Intelligent monitoring method and system based on multi-source data

Abstract

The application provides an intelligent monitoring method and system based on multi-source data, relates to the technical field of equipment fault monitoring and diagnosis, and solves the technical problems that weak fault signals are difficult to extract, the false alarm rate of variable working conditions is high, historical fault samples are relied on, and coupled faults cannot be automatically separated and traced. The method comprises the following steps: collecting multi-measurement-point vibration and acoustic emission signals; calculating an abnormal diffusion index representing the thick-tail characteristics of the signal increment distribution; intercepting a multi-channel signal segment when the index exceeds a dynamic threshold; performing multi-scale decomposition on the signal segment and calculating conditional transfer entropy to construct a causal information tensor; obtaining causal components through tensor decomposition, and monitoring their generation and evolution trend to identify early faults. The application is used in the online early fault monitoring and early warning process of industrial equipment.

Description

Technical Field

[0001] This invention belongs to the field of equipment fault monitoring and diagnosis technology, specifically an intelligent monitoring method and system based on multi-source data. Background Technology

[0002] Equipment fault monitoring and diagnosis are crucial for ensuring the safe operation of industrial equipment and preventing unplanned downtime. As industrial production moves towards intelligent and unmanned operations, increasingly higher demands are being placed on the early detection, accuracy, and automation of equipment fault monitoring. Early-stage faults refer to a transitional phase between normal operation and a clearly malfunctioning state. These faults are extremely subtle and easily masked by system noise and operating condition fluctuations; therefore, reliable detection of early-stage faults has always been a technical challenge in this field.

[0003] Currently, commonly used equipment fault monitoring methods are mainly divided into three categories: signal analysis-based methods, mechanism model-based methods, and data-driven methods. Signal analysis-based methods typically extract time-domain, frequency-domain, or time-frequency-domain features from collected vibration, temperature, and other signals, such as calculating signal kurtosis, root mean square (RMS) values, and envelope spectra. An alarm is triggered when the feature values ​​exceed preset thresholds. However, the feature indicators relied upon by these methods have limited sensitivity to early, minor faults. When a fault is still in its nascent stage, the resulting signal energy change is minimal, and traditional time-domain and frequency-domain features often cannot be effectively identified in noisy environments. Mechanism model-based methods establish dynamic or finite element models of the equipment and compare the residuals between the model output and actual measurements to determine the fault. However, these methods have high modeling costs, poor adaptability, and are difficult to widely apply in complex industrial environments. Data-driven methods, especially deep learning methods, have developed rapidly in recent years. They learn failure modes by training on a large amount of historical data. However, such methods usually require a sufficient number of failure samples of various types. In actual industrial scenarios, early failure samples and rare failure samples are extremely scarce and severely imbalanced. At the same time, the generalization ability of the model is often insufficient when facing new equipment or new operating conditions.

[0004] Furthermore, existing fault monitoring methods have significant shortcomings when dealing with variable operating conditions. The operating conditions of industrial equipment, such as speed and load, frequently change. These changes themselves cause variations in signal characteristics. Most existing methods employ fixed thresholds or simple compensation strategies, which easily generate numerous false alarms under variable operating conditions. Alternatively, increasing the threshold to reduce false alarms may lead to missed alarms. Another prominent problem is that most existing methods target single fault sources for detection and diagnosis. When multiple concurrent or coupled faults exist, the signals generated by each fault superimpose and interfere with each other along their propagation paths. Traditional methods struggle to effectively separate and independently identify coupled faults, and even more so to trace the propagation paths and impact ranges of each fault, resulting in ambiguous fault location and unreliable diagnostic results.

[0005] In summary, existing technologies for early-stage equipment fault monitoring suffer from several technical problems, including difficulty in effectively extracting weak fault signals, difficulty in balancing false alarm and false negative rates under varying operating conditions, over-reliance on historical fault samples, and the inability to automatically separate coupled faults and trace their propagation paths. A new monitoring method that can overcome these shortcomings is urgently needed. Summary of the Invention

[0006] This application provides an intelligent monitoring method and system based on multi-source data, which solves the technical problems of existing technologies in early equipment fault monitoring, such as difficulty in effectively extracting weak fault signals, high false alarm rate under varying operating conditions, reliance on a large number of historical fault samples for training, and inability to automatically separate coupled faults and trace their propagation paths.

[0007] To achieve the above objectives, this application adopts the following technical solution: Firstly, it provides intelligent monitoring methods based on multi-source data, including: Vibration and acoustic emission signal data from several measuring points on the equipment are collected to obtain signal data for each measuring point. Time-varying anomaly diffusion spectrum analysis was performed on the signal data at each measurement point to calculate the anomaly diffusion index, which characterizes the thick-tailed characteristics of the signal increment distribution. When the abnormal diffusion index of any measuring point exceeds the dynamic threshold, the multi-channel signal segment containing all measuring points before and after the current moment is extracted. The multi-channel signal segment is decomposed into multiple scales, the conditional transfer entropy between channels at each scale is calculated, and a causal information tensor is constructed based on the conditional transfer entropy. Tensor decomposition is performed on the causal information tensor to obtain multiple causal components; Monitor the generation and temporal evolution trend of the causal components to identify early failure states of the equipment.

[0008] Based on the above technical solutions, the intelligent monitoring method based on multi-source data provided in this application uses the time-varying anomaly diffusion spectrum analysis to calculate the anomaly diffusion index as the trigger criterion for early anomalies. It utilizes the statistical physical characteristics of signal increment distribution rather than traditional waveform amplitude characteristics to perceive weak faults, enabling the system to have extremely high detection sensitivity to early weak impacts. By intercepting multi-channel signal segments when the anomaly diffusion index exceeds the limit and performing multi-scale decomposition and conditional transfer entropy calculation, the anomaly perception of a single channel is transformed into causal analysis of information flow between multiple channels, thereby revealing the propagation path and coupling relationship of the fault without the need for historical fault samples. By constructing the multi-scale conditional transfer entropy as a causal information tensor and performing tensor decomposition, unsupervised blind extraction of potential fault information flow patterns from high-dimensional information flow data is achieved, enabling multiple concurrent or coupled early faults to be automatically separated and independently tracked at the information flow level. By monitoring the weight generation and temporal evolution trend of causal components, and using the dynamic change of information flow intensity as a measure of the degree of fault development and deterioration, accurate determination of the early state is achieved.

[0009] Furthermore, the signal data from each measurement point undergoes time-varying anomaly diffusion spectrum analysis, specifically including: For each measurement point signal, a sliding window is used to extract a signal segment of a preset length, and the detrended increment sequence of the signal within the window is calculated; wherein, the detrended increment sequence refers to the sequence of signal increments calculated for different delay intervals after removing the trend term from the original signal; Based on the aforementioned trend increment sequence, construct a q-order structure function: By fitting Estimate the generalized Hurst exponent for each order q. ;in, For delay, For signal delay The incremental change This represents the time average, where q is the real number order. The Anomalous Diffusion Index (ADI) is calculated based on the generalized Hurst index, using the following formula: Where q takes values ​​in the range [-5, 5]. This represents the function for calculating standard deviation. It is the generalized Hurst exponent when q=2.

[0010] Furthermore, the calculation process of the detrended increment sequence includes: A trend line is obtained by performing polynomial fitting on the original signal segment within the sliding window, and the trend line is subtracted from the original signal segment to obtain the detrended signal. In the detrending signal, for each time point t and a given delay τ, the difference between the signal value at time point t and after the delay τ is calculated to obtain the detrending increment sequence.

[0011] Furthermore, the method for obtaining the dynamic threshold includes: During the confirmed normal operating period of the equipment, the abnormal diffusion index sequence of each measuring point is continuously calculated, the probability density distribution of the abnormal diffusion index sequence is fitted, and the static benchmark threshold is determined according to the preset false alarm rate; wherein, the false alarm rate represents the maximum permissible probability of misjudging the equipment status as an abnormal state. The operating parameters of the equipment are collected synchronously, and a mapping model is constructed with the operating parameters as input and the anomaly diffusion index as output; wherein, the operating parameters include, but are not limited to, speed, load and temperature; During actual monitoring, the mapping model is queried based on real-time operating parameters to obtain the predicted value of the anomaly diffusion index under real-time operating conditions. The difference between the predicted value and the predicted value of the abnormal diffusion index of the equipment under the baseline operating conditions is used as the index offset. The dynamic threshold is obtained by adding the offset to the static baseline threshold.

[0012] Furthermore, the step of performing multi-scale decomposition on the multi-channel signal segments and calculating the conditional transfer entropy between channels at each scale specifically includes: Noise-assisted multivariate empirical mode decomposition is performed on the multi-channel signal segment to obtain multiple scale-aligned and channel-comparable mode function components; At each scale component, for any two channels i and j, the conditional transfer entropy from channel i to channel j is calculated using the signals of the other channels at that scale as conditions; the conditional transfer entropy reflects the intensity of the directional information flow from channel i to channel j at that scale after excluding the influence of common driving variables.

