A Method and System for Efficient Storage of Vibration Data Based on Machine Learning-Based Sensitive Monitoring

Abstract

This invention relates to the field of machine learning and equipment health monitoring technology, and discloses an efficient method and system for storing vibration data based on machine learning-based sensitive monitoring. The method includes: setting up a parameter-locked anchor point model and an online learning buoy model in parallel; simultaneously inputting real-time vibration data into both models, calculating the benchmark drift based on the difference in their errors, and simultaneously calculating the instability index based on the dispersion of the anchor point model error sequence; triggering data storage when any index exceeds a preset threshold. This invention solves the technical problem that monitoring models cannot detect long-term slow degradation and early instability precursors of equipment due to continuous adaptation by constructing a differential reference system between a static health benchmark and a dynamic current benchmark, combined with the analysis of error sequence stability. It can distinguish between real fault trends and normal operating condition fluctuations, and avoids contamination of the monitoring benchmark.

Description

Technical Field

[0001] This invention relates to a method and system for efficient storage of vibration data based on machine learning-based sensitive monitoring, belonging to the field of machine learning and equipment health monitoring technology. Background Technology

[0002] In the health status monitoring of large rotating machinery, such as steam turbine generator sets, deploying machine learning-based online monitoring models has become a common technique. These models, such as autoencoder neural networks or long short-term memory networks, continuously learn from vibration data during normal equipment operation to establish a dynamic baseline of normal behavior. They then use deviations between newly acquired data and this baseline, such as reconstruction or prediction errors, to determine the occurrence of abnormal events. This approach is recognized for its effectiveness in capturing sudden shocks or failures. However, when this continuously adaptive monitoring method is applied to scenarios requiring long-term, high-reliability operation, an inherent property of its design principle—that the model adapts to the current state of the equipment—is challenged. Continuous online learning, however, exhibits an inherent limitation at the principle level. For many highly hazardous failure modes, such as early fatigue pitting of bearings or aging of insulation materials, the initial stage does not manifest as a violent vibration change, but rather as an extremely slow performance degradation. This degradation is reflected in the vibration signal as a continuous, small drift in statistical characteristics. When the monitoring model encounters new data carrying this small drift, it will not trigger an alarm because the difference from the known normal is too small. More importantly, the model's online learning mechanism will learn and absorb these small drift data as new normal, causing the model's own definition of normal to drift synchronously with the slow decline in equipment health.

[0003] To avoid this synchronous drift of the monitoring benchmark, one direct approach is to significantly reduce the online learning rate of the model. However, this would prevent it from adapting to benign vibration pattern changes caused by normal operating conditions such as load and speed variations, leading to numerous false alarms and rendering it impractical for engineering applications. Another approach is to use a fixed, offline-trained health model. However, this, too, cannot distinguish between normal operating condition fluctuations and early fault symptoms, making it overly sensitive to complex and ever-changing real-world operating conditions. Furthermore, existing technologies, when improving monitoring accuracy, often focus on optimizing the physical environment of sensor input, neglecting the deep-seated cognitive deficiencies in the monitoring model's algorithmic logic. For example, [authorization notice number] Chinese invention patent CN215261882U discloses a high-precision air monitor with high detection sensitivity. Its technical solution is to use a cooling fan to blow away dust from the sensor probe and to balance the internal and external air pressure through the air intake pipe and filter window, thereby maintaining the accuracy of the sensor's original data. This approach is essentially a direct intervention in the physical front end of the monitoring. Although it can ensure the fidelity of the input signal to a certain extent, it is completely powerless to deal with data information drift caused by the slow degradation of the equipment itself, or cognitive drift caused by the continuous adaptive learning of the monitoring model. It solves the problem of how the signal is accurately collected, but avoids the core problem of how the signal should be interpreted stably and over a long period of time.

[0004] Existing monitoring methods present a trade-off between adaptability to the current state and retention of the initial health state, resulting in two main problems: 1. For slow, gradual system degradation, monitoring models, due to their continuous adaptive learning, have fundamental cognitive limitations and cannot perceive the long-term drift of their knowledge base relative to an absolute health benchmark. 2. Existing monitoring methods primarily focus on instantaneous anomalies in vibration signals, lacking an effective means to directly extract and quantify the long-term evolution trend of equipment health status from the data, leading to the storage of a large amount of redundant data that does not contain information on trend changes in status. Therefore, the technical problem to be solved by this invention is how to establish a method that can stably anchor the absolute health state of equipment while dynamically adapting to changes in its normal operating conditions, thereby sensitively identifying characteristics representing long-term systematic drift in equipment health status from continuous vibration data streams, and efficiently storing this data. Summary of the Invention

[0005] This invention provides a method and system for efficient storage of vibration data based on machine learning-based sensitive monitoring. Its main purpose is to solve the problem that existing machine learning monitoring methods cannot effectively detect the long-term slow deterioration of equipment health status due to the synchronous drift of the monitoring benchmark caused by continuous online learning.

[0006] To achieve the above objectives, this invention provides an efficient method for storing vibration data based on machine learning-based sensitive monitoring, comprising the following steps: The current vibration data is simultaneously input into the anchor point model and the buoy model to obtain the first error value and the second error value, respectively. The reference drift is calculated based on the difference between the first error value and the second error value. For multiple first error values ​​generated consecutively within a preset time window, calculate the instability index that characterizes the degree of discreteness of their sequence; When the reference drift exceeds a preset first threshold, or when the instability index exceeds a preset second threshold, the storage of the current vibration data is triggered. The buoy model continuously adapts its internal parameters to the real-time vibration data from the monitored device through an online learning mechanism. The online learning mechanism has the inherent characteristic that the knowledge base of the buoy model drifts synchronously with the slow deterioration of the device. The internal parameters of the anchor point model are trained offline using only historical health vibration data from the device under monitoring and are locked to form a static health benchmark for measuring the drift of the buoy model's knowledge benchmark. Furthermore, the instability index is the statistical variance of the first error value sequence. The calculation method is as follows:

[0007] in, The length of the preset time window, For the first time window The first error value, It is the arithmetic mean of all first error values ​​within the time window.

