Equipment fault prediction method and system based on vibration mechanism and machine learning

Abstract

The invention discloses an equipment fault prediction method and system based on a vibration mechanism and machine learning, and the method comprises the steps: firstly building an initial vibration characteristic frequency spectrum library containing various state characteristic frequency spectrums and corresponding early warning values according to an equipment mechanism, and achieving the quick deployment under the condition of no historical data; then, carrying out personalized calibration on the initial vibration characteristic frequency spectrum library by utilizing actual fault-free operation data of the equipment, and realizing high-precision adaptation from a universal model to specific equipment based on a training machine learning model; in the operation stage, the machine learning model is awakened to perform deep diagnosis only when the actual vibration characteristic frequency spectrum exceeds the corresponding early warning value, and otherwise, the machine learning model is dormant; according to the method, high diagnosis precision is guaranteed, meanwhile, continuous consumption of data storage and computing resources of the system is greatly reduced, and unification of rapid deployment, accurate prediction and low-cost operation is achieved.

Description

Technical Field

[0001] This application relates to the field of equipment maintenance, specifically to an equipment fault prediction and system based on vibration mechanism and machine learning. Background Technology

[0002] In the field of predictive maintenance of equipment, vibration signal analysis and diagnosis are the core technologies. The current mainstream technical solutions are mainly divided into two categories, but each of them has inherent defects that are difficult to overcome, and they cannot achieve a good balance between practicality, economy and accuracy.

[0003] The first category is diagnostic methods based on pure physical mechanism models. These methods establish theoretical vibration models based on equipment dynamics principles and preset fixed fault characteristic spectra and alarm thresholds. Their advantage lies in the fact that they do not require historical fault data and can be deployed quickly. However, their drawbacks are significant: First, the theoretical model cannot cover individual differences in equipment due to manufacturing tolerances, installation conditions, foundation characteristics, and operating loads, leading to a significant deviation between the preset normal baseline and the actual health state of a specific piece of equipment, resulting in numerous false alarms. Second, the mechanism model is fixed and lacks the ability to learn from data and adapt to slow changes in operating conditions. When the vibration characteristics of the equipment drift due to normal aging or process adjustments, the original fixed thresholds become inapplicable. The system loses its monitoring value due to continuous false alarms, or the thresholds may be artificially increased to avoid false alarms, thus losing sensitivity to real faults and causing missed alarms.

[0004] The second category is diagnostic methods based on pure artificial intelligence (AI) or machine learning, such as the Chinese patent application CN202010278878.7. This method trains a model by collecting a large amount of historical operating data of equipment (including normal and various fault states), which theoretically can achieve high-precision fault identification and prediction. However, it faces severe challenges in implementation: model training heavily relies on sufficient, high-quality, and labeled fault data. For most industrial sites, acquiring complete data covering various faults is extremely costly and time-consuming, resulting in a lengthy pre-launch phase that prevents rapid deployment. In addition, to achieve real-time monitoring, such systems typically require continuous collection of vibration data and uninterrupted online AI inference analysis, which brings huge data storage pressure, network transmission load, and computing resource consumption, resulting in high implementation and maintenance costs and making large-scale deployment difficult on resource-constrained edge devices. Summary of the Invention

[0005] This application provides a method and system for predicting equipment failures based on vibration mechanism and machine learning, which can solve the technical problems in the prior art where pure mechanism models are inaccurate in diagnosis and pure AI models have long deployment cycles and high resource consumption during continuous operation.

[0006] In a first aspect, embodiments of this application provide a method for predicting equipment failures based on vibration mechanisms and machine learning, comprising: Based on the equipment type, operating conditions, and dynamic principles, vibration mechanism models corresponding to each component of the equipment are constructed; vibration characteristic spectra simulated by each vibration mechanism model under normal conditions, various deterioration conditions, and various fault conditions are obtained, and corresponding warning values ​​are set with the vibration characteristic spectra under each fault condition as a reference to obtain an initial vibration characteristic spectrum library. The initial vibration feature spectrum library is calibrated using the vibration feature spectrum of each component of the equipment during the initial learning period when it operates without faults, thereby generating a personalized vibration feature spectrum library; then, a machine learning model for degradation prediction and fault diagnosis is trained based on the personalized vibration feature spectrum library. The actual vibration characteristic spectrum of each component of the device after the initial learning period is obtained and compared with the personalized vibration characteristic spectrum library. If the actual vibration characteristic spectrum exceeds the corresponding warning value, the current component triggers an alarm. In response to the alarm, the machine learning model is called to perform degradation prediction and fault diagnosis for the current component. Otherwise, the machine learning model is put into a dormant state.

[0007] Preferably, a frequency-controlled temperature and vibration integrated sensor deployed at a preset key part of the device is used to collect vibration time-domain waveform data and convert it into vibration characteristic spectrum; the frequency response range of the frequency-controlled temperature and vibration integrated sensor covers 0.4kHz to 12kHz; The vibration characteristic spectrum includes a high-frequency spectrum, a mid-frequency spectrum, and a low-frequency spectrum; the high-frequency spectrum is used to monitor early-stage equipment failures, the mid-frequency spectrum is used to monitor mid-stage equipment failure characteristic frequencies, and the low-frequency spectrum is used to monitor severe equipment failures and structural vibration characteristic frequencies.

[0008] Preferred options also include: Synchronously collect temperature data from preset key components of the equipment; A machine learning model for degradation prediction and fault diagnosis is obtained by training a personalized vibration feature spectrum library and the temperature data.