[0013] Furthermore, the construction of the causal information tensor based on the conditional transitive entropy specifically includes: Using a sliding time window as the unit, the conditional transfer entropy at each scale within each time window is calculated; For any scale, the conditional transfer entropy between all pairs of channels is arranged into an M×M causal matrix; where M is the total number of measurement point channels; the measurement point channel refers to the signal acquisition channel corresponding to each independently arranged sensor; By sequentially stacking the causal matrices of each scale for T consecutive time windows along the time dimension, a four-dimensional tensor is formed, resulting in a causal information tensor. The four-dimensional tensor corresponds to the time window index, the sending channel index, the receiving channel index, and the scale index, respectively. The sending channel index and the receiving channel index identify the measurement point channel as the information outflow and the measurement point channel as the information inflow, respectively.

[0014] Furthermore, the tensor decomposition of the causal information tensor to obtain multiple causal components specifically includes: Using the CANDECOMP / PARAFAC decomposition method, the four-dimensional causal information tensor is approximated as a weighted linear combination of a finite number of rank-1 components. Each rank-1 component consists of the outer product of a time factor vector, a transmission channel factor vector, a reception channel factor vector, and a scale factor vector. The factor vectors are feature vectors extracted from each dimension by the tensor decomposition and are obtained through alternating least squares iterative updates. During the decomposition process, when a new time window slice is obtained, the existing rank-one component factor vector is recursively adjusted through an online incremental update algorithm; If the reconstruction error of the new time window slice exceeds a preset threshold, a new rank-one component is added and initialized to obtain a causal component; wherein, the rank-one component is used to characterize a fault causal propagation mode.

[0015] Furthermore, the monitoring of the generation and temporal evolution trend of the causal components specifically includes: The weight of each causal component is continuously recorded, and the time factor vector of each causal component is tracked element by element; wherein, the weight is obtained by the CANDECOMP / PARAFAC decomposition method; When the weight of a certain causal component exceeds the preset noise threshold, and each element in the time factor vector of the causal component shows a monotonically increasing trend over time, the causal component is determined to be a newly generated early fault causal pattern. For each existing early failure causal mode, a sliding window linear regression is performed on the time factor vector to extract the regression slope as an indicator of the development direction of the failure mode, and the cumulative change in the regression is extracted as an indicator of the cumulative influence intensity of the failure mode, thus obtaining a time evolution trend indicator.

[0016] Furthermore, the process for identifying the early fault state specifically includes: For the newly generated early fault causal pattern, the channel with the largest load in the transmitting channel factor vector is extracted as the fault influence source measurement point, the channel with the largest load in the receiving channel factor vector is extracted as the main propagation path, and the scale with the largest load in the scale factor vector is extracted as the fault feature frequency band. When two or more early fault causal modes exist simultaneously, it is determined that the device has a coupled fault. Based on the load distribution differences of each mode on the transmit channel factor, receive channel factor and scale factor, the coupled fault is separated, and the time evolution trend index of each separated fault mode is calculated. For any isolated fault mode, if the development direction index is greater than zero and the cumulative influence intensity index exceeds the preset limit, then the equipment component where the fault influence source measuring point corresponding to the fault mode is located is determined to have entered an early fault state. When the fault impact source measurement point is consistent with the measurement point where the abnormal diffusion index exceeds the limit, the fault impact source, main propagation path, fault characteristic frequency band and time evolution trend index are output to obtain early fault diagnosis results.

[0017] Secondly, this application provides an intelligent monitoring system based on multi-source data, including: an anomaly analysis module, an interception module, a causal relationship construction module, and a fault identification module; wherein, The anomaly analysis module is used to collect vibration and acoustic emission signal data from several measuring points on the equipment, obtain signal data from each measuring point, perform time-varying anomaly diffusion spectrum analysis on the signal data from each measuring point, and calculate the anomaly diffusion index that characterizes the thick-tailed characteristics of the signal increment distribution. The interception module is used to intercept a multi-channel signal segment containing all measurement points before and after the current moment when the abnormal diffusion index of any measurement point exceeds the dynamic threshold. The causal construction module is used to perform multi-scale decomposition on the multi-channel signal segment, calculate the conditional transit entropy between channels at each scale, and construct a causal information tensor based on the conditional transit entropy. The fault identification module is used to perform tensor decomposition on the causal information tensor to obtain multiple causal components, and monitor the generation and temporal evolution trend of the causal components to identify the early fault state of the equipment.

[0018] Compared with the prior art, the beneficial effects of this application are: This application utilizes time-varying anomaly diffusion spectrum analysis to detect early faults based on the statistical characteristics of signal increment distribution, rather than directly relying on traditional features such as signal amplitude or energy. This allows for a sensitive response even when the fault impact is extremely weak and insufficient to cause significant waveform distortion, significantly advancing fault detection time. The calculation of the detrended increment sequence eliminates the influence of slow baseline drift in the signal, enabling the anomaly diffusion index to more purely reflect the statistical regularity of transient impacts, thus improving the stability and anti-interference capability of early fault characterization. The dynamic threshold acquisition method establishes a mapping relationship between operating parameters and the anomaly diffusion index, allowing the alarm threshold to adaptively adjust with operating conditions such as speed and load. This effectively reduces false alarms caused by operating condition fluctuations and ensures the reliability of early fault triggering under complex and variable operating conditions.

[0019] This application, after the anomaly propagation index is triggered, utilizes noise-assisted multivariate empirical mode decomposition to perform multi-scale aligned decomposition of multi-channel signal segments and calculates the conditional propagation entropy at each scale, conditioned on the remaining channels. This allows for the extraction of the true directional information flow between channels, excluding interference from common driving sources, thereby accurately capturing the causal propagation relationship of faults between equipment components. Based on this, a causal information tensor is constructed, organizing the multi-channel information flow into a four-dimensional structure including time, transmitting channel, receiving channel, and scale. Online tensor decomposition dynamically extracts causal components representing different fault causal propagation modes, achieving unsupervised fault mode discovery even without historical fault samples. When multiple causal components exist simultaneously, this application can automatically separate and independently track different fault modes based on the differences in the loading of each component on the channel and scale factors, effectively solving the early diagnosis problem under multi-fault coupling conditions. Combined with a quantitative assessment of the temporal evolution trend of each causal component, this application can not only determine the occurrence of early faults but also provide the fault's impact source, propagation path, characteristic frequency band, and severity trend, providing a more comprehensive and operable decision-making basis for equipment maintenance. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This application provides a system architecture diagram of an intelligent monitoring system based on multi-source data, as illustrated in the embodiments of this application. Figure 2 A flowchart illustrating the intelligent monitoring method based on multi-source data provided in this application embodiment; Figure 3 A flowchart illustrating another intelligent monitoring method based on multi-source data provided in this application embodiment; Figure 4 A flowchart illustrating another intelligent monitoring method based on multi-source data provided in this application embodiment; Figure 5 This is a flowchart illustrating another intelligent monitoring method based on multi-source data provided in an embodiment of this application. Detailed Implementation

[0022] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0023] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0024] The intelligent monitoring method based on multi-source data provided in this application can be applied to, for example... Figure 1 In the intelligent monitoring system based on multi-source data shown, such as Figure 1 As shown, the system includes: an anomaly analysis module, an interception module, a cause-effect construction module, and a fault identification module; among which, The anomaly analysis module is used to collect vibration and acoustic emission signal data from several measuring points on the equipment, obtain signal data from each measuring point, perform time-varying anomaly diffusion spectrum analysis on the signal data from each measuring point, and calculate the anomaly diffusion index that characterizes the thick-tailed characteristics of the signal increment distribution. The interception module is used to intercept multi-channel signal segments containing all measurement points before and after the current moment when the abnormal diffusion index of any measurement point exceeds the dynamic threshold. The causal construction module is used to perform multi-scale decomposition of multi-channel signal segments, calculate the conditional transit entropy between channels at each scale, and construct a causal information tensor based on the conditional transit entropy. The fault identification module is used to perform tensor decomposition on the causal information tensor to obtain multiple causal components, and monitor the generation and temporal evolution trend of the causal components to identify the early fault state of the equipment.

[0025] To address the technical problems in existing technologies, such as difficulty in extracting weak fault signals, high false alarm rates under varying operating conditions, reliance on large amounts of historical fault samples for training, and the inability to automatically separate coupled faults and trace their propagation paths, embodiments of this application provide an intelligent monitoring method and system based on multi-source data. The method includes: Vibration and acoustic emission signal data from several measuring points on the equipment are collected to obtain signal data for each measuring point. Time-varying anomaly diffusion spectrum analysis was performed on the signal data at each measurement point to calculate the anomaly diffusion index, which characterizes the thick-tailed characteristics of the signal increment distribution. When the abnormal diffusion index of any measuring point exceeds the dynamic threshold, the multi-channel signal segment containing all measuring points before and after the current moment is extracted. Multi-scale decomposition of multi-channel signal segments is performed, the conditional transfer entropy between channels at each scale is calculated, and a causal information tensor is constructed based on the conditional transfer entropy. Tensor decomposition is performed on the causal information tensor to obtain multiple causal components; Monitor the generation and temporal evolution of causal components to identify early failure states of equipment.