[0008] Furthermore, it also includes: when the storage of the current vibration data is triggered, a drift residual signal is generated based on the reconstruction results of the current vibration data using the anchor point model and the buoy model, and a drift fingerprint vector of a predetermined dimension is extracted from the drift residual signal using a preset feature extraction module, and the current vibration data and the drift fingerprint vector are jointly stored. The drift residual signal is obtained by calculating the difference between the reconstruction results of the anchor point model and the reconstruction results of the buoy model.

[0009] Furthermore, both the anchor model and the buoy model are autoencoder neural network models, and the first error value and the second error value are the reconstruction errors of the corresponding autoencoder neural network models.

[0010] Furthermore, the steps of obtaining the first error value and the second error value also include: measuring the first inference delay required for the anchor model to generate the first error value and the second inference delay required for the buoy model to generate the second error value, and calculating the drift complexity index defined by the difference between the first inference delay and the second inference delay; when storing the current vibration data, it also includes: jointly storing the current vibration data, the reference drift degree, and the drift complexity index as data records.

[0011] Furthermore, when storing the current vibration data, the method also includes: if the drift complexity index exceeds a preset complexity threshold when the storage of the current vibration data is triggered, the storage event is marked as a complex drift event.

[0012] Furthermore, when an instruction is received indicating that the device to be monitored has entered a new benign operating state, vibration data samples characterizing the new benign operating state are acquired, and the online fine-tuning process of the buoy model is initiated; in the online fine-tuning process, a composite loss function is used to update the internal parameters of the buoy model.

[0013] Furthermore, the composite loss function includes a reconstruction loss term used to drive the buoy model to learn the vibration data sample patterns, and a knowledge distillation loss term used to constrain the buoy model's understanding of historical healthy vibration data from deviating from the anchor model. The knowledge distillation loss term achieves the constraint by calculating the difference in the output of the internal activation layers of the buoy model and the anchor model when processing the same historical healthy vibration data.

[0014] Furthermore, the storage of current vibration data also includes: when the second error value exceeds the preset third threshold, the storage of current vibration data is also triggered, and the triggering cause is marked as a sudden impact; after the monitored equipment undergoes major maintenance or a fundamental change in its condition, in response to external instructions, new healthy vibration data is collected, a new anchor point model is retrained in the cloud using the new healthy vibration data, and then sent to the monitoring equipment to update the static health benchmark.

[0015] This invention also provides a high-efficiency vibration data storage system based on machine learning-based sensitive monitoring, comprising the following steps for implementing the aforementioned high-efficiency vibration data storage method based on machine learning-based sensitive monitoring: The buoy model module is used to continuously adapt to the current vibration data acquired in real time through an online learning mechanism. The online learning mechanism has the inherent characteristic that the knowledge base of the buoy model module drifts synchronously with the slow deterioration of the equipment. The anchor point model module, whose internal parameters are locked, is used to form a static health baseline; The data processing module is used to receive the current vibration data and simultaneously input the current vibration data into the anchor point model module and the buoy model module to obtain the first error value and the second error value respectively. The health status quantification module is used to calculate the baseline drift based on the difference between the first error value and the second error value; and to calculate the instability index, which characterizes the degree of discreteness of the sequence, for multiple first error values ​​generated continuously within a preset time window. The storage decision module is used to generate a storage trigger command when the baseline drift exceeds a preset first threshold or when the instability index exceeds a preset second threshold. The data storage module is used to store the current vibration data in response to the storage trigger command generated by the storage decision module.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention provides an efficient vibration data storage method based on machine learning-based sensitive monitoring. By setting up a parameter-locked anchor point model and a buoy model with online learning capabilities in parallel and continuously calculating the difference in their error values, a long-term stable state reference system is established in the field of mechanical vibration monitoring. Unlike existing monitoring methods that typically allow the model to continuously adapt to the current state of the equipment, the anchor point model in this invention always adheres to the data characteristics of the equipment in its initial healthy state, while the buoy model continuously fits the details of the equipment's current operation. When the equipment experiences slow and gradual performance degradation, the error of the buoy model remains low due to its adaptability, while the error of the anchor point model continues to increase as the equipment state deviates from its initial healthy state. The difference in their errors becomes a quantitative indicator that directly represents the drift of the equipment's health status benchmark. This approach enables the monitoring system to distinguish between vibration fluctuations caused by short-term changes in operating conditions and systemic degradation caused by long-term component wear, avoiding the technical problem of the monitoring benchmark being contaminated by degraded data during continuous learning, thus ignoring early slow faults.

[0017] When determining the need for data storage, the different reconstruction results of the same vibration data generated by the anchor point model and the buoy model are further utilized. By calculating the difference between the two reconstruction results, a residual signal that can characterize the current drift pattern is generated. Because the reconstruction result of the anchor point model tends to filter out its incomprehensible anomalous components and retain the healthy pattern, while the reconstruction result of the buoy model is closer to the original signal containing anomalous components, the residual signal obtained by subtracting the two is, in a physical sense, an effective extraction of the purely systematic anomalous drift component in the original signal. By storing the feature information representing this residual signal, this invention not only records the fact that the equipment state has deviated, but also attaches structured information about the deviation pattern to each record. This allows subsequent fault root cause analysis to be performed directly based on this pattern-discriminative information for clustering or identification, without the need for complex feature mining from massive amounts of original vibration data.