[0009] Preferably, when the warning value is a preset amplitude, the actual amplitude of the actual vibration characteristic spectrum at the preset characteristic frequency is obtained and compared with the preset amplitude for analysis; When the warning value is a preset growth rate, the actual growth rate of the amplitude of the actual vibration characteristic spectrum at the preset characteristic frequency within a preset window time is obtained and compared with the preset growth rate for analysis. When the warning value is a preset growth rate, the actual growth rate of the amplitude of the actual vibration characteristic spectrum at the preset characteristic frequency within a preset window time is obtained and compared with the preset growth rate.

[0010] Preferably, the initial vibration characteristic spectrum library is calibrated using the vibration characteristic spectrum of each component of the equipment during the initial learning period when they are operating without faults. This specifically includes the following steps: Obtain the spectrum of the first target vibration feature that is persistent and has a stable amplitude from the vibration data of the current component under fault-free operation; Extract the amplitude at a preset normal characteristic frequency from the vibration characteristic spectrum of the first target, and use it as the current normal amplitude; Obtain the vibration characteristic spectrum of the current component in its normal state from the initial vibration characteristic spectrum library; extract the amplitude at a preset normal characteristic frequency from the vibration characteristic spectrum in the normal state as a reference normal amplitude; Based on the current normal amplitude and the reference normal amplitude, an amplitude calibration coefficient is calculated; using the amplitude calibration coefficient, the vibration characteristic spectrum of the current component in the normal state, each deterioration state, and each fault state in the initial vibration characteristic spectrum library is calibrated; then, the warning values ​​in each fault state are synchronously calibrated. Repeat the above steps to complete the calibration of the initial vibration characteristic spectrum library.

[0011] Preferably, after responding to the alarm and invoking the machine learning model to predict device faults, the method further includes the following steps: Obtain the vibration characteristic spectrum of the second target after the current component alarms; If the output of the machine learning model is a deteriorated state, and the amplitude of the second target vibration characteristic spectrum at the fault characteristic frequency is stable and greater than the original warning value, then the warning value is raised. If the output of the machine learning model indicates a fault state, the warning value remains unchanged.

[0012] Preferably, after raising the warning value and performing maintenance on the equipment, the third target vibration characteristic spectrum of the current component is obtained; If the amplitude of the vibration characteristic spectrum of the third target at the fault characteristic frequency is stable and less than the upward-adjusted warning value, then the upward-adjusted warning value will be lowered based on the amplitude of the vibration characteristic spectrum of the third target at the fault characteristic frequency at this time.

[0013] Preferably, raising the warning value specifically includes the following steps: Calculate the amplitude of the second target vibration characteristic spectrum at the fault characteristic frequency, and the amplitude deviation ratio between it and the warning value before adjustment; determine the adjustment coefficient corresponding to the amplitude deviation ratio according to a preset mapping relationship; use the adjustment coefficient to calculate the warning value before adjustment to obtain the warning value after adjustment; or, Raising the warning threshold specifically includes the following steps: Obtain the confidence level when the output of the machine learning model is in a deteriorated state; determine the upward adjustment range of the warning value based on the comparison result between the confidence level and the preset confidence threshold; wherein, the higher the confidence level, the greater the corresponding upward adjustment range.

[0014] Preferably, the machine learning model is invoked to predict device degradation and diagnose faults, specifically including: In response to the alarm, the machine learning model is activated; The vibration characteristic spectrum of the component that triggered the alarm after the alarm was collected is used as the vibration characteristic spectrum of the second target. The vibration characteristic spectrum of the second target is input into the machine learning model to output diagnostic results about the fault location, fault type and degree of deterioration; Based on the diagnostic results, determine and update the warning value corresponding to the vibration characteristic spectrum of the second target; The second target vibration characteristic spectrum, its corresponding diagnostic results, and the updated warning value are added to the personalized vibration characteristic spectrum library.

[0015] Secondly, embodiments of this application provide a device fault prediction system based on vibration mechanisms and machine learning, comprising: The first module is used to construct vibration mechanism models corresponding to each component of the equipment based on the equipment type, operating conditions and dynamic principles; obtain the vibration characteristic spectrum of each vibration mechanism model under normal conditions, various deterioration conditions and various fault conditions, and set corresponding warning values ​​with the vibration characteristic spectrum under each fault condition as a reference to obtain the initial vibration characteristic spectrum library. The second module is used to calibrate the initial vibration feature spectrum library by using the vibration feature spectrum of each component of the equipment during the initial learning period when it operates without faults, and generate a personalized vibration feature spectrum library; then, a machine learning model for degradation prediction and fault diagnosis is trained based on the personalized vibration feature spectrum library. The third module is used to obtain the actual vibration characteristic spectrum of each component of the equipment after the initial learning period, and compare and analyze it with the personalized vibration characteristic spectrum library; if the actual vibration characteristic spectrum exceeds the corresponding warning value, the current component triggers an alarm, and in response to the alarm, the machine learning model is called to perform degradation prediction and fault diagnosis for the current component; otherwise, the machine learning model is put into a dormant state.

[0016] The beneficial effects of the technical solutions provided in this application include: By constructing vibration mechanism models for each component of the equipment based on its inherent mechanisms, and obtaining simulated vibration characteristic spectra for various states, an initial vibration characteristic spectrum library is established. This system completely disregards actual fault data collected during the equipment's historical operation, instead relying on publicly available physical laws, equipment drawings, and dynamic principles for prior modeling. This allows the system to immediately possess basic fault diagnosis and early warning capabilities during the initial deployment phase, even before any historical fault data has been accumulated for the equipment. The system avoids the lengthy data waiting and training cycle required for pure AI models, thus shortening deployment and debugging time.

[0017] The initial library is calibrated using the vibration characteristic spectrum of the equipment during its fault-free operation in the initial learning period, generating a personalized vibration characteristic spectrum library. A machine learning model is then trained based on this library. The calibration process essentially quantifies and eliminates systematic deviations caused by individual differences in installation, manufacturing, and foundation of the equipment. Subsequently, the machine learning model is trained using the calibrated, more realistic personalized vibration characteristic spectrum library. This establishes the feature-state mapping relationship of the machine learning model on an accurate benchmark that has been individually corrected, improving the accuracy of fault identification and degradation judgment and significantly reducing the false alarm rate caused by individual differences.

[0018] For the majority of the device's normal uptime, the system only performs comparisons between spectral feature values ​​and warning thresholds, without continuously storing highly sampled raw waveform data or calling complex machine learning models for inference. This fundamentally eliminates the enormous burden of full data storage and uninterrupted computation inherent in traditional solutions; the resource-intensive machine learning model is only activated for computation when an alarm signal is detected, resulting in a reduction of the system's average data storage, computational resource utilization, and energy consumption by an order of magnitude.

[0019] In summary, the above is not a simple superposition of mechanism models and machine learning models, but rather a combination of the advantages of both in a temporal and conditional manner, while effectively avoiding their respective disadvantages. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the equipment failure prediction method based on vibration mechanism and machine learning proposed in this application. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0023] This application aims to overcome the shortcomings of the prior art and solve the following core problems: How can we provide a method for predicting equipment failures that can be quickly deployed and have initial diagnostic capabilities in the absence of a large amount of historical failure data, while also being able to accurately adapt to the individual differences and state changes of specific equipment through autonomous learning, and ultimately achieve high-precision failure early warning? Throughout the entire operation, how can we significantly reduce the continuous consumption of data storage space and computing resources, thereby forming a predictive maintenance solution that is both efficient, accurate, and cost-effective?

[0024] Firstly, reference Figure 1 This application provides a method for predicting equipment failures based on vibration mechanisms and machine learning, comprising: Step S1: Based on the equipment type, operating conditions and dynamic principles, construct vibration mechanism models corresponding to each component of the equipment; obtain the vibration characteristic spectrum simulated by each vibration mechanism model under normal conditions, various deterioration conditions and various fault conditions, and set corresponding warning values ​​with the vibration characteristic spectrum under each fault condition as a reference to obtain the initial vibration characteristic spectrum library. Step S2: Use the vibration characteristic spectrum of each component of the equipment during the initial learning period (20-30 days after the system is put into use) to calibrate the initial vibration characteristic spectrum library and generate a personalized vibration characteristic spectrum library; then train a machine learning model (AI model) for degradation prediction and fault diagnosis based on the personalized vibration characteristic spectrum library; the personalized vibration characteristic spectrum library can be used as a diagnostic standard, which is the diagnostic capability of this solution in the early stage. Step S3: Obtain the actual vibration characteristic spectrum of each component of the equipment after the initial learning period, and compare and analyze it with the personalized vibration characteristic spectrum library; if the actual vibration characteristic spectrum exceeds the corresponding warning value, the current component triggers an alarm, and in response to the alarm, the machine learning model is called to perform degradation prediction and fault diagnosis for the current component; otherwise, the machine learning model is put into a dormant state.

[0025] Through prior mechanistic modeling, the system can establish preliminary diagnostic capabilities even in the absence of a large amount of historical fault data, solving the problems of long modeling cycles and large data requirements in the initial stage of pure AI methods. Subsequently, through actual data calibration and AI training, the model is made to fit the specific equipment operating conditions, making up for the poor generalization ability and low accuracy of pure mechanistic models. Ultimately, a balance between rapid deployment and high accuracy is achieved. The warning value triggering and hibernation mechanism ensures that the computationally intensive AI diagnostic module only operates when an alarm is triggered, and only performs lightweight threshold monitoring under normal conditions. This directly reduces the system's continuous data storage load, computing resource consumption, and energy consumption, making the system easier to deploy and operate long-term in resource-constrained industrial sites.

[0026] Specifically, this can be understood as: First, based on the initial vibration characteristic spectrum library for modeling, the system can quickly establish a diagnostic baseline by relying on physical principles and prior knowledge, even when a large amount of historical fault data is lacking. This directly solves the pain points of data scarcity and long modeling cycles in the early stages of pure AI methods, enabling rapid system deployment.

[0027] Secondly, by using actual operational data to calibrate and train the machine learning model, and incorporating information on individual equipment differences and real-world operating conditions, the theoretical model was fine-tuned. This step significantly improved the model's relevance and prediction accuracy, overcoming the problems of poor generalization and high false alarm rates caused by the neglect of variables such as actual installation and load in pure mechanistic models. This fostered a positive interaction where theory guides practice, and practice corrects theory.

[0028] Finally, the linkage mechanism of triggering AI diagnosis with a threshold alarm and otherwise going into sleep mode is the core of this method's resource optimization. Under normal circumstances, only lightweight threshold comparisons are performed, avoiding the enormous data storage pressure and computing power consumption caused by the 24 / 7 uninterrupted complex spectrum analysis and AI inference of traditional solutions. The high-energy-consuming AI diagnostic module is only awakened when abnormal symptoms appear. This on-demand computing mode significantly reduces the overall operating cost and hardware threshold of the system, making this predictive maintenance solution highly feasible and economical in resource-constrained industrial environments.

[0029] In summary, the above is not a simple superposition of mechanism models and machine learning models, but rather a temporal and conditional combination of the advantages of both, effectively avoiding their respective disadvantages.

[0030] In some embodiments, a frequency-controlled temperature and vibration integrated sensor deployed at a preset key part of the device is used to collect vibration time-domain waveform data and convert it into vibration characteristic spectrum; the frequency response range of the frequency-controlled temperature and vibration integrated sensor covers 0.4kHz to 12kHz; The vibration characteristic spectrum includes high-frequency, mid-frequency, and low-frequency bands. The high-frequency band is used to monitor early-stage equipment failures, the mid-frequency band is used to monitor mid-stage equipment failure characteristic frequencies, and the low-frequency band is used to monitor severe equipment failures and structural vibration characteristic frequencies. The high-frequency band ranges from 5 kHz to 12 kHz, the mid-frequency band from 2 kHz to 5 kHz, and the low-frequency band from 0.4 kHz to 2 kHz. The specific division of these ranges can be set according to different needs.

[0031] In this embodiment, a frequency response sensor covering 0.4kHz to 12kHz is used as the hardware foundation for achieving early fault warning. Many early equipment faults (such as initial spalling of bearings or microscopic cavitation of gears) generate high-frequency impact signals with weak energy and rich frequency components. Traditional low-frequency or narrow-band sensors cannot effectively capture these signals. Wide-band coverage ensures that the sensor can collect raw vibration information across the entire frequency range containing these early weak features, providing complete data raw materials for subsequent in-depth analysis. Based on this, the spectrum is divided into three frequency bands and their monitoring purposes are clearly defined, providing a software analysis strategy. It structures the complex wide-spectrum information: using the high-frequency band (third band) to specifically capture and amplify the weak high-frequency features of early faults to achieve early detection; using the mid-frequency band (second band) to monitor the characteristic frequencies of energy enhancement in the middle stage of fault development for mid-stage diagnosis; and using the low-frequency band (first band) to identify late-stage faults such as loose overall equipment structure and severe imbalance, as well as macroscopic mechanical problems. This frequency-band and target-based monitoring strategy avoids signal confusion or feature overload caused by a one-size-fits-all analysis, thereby significantly improving the accuracy, timeliness, and comprehensiveness of fault identification.

[0032] In some embodiments, step S1 further includes: synchronously collecting temperature data of preset key parts of the device; Step S2 also includes training a machine learning model for degradation prediction and fault diagnosis based on a personalized vibration feature spectrum library and temperature data.

[0033] In this embodiment, although vibration signals are sensitive, they can sometimes be affected by non-fault factors (such as instantaneous load impacts or process parameter adjustments), and false alarms may occur based solely on excessive vibration. Temperature, as a direct reflection of the equipment's thermodynamic state, is usually strongly correlated with fault processes that generate heat, such as friction, wear, and lubrication failure. By forcibly integrating these two types of heterogeneous information during the model training phase, the trained AI model can learn the deep coupling relationship between vibration and temperature and the mapping of fault modes. For example, the model can learn that when the vibration amplitude rises slightly while the temperature remains stable, it may be an acceptable normal fluctuation or external disturbance; while when vibration and temperature show a coordinated upward trend, it is highly likely to be a real deterioration fault such as insufficient bearing lubrication or worsened misalignment. This multi-parameter cross-validation mechanism greatly enhances the credibility of the diagnostic conclusions. At the same time, the rate and absolute level of temperature rise also provide a quantitative basis for judging the urgency and severity of the fault.

[0034] In some embodiments, in step S1, when the warning value is a preset amplitude, the actual amplitude of the actual vibration characteristic spectrum at the preset characteristic frequency is obtained and compared with the preset amplitude for analysis. When the warning value is a preset growth rate, the actual growth rate of the amplitude of the actual vibration characteristic spectrum at the preset characteristic frequency within the preset window time is obtained and compared with the preset growth rate for analysis. When the warning value is the preset growth rate, the actual growth rate of the amplitude of the actual vibration characteristic spectrum at the preset characteristic frequency within the preset window time is obtained and compared with the preset growth rate.

[0035] Several types are given in this embodiment, and the fundamental reason is as follows: Preset amplitude over-limit alarms are designed for sudden or abrupt faults, such as sudden component breakage or severe impacts. These faults instantly cause the amplitude of the vibration characteristic frequency to exceed the safety threshold. This is the most traditional and basic alarm method, capable of quickly responding to emergencies. However, many equipment faults (such as uniform wear or fatigue crack propagation) are slow, gradual processes. Their vibration amplitude may remain within the normal range for a long time, but a clear upward trend is emerging. In this case, relying solely on amplitude alarms will be severely lagging. Trend increase alarms and trend acceleration alarms are designed to address this issue. Trend increase alarms can identify states where the absolute value is not high but continues to deteriorate. Trend acceleration alarms go a step further, focusing on the acceleration of amplitude changes (e.g., the amplitude increases at an increasingly faster rate), providing earlier warnings of risk points where the degradation process is accelerating and instability may be imminent. Acceleration alarms provide the earliest, forward-looking risk alerts; increase alarms confirm the trend of continuous degradation; and amplitude alarms issue a final warning when degradation becomes clear. This combined strategy greatly extends the early warning window, enabling maintenance personnel to be notified at the nascent stage and early development stage of a fault, thus allowing sufficient time for planned maintenance.