[0026] Based on this, this application uses the statistical physical properties of the abnormal diffusion index to detect weak anomalies, triggers multi-channel causal information flow analysis, uses tensor decomposition to extract fault propagation patterns in an unsupervised manner and automatically separates coupled faults, thereby accurately identifying faults in the very early stages and providing the source of influence and propagation path, significantly reducing the false alarm rate under varying operating conditions and eliminating the dependence on historical fault samples.

[0027] like Figure 2 As shown in the embodiments of this application, the intelligent monitoring method based on multi-source data includes: S1. Perform time-varying anomaly diffusion spectrum analysis on the signal data at each measurement point, and calculate the anomaly diffusion index, which characterizes the thick-tailed characteristics of the signal increment distribution.

[0028] The signal data at each measuring point includes, but is not limited to, vibration and acoustic emission signal data from several measuring points on the equipment. Signal increment refers to the change in signal over a given time interval, and its distribution characteristics reflect the dynamic behavior of the system. Under normal conditions, the increment of the equipment vibration signal approximately follows a Gaussian distribution, i.e., without a thick tail. When an early fault generates a weak transient impact, the increment distribution deviates from the Gaussian shape, with a significantly thicker tail, indicating a significantly increased probability of large-amplitude impacts. The anomaly diffusion index is precisely the indicator that quantifies this thick tail; its magnitude directly reflects the strength of the non-stationary impact component in the signal and early signs of fault evolution.

[0029] In some implementations, time-varying anomaly diffusion spectrum analysis can be achieved by estimating the fractal characteristics or self-similarity parameters of the signal increment within a sliding time window. For example, scaling exponents can be obtained using detrended fluctuation analysis, or the degree of heavy tails can be measured by estimating the stable distribution characteristic exponent of the increment sequence. Alternatively, multifractal detrended fluctuation analysis can be used to obtain the generalized Hearst exponent spectrum, which can then be used to synthesize the anomaly diffusion index. Furthermore, methods such as recursive graphs or sorting pattern entropy can be used to capture changes in signal dynamics.

[0030] It should be noted that the abnormal diffusion index obtained in this step is not only sensitive to weak shocks, but also naturally immune to stationary random noise, because stationary noise always maintains Gaussian diffusion characteristics and will not cause significant drift of the index. This allows the index to work reliably even at low signal-to-noise ratios.

[0031] For example, the ratio of the fourth moment to the second moment of the incremental sequence can be estimated piecewise as a simple measure of the fat tail, or an online recursive algorithm can be introduced to update the exponent in real time to adapt to high-speed data flow scenarios.

[0032] S2. When the abnormal diffusion index of any measuring point exceeds the dynamic threshold, extract the multi-channel signal segment containing all measuring points before and after the current moment.

[0033] The multi-channel signal segment refers to a multi-dimensional time series segment composed of signals synchronously acquired by all measurement point channels within the same time period. Each row corresponds to one measurement point channel, and each column corresponds to the sampled value at a specific moment. This signal segment completely preserves the spatiotemporal correlation information of multi-source signals before and after the occurrence of fault symptoms, providing a complete data foundation for subsequent analysis of the causal transmission of faults among multiple components.

[0034] In some implementations, the dynamic threshold can be obtained by estimating the statistical distribution of the abnormal diffusion index under normal conditions online, for example, by fitting its probability density using kernel density estimation or a Gaussian mixture model, and determining the upper limit based on a given significance level. Alternatively, support vector regression or neural networks can be used to establish a mapping model from operating parameters to the abnormal diffusion index, compensating for deviations caused by changes in operating conditions in real time, thereby achieving adaptive threshold adjustment. Furthermore, the dynamic threshold can also employ sequential probability ratio tests or cumulative sum control chart methods to detect the index sequence online, quickly identifying persistent deviations.

[0035] It should be noted that the time length of the intercepted multi-channel signal segment should at least cover the time scale of the early fault impact propagation. A fixed length can be set according to the transmission characteristics of the equipment structure or experience, or it can be adaptively determined according to the duration of the abnormality.

[0036] For example, when the abnormal diffusion index of a certain measuring point exceeds the dynamic threshold for several consecutive calculation cycles, data segments of predetermined length are taken before and after the first time the limit is exceeded, and spliced ​​together to form a multi-channel signal segment to be analyzed.

[0037] S3. Perform multi-scale decomposition on the multi-channel signal segment, calculate the conditional transit entropy between channels at each scale, and construct the causal information tensor based on the conditional transit entropy.

[0038] Multi-scale decomposition unfolds multi-channel signals across multiple time scales to reveal oscillation modes at different frequencies or time scales. Commonly used methods include empirical mode decomposition, variational mode decomposition, and wavelet packet decomposition. Conditional Transfer Entropy (CTE) is an asymmetric metric based on information theory. It quantifies the degree to which the historical information of one channel reduces the uncertainty of the future state of another channel. During calculation, all other channels are used as conditions, eliminating spurious influences from common driving sources and indirect coupling, thus extracting the true directional information flow intensity between channels. The causal information tensor is a multidimensional array that organizes the conditional transfer entropy between channel pairs at multiple scales according to four dimensions: time, transmitting channel, receiving channel, and scale. It comprehensively characterizes the spatiotemporal multi-scale dynamic structure of the device information flow network.

[0039] In some implementations, multi-scale decomposition can also employ methods such as harmonic decomposition and adaptive scale separation. The calculation of conditional transfer entropy can be replaced by approximate entropy, symbolic transfer entropy, or kernel-based transfer entropy to reduce computational complexity. When constructing the causal information tensor, the time dimension can be calculated sequentially using a sliding window, with the window step size set according to dynamic requirements. In addition to using conditional transfer entropy, the elements of the causal matrix can also use nonlinear Granger causality exponents or directed transfer functions in the frequency domain, as long as they can reflect the directional information flow.

[0040] It should be noted that by introducing a condition set, the conditional transitivity in this application can effectively distinguish between direct causality and indirect association, avoiding the spurious connection problem of traditional transitivity in multivariate systems, thereby ensuring the accuracy of subsequent causal analysis.

[0041] For example, noise-assisted multivariate empirical mode decomposition can be applied to the multi-channel signal segment to obtain multiple scale-aligned intrinsic mode functions. Then, the transfer entropy conditioned on the remaining channels can be calculated on each scale component to form a causal matrix, which can then be stacked into a four-dimensional tensor in time order.

[0042] S4. Perform tensor decomposition on the causal information tensor to obtain multiple causal components, and monitor the generation and temporal evolution trend of the causal components to identify early fault states of the equipment.

[0043] Here, causal components refer to the structural components obtained by performing low-rank decomposition on the causal information tensor. Each component is represented by a set of factor vectors corresponding to time, transmission channel, reception channel, and scale, respectively. Their outer product approximately reconstructs the component's contribution to the tensor. Each causal component can be interpreted as an independent fault causal propagation mode, where the transmission channel factor indicates the source of the fault information flow, the reception channel factor indicates the affected path, the scale factor indicates the characteristic frequency band where the fault information is located, and the time factor describes the evolution process of the mode from non-existent to present and from weak to strong.

[0044] In some implementations, tensor decomposition can employ CANDECOMP / PARAFAC decomposition, or variants of Tucker decomposition, tensor singular value decomposition, or nonnegative tensor decomposition. Dynamic decomposition over time can be achieved through online incremental update algorithms to adapt to streaming data. Monitoring the generation of causal components can be achieved by detecting the emergence of new pattern factors after decomposition and their weights exceeding noise levels. Tracking the temporal evolution trend can be done using methods such as linear regression, Mann-Kendall trend test, or Bayesian change point detection, extracting slope and cumulative change as trend indicators.

[0045] It should be noted that by utilizing the low-rank sparsity property of tensor decomposition, this application can automatically mine potential fault propagation patterns from high-dimensional information flow data without any historical fault sample annotations, and separate multiple fault modes that coexist based on factor loading distribution, thus solving the confusion problem of traditional methods under multiple fault coupling.

[0046] For example, when the online CP decomposition detects a new component with a steadily increasing weight and a monotonically increasing time factor, it can be determined as a newly emerging early fault causal pattern. Then, the fault source is located based on the channel with the largest load in the transmission channel factor, and the early fault state is confirmed by combining the consistency with the measurement point that triggered the abnormal diffusion index to exceed the limit.

[0047] Based on the above technical solutions, the intelligent monitoring method based on multi-source data provided in this application realizes the triggering of weak faults in the background of strong noise through statistical physics-driven abnormal diffusion analysis, realizes unsupervised blind separation of coupled faults and propagation path tracing by using multi-channel causal information flow tensor decomposition, and comprehensively improves the sensitivity, accuracy and interpretability of early fault monitoring by combining it with working condition adaptive dynamic threshold.

[0048] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3 As shown, the above S1 can be implemented through the following S101, S102 and S103, which are explained in detail below: S101. Perform detrending processing on the signal at each measurement point to obtain a detrended signal, and construct a detrended incremental sequence based on the detrended signal.

[0049] Detrending processing refers to removing slowly changing trend components from the original signal, making the remaining signal better reflect rapidly changing dynamic characteristics. Trend components may originate from sensor drift, changes in ambient temperature, or slow adjustments in equipment operating status. These factors are unrelated to fault impacts and, if not removed, will interfere with subsequent statistical analysis. The detrended increment sequence is obtained by calculating the difference between the signal value at each time point and a given delay interval based on the detrended signal. This sequence reflects the distribution of the signal's amplitude changes at different time scales.