[0018] Furthermore, the method of this invention integrates statistical analysis of the anchor model error value sequence with measurement of the difference in inference delay between the anchor model and the buoy model into a unified monitoring process. It calculates the dispersion of the anchor model error values ​​over a period of time to identify early instability characterized by intermittent impacts before the equipment enters a clear unidirectional drift. Simultaneously, it measures and compares the time required for the two models to process the same data frame, utilizing the increased computational burden of the anchor model when processing unfamiliar complex signals to quantify changes in the intrinsic complexity of the vibration signal. These two pieces of information, together with the aforementioned baseline drift, constitute a multi-dimensional, complementary description of the equipment's health status. This allows a single storage decision to no longer be based solely on the drift amplitude but to simultaneously include a comprehensive assessment of state stability and signal complexity. Ultimately, this forms a context-rich, self-interpretive, high-value dataset, providing direct and effective input for downstream machine learning fault diagnosis models. Attached Figure Description

[0019] Figure 1 This is a diagram of the dual-model collaborative monitoring and high-efficiency data storage system architecture of the present invention; Figure 2 This is a schematic diagram illustrating the system deployment scenario and human-computer interaction management of the present invention; Figure 3 This is a timing diagram of the interaction between data value enhancement and conditional storage in this invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0021] This invention provides a method and system for efficient storage of vibration data based on machine learning-based sensitive monitoring. It can be deployed in edge monitoring equipment for large rotating machinery such as steam turbine generator sets. Logically, the execution entity of this method may include a bimodal model inference module, a dual-channel health state quantization module, and a storage decision module. The bimodal model inference module runs a parameter-locked anchor point model and an online-learning buoy model in parallel. The dual-channel health state quantization module receives the outputs of the two models and calculates the baseline drift degree and instability index respectively. Finally, the storage decision module triggers data storage instructions based on the state of these two indicators, thus forming a complete data monitoring and efficient storage process. In one specific implementation, both the anchor point model and the buoy model are configured as autoencoder neural network models. This model structure consists of an encoder and a decoder connected in series. The encoder compresses the input high-dimensional vibration data frame into a low-dimensional latent space vector, and the decoder reconstructs the original data frame from this latent space vector. The training objective of the model is to minimize the difference between the original input data and the decoder's reconstructed output data; this difference value, the reconstruction error, is used as the basis for subsequent analysis.

[0022] During the system initialization phase, the first step is to establish an anchor point model. This step is performed in an offline environment. Considering that large rotating machinery generally undergoes a mechanical break-in period after its initial commissioning or after a comprehensive overhaul, during which its vibration characteristics are unstable and not in an ideal healthy state, the procedure for data acquisition points used to train the anchor point model should be as follows: First, wait for the unit to complete the necessary break-in process. After maintenance experts confirm, based on equipment performance indicators, that it has entered a long-term, stable operating state where all parameters meet the design standards, data acquisition should be initiated. The acquisition process should cover multiple typical operating conditions of the equipment under this stable and healthy state. For example, different loads or speeds are used, and the total data duration (e.g., 72 hours or more of cumulative operation) is sufficient to capture vibration patterns with sufficient statistical representativeness. These verified and screened historical data that truly represent a stable and healthy state are used as the sole training set to train an initial autoencoder network until its reconstruction error converges to a preset low level. After training, the model's internal parameters, namely the weights and biases of its encoder and decoder, are completely locked and solidified, and then deployed as a static module to the edge monitoring device. This is the anchor model, which constitutes the static health benchmark for all subsequent judgments. In parallel with the anchor model, the system also provides a buoy model. During initial deployment, the buoy model can directly replicate the already trained anchor model to ensure both have the same initial knowledge baseline. After deployment to edge devices, the buoy model is configured to allow continuous online learning. To ensure high sensitivity to the long-term, slow degradation trend of the device, the buoy model adopts a preset baseline low learning rate (e.g., during normal stable operation of the device). ) Perform continuous and gradual parameter fine-tuning; Considering the varying rates of change in vibration characteristics of equipment under different operating states (such as start-up, shutdown, and load adjustment), this method also includes an adaptive adjustment mechanism to address the potential issues of overly rapid adaptation to normal operating conditions or overly slow adaptation to transitions to benign operating conditions that may result from a single fixed learning rate. When responding to maintenance personnel's instructions and confirming that the equipment has entered a new benign operating state, the system can initiate a brief online rapid fine-tuning process. During this process, a specific composite loss function is used to accelerate the buoy model's adaptation to the new operating condition data pattern. Simultaneously, the anchor point model is used to constrain its knowledge baseline, preventing it from deviating from its understanding of the initial healthy state. Through this combination of steady-state slow adaptation and transient fast adjustment strategy, the buoy model's knowledge baseline can effectively capture long-term slow degradation while also ensuring rapid adaptation to changes in normal and benign operating conditions, thereby avoiding false alarms.

[0023] During the continuous operation of the system, the monitoring process is executed cyclically according to the following steps: First, the data acquisition unit acquires a current vibration data frame at fixed time intervals (e.g., 100 milliseconds). The data frame is a one-dimensional time series vector containing, for example, 1024 sampling points; subsequently, the data frame The data is simultaneously input into both the deployed anchor model and the buoy model; after processing, the anchor model outputs reconstructed data. And calculate the first error value. Its calculation method is as follows and The mean square error between them; similarly, the buoy model outputs a reconstructed data. And calculate the second error value. Because the parameters of the anchor model are locked, it cannot fit any signal changes that deviate from the initial healthy pattern; therefore, the first error value... It objectively reflects the deviation of the current equipment status from the static health baseline; while the buoy model, due to its continuous adaptability, has a second error value. This reflects the deviation of the current state from the recent normal state; after obtaining these two error values, the dual-channel health status quantification process is started in parallel; In the first channel, to capture long-term, gradual degradation of the device's health status, the system performs a baseline drift calculation, which is achieved through a subtraction operation, namely the baseline drift. In a specific numerical example, assuming that in the early stages of device health, for an input data frame, It is 0.012. If the value is 0.011, then the calculated value is... The value is 0.001, which is close to 0. After a period of operation, a certain component of the equipment experiences slow wear. For a data frame under similar operating conditions, the anchor point model cannot fit this degradation pattern. It rose to 0.085, while the buoy model, having adapted to this degradation, The value is only 0.015, at which point the calculated value is... The value was 0.070, showing a significant increase.