[0036] In some embodiments, step S2 involves calibrating the initial vibration characteristic spectrum library using the vibration characteristic spectrum of each component of the device during its fault-free operation in the initial learning period. This specifically includes the following steps: Obtain the spectrum of the first target vibration feature that is persistent and has a stable amplitude from the vibration data of the current component under fault-free operation; Extract the amplitude at the preset normal characteristic frequency from the vibration characteristic spectrum of the first target, and use it as the current normal amplitude; Obtain the vibration characteristic spectrum of the current component in its normal state from the initial vibration characteristic spectrum library; extract the amplitude at the preset normal characteristic frequency from the vibration characteristic spectrum of the normal state as the reference normal amplitude; Calculate the amplitude calibration coefficient based on the current normal amplitude and the reference normal amplitude; use the amplitude calibration coefficient to calibrate the vibration characteristic spectrum of the current component in the initial vibration characteristic spectrum library for normal state, various deterioration states and various fault states; then, synchronously calibrate the warning values ​​for each fault state. Repeat the above steps to complete the calibration of the initial vibration characteristic spectrum library.

[0037] In this embodiment, obtaining the characteristic spectrum of the device during its initial learning period without failure is a prerequisite for calibration. Extracting the current normal amplitude and the reference normal amplitude, calculating the calibration coefficient, and synchronously calibrating the entire spectrum library and warning values ​​are key algorithms for achieving precise personalization. The advantage lies in its discovery and utilization of a core assumption: vibration response changes caused by individual differences (such as varying installation stiffness) may have a systematic and proportional impact on different frequency components and different fault states. Therefore, it calculates the ratio of the actual to the theoretical amplitude at key characteristic frequencies (i.e., the calibration coefficient) and applies this coefficient to the warning values ​​for characteristic spectra (normal, various degradations, various faults) in the component's theoretical library. This process is not a simple threshold shifting but a systematic scaling of the entire state judgment scale. For example, a device with excessive overall vibration due to a soft installation foundation will have a normal spectrum amplitude higher than the theoretical value; the corresponding fault spectrum threshold should also be adjusted proportionally, otherwise, false alarms will occur continuously. This method aligns all the system's judgment benchmarks with the actual situation of the device through a one-time calibration, ensuring that the fairness and accuracy of the monitoring are established from day one. This avoids the painful process of adjusting thresholds through trial and error that traditional methods require over a long period of time, and greatly improves the system's practicality and intelligence level in the early stages.

[0038] In some embodiments, after responding to an alarm and calling a machine learning model to predict device faults in step S3, the following steps are also included: Obtain the vibration characteristic spectrum of the second target after the current component alarms; If the output of the machine learning model is a deteriorated state, and the amplitude of the second target vibration characteristic spectrum at the fault characteristic frequency is stable and greater than the original warning value, then the warning value is raised. If the output of the machine learning model indicates a fault state, the warning value remains unchanged.

[0039] This embodiment embodies an intelligent decision-making logic based on diagnostic results. It doesn't simply follow data fluctuations but integrates AI's deep understanding of the device's state. The logical chain is as follows: when a threshold is triggered for the first time, the system activates AI for deep diagnosis; if the AI ​​determines that the device has indeed experienced some irreversible degradation (such as slight wear on bearings or pitting on gear teeth), rather than a temporary disturbance, and monitoring data confirms that this degradation has stabilized the device's vibration level at a new, higher platform (new steady state), then the system understands that the device's current health standard has changed. At this point, continuing to use the old threshold set based on the completely new device state is unreasonable, as it would cause the system to continuously alarm the device, which is already operating with defects but is in a stable and controllable state. By raising the warning value to a level that matches the new steady state, the system effectively acknowledges and adapts to this state transition of the device. This simulates the decision-making of a human expert: for an old device known to have wear but operating smoothly, the vibration standard can be appropriately relaxed. This mechanism greatly reduces invalid alarms, ensuring that alarm information truly focuses on new abnormal changes rather than known state degradation, thus guaranteeing the seriousness and guiding significance of alarms and enabling maintenance personnel to concentrate on risks that truly require attention.

[0040] Furthermore, after raising the warning value and performing maintenance on the equipment, the third target vibration characteristic spectrum of the current component is obtained; If the amplitude of the vibration characteristic spectrum of the third target at the fault characteristic frequency is stable and less than the upward-adjusted warning value, then the upward-adjusted warning value will be lowered based on the amplitude of the vibration characteristic spectrum of the third target at the fault characteristic frequency at this time. The specific amount of the downward adjustment can be referred to the steps for the specific amount of upward adjustment below.

[0041] This embodiment addresses the issue of missed alarms caused by the failure of warning values ​​to decrease after equipment repair. For example, consider a piece of equipment with worn bearings, where the threshold has been raised from A to B. After replacing the bearing, the vibration level should ideally return to near its initial state (below A). If the threshold remains at B, the system will lose its ability to detect early faults in this newly repaired equipment, as any new minor degradation requires reaching a high level (B) before triggering an alarm, thus missing the optimal repair opportunity. By monitoring stable operating data after repair, the system automatically determines that the equipment's condition has recovered and executes a threshold callback. This achieves dynamic binding and real-time synchronization between the threshold and the equipment's actual health status. This closed loop of raising and lowering the threshold enables the entire warning system to possess state-following intelligence. It is no longer a rigid, one-time set of rules, but a system capable of understanding the changes in the equipment's lifecycle state and adjusting its judgment criteria accordingly. This maintains the effectiveness and sensitivity of the predictive maintenance system over the long term, representing a high level of practicality and intelligence.

[0042] Furthermore, two options for raising the warning value are provided: The first method, raising the warning value, specifically includes the following steps: Calculate the amplitude of the vibration characteristic spectrum of the second target at the fault characteristic frequency and the amplitude deviation ratio between it and the warning value before adjustment; determine the adjustment coefficient corresponding to the amplitude deviation ratio according to the preset mapping relationship; use the adjustment coefficient to calculate the warning value before adjustment to obtain the warning value after adjustment.

