Fault early warning method and device

Abstract

The invention discloses a fault early warning method and device, and relates to the technical field of mechanical equipment health management. The method comprises the following steps: acquiring vibration signal data of mechanical equipment, and performing health degree marking and balance processing on a sample through a multi-stage degradation state feature distribution smoothing strategy; health factors with high consistency and high sensitivity are extracted by using a channel shuffling attention mechanism feature smoothing network, and the network realizes multi-scale degradation feature extraction through a convolutional layer, a pooling layer and a channel shuffling attention module; and inputting the extracted health factors into an isolated forest algorithm model, and determining an alarm threshold through anomaly detection to realize early fault early warning. According to the method, the problems of sample imbalance, inflexible attention distribution and early warning hysteresis in an existing method are solved, and the accuracy and the real-time performance of health index extraction and early fault early warning are remarkably improved.

Description

Technical Field

[0001] This application relates to the technical field of mechanical equipment health management, and particularly to a fault early warning method and device. Background Art

[0002] In the current rapid development of intelligent manufacturing, the stable operation of equipment plays a decisive role in the efficient production of various industries. Mechanical failures not only lead to production interruptions and increased maintenance costs, but may also cause safety accidents in severe cases. Therefore, mechanical condition monitoring and fault early warning technologies have become the core means to ensure the reliable operation of equipment and have been widely applied in many fields.

[0003] Taking the automotive manufacturing industry as an example, a large number of high-precision and high-value mechanical equipment in the production line cooperate. Any equipment failure may cause the entire production line to paralyze. To ensure production continuity and product quality, enterprises deploy various sensors on the equipment to collect multi-dimensional data such as vibration, pressure, and temperature in real time, which is used as the basis for judging the health status of the equipment. In the energy industry, whether it is the steam turbine in a thermal power station or the drilling equipment in oil extraction, their stable operation is related to the safety and stability of energy supply. Through mechanical condition monitoring technology, potential equipment failure hidden dangers can be detected in advance, and maintenance plans can be reasonably arranged to avoid energy supply interruptions and huge economic losses caused by sudden failures.

[0004] Although mechanical condition monitoring technology has achieved remarkable results in its development process, currently, it still encounters difficult challenges. In terms of fault early warning, the existing early warning mechanisms usually trigger alarms only after obvious external signs of faults appear or the performance of the equipment has declined to a certain extent. This means that potential hidden dangers inside the equipment cannot be detected at an early stage, resulting in obvious delays in early warning. And this lag is very likely to cause the best opportunity for equipment maintenance and intervention to be missed, increasing the risk of equipment failure and maintenance costs.

[0005] Therefore, there is an urgent need for a fault early warning method to timely give fault early warnings for mechanical equipment. Summary of the Invention

[0006] Based on this, it is necessary to provide a fault early warning method for the above technical problems to timely give fault early warnings for mechanical equipment.

[0007] This specification adopts the following technical solutions:

[0008] This specification provides a fault early warning method, including:

[0009] Obtain the historical vibration data of the mechanical equipment;

[0010] Input the degradation data of historical data into the trained channel shuffle attention mechanism feature smoothing network model; the channel shuffle attention mechanism feature smoothing network includes: a convolutional layer, a max pooling layer, a channel shuffle module, a feature processing module, and a linear output layer;

[0011] In the convolutional layer, extract the preliminary features of the historical vibration data; in the max pooling layer, perform a pooling operation on the preliminary features to obtain the compressed preliminary features;

[0012] In the channel shuffle module, group the compressed preliminary features into multiple sub-features according to the channel dimension; perform global average pooling on each sub-feature, and through a simple gated mechanism with parameters and an activation function, obtain the shuffled channel attention weights; perform group normalization on each sub-feature, and through parameter-enhanced representation and an activation function, obtain the shuffled spatial attention weights; according to the shuffled channel attention weights and the shuffled spatial attention weights, perform a channel shuffle operation on all sub-features to obtain the shuffled features; repeat the convolutional operation, max pooling operation, and channel shuffle operation on the shuffled features to obtain the deep features of the historical vibration data;

[0013] In the feature processing module, flatten the deep features of the historical data, perform momentum update on the mean and covariance of the flattened features to obtain the running statistics of the historical vibration data, and according to the running statistics, calibrate the deep feature distribution of the historical data to obtain the smoothed features;

[0014] In the linear output layer, perform a linear operation on the smoothed features and output the health factors at multiple future moments;

[0015] Input the health factors at multiple future moments into the isolation forest warning model for anomaly detection to obtain the mechanical equipment fault alarm threshold and the first fault time.

[0016] Preferably, obtaining the historical vibration data of the mechanical equipment specifically includes:

[0017] Deploy vibration sensors in the X and Y directions on the test components of the mechanical equipment, and continuously obtain the original state data during the entire life cycle of the bearing from normal operation to failure;

[0018] Input the original state data at each moment into the normalization algorithm, and normalize the original state data at each moment through the normalization algorithm to obtain the state data at multiple moments, that is, the historical vibration data.

[0019] Preferably, training the channel shuffle attention mechanism feature smoothing network model includes:

[0020] Determine the degradation label division range of historical vibration data, and define the historical vibration data as multiple discrete values as the degradation labels of the historical vibration data. The quadratic degradation curve formula is:

[0021]

[0022] In the formula, t n is the maximum life of the device, and t i is the current time.

[0023] By rounding the quadratic degradation function value of the current time of the historical vibration data, the corresponding label value is obtained. The formula is:

[0024]

[0025] In the formula, Lable represents the label value of the historical vibration data, and [] is the rounding function.

[0026] Preferably, global average pooling is performed on each sub-feature, and the channel attention weight is obtained through a simple gating mechanism with parameters and an activation function. The formula is:

[0027] X' k1 = σ(F c (F gp (X k1 ))) × X k1 = σ(W1(F gp (X k1 )) + b1) × X k1 ;

[0028] In the formula, X' k1 is the channel attention weight after shuffling the sub-features, X k1 is the channel attention of the sub-feature, σ is the activation function, F gp (·) is global average pooling, and W1 and b1 are used to scale and translate the vector in the channel dimension.

[0029] Preferably, group normalization is performed on each sub-feature. After parameter-enhanced representation, the shuffled spatial attention weight is obtained through the activation function. The formula is:

[0030] X' k2 = σ(W2 × GN(X k2 )) + b2) × X k2 ;

[0031] In the formula, X' k2 is the channel attention weight after shuffling, X k2 is the spatial attention of the sub-feature, GN(X k2 ) represents performing group normalization on X k2Perform a normalization operation and linearly enhance the representation with parameters W2 and b2.

[0032] Preferably, smooth features are obtained, specifically including:

[0033] Calculate the mean and covariance of the flattened features to characterize the feature distribution of historical vibration data. The calculation formulas for the mean and covariance of the flattened features are:

[0034]

[0035] where, u i and V i are the mean and covariance of the flattened features respectively, N i is the total number of flattened feature samples, i represents the grouping number of the flattened feature samples, i = 1, 2, 3, …, z j represents the feature of the jth sample in the flattened features where j = 1, 2, 3, …, N i ;

[0036] Use the kernel symmetry statistic to perform weighted summation on the mean and covariance to obtain the smoothed mean and covariance. The smoothing formula is:

[0037]

[0038] In the formula, is the smoothed mean, is the smoothed covariance, k(y i , y i' ) is the kernel function, where y i is the predicted value of a sample in the historical vibration data, y i' is the predicted value of the remaining samples in the historical data, u i' is the mean of the remaining samples in the historical data, V i' is the covariance of the remaining samples in the historical data, and B is the number of groups of the historical vibration data;

[0039] Update the mean u i' and covariance V i' of the remaining samples in the historical vibration data by momentum. The update formulas are:

[0040] u i' ← αu i' + (1 - α)u i ;

[0041] V i' ← αV i' + (1 - α)V i ;

[0042] where α is the momentum coefficient;

[0043] The mean u of the remaining samples in the data after momentum update i' and the covariance V i' are the running statistics of the historical vibration data;

[0044] Calibrate the flattened features according to the running statistics of the historical vibration data, and the formula is:

[0045]

[0046] where is the calibrated feature.

[0047] Preferably, the shape of the preliminary feature is determined by the convolution kernel size, the stride size, and the padding size.

[0048] Preferably, the channel shuffle operation is to perform parallel operations on sub-features through multiple shuffle units.

[0049] Preferably, the isolation forest warning model includes multiple isolation trees; the specific steps of inputting the health factor into the isolation forest warning model for anomaly detection include:

[0050] Input the health factor into multiple isolation trees;

[0051] Calculate the expected isolation depth of each health factor in the isolation tree, and the formula is:

[0052]

[0053] where is the expected isolation depth of the health factor x in a tree, and c(n) is a constant related to the number n of input health factors, usually 2log2(n), which determines the maximum depth of the isolation tree;

[0054] Evaluate the anomaly degree at each moment according to the expected isolation depth, mark the anomaly moments, and determine the alarm threshold;

[0055] Determine the first failure time according to the chronological order of the occurrence of the anomaly moments.

[0056] The present invention also provides a health index extraction and early fault warning device, which is characterized by including:

[0057] An acquisition module, configured to acquire historical vibration data of mechanical equipment and divide it into a fixed length.

[0058] An extraction module, configured to generate a quadratic degradation curve of historical vibration data, divide the historical vibration data into labels according to the distribution trend of the quadratic degradation curve, perform feature learning according to time to obtain degradation data, and input the degradation data into a channel shuffle attention mechanism feature smoothing network model to obtain a health factor;

[0059] An early warning module, configured to input the health factor into an isolation forest early warning model for anomaly detection to obtain an anomaly label of the mechanical equipment, and obtain a mechanical equipment fault alarm threshold and a first fault time according to the anomaly label.