[0050] In some implementations, detrending can be achieved using polynomial fitting, where a trend line is determined using least squares and subtracted from the original signal. Specifically, for the original signal segment x(t) within the sliding window, a second-order polynomial can be used. After fitting, the detrended signal is Delay interval The selection of the frequency range and sampling rate can be set according to the characteristic frequency range of the device, typically covering multiple values ​​from short to long delays to capture incremental statistical characteristics at different time scales. Based on detrending signals... The constructed detrended increment sequence consists of increments at various delays. composition.

[0051] It should be noted that the construction of the detrended incremental sequence is the data foundation for the entire anomaly diffusion analysis, and its quality directly affects the accuracy of the subsequent generalized Hurst exponent estimation. Therefore, the selection of detrending methods should take into account both computational efficiency and the thoroughness of trend removal.

[0052] For example, a trend line is obtained by performing a second-order polynomial fitting on the original signal segment within each sliding window. The trend line is then subtracted from the original signal to obtain the detrended signal. Finally, the detrended signal is processed for each time point t and each preset delay. According to the formula Calculate the signal difference to form a detrended increment sequence. The delay value can be selected at intervals that are integer multiples of the sampling interval.

[0053] S102. Construct a multi-order structure function based on the detrended incremental sequence, and estimate the generalized Hurst exponent corresponding to each order by fitting the scaling relationship between the structure function and the delay.

[0054] The structure function is a statistical measure of the moments of each order in the detrended increment sequence. It is defined as the average of the absolute values ​​of the increments at different orders q, reflecting the growth pattern of the signal increments at different moment orders.

[0055] Specifically, for a given order q and delay Structure function It can be represented as ,in This represents averaging over time t. When the signal exhibits self-similarity, the structure function and the delay satisfy a power-law relationship. , where H(q) is the generalized Hurst exponent.

[0056] The generalized Hurst exponent is a set of parameters that describes the self-similarity and long-range dependence of a signal. When H(q) varies with the order q, it indicates that the signal has multifractal characteristics, that is, fluctuations of different amplitudes follow different scaling rules. When the equipment is running normally, the generalized Hurst exponent remains constant and close to 0.5 throughout the entire order range, exhibiting the characteristics of a single fractal Brownian motion; when a fault occurs, the scaling behavior of large increments changes, and the generalized Hurst exponent becomes dispersed, especially the second-order exponent H(2) deviates from 0.5 and the fluctuations of exponents of all orders increase.

[0057] In some implementations, the structure function can be constructed by using the expected value of the squared increment instead of the absolute value, or by introducing weighting coefficients to give different considerations to near and far delays. Besides least-squares fitting, the estimation of the generalized Hurst exponent can also utilize maximum likelihood estimation or Bayesian estimation to improve the estimation accuracy for small samples. Furthermore, multifractal analysis can be directly performed on the increment sequence to obtain the complete singularity spectrum, from which the generalized Hurst exponent can be extracted.

[0058] It should be noted that the order range of the structure function should be wide enough to fully characterize the heavy-tailed property, typically covering both negative and positive orders. Negative orders are mainly affected by small increments, while positive orders are mainly affected by large increment shocks. Only by comprehensively considering the generalized Hurst exponents of all orders can the anomalous diffusion changes caused by weak shocks be fully captured.

[0059] For example, construct a structure function of order q in the range [-5, 5], for each delay Calculate the average of the absolute values ​​of the increments at the corresponding order, and then fit the linear relationship between the structure function and the delay at each order in a double logarithmic coordinate system. The slope qH(q) divided by q gives the generalized Hurst exponent H(q) of that order.

[0060] S103. Calculate the anomaly diffusion index based on the generalized Hurst exponent, and use it as a quantitative index to characterize the heavy-tailed characteristics of the signal increment distribution.

[0061] Among them, the abnormal diffusion index is a scalar indicator that combines the degree of deviation of the generalized Hurst exponent from Gaussian diffusion and the degree of dispersion of exponents of various orders.

[0062] Specifically, when the signal follows Gaussian diffusion, the generalized Hurst exponent always takes the value of 0.5 at the second order, and the exponents of all orders remain consistent. When a fault impact causes a heavy tail in the incremental distribution, the second-order generalized Hurst exponent will deviate from 0.5, and the differences between the exponents of all orders will increase. Based on this, the anomaly diffusion index (ADI) can be calculated using the following formula: Where H(2) is the generalized Hurst exponent when order q=2, H(q) represents the standard deviation of H(q) as q varies within its range.

[0063] An increase in the Anomalous Diffusion Index (ADI) directly reflects the appearance and enhancement of non-stationary impulse components in the signal, and its time-series evolution trend can be used for online early warning of faults. This index is naturally immune to stationary noise because the incremental statistical characteristics of stationary noise do not change over time; the ADI will remain near a stable baseline. It only produces a significant response when intermittent, large-amplitude impulses caused by faults occur, which allows the index to maintain reliable detection capabilities even under extremely low signal-to-noise ratio conditions.

[0064] In some implementations, the anomaly diffusion index can be calculated using only the generalized Hurst exponent of positive or even order, or different weights can be assigned to different orders to highlight the orders sensitive to specific fault types. For scenarios with high computational efficiency requirements, only a few representative orders can be selected for simplified calculation.

[0065] For example, take each order The generalized Hurst exponent is used to calculate the absolute deviation of H(2) from 0.5, and the standard deviation of H(q) at all orders is added to obtain the abnormal diffusion index value at that moment. The abnormal diffusion index sequence is output in chronological order for subsequent threshold comparison and trigger judgment.

[0066] Based on the above technical solution, by constructing the detrended incremental sequence, scaling analysis of the structure function, and comprehensive calculation of the anomaly diffusion index (ADI), this application transforms the problem of early fault perception from traditional waveform amplitude detection to statistical physical characteristic analysis of signal increment distribution, fundamentally improving the sensitivity to weak fault impacts and the ability to resist noise interference, and providing an accurate and reliable trigger source for subsequent multi-channel causal analysis.

[0067] In one possible implementation of this application embodiment, the above-mentioned S2 can be specifically implemented by the following S201, S202 and S203, which are described in detail below: S201, Obtain the dynamic threshold.

[0068] The dynamic threshold is a criterion used to determine whether the abnormal diffusion index is abnormal. Unlike a fixed threshold, it can be adjusted in real time according to the current operating conditions of the equipment to reduce the impact of operating condition fluctuations on the monitoring results. Equipment operating conditions mainly include parameters such as speed, load, and temperature. Changes in these parameters directly affect the statistical characteristics of the vibration signal, causing the abnormal diffusion index to fluctuate even within the normal range. The core idea of ​​the dynamic threshold is to decouple operating condition factors from the fault judgment logic, allowing the threshold to always fluctuate with changes in normal operating conditions.

[0069] In some implementations, obtaining the dynamic threshold first requires continuously calculating the anomaly diffusion index sequence for each measuring point during the confirmed normal operating period of the equipment, and then fitting the probability density distribution of this sequence. The probability density distribution fitting can employ parametric methods such as Gaussian distribution fitting, or non-parametric methods such as kernel density estimation. After completing the distribution fitting, a confidence upper limit is determined based on a preset false alarm rate as the static baseline threshold. The false alarm rate represents the maximum permissible probability of misclassifying a normal state as an abnormal state, and is typically set according to the tolerance for false alarms in the industrial environment, for example, 1 / 100 or 1 / 1000. A value of 1 / 100 means that when the system monitors 100 normal equipment operating states, at most one instance is allowed to be incorrectly classified as an abnormal state, representing a relatively moderate false alarm tolerance.

[0070] Furthermore, operating parameters of the equipment during normal operation are collected synchronously to establish a mapping model between these parameters and the anomaly diffusion index. This mapping model is trained using operating parameters such as speed, load, and temperature as inputs and the anomaly diffusion index at the corresponding time point as output. It can be constructed using methods such as multiple linear regression, support vector regression, or shallow neural networks. The training data for the mapping model should cover the typical operating conditions of the equipment during normal operation to ensure the model's generalization ability under varying operating conditions.

[0071] In actual monitoring, real-time collected operating parameters are input into the mapping model to obtain the predicted value of the anomaly diffusion index under real-time operating conditions. Simultaneously, the baseline predicted value output by the mapping model under the baseline operating conditions is calculated, and the difference between the two is used as the index offset. Adding the index offset to the static baseline threshold yields the dynamic threshold that adaptively adjusts with the operating conditions. The baseline operating condition can be selected from the most common operating condition or the rated operating condition of the equipment, and the corresponding output value of the mapping model is the baseline predicted value.

[0072] It should be noted that the use of dynamic thresholds effectively avoids the abnormal increase in diffusion index caused by normal operations such as device acceleration and loading being misjudged as a fault, while maintaining sensitive detection capability when a fault actually occurs, significantly balancing the contradiction between detection sensitivity and false alarm rate.

