A method, system and related devices for failure monitoring of industrial equipment

Abstract

The application provides a fault monitoring method and system for industrial equipment and related equipment, and relates to the technical field of industrial equipment monitoring. The method comprises the following steps: acquiring multi-parameter sensing data of the industrial equipment; forming multi-source time series data after time-space correlation of the multi-parameter sensing data; extracting a multi-dimensional feature vector corresponding to each parameter from a parameter channel of the multi-source time series data; inputting the multi-dimensional feature vector into a multi-level fusion model; fusing the multi-dimensional feature vector through the multi-level fusion model to generate a global fusion feature vector; inputting the global fusion feature vector into a trained diagnosis model to output a local health score and a corresponding confidence; generating a fault probability of the industrial equipment in a Bayesian inference framework by combining historical health baseline data and real-time operating conditions of the industrial equipment; and determining an abnormal parameter and a contribution degree of the abnormal parameter to a current abnormal state based on a comparison result of the fault probability and a decision threshold.

Description

Technical Field

[0001] This application relates to the field of industrial equipment monitoring technology, and in particular to a method, system and related equipment for fault monitoring of industrial equipment. Background Technology

[0002] With the rapid development of the Internet of Things (IoT) and intelligent sensing technologies, multi-dimensional and panoramic sensing of the operating status of critical industrial equipment has become possible. By deploying various heterogeneous sensors such as vibration, temperature, pressure, acoustics, and current sensors, the system can generate massive, high-dimensional, and asynchronous time-series data streams. This multi-source heterogeneous data contains deep information about the health status of the equipment, and its effective fusion and intelligent analysis are the core of achieving predictive maintenance and avoiding catastrophic failures.

[0003] Currently, development in this technology field mainly revolves around multi-sensor data fusion, aiming to improve the accuracy, robustness, and intelligence of perception and decision-making. Related technical solutions can be broadly categorized into the following levels: In terms of data preprocessing and spatiotemporal alignment, the industry consensus is that resolving spatiotemporal discrepancies between sensor data is a primary prerequisite for effective fusion. While methods like adjacent timestamp region matching and flow estimation correct feature misalignment caused by asynchronous data, these approaches overlook the fact that spatiotemporal discrepancies between sensor data can lead to insufficient synchronization during fusion, thus affecting map construction accuracy. Solutions address this by constructing time alignment modules to improve spatial registration accuracy. These solutions primarily address the synchronization and registration of spatial sensing data such as visible light and LiDAR, but they lack in-depth correlation modeling for the complex, dynamically changing temporal and frequency domain coupling relationships between physical parameters such as vibration and temperature in industrial applications.

[0004] At the feature extraction and fusion level, efforts are made to mine more discriminative features from multi-source data, filter key feature subsets through feature selection techniques, and integrate them into a fusion feature vector through fusion algorithms. However, such methods often employ weighted averaging, principal component analysis, or simple neural network concatenation. Their feature weights are often static or determined through global optimization, failing to adaptively adjust the contribution of each parameter and feature dimension based on the real-time operating conditions of the equipment (such as load and speed). For the weak and sparse fault feature signals hidden in massive amounts of normal data exhibited by early anomalies, this "averaging" or "fixed weight" fusion strategy easily leads to the submergence of key anomaly information, resulting in severely insufficient sensitivity.

[0005] At the level of decision fusion and intelligent diagnosis, researchers are trying to introduce more advanced machine learning and probabilistic models, using LSTM models for time series prediction, and after blurring the prediction error, combining it with Bayesian networks to calculate the posterior probability to trigger decision-making, or using target intent recognition methods based on dynamic Bayesian networks to process time series information and uncertain information.

[0006] These technologies demonstrate a trend towards combining time-series modeling, uncertainty reasoning, and decision-making. However, they still have significant limitations: First, their decision fusion is mostly at a single-level stage of "feature-decision" or "error-probability," lacking a multi-level, progressive fusion framework that progresses from fine-grained feature adaptive weighting to preliminary multi-model diagnosis, and then to dynamic global probability updates, making it difficult to refine anomalous information layer by layer. Second, existing methods such as Bayesian networks often require pre-defined, complete network structures or rule bases. For the nonlinear coupling relationships between multiple physical quantities in industrial equipment that change with operating conditions, the models struggle to learn and represent them dynamically, leading to diagnostic failures under varying operating conditions.

[0007] Furthermore, these technologies generally lack interpretability in decision-making. After the system provides an anomaly probability or alarm, operations and maintenance personnel cannot quickly and intuitively understand which one or more key parameters led to the decision, and how much influence they each have. This seriously hinders the rapid verification of alarms and subsequent maintenance actions, reducing the practical value of intelligent systems. Summary of the Invention

[0008] The main objective of this application is to propose a fault monitoring method, system, and related equipment for industrial equipment, aiming to solve the problems in the background art.