[0024] In the second channel, to detect the early unstable state characterized by intermittent shocks that may occur before a definite unidirectional drift occurs in the sensing device, the system performs an instability index calculation; this calculation is performed within a preset fixed-length sliding time window, the length of which is... It is not an arbitrarily set empirical value, but a key parameter to ensure the appropriateness of its selection. Before system deployment, the values ​​are optimized using a standardized offline calibration procedure to adapt to the dynamic characteristics of the specific monitoring object. This procedure tests different... The monitoring signal-to-noise ratio at the given value was ultimately selected to maximize the signal-to-noise ratio. The value is used as the optimal configuration for this monitoring scenario. This data-driven optimization method ensures that the choice of window length N has the best monitoring sensitivity for the target fault mode. For example, in a specific calibration implementation, the system maintains a queue in memory to store the most recently generated N=100 consecutive first error values. For this length is For time series data, the system calculates its statistical variance as an instability index, and the calculation method is as follows:

[0025] in, The length of the preset time window, For the first time window The first error value, This is the arithmetic mean of all first error values ​​within the time window; During stable operation of the equipment, continuous The value fluctuates very little, with its variance close to 0; however, when the equipment experiences early pitting and other failures, and generates random, minor impacts, The sequence exhibits abrupt changes; even if its mean does not change significantly, its variance increases substantially, thus capturing the deterioration of dynamic stability. Ultimately, a storage decision rule is established and executed, which integrates three independent trigger channels: First, when the reference drift degree calculated by the first channel Exceeding a preset first threshold (For example, 0.05) indicates that the device has experienced a long-term systematic drift, triggering storage; Second, when the instability index calculated by the second channel Exceeding a preset second threshold (For example, 0.002) indicates that the device has entered an early unstable state, triggering storage; Third, when the second error value of the buoy model It exceeds a preset third threshold used to characterize sudden impact. At the same time, storage is also triggered, and the triggering reason is marked as a sudden impact; The determination of these three thresholds was accomplished through a standardized offline calibration procedure. This involved replay testing using historical datasets containing known fault evolution processes, analyzing receiver operating characteristic curves to find the optimal threshold combination that maximizes the detection rate of various faults while keeping the false alarm rate under normal operating conditions within an acceptable range. In the absence of any storage conditions, the current vibration data... They are then discarded directly without occupying any storage resources.

[0026] To further enhance the diagnostic value of the stored data, this method may also include a deep characterization step of the drift pattern; when storage is triggered, the system not only saves the original data, but also utilizes the reconstruction results generated by the two models. and A drift residual signal is generated by calculating the difference between the two. ;because The tendency is to retain the healthy pattern components in the signal, while It is closer to the original signal containing anomalous components, therefore This can be viewed as the extraction of abnormal drift components from the signal; subsequently, the residual signal... The data is fed into a pre-defined feature extraction module, such as the encoding end of a miniature one-dimensional convolutional network, from which a drift fingerprint vector of a predetermined dimension (e.g., 32-dimensional) is extracted. Finally, the system will display the current vibration data. With the drift fingerprint vector Perform joint storage to provide structured information for subsequent root cause analysis of faults; To quantify the inherent complexity of the detected drift signal, this method can also measure the computational delay of the bimodal model inference process within a parallel processing flow; specifically, when processing the current vibration data frame... When input is simultaneously fed into both the anchor point model and the buoy model, the computation time required for each model to go from receiving the input to outputting the reconstruction result is measured using the high-precision clock of the system processor. This time is denoted as the first inference delay. Second reasoning delay Furthermore, by calculating the difference between the two, a drift complexity index is obtained. The calculation principle of this index is that when the degradation mode of the vibration signal is relatively simple, the two models have similar calculation paths. The value is close to zero, and when new and complex harmonics or modulation components appear in the signal, the anchor model that has not learned this mode needs to expend more computational resources, resulting in its inference delay. Significantly increased, thus making The value increases; correspondingly, the storage decision rule can be further extended to include the current vibration data when data storage is triggered. Reference drift and the drift complexity index Stored as a single data record in a unified manner, and when When the value exceeds a preset complexity threshold, the storage event can be marked as a complex drift event in the data record.

[0027] To stabilize and quantify the drift complexity index, in the specific implementation, within a preset computation cycle, such as the processing cycle of 10 consecutive data frames, to reduce the computational overhead caused by continuous measurement, the system does not record every single data frame. Instead, it randomly samples or selects several representative data frames at fixed intervals within that cycle, records their first inference delay and second inference delay respectively, and removes the maximum and minimum delay differences within that cycle. The remaining delay differences are then arithmetically averaged, and this average is used as the final drift complexity index value output for that computation cycle. Simultaneously, to reduce the impact of operating system task scheduling on delay measurement, the inference calculations of the anchor model and the buoy model are configured to be executed in independent high-priority processor threads, thereby obtaining a more stable and accurate drift complexity index. This method provides a stability index that reflects the inherent structural complexity of vibration data. Furthermore, to address known and benign state changes that may occur during actual equipment operation, such as a planned component replacement or operating mode switch, this method may also include a rapid adaptation mechanism for benign states. When maintenance personnel perform such operations, they can issue a command to the system. Upon receiving the command, the system immediately enters a benign state adaptation mode. In this mode, storage decisions based on the baseline drift are temporarily suspended to prevent misjudgments due to state transitions. The system waits for maintenance personnel to provide a vibration data sample representing the new healthy state. Once the sample data is received, a brief online fine-tuning process for the buoy model is initiated, using a composite loss function for its update. The weighting coefficient here This is a key hyperparameter, and its value needs to be determined through offline calibration. The calibration method typically involves using a validation dataset containing old healthy states and new benign states, and testing a series of parameters using methods such as grid search. The candidate values ​​were then selected, and the one that minimized the performance degradation of the buoy model when processing old health data was chosen while rapidly reducing the reconstruction error of new state data. Value, of which, It is a conventional reconstruction loss term, used to drive the buoy model to learn the features of new data samples; It is a knowledge distillation loss term that constrains the buoy model by calculating the difference in the output of the activation layer inside the buoy model and the anchor model when processing some typical old health data, so as to prevent the buoy model from deviating from the representation of the initial health state when learning new knowledge. Through this fine-tuning constrained by the anchor model, the buoy model can quickly and stably adapt to new benign states.

[0028] Example 1: This embodiment describes a specific application of a machine learning-based, sensitive vibration data storage method in a particular industrial scenario. In a long-term condition monitoring application of a 600 MW supercritical steam turbine generator unit, the monitoring system faces the challenge of distinguishing between two vibration changes of different natures but similar signal morphology. One is a fluctuation in operating conditions caused by grid load scheduling, falling within the normal operating range; the other is an extremely slow, gradual wear and deterioration of a thrust bearing due to minor manufacturing defects that begins after several months of operation. In the initial stage of this scenario, after the unit has completed commissioning and entered stable operation, the system uses the collected vibration data, following the procedures disclosed in the aforementioned embodiment, to offline train and lock an anchor point model, while simultaneously deploying a buoy model isomorphic and possesses online learning capabilities. During the first three months of operation, the unit's condition is stable, with vibration changes only caused by normal load fluctuations between 50% and 90% of rated power. During this period, whenever the load changes rapidly, the first error value of the anchor point model... The second error value of the buoy model Both will exhibit brief, unidirectional pulse-like increases simultaneously, but since their trends are basically the same, the difference between them represents the baseline drift. It fluctuated within a low range close to zero, never reaching the preset first threshold. .

[0029] Starting from the fourth month, the slow wear of the thrust bearing began to have a persistent, minor impact on the statistical characteristics of the vibration signal. Because this change was gradual, at any given time, the new vibration data differed very little from the recent normal state to which the buoy model had adapted. Therefore, the buoy model's online learning mechanism absorbed these data points carrying minor degradation information as new normal samples, leading to its second error value. It continues to remain at the previous low level; however, for anchor-point models with locked parameters, this continuous deviation from its initial healthy baseline is unfittable, resulting in the first error value. It begins to deviate from the original baseline, exhibiting a continuous upward trend; at this point, the baseline drift... As a direct measure of the cognitive differences between the two models, starting with The synchronous and continuous growth effectively separates normal operating condition fluctuations from the true deterioration trend, solving the technical problem of the mutual constraint between adaptability and memory in the monitoring field. In the early stage of this process, i.e., the fifth month, wear has not yet formed a clear macro trend, but rather manifests as intermittent minor impacts, which makes the first error value... Irregular and random fluctuations appeared in the time series; although at this time The arithmetic mean has not increased significantly, but its statistical variance over a fixed time window, i.e., the instability index, is significant. The number of fluctuations increased significantly due to these fluctuations and exceeded the preset second threshold first. The system triggered the first data storage based on this. The stored data not only contained the original vibration signals indicating a decrease in the system's dynamic stability, but also included a label indicating that the triggering cause was instability, providing crucial prior information for subsequent analysis. This synergistic dual-channel monitoring mechanism enabled the system to not only identify established long-term drift, but also to capture the precursory signs of instability before drift formation. By the seventh month, with the accumulation of wear... The mean is significantly higher than This causes the reference drift degree It stably exceeded the first threshold. The system began triggering data storage at a higher frequency; each time by Exceeding limits and triggering storage represent a quantifiable and significant deviation of the equipment's health status from its initial baseline. By performing trend analysis on this series of stored data, maintenance personnel can obtain objective evidence of the equipment's degradation rate, thereby replacing the thrust bearing within a planned maintenance window and avoiding unplanned downtime due to sudden failures. This method constructs a stable state reference system by setting up a static anchor point model and a dynamic buoy model in parallel, enabling the monitoring system to penetrate the interference of short-term operating condition fluctuations, accurately identify the long-term evolution trend of the equipment's health status, and precisely use storage resources to record the key data nodes that truly characterize state changes.