[0073] For example, during normal equipment operation, assume the rotational speed fluctuates between 900 rpm and 1500 rpm, the load varies between 30% and 100% of the rated load, and the temperature varies between 25°C and 65°C. Using rotational speed and load as input variables and the anomaly diffusion index at the corresponding time point as the output variable, a support vector regression mapping model is constructed using several sets of normal samples. Next, the equipment's rated operating condition is selected as the baseline operating condition, i.e., a rotational speed of 1500 rpm and a load of 100%. The baseline operating condition parameters are input into the trained support vector regression mapping model, yielding a baseline predicted value of 0.50. In actual monitoring, assume the real-time operating condition parameters collected at a certain moment are a rotational speed of 1200 rpm and a load of 60%. Inputting this set of operating condition parameters into the mapping model yields a predicted value of 0.46 for the anomaly diffusion index under the real-time operating condition. The index offset is calculated as the real-time predicted value of 0.46 minus the baseline predicted value of 0.50, resulting in an offset of -0.04. Adding the offset of -0.04 to the static baseline threshold of 0.72 yields a dynamic threshold of 0.68 for the current moment. If at another moment the equipment speed increases to 1500 rpm and the load increases to 95%, the real-time predicted value output by the mapping model is 0.52, with an offset of +0.02 from the baseline predicted value of 0.50. In this case, the dynamic threshold is adjusted to 0.72 plus 0.02, which equals 0.74.

[0074] S202. Compare the abnormal diffusion index of each measuring point with the dynamic threshold. When the abnormal diffusion index of any measuring point exceeds the dynamic threshold, it is determined that an abnormal diffusion event has occurred.

[0075] An abnormal diffusion event refers to a significant deviation of the incremental distribution characteristics of a signal at a certain measuring point from the normal state, indicating that the component at that measuring point may experience an early failure. Because the abnormal diffusion index is extremely sensitive to weak impacts and can respond at the fault initiation stage, the trigger time of an abnormal diffusion event is usually much earlier than the alarm time of traditional amplitude or energy indicators.

[0076] In some implementations, to avoid false triggering due to momentary interference, the abnormal diffusion index may need to exceed the dynamic threshold for several consecutive calculation cycles before being considered an abnormal diffusion event. The number of consecutive exceedances can be determined by a trade-off between the device's dynamic characteristics and response speed requirements, typically ranging from 3 to 5 cycles. Furthermore, the rate of increase of the abnormal diffusion index can be used as an auxiliary indicator; when the index not only exceeds the limit but also its rate of increase exceeds a certain slope, the likelihood of a fault is higher.

[0077] It should be noted that the abnormal diffusion event is a key bridge connecting single-channel anomaly perception and multi-channel causal analysis. It provides a timing signal for extracting key analysis segments from massive real-time monitoring data, so that subsequent computationally intensive multi-channel causal analysis is only activated when there are abnormal signs, which can save overall computing resources.

[0078] S203. Taking the trigger time of the abnormal diffusion event as the center, extract the multi-channel signal segment containing all measurement points before and after the current time.

[0079] The multi-channel signal segment refers to a multi-dimensional time series segment composed of signals synchronously acquired by all measuring point channels within the same time period. This signal segment is organized in matrix form, with each row corresponding to the sampling sequence of one measuring point channel and each column corresponding to the synchronous sampling value at a given moment, completely preserving the spatiotemporal correlation information of multi-source signals before and after the occurrence of fault symptoms. Extracting a signal segment within a certain time range before and after the current moment ensures that it includes both the evolution process before and after the initial occurrence of the fault impact and the dynamic process of fault information propagating from the source to other channels through the transmission path, providing a complete data foundation for subsequent analysis of the causal transmission of the fault among multiple components.

[0080] Specifically, let M be the total number of measurement point channels arranged on the equipment, N be the number of sampling points for the intercepted signal segment, and the sampling period be... The corresponding sampling frequency ,and Not lower than 25.6kHz. The zero point is set at the moment the anomalous diffusion event is triggered. , cut to Synchronous sampling data within a time range can be represented as a two-dimensional matrix: The i-th row of the matrix corresponds to the sampling sequence of the i-th measurement point channel, which can be denoted as a row vector. The j-th column corresponds to the synchronous sampled values ​​of all channels at the j-th sampling time, and can be denoted as a column vector. .

[0081] element This represents the amplitude of the vibration or acoustic emission signal acquired by the i-th measuring point channel at the j-th sampling time, and its corresponding timestamp is: , .

[0082] In some implementations, the intercepted time length can be estimated based on the equipment's geometry, material properties, and vibration propagation speed, or a fixed length can be set empirically, such as taking data from 0.5 to 2 seconds before and after the trigger moment. For rotating machinery, the intercepted length can correspond to several rotation cycles to ensure that the complete impact propagation process is included. The interception method can employ hardware buffer backtracking or maintain a circular buffer in software, locking the start and end positions of the data segment through a trigger signal.

[0083] Based on the above technical solutions, by constructing a dynamic threshold that adapts to operating conditions, triggering abnormal diffusion events based on continuous over-limit triggering of the abnormal diffusion index, and extracting multi-channel signal segments containing synchronization signals from all measuring points, this application achieves a precise transition from single-channel abnormal perception to multi-channel joint analysis. This ensures the sensitivity of early triggering of weak faults while avoiding false triggering due to changing operating conditions and high computational overhead throughout the entire period, laying a high-quality data foundation for subsequent causal information flow extraction.

[0084] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the above S3 can be implemented through the following S301, S302 and S303, which are explained in detail below: S301. Perform multi-scale decomposition on the intercepted multi-channel signal segments to obtain multiple scale-aligned and channel-comparable modal function components.

[0085] Multi-scale decomposition refers to unfolding multi-channel signals across multiple time scales to reveal oscillation modes at different frequencies or time scales. The impact signals generated by equipment failures are often not of a single frequency but are distributed across a wide frequency band, and different fault types and propagation paths may be prominent at different characteristic scales. Through multi-scale decomposition, the original broadband signal can be separated into several narrowband mode function components, each corresponding to a dominant oscillation scale, thus providing the conditions for subsequent subscale analysis of the fault information flow.

[0086] In some implementations, multi-scale decomposition employs Noise-Assisted Multivariate Empirical Mode Decomposition (NA-MEMD). This method, based on multivariate empirical mode decomposition, superimposes several independent Gaussian white noise channels onto the original multi-channel signal. Utilizing the statistical characteristics of the uniform distribution of white noise in the time-frequency space, it guides the decomposition process to adaptively separate the signal according to local oscillation scales, effectively suppressing mode aliasing.

[0087] The specific operation process of NA-MEMD is as follows: Suppose the extracted multi-channel signal segment contains M channels, and the signal length of each channel is N sampling points, denoted as... , where t=1,2,...,N.

[0088] First, generate L independent Gaussian white noise signals with L channels. The noise power is typically set to 5% to 10% of the original signal power, and L is generally a number equivalent to M, for example, L=M. The original M channels and the L noise channels are then combined to form a multivariable signal with dimensions P=M+L. Performing multivariable empirical mode decomposition on a P-dimensional multivariable signal Y(t) involves the following steps: Step a: Initialize the residual signal R(t) = Y(t); Step b: Generate Q groups of direction vectors, each group of direction vectors being a random point on a P-dimensional unit sphere, used to project multivariable signals onto different directions; Step c: For each set of direction vectors, project the residual signal onto that direction to obtain a one-dimensional projection signal, extract the local maxima and local minima of the projection signal, calculate the mean of the upper and lower envelopes, and average the mean over all projection directions to obtain the P-dimensional multivariate local mean M(t). Step d: Subtract the local mean from the residual signal to obtain the candidate mode function D(t) = R(t) - M(t); Step e: Determine whether D(t) satisfies the stopping criterion of the intrinsic mode function, usually based on the local mean approaching zero or the number of extreme points being equal to the number of zero-crossing points. If satisfied, output D(t) as a mode function component and let R(t) = R(t) - D(t), then return to step b to continue extracting the next layer of mode functions. If not satisfied, return to step c as the new residual signal to continue the filtering iteration. Step f: Repeat steps b to e until the residual signal R(t) becomes a monotonic function or a constant, at which point the decomposition ends.

[0089] After the above decomposition, K P-dimensional intrinsic mode function components are obtained, each component is denoted as . k=1,2,...,K, each The signal contains P channels. Components corresponding to L noise channels are discarded, retaining the components corresponding to the original M channels, thus obtaining the modal function components of the M measurement point channels at K scales. Since all channels share a unified local mean estimate during the decomposition process, the components of each channel in the same IMF layer are naturally aligned in scale. This means that the instantaneous frequency and bandwidth of the IMF components of each channel in the same IMF layer are basically consistent, providing comparability assurance for subsequent cross-channel information flow analysis.

[0090] S302. At each scale component, calculate the conditional transfer entropy between any two channels with the other channels as conditions, and obtain the directional information flow intensity between channels at each scale.

[0091] Conditional Transfer Entropy (CTE) is an asymmetric statistic based on information theory. It measures the degree to which the historical information of one variable reduces the uncertainty of the future state of another variable. In the calculation process, other variables are used as conditions to eliminate the spurious effects brought about by common driving sources and indirect coupling.

[0092] For a modal function component at a fixed scale k, let the time series of channel i be a(t), the time series of channel j be b(t), and the time series of the remaining M-2 channels form the conditional vector sequence Z(t). The conditional transitive entropy from channel i to channel j is defined as: Where I represents mutual information, b(t+1) represents the state of channel j at time t+1, and the superscript (d) represents the historical state vector of length d.

[0093] Specifically, the conditional transitive entropy based on historical states can be expressed as the Shannon entropy form of the joint probability distribution: Where H represents conditional entropy. The first term represents the uncertainty of channel j at the next moment when its own history and the history of the condition set are known; the second term represents the remaining uncertainty of channel j at the next moment after the history of channel i is further known. The difference between the two is the additional predictive power of the historical information of channel i for the future state of channel j, that is, the conditional transfer entropy value from channel i to channel j. The larger the value, the stronger the directional information flow from channel i to channel j, meaning that the past state of channel i has a significant predictive ability for the future state of channel j.

[0094] In some implementations, calculating the conditional transitive entropy first requires reconstructing the state space of the continuous signal. For the time series of each channel, the embedding theorem is used to map it to a high-dimensional phase space to capture the dynamic information of the system. (Historical state vector) It consists of the sampled values ​​from the current time and the previous d-1 time steps: The embedding dimension *d* and embedding delay can be determined using mutual information to determine the optimal embedding delay, and the minimum embedding dimension using the pseudo-nearest neighbor method. (Condition set historical state vector) It consists of the cascaded historical states of each condition channel.

[0095] For estimating discrete probability distributions, since conditional transitive entropy involves high-dimensional joint probability distributions, directly using the histogram method will face the curse of dimensionality. This embodiment uses the k-nearest neighbor method for probability density estimation. This method, proposed by Kraskov, Stögbauer, and Grassberger, is abbreviated as the KSG estimator. The core idea of ​​the KSG estimator is to adaptively adjust the estimation range using the local neighborhood distances of sample points in the joint space. The specific calculation process is as follows: For a given sample set, b(t+1), , and The sample combination is a point in a high-dimensional joint space. For each sample point i, find its ε-th nearest neighbor in the joint space and record the distance of that nearest neighbor. Then, in two conditional spaces (containing and not containing)... The KSG estimator counts the number of neighboring points that fall within the corresponding distance range, and uses the asymptotic property of the gamma function to estimate the conditional entropy difference, thus obtaining an estimate of the conditional transitive entropy. The KSG estimator can achieve good estimation accuracy when the data volume is small or the distribution is irregular, and it does not require a preset number of bins, making it a practical choice for industrial applications.

[0096] It is important to note that the introduction of condition sets is the key difference between this step and traditional transfer entropy calculation. In complex multi-component systems, multiple channels may share a common driving source; for example, the vibration of the entire equipment's foundation may simultaneously affect multiple measuring points. Without using other channels as conditions, directly calculating the transfer entropy between pairs of channels would generate numerous spurious causal connections, misclassifying common drivers as direct causality. Conditional transfer entropy, by incorporating the historical information of all other channels into the known conditions, effectively eliminates this common-cause effect, ensuring that the extracted directional information flow reflects the true causal transfer relationship. Furthermore, the asymmetry of conditional transfer entropy naturally provides directional information about the information flow. and They are usually not equal, which can be used to distinguish the source and sink of fault information.

[0097] For example, for the M modal function components at the k-th scale, the optimal embedding dimension and embedding delay of each channel are first determined using the mutual information method and the pseudo nearest neighbor method, completing the state space reconstruction and obtaining the historical state vector of each channel. Then, for each pair of ordered channels i and j, the signals of the remaining M-2 channels at this scale are used to form a condition set, and the KSG estimator is used to calculate... The k-nearest neighbor parameter ε is typically taken as 10 to 20. By traversing all ordered channel pairs (a total of M×(M-1) combinations), an M×M conditional transfer entropy matrix at this scale is obtained, with matrix elements... This represents the directional information flow intensity from channel i to channel j. Diagonal elements are set to zero or left undefined.

[0098] S303. Stack causal matrices of various scales for multiple consecutive time windows along the time dimension to construct a four-dimensional causal information tensor.

[0099] The causality matrix, at a fixed scale and time window, is an M×M matrix formed by arranging the conditional transit entropies between all pairs of channels. For the t-th time window and the s-th scale, the causality matrix is ​​denoted as... Its element in the i-th row and j-th column is , which represents the intensity of the directional information flow from channel i to channel j within this time window and at this scale.

[0100] The causal information tensor is a four-dimensional data structure formed by stacking the causal matrices of all scales across multiple consecutive time windows along the time dimension. Let the time window index be t, ranging from 1 to T; the transmitting channel index be i, ranging from 1 to M; the receiving channel index be j, ranging from 1 to M; and the scale index be s, ranging from 1 to S. Then the causal information tensor is denoted as: Its elements This represents the conditional transfer entropy value from channel i to channel j at the t-th time window and the s-th scale. This tensor structure fully and dynamically records the spatiotemporal multi-scale evolution process of the device's multi-channel information flow network, including both snapshots of the information flow topology at various scales at a certain moment and the continuous change trajectory of the topology over time.

[0101] It should be noted that the construction of the four-dimensional causal information tensor integrates the originally scattered multi-scale, multi-channel, and multi-time-point conditional propagation entropy information into a holistic tensor object. This allows the analysis of fault propagation patterns to no longer be limited to a single time segment or a single scale, but to be globally optimized and extracted within a spatiotemporal joint framework. This multi-dimensional structured processing method also naturally supports online incremental updates, adapting to the needs of long-term uninterrupted monitoring.

[0102] For example, for each sliding time window, the causal matrix at S scales is obtained according to methods S301 and S302, with each causal matrix having a dimension of M×M. After accumulating T consecutive time windows, all causal matrices are arranged in the order of time window index t, transmission channel index i, reception channel index j, and scale index s, constructing a four-dimensional tensor of dimension T×M×M×S, which is the causal information tensor T. As monitoring continues, a new causal matrix is ​​generated for each new time window. This is added to the time dimension of the tensor, and the oldest causal matrix is ​​removed according to the sliding window strategy, so that the tensor always maintains a size of T time windows for subsequent online tensor decomposition.

[0103] Based on the above technical solution, frequency band separation and inter-channel scale alignment of the signal are achieved through noise-assisted multivariate empirical mode decomposition. The true directional information flow between channels is extracted by conditional transfer entropy calculation with the other channels as conditions, under the premise of eliminating common cause interference. By constructing a four-dimensional causal information tensor, the scattered information flow data is integrated into a spatiotemporal multi-scale structure. This application realizes the system transformation from the original multi-channel signal to the structured fault information flow representation, laying a solid data foundation for subsequent tensor decomposition and unsupervised fault mode discovery.

[0104] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 5 As shown, the above S4 specifically includes the following S401 to S403: S401. Perform tensor decomposition on the causal information tensor to obtain multiple causal components.

[0105] Among them, the causal information tensor is a four-dimensional tensor constructed in step S303, denoted as... The purpose of tensor decomposition is to extract potential low-dimensional factors from this high-dimensional data structure, decomposing the complex global information flow network into several simple components, each corresponding to an independent fault causal propagation pattern.

[0106] In some implementations, tensor decomposition employs the nonnegative CANDECOMP / PARAFAC decomposition method, abbreviated as CP decomposition. CP decomposition approximates the four-dimensional causal information tensor as a weighted linear combination of R rank-1 components, and its mathematical expression is: ; Where R is a positive integer, representing the total number of components obtained from the decomposition, which can be automatically determined or preset based on data characteristics; Let be the weight of the r-th component, representing the strength of that component's contribution to the overall tensor; symbol This represents the outer product operation of vectors; Let r be the time factor vector of the r-th component, whose element values ​​reflect the change in information flow intensity of the fault mode in different time windows; Let be the transmission channel factor vector of the r-th component, whose element values ​​characterize the degree of contribution of each channel as a source of fault information transmission; Let be the receiving channel factor vector of the r-th component, whose element values ​​characterize the degree to which each channel is affected as a receiver of fault information; Let be the scaling factor vector of the r-th component, whose element values ​​indicate the distribution of the fault mode at different characteristic time scales. All factor vectors are subject to non-negativity constraints, i.e. , , , This is to ensure the physical interpretability of the decomposition results.

[0107] Solving the CP decomposition problem is usually transformed into an optimization problem, with the objective function being to minimize the reconstruction error between the original tensor and the reconstructed tensor, and the square of the Frobenius norm being used as the loss metric. ; in This represents the Frobenius norm, which is the square root of the sum of the squares of all elements in the tensor. This optimization problem is solved iteratively using the Alternating Least Squares (ALS) method. The specific operation process of ALS is as follows: Initialize all factor vectors to small positive random numbers; in each iteration, fix the factor vectors of the other three dimensions, optimize only the factor vector of one dimension, and update them in turn. To update For example, the tensor is expanded into a matrix along the time dimension, using known... , and Construct the design matrix and update it by solving a linear least squares problem under non-negativity constraints. And will update Negative elements are set to zero to satisfy the non-negativity constraint. Similarly, updates are performed sequentially. , and ; Repeat the alternating update process described above until the relative change in reconstruction error is less than a preset convergence threshold, such as one ten-thousandth, or until the maximum number of iterations is reached. The number of components R can be determined using a core consistency diagnostic method or by selecting an inflection point value that significantly slows down the reduction of reconstruction error through cross-validation.