[0009] To achieve the above objectives, one aspect of this application proposes a fault monitoring method for industrial equipment, the method comprising: Acquire multi-parameter sensing data of industrial equipment, perform spatiotemporal correlation on the multi-parameter sensing data to form multi-source time series data, and extract multi-dimensional feature vectors corresponding to each parameter from the parameter channels of the multi-source time series data. The multidimensional feature vector is input into a multi-level fusion model, and the multi-dimensional feature vector is fused through the multi-level fusion model to generate a global fusion feature vector; The global fusion feature vector is input into the trained diagnostic model, and the diagnostic model outputs a local health score and the corresponding confidence level. Based on the local health score and the corresponding confidence level, and combined with the historical health baseline data and real-time operating conditions of the industrial equipment, the failure probability of the industrial equipment is generated within a Bayesian inference framework. Based on the comparison between the fault probability and the decision threshold, the abnormal parameters and their contribution to the current abnormal state are determined.

[0010] In some embodiments, the multidimensional feature vector is input into a multi-level fusion model, and the multidimensional feature vector is fused through the multi-level fusion model to generate a global fusion feature vector; The attention mechanism network built into the multi-level fusion model is used to incorporate the multi-dimensional feature vectors; The attention mechanism network adaptively learns and outputs a weight vector based on the current working conditions. Each weight element in the weight vector corresponds to a feature dimension in the multi-dimensional feature vector. The feature dimensions are pre-labeled with corresponding physical parameter labels. The multidimensional feature vector is weighted and fused element by element using the weight vector to generate a weighted global fused feature vector.

[0011] In some embodiments, determining the abnormal parameters and their contribution to the current abnormal state based on the comparison between the fault probability and the decision threshold specifically includes: The failure probability is compared with a dynamic decision threshold; When the failure probability exceeds the decision threshold, an early warning signal is triggered. Based on the warning signal, according to the physical parameter labels to which each feature dimension belongs, the weight vector is divided into several weight sub-vectors, where each weight sub-vector corresponds to a physical parameter; Calculate the magnitude or average weight value of the weight subvector corresponding to each physical parameter; The physical parameter with the highest modulus or average weight value is identified as an abnormal parameter, and the contribution of the abnormal parameter is represented by the proportion of its weight value to the total weight.

[0012] In some embodiments, the step of forming multi-source time-series data by performing spatiotemporal correlation on the multi-parameter sensing data specifically includes: The sensor data stream corresponding to the sensor with the highest sampling rate among all sensors or a pre-specified reference clock source is used as the reference time axis; Interpolate or resample the data streams from the remaining sensors to ensure that the data points of all parameters are time-aligned on a unified timestamp sequence. After time alignment is completed, the physical installation location of the sensor and the geometric model of the industrial equipment are obtained; Using the unified three-dimensional coordinate system of the industrial equipment as a spatial reference, the data streams of each sensor are mapped to their respective monitoring points in the unified three-dimensional coordinate system to complete spatial alignment.

[0013] In some embodiments, generating the failure probability of the industrial equipment within a Bayesian inference framework specifically includes: The local health score and its corresponding confidence level are converted into a likelihood function, and the confidence level is used to adjust the strength of the likelihood function. Using historical health baseline data and real-time operating conditions as priors, the local health score output by the diagnostic model is fused according to Bayesian formula, and the failure probability of the industrial equipment is output after normalization.

[0014] In some embodiments, before determining the abnormal parameters and their contribution to the current abnormal state based on the comparison result of the fault probability and the decision threshold, the method further includes: Obtain the overall abnormal probability distribution of the industrial equipment under long-term normal operation in historical health baseline data, and use the preset quantile of the overall abnormal probability distribution as the basic threshold. The real-time operating condition parameters of the industrial equipment are obtained, and the real-time operating condition parameters include at least the load level and the environmental stress level. Based on the real-time operating condition parameters, determine the corresponding adjustment factor; The basic threshold is multiplied or added to the adjustment factor to generate a dynamic decision threshold.

[0015] In some embodiments, before inputting the globally fused feature vector into the trained diagnostic model, the method further includes: Acquire raw multi-parameter sensing data of industrial equipment under historical normal operating conditions and various known abnormal conditions; After performing spatiotemporal correlation and feature extraction on the original multi-parameter sensing data, a training sample set is generated; The training sample set is input into multiple different types of diagnostic models for preliminary training, including support vector machine models, random forest models, and deep neural network models. The support vector machine model is trained using radial basis function as kernel function, the random forest model is trained by constructing and integrating multiple decision trees, and the deep neural network model is trained using cross-entropy as loss function and backpropagation algorithm. After initial training of the diagnostic models, the output of each diagnostic model on the validation set is calibrated to complete the final training of the diagnostic models.