[0030] Example 2: To objectively verify the effectiveness of the vibration data efficient storage method based on machine learning-based sensitive monitoring claimed in this invention, particularly its performance in distinguishing between normal operating condition fluctuations and early slow degradation, the following comparative experiment was designed and executed. The experimental data came from a publicly available dataset of rolling bearing vibration signals throughout their entire lifecycle. This dataset records the complete evolution of the bearing from a normal state, through varying degrees of pitting, to eventual failure. Using this dataset, the experiment constructed a simulation sequence capable of simulating progressive wear of a bearing over a 1000-hour operating cycle by chronologically piecing together data segments from different health states. The initial 800 hours of this sequence represented healthy operating data. Starting from hour 801, early weak fault characteristics were gradually mixed in linearly, transitioning to a significant fault state after 950 hours. Two test groups were set up: an experimental group using the method of this invention and a control group using existing conventional monitoring methods. The monitoring system in the experimental group strictly followed the configuration described above, deploying in parallel an anchor point model trained and locked based on the initial 24 hours of health data, and a buoy model with online learning capabilities. Its storage trigger condition was determined by the baseline drift degree. With instability index The control group's monitoring system adopted a common single online learning model structure in the field. This model was consistent with the experimental group's buoy model in terms of initial state and learning rate, but its storage trigger condition was based solely on whether its own reconstruction error exceeded an optimized single threshold. After the experiment started, the simulation data sequence was synchronously input into the monitoring systems of both the experimental and control groups at a uniform time step. During the 0 to 800-hour healthy operation phase of the sequence, both systems performed stably, and the reconstruction error of the control group was consistent with the first error value of the experimental group. Second error value All values ​​remained below 0.02, without triggering any data storage. When the simulation sequence began to inject weak, gradual fault characteristics at 801 hours, the control group's online learning model, due to its continuous adaptability, did not show a significant increase in reconstruction error, remaining consistently at around 0.03 and failing to trigger storage. In contrast, the experimental group's anchor point model, due to its parameter locking, had a first error value... It began to react to this deviation from the initial healthy pattern, exhibiting a slow and continuous increase, while the second error value of the buoy model... The difference between the two, i.e., the baseline drift, is due to the fact that the data has adapted to the new data model and remains essentially unchanged. The unidirectional accumulation begins, and the results are shown in Table 1.

[0031] Table 1: Quantitative comparison of monitoring indicators at key time points during the experiment.

[0032]

[0033] As shown in Table 1, after 910 hours of operation, the reconstruction error of the control group remained below its trigger threshold, failing to identify any anomalies, while the instability index of the experimental group... The test group first exceeded its threshold of 0.002 due to the intermittent impact characteristics of early failures, triggering the first effective data storage; thereafter, as the degradation deepened, the baseline drift of the test group reached a certain level after 950 hours. The error rate increased to 0.056, exceeding its threshold of 0.05, and began to continuously trigger data storage. The monitoring model of the control group, whose knowledge base had been adapted to the deteriorated data, failed to adequately respond to this slow change in its reconstruction error. The experimental results show that the method claimed in this invention, by constructing a differential reference system between the static health benchmark and the dynamic current benchmark, and combining it with the analysis of the dispersion of the error sequence, can distinguish between systematic drift and background noise caused by the slow deterioration of the equipment, and trigger the storage of high-value data in the early stage of the fault. In contrast, conventional methods using a single adaptive model have a monitoring blind spot for such faults.

[0034] To further verify the necessity of the dual-channel quantization index proposed in this invention from a reverse perspective, the following comparative example is set up.

[0035] Comparative Example 1: To objectively verify the necessity and superiority of the dual-channel quantification index of benchmark drift degree and instability index in the method claimed in this invention, compared with conventional monitoring methods that rely solely on the error of a single model, in identifying slowly deteriorating faults, this comparative example was specifically designed and executed. The experimental configuration of this comparative example aims to simulate conventional technical solutions in the background section of this art. Specifically, the monitoring system uses only one online-learning autoencoder neural network model, whose network structure, initial parameters, and online learning rate (…) are described. All of them are completely consistent with the buoy models of the test group in the aforementioned embodiments. The trigger condition for their data storage is set as follows: when the reconstruction error of the model itself exceeds a preset, optimized single threshold (set to 0.05, which is the same as the threshold of the reference drift degree in the aforementioned embodiments). When maintaining consistency for fair comparison, data storage is triggered; the test process uses the same simulation data sequence as the aforementioned embodiment to simulate the progressive wear of the bearing within a 1000-hour operating cycle. The initial 800 hours of this sequence are healthy operating data. Starting from 801 hours, early weak fault characteristics are gradually mixed in in a linear manner, and after 950 hours, it transitions to a significant fault state. This simulation data sequence is input into the monitoring system of this comparative example with a uniform time step. The monitoring indicators and storage decision results at key time nodes are recorded in Table 2.

[0036] Table 2: Quantitative comparison of monitoring indicators at key time points in this comparative example.

[0037]

[0038] The experimental results show that during the entire 1000-hour simulation test cycle, the conventional monitoring method using a single online learning model never exceeded the preset threshold of 0.05 in reconstruction error, and therefore never triggered any data storage. The fundamental reason is that for gradual and slow degradation, the model's online learning mechanism learns and absorbs new data carrying weak fault characteristics as the new normal, causing its knowledge base to drift synchronously with the equipment degradation, thus keeping its reconstruction error at a low level. This experimental result, from the opposite perspective, confirms that without the static health benchmark composed of the anchor point model proposed in this invention, and the differential reference and instability analysis based on it, the conventional monitoring method has a fundamental monitoring blind spot for slow and gradual equipment degradation.

[0039] Example 3: This embodiment combines Figures 1 to 3 This section describes an efficient storage method and system for vibration data based on machine learning-based sensitive monitoring, such as... Figure 1 As shown, continuous real-time vibration data streams are fed in parallel into a parameter-locked anchor point model module and an online-learning buoy model module. The former generates a first error value based on a static health benchmark. The latter generates a second error value based on the current dynamic benchmark. These two error values ​​are then fed into a dual-channel health status quantification module. In one channel, the module quantifies long-term systematic drift by calculating the baseline drift degree, and in the other channel, it detects early signs of instability by calculating the instability index. Subsequently, a storage decision module determines whether to trigger data storage based on whether any indicator exceeds a threshold. If the threshold is not exceeded, redundant data is discarded directly. If the threshold is exceeded, a data storage module is instructed to store high-value data. At the same time, a key module can be triggered to enhance data value, including drift fingerprint vector extraction to extract structured information of drift patterns for fault root cause analysis, and drift complexity index calculation to quantify changes in the intrinsic complexity of vibration signals. This enhanced information is then stored together with the original data in the data storage module.

