenance.

[0139] For aging target components, a dynamically updated reinforcement learning model strategy is adopted. By dynamically sliding the data window to incorporate new data in real time and using a health forgetting factor to weaken the influence of outdated health data, the model can adapt to the non-stationary evolution trend of component aging. At the same time, through dynamic training and parameter updates of reinforcement learning, the prediction lag problem of traditional static models for non-stationary data is optimized, improving the prediction accuracy of failure occurrence time and severity, providing scientific decision support for the preventive maintenance of centrifugal fans, and reducing the risk of failure escalation.

[0140] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the failure trend of centrifugal fans based on multi-source information fusion, characterized in that: Includes the following steps: During the operation of the centrifugal fan, by monitoring multi-source data and analyzing statistical characteristics, the changing trend of data distribution can be captured, and it can be identified whether the centrifugal fan causes non-stationary changes in data due to equipment aging. If the non-stationary change is caused by equipment aging, then the initial health status data of the equipment is used as the baseline value, and a two-dimensional comparative analysis of time and space is performed. The process of comparing and analyzing the time and space dimensions is as follows: First, for the features corresponding to non-stationary changes, the relative deviation of each data point is calculated within a preset time window. The aging trend index is output by analyzing the relative deviation. Suspicious aging features are extracted by the aging trend index. Then, for the suspicious aging components mapped by the suspicious aging features, the correlation consistency index of the sensor in the healthy state is calculated. The reliability of the suspicious aging features corresponding to the suspicious aging components is identified by the correlation consistency index. The process for obtaining the suspected aging components is as follows: Within the time window, calculate the absolute value of the difference between the current value and the benchmark value for each data point, and calculate the ratio of the absolute value of the difference to the benchmark value as the relative deviation; By analyzing the relative deviation, the percentage of deviations exceeding the standard and the percentage of deviations showing an increasing trend are calculated, and the product is processed to output the aging trend index. If the aging trend index exceeds the standard, the corresponding non-stationary feature is determined to be a suspicious aging feature. The components corresponding to the suspicious aging features are listed through the feature-component mapping relationship. The process of identifying the reliability of suspicious aging features corresponding to suspicious aging components through the correlation consistency index is as follows: For any two different sensor combinations in the same suspected aging component, calculate the Pearson correlation coefficient in the healthy state, and average the Pearson correlation coefficients of all combinations to obtain the average correlation coefficient. The trend consistency is obtained and multiplied with the average correlation coefficient to calculate the association consistency index. If the correlation consistency index is greater than the limit, the suspected aging characteristics corresponding to the component are determined to be reliable. Combine physical mechanisms to verify the target components that have aged; The physical mechanisms include: bearing aging, impeller aging, and motor aging; For target components that are aging, a fault trend prediction strategy is automatically switched to a dynamically updated reinforcement learning model. The reinforcement learning model includes setting a dynamic sliding data window and introducing a health forgetting factor.

2. The centrifugal fan fault trend prediction method based on multi-source information fusion according to claim 1, characterized in that: The process for identifying whether the centrifuge's data non-stationary changes are due to equipment aging is as follows: The ADF unit root test method was used to analyze time series data from multiple sources and to identify non-stationary changes. A regression model based on the ADF test was established for the statistical characteristic sequence of multi-source data of centrifugal fans. The parameters of the regression model are estimated using the least squares method, and the ADF statistic is calculated. Compare the ADF statistic with the critical value at the pre-set significance level, which is determined by the theoretical distribution of the ADF test; If the ADF statistic is greater than or equal to the critical value, the sequence is determined to be non-stationary. If the sequence is determined to be non-stationary and the trend term is significant, i.e., the t-test shows a p-value < 0.05, it is determined to be a non-stationary change caused by the equipment.

3. The centrifugal fan fault trend prediction method based on multi-source information fusion according to claim 1, characterized in that: The calculation process for the percentage of deviations exceeding the standard is as follows: The deviation rate between the relative deviation and the preset relative deviation threshold is used to obtain the deviation severity from the baseline. The proportion of data points with a positive deviation severity from the baseline is calculated as the percentage of data points exceeding the deviation threshold.

4. The centrifugal fan fault trend prediction method based on multi-source information fusion according to claim 1, characterized in that: The calculation process for the percentage of the trend growth is as follows: The relative deviations of adjacent data points are calculated to extract data points with positive differences. The proportion of data points with positive differences in the total number of data points is then used as the percentage of trend growth.

5. The centrifugal fan fault trend prediction method based on multi-source information fusion according to claim 1, characterized in that: The calculation process for the trend consistency is as follows: The number of sensors containing suspicious aging characteristics in a suspected aging component is counted, and the ratio is calculated with the total number of sensors in that component to obtain the trend consistency.

6. The centrifugal fan fault trend prediction method based on multi-source information fusion according to claim 1, characterized in that: The dynamically updated reinforcement learning model strategy includes: For the target component that is aging, a reinforcement learning model framework is constructed, which includes: state space, action space and reward function; Set the initial data window length based on the historical aging data of the target component; When newly collected multi-source data is incorporated into the time series database, if the average growth rate of the relative deviation of the target component exceeds the standard within the initial data window length, the data window will automatically slide forward by 1 time unit to form a dynamic sliding data window, removing the oldest data and retaining the latest data. Introducing a healthy forgetting factor, the formula for calculating the healthy forgetting factor is as follows: ;in, , where k is the decay coefficient and m is the aging trend index; The weight of the data at time n within the sliding data window is N represents the current time, thus making the weight of newly collected multi-source data higher than that of previously collected multi-source data, thereby enhancing the model's ability to capture aging trends.

7. The centrifugal fan fault trend prediction method based on multi-source information fusion according to claim 6, characterized in that: The dynamically updated reinforcement learning model strategy also includes: dynamic training and parameter updating of the reinforcement learning model, as well as output of fault severity level; Multi-source data within a dynamically sliding data window is mapped to state-space parameters and input into the reinforcement learning agent. The agent selects the optimal strategy from the action space based on the current state and outputs the fault trend prediction result, including: Fault occurrence time window, fault severity level; The reward value is calculated by comparing the actual new data collected with the prediction results. Based on the reward value, the agent's policy network parameters are updated using a temporal difference algorithm; The prediction results are compared with the physical mechanism of the target component. If they match the physical mechanism, the prediction is confirmed to be valid. If they do not match the physical mechanism, the model is retrained and the weight of the physical constraint features is increased.

Images