A hoisting mechanism bearing fault diagnosis method based on causal feature learning

Abstract

The present application relates to hoisting mechanism bearing fault diagnosis technical field, especially in kind based on causal feature learning hoisting mechanism bearing fault diagnosis method. Including: the bearing source domain data is acquired;With bearing source domain data training deep learning fault diagnosis model;Acquire bearing target domain data;The bearing target domain data is input into the trained deep learning fault diagnosis model, and the bearing fault diagnosis result is output. The hoisting mechanism bearing fault diagnosis method based on the field generalization technology proposed in the present application can solve the problem that the original model cannot perform high-quality and high-efficiency fault diagnosis on new data due to working condition changes in real environment, a new fault diagnosis method is proposed, which can significantly reduce the data acquisition cost, greatly improve the safety and reliability of hoisting mechanism bearing operation.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology for hoisting mechanisms, and in particular to a method for diagnosing bearing faults in hoisting mechanisms based on causal feature learning. Background Technology

[0002] Lifting mechanisms are widely used in various types of lifting equipment and material handling systems to perform different tasks of vertically lifting and lowering heavy objects. In environments such as factories, construction sites, and ports, they are responsible for moving heavy materials or finished products and installing and relocating large structural components. These tasks typically require lifting mechanisms to safely and reliably handle loads of varying weights and sizes, ensuring that materials are accurately placed in designated locations. In this process, bearings, as key components of the lifting mechanism, play an indispensable role. They support and guide moving parts, reduce friction, and ensure smooth operation of the transmission system. Especially under high loads, the quality and performance of the bearings directly affect the stability and efficiency of the entire system. Healthy bearings can withstand pressure from all directions and adapt to different temperature changes and other harsh working conditions, ensuring stable operation of the equipment for extended periods. However, during long-term high-load use, especially under non-standard operating conditions, jogging phenomena frequently occur in lifting mechanisms, and bearings are prone to varying degrees of fatigue or early minor failures. Due to limitations in the bearing's installation location and the overall shape of the lifting mechanism, it is usually difficult to directly inspect the bearing's health condition. Current health management methods typically employ periodic replacement, meaning equipment is replaced at regular intervals. Unfortunately, even bearings from the same manufacturer and batch can have significantly different remaining service lives under varying loads and usage habits. If bearings are not replaced before their lifespan reaches its threshold, it can lead to increased mechanical vibration, affecting operational accuracy, potentially damaging the entire equipment, and even endangering the lives of on-site personnel. Conversely, replacing bearings well before they reach their remaining service life threshold results in severe resource waste and increased production costs. Therefore, effective condition monitoring and intelligent diagnostics for lifting mechanisms, especially their bearings, are crucial for ensuring operational safety and economic efficiency.

[0003] With the rapid development of technologies such as computers, sensors, artificial intelligence, embedded systems, and communications, fault diagnosis technology based on deep learning has advanced rapidly. However, for bearings in hoisting mechanisms, intelligent fault diagnosis technology still faces many challenges. Firstly, the number of fault samples for hoisting mechanism bearings is relatively small, and due to changes in operating conditions, there is significant domain drift between training and testing samples. For example, data collected under heavy load and low speed conditions, and the neural network trained based on this data, may not accurately determine the actual health status of light load and high speed samples. To address this issue, some scholars have proposed different diagnostic methods, such as using domain adaptation technology. This involves training a neural network to extract common features from source and target domain data and training these features to make them discriminative. Existing literature shows that this method can indeed alleviate the problem of reduced fault diagnosis accuracy caused by domain drift. However, it is important to note that this method can only extend diagnostic knowledge from source domain data to the target domain that can participate in training. Once the operating conditions of the target domain change again, the fault diagnosis accuracy based on the domain adaptation method will decrease to varying degrees, and retraining the network for newly collected data still has a certain lag. Secondly, domain-adaptive methods still require the collection of target domain data for training, and cannot perform online diagnosis of data from actual operation. This results in a certain lag in intelligent diagnosis, and the equipment often operates with defects.