[0060] The above at least one technical solution adopted in this specification can achieve the following beneficial effects:

[0061] Through the constructed channel shuffle attention mechanism feature smoothing network model, the present invention generates shuffled channel attention and spatial attention through the channel shuffle module to shuffle the data features, and calibrates the shuffled features through the feature processing module, accurately captures the key features in the original vibration signal, extracts health factors with high consistency and high sensitivity, effectively reflects the change of the health state of the bearing, enables the health factor to quickly identify the sub-healthy state in the isolation forest model, so as to achieve efficient and timely early fault warning, solves the problem of lag in fault warning of mechanical equipment, thus detecting equipment anomalies in advance, gaining time for equipment maintenance, and reducing the losses caused by equipment failures. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The drawings described herein are used to provide a further understanding of the present application, and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation to the present application. In the drawings:

[0063] Figure 1 is a schematic flow chart of a fault early warning method provided by the present invention;

[0064] Figure 2 is a schematic diagram of the principle of the channel attention mechanism of a fault early warning method provided by the present invention;

[0065] Figure 3 is a schematic diagram of the principle of the feature processing module of a fault early warning method provided by the present invention;

[0066] Figure 4 is a schematic structural diagram of a channel shuffle attention mechanism feature smoothing network model of a fault early warning method provided by the present invention;

[0067] Figure 5 is a schematic diagram of the principle of isolation forest early warning of a fault early warning method provided by the present invention;

[0068] Figure 6 It is the structural block diagram of a health index extraction and early fault warning device provided by the present invention;

[0069] Figure 7 It is the schematic diagram of a computer device for an early fault warning method provided by the present invention;

[0070] Figure 8 It is the experimental bench diagram of the data set used for an early fault warning method provided by the present invention. Specific embodiments

[0071] To make the purpose, technical solutions and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this specification and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all of them. Based on the embodiments in the specification, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of this application.

[0072] The following will describe in detail the technical solutions provided by each embodiment of this application in conjunction with the drawings.

[0073] Figure 1 It is the flow schematic diagram of an early fault warning method in this specification, specifically including the following steps:

[0074] S101: Obtain the historical vibration data of mechanical equipment;

[0075] Deploy vibration sensors in the X and Y directions on the test components of mechanical equipment, and continuously obtain the original state data during the entire life cycle of the bearing from normal operation to failure;

[0076] Input the original state data at each moment into the normalization algorithm, normalize the original state data at each moment through the normalization algorithm, obtain the state data at multiple moments and divide them according to a fixed length to obtain the historical vibration data.

[0077] Specifically, the mechanical equipment can be an aeroengine, military equipment, industrial manufacturing equipment, transportation equipment, energy equipment, etc.

[0078] Specifically, the vibration data in the entire life cycle is expressed as X = [x1, x2,..., x LD T , and LD is the service life of the equipment.

[0079] Specifically, the normalization algorithm can be processed by the Min-max scaling method.

[0080] ​S102: Input the degradation data of historical data into the trained channel shuffle attention mechanism feature smoothing network model to obtain the health factors at multiple future moments; the channel shuffle attention mechanism feature smoothing network includes: a convolutional layer, a max pooling layer, a channel shuffle module, a feature processing module, and a linear output layer;

[0081] Among them, training the channel shuffle attention mechanism feature smoothing network model includes:

[0082] Determine the degradation label division range of historical vibration data, and define the historical vibration data as multiple discrete values as the degradation labels of the historical vibration data. The quadratic degradation curve formula is:

[0083]

[0084] In the formula, t n is the maximum life of the device, and t i is the current time.

[0085] By taking the integer of the quadratic degradation function value of the historical vibration data at the current time, the corresponding label value is obtained. The formula is:

[0086]

[0087] In the formula, Lable represents the label value of the historical vibration data, and [] is the integer function.

[0088] Among them, the obtaining of the health factors specifically includes: in the convolutional layer, extracting the preliminary features of the historical vibration data; in the max pooling layer, performing a pooling operation on the preliminary features to obtain the compressed preliminary features; in the channel shuffle module, grouping the compressed preliminary features into multiple sub-features according to the channel dimension; performing global average pooling on each sub-feature to obtain the channel attention weights; performing group normalization on each sub-feature to obtain the spatial attention weights; according to the channel attention weights and spatial attention weights, performing a channel shuffle operation on each sub-feature to obtain the shuffled features; repeating the convolutional operation, max pooling operation, and channel shuffle operation on the shuffled features to obtain deep features; in the feature processing module, flattening the deep features, performing momentum update on the mean and covariance of the flattened features to obtain the running statistics; calibrating the deep feature distribution according to the running statistics to obtain the smoothed features; in the linear output layer, outputting the health factors through linear operations.

[0089] Performing global average pooling on each sub-feature, and obtaining the channel attention weights through a simple gated mechanism with parameters and an activation function. The formula is:

[0090] X' k1 =σ(F c (Fgp (X k1 )))×X k1 =σ(W1(F gp (X k1 ))+b1)×X k1 ;

[0091] In the formula, X' k1 is the channel attention weight after sub - feature shuffling, X k1 is the channel attention of the sub - feature, σ is the activation function, F gp (·) is global average pooling, and W1 and b1 are used to scale and translate the vector in the channel dimension.

[0092] Perform group normalization on each sub - feature. After parameter - enhanced representation, obtain the shuffled spatial attention weight through the activation function. The formula is:

[0093] X' k2 =σ(W2×GN(X k2 )+b2)×X k2 ;

[0094] In the formula, X' k2 is the channel attention weight after shuffling, X k2 is the spatial attention of the sub - feature, GN(X k2 ) represents the normalization operation on X k2 , and it is linearly enhanced by parameters W2 and b2.

[0095] Obtain the smoothed feature, specifically including:

[0096] Calculate the mean and covariance of the flattened feature to characterize the feature distribution of historical vibration data. The calculation formulas for the mean and covariance of the flattened feature are:

[0097]

[0098] where, u i and V i are the mean and covariance of the flattened feature respectively, N i is the total number of flattened feature samples, i represents the grouping number of flattened feature samples, i = 1, 2, 3, …, z j represents the feature of the j - th sample in the flattened feature , j = 1, 2, 3, …, N i ;

[0099] Use the kernel symmetry statistic to perform weighted summation on the mean and covariance to obtain the smoothed mean and covariance. The smoothing formula is:

[0100]

[0101] In the formula, is the smoothed mean value, is the smoothed covariance, k(y i , y i' ) is the kernel function, where y i is the predicted value of a sample in the historical vibration data, y i' is the predicted value of the remaining samples in the historical data, u i' is the mean value of the remaining samples in the historical data, V i' is the covariance of the remaining samples in the historical data; B is the number of groups of the historical vibration data;

[0102] The momentum updates the mean value u i' and covariance V i' of the remaining samples in the historical vibration data, and the update formula is:

[0103] u i' ←αu i' +(1 - α)u i ;

[0104] V i' ←αV i' +(1 - α)V i ;

[0105] In the formula, α is the momentum coefficient;

[0106] The mean value u i' and covariance V i' of the remaining samples in the data after momentum update are the running statistics of the historical vibration data;

[0107] According to the running statistics of the historical vibration data, the flattened features are calibrated, and the formula is:

[0108]

[0109] In the formula, is the calibrated feature.

[0110] Specifically, the pre - processed data is input into the SAFS - Net module, and hyperparameters such as the weight coefficient λ, learning rate α, number of iterations T, and the module structure of the feature processing module are set. According to the multi - level degradation state feature distribution smoothing strategy, the health degree of the training set data is marked, and the data is input into the SAFS - Net network module in batches for training. During the training process, the model performs forward propagation on the batch data, calculates the gradient loss, and updates the parameters using the optimization algorithm until the set number of iterations is completed or the stop condition is met. The pseudo - code for training the SAFS - Net model is:

[0111]

[0112]

[0113] Specifically, the SAFS-Net module adopts an advanced and efficient channel attention mechanism. In the application scenario of deep convolutional neural networks, as a unique attention module, this channel attention mechanism has significant advantages. It can significantly improve the overall performance of the network while effectively reducing the computational cost. See Figure 2 , which is the schematic diagram of the channel attention mechanism. The specific operation process is as follows: First, for the input feature map Implement a grouping strategy along the channel dimension and divide it into several sub-features with a scale of . For each sub-feature, make full use of the shuffle unit for parallel processing. During this process, synchronously construct the channel attention X k1 and the spatial attention X k2 . Through this carefully designed mechanism, the model can simultaneously and keenly focus on the "content" and "position" information of the features, thereby achieving a more accurate understanding and processing of the input data.

[0114] Specifically, the data corresponding to each label shows an unbalanced state. To further improve the training effect of the model, the feature processing module of the present invention performs a feature distribution smoothing operation on the training samples. See Figure 3 , which is the schematic diagram of the feature processing module. This operation can optimize the data distribution to a certain extent and improve the stability and accuracy of model training. The feature processing module adopts the method of inserting a feature calibration layer after the final feature layer of the deep network. When training the features, the momentum update method is used to calculate the running statistics. Specifically, in each round of the training process, the smoothing statistics are fixed, while between different training rounds, the smoothing statistics are updated. In principle, feature distribution smoothing is based on such a setting: Since the target space is continuous, the feature space should also exhibit corresponding continuity. Based on this, it performs distribution smoothing on the feature statistics to calibrate the possibly biased feature distribution estimation. Especially for the target values with insufficient representation in the training data, this method can effectively improve this situation. Through the above operations, the model can more accurately learn the internal relationship between the features and the targets, enhancing the ability to process unbalanced data. Finally, in the deep unbalanced regression task, the overall performance of the model is significantly improved.

[0115] Specifically, see Figure 4 , which is the schematic diagram of the feature smoothing network model structure of the channel shuffle attention mechanism. The input is the original data, which is the original vibration signal collected from the bearing accelerated degradation test. Record a set of degradation data as X ∈ R N′L, where N represents the number of samples and L represents the sample length. In the feature extraction stage, the input data first enters the first convolutional module. In this module, features are initially extracted through convolutional operations to obtain The data shape here is determined by the convolutional kernel size K, the stride size S, and the padding size P. Subsequently, X1 undergoes a pooling operation through the max-pooling layer, and the compressed features are obtained after pooling Next, X2 is input into the channel shuffle module, and the features are obtained after the shuffle operation where C1 represents the number of channels and g is the number of groups for shuffling. To obtain deeper features, multiple convolutional, pooling, and shuffle operations are performed on the data. At different stages, the number of channels and the number of groups will change, such as settings of channel = 20, groups = 2; channel = 40, groups = 4; channel = 80, groups = 8, etc. In the feature processing and output stage, deep features X final are obtained after five convolutional operations. After flattening it, smoothing processing is performed through the feature processing module, and then through the linear layer (Linear), and finally the extracted health factors are obtained. In terms of model optimization, by optimizing the improved loss function L, the health factors obtained from model training are made to conform to the degradation trend of the health level as much as possible. The expression of the loss function is where y is the defined label and λ is a regularization coefficient used to control the influence of the weighted term on the loss function.