[0016] To achieve the above objectives, another aspect of this application proposes a fault monitoring system for industrial equipment, the system comprising: The feature vector construction module is used to acquire multi-parameter sensing data of industrial equipment, perform spatiotemporal correlation on the multi-parameter sensing data to form multi-source time series data, and extract multi-dimensional feature vectors corresponding to each parameter from the parameter channels of the multi-source time series data. The feature fusion module is used to input the multi-dimensional feature vector into a multi-level fusion model, and fuse the multi-dimensional feature vector through the multi-level fusion model to generate a global fusion feature vector; The health scoring module is used to input the global fusion feature vector into the trained diagnostic model, and output a local health score and corresponding confidence level through the diagnostic model. The probability calculation module is used to generate the failure probability of the industrial equipment within a Bayesian inference framework based on the local health score and the corresponding confidence level, combined with the historical health baseline data and real-time operating conditions of the industrial equipment. The fault monitoring module is used to determine the abnormal parameters and their contribution to the current abnormal state based on the comparison result between the fault probability and the decision threshold.

[0017] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0018] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0019] The embodiments of this application include at least the following beneficial effects: This application proposes a fault monitoring method, system, and related equipment for industrial equipment. By fusing multi-dimensional feature vectors and inputting the resulting global fused feature vectors in parallel into diagnostic models, each model independently performs preliminary diagnosis and outputs a local health score with confidence. This achieves feature-level fusion and decision-level fusion. This application constructs a progressive analysis chain from data feature purification to multivariate preliminary diagnosis. Feature-level fusion provides high signal-to-noise ratio input for subsequent diagnosis, while the parallel processing of decision-level decision models constitutes an integrated learning system that can perform cross-validation and complementarity from different algorithm perspectives. This effectively improves the potential problems of insufficient generalization ability or overfitting of a single model and enhances the robustness and accuracy of the overall diagnosis. This application incorporates the uncertainties of different diagnostic models into a unified Bayesian inference framework, realizing the quantitative fusion and dynamic evolution tracking of multi-source uncertain information, enabling early and sensitive capture of abnormal evolution trends, and effectively suppressing false alarms caused by fluctuations in single-point data. The parameters output by this application include abnormal parameters and their corresponding contribution, making the decision-making basis transparent. This greatly assists operation and maintenance personnel in quickly verifying alarms and analyzing root causes, transforming traditional black-box alarms into white-box decision support, and improving the practicality and reliability of predictive maintenance systems. Attached Figure Description

[0020] Figure 1 This is a flowchart of a fault monitoring method for industrial equipment provided in an embodiment of this application; Figure 2 This is a diagram of a fault monitoring system for industrial equipment provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0022] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0023] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0025] In related technologies, the sensitivity to early and weak anomalies is low: due to the lack of refined modeling of the dynamic coupling relationship of multiple parameters and adaptive feature focusing mechanism, traditional static or shallow fusion methods cannot effectively separate and enhance weak early fault features from strong background noise, resulting in missed detection or alarm delay.

[0026] Poor generalization ability and robustness under varying operating conditions: The parameters or rules of related fusion and diagnostic models are often trained based on fixed operating conditions. When the operating load, environment and other conditions change, the normal correlation pattern between parameters changes, the original model is prone to false alarms, and the system reliability decreases.

[0027] The decision-making process is opaque and lacks interpretability: related intelligent diagnostic systems usually give a "black box" conclusion that is difficult to trace. Operation and maintenance personnel cannot know the specific physical basis of the alarm and the ranking of parameter contributions, resulting in low trust in automated decision-making and failure to form a closed-loop management of "perception-diagnosis-decision-verification".

[0028] The aforementioned technical issues will have serious impacts and consequences in practical work: Missed or delayed alarms of early faults will cause the best maintenance window to be missed, allowing potential defects to develop into functional failures or even catastrophic accidents, greatly increasing maintenance costs and safety risks. Frequent false alarms under varying operating conditions will trigger a "boy who cried wolf" effect, causing genuine alarms to be ignored, while consuming a large amount of unnecessary inspection and troubleshooting resources. The lack of interpretability in decision-making prevents high-value monitoring data from being transformed into clear operational knowledge, hindering the effective collaboration between human experience and artificial intelligence, and restricting the maturity and development of predictive maintenance systems.

[0029] In view of this, this application provides a method, system and related equipment for fault monitoring of industrial equipment.

[0030] refer to Figure 1 As shown, one embodiment of this application proposes a fault monitoring method for industrial equipment, wherein the method includes: S101: Acquire multi-parameter sensing data of industrial equipment, perform spatiotemporal correlation on the multi-parameter sensing data to form multi-source time series data, and extract multi-dimensional feature vectors corresponding to each parameter from the parameter channels of the multi-source time series data. S102: Input the multi-dimensional feature vector into the multi-level fusion model, and fuse the multi-dimensional feature vector through the multi-level fusion model to generate a global fusion feature vector; S103: Input the global fusion feature vector into the trained diagnostic model, and output the local health score and corresponding confidence level through the diagnostic model; S104: Based on the local health score and the corresponding confidence level, and combined with the historical health baseline data and real-time operating conditions of the industrial equipment, generate the failure probability of the industrial equipment within the Bayesian inference framework. S105: Based on the comparison result between the fault probability and the decision threshold, determine the abnormal parameters and the contribution of the abnormal parameters to the current abnormal state.