[0004] The existing technology has the following drawbacks:

[0005] 1. Existing technologies cannot effectively utilize large and complex data, requiring operators to rely on their subjective experience to judge the health status of the hoisting mechanism bearings, resulting in low diagnostic efficiency and success rate.

[0006] 2. Fault diagnosis based on deep learning methods cannot meet reliability requirements in the case of data drift.

[0007] 3. Existing technologies rely on large amounts of labeled, high-quality datasets, but the cost of data collection is too high, and the collected data usually does not contain a large number of labels, resulting in low diagnostic accuracy of existing methods.

[0008] 4. Fault diagnosis based on domain-adaptive methods cannot cope with real-time data generated during industrial operation, and the diagnostic accuracy will still decrease to varying degrees once the operating conditions change.

[0009] 5. Traditional domain-based generalization methods rely solely on operational data and statistical information of their corresponding health status for modeling, without extracting causal information that determines the specific health status of the sample data, resulting in insufficient diagnostic accuracy. Summary of the Invention

[0010] This invention discloses a bearing fault diagnosis method for hoisting mechanisms based on causal feature learning. The specific method is as follows:

[0011] Obtain bearing source domain data;

[0012] A deep learning fault diagnosis model was trained using bearing source domain data;

[0013] Obtain bearing target domain data;

[0014] Input the bearing target domain data into the trained deep learning fault diagnosis model, and output the bearing fault diagnosis results.

[0015] Furthermore, when training a deep learning fault diagnosis model using bearing source domain data, a custom total loss function L is defined. Total Specifically, it includes:

[0016] Deep learning fault diagnosis model diagnostic accuracy evaluation item L ce ;

[0017] The similarity evaluation term L of source domain data before and after the introduction of interference signals sim ;

[0018] Before and after the introduction of interference signals, the data from the same source domain, after being processed by the deep learning fault diagnosis model, outputs a similarity evaluation term L in the feature dimension. dep ;

[0019] The similarity evaluation term L is obtained after feature extraction from the bearing source domain data and bearing target domain data by a deep learning fault diagnosis model. ang .

[0020] Furthermore, define a custom total loss function L ce The specific formula is as follows:

[0021]

[0022] In the formula, α, β, and γ are all hyperparameters, and N is the number of samples.

[0023] Furthermore, the evaluation term L for the diagnostic accuracy of the deep learning fault diagnosis model... ce The Softmax classification layer uses the cross-entropy loss function, and the specific formula is as follows:

[0024]

[0025] in, y is the probability distribution calculated from the model's raw output using the Softmax function. c R represents the true label, and R is the number of categories in the sample.

[0026] Furthermore, the source domain data similarity evaluation term Lsim The specific method for obtaining it is as follows:

[0027] Fourier transform of bearing source domain data;

[0028] Perturb the Fourier transformed bearing source domain data;

[0029] Perform inverse Fourier transform on the bearing source domain data after the disturbance operation;

[0030] The similarity between the original bearing source domain data and the bearing source domain data after inverse Fourier transform is calculated using the Pearson correlation coefficient, and is used as the source domain data similarity evaluation term L. sim .

[0031] Furthermore, the similarity evaluation term L in the output feature dimension is... dep The specific method for obtaining it is as follows:

[0032] Fourier transform of bearing source domain data;

[0033] Perturb the Fourier transformed bearing source domain data;

[0034] Perform inverse Fourier transform on the bearing source domain data after the disturbance operation;

[0035] The similarity of the output features of the fully connected layer of the deep learning fault diagnosis model before and after interference is calculated using the Wilson correlation coefficient for data from the same bearing source domain. This similarity is used as the L-value for the output feature dimension. dep .

[0036] Furthermore, the similarity evaluation term L is obtained by extracting features from the bearing source domain data and bearing target domain data using a deep learning fault diagnosis model. ang The specific method for obtaining it is as follows:

[0037] Feature vectors of bearing source domain data are extracted using a deep learning fault diagnosis model;

[0038] Feature vectors of bearing target domain data are extracted using a deep learning fault diagnosis model;

[0039] The similarity between two feature vectors is calculated using the angular distance metric formula. This similarity is then used as the L-value for feature extraction from the source and target domain data of the bearing using a deep learning fault diagnosis model. ang .