[0116] Specifically, the trained SAFS-Net module is used to test the test set data. The test set data should be in the same format as the training data and is the original vibration signal collected in the bearing accelerated degradation test where N test is the number of test samples, and L test is the length of the test samples. The prepared test set data X testFeed the trained SAFS-Net module. The data first enters the first convolutional module for preliminary feature extraction, and is operated according to the parameters such as the convolutional kernel size, stride, and padding set by the module to obtain preliminary features. The preliminary features are successively passed through the max pooling layer, channel shuffle module, etc., and multiple convolution, pooling, and channel shuffle operations are performed to continuously extract the deep features in the data. In this process, the number of channels and the number of groups in different stages change dynamically according to the module settings. After obtaining the deep features through five convolution operations, they are flattened, smoothed through the feature processing module, and then operated through the linear layer to finally output the health factor. For each sample in the test set, the corresponding health factor value is obtained according to the above process. The health factor values corresponding to all samples in the test set are connected in sequence according to the sample order to draw the health factor curve. This curve reflects the change of the health degree of the bearing in different states in the test set, and can be used to evaluate the performance of the model in practical applications, as well as to conduct subsequent analyses such as real-time monitoring and fault warning of the bearing operation status.

[0117] S103: Input the health factors at multiple future moments into the isolation forest warning model for anomaly detection to obtain the mechanical equipment fault alarm threshold and the first fault time.

[0118] Input the health factor into the isolation forest warning model for anomaly detection, specifically including: input the health factor into multiple isolation trees; calculate the expected isolation depth of each data point in the isolation tree, and the formula is:

[0119]

[0120] In the formula, is the expected isolation depth of the data point x in a tree, c(n) is a constant about the number n of input health factors, usually 2log2(n), which determines the maximum depth of the isolation tree; evaluate the anomaly degree at each moment according to the expected isolation depth, mark the anomaly moment, and determine the alarm threshold; determine the first fault time according to the time sequence of the occurrence of the anomaly moment.

[0121] Specifically, the isolation forest algorithm, as an efficient data processing and anomaly detection method, its core lies in adopting the way of random segmentation to process data. In actual operation, it orderly divides the entire data space into multiple subspaces by randomly selecting a feature and a segmentation value on this feature. This unique segmentation mechanism enables outliers to be isolated faster than normal data points due to their inherent sparsity characteristics. Based on such a principle, the isolation forest algorithm can effectively detect anomalies in the data.

[0122] When the obtained health factor data is input into the isolation forest early warning model, the model construction process officially starts. In this process, t isolation trees will be constructed. For each input data point x, the expected value E(H(x)) of its isolation depth h(x) is not obtained out of thin air, but jointly determined by the number of trees and the number of samples. The specific calculation formula is where represents the expected isolation depth of the data point x on a single tree, and c(n) is a function closely related to the number of input health factors. In the calculation process of this function, factors such as the quantity scale of the input data are fully considered, thus ensuring the accuracy and reliability of the calculation of the expected isolation depth.

[0123] During the operation of the model, it will strictly follow the principle of the isolation forest algorithm to conduct meticulous outlier identification on each input health factor data. Once a data point shows the characteristics of an outlier, the model will quickly mark it as an anomaly. These anomaly marks are not just simple identifications; they carry important information about the abnormal operation state of the device. Based on these anomaly marks, we can further determine the alarm threshold. The determination of the alarm threshold is a process that comprehensively considers various factors. It is not only related to the number and distribution of anomaly marks, but also closely related to factors such as the normal operation parameter range of the device and historical fault data. By scientifically and reasonably determining the alarm threshold, we can issue an alarm in a timely manner when the operation state of the device is abnormal, reminding relevant personnel to take corresponding measures. In addition, by deeply analyzing information such as the time sequence, frequency, and severity of the occurrence of anomaly marks, we can accurately obtain the time of the first fault. The determination of the time of the first fault is of crucial significance for the maintenance and fault prevention of the device. It can help us formulate maintenance plans in advance, reasonably arrange maintenance resources, and take effective measures before the device suffers a serious fault, thus avoiding problems such as production stagnation and economic losses caused by device failures. In summary, the process of inputting health factors into the isolation forest early warning model for anomaly detection is a complete and rigorous process from data input, model calculation, anomaly identification to threshold determination and fault time prediction. This process provides strong technical support for us to achieve the efficient operation and reliable maintenance of the device.

[0124] To fully verify the actual effectiveness of the health index extraction and early fault warning method proposed in the present invention, in the research of the present invention, the bearing degradation data set of the PHM 2012 Challenge was specifically selected to carry out the verification work of this method. The platform structure relied on by this experiment is as Figure 8As shown, it mainly consists of a rotating part, a degradation generation part (simulating the degradation process of the bearing by applying radial force), and a measurement part. It is worth mentioning that this platform can cause obvious degradation of the bearing within just a few hours. The data used in the experiment are all from the degradation process records of normal bearings (at the beginning of the experiment, the bearings have no defects) under different operating conditions. And the finally detected degraded bearings cover almost all types of common defects, such as ball defects, inner ring defects, outer ring defects, and cage defects, etc. In terms of the acquisition of vibration signals, two miniature accelerometers (model DYTRAN 3035B) placed perpendicular to each other at 90° are used. These two accelerometers are respectively installed on the vertical axis and the horizontal axis, and are radially installed at the outer ring position of the bearing. The sampling frequency of acceleration measurement is set at 25.6 kHz, and the vibration signal is collected every 10 seconds, with the signal duration of each collection being 0.1 second, and including data from both the horizontal and vertical channels. When the vibration acceleration reaches 20g, the experiment stops running, and at this time the bearing is considered to be in a failure state. In the experiment of this paper, the selected samples completely cover the entire life cycle of the bearing from normal operation to final failure. The specific experimental settings are shown in Table 1.