[0031] Specifically, the multi-parameter sensing data in step S101 is acquired through multiple sensors, and the multi-source time-series data is formed by spatiotemporal correlation of the multi-parameter sensing data, including: The sensor data stream corresponding to the sensor with the highest sampling rate among all sensors or a pre-specified reference clock source is used as the reference time axis; Interpolate or resample the data streams from the remaining sensors to ensure that the data points of all parameters are time-aligned on a unified timestamp sequence. After time alignment is completed, the physical installation location of the sensor and the geometric model of the industrial equipment are obtained; Using the unified three-dimensional coordinate system of the industrial equipment as a spatial reference, the data streams of each sensor are mapped to their respective monitoring points in the unified three-dimensional coordinate system to complete spatial alignment.

[0032] In this application, the multidimensional feature vector includes time-domain features, frequency-domain features, and time-frequency features. The time-domain features include mean, standard deviation, peak-to-peak value, skewness, and kurtosis. The frequency-domain features include the energy proportion or centroid frequency of the signal within a preset frequency band, calculated by fast Fourier transform. The time-frequency features include the instantaneous amplitude and frequency of wavelet energy or intrinsic mode function at each scale, obtained by wavelet transform or Hilbert-Huang transform.

[0033] In this application, parameter channels or parameters are used to refer to physical sources, and in this application, they are used to refer to sensors, while multidimensional feature vectors are sensor data streams obtained through parameters.

[0034] Specifically, in step S102, the multi-dimensional feature vector is input into a multi-level fusion model, and the multi-dimensional feature vector is fused through the multi-level fusion model to generate a global fused feature vector, including: The attention mechanism network built into the multi-level fusion model is used to incorporate the multi-dimensional feature vectors; The attention mechanism network adaptively learns and outputs a weight vector based on the current working conditions. Each weight element in the weight vector corresponds to a feature dimension in the multi-dimensional feature vector. The feature dimensions are pre-labeled with corresponding physical parameter labels. The weight vector is used to perform element-wise weighted fusion of the multidimensional feature vector to generate a weighted global fusion feature vector. That is, the weight vector and the multidimensional feature vector are fused element-wise, or each feature dimension is multiplied by the corresponding weight element to obtain the weighted global fusion feature vector. Specifically, the expression for the weight vector is as follows: ,in For weighted sub-vectors, The number of feature dimensions The expression for the global fusion feature vector is: ; in For global fusion feature vectors, For the weight vector, For multidimensional feature vectors, For Hadama accumulation.

[0035] The attention mechanism network is a feedforward neural network, whose input is a concatenated vector of the multidimensional feature vector and the condition encoding vector representing the real-time operating conditions. The network includes a fully connected hidden layer and a Softmax output layer. The hidden layer performs a nonlinear transformation on the concatenated vector, and the Softmax output layer normalizes the output of the hidden layer into a weight vector that corresponds to the total dimension of the input features and has a sum of 1. The weighted global fusion feature vector is obtained by performing a Hadamard product operation between the weight vector and the feature part in the concatenated vector.

[0036] Furthermore, the attention mechanism network dynamically adjusts the contribution of different physical parameters and different feature dimensions in the fusion process based on the load, speed or environmental stress information of the device in real time, so as to enhance the key features related to the current abnormal state.

[0037] This application achieves feature-level fusion and decision-level fusion by fusing multi-dimensional feature vectors and inputting the resulting global fused feature vectors into the diagnostic model in parallel. Each model independently performs a preliminary diagnosis and outputs a local health score with confidence.

[0038] In the feature-level fusion stage, this application introduces an attention mechanism network implemented using a feedforward neural network and a Softmax function. This attention mechanism network takes the concatenated multidimensional feature vector and the real-time operating condition encoding vector as input. Through nonlinear transformation and normalization, it dynamically generates a weight vector corresponding to the input feature dimensions, with a sum of 1. This weight vector is then used to weight the original features using a Hadamard product (element-wise multiplication) to generate a global fusion feature vector. This approach enables the system to adaptively adjust the contribution of different physical parameters and feature dimensions in the fusion process based on the real-time operating conditions of the equipment (such as load and speed), achieving dynamic focusing and enhancement of key abnormal signals. This overcomes the shortcomings of traditional weighted averaging or fixed-weight fusion methods in terms of insufficient sensitivity under varying operating conditions, significantly improving the ability to identify early and weak abnormal features.

[0039] Specifically, prior to step S103, the process includes pre-training of the diagnostic model, which includes: Acquire raw multi-parameter sensing data of industrial equipment under historical normal operating conditions and various known abnormal conditions; After performing spatiotemporal correlation and feature extraction on the original multi-parameter sensing data, a training sample set is generated; The training sample set is input into multiple different types of diagnostic models for preliminary training, including support vector machine models, random forest models, and deep neural network models. The support vector machine model is trained using radial basis function as kernel function, the random forest model is trained by constructing and integrating multiple decision trees, and the deep neural network model is trained using cross-entropy as loss function and backpropagation algorithm. After initial training of the diagnostic models, the output of each diagnostic model on the validation set is calibrated to complete the final training of the diagnostic models, so that the output score can represent the probability estimate of whether the sample belongs to the abnormal category.