[0040] Furthermore, the bearing source domain data, obtained through bearing failure tests, is labeled data;

[0041] The bearing target domain data is obtained by sensors detecting the actual working site of the hoisting mechanism and is unmarked data.

[0042] Furthermore, the bearing failures of the hoisting mechanism include: inner ring failure and outer ring failure.

[0043] Due to the adoption of the above technical solutions, the present invention has the following beneficial effects:

[0044] 1. The hoisting mechanism bearing fault diagnosis method based on domain generalization technology proposed in this invention can solve the problem that the original model cannot perform high-quality and high-efficiency fault diagnosis on new data due to changes in working conditions in real-world environments. It proposes a new fault diagnosis method that can significantly reduce data acquisition costs and greatly improve the safety and reliability of hoisting mechanism bearing operation.

[0045] 2. The domain-generalized fault diagnosis method based on causal feature consistency extraction can extract the most essential features affecting the health status of the hoisting mechanism bearings, rather than statistical features. It no longer requires domain-invariant features to be absolutely invariant, but allows features to fluctuate in value, as long as the direction is as similar as possible. This is more in line with the actual operating logic of the hoisting mechanism bearings and results in high diagnostic accuracy.

[0046] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0047] The accompanying drawings of this invention are described below.

[0048] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0050] A method for diagnosing bearing faults in hoisting mechanisms based on causal feature learning, such as... Figure 1 As shown, the specific steps are as follows:

[0051] S1. Obtain bearing source domain data.

[0052] In this embodiment, the source domain dataset originates from laboratory bearing operation data collected and recorded using an accelerometer. This data is labeled. Based on the types of frequent bearing failures, it is divided into inner ring failure data and outer ring failure data. Operating data under normal conditions is also collected to form the source domain dataset.

[0053] S2. Train a deep learning fault diagnosis model using bearing source domain data.

[0054] In step S2, the main parameters of the deep learning fault diagnosis model are shown in the table below:

[0055]

[0056] When training a deep learning fault diagnosis model using bearing source domain data, a custom total loss function L is defined. Total Specifically, it includes:

[0057] S21, Evaluation item for the diagnostic accuracy of deep learning fault diagnosis model (L) ce .

[0058] Deep learning fault diagnosis model diagnostic accuracy evaluation item L ce The Softmax classification layer employs the cross-entropy loss function to calculate the health status classification loss of the source domain samples, as detailed below:

[0059]

[0060] in, y is the probability distribution calculated from the model's raw output using the Softmax function. c R represents the true label, and R is the number of categories in the sample.

[0061] S22. Source domain data similarity evaluation term L before and after introducing interference signal. sim ;

[0062] In order to extract the causal factors that determine the health status of a sample from the laboratory's operational data, ignore the interference of non-causal information on the model, and retain as much domain-specific information about the working state as possible so that the features extracted by the network are consistent with the changes in the working state of the hoisting mechanism, the causal information extraction module is first used to model the inherent properties of ideal causal information so that the extracted features have ideal causal properties.

[0063] Since the phase information of the bearing vibration signal after Fourier transform can better reflect the internal structural pattern of the mechanical system, it has higher stability and robustness, and can capture more complex nonlinear interaction effects. Therefore, it is believed that the phase signal can better characterize causal information, while the amplitude signal contains a large amount of low-level statistical information.

[0064] This embodiment relies on suppressing amplitude signals to highlight the role of phase signals in fault diagnosis, thereby extracting essential information characterizing the health status. To capture the relative temporal relationships and coordinated motion patterns among the components inside the hoisting mechanism bearing, reveal the essential characteristics of the bearing's operation, extract more high-level semantic features, and suppress statistical interference, the source domain data is transformed using Fourier interference as follows:

[0065]

[0066] Where, x 0 For the original sample, F represents the Fourier transform, A is the amplitude after the Fourier transform, and P is the phase signal after the Fourier transform.