[0125] Table 1

[0126]

[0127] Build a network model and use the training sample set for network model training. The sample length is set to 2560, representing the amount of data contained in each sample to extract the operating state of the bearing. The batch size is 32, that is, 32 samples are selected for each training to update the parameters, which affects the stability and efficiency of training. The number of training epochs reaches 100, enabling the model to learn the training data set 100 times to optimize the parameters. The learning rate is 0.001, which determines the step size of parameter update and is conducive to model convergence. When using the Isolation Forest to achieve early fault warning, the number of constructed isolation trees is 100. In the construction process of each tree, the number of samples randomly selected from the data set is 64 to ensure the stability of threshold selection during calculation.

[0128] To verify the effectiveness of the health factors constructed by the proposed method under diverse operating conditions, it is crucial to understand the degradation characteristics of bearings. When a bearing is in a good working environment, its degradation process is stable, the service life is long, the signs of performance degradation are significant, it often shows a sub-healthy state, and the degradation process is more in line with the theoretical degradation curve. However, under harsh operating conditions, the bearing degradation is sudden, the service life is short and the state fluctuates greatly, and the occurrence of faults is often unexpected. In view of this, this study selects datasets under three different operating conditions, aims to examine the degree of fit between the extracted health factors and the original signals, and uses three indicators, Mon, Cor, and Rob, to quantitatively evaluate the health factors, so as to comprehensively measure the applicability of this method under different operating conditions.

[0129] The experimental results are shown in Table 2. The experimental results show that SAFS-Net effectively addresses the problem of sample imbalance through the feature distribution smoothing strategy and accurately captures key features using the channel shuffle attention mechanism. The extracted health factors can excellently characterize the health state of the bearing at each degradation stage. From Table 2, under the three operating conditions, it has good performance in indicators such as monotonicity, monotonicity, and robustness, and has significant advantages compared with other methods, which strongly proves the effectiveness of SAFS-Net in extracting health factors and the good applicability of this method in bearing health state monitoring.

[0130] Table 2

[0131] Operating condition Monotonicity Monotonicity Robustness Condition 1 0.8828 0.7897 0.9680 Condition 2 0.5400 0.7751 0.9384 Condition 3 0.2240 0.8450 0.8850

[0132] To verify the effectiveness of the threshold obtained by the proposed fault warning algorithm, the health factors extracted by SAFS-Net are used for early fault warning to achieve equipment health state monitoring. Early fault warning is of great significance for real-time monitoring of the bearing operation state, timely capturing potential fault signs, effectively preventing downtime losses caused by sudden equipment failures, and ensuring the continuity and stability of industrial production. By accurately setting the fault warning threshold, it is expected to issue an alarm when the bearing fault is still in its infancy, providing a sufficient time window for formulating scientific and reasonable maintenance strategies. The first warning time (FPT), as a key indicator of early fault warning, can effectively evaluate the effect of the fault prediction model.

[0133] The experimental results are shown in Table 3. For different bearings (Bearing1_1, Bearing1_2, Bearing2_7, Bearing3_3), there are obvious differences in their life lengths and the first warning times. Among them, the life length of Bearing1_1 is 2374, and the FPT is 2129; the life length of Bearing1_2 is 870, and the FPT is 713; the life length of Bearing2_7 is only 229, and the FPT is 10; the life length of Bearing3_3 is 433, and the FPT is 68. The Isolation Forest adopts the strategy of randomly splitting data, which can quickly identify and isolate outliers and performs excellently in the early fault warning of most working conditions. In most working conditions, it can accurately capture the abnormal signals of the bearing operation state and provide strong support for fault warning. In severe working conditions like Bearing2_7, the bearing life is extremely short, and the degradation process is rapid and complex. Although the iForest fails to accurately lock the true starting moment of the fault, due to its high sensitivity to abnormal data, it timely detects the drastic changes in degradation, realizes a practical early warning, reserves valuable time for maintenance decisions, and highlights its effectiveness and application value in the early fault warning under complex working conditions.

[0134] Table 3

[0135] Bearing1_1 Bearing1_2 Bearing2_7 Bearing3_3 Life length 2374 870 229 433 FPT 2129 713 10 68

[0136] The content of the present invention is mainly divided into two parts. The first part is the extraction of health factors, aiming to accurately obtain the key information reflecting the equipment health status from the collected mechanical operation data; the second part is the warning part, which uses the extracted health factors to build a model to timely discover the potential fault risks of the equipment and issue an alarm.

[0137] In the part of health factor extraction, first, a multi-level degradation state strategy based on the quadratic degradation curve is adopted to process the full life cycle data of the equipment, so that the model can be fully trained throughout the degradation cycle. Then, the SAFS-Net is used to perform deep feature extraction on the original vibration signal through operations such as convolution, pooling, and channel shuffling, so as to obtain health factors with high consistency and high sensitivity.