[0040] Specifically, the deep neural network model is a multilayer perceptron, whose input layer receives the weighted global fusion feature vector, and after passing through at least two fully connected hidden layers, the output layer outputs the local health score through the Sigmoid activation function.

[0041] The local health score output by the deep neural network model is given by the Sigmoid activation function, with the following formula: ; in For local health scoring, This is the output of the fully connected hidden layer.

[0042] Furthermore, support vector machines and random forest models directly output calibrated failure probabilities.

[0043] Specifically, generating the failure probability of the industrial equipment within the Bayesian inference framework includes: The local health score and its corresponding confidence level are converted into a likelihood function. The confidence level is used to adjust the strength of the likelihood function, where This indicates an assumption of an anomaly. For local health scoring, This represents the probability of failure.

[0044] Using historical health baseline data and real-time operating conditions as priors, the local health score output by the diagnostic model is fused according to Bayesian formula, and the failure probability of the industrial equipment is output after normalization.

[0045] Higher confidence levels correspond to a sharper likelihood distribution; let the prior anomaly probability at time k be... The posterior probability after fusing multiple diagnostic models Calculate using the following formula: ; in This indicates that the posterior probability, after being normalized, is used as the overall anomaly probability output at time k.

[0046] For support vector machine and random forest models, confidence level The calculation formula is the absolute difference between the model's predicted probability for the current input sample and 0.5, multiplied by 2. ; For deep neural network models, the confidence score is the reciprocal of the standard deviation of the scores generated by multiple forward propagations of the Dropout layer during prediction, after being normalized to the minimum and maximum values.

[0047] The formula for calculating the confidence score of a deep neural network model is: The scoring sequence is obtained through the second Dropout forward propagation. Calculate its standard deviation: ; in For the first The local health score output by the Dropout forward propagation, where t is the index of the forward propagation number. This represents the total number of forward propagations in Dropout. The arithmetic mean of the rating sequence. denoted as the standard deviation of the scoring sequence.

[0048] Then take the reciprocal and perform maximum / minimum normalization, that is... ; in For confidence level, The minimum standard deviation. This represents the maximum standard deviation.

[0049] Specifically, in step S105, based on the comparison result between the fault probability and the decision threshold, the abnormal parameters and their contribution to the current abnormal state are determined, including: The failure probability is compared with a dynamic decision threshold; When the failure probability exceeds the decision threshold, an early warning signal is triggered. Based on the warning signal, according to the physical parameter labels to which each feature dimension belongs, the weight vector is divided into several weight sub-vectors, where each weight sub-vector corresponds to a physical parameter; Calculate the magnitude or average weight value of the weight subvector corresponding to each physical parameter; The physical parameter with the highest modulus or average weight value is identified as an abnormal parameter, and the contribution of the abnormal parameter is represented by the proportion of its weight value to the total weight.

[0050] Furthermore, the acquisition of dynamic decision thresholds includes: Obtain the overall anomaly probability distribution of the industrial equipment under long-term normal operation in historical health baseline data, and use the preset quantile of the overall anomaly probability distribution as the basic threshold. That is, take its 1-α quantile as the basic threshold. Where α is the preset upper limit of the false alarm rate. The real-time operating condition parameters of the industrial equipment are obtained, and the real-time operating condition parameters include at least the load level and the environmental stress level. Based on the real-time operating condition parameters, determine the corresponding adjustment factor; The basic threshold is multiplied or added to the adjustment factor to generate a dynamic decision threshold.

[0051] The formula for calculating the dynamic decision threshold is: ; in For dynamic decision thresholds, Based on the threshold, To adjust the factor, This is the preset threshold offset.

[0052] Taking bearing failure as an example: the kurtosis value of a normal bearing is close to 3. When early pitting or cracking occurs in the bearing, the kurtosis value can significantly rise to above 5. For example, in a wind turbine gearbox bearing monitoring case, the kurtosis index detected inner ring pitting failure 3 months in advance, while the effective value (RMS) only significantly increased after the failure worsened; the kurtosis value of inner ring failure is usually higher than that of outer ring failure. Remedial measures: Set a kurtosis threshold (e.g., ≥4) to trigger an early warning, promptly arrange for shutdown and bearing replacement to avoid chain damage to the gearbox; simultaneously, reduce the load and closely monitor the bearing temperature and vibration trends.

[0053] refer to Figure 2 As shown, another aspect of this application embodiment also proposes a fault monitoring system for industrial equipment, the system comprising: The feature vector construction module is used to acquire multi-parameter sensing data of industrial equipment, perform spatiotemporal correlation on the multi-parameter sensing data to form multi-source time series data, and extract multi-dimensional feature vectors corresponding to each parameter from the parameter channels of the multi-source time series data. The feature fusion module is used to input the multi-dimensional feature vector into a multi-level fusion model, and fuse the multi-dimensional feature vector through the multi-level fusion model to generate a global fusion feature vector; The health scoring module is used to input the global fusion feature vector into the trained diagnostic model, and output a local health score and corresponding confidence level through the diagnostic model. The probability calculation module is used to generate the failure probability of the industrial equipment within a Bayesian inference framework based on the local health score and the corresponding confidence level, combined with the historical health baseline data and real-time operating conditions of the industrial equipment. The fault monitoring module is used to determine the abnormal parameters and their contribution to the current abnormal state based on the comparison result between the fault probability and the decision threshold.

[0054] The present application will be described below with reference to specific embodiments: Application scenario: Monitoring the condition and providing early warning of anomalies for the high-speed shaft bearing of a 1.5MW wind turbine gearbox, specifically: In this scenario, the protected object is the high-speed shaft of the gearbox, with a rated speed of 1800 rpm and a rotational frequency of 30 Hz. The sensor deployment scheme is as follows: a piezoelectric accelerometer (PCB 608A11) is installed radially and horizontally on both the drive-end bearing housing B1 and the non-drive-end bearing housing B2 of the high-speed shaft, with a range of ±50g, sensitivity of 100 mV / g, and a sampling rate of 2560 Hz; the same type of accelerometer is installed at measuring point G1 on the top of the gearbox housing, with the same sampling rate of 2560 Hz; a PT100 platinum resistance temperature sensor (accuracy ±0.3℃) is installed on both the drive-end and non-drive-end bearing housings and on the gearbox housing, with a sampling rate of 1 Hz; an oil particle size sensor is installed at the lubricating oil pipeline outlet to measure the concentration of ferromagnetic abrasive particles in the lubricating oil, in ppm, with a sampling rate of 1 Hz; a key phase sensor is set at the output end of the high-speed shaft to provide a speed signal of one pulse per revolution, with the measured speed fluctuating between 1780 and 1820 rpm. All sensors acquire data synchronously via a data acquisition unit, with the vibration sensor clock serving as the reference time axis. In this embodiment, the unit operates under full load conditions, corresponding to a condition encoding vector of [1,0,0].

[0055] In practice, the sensor data is first spatiotemporally aligned and features are extracted. Using the vibration data stream as a reference time axis, the 1Hz sampled data streams of temperature and oil abrasive particle concentration are linearly interpolated and resampled to align all parameters on a unified 2560Hz timestamp sequence, forming spatiotemporally aligned multi-source time-series data. Then, data is extracted using a sliding window with a time window length of 1 second and a sliding step size of 0.5 seconds.

[0056] For each parameter channel within each window, 15-dimensional features are extracted. Taking a window containing early anomalies as an example, the time-domain features extracted from the vibration channel of the drive-end bearing housing are: mean acceleration 0.12 m / s², standard deviation 2.83 m / s², peak-to-peak value 18.7 m / s², skewness 0.41, and kurtosis 5.82. The frequency-domain features are calculated using a 2560-point Fast Fourier Transform, yielding an energy percentage of 32% in the 0-500 Hz band, 47% in the 501-1000 Hz band, and 18% in the 1001-2000 Hz band, with a centroid frequency of 612 Hz. The time-frequency domain features are decomposed into four levels using the db4 wavelet, with the normalized energy values ​​of the detail coefficients d1 to d4 being 0.23, 0.35, 0.28, and 0.14, respectively. The characteristics of the vibration channel of the non-drive end bearing housing are: mean 0.09 m / s², standard deviation 2.15 m / s², peak-to-peak value 14.3 m / s², skewness 0.28, kurtosis 4.56, with frequency band energy proportions of 35%, 42%, and 20% respectively, a centroid frequency of 587 Hz, and wavelet energies of 0.19, 0.31, 0.30, and 0.20. The characteristics extracted from the temperature channel of the drive end bearing housing are: mean 76.8℃, standard deviation 0.35℃, peak-to-peak value 1.2℃, skewness 0.05, kurtosis 2.95, and low-frequency energy proportion of 92%. The characteristics of the temperature channel of the non-drive end are: mean 72.1℃, standard deviation 0.28℃, peak-to-peak value 1.0℃, skewness 0.03, kurtosis 2.88, and low-frequency energy proportion of 93%. The features extracted from the oil abrasive particle concentration channel are: mean 18.2 ppm, standard deviation 2.4 ppm, with zeros padded in the remaining dimensions. The features from the five parameter channels are concatenated to obtain a 75-dimensional feature vector, which is then concatenated with the 3-dimensional operating condition encoding vector to form a 78-dimensional input vector.

[0057] Next, the feature-level fusion step is performed. The attention mechanism network is a feedforward neural network containing a fully connected hidden layer with 128 neurons, using ReLU activation function, and a 78-dimensional softmax output layer. This network receives a 78-dimensional concatenated vector, outputs a weight vector W after forward propagation, where each dimension is non-negative and sums to 1. During the calculation, the weight sub-vectors corresponding to each parameter channel are extracted. The average weight for the vibration B1 channel is 0.312, for vibration B2 it is 0.198, for temperature B1 it is 0.247, for temperature B2 it is 0.131, and for oil abrasive particles it is 0.112. It can be seen that the network automatically assigns higher weights to the vibration B1 channel with abnormally high kurtosis and the temperature B1 channel with a relatively high average temperature. The Hadamard product is then applied to the feature parts using the weight vector W to obtain the global fused feature vector. .

