An online industrial bearing streaming data fault diagnosis method and system

Abstract

The application discloses an online industrial bearing streaming data fault diagnosis method and system, and relates to the technical field of mechanical fault diagnosis and intelligent maintenance. The method firstly performs local training on source domain data in the source domain, aligns the category features through a source domain prototype alignment module, so as to obtain discriminative feature prototypes. Then, a pre-trained model is deployed to the target domain, a dynamic threshold diagnosis module is applied to unlabeled online target data, the sample confidence is calculated according to multiple random MC dropout predictions, a multi-threshold strategy is adopted to dynamically classify and filter the samples, and reliable pseudo-labels are extracted for model updating. Meanwhile, a Gaussian mixture distance loss constraint is constructed based on the feature distribution of the source domain and the target domain, the cross-domain distribution difference is measured and reduced. The model is continuously optimized, the feature space can be effectively aligned without the source domain data, the accuracy and stability of bearing fault diagnosis are greatly improved, and the method is suitable for real-time fault monitoring and intelligent maintenance in industrial fields.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical fault diagnosis and intelligent maintenance, and particularly relates to an online industrial bearing streaming data fault diagnosis method and system. BACKGROUND

[0002] In the industrial fields of electric power, energy, high-end manufacturing, etc., rotating machinery is widely used in various continuous production processes as a key power equipment, and bearings, as its core components, directly determine the reliability, safety and production efficiency of the whole machine equipment. Therefore, timely and accurate fault diagnosis of rotating machinery bearings is of great significance for realizing predictive maintenance, avoiding unplanned downtime and ensuring production continuity. However, in real industrial operating environments, the fault diagnosis model faces a series of severe challenges from data distribution shift, signal streaming input and on-site deployment conditions from development to deployment.

[0003] Due to the influence of different equipment models, running loads, environmental conditions and installation methods, etc., the vibration signals collected from different working conditions often have significant differences in statistical distribution, i.e., there is a domain distribution mismatch problem. If the model trained offline under a certain source working condition is directly deployed in a new target working condition, the model performance will usually decrease significantly, and the fundamental reason is that the domain shift in the feature space makes it difficult for the model to achieve effective generalization.

[0004] To alleviate the field difference, traditionally, field adaptation or field generalization methods are used: field generalization attempts to construct a more robust model in the training stage to cope with unknown working conditions, however, when the equipment operating conditions continue to change dynamically, the offline model based on static assumptions is still difficult to maintain recognition accuracy in the long-term drifting working conditions; while the supervised domain adaptation method usually needs to access the source domain and target domain data at the same time, and aligns the distributions of the two domains explicitly to maintain the model performance, but in industrial field, due to the reality restrictions of data privacy, commercial secrets and transmission cost, etc., the source domain data often cannot be called in the deployment stage, which greatly limits the practical feasibility of this kind of domain adaptation method.

[0005] Industrial sites generally face the reality constraint of unavailability of source domain data. On the one hand, because the original training data may involve sensitive information such as equipment parameters and running processes, for the consideration of intellectual property rights and industrial safety, it is usually not allowed to leave the local environment; on the other hand, because the source domain data is large in volume, high in transmission and storage cost, it is difficult to frequently call in the edge side or on-site system. Under this background, the test-time adaptation technology without source data has attracted widespread attention in recent years, which only relies on the pre-trained source domain model and the real-time collected unlabeled data of the target domain, and fine-tunes the model parameters online in the inference stage, thereby improving the adaptation ability of the model to the target working condition without touching the source data, which has significant industrial application potential.

[0006] In addition, in real running scenarios, the monitoring data of industrial rotating machinery naturally has the characteristics of streaming, continuous, and non-stationary. The vibration signals continuously collected by sensors constitute a constantly evolving time series data stream, and the traditional offline training mode based on complete batch data cannot meet the needs of real-time diagnosis. The model must have the ability of online and batch-by-batch learning, and the parameters must be updated in time in the process of continuous data arrival. At the same time, since the working condition itself may slowly drift or even step change over time, the model also needs to have the ability to continuously adapt to new class distribution. In addition, in the industrial scene, it is difficult to perform expert labeling on a large amount of data, and the model must make full use of unlabeled data to achieve iterative evolution through self-supervised and self-training, which puts higher requirements on the stability and efficiency of the learning strategy.

[0007] To avoid the interference of false pseudo-labels on model updating, existing adaptive methods often use a fixed global confidence threshold, and only samples with a prediction confidence higher than the threshold are retained for self-training, while samples with a low confidence are directly discarded. However, in the real streaming industrial data scenario, this strategy has obvious defects: due to the change of data distribution over time, the sample confidence distribution often presents a scattered and non-uniform characteristic, and a single high threshold will exclude a large number of boundary samples with slightly lower confidence than the threshold but still have information, resulting in a serious shortage of effective samples for model updating.

[0008] In addition, the running process of rotating machinery has typical non-stationary characteristics, and the feature distribution of known fault types will continuously drift with running time, load state, and wear degree. More complexly, there may be fault categories that have never been seen in the training phase on site. That is, in the closed set diagnosis task, the model still needs to effectively respond to the distribution changes within the known classes; while in the open set scenario, the model should not only accurately identify known faults, but also have sensitivity and rejection ability to unknown abnormal patterns, to avoid misjudging new faults as known classes, thereby ensuring the reliability of the diagnosis conclusion and providing a reliable basis for operation and maintenance decisions.

[0009] In summary, existing rotating machinery fault diagnosis methods generally rely on source domain historical data and are difficult to fully utilize target domain unlabeled streaming monitoring data for rapid adaptive updating, and most of them are for closed set scenarios, lacking effective recognition ability for potential new fault patterns or new working condition patterns. SUMMARY

[0010] In view of the deficiencies of the prior art, the present application provides an online industrial bearing streaming data fault diagnosis method and system to solve the problem of model adaptation to target domain streaming data under the condition that source domain data is not available, which aims to fully exploit the unlabeled data in the target domain without using the original data in the source domain, realize continuous updating of the model during deployment, and improve the accuracy and stability of cross-domain fault diagnosis.

[0011] The technical scheme of the present application is:

[0012] An online industrial bearing streaming data fault diagnosis method, comprising the following steps:

[0013] Obtain source domain historical operation data, including samples of several bearing fault states and corresponding fault category labels and source domain identifiers;

[0014] Preprocess the source domain historical operation data to obtain preprocessed source domain historical operation data;

[0015] Construct a fault diagnosis model and use the preprocessed source domain historical operation data to perform source domain offline training on it to obtain a source domain model, and save the statistical representation of the source domain feature distribution obtained during the source domain offline training process; the fault diagnosis model is used to diagnose the fault of the input bearing fault state sample to obtain the fault category;

[0016] Obtain the vibration monitoring data of the target domain, and perform target domain online adaptation on the source domain model according to the statistical representation of the source domain feature distribution to obtain a final fault diagnosis model;

[0017] Obtain real-time vibration monitoring data and use the final fault diagnosis model to perform online fault diagnosis.

[0018] Further, the source domain historical operation data is represented as Each sample of the bearing fault state therein is a vibration signal collected by a vibration sensor installed on the bearing The source domain identifier is the running condition in the source domain , is the fault category label, represents the source domain, represents the number of bearing fault state samples, represents the number of fault categories, represents the total number of fault categories, represents the number of running conditions, represents the total number of running conditions.

[0019] Further, the preprocessing includes segmