Bearing running state real-time monitoring method combining edge calculation and adaptive threshold judgment

Abstract

The invention relates to the technical field of state real-time monitoring, in particular to a bearing running state real-time monitoring method combining edge calculation and self-adaptive threshold judgment, which comprises the following steps: respectively acquiring original running signals of corresponding bearings, and obtaining a state initial feature set; generating a preliminary state judgment label based on the basic threshold set; receiving an initial feature set and an initial state judgment label, analyzing the state consistency of multiple edge nodes by a central coordination node, establishing a threshold drift trend model, and predicting a future evolution direction; meanwhile, key working condition factors are extracted, a working condition linkage factor set is generated, and a drifting response model of prediction driving is formed; issuing the model to each edge node, correcting threshold offset, and generating a state monitoring conclusion; and if the fusion result triggers an abnormal standard, immediately outputting real-time early warning information to a field system or an upper platform. According to the method, the comprehensive requirements on real-time performance, accuracy and disposition performance in a complex industrial scene are met, and the overall reliability and the intelligent level of the system are improved.

Description

Technical Field

[0001] This invention relates to the field of real-time condition monitoring technology, and in particular to a method for real-time monitoring of bearing operating conditions that combines edge computing and adaptive threshold judgment. Background Technology

[0002] With the continuous improvement of industrial intelligence, real-time monitoring and fault early warning of critical equipment operation status have become core links in ensuring production safety and improving operation and maintenance efficiency. Traditional status monitoring methods mainly rely on central servers to centrally collect and analyze data from various sensors. Although they have a certain degree of accuracy, they still have the following problems in actual complex industrial sites: insufficient full utilization of edge nodes, resulting in a significant increase in response latency and network pressure.

[0003] Existing methods typically upload raw multi-dimensional operational data such as vibration, temperature, and current to the cloud for unified processing. In scenarios with limited network bandwidth or widely distributed equipment deployments, this can easily lead to data congestion, delayed judgment, or even data loss, hindering the timely detection of potential anomalies. State judgment algorithms lack adaptability to dynamic changes in operating conditions, resulting in rigidly fixed threshold settings and high false positive rates. Most methods employ statically set feature threshold judgment strategies, failing to dynamically adjust judgment criteria based on the real-time operating environment of the equipment, such as load, speed, and temperature. This easily leads to normal states being misjudged as abnormalities or early faults going undetected, reducing the accuracy and practicality of fault detection. Furthermore, they neglect the correlation and collaborative judgment mechanisms between multiple monitoring nodes, making it difficult to identify hidden risks such as threshold drift. Industrial equipment often has structural coupling or process linkage relationships, but most existing systems treat edge nodes as independent units, lacking the ability to fuse and analyze information such as synchronous feature change trends and topology contagion effects, failing to promptly identify overall feature drift or local fault evolution caused by co-variation of operating conditions. Summary of the Invention

[0004] This invention provides a real-time monitoring method for bearing operating status that combines edge computing and adaptive threshold judgment. It is an industrial equipment status monitoring method with edge intelligence, collaborative modeling and dynamic threshold adaptation capabilities, so as to meet the needs of efficient monitoring and accurate early warning in multi-node distributed deployment and dynamic operating conditions.

[0005] A method for real-time monitoring of bearing operating status combining edge computing and adaptive threshold judgment includes the following steps:

[0006] S1: At edge nodes distributed at different monitoring points, the original operating signals of the corresponding bearings are collected respectively, and a lightweight feature extraction algorithm is executed locally to obtain an initial state feature set including vibration amplitude features, frequency domain impact features and short-time energy distribution features; based on the locally set basic threshold set, a preliminary state judgment label is generated.

[0007] S2: Receive the initial state feature set and preliminary state judgment label of the output. The central coordination node analyzes the consistency of the states of multiple edge nodes, identifies the node groups with drift, establishes a threshold drift trend model based on the change trend within the time window, and predicts the future evolution direction. At the same time, it extracts the key working condition factors that cause drift, generates a set of working condition linkage factors, and adjusts the threshold correction strategy of each edge node accordingly to form a prediction-driven drift response model.

[0008] S3: The drift response model is sent to each edge node to correct the threshold offset in the local state judgment process, and a fused state monitoring conclusion is generated based on the corrected judgment result; if the fused result triggers the abnormal standard, real-time early warning information is immediately output to the field system or upper-level platform.

[0009] Optionally, S1 specifically includes:

[0010] S11: Each edge node periodically collects the vibration acceleration signal of the corresponding bearing at a fixed sampling frequency and duration, which serves as the original operating signal;

[0011] S12: Execute the lightweight feature extraction algorithm on the original running signal locally to obtain an initial feature set of vibration amplitude features, frequency domain impact features, and short-time energy distribution features.

[0012] Optionally, the lightweight feature extraction algorithm includes processing the original running signal after each edge node completes the buffering of the original running signal, and extracting vibration amplitude features, frequency domain impact features, and short-time energy distribution features, wherein:

[0013] The effective value of the vibration signal is calculated as the vibration amplitude characteristic.

[0014] The signal is envelope demodulated and the amplitude at the bearing fault characteristic frequency is analyzed as the frequency domain impact characteristic.

[0015] Wavelet packet decomposition is performed on the signal, and the entropy of the proportion of energy in each sub-band to the total energy is calculated as a short-time energy distribution characteristic.

[0016] Optionally, each feature in the initial state feature set is compared with the corresponding threshold range in the locally stored basic threshold set, and the initial state judgment label, identified as normal, attention, or warning, is output according to the preset multi-level judgment logic.

[0017] Optionally, the central coordination node receives the initial state feature set and the preliminary state determination label from all edge nodes; S2 specifically includes:

[0018] S21: Based on the physical subsystem or topological proximity of the bearings corresponding to each edge node, divide all edge nodes into one or more logical node groups;

[0019] S22: For each node group, calculate the correlation coefficient of similar feature sequences in the initial state feature set of each edge node in the node group within the same time period, as the group synergy index; when the group synergy index is lower than the preset correlation threshold, determine that there is a node group with correlation threshold drift.

[0020] S23: For the identified node group, extract the initial state feature set sequence of the identified node group within a preset historical time window, and use a time series prediction algorithm to fit and predict the initial state feature set sequence to establish a threshold drift trend model describing the overall drift direction and rate of features within the group.

[0021] S24: Synchronously acquire historical operating condition parameters shared by the equipment or area where the node group is located, including load, speed and ambient temperature; through correlation analysis, select the subset of parameters that are most strongly correlated with the characteristic drift trend of the node group from the operating condition parameters to form a key operating condition factor set.

[0022] S25: Based on the predicted values ​​of the characteristics of future periods by the threshold drift trend model, and the quantitative influence relationship of the key working condition factor set on the feature drift, construct a function model with real-time working condition factors as input and the correction amount of specific thresholds in the basic threshold set of each edge node as output, as the drift response model.

[0023] Optionally, the time series prediction uses a sliding window regression prediction algorithm, which specifically includes: extracting a continuous sequence of initial state features within a fixed-length historical time window as a sample window; constructing a regression model based on the feature change data within the sample window to fit the overall trend; and performing short-term extrapolation of the feature values ​​for future periods at the end of the window. The resulting prediction results are used to determine whether there is a trend drift in the features.

[0024] Optionally, the correlation coefficient includes: extracting the initial feature sequences of the same state of all edge nodes in the same node group within the same time period, calculating the Pearson correlation coefficient between feature sequences pair by pair to measure the consistency of feature change trends, and after completing the correlation calculation of all node pairs, averaging all correlation coefficients to generate the average correlation index of the node group under the feature dimension, which is used to describe the consistency of the overall feature fluctuation direction.

[0025] Optionally, S3 specifically includes:

[0026] S31: The central coordination node encapsulates the drift response model into configuration instructions that include model parameters, calculation rules, and applicable operating conditions, and sends them to each edge node in the node group;

[0027] S32: Each edge node receives and parses the configuration instruction, updates the drift response model stored locally; in each subsequent local state judgment process, it obtains the current set of key working condition factors in real time, inputs the updated drift response model, calculates the correction amount for each threshold in the local basic threshold set, and applies the correction amount to generate the corrected dynamic threshold set.

[0028] S33: Each edge node uses the modified dynamic threshold set to re-judge the initial state feature set collected and extracted in real time, generates a local state judgment result, and reports the local state judgment result and the feature values ​​used to the central coordination node.

[0029] S34: The central coordination node receives the local state determination results and feature values ​​of each edge node in the node group, and executes the fusion logic to generate the fused state monitoring conclusion;

[0030] S35: Compare the fused status monitoring conclusion with the preset anomaly standard. If the conclusion indicates that there is an anomaly or the fault risk level exceeds the threshold, generate real-time early warning information.

[0031] Optionally, the fused state monitoring conclusions include voting on the judgment results, weighted evaluation based on the feature values ​​of each node, and fault propagation reasoning based on the device topology.

[0032] Optionally, the real-time early warning information includes the location of the edge node to which the anomaly belongs, the corresponding risk level identifier, the cause description, and recommended handling measures.

[0033] The beneficial effects of this invention are:

[0034] 1. This invention introduces a lightweight multi-dimensional state feature extraction algorithm locally at the edge node, integrating vibration amplitude features, frequency domain impact features, and short-time energy distribution features to achieve rapid modeling and preliminary local judgment of equipment operating status, reducing the computational burden on edge devices. Simultaneously, it proposes a threshold correction mechanism based on dynamic adjustment of operating conditions, supporting edge nodes to adaptively adjust the basic discrimination threshold, improving the dynamic adaptability and accuracy of state monitoring.

[0035] 2. This invention divides edge nodes into logical analysis node groups based on physical or topological correlation through a central coordination node, and uses a coordination degree index to determine the consistency of features within the group, thus achieving accurate identification of correlation threshold drift. Furthermore, a time series prediction algorithm is used to construct a feature change trend model, and a threshold response model is constructed in conjunction with key operating conditions. This avoids the false alarms or missed alarms caused by unreasonable fixed threshold settings in traditional methods, enhancing the early warning capability for weak faults and potential risks.

[0036] 3. This invention designs an edge-middle collaborative closed-loop structure, supporting edge nodes to receive drift response models issued by the coordinating node and generate dynamic threshold sets based on local key operating condition inputs, thereby re-judging and feeding back the state characteristics. The coordinating node can generate fused judgment results based on various strategies such as voting, feature weighting, or topology propagation reasoning, and output early warning information including location, level, and handling suggestions in conjunction with the anomaly level, meeting the comprehensive requirements of real-time performance, accuracy, and manageability in complex industrial scenarios, and improving the overall reliability and intelligence level of the system. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of the logical framework of an embodiment of the present invention. Detailed Implementation

[0040] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0041] like Figure 1 and Figure 2 As shown, a method for real-time monitoring of bearing operating status combining edge computing and adaptive threshold judgment includes the following steps:

[0042] S1: At edge nodes distributed at different monitoring points, the original operating signals of the corresponding bearings are collected respectively, and a lightweight feature extraction algorithm is executed locally to obtain an initial state feature set including vibration amplitude features, frequency domain impact features and short-time energy distribution features; based on the locally set basic threshold set, a preliminary state judgment label is generated.