[0058] Then perform decision-level fusion, The diagnostic model is fed in parallel into three independently trained diagnostic models: a support vector machine (SVM) model, a random forest model, and a deep neural network model. The SVM model uses a radial basis function kernel with parameters C=1.0 and γ=0.01. Its output is then scaled and calibrated using Platt to obtain a local health score. =0.78, confidence level Multiplying this score by 2 by the absolute difference between 0.5 and 0.5 gives 0.56. The random forest model consists of 100 decision trees with a maximum depth of 10, and its output is the mean probability that all decision trees voted to predict an anomaly. =0.72, confidence level =0.44. The deep neural network model is a multilayer perceptron, containing two fully connected hidden layers with 128 and 64 neurons respectively. The output layer uses the sigmoid activation function to output the score. =0.81; Perform 30 Dropout forward propagations on the model, with a dropout hold probability of 0.8, and obtain the standard deviation of the score sequence. The value is 0.042. Calculate its reciprocal and perform maximum-minimum normalization to obtain the confidence level. ==0.72.

[0059] The local health scores of the support vector machine model, random forest model, and deep neural network model were all greater than 0.5, and were concentrated between 0.72 and 0.81, which cross-validated the existence of the anomaly from the perspective of different algorithms.

[0060] In the Bayesian inference framework, the prior anomaly probability is first obtained. This prior anomaly probability, calculated based on historical health baseline data, is obtained for the unit under full-load operating conditions. Set to 0.03, the likelihood function strength of each model is weighted by its confidence level; high-confidence deep neural network models have stronger evidence. Logarithmic domain computation is used to avoid numerical underflow. The prior is multiplied by the three likelihoods, and after normalization, the overall anomaly probability at the current time step is obtained. =0.86, which is 0.83 greater than the prior.

[0061] Finally, tiered early warnings and interpretable outputs are executed. Based on the distribution of the overall anomaly probability under historical normal operating conditions, the 99.5th percentile is taken as the basic threshold. =0.15, preset false alarm rate upper limit α=0.005.

[0062] Adjustment factor corresponding to current full load condition =1.0, preset threshold offset =0.20, dynamic threshold =0.35.

[0063] because The value of 0.86 exceeds the dynamic threshold, triggering a level-two warning. Simultaneously with the warning signal output, the system backtracks the attention weight vector, calculates the average weight based on the parameters, and identifies the three parameters with the highest anomaly contribution: Drive-end bearing housing vibration (B1), contributing 31.2%, characterized by a kurtosis of 5.82 far exceeding the normal baseline of 3.0, indicating an impact fault; drive-end bearing housing temperature (B1), contributing 24.7%, with a mean of 76.8℃ significantly higher than the normal baseline of 55-62℃, indicating increased friction or lubrication deterioration; and non-drive-end bearing housing vibration (B2), contributing 19.8%, with a kurtosis of 4.56 further confirming the fault path. The oil abrasive particle concentration of 18.2 ppm also exceeds the normal baseline by 8 ppm, further confirming accelerated wear.

[0064] Based on the contribution ranking and characteristic performance, the following maintenance recommendations are generated: immediately reduce the load to 70% of the rated value and intensify monitoring of vibration kurtosis and temperature trends; complete oil ferrography analysis within 24 hours to confirm abrasive morphology and replenish or replace grease; within 72 hours, utilize a low wind speed window to shut down and inspect the inner raceway of the high-speed shaft drive end bearing, replacing the bearing as needed to prevent cascading damage to the gearbox. This closed-loop process fully demonstrates the complete link from adaptive feature fusion of multi-source sensing data to interpretable early warning decisions.

[0065] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0066] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0067] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0068] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0069] The methods provided in this application relate to the field of information technology. The methods provided in this application can be applied to terminals, servers, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or in-vehicle terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the method, but is not limited to the above forms.

[0070] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0071] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirection to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0072] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0073] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0074] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0075] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0076] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0077] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0078] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0079] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0080] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0081] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0082] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0083] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A fault monitoring method for industrial equipment, characterized in that, The method includes: Acquire multi-parameter sensing data of industrial equipment, perform spatiotemporal correlation on the multi-parameter sensing data to form multi-source time series data, and extract multi-dimensional feature vectors corresponding to each parameter from the parameter channels of the multi-source time series data. The multidimensional feature vector is input into a multi-level fusion model, and the multi-dimensional feature vector is fused through the multi-level fusion model to generate a global fusion feature vector; The global fusion feature vector is input into the trained diagnostic model, and the diagnostic model outputs a local health score and the corresponding confidence level. Based on the local health score and the corresponding confidence level, and combined with the historical health baseline data and real-time operating conditions of the industrial equipment, the failure probability of the industrial equipment is generated within a Bayesian inference framework. Based on the comparison between the fault probability and the decision threshold, the abnormal parameters and their contribution to the current abnormal state are determined.