[0108] During continuous monitoring, as the causal matrix of new time windows is continuously generated and added to the causal information tensor, an online incremental update strategy is required to adapt to the streaming data scenario. Specifically, after the causal matrix of the T+1th time window is generated, exponential smoothing prediction can be used first, utilizing existing factor vectors. The initial values ​​of the factor elements corresponding to the new time window are extrapolated. Then, the causal matrix slices of the new time window are locally iteratively optimized, and the elements in each factor vector affected by the new data are fine-tuned while the remaining elements remain basically unchanged to reduce the computational cost of online updates. If the reconstruction error exceeds the preset error threshold after adding the new time window, it is determined that a new information flow pattern has appeared that cannot be represented by the existing components. At this time, a new rank-one component R is added and incremented by 1, and the factor vector of the new component is initialized with the CP decomposition result based on the new data slice, thereby dynamically obtaining multiple causal components that correspond one-to-one with different fault causal propagation patterns.

[0109] It should be noted that CP decomposition compresses the complex four-dimensional information flow tensor into several sets of one-dimensional factor vectors. Each rank-1 component independently characterizes a complete fault propagation pattern, including the temporal evolution of fault impact, spatial propagation path, and frequency band characteristics. This low-rank representation greatly simplifies subsequent pattern monitoring and diagnostic reasoning, while the non-negativity constraint ensures the interpretable physical meaning of the decomposition results.

[0110] For example, a causal information tensor of dimension T × M × M × S is input into the CP decomposition module, with the initial number of components R set to 2 or 3. The ALS algorithm is used iteratively to solve the problem, and the convergence threshold is set to 0.01%. During online operation, for each new time window slice of the causal matrix, an incremental update algorithm is used to adjust the existing factor vectors and determine whether new components are needed. After decomposition, the weights of the R rank-1 components are output. and the corresponding factor vector , , and Each component constitutes an independent causal component.

[0111] S402. Monitor the generation and temporal evolution trend of causal components, determine newly generated early failure causal patterns, and track the development trend of existing patterns.

[0112] Among them, the monitoring of the generation of causal components refers to the continuous detection of whether new components appear and remain stable in the decomposition results. When the weight of a certain causal component... When a component is introduced from scratch—meaning it was not present in the decomposition results of the previous time period, is introduced for the first time in the current time period, and its weight exceeds a preset noise base threshold—and simultaneously, the time factor vector of this component is... If each element in the table shows a monotonically increasing trend with the window index, then the component is determined to be a newly generated early fault causal pattern. The noise floor threshold can be set based on the statistical fluctuation range of the weights of existing components under normal conditions, and is usually taken as several times the standard deviation of the mean of the normal component weights.

[0113] Monitoring the temporal evolution trend is achieved by analyzing the time factor vector of each existing early failure causal pattern. This is achieved through quantitative analysis. For the time factor vector of the r-th component... Its elements This represents the relative intensity of the fault mode information flow in the t-th time window. Perform sliding window linear regression with the time window index t as the independent variable. Establish a linear model with the dependent variable as the dependent variable: ;in The intercept is... The regression slope, The residuals. Regression slope. The following is estimated using the least squares method: , The mean of the time window index. The mean of the time factor vector. Regression slope. As an indicator of the development direction of this failure mode: if A value greater than zero and statistically significant indicates that the information flow intensity of this failure mode continues to increase over time, and the failure is developing and worsening; if Approaching zero indicates that the fault is in a stable state; if A value less than zero indicates a weakening of the fault information flow, suggesting that the fault may be in a self-healing phase or entering a new evolutionary stage.

[0114] The cumulative index of influence intensity is defined as the difference between the first and last elements of the time factor vector or the cumulative integral of the regression slope: ,or This indicator reflects the total cumulative change in the information flow intensity from the occurrence of the fault mode to the current moment, and directly measures the development range of the fault severity.

[0115] In some implementations, the window length of sliding window linear regression can be set according to the monitoring sensitivity requirements. A shorter window allows for a faster response to trend changes, while a longer window provides a more stable slope estimate. Besides linear regression, the Mann-Kendall trend test can be used for nonparametric trend assessment. This test does not rely on data distribution assumptions and is highly robust to outliers. Alternatively, Bayesian change point detection methods can be used to identify structural change points in the trend to determine whether the fault has entered an accelerated degradation phase.

[0116] It should be noted that, through quantitative analysis of the time factor vector of causal components, this application can not only detect the emergence of failure modes in the early stage, but also continuously track the evolution direction and speed of the mode, providing a quantitative basis for maintenance decisions on the failure development trend, and realizing the upgrade from passive alarm to proactive trend prediction capability.

[0117] For example, the noise floor threshold is set to the upper bound corresponding to twice the standard deviation of the mean of the existing component weights under normal conditions. For each causal component, its weight is continuously recorded. and time factor vector When a new component weight is detected to exceed the threshold and The elements in the middle monotonically increase with the window index, indicating a new early fault causal pattern. For existing patterns, linear regression is performed using the ten most recent time windows as a sliding window to extract... and As an indicator of time-varying trends, the output is updated regularly.

[0118] S403. Based on the causal component, the channel and scale load distribution identifies the source of fault influence, propagation path and characteristic frequency band, and combines the time evolution trend index to determine the early fault state.

[0119] Specifically, for each early fault causal pattern, the physical attribute information of the fault is extracted based on the load distribution of its factor vector. (Transmission channel factor vector) The element values ​​represent the probability that each channel is a source of fault information; the channel with the highest load is identified as the measurement point affecting the fault. (Receive channel factor vector) The element values ​​represent the degree to which each channel is affected by fault information. The channel with the highest load is identified as the terminal of the main propagation path. By combining the pairing relationship between the transmitting and receiving channels, a complete propagation path of fault information from the source to the affected components can be constructed. (Scale factor vector) The element values ​​represent the activity level of the fault mode at different characteristic time scales. The scale with the largest load is determined as the fault characteristic frequency band, which corresponds to the energy concentration frequency band of the fault impact signal.

[0120] When two or more early fault causal modes exist simultaneously, i.e., the number of components R obtained by tensor decomposition is greater than or equal to 2, this application automatically determines that the device has a coupled fault, i.e., a situation where multiple fault sources arise simultaneously or a fault source induces a cascading fault. Based on each mode in the transmission channel factor... Receive channel factor and scale factor Differences in load distribution enable unsupervised blind separation of different coupled faults.

[0121] Specifically, if the maximum loads of the transmit channel factor vectors of the two components point to different measurement points, they are determined to be independent fault sources at different locations; if the transmit channel factors of the two components point to the same measurement point but the peak values ​​of the scale factor distribution are located in different frequency bands, they are determined to be different types of faults at the same location; if the receive channel factor of one component highly overlaps with the transmit channel factor of another component, it is determined to be a cascading propagation relationship of faults. After separation, subsequent trend determination and state diagnosis are performed independently for each fault mode.

[0122] Early fault state determination is based on a combination of temporal evolution trend indicators and physical attribute information. For each separated fault mode, if its development direction indicator... The value remains consistently greater than zero, and the cumulative effect intensity index If the severity limit is exceeded, the equipment component where the fault-affecting source measurement point corresponding to the fault mode is located is determined to have entered an early fault state. To further improve diagnostic reliability, this application introduces a consistency verification mechanism: the fault-affecting source measurement point determined to have entered an early fault state is compared with the measurement point in step S202 that triggered the abnormal diffusion event and caused the abnormal diffusion index of the multi-channel signal segment to exceed the limit. If the two are consistent, the physical source of the early fault is confirmed to be reliable, indicating that the diagnostic conclusions of single-channel abnormal diffusion analysis and multi-channel causal information flow analysis corroborate each other. If they are inconsistent, the fault mode is marked as an observation state, and its development trend is continued to be tracked to avoid false alarms caused by system disturbances or model fluctuations.

[0123] The output of early fault diagnosis results includes: the location of the fault-affecting measurement point, based on... The maximum load value is determined; the main fault propagation path is based on... and The correspondence between loads is determined; the fault characteristic frequency band is based on... The scale corresponding to the maximum load is determined; the trend of fault severity is determined by the cumulative impact strength index. Current value and its growth rate Characterization; fault type identification, determined based on the comprehensive characteristics of characteristic frequency bands and propagation paths, such as typical rotating machinery fault types like inner ring fault, outer ring fault, or rolling element fault.

[0124] In some implementations, the preset severity limit can be set based on historical fault data statistics or expert experience, or a conservative value can be initially set and adaptively adjusted based on verification feedback during operation. Fault type discrimination can establish a knowledge base mapping characteristic frequency bands and propagation paths to fault types, and utilize the prior physical laws of typical faults to assist in classification.