[0067] Next, the source domain data is perturbed to generate lead samples, as shown in the following formula:

[0068]

[0069] Where x 0 ' represents a sample from another source domain, λ represents the degree of perturbation to the two source domain signals, and λ ~ (-1, 1). The amplitude of the newly generated data after its Fourier transform is represented, while the phase remains the same as the previous sample. Then, an inverse Fourier transform is performed on the newly generated data, as shown in the following formula:

[0070]

[0071] Among them, F -1 This is the inverse Fourier transform.

[0072] Next, the similarity between the new data and the original data is calculated using the Pearson correlation coefficient. This similarity is then used as the loss function to induce the deep learning fault diagnosis model to extract features, which can maximize the preservation of phase information and suppress amplitude information. Specifically:

[0073]

[0074] Where PEA(·,·) represents the calculation of the Pearson correlation coefficient, and N is the dimension of the features extracted by the neural network. These represent C under specific working conditions. j The original data and the newly generated data are processed in the high-dimensional space of the network using the MaxAbs Scaling method to obtain the feature values ​​of the i-th dimension, aiming to maximize the similarity between the features of the generated data and the features of the original data. The purpose of this formula is to enable the network to retain as many high-level physical features as possible related to the operating state of the hoisting mechanism bearings, while eliminating non-causal features.

[0075] S23. Before and after the introduction of interference signals, the data from the same source domain, after being processed by the deep learning fault diagnosis model, outputs a similarity evaluation term L in the feature dimension. dep .

[0076] Based on the fundamental properties of ideal causal features, the feature vectors of causal factors should be independent of each other. To simulate this property, the Pearson correlation coefficient method is used to calculate the similarity between different dimensions of the same vector before and after Fourier interferometry. This similarity function is then used as the loss function, aiming to minimize the feature similarity between different dimensions of the features extracted by the neural network, as detailed below:

[0077]

[0078] S24. Similarity evaluation term L after feature extraction from bearing source domain data and bearing target domain data by a deep learning fault diagnosis model. ang .

[0079] The operating signals of the hoisting mechanism bearings are determined by both the bearing model and the operating conditions. Existing transfer learning methods that use feature distribution difference metrics to extract domain-invariant features may overlook individual differences between different domains, thus limiting the performance of this approach. Therefore, this embodiment proposes a novel method that no longer emphasizes extracting absolute numerical consistency from personalized domains. Instead, it uses angular distance metrics to measure the similarity of features between two source domains, as detailed below:

[0080]

[0081] Among them, ||R A ||and||R B || represents the Euclidean norm of source domain A and source domain B, respectively. Its value ranges from [0, Π]. When its value is 0, it means that the characteristic directions of the two source domains are completely consistent, while Π represents that the directions are completely opposite.

[0082] S25. To ensure that all loss terms are on the same scale and to avoid some loss terms dominating the optimization process due to numerical differences, all losses are normalized. First, for the Pearson similarity loss, it is converted to 1-|r|, where r is the absolute value of the Pearson similarity. Second, for the diagonal distance loss, it is planned to [0,1] by dividing by Π.

[0083] S26. Using a weighted summation method, the three similarity functions are combined to obtain the causal feature loss function, as follows:

[0084]

[0085] Where α, β, and γ are hyperparameters, and N is the number of samples.

[0086] S27. The overall optimization objective is that the extracted features possess the basic properties of ideal causal features, namely, causal factors are independent of each other, causal factors can be separated from non-causal factors, and causal factors conform to the actual operating mechanism of the hoisting mechanism bearing. That is, different fields have both individualized factors and certain common characteristics. Therefore, the overall loss of the network is as follows:

[0087] L Total =L ce +nL cau

[0088] Where n is a hyperparameter that determines the degree of generalization to the domain, and after clarifying the above loss function, i.e., the optimization objective, let θ f θ c Let represent the parameters of the domain-shared feature extractor and the health status classifier, respectively. Then, the total loss function can be re-given by the following formula:

[0089]

[0090] The deep learning fault diagnosis model is trained using stochastic gradient descent and updates the network parameters through backpropagation. The stochastic gradient descent method is used to update θ. f θ c The formula for the process is as follows:

[0091]

[0092] Where ε is the learning rate.