2. The fault monitoring method for industrial equipment according to claim 1, characterized in that, The multidimensional feature vector is input into a multi-level fusion model, and the multi-dimensional feature vector is fused through the multi-level fusion model to generate a global fusion feature vector; The attention mechanism network built into the multi-level fusion model is used to incorporate the multi-dimensional feature vectors; The attention mechanism network adaptively learns and outputs a weight vector based on the current working conditions. Each weight element in the weight vector corresponds to a feature dimension in the multi-dimensional feature vector. The feature dimensions are pre-labeled with corresponding physical parameter labels. The multidimensional feature vector is weighted and fused element by element using the weight vector to generate a weighted global fused feature vector.

3. The fault monitoring method for industrial equipment according to claim 2, characterized in that, The determination of abnormal parameters and their contribution to the current abnormal state based on the comparison between the fault probability and the decision threshold specifically includes: The failure probability is compared with a dynamic decision threshold; When the failure probability exceeds the decision threshold, an early warning signal is triggered. Based on the warning signal, according to the physical parameter labels to which each feature dimension belongs, the weight vector is divided into several weight sub-vectors, where each weight sub-vector corresponds to a physical parameter; Calculate the magnitude or average weight value of the weight subvector corresponding to each physical parameter; The physical parameter with the highest modulus or average weight value is identified as an abnormal parameter, and the contribution of the abnormal parameter is represented by the proportion of its weight value to the total weight.

4. The fault monitoring method for industrial equipment according to claim 1, characterized in that, The process of spatiotemporally correlating the multi-parameter sensing data to form multi-source time-series data specifically includes: The sensor data stream corresponding to the sensor with the highest sampling rate among all sensors or a pre-specified reference clock source is used as the reference time axis; Interpolate or resample the data streams from the remaining sensors to ensure that the data points of all parameters are time-aligned on a unified timestamp sequence. After time alignment is completed, the physical installation location of the sensor and the geometric model of the industrial equipment are obtained; Using the unified three-dimensional coordinate system of the industrial equipment as a spatial reference, the data streams of each sensor are mapped to their respective monitoring points in the unified three-dimensional coordinate system to complete spatial alignment.

5. The fault monitoring method for industrial equipment according to claim 1, characterized in that, The generation of the failure probability of the industrial equipment within the Bayesian inference framework specifically includes: The local health score and its corresponding confidence level are converted into a likelihood function, and the confidence level is used to adjust the strength of the likelihood function. Using historical health baseline data and real-time operating conditions as priors, the local health score output by the diagnostic model is fused according to Bayesian formula, and the failure probability of the industrial equipment is output after normalization.

6. The fault monitoring method for industrial equipment according to claim 1, characterized in that, Before determining the abnormal parameters and their contribution to the current abnormal state based on the comparison result between the fault probability and the decision threshold, the method further includes: Obtain the overall abnormal probability distribution of the industrial equipment under long-term normal operation in historical health baseline data, and use the preset quantile of the overall abnormal probability distribution as the basic threshold. The real-time operating condition parameters of the industrial equipment are obtained, and the real-time operating condition parameters include at least the load level and the environmental stress level. Based on the real-time operating condition parameters, determine the corresponding adjustment factor; The basic threshold is multiplied or added to the adjustment factor to generate a dynamic decision threshold.

7. The fault monitoring method for industrial equipment according to claim 1, characterized in that, Before inputting the globally fused feature vector into the trained diagnostic model, the method further includes: Acquire raw multi-parameter sensing data of industrial equipment under historical normal operating conditions and various known abnormal conditions; After performing spatiotemporal correlation and feature extraction on the original multi-parameter sensing data, a training sample set is generated; The training sample set is input into multiple different types of diagnostic models for preliminary training, including support vector machine models, random forest models, and deep neural network models. The support vector machine model is trained using radial basis function as kernel function, the random forest model is trained by constructing and integrating multiple decision trees, and the deep neural network model is trained using cross-entropy as loss function and backpropagation algorithm. After initial training of the diagnostic models, the output of each diagnostic model on the validation set is calibrated to complete the final training of the diagnostic models.

8. A fault monitoring system for industrial equipment, characterized in that, The system includes: The feature vector construction module is used to acquire multi-parameter sensing data of industrial equipment, perform spatiotemporal correlation on the multi-parameter sensing data to form multi-source time series data, and extract multi-dimensional feature vectors corresponding to each parameter from the parameter channels of the multi-source time series data. The feature fusion module is used to input the multi-dimensional feature vector into a multi-level fusion model, and fuse the multi-dimensional feature vector through the multi-level fusion model to generate a global fusion feature vector; The health scoring module is used to input the global fusion feature vector into the trained diagnostic model, and output a local health score and corresponding confidence level through the diagnostic model. The probability calculation module is used to generate the failure probability of the industrial equipment within a Bayesian inference framework based on the local health score and the corresponding confidence level, combined with the historical health baseline data and real-time operating conditions of the industrial equipment. The fault monitoring module is used to determine the abnormal parameters and their contribution to the current abnormal state based on the comparison result between the fault probability and the decision threshold.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images