[0138] In the warning part, a warning model based on the Isolation Forest algorithm is built. The extracted health factors are input into the model. According to the principle of randomly splitting data, the model quickly identifies the outliers in the data, and then evaluates the abnormal degree of the equipment at each moment. Once an abnormality is detected, the abnormal points are marked and the alarm threshold is divided to achieve timely early warning of early faults.

[0139] The above is a method for early fault warning provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding device for health index extraction and early fault warning, such asFigure 6 as shown

[0140] Figure 6 The figure is a schematic diagram of a health index extraction and early fault warning device provided by the present invention. The device 600 includes:

[0141] A collection module 601, which acquires multiple groups of historical vibration data of a mechanical device under a certain working condition, divides them into samples of a fixed length, and then divides the training set and the test set according to groups;

[0142] An extraction module 602, which performs offline training and online testing of the model, marks the health degree of the training set data according to the multi-level degradation state strategy, inputs it into the SAFS-Net module in batches for training, and then sends the test set data into the trained SAFS-Net module to output a health factor curve;

[0143] An early warning module 603, which inputs the obtained health factor into an isolation forest early warning model for anomaly detection, and obtains an alarm threshold and the first fault time according to the anomaly mark.

[0144] For the specific limitations of a health index extraction and early fault warning device, reference can be made to the limitations of an early fault warning method in the above text, which will not be elaborated here. Each module in the above-mentioned health index extraction and early fault warning device can be implemented in whole or in part by software, hardware and their combination. The above-mentioned modules can be embedded in the processor of the computer device in the form of hardware or independent of it, or stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

[0145] The beneficial effects of the present invention on the method for extracting health indicators driven by sensors and early fault warning are as follows: By means of the vibration data collected by sensors, combining the deep feature mining ability of SAFS-Net and the efficient anomaly detection ability of the Isolation Forest algorithm, the original sensor data is transformed into the key basis for equipment health assessment and fault warning. Through the multi-level degradation state feature distribution smoothing strategy, it can effectively handle the sample imbalance problem in the bearing full life cycle data, perform accurate feature extraction and multi-state degradation state evaluation, and solve the problem that traditional methods are difficult to accurately describe the equipment degradation process. It can be seen from the experimental results that this method shows good adaptability and stability under different working conditions and complex operating environments. The extracted health indicators can be highly consistent with the changes in the original signal state in the face of various situations such as long-term stable degradation, sudden faults, and short-term stable degradation. The monotonicity, trend, and robustness of the quantitative indicators are significantly superior, and can accurately reflect the equipment health state. In addition, this method uses the channel shuffle attention mechanism to fully mine the potential health state information in the data, greatly improving the accuracy and effectiveness of feature extraction. Moreover, the early warning model constructed based on the Isolation Forest algorithm, after combining the extracted health factors, can quickly and accurately identify and issue an early warning when the equipment shows initial fault signs. Even in the stage when the fault characteristics are not obvious, it can effectively detect abnormalities. In the case of limited data volume, it can still maintain a high fault early warning accuracy and stability by virtue of its unique algorithm design, providing strong technical support for equipment health management and effectively reducing the losses caused by equipment failures. Although the implementation embodiments of the present invention are disclosed as above, they are not limited to the applications listed in the specification and the implementation embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the field, additional modifications can be easily made. Therefore, without departing from the general concept defined by the claims and the equivalent scope, the present invention is not limited to the specific details and the illustrated examples shown and described herein.

[0146] The present invention also provides a computer-readable storage medium, which stores a computer program that can be used to execute the above Figure 7 provided method for predictive maintenance of a mechanical device.

[0147] The present invention also provides Figure 7 the structural schematic diagram of the computer device shown in Figure 7 As shown, at the hardware level, this computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include other hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the above Figure 1 provided method for predictive maintenance of a mechanical device.

[0148] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0149] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope described in this specification.

Claims

1. A method for early warning of faults, characterized in that, Including: Obtain the historical vibration data of mechanical equipment; Input the degradation data of the historical data into the trained channel shuffle attention mechanism feature smoothing network model; The channel shuffle attention mechanism feature smoothing network includes: a convolutional layer, a max pooling layer, a channel shuffle module, a feature processing module, and a linear output layer; In the convolutional layer, extract the preliminary features of the historical vibration data; in the max pooling layer, perform a pooling operation on the preliminary features to obtain the compressed preliminary features; In the channel shuffle module, group the compressed preliminary features into multiple sub-features according to the channel dimension; perform global average pooling on each sub-feature, and obtain the shuffled channel attention weights through a simple gated mechanism with parameters and an activation function; perform group normalization on each sub-feature, and obtain the shuffled spatial attention weights through parameter enhancement representation and an activation function; according to the shuffled channel attention weights and the shuffled spatial attention weights, perform a channel shuffle operation on all sub-features to obtain the shuffled features; repeat the convolutional operation, the max pooling operation, and the channel shuffle operation on the shuffled features to obtain the deep features of the historical vibration data; In the feature processing module, flatten the deep features of the historical data, perform momentum update on the mean and covariance of the flattened features to obtain the running statistics of the historical vibration data, and calibrate the deep feature distribution of the historical data according to the running statistics to obtain the smoothed features; In the linear output layer, perform a linear operation on the smoothed features and output the health factors at multiple future moments; Input the health factors at multiple future moments into the isolation forest warning model for anomaly detection to obtain the mechanical equipment fault alarm threshold and the first fault time.