[0043] The second method, raising the warning value, specifically includes the following steps: Obtain the confidence level when the machine learning model outputs a deteriorated state; determine the upward adjustment range of the warning value based on the comparison between the confidence level and the preset confidence threshold; where the higher the confidence level, the greater the corresponding upward adjustment range.

[0044] Both of the above-described adjustment calculation paths embody the concept of data-driven precision control. The first path, determining the adjustment coefficient based on the amplitude deviation ratio, is a direct feedback control. It establishes a simple mapping relationship: the larger the proportion by which the new steady-state amplitude exceeds the old threshold, the more significant the state transition, and therefore the larger the threshold adjustment. This method is intuitive, computationally simple, and can quickly respond to state changes. The second path, determining the adjustment magnitude based on the confidence level of the AI ​​diagnostic results, represents a higher-order intelligent decision-making approach. When the AI ​​model outputs a deteriorated state, it typically includes a confidence score reflecting its degree of certainty in its judgment. Higher confidence indicates that the AI ​​is more certain that the device has experienced irreversible degradation, rather than an ambiguous edge case, thus allowing for a more aggressive threshold adjustment. Lower confidence may necessitate a more conservative, smaller adjustment, or even temporarily maintaining the original threshold for continued observation. This method bases the authority of the threshold adjustment on the reliability of the AI ​​diagnosis, making the entire decision-making chain highly logically consistent.

[0045] Both methods make threshold adaptation no longer a simple matter of raising it slightly, but a refined operation that is based on evidence and measurable, thus enhancing the reliability of the system.

[0046] In some embodiments, step S3 involves calling a machine learning model to predict device degradation and diagnose faults, specifically including: In response to an alarm, the machine learning model is activated. The vibration characteristic spectrum of the component that triggered the alarm after the alarm was collected is used as the vibration characteristic spectrum of the second target. The vibration characteristic spectrum of the second target is input into the machine learning model to output diagnostic results about the fault location, fault type and degree of deterioration; Based on the diagnostic results, determine and update the warning values ​​corresponding to the vibration characteristic spectrum of the second target; The vibration characteristic spectrum of the second target, its corresponding diagnostic results, and the updated warning values ​​are supplemented and updated into the personalized vibration characteristic spectrum library.

[0047] The vibration spectrum, diagnostic results, and updated warning values ​​for this event are then added to the personalized knowledge base. This step is the core mechanism for the system's autonomous learning. It means that each real-world fault (or degradation) case, including its unique vibration spectrum characteristics, the AI's final diagnostic conclusion, and the alarm standards adjusted through practical testing, is archived as a complete case in the device's dedicated knowledge base. This continuously expanding case database has multiple values: First, it can supplement subsequent training data, directly used for retraining the AI ​​model and improving the model's accuracy in identifying similar faults; second, it can serve as a reference for judging similar faults, enhancing the system's ability to recognize patterns in rare or complex faults; third, it essentially represents the digital accumulation of expert experience.

[0048] Furthermore, the diagnostic results of this application regarding the location, type, and degree of deterioration of the fault can guide maintenance personnel in developing equipment maintenance and repair plans. The diagnostic results regarding the location, type, and degree of deterioration of the fault require manual confirmation and feedback, and the diagnostic results are marked. After the fault is confirmed, the manual personnel feed back the equipment inspection results to the system to confirm whether the prediction results are correct. The system learns and records, accumulates experience, and uses it as a basis for the next judgment, optimizes the weight score of such fault judgment decisions, and adjusts the model at the same time.

[0049] The training and analysis of the machine learning model in this application are existing technologies, and therefore no specific limitations are specified.

[0050] The beneficial effects of this application are: (1) Traditional pure AI methods require a long data accumulation period, while pure mechanistic models suffer from insufficient accuracy and poor generalization. This invention effectively solves this contradiction by integrating mechanistic modeling first, actual data calibration, and AI model training. The mechanistic model provides an initial diagnostic capability that can be quickly deployed, significantly shortening the system deployment cycle; subsequent calibration and training based on individual device data enable continuous optimization of model accuracy. This method reduces reliance on massive amounts of historical fault data, reduces the complexity and time cost of model debugging, and enables predictive maintenance systems to be quickly put into use and generate value immediately.

[0051] (2) This invention achieves full-cycle coverage of equipment from early deterioration to severe failure through wideband sensors, spectrum partitioning strategies, and multi-dimensional alarm logic. This brings multiple benefits: First, it enables early minor faults to be detected in a timely manner (with warnings several months in advance), avoiding sudden downtime and major losses caused by fault escalation, and reducing unplanned maintenance. Second, accurate fault prediction allows maintenance departments to formulate precise spare parts plans and window-of-care maintenance schemes, upgrading from preventive maintenance to predictive maintenance, avoiding over-maintenance and spare parts inventory backlog, and directly reducing spare parts and labor costs. Third, through early corrective action, it slows down the wear rate of core equipment components from the root, thereby effectively extending the overall service life of the equipment.

[0052] (3) One of the core innovations of this invention is the linkage mechanism that triggers AI diagnosis with a threshold alarm, otherwise it goes into sleep mode. When the device is running normally, the system only performs real-time threshold comparison with low computational complexity and does not need to continuously store high-frequency raw waveform data. Only when the threshold is triggered will the computationally intensive AI diagnosis module be awakened to perform in-depth analysis. Compared with the traditional solution of uninterrupted full data storage and AI inference, this design greatly reduces the total amount of data storage, reduces network transmission pressure, and significantly saves computing resources. This enables this solution to run efficiently on edge computing devices or resource-constrained industrial servers, reducing the overall hardware cost of implementation and maintenance.