[0093] S3. Obtain bearing target domain data.

[0094] In this embodiment, the target domain data is the actual operating data of the hoisting mechanism collected on-site using an accelerometer. The target domain data is automatically collected by the hoisting mechanism's moving equipment monitoring system. However, during data collection, the moving equipment monitoring system may be affected by wind disturbances, abnormal operations by on-site personnel, or abnormally bumpy working environments, causing the sensing equipment to collect noise signals of varying degrees. It may also be subject to electromagnetic interference or data loss during data transmission and storage, affecting the performance of the intelligent fault diagnosis model. Therefore, before inputting the data into the model, the bearing vibration signal is first cleaned to reduce noise and fill in missing values.

[0095] S4. Input the bearing target domain data into the trained deep learning fault diagnosis model and output the bearing fault diagnosis results.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for diagnosing bearing faults in hoisting mechanisms based on causal feature learning, characterized in that, The specific method is as follows: Obtain bearing source domain data; A deep learning fault diagnosis model was trained using bearing source domain data; Obtain bearing target domain data; Input the bearing target domain data into the trained deep learning fault diagnosis model and output the bearing fault diagnosis results. When training a deep learning fault diagnosis model using bearing source domain data, a custom total loss function is defined. Specifically, it includes: Evaluation Items for Diagnostic Accuracy of Deep Learning Fault Diagnosis Model ;in, This is the probability distribution calculated from the model's raw output using the Softmax function. R represents the true labels, and R is the number of categories in the sample. Similarity evaluation items of source domain data before and after the introduction of interference signals ; Before and after the introduction of interference signals, the data from the same source domain, after being processed by the deep learning fault diagnosis model, output similarity evaluation items in the feature dimension. ; After data from different source domains are processed by a deep learning fault diagnosis model, the output feature dimension similarity evaluation item is generated. ; Transform the source domain data: ;in, For the original sample, Represents Fourier transform, The amplitude after Fourier transform. The phase signal after Fourier transform; The source domain data is perturbed to generate induced samples. The specific formula is as follows: ;in, A sample representing data from another source domain. This represents the degree of perturbation to the signals from the two source domains, and ~(-1,1) The amplitude of the Fourier transform of the newly generated data remains the same as that of the previous sample; Perform an inverse Fourier transform on the newly generated data. The specific formula is as follows: ;in, This is the inverse Fourier transform; Source domain data similarity evaluation items The calculation formula is: in, This represents the calculation of the Pearson correlation coefficient, where N is the dimension of the features extracted by the neural network. , They represent specific working conditions. The original data and newly generated data are then processed in the high-dimensional space of the network using the MaxAbsScaling method. The goal is to maximize the similarity between the features of the generated data and the features of the original data after dimensional processing. Output feature dimension similarity evaluation item The Pearson correlation coefficient method is used to calculate the similarity between different dimensions of the same vector before and after Fourier interference, and this similarity function is used as the loss function. The goal is to minimize the feature similarity between different dimensions of the features extracted by the neural network, as detailed below: ; The similarity between two source neighborhood features is measured using angular distance as follows: ;in, and represents the Euclidean norm of source domain A and source domain B respectively; its value ranges from [0, Π]. When its value is 0, it means that the characteristic directions of the two source domains are completely consistent, while Π represents that the directions are completely opposite.

2. The bearing fault diagnosis method for hoisting mechanisms based on causal feature learning as described in claim 1, characterized in that, The bearing source domain data is obtained through bearing failure tests and is labeled data; The bearing target domain data is obtained by sensors detecting the actual working site of the hoisting mechanism and is unmarked data.

3. The bearing fault diagnosis method for hoisting mechanisms based on causal feature learning as described in claim 1, characterized in that, The bearing failures of the hoisting mechanism include: inner ring failure and outer ring failure.

Images