[0043] S1 specifically includes:

[0044] S11: Each of the aforementioned edge nodes is sampled at a fixed frequency. and sampling duration The vibration acceleration signals of the corresponding bearings are periodically collected and recorded as the original operating signal sequence: ;

[0045] in, This represents the original vibration signal; the specific steps are as follows:

[0046] Sampling parameter configuration:

[0047] During the system initialization phase, a uniform sampling frequency is set for each edge node. and sampling duration The sampling frequency is usually set to be more than 10 times higher than the upper limit of the main fault characteristic frequency of the bearing to meet the requirements of the sampling theorem, with a range of 5 to 20, and the sampling duration is 0.5 to 2 seconds.

[0048] Sensor activation and data preparation: Each edge node is connected to an accelerometer, which is fixedly mounted on or near the bearing housing. The node control module activates the sensor according to a periodic scheduling signal and turns on the analog-to-digital conversion circuit to prepare for the acquisition of analog vibration signals.

[0049] Vibration signal acquisition process: Within a set time window, edge nodes are sampled at a set frequency. Acceleration signals are continuously acquired and converted from analog to digital signals in real time, forming a sequence of sampling points. The acquisition duration is... During this time period, a complete record of vibration acceleration signal is formed.

[0050] Signal buffering and labeling: After acquisition, the digitized vibration signal segment is labeled as the original operating signal of the current cycle, denoted as... ,in This signal records the complete operating vibration behavior of the bearing at the current moment; the edge node temporarily stores this signal in its local buffer, along with auxiliary information such as timestamp, node number, and sensor number.

[0051] Periodic execution mechanism: The above data collection process is executed cyclically at a fixed period. The sampling period is typically slightly longer than [a certain value]. This allows for signal processing time and ensures the decoupled operation of the data acquisition and feature extraction modules. Edge nodes can dynamically adjust their sampling period according to system strategies to match the operating rhythm and anomaly detection requirements of different devices.

[0052] S12, regarding the original operating signal A lightweight local feature extraction algorithm is executed to obtain three feature dimensions:

[0053] S121, the vibration amplitude characteristic specifically includes: squaring the instantaneous acceleration values ​​of the entire signal within the time window, calculating their mean, and then taking the square root of this mean to obtain the effective value representing the overall intensity of the vibration signal segment. This characteristic reflects the overall stress state and excitation degree of the bearing, characterizes the overall energy level of the vibration signal, and reflects the overall response intensity of the bearing structure. The vibration amplitude characteristic is expressed as:

[0054] ;

[0055] in, Indicates the sampling duration Within, the integral of the instantaneous energy of the signal is summed. This represents the time averaging factor, used to normalize the integral value.

[0056] S122, Frequency Domain Impact Characteristics: To identify the periodic impact response of a bearing at specific frequencies, envelope analysis is used to extract key frequency components. Specifically, this involves: first, demodulating the original vibration signal using its envelope to extract the envelope curve showing the amplitude variation over time; then, performing a Fast Fourier Transform on this envelope signal to obtain its spectral distribution. Within the spectrum, the focus is on observing the typical fault characteristic frequencies of the bearing, such as the structural frequencies of the inner ring, outer ring, and rolling elements, and extracting the amplitude information at the corresponding frequencies. This is represented as:

[0057] S1221, for signal Constructing the envelope signal by performing Hilbert transform :

[0058] ;

[0059] in, Represents the imaginary unit, satisfying , Represents the result of the Hilbert transform, indicating The orthogonal analytic signal; the Hilbert transform is a signal analysis method used to extract phase information and amplitude envelope from real signals, applied in fields such as mechanical vibration analysis and impact fault detection. In this invention, the Hilbert transform converts the bearing vibration signal into its corresponding analytic signal, thereby obtaining the signal envelope for identifying potential periodic impact components. Specifically, for the acquired original vibration acceleration signal, the Hilbert transform constructs an imaginary part signal orthogonal to the original signal, forming a complex analytic signal with the original real part signal. The amplitude of this analytic signal is the envelope signal. The specific steps are as follows:

[0060] Input real signal: Read a segment of the original vibration signal acquired by the edge node and use it as the transformation input;

[0061] Constructing orthogonal components: The orthogonal components of the original signal are generated through Fourier transform, filtering and inverse transform, which are the results after Hilbert transform;

[0062] Generating analytic signals: Combining the original signal with its orthogonal components to form a complex analytic signal;

[0063] Extracting envelope information: Calculate the amplitude of the analytical signal to obtain the envelope curve.

[0064] S1222, for Perform a Fourier transform to obtain the envelope spectrum. In the known set of bearing characteristic frequencies Extract its amplitude:

[0065] ;

[0066] in, For the inner ring fault frequency, For the outer ring fault frequency, To accurately capture the periodic impact characteristics caused by early damage during bearing operation, a frequency domain feature extraction method based on envelope analysis is introduced at the edge nodes to identify the response intensity at key fault frequencies from the vibration signal, representing the rolling element failure frequency. This process mainly focuses on analyzing the maximum value of the envelope amplitude at the characteristic frequency, and the specific steps are as follows:

[0067] Envelope signal acquisition: Perform Hilbert transform on each segment of the original vibration acceleration signal to extract the amplitude envelope sequence of the signal.

[0068] Spectrum analysis and processing: Perform a Fast Fourier Transform on the envelope signal to obtain the envelope spectrum. This spectrum reflects the frequency components included in the envelope and their corresponding amplitude information.

[0069] Characteristic frequency selection: Based on the structural parameters and rotational speed conditions of the bearing being monitored, its key fault characteristic frequencies are pre-calculated, including the inner ring fault frequency, outer ring fault frequency, and rolling element fault frequency. These frequencies can be obtained through empirical formulas or provided by the manufacturer.

[0070] Frequency domain amplitude comparison: Find the amplitude information at the above characteristic frequency positions in the envelope spectrum, and take the maximum value among the three as the frequency domain impulse characteristic value of this period.

[0071] Feature value output: The maximum amplitude extracted at the end is used as the frequency domain impulse feature of this period to construct the initial feature set.