[0053] (4) Possesses self-learning and self-adaptive capabilities, significantly improving the system's intelligence level and reducing manual workload; the system of this invention is not a static tool, but a self-optimizing intelligent agent. The calibration algorithm enables the system to automatically adapt to the individual differences of each device, eliminating initial false alarms. Closed-loop adaptive threshold management can dynamically adjust alarm standards according to the device status, fundamentally solving the persistent problem of repeated alarms or missed alarms caused by normal aging of equipment or changes in operating conditions, greatly reducing the workload of equipment managers in frequently handling invalid alarms and manually adjusting thresholds. Each diagnostic result is used to update the personalized knowledge base, realizing the digital accumulation and inheritance of experience, enabling the system's diagnostic capabilities to continuously evolve autonomously, constantly reducing dependence on senior experts, and forming accumulative intelligent assets.

[0054] Secondly, embodiments of this application also provide a device fault prediction system based on vibration mechanisms and machine learning, comprising: The first module is used to construct vibration mechanism models corresponding to each component of the equipment based on the equipment type, operating conditions and dynamic principles; obtain the vibration characteristic spectrum of each vibration mechanism model under normal conditions, various deterioration conditions and various fault conditions, and set corresponding warning values ​​with the vibration characteristic spectrum under each fault condition as a reference to obtain the initial vibration characteristic spectrum library. The second module is used to calibrate the initial vibration feature spectrum library by utilizing the vibration feature spectrum of each component of the equipment during the initial learning period when it operates without faults, and generate a personalized vibration feature spectrum library; then, a machine learning model for degradation prediction and fault diagnosis is trained based on the personalized vibration feature spectrum library. The third module is used to obtain the actual vibration characteristic spectrum of each component of the equipment after the initial learning period, and compare and analyze it with the personalized vibration characteristic spectrum library. If the actual vibration characteristic spectrum exceeds the corresponding warning value, the current component triggers an alarm. In response to the alarm, the machine learning model is called to perform degradation prediction and fault diagnosis for the current component. Otherwise, the machine learning model is put into a dormant state.

[0055] Thirdly, embodiments of this application provide a device fault prediction device based on vibration mechanism and machine learning. The device fault prediction device based on vibration mechanism and machine learning can be a device with data processing capabilities, such as a personal computer (PC), a laptop, or a server.

[0056] In this embodiment of the application, the device fault prediction device based on vibration mechanism and machine learning may include a processor, a memory, a communication interface, and a communication bus.

[0057] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0058] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces. These interfaces are used for interconnecting internal components of the vibration mechanism and machine learning-based equipment fault prediction device, as well as for interconnecting the device with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0059] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0060] The processor can be a general-purpose processor, which can call a device fault prediction program based on vibration mechanism and machine learning stored in memory and execute the device fault prediction method based on vibration mechanism and machine learning provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the device fault prediction program based on vibration mechanism and machine learning is called can be referred to the various embodiments of the device fault prediction method based on vibration mechanism and machine learning in this application, and will not be repeated here.

[0061] Fourthly, embodiments of this application also provide a computer-readable storage medium.

[0062] This application provides a computer-readable storage medium storing a device failure prediction program based on vibration mechanism and machine learning, wherein when the device failure prediction program based on vibration mechanism and machine learning is executed by a processor, it implements the steps of the device failure prediction method based on vibration mechanism and machine learning as described above.

[0063] The method implemented when the equipment failure prediction program based on vibration mechanism and machine learning is executed can be referred to in various embodiments of the equipment failure prediction method based on vibration mechanism and machine learning in this application, and will not be repeated here.

[0064] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0065] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0066] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0067] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0068] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0069] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0070] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for predicting equipment failures based on vibration mechanisms and machine learning, characterized in that, It includes: Based on the equipment type, operating conditions, and dynamic principles, construct vibration mechanism models for each component of the equipment; Obtain the vibration characteristic spectrum of each vibration mechanism model under normal, various deterioration and fault conditions, and set the corresponding warning value with the vibration characteristic spectrum under each fault condition as a reference to obtain the initial vibration characteristic spectrum library. The initial vibration characteristic spectrum library is calibrated using the vibration characteristic spectrum of each component of the equipment during the initial learning period when it operates without faults, thereby generating a personalized vibration characteristic spectrum library. Then, a machine learning model for degradation prediction and fault diagnosis is trained based on the personalized vibration feature spectrum library. The actual vibration characteristic spectrum of each component of the device after the initial learning period is obtained and compared with the personalized vibration characteristic spectrum library. If the actual vibration characteristic spectrum exceeds the corresponding warning value, the current component triggers an alarm. In response to the alarm, the machine learning model is called to perform degradation prediction and fault diagnosis for the current component. Otherwise, the machine learning model is put into a dormant state.

2. The equipment fault prediction method based on vibration mechanism and machine learning as described in claim 1, characterized in that: The integrated frequency-pass temperature and vibration sensor is deployed at key locations in the device to collect vibration time-domain waveform data and convert it into a vibration characteristic spectrum; the frequency response range of the integrated frequency-pass temperature and vibration sensor covers 0.4kHz to 12kHz. The vibration characteristic spectrum includes a high-frequency spectrum, a mid-frequency spectrum, and a low-frequency spectrum; the high-frequency spectrum is used to monitor early-stage equipment failures, the mid-frequency spectrum is used to monitor mid-stage equipment failure characteristic frequencies, and the low-frequency spectrum is used to monitor severe equipment failures and structural vibration characteristic frequencies.