[0040] like Figure 2As shown, the edge monitoring device is responsible for the core technical process of sensitively monitoring vibration data and triggering high-value data storage when preset conditions are met. Maintenance personnel can then issue instructions to the system to update the static health baseline after the equipment undergoes major repairs or a fundamental change in its condition, or instruct the system to quickly adapt to the new benign operating state when the equipment switches to a new benign operating state. For example... Figure 3 As shown, firstly, the storage decision module obtains the reconstruction results from the anchor model and the buoy model respectively, and calculates the drift residual signal accordingly. Then, the signal is transmitted to the feature extraction module, which extracts the drift fingerprint vector. In parallel, the storage decision module measures the difference in inference latency between the two models and calculates the drift complexity index. Finally, based on whether the index exceeds a preset threshold, the system executes a differentiated storage strategy. If the threshold is exceeded, the original data and drift features are stored together, and the event is marked as a complex drift event. If the threshold is not exceeded, the normal process of storing the original data and drift features is executed. All data is finally transmitted to the data storage module for storage.

[0041] Example 4: This embodiment discloses a specific procedure for offline calibration of key parameters; in the initial deployment of a monitoring system for a large centrifugal compressor, its first threshold... Second threshold The length of the time window for calculating the instability index The online learning rate of the buoy model was determined through the following standardized steps: The initial step of the calibration process was the establishment of the anchor model. This step involved collecting vibration data from an accelerometer deployed at a key bearing of the compressor, which had been continuously running for at least 24 hours under stable operating conditions. This data was confirmed as representing the ideal health state of the equipment, and its sampling frequency was 10.24 kHz. The calibration environment was a computing platform with a standard deep learning framework. Based on this, an autoencoder neural network was constructed as the anchor model. Its network structure was set to include five fully connected layers, with the number of neurons in each layer being 1024-256-64-256-1024, and the activation function was a modified linear unit. The network was trained using the aforementioned collected healthy vibration data until its reconstruction error on the validation set changed less than a predetermined tolerance over several consecutive training cycles. Then, the network parameters were locked, completing the offline construction of the anchor model.

[0042] To determine the length of the time window for calculating the instability index and the corresponding second threshold A test procedure based on injected synthetic signals was executed. First, a test dataset was prepared, consisting of two parts: one part was the previously collected healthy vibration data for one hour, and the other part was based on the healthy data, with a series of low-amplitude pulse signals injected at random intervals for 50 milliseconds to simulate the intermittent impacts caused by early pitting corrosion of the rolling element. Second, a selectable time window length was set. The range is from 50 to 200, with a step size of 10; for each candidate... The test dataset is input into the fixed anchor point model to calculate the first error value. Sequences, and then calculate in different Instability index under value Through analysis, the position of the injected pulse signal was selected. Response value and background health data segment The signal-to-noise ratio between values ​​reaches its maximum. Value, that is , as the length of the time window; in this Once the value is determined, it is calculated using only the health data portion of the test dataset. The value was subjected to statistical distribution analysis, and the value obtained by adding three times the standard deviation to the mean was set as the second threshold. To determine the online learning rate and first threshold of the buoy model. A test procedure based on simulated slow degradation data is executed. The simulation dataset used in this procedure is generated by superimposing healthy data with a weak noise signal whose amplitude increases linearly over time, representing wear characteristics. A list of candidate learning rates is set, containing... , and Three values; for each candidate learning rate, initialize a buoy model and process the simulation dataset in parallel with the anchor model, recording the baseline drift throughout the process. The curves evolve over time; through comparative analysis, selection is made that allows... coefficient of determination for linear fit of curve The learning rate that is the largest and has the smallest short-term standard deviation of volatility. The online learning rate of the buoy model; once this learning rate is determined, it will be... On the evolution curve, the point where the slope first continuously exceeds a certain predetermined slope threshold corresponds to... The value is set as the first threshold. By executing the above series of calibration procedures, all key parameters in the system are set to definite values ​​based on data and established procedures, so that the monitoring method completes all quantitative configuration before being put into actual operation.

[0043] Example 5: This embodiment aims to supplement the explanation of the specific response procedures of the aforementioned method when dealing with two foreseeable non-faulty major state changes during the equipment's operational lifecycle. The first scenario is when the monitored equipment undergoes a planned overhaul or replacement of a critical component, resulting in a permanent change in its inherent vibration baseline. In this case, the previously established anchor point model can no longer represent the ideal health state of the current equipment. To rebuild the static health baseline, the system is configured to respond to a baseline update command issued by an external maintenance management system. Upon receiving the command, the monitoring system suspends the storage decision based on the baseline drift and enters a new health data acquisition cycle. During this cycle, the system continuously collects and uploads a preset duration of vibration data marked as the new health state to the cloud server. Using this batch of new health vibration data in the cloud, following the same technical procedures as the initial model construction, a new anchor point model is retrained and generated. Finally, this new anchor point model is deployed to the edge monitoring device to replace the original anchor point model, thereby completing the update of the static health baseline.

[0044] The second scenario involves equipment switching to a known but benign operating condition due to production plan adjustments. For example, a multi-purpose mixer might switch from processing low-viscosity materials to processing high-viscosity materials, resulting in an immediate and significant change in its vibration pattern. To prevent the system from misinterpreting this benign change as an abnormal drift requiring data storage, the system is configured to receive a benign state switching command associated with this specific operating condition. Upon receiving this command, the system triggers an online rapid fine-tuning process for the buoy model. During this process, the system acquires a vibration data sample representing the new benign operating condition and updates the buoy model's internal parameters using a composite loss function that includes a reconstruction loss term and a knowledge distillation loss term. The knowledge distillation loss term here serves to constrain the buoy model's representation of the initial healthy vibration data while accelerating its adaptation to the new operating condition data pattern, using a fixed anchor point model as a reference. By executing this briefly constrained fine-tuning process, the buoy model's knowledge baseline is transferred to the new benign state, reducing the baseline drift. It can quickly return to near zero.