[0072] S123, Short-time energy distribution characteristics, i.e., wavelet packet energy entropy. Considering that under certain fault conditions, vibration energy may be concentrated in a specific frequency band or exhibit a non-uniform distribution over time, an energy entropy feature based on wavelet packet decomposition is introduced. Specifically, the original vibration signal is decomposed into several sub-frequency bands using multi-level wavelet packet decomposition, and the proportion of vibration energy in each sub-frequency band is calculated. Then, based on the distribution of these proportions, the information entropy is calculated, representing the degree of dispersion of signal energy across different frequency bands, expressed as:

[0073] S1231, for Perform wavelet packet decomposition to the th Layer, to obtain Each sub-band signal is used to calculate the energy ratio of each sub-band. :

[0074] ;

[0075] in, Indicates the first To more sensitively capture the frequency band concentration and uneven distribution characteristics of vibration energy under bearing fault conditions, wavelet packet decomposition is introduced at the edge nodes after acquiring the vibration signal. This allows for fine-grained analysis of the signal's energy distribution, and energy distribution features are constructed based on the energy proportion of each frequency band to assist in state judgment. The specific process is as follows:

[0076] Wavelet packet decomposition: Edge nodes perform multi-level wavelet packet decomposition on the current segment of the original running signal. Wavelet packets are an analysis method with strong time-frequency localization capabilities, capable of dividing the signal into full frequency bands at different frequency resolutions. The number of decomposition levels can be set. The signal can be decomposed into Each frequency band covers a different frequency range.

[0077] Band energy calculation: For each sub-band, extract its corresponding signal component and calculate its energy value within the current time period. Energy is typically defined based on the cumulative square of the signal amplitude, representing the energy intensity of that band within the entire signal.

[0078] Energy normalization processing: In order to facilitate the comparison of the relative distribution of energy between different frequency bands, the energy of each sub-band is divided by the sum of the energies of all frequency bands, and its energy proportion in the overall signal is calculated.

[0079] Constructing an energy percentage vector: After normalization, an energy percentage vector is formed, where each component represents the proportion of energy in a sub-band in the total energy.

[0080] Feature output and application: The energy proportion of each sub-band can be further used to calculate entropy-type features, such as energy entropy and information entropy.

[0081] S1232, and then calculate the distribution entropy: ;

[0082] in, For the first The energy of a sub-band signal Its proportion, Entropy represents the energy distribution complexity, which reflects the complexity of the energy distribution. Entropy represents the energy distribution, i.e., the complexity or discreteness of the energy distribution. It is a statistical characteristic that reflects whether the energy in a signal is concentrated in a few frequency bands or exhibits a uniform diffusion pattern. It is suitable for determining whether a bearing has potential irregular excitation sources or abnormal frequency structure. The specific processing flow is as follows:

[0083] Wavelet packet decomposition: The original vibration signal undergoes multi-level wavelet packet decomposition, with the number of decomposition levels set to [value missing]. For example, with 3 to 5 layers, the entire frequency domain can be divided into Each sub-band has a non-overlapping width. Each sub-band corresponds to a local frequency range, which is used to capture local vibration response characteristics.

[0084] Sub-band energy calculation: Extract the reconstructed signal corresponding to each sub-band, and calculate the time-domain energy of each signal segment to obtain the energy value of each sub-band.

[0085] Energy proportion normalization: Normalize the energy of all sub-bands, calculate the proportion of each sub-band's energy to the total energy, and obtain a set of normalized energy proportion values.

[0086] Energy entropy calculation: Based on the distribution of energy proportions, calculate the energy distribution entropy of the entire signal at this decomposition level. If the energy is highly concentrated in a few frequency bands, the entropy value is low, indicating that the signal has a clear dominant frequency structure; if the energy is evenly distributed across multiple frequency bands, the entropy value is high, indicating that the signal spectrum exhibits a diffuse state or enhanced background disturbance.

[0087] Feature value output: The final energy distribution entropy is used as the short-time energy distribution feature value under this sampling period, recorded in the initial feature set of the state, and participates in the subsequent state evaluation and threshold judgment process.

[0088] The above three features together constitute the initial state feature set: ;

[0089] S13, Initial state determination label output: Each feature in the initial state feature set is compared with the basic threshold set stored locally by the edge nodes. Perform a one-to-one comparison.

[0090] S131, the basic threshold set includes a normal upper limit, an attention upper limit, and a warning upper limit for each type of feature:

[0091] ;

[0092] The basic threshold set refers to a set of multi-level feature threshold standards that are preset locally or generated by self-learning at the edge nodes. It is used to classify and judge the initial feature set of the currently collected bearing state. The basic threshold set is organized according to feature categories and divided into three levels in each feature dimension: normal, attention, and warning. The content of the basic threshold set mainly includes the graded thresholds corresponding to the following three feature dimensions.

[0093] The three-segment threshold group for vibration effective value characteristics includes: the upper limit of the normal threshold, used to distinguish between normal and attention; the upper limit of the attention threshold, used to distinguish between attention and warning; and the upper limit of the warning threshold, which triggers a warning if the threshold is exceeded.

[0094] The three-stage threshold set for frequency domain impulse characteristics includes: upper limit of normal threshold, upper limit of attention threshold, and upper limit of warning threshold, used to determine whether a periodic impulse response exists.

[0095] The three-segment threshold group for short-time energy distribution characteristics includes: upper limit of normal threshold, upper limit of attention threshold, and upper limit of warning threshold for judging whether the spectral energy is abnormally dispersed.

[0096] S132, based on the preset multi-level judgment logic, output the preliminary state judgment label. The specific rules are as follows:

[0097] If all features are below the corresponding normal threshold: the output label is normal;

[0098] If any feature exceeds the corresponding attention threshold but does not exceed the warning threshold: output the label "Attention";

[0099] If any feature exceeds the corresponding warning threshold: output a warning label.