2. The fault early warning method according to claim 1, characterized in that, The obtaining of the historical vibration data of the mechanical equipment specifically includes: Deploy vibration sensors in the X and Y directions on the test components of the mechanical equipment, and continuously obtain the original state data during the entire life cycle of the bearing from normal operation to failure; Input the original state data at each moment into the normalization algorithm, and normalize the original state data at each moment through the normalization algorithm to obtain the state data at multiple moments, that is, the historical vibration data.

3. The early fault warning method according to claim 1, characterized in that, The training of the channel shuffle attention mechanism feature smoothing network model includes: Determine the degradation label division range of the historical vibration data, and define the historical vibration data as multiple discrete values as the degradation labels of the historical vibration data. The quadratic degradation curve formula is: where t n is the maximum lifespan of the device, and t i is the current time. Obtain the corresponding label value by taking the integer of the quadratic degradation function value of the historical vibration data at the current time. The formula is: In the formula, Lable represents the label value of the historical vibration data, and [] is the integer function.

4. The early fault warning method according to claim 1, characterized in that, The global average pooling is performed on each sub-feature, and the channel attention weights are obtained through a simple gated mechanism with parameters and an activation function. The formula is: X' k1 = σ(F c (F gp (X k1 ))) × X k1 = σ(W1(F gp (X k1 )) + b1) × X k1 ; where, X' k1 is the channel attention weight after sub-feature shuffling, X k1 is the channel attention of the sub-feature, σ is the activation function, F gp (·) is global average pooling, and W1 and b1 are used to scale and translate the vector in the channel dimension.

5. A method for early warning of faults according to claim 1, characterized in that, The group normalization is performed on each sub-feature, and after the parameter enhancement representation, the shuffled spatial attention weights are obtained by the activation function. The formula is: X' k2 = σ(W2 × GN(X k2 ) + b2) × X k2 ; where, X' k2 is the channel attention weight after shuffling, X k2 is the spatial attention of the sub-feature, GN(X k2 ) represents performing a normalization operation on X k2 and linearly enhancing it with parameters W2 and b2.

6. The early fault warning method according to claim 1, wherein, The obtaining of the smoothed features specifically includes: Calculate the mean and covariance of the flattened features to characterize the feature distribution of historical vibration data. The calculation formulas for the mean and covariance of the flattened features are as follows: Among them, u i and V i are the mean and covariance of the flattened features respectively, N i is the total number of flattened feature samples, i represents the flattened feature sample group number, i = 1, 2, 3, …, z j represents the feature of the j-th sample in the flattened feature , j = 1, 2, 3, …, N i ; Use the kernel symmetry statistic to perform weighted summation on the mean and covariance to obtain the smoothed mean and covariance. The smoothing formula is: In the formula, is the smoothed mean value, is the smoothed covariance, k(y i , y i' ) is the kernel function, where y i is the predicted value of a sample in the historical vibration data, y i' is the predicted value of the remaining samples in the historical data, u i' is the mean value of the remaining samples in the historical data, V i' is the covariance of the remaining samples in the historical data, and B is the number of groups of the historical vibration data; The mean u of the remaining samples in the momentum update history vibration data i' and the covariance V i' , and the update formula is: u i' ←αu i' +(1-α)u i ; V i' ←αV i' +(1 - α)V i ; where α is the momentum coefficient; The mean u of the remaining samples in the data after momentum update i' and the covariance V i' are the running statistics of the historical vibration data; Calibrate the flattened features according to the running statistic of the historical vibration data. The formula is: In the formula, is the calibrated feature.

7. The fault early warning method according to claim 1, wherein The isolated forest warning model includes multiple isolated trees; The step of inputting the health factor into the isolated forest warning model for anomaly detection specifically includes: Input the health factor into multiple isolated trees; Calculate the expected isolation depth of each health factor in the isolated tree. The formula is: wherein, is the expected isolation depth of the health factor x in a tree, and c(n) is a constant related to the number n of input health factors, usually 2log2(n), which determines the maximum depth of the isolation tree; Evaluate the anomaly degree at each moment according to the expected isolation depth, mark the anomaly moments, and determine the alarm threshold; Determine the first failure time according to the chronological order of the occurrence of the anomaly moments.

8. The fault early warning method according to claim 1, wherein The shape of the preliminary features is determined by the convolution kernel size, the stride size, and the padding size.

9. The fault early warning method according to claim 1, characterized in that The channel shuffle operation is to perform parallel operations on the sub-features through multiple shuffle units.

10. A device for early warning of faults, characterized in that, It includes: An acquisition module, which is used to obtain the historical vibration data of the mechanical equipment and divide it into a fixed length. An extraction module, which is used to obtain the secondary degradation curve of the historical vibration data, perform label division on the historical vibration data according to the distribution trend of the secondary degradation curve, and perform feature learning according to time to obtain the degraded data; And input the degraded data into the channel shuffle attention mechanism feature smoothing network model to obtain the health factor; A warning module, which is used to input the health factor into the isolated forest warning model for anomaly detection to obtain the mechanical equipment fault alarm threshold and the first failure time.

Images