[0045] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for efficient storage of vibration data based on machine learning-based sensitive monitoring, characterized in that, Includes the following steps: The current vibration data is simultaneously input into the anchor point model and the buoy model to obtain the first error value and the second error value, respectively. The reference drift is calculated based on the difference between the first error value and the second error value. For multiple first error values ​​generated consecutively within a preset time window, calculate the instability index that characterizes the degree of discreteness of their sequence; When the reference drift exceeds a preset first threshold, or when the instability index exceeds a preset second threshold, the storage of the current vibration data is triggered. The buoy model continuously adapts its internal parameters to the real-time vibration data from the monitored device through an online learning mechanism. The online learning mechanism has the inherent characteristic that the knowledge base of the buoy model drifts synchronously with the slow deterioration of the device. The internal parameters of the anchor point model are trained offline using only historical health vibration data from the device under monitoring and are locked to form a static health benchmark for measuring the drift of the buoy model's knowledge benchmark.

2. The method for efficient storage of vibration data based on machine learning-based sensitive monitoring according to claim 1, characterized in that, The instability index is the statistical variance of the first error value sequence. The calculation method is as follows: in, The length of the preset time window, For the first time window The first error value, It is the arithmetic mean of all first error values ​​within the time window.

3. The method for efficient storage of vibration data based on machine learning-based sensitive monitoring according to claim 1, characterized in that, Also includes: When the storage of the current vibration data is triggered, a drift residual signal is generated based on the reconstruction results of the current vibration data using the anchor point model and the buoy model. A drift fingerprint vector of a predetermined dimension is extracted from the drift residual signal using a preset feature extraction module. The current vibration data and the drift fingerprint vector are then stored together. The drift residual signal is obtained by calculating the difference between the reconstruction results of the anchor point model and the reconstruction results of the buoy model.

4. The method for efficient storage of vibration data based on machine learning-based sensitive monitoring according to claim 1, characterized in that, Both the anchor model and the buoy model are autoencoder neural network models, and the first error value and the second error value are the reconstruction errors of the corresponding autoencoder neural network models.

5. The method for efficient storage of vibration data based on machine learning-based sensitive monitoring according to claim 1, characterized in that, The steps of obtaining the first error value and the second error value also include: measuring the first inference delay required for the anchor model to generate the first error value and the second inference delay required for the buoy model to generate the second error value, and calculating the drift complexity index defined by the difference between the first inference delay and the second inference delay; when storing the current vibration data, the steps also include: jointly storing the current vibration data, the reference drift degree and the drift complexity index as data records.

6. The method for efficient storage of vibration data based on machine learning-based sensitive monitoring according to claim 5, characterized in that, The storage of current vibration data also includes: when the storage of current vibration data is triggered, if the drift complexity index exceeds a preset complexity threshold, the storage event is marked as a complex drift event.

7. The method for efficient storage of vibration data based on machine learning-based sensitive monitoring according to claim 1, characterized in that, When an instruction is received indicating that the device to be monitored has entered a new benign operating state, vibration data samples characterizing the new benign operating state are acquired, and the online fine-tuning process of the buoy model is initiated; in the online fine-tuning process, a composite loss function is used to update the internal parameters of the buoy model.

8. The method for efficient storage of vibration data based on machine learning-based sensitive monitoring according to claim 7, characterized in that, The composite loss function includes a reconstruction loss term that drives the buoy model to learn the vibration data sample patterns, and a knowledge distillation loss term that constrains the buoy model's understanding of historical healthy vibration data from deviating from the anchor model. The knowledge distillation loss term achieves the constraint by calculating the difference in the output of the internal activation layer between the buoy model and the anchor model when processing the same historical healthy vibration data.

9. The method for efficient storage of vibration data based on machine learning-based sensitive monitoring according to claim 1, characterized in that, The storage of current vibration data also includes: when the second error value exceeds the preset third threshold, the storage of current vibration data is also triggered, and the triggering cause is marked as a sudden impact; after the monitored equipment undergoes major maintenance or a fundamental change in its condition, in response to external instructions, new healthy vibration data is collected, a new anchor point model is retrained in the cloud using the new healthy vibration data, and then sent to the monitoring equipment to update the static health benchmark.

10. A high-efficiency vibration data storage system based on machine learning-based sensitive monitoring, used to implement the high-efficiency vibration data storage method based on machine learning-based sensitive monitoring as described in claim 1, characterized in that, include: The buoy model module is used to continuously adapt to the current vibration data acquired in real time through an online learning mechanism. The online learning mechanism has the inherent characteristic that the knowledge base of the buoy model module drifts synchronously with the slow deterioration of the equipment. The anchor point model module, whose internal parameters are locked, is used to form a static health baseline; The data processing module is used to receive the current vibration data and simultaneously input the current vibration data into the anchor point model module and the buoy model module to obtain the first error value and the second error value respectively. The health status quantification module is used to calculate the baseline drift based on the difference between the first error value and the second error value; and to calculate the instability index, which characterizes the degree of discreteness of the sequence, for multiple first error values ​​generated continuously within a preset time window. The storage decision module is used to generate a storage trigger command when the baseline drift exceeds a preset first threshold or when the instability index exceeds a preset second threshold. The data storage module is used to store the current vibration data in response to the storage trigger command generated by the storage decision module.

Images