[0125] It should be noted that this application achieves a dual verification loop of single-channel statistical analysis and multi-channel causal analysis by verifying the consistency between the factor load distribution of tensor decomposition and the abnormal diffusion trigger source, which significantly improves the credibility of early fault diagnosis results and avoids misjudgments that may occur with a single analysis method.

[0126] Based on the above technical solution, the causal components of the fault are extracted from the causal information tensor in an unsupervised manner through non-negative CP decomposition, the generation and evolution trend of the fault mode are monitored through quantitative analysis of the time factor vector, the fault impact source, propagation path and characteristic frequency band are identified through the channel and scale factor load distribution, and consistency verification is performed in combination with the abnormal diffusion trigger source. This application realizes the full automation and intelligence of the entire process from early fault perception, coupled fault separation, propagation path tracking to fault state diagnosis and severity assessment.

[0127] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0128] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely illustrative descriptions of the application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. An intelligent monitoring method based on multi-source data, characterized in that, include: Vibration and acoustic emission signal data from several measuring points on the equipment are collected to obtain signal data for each measuring point. Time-varying anomaly diffusion spectrum analysis was performed on the signal data at each measurement point to calculate the anomaly diffusion index, which characterizes the thick-tailed characteristics of the signal increment distribution. When the abnormal diffusion index of any measuring point exceeds the dynamic threshold, the multi-channel signal segment containing all measuring points before and after the current moment is extracted. The multi-channel signal segment is decomposed into multiple scales, the conditional transfer entropy between channels at each scale is calculated, and a causal information tensor is constructed based on the conditional transfer entropy. Tensor decomposition is performed on the causal information tensor to obtain multiple causal components; Monitor the generation and temporal evolution trend of the causal components to identify early failure states of the equipment.

2. The intelligent monitoring method based on multi-source data according to claim 1, characterized in that, The signal data from each measurement point are subjected to time-varying anomaly diffusion spectrum analysis, specifically including: For each measurement point signal, a sliding window is used to extract a signal segment of a preset length, and the detrended increment sequence of the signal within the window is calculated; wherein, the detrended increment sequence refers to the sequence of signal increments calculated for different delay intervals after removing the trend term from the original signal; Based on the aforementioned trend increment sequence, construct a q-order structure function: By fitting Estimate the generalized Hurst exponent for each order q. ;in, For delay, For signal delay The incremental change This represents the time average, where q is the real number order. The Anomalous Diffusion Index (ADI) is calculated based on the generalized Hurst index, using the following formula: Where q takes values ​​in the range [-5, 5]. This represents the function for calculating standard deviation. It is the generalized Hurst exponent when q=2.

3. The intelligent monitoring method based on multi-source data according to claim 2, characterized in that, The calculation process of the detrending increment sequence includes: A trend line is obtained by performing polynomial fitting on the original signal segment within the sliding window, and the trend line is subtracted from the original signal segment to obtain the detrended signal. In the detrending signal, for each time point t and a given delay τ, the difference between the signal value at time point t and after the delay τ is calculated to obtain the detrending increment sequence.

4. The intelligent monitoring method based on multi-source data according to claim 1, characterized in that, The method for obtaining the dynamic threshold includes: During the confirmed normal operating period of the equipment, the abnormal diffusion index sequence of each measuring point is continuously calculated, the probability density distribution of the abnormal diffusion index sequence is fitted, and the static benchmark threshold is determined according to the preset false alarm rate; wherein, the false alarm rate represents the maximum permissible probability of misjudging the equipment status as an abnormal state. The operating parameters of the equipment are collected synchronously, and a mapping model is constructed with the operating parameters as input and the anomaly diffusion index as output; wherein, the operating parameters include, but are not limited to, speed, load and temperature; During actual monitoring, the mapping model is queried based on real-time operating parameters to obtain the predicted value of the anomaly diffusion index under real-time operating conditions. The difference between the predicted value and the predicted value of the abnormal diffusion index of the equipment under the baseline operating conditions is used as the index offset. The dynamic threshold is obtained by adding the offset to the static baseline threshold.

5. The intelligent monitoring method based on multi-source data according to claim 1, characterized in that, The step of performing multi-scale decomposition on the multi-channel signal segments and calculating the conditional transfer entropy between channels at each scale specifically includes: Noise-assisted multivariate empirical mode decomposition is performed on the multi-channel signal segment to obtain multiple scale-aligned and channel-comparable mode function components; At each scale component, for any two channels i and j, the conditional transfer entropy from channel i to channel j is calculated using the signals of the other channels at that scale as conditions; the conditional transfer entropy reflects the intensity of the directional information flow from channel i to channel j at that scale after excluding the influence of common driving variables.

6. The intelligent monitoring method based on multi-source data according to claim 1, characterized in that, The construction of the causal information tensor based on the conditional transitive entropy specifically includes: Using a sliding time window as the unit, the conditional transfer entropy at each scale within each time window is calculated; For any scale, the conditional transfer entropy between all pairs of channels is arranged into an M×M causal matrix; where M is the total number of measurement point channels; the measurement point channel refers to the signal acquisition channel corresponding to each independently arranged sensor; By sequentially stacking the causal matrices of each scale for T consecutive time windows along the time dimension, a four-dimensional tensor is formed, resulting in a causal information tensor. The four-dimensional tensor corresponds to the time window index, the sending channel index, the receiving channel index, and the scale index, respectively. The sending channel index and the receiving channel index identify the measurement point channel as the information outflow and the measurement point channel as the information inflow, respectively.

7. The intelligent monitoring method based on multi-source data according to claim 1, characterized in that, The tensor decomposition of the causal information tensor to obtain multiple causal components specifically includes: Using the CANDECOMP / PARAFAC decomposition method, the four-dimensional causal information tensor is approximated as a weighted linear combination of a finite number of rank-1 components. Each rank-1 component consists of the outer product of a time factor vector, a transmission channel factor vector, a reception channel factor vector, and a scale factor vector. The factor vectors are feature vectors extracted from each dimension by the tensor decomposition and are obtained through alternating least squares iterative updates. During the decomposition process, when a new time window slice is obtained, the existing rank-one component factor vector is recursively adjusted through an online incremental update algorithm; If the reconstruction error of the new time window slice exceeds a preset threshold, a new rank-one component is added and initialized to obtain a causal component; wherein, the rank-one component is used to characterize a fault causal propagation mode.

8. The intelligent monitoring method based on multi-source data according to claim 1, characterized in that, The monitoring of the generation and temporal evolution trend of the causal components specifically includes: The weight of each causal component is continuously recorded, and the time factor vector of each causal component is tracked element by element; wherein, the weight is obtained by the CANDECOMP / PARAFAC decomposition method; When the weight of a certain causal component exceeds the preset noise threshold, and each element in the time factor vector of the causal component shows a monotonically increasing trend over time, the causal component is determined to be a newly generated early fault causal pattern. For each existing early failure causal mode, a sliding window linear regression is performed on the time factor vector to extract the regression slope as an indicator of the development direction of the failure mode, and the cumulative change in the regression is extracted as an indicator of the cumulative influence intensity of the failure mode, thus obtaining a time evolution trend indicator.

9. The intelligent monitoring method based on multi-source data according to claim 8, characterized in that, The process for identifying early fault states specifically includes: For the newly generated early fault causal pattern, the channel with the largest load in the transmitting channel factor vector is extracted as the fault influence source measurement point, the channel with the largest load in the receiving channel factor vector is extracted as the main propagation path, and the scale with the largest load in the scale factor vector is extracted as the fault feature frequency band. When two or more early fault causal modes exist simultaneously, it is determined that the device has a coupled fault. Based on the load distribution differences of each mode on the transmit channel factor, receive channel factor and scale factor, the coupled fault is separated, and the time evolution trend index of each separated fault mode is calculated. For any isolated fault mode, if the development direction index is greater than zero and the cumulative influence intensity index exceeds the preset limit, then the equipment component where the fault influence source measuring point corresponding to the fault mode is located is determined to have entered an early fault state. When the fault impact source measurement point is consistent with the measurement point where the abnormal diffusion index exceeds the limit, the fault impact source, main propagation path, fault characteristic frequency band and time evolution trend index are output to obtain early fault diagnosis results.

10. An intelligent monitoring system based on multi-source data, characterized in that, include: The module includes anomaly analysis, interception, cause-effect construction, and fault identification modules; among them, The anomaly analysis module is used to collect vibration and acoustic emission signal data from several measuring points on the equipment, obtain signal data from each measuring point, perform time-varying anomaly diffusion spectrum analysis on the signal data from each measuring point, and calculate the anomaly diffusion index that characterizes the thick-tailed characteristics of the signal increment distribution. The interception module is used to intercept a multi-channel signal segment containing all measurement points before and after the current moment when the abnormal diffusion index of any measurement point exceeds the dynamic threshold. The causal construction module is used to perform multi-scale decomposition on the multi-channel signal segment, calculate the conditional transit entropy between channels at each scale, and construct a causal information tensor based on the conditional transit entropy. The fault identification module is used to perform tensor decomposition on the causal information tensor to obtain multiple causal components, and monitor the generation and temporal evolution trend of the causal components to identify the early fault state of the equipment.

Images