3. The equipment fault prediction method based on vibration mechanism and machine learning as described in claim 2, characterized in that, Also includes: Synchronously collect temperature data from preset key components of the equipment; A machine learning model for degradation prediction and fault diagnosis is obtained by training a personalized vibration feature spectrum library and the temperature data.

4. The equipment failure prediction method based on vibration mechanism and machine learning as described in claim 1, characterized in that: When the warning value is a preset amplitude, the actual amplitude of the actual vibration characteristic spectrum at the preset characteristic frequency is obtained and compared with the preset amplitude for analysis. When the warning value is a preset growth rate, the actual growth rate of the amplitude of the actual vibration characteristic spectrum at the preset characteristic frequency within a preset window time is obtained and compared with the preset growth rate for analysis. When the warning value is a preset growth rate, the actual growth rate of the amplitude of the actual vibration characteristic spectrum at the preset characteristic frequency within a preset window time is obtained and compared with the preset growth rate.

5. The equipment fault prediction method based on vibration mechanism and machine learning as described in claim 1, characterized in that, The initial vibration characteristic spectrum library is calibrated using the vibration characteristic spectrum of each component of the equipment during the initial learning period when it operates without faults. This process includes the following steps: Obtain the spectrum of the first target vibration feature that is persistent and has a stable amplitude from the vibration data of the current component under fault-free operation; Extract the amplitude at a preset normal characteristic frequency from the vibration characteristic spectrum of the first target, and use it as the current normal amplitude; Obtain the vibration characteristic spectrum of the current component in its normal state from the initial vibration characteristic spectrum library; extract the amplitude at a preset normal characteristic frequency from the vibration characteristic spectrum in the normal state as a reference normal amplitude; Based on the current normal amplitude and the reference normal amplitude, an amplitude calibration coefficient is calculated; using the amplitude calibration coefficient, the vibration characteristic spectrum of the current component in the normal state, each deterioration state, and each fault state in the initial vibration characteristic spectrum library is calibrated; then, the warning values ​​in each fault state are synchronously calibrated. Repeat the above steps to complete the calibration of the initial vibration characteristic spectrum library.

6. The equipment fault prediction method based on vibration mechanism and machine learning as described in claim 1, characterized in that, After responding to the alarm and invoking the machine learning model to predict device malfunctions, the process further includes the following steps: Obtain the vibration characteristic spectrum of the second target after the current component alarms; If the output of the machine learning model is a deteriorated state, and the amplitude of the second target vibration characteristic spectrum at the fault characteristic frequency is stable and greater than the original warning value, then the warning value is raised. If the output of the machine learning model indicates a fault state, the warning value remains unchanged.

7. The equipment fault prediction method based on vibration mechanism and machine learning as described in claim 6, characterized in that: After raising the warning value and performing equipment maintenance, the third target vibration characteristic spectrum of the current component is obtained; If the amplitude of the vibration characteristic spectrum of the third target at the fault characteristic frequency is stable and less than the upward-adjusted warning value, then the upward-adjusted warning value will be lowered based on the amplitude of the vibration characteristic spectrum of the third target at the fault characteristic frequency at this time.

8. The equipment fault prediction method based on vibration mechanism and machine learning as described in claim 6, characterized in that: Raising the warning threshold specifically includes the following steps: Calculate the amplitude of the second target vibration characteristic spectrum at the fault characteristic frequency, and the amplitude deviation ratio between it and the warning value before adjustment; determine the adjustment coefficient corresponding to the amplitude deviation ratio according to a preset mapping relationship; use the adjustment coefficient to calculate the warning value before adjustment to obtain the warning value after adjustment; or, Raising the warning threshold specifically includes the following steps: Obtain the confidence level when the output of the machine learning model is in a deteriorated state; determine the upward adjustment range of the warning value based on the comparison result between the confidence level and the preset confidence threshold; wherein, the higher the confidence level, the greater the corresponding upward adjustment range.

9. The equipment fault prediction method based on vibration mechanism and machine learning as described in claim 1, characterized in that, The machine learning model is invoked to predict device degradation and diagnose faults, specifically including: In response to the alarm, the machine learning model is activated; The vibration characteristic spectrum of the component that triggered the alarm after the alarm was collected is used as the vibration characteristic spectrum of the second target. The vibration characteristic spectrum of the second target is input into the machine learning model to output diagnostic results about the fault location, fault type and degree of deterioration; Based on the diagnostic results, determine and update the warning value corresponding to the vibration characteristic spectrum of the second target; The second target vibration characteristic spectrum, its corresponding diagnostic results, and the updated warning value are added to the personalized vibration characteristic spectrum library.

10. A device fault prediction system based on vibration mechanism and machine learning, characterized in that, It includes: The first module is used to construct vibration mechanism models for each component of the equipment based on the equipment type, operating conditions, and dynamic principles. Obtain the vibration characteristic spectrum of each vibration mechanism model under normal, various deterioration and fault conditions, and set the corresponding warning value with the vibration characteristic spectrum under each fault condition as a reference to obtain the initial vibration characteristic spectrum library. The second module is used to calibrate the initial vibration feature spectrum library by using the vibration feature spectrum of each component of the equipment during the initial learning period when it operates without faults, and generate a personalized vibration feature spectrum library; then, a machine learning model for degradation prediction and fault diagnosis is trained based on the personalized vibration feature spectrum library. The third module is used to obtain the actual vibration characteristic spectrum of each component of the equipment after the initial learning period, and compare and analyze it with the personalized vibration characteristic spectrum library; if the actual vibration characteristic spectrum exceeds the corresponding warning value, the current component triggers an alarm, and in response to the alarm, the machine learning model is called to perform degradation prediction and fault diagnosis for the current component; otherwise, the machine learning model is put into a dormant state.

Images