[0100] The reference ranges for the upper limits of the normal, attention, and warning thresholds for the three key features used are as follows:

[0101] 1. Characteristics of effective vibration value :

[0102] Normal threshold upper limit: 0.2 to 0.4; attention threshold upper limit: 0.4 to 0.7; warning threshold upper limit: greater than 0.7.

[0103] 2. Frequency domain impulse characteristics :

[0104] Normal threshold upper limit: 3 to 5, attention threshold upper limit: 5 to 8, warning threshold upper limit: greater than 8;

[0105] 3. Short-time energy distribution characteristics :

[0106] Normal threshold upper limit: 1.5 to 2.2, attention threshold upper limit: 2.2 to 2.6, warning threshold upper limit: greater than 2.6.

[0107] S2: Receive the initial state feature set and preliminary state judgment label of the output. The central coordination node analyzes the consistency of the states of multiple edge nodes, identifies the node groups with drift, establishes a threshold drift trend model based on the change trend within the time window, and predicts the future evolution direction. At the same time, it extracts the key working condition factors that cause drift, generates a set of working condition linkage factors, and adjusts the threshold correction strategy of each edge node accordingly to form a prediction-driven drift response model.

[0108] S2 specifically includes:

[0109] The central coordinating node in S2 receives the initial state feature set from all edge nodes. and the corresponding preliminary status determination label ,in Indicates the edge node number.

[0110] S21, The central coordinating node maintains an adjacency matrix representing the physical or logical connection relationships between edge nodes. Its construction rules are as follows:

[0111] ;

[0112] This adjacency relationship can be constructed based on one of the following information: the structural hierarchy of the equipment module in which the bearing is located, such as fan group, compressor group, etc.;

[0113] Are the physical distances between nodes less than a set threshold? Is there functional coupling in the process flow diagram?

[0114] Logical node group partitioning: based on adjacency matrix The connectivity clustering algorithm, which divides all edge nodes into several non-overlapping logical node groups, uses a graph connectivity component extraction algorithm based on the adjacency matrix to divide nodes into multiple node groups.

[0115] ;

[0116] Each of them Represents a group of nodes. ;

[0117] Each node group includes a set of edge nodes that are physically close to each other or have a coupled relationship in their operating status, which are used for subsequent execution of consistency measurement within the group and establishment of threshold drift trends;

[0118] S22, Calculation of Coordination Index and Identification of Drift Node Groups:

[0119] S221, in each node group In this context, let the group include... There are several edge nodes. For each type of initial state feature, namely the effective value of vibration, frequency domain impact characteristics, and short-time energy distribution characteristics, the central coordination node extracts the data for each node within a specified time window. A continuous sequence of eigenvalues ​​within a given range is represented as:

[0120] ;

[0121] in, Represents a node group The Middle Time series of similar features of each node.

[0122] S222, Method 1 for calculating the synergy index: Correlation coefficient method:

[0123] For any two nodes within the group Calculate the Pearson correlation coefficient between their characteristic sequences:

[0124] ;

[0125] in, Represents a node The mean of the characteristic sequence, The Pearson correlation coefficient is a method for measuring the consistency of trends between two time series. In this invention, it is used to analyze whether the changing trends of different edge nodes within the same node group are synchronous in a certain type of feature. Specifically, for any two edge nodes within a node group, continuous value sequences of a certain initial state feature are extracted within the same time window, and the changing patterns of these two sequences in the time dimension are compared. If the feature values ​​of the two nodes show the same or similar trends in their fluctuations, i.e., they rise and fall together, their Pearson correlation coefficient will be high, close to 1; if the trends are opposite, the coefficient will be close to -1; if there is no obvious trend relationship, the coefficient will be close to 0. By averaging the correlation coefficients between all pairs of nodes within the same group, the overall synergy of the group in that feature dimension can be obtained. If the average correlation is lower than a set threshold, it indicates that there is a trend divergence in the state features of the nodes within the group, which may be a drift phenomenon caused by changes in operating conditions, local anomalies, or threshold imbalance.

[0126] The average correlation coefficient and synergy index within the group are obtained by averaging across all node pairs.

[0127] ;

[0128] S223, Method 2 for calculating the synergy index: Euclidean distance method:

[0129] Directly measure the average Euclidean distance between feature sequences of different nodes:

[0130] ;

[0131] The corresponding average distance index within the group is:

[0132] ;

[0133] in, The Euclidean distance method measures the relative difference in numerical magnitude between two time series. In this invention, it is used to analyze whether there is a shift in the numerical level of similar features among different edge nodes within the same node group. Specifically, two edge nodes within the node group are selected, and continuous numerical sequences of a certain feature are extracted within the same time window. The two sets of values ​​are subtracted point by point, squared, and summed. The square root is then taken to obtain the overall degree of difference between the two sequences. If the feature sequence values ​​of the two nodes are close, the Euclidean distance is smaller, indicating consistent state behavior; conversely, the larger the distance, the more likely there is a shift in state perception. Similarly, the average Euclidean distance between all pairs of nodes within the group is taken to obtain the overall degree of difference of the group in that feature dimension. If the average distance exceeds a preset upper limit, it indicates that the state perception values ​​of some nodes within the group have experienced numerical-level drift. The synergy analysis process in this step is performed independently for each type of feature. Each type of feature has the following significance and role in the analysis:

[0134] The effective vibration value characteristic mainly reflects the overall excitation level of the bearing where the node is located. If the trend of this characteristic varies greatly among nodes in the same group, it may indicate uneven equipment load or sensor calibration differences, providing a high-level drift indication capability. Frequency domain impact characteristics are used to identify periodic local impact behavior. The synergy of this characteristic can reveal whether asynchronous micro-damage evolution exists within the same structural region, and it is highly sensitive to early damage drift. Short-time energy distribution characteristics, i.e., energy entropy, are used to describe the degree of diffusion of signal spectral energy. If the nodes within a group have poor synergy on this characteristic, it may reflect the alienation process of the spectral energy structure, and is an indirect indicator of systemic drift caused by abnormal operating conditions. Therefore, these three types of characteristics provide a foundation for drift detection at three levels: macroscopic intensity, local impact, and spectral complexity. This step, through their synergy analysis, can achieve early identification of drift risks within the node group.

[0135] S223, when Below the set correlation threshold or Distance above the set threshold If the correlation threshold is considered to have shifted within the group, it is marked as a shift-sensitive node group.

[0136] In the scheme described in this invention, the two collaboration degree judgment thresholds used to identify whether there is correlation threshold drift within the node group are specifically set as follows:

[0137] Correlation threshold When using characteristic correlation coefficients, such as the Pearson correlation coefficient, as an indicator of synergy, if the average correlation coefficient within the node group is... If the value is below 0.85, it is considered that the consistency of the characteristic behavior between nodes has decreased, and there is a risk of threshold drift.

[0138] Euclidean distance threshold When using the average Euclidean distance between features When used as an indicator of synergy, if the average Euclidean distance is higher than 0.12 (after normalization), it is considered that the eigenvalue distribution deviates from the normal synergy range and there is potential drift.

[0139] S23, for the identified drift-sensitive node group, extract its data within a historical time window. The sequence of initial feature sets within a given time series is denoted as the time series. Using time series forecasting algorithms, due to their low computational complexity, low deployment cost, and efficient operation at edge or central coordination nodes, they are suitable for trend fitting and short-term forecasting of state characteristic values ​​within short periods. Specifically, a sliding window regression forecasting algorithm is used to model the trend of this characteristic sequence, fitting its drift direction and rate of change to construct a threshold drift trend model.

[0140] ;

[0141] in, Represents the predicted future feature value. Represents a node The feature trend prediction function is a function model used to simulate and predict the changing trend of initial features of edge nodes over a future time period. This function takes the feature sequence within a historical time window as input and outputs an estimate of future feature values, reflecting the direction, rate, and degree of feature change. The process of constructing a threshold drift trend model, based on the feature trend prediction function, mainly includes the following steps:

[0142] Extracting historical feature sequences: From the group of nodes identified as drift-sensitive, collect the initial state features of each node within a fixed historical time window, namely, the effective value of vibration, frequency domain impact features, or short-term energy distribution features, and construct a continuous feature change trajectory.

[0143] Fitting the trend of feature changes: Apply the feature trend prediction function to fit the above historical feature trajectory to extract the dominant change direction and rate of change of the sequence; for example, whether it continues to rise, fall or fluctuates periodically, so as to obtain a trend line describing the evolution direction of the current state features.

[0144] Predicting future feature changes: Extrapolating the trend function by one or more time steps yields the predicted feature value for that node in the short term. These predicted values ​​are used to determine whether the feature will exceed the current baseline threshold boundary, or whether there is a risk of continued drift when approaching the warning range;

[0145] Construct a drift trend model: Compare the extrapolated trend results with the basic threshold set to determine the probability and extent of future feature values ​​deviating from the basic threshold. Based on information such as the magnitude, rate, and persistence of the deviation, construct a threshold drift trend model describing the node under the current operating conditions. This model does not directly adjust the threshold but provides a basis for judging whether the current threshold setting is about to become invalid.

[0146] Supports dynamic correction and feedback: This trend model continuously receives updated data, constantly refreshing the prediction results. When the drift trend is confirmed and continues to strengthen, it will trigger the threshold correction function model in subsequent steps to adjust the basic threshold and provide response compensation.

[0147] The steps for time series prediction of the initial feature set sequence are specifically implemented using the sliding window regression prediction algorithm, which includes:

[0148] Within each node group, for drift-sensitive nodes, a continuous sequence of initial state features is extracted within a fixed-length historical time window as the current sample window;

[0149] Within the sample window, construct regression training data including time indices and corresponding feature values, and fit the overall trend direction of current feature changes using the least squares method;

[0150] At the end of the window, based on the fitted trend of change, the feature values ​​for several future time steps are extrapolated in the short term to obtain the prediction results of the feature values ​​at future times.

[0151] The above prediction results are used to determine whether the features exhibit obvious upward, downward, or fluctuating patterns, thus serving as the input basis for subsequently constructing a threshold drift trend model.

[0152] S24, synchronously acquire historical data of operating parameters of the corresponding equipment in the drift node group within the same historical time window to form a set of operating parameters: ;

[0153] Common parameters include: load, speed, ambient temperature, lubrication status indicators such as oil temperature and oil pressure, and external disturbance signals such as shocks and starting frequency. All parameters are time-aligned, covering the same historical time window as the drift trend analysis.

[0154] Through correlation analysis, The coupling relationship between the node feature drift trend and the parameter subset that has the most significant impact on the drift is evaluated, denoted as:

[0155] ;

[0156] in, Represents the set of all operating condition parameter sequences. Represents the set of critical operating conditions. Indicates the first Each operating parameter at time The value of .

[0157] S25, Predicted value based on drift trend model With key operating condition factors set Joint modeling to construct a function model This refers to the drift response model, which dynamically outputs the correction amount of the base threshold for each edge node under specific operating conditions.

[0158] ;

[0159] in, This represents the amount of correction to the base threshold. This represents the drift response function model, i.e., at the node level. The output of this function model is the response to the nodes. The dynamic correction amount for a specific threshold within the base threshold set can be positively or negatively offset from the current set value, achieving personalized adaptive adjustment of the base threshold. Constructed using piecewise linear mapping functions or polynomial response functions, it possesses the following characteristics:

[0160] 1. When the predicted feature value is close to the normal threshold boundary and the operating conditions fluctuate drastically, output a positive correction amount, and the threshold is appropriately moved up or down; when the predicted feature value is stable or the operating conditions change slightly, output a correction amount close to zero; when the predicted value has exceeded the attention threshold and the trend is strengthening, output a larger correction amount to push the basic threshold into the transition zone ahead of time.

[0161] 2. Response parameter adjustment: Different slope factors and weighting coefficients can be set in the function to control the sensitivity of characteristic trends to the correction magnitude, as well as the proportion of operating conditions in the overall decision. The system allows optimization of function parameters through empirical tuning or historical learning.

[0162] 3. Function model deployment method: The function model can be generated uniformly at the central coordination node, or a lightweight version can be executed locally on the edge node, depending on computing resources and response time requirements.

[0163] S3: The drift response model is sent to each edge node to correct the threshold offset in the local state judgment process, and a fused state monitoring conclusion is generated based on the corrected judgment result; if the fused result triggers the abnormal standard, real-time early warning information is immediately output to the field system or upper-level platform.

[0164] S3 specifically includes:

[0165] S31: The central coordinating node encapsulates the drift response model into configuration instructions, which include:

[0166] The configuration instructions are: a set of model parameters used to describe the model structure and calculation path; calculation rules used to indicate the threshold correction method; and a working condition adaptation range for specifying the applicable scenarios of drift compensation. The configuration instructions are sent to each edge node in the corresponding node group through an industrial control bus or edge cluster communication protocol.

[0167] S32: The edge node receives and parses the configuration instructions, replaces and updates the drift response model stored locally.

[0168] S321: During each execution of the local status judgment process, the edge node collects the current key operating condition factor set in real time. The updated drift response model is then used as input to the model. :

[0169] ;

[0170] in, This represents the real-time adjustment amount for each threshold in the local base threshold set. This is the set of key operating condition factors currently collected. This section describes a multi-input single-output mapping function constructed based on key operating condition factors. This refers to a function model used to dynamically calculate the correction amounts of various thresholds based on the current set of key operating conditions, enabling adaptive drift compensation of the basic threshold set; this function model... The specific plan is as follows:

[0171] The model first uses a set of key operating condition factors. , As input, each This represents a real-time acquired environmental or operating condition parameter, such as load, current, speed, or ambient temperature. To accurately reflect the different weights of each factor's influence on changes in state characteristics, Typically, a machine learning regression model with parameter self-learning capabilities is used for modeling.

[0172] This invention preferably employs a combination of a sliding window regression prediction model and a linear weighted combination mechanism as its core. The specific approach involves training a set of regression weights or sub-models by maintaining a historical sliding window of data, including operating conditions and characteristic changes, locally at each node. This data is used to predict the impact trend of different combinations of operating conditions on a given basic threshold term. The correction amount for each basic threshold term is calculated independently by its corresponding sub-model.

[0173] To improve the real-time performance and interpretability of the model, the following mechanism can be introduced:

[0174] Standardization and normalization preprocessing: Normalize each input factor to make them comparable;

[0175] Factor importance screening: During the model training phase, principal component analysis is used to select factors, retaining only the most significant principal factors.

[0176] Dynamic weight update mechanism: Periodically re-evaluate the fit of the regression model and dynamically adjust factor weights or replace sub-model parameters.

[0177] final, Output for each basic threshold term Correction amount This data is then used to construct the corrected dynamic threshold set. This multi-input single-output mapping method ensures that the interactive effects between different operating conditions can be reasonably absorbed, thus making it more adaptable and robust in the state judgment process.

[0178] S322: Adjustment amount Acting on the original basic threshold set Generate a corrected dynamic threshold set It is used to update the discrimination criteria in the state judgment process.

[0179] S33: Edge nodes are based on the corrected dynamic threshold set. The initial feature set of the state collected and extracted at the current moment Perform a re-evaluation and regenerate the refined local state determination result. And the determination result and characteristic value Report to the central coordination node.

[0180] S34: The central coordination node summarizes the refined judgment results and feature value data uploaded by each edge node in the node group, and uses at least one of the following fusion logics to generate the fused state monitoring conclusion. :

[0181] Voting logic: Calculate the judgment results of each edge node and determine the fusion result based on the majority principle of frequency of occurrence;

[0182] Feature-weighted evaluation logic: Weights are set based on the relative magnitude or trend of each node's feature values ​​to make a weighted judgment on the state;

[0183] Topological reasoning logic: Combining the physical connections or functional dependencies between the devices where the nodes are located, it performs causal propagation analysis of potential faults to help locate and confirm state risks.

[0184] S35: The central coordination node will integrate post-state monitoring conclusions. The system compares the data with a pre-defined anomaly detection standard library. If an anomaly is detected, a warning or fault risk level is issued that exceeds a set threshold. The system generates a real-time update including the location of the edge node to which the anomaly belongs, the corresponding risk level identifier, a description of possible causes, and recommended handling measures. This update is then output to the on-site monitoring system or uploaded to the upper-level cloud management platform via industrial communication protocols to support timely intervention and maintenance decisions by operations and maintenance personnel.

[0185] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0186] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for real-time monitoring of bearing operating status combining edge computing and adaptive threshold judgment, characterized in that, Includes the following steps: S1: At edge nodes distributed at different monitoring points, the original operating signals of the corresponding bearings are collected respectively, and a lightweight feature extraction algorithm is executed locally to obtain an initial state feature set including vibration amplitude features, frequency domain impact features and short-time energy distribution features; based on the locally set basic threshold set, a preliminary state judgment label is generated. S2: Receive the initial state feature set and preliminary state judgment label of the output. The central coordination node analyzes the consistency of the states of multiple edge nodes, identifies the node groups with drift, establishes a threshold drift trend model based on the change trend within the time window, and predicts the future evolution direction. At the same time, it extracts the key working condition factors that cause drift, generates a set of working condition linkage factors, and adjusts the threshold correction strategy of each edge node accordingly to form a prediction-driven drift response model. S3: The drift response model is sent to each edge node to correct the threshold offset in the local state judgment process, and a fused state monitoring conclusion is generated based on the corrected judgment result. If the fusion result triggers an abnormal standard, a real-time early warning message will be immediately output to the on-site system or the upper-level platform.

2. The method for real-time monitoring of bearing operating status combining edge computing and adaptive threshold judgment according to claim 1, characterized in that, S1 specifically includes: S11: Each edge node periodically collects the vibration acceleration signal of the corresponding bearing at a fixed sampling frequency and duration, which serves as the original operating signal; S12: Execute the lightweight feature extraction algorithm on the original running signal locally to obtain an initial state feature set including vibration amplitude features, frequency domain impact features, and short-time energy distribution features.

3. The method for real-time monitoring of bearing operating status combining edge computing and adaptive threshold judgment according to claim 2, characterized in that, The lightweight feature extraction algorithm includes processing the original running signal after each edge node completes buffering, extracting vibration amplitude features, frequency domain impact features, and short-time energy distribution features, wherein: The effective value of the vibration signal is calculated as the vibration amplitude characteristic. The signal is envelope demodulated and the amplitude at the bearing fault characteristic frequency is analyzed as the frequency domain impact characteristic. Wavelet packet decomposition is performed on the signal, and the entropy of the proportion of energy in each sub-band to the total energy is calculated as a short-time energy distribution characteristic.

4. The method for real-time monitoring of bearing operating status combining edge computing and adaptive threshold judgment according to claim 2, characterized in that, Each feature in the initial state feature set is compared with the corresponding threshold range in the locally stored basic threshold set. Based on the preset multi-level judgment logic, the initial state judgment label is output as normal, attention, or warning.

5. The method for real-time monitoring of bearing operating status combining edge computing and adaptive threshold judgment according to claim 1, characterized in that, The central coordination node receives the initial state feature set and the preliminary state determination label from all edge nodes; S2 specifically includes: S21: Based on the physical subsystem or topological proximity of the bearings corresponding to each edge node, divide all edge nodes into one or more logical node groups; S22: For each node group, calculate the correlation coefficient of similar feature sequences in the initial state feature set of each edge node in the node group within the same time period, as the group synergy index; when the group synergy index is lower than the preset correlation threshold, determine that there is a node group with correlation threshold drift. S23: For the identified node group, extract the initial state feature set sequence of the identified node group within a preset historical time window, and use a time series prediction algorithm to fit and predict the initial state feature set sequence to establish a threshold drift trend model describing the overall drift direction and rate of features within the group. S24: Synchronously acquire historical operating condition parameters shared by the equipment or area where the node group is located, including load, speed and ambient temperature; through correlation analysis, select the subset of parameters that are most strongly correlated with the characteristic drift trend of the node group from the operating condition parameters to form a key operating condition factor set. S25: Based on the predicted values ​​of the characteristics of future periods by the threshold drift trend model, and the quantitative influence relationship of the key working condition factor set on the feature drift, construct a function model with real-time working condition factors as input and the correction amount of specific thresholds in the basic threshold set of each edge node as output, as the drift response model.

6. The method for real-time monitoring of bearing operating status combining edge computing and adaptive threshold judgment according to claim 5, characterized in that, The time series prediction uses a sliding window regression prediction algorithm, which specifically includes: extracting a continuous sequence of initial state features within a fixed-length historical time window as a sample window; constructing a regression model based on the feature change data within the sample window to fit the overall trend; and performing short-term extrapolation of the feature values ​​for future periods at the end of the window. The resulting prediction results are used to determine whether there is a trend drift in the features.

7. The method for real-time monitoring of bearing operating status combining edge computing and adaptive threshold judgment according to claim 5, characterized in that, The correlation coefficient includes: extracting the initial feature sequences of the same state of all edge nodes in the same node group within the same time period, calculating the Pearson correlation coefficient between feature sequences pair by pair to measure the consistency of feature change trends, and after completing the correlation calculation of all node pairs, averaging all correlation coefficients to generate the average correlation index of the node group under the feature dimension, which is used to describe the consistency of the overall feature fluctuation direction.

8. The method for real-time monitoring of bearing operating status combining edge computing and adaptive threshold judgment according to claim 1, characterized in that, S3 specifically includes: S31: The central coordination node encapsulates the drift response model into configuration instructions that include model parameters, calculation rules, and applicable operating conditions, and sends them to each edge node in the node group; S32: Each edge node receives and parses the configuration instruction, updates the drift response model stored locally; in each subsequent local state judgment process, it obtains the current set of key working condition factors in real time, inputs the updated drift response model, calculates the correction amount for each threshold in the local basic threshold set, and applies the correction amount to generate the corrected dynamic threshold set. S33: Each edge node uses the modified dynamic threshold set to re-judge the initial state feature set collected and extracted in real time, generates a local state judgment result, and reports the local state judgment result and the feature values ​​used to the central coordination node. S34: The central coordination node receives the local state determination results and feature values ​​of each edge node in the node group, and executes the fusion logic to generate the fused state monitoring conclusion; S35: Compare the fused status monitoring conclusion with the preset anomaly standard. If the conclusion indicates that there is an anomaly or the fault risk level exceeds the threshold, generate real-time early warning information.

9. The method for real-time monitoring of bearing operating status combining edge computing and adaptive threshold judgment according to claim 8, characterized in that, The fusion-based state monitoring conclusions include voting on the judgment results, weighted evaluation based on the feature values ​​of each node, and fault propagation reasoning based on the device topology.

10. A method for real-time monitoring of bearing operating status combining edge computing and adaptive threshold judgment according to claim 8, characterized in that, The real-time early warning information includes the location of the edge node to which the anomaly belongs, the corresponding risk level identifier, the cause description, and recommended handling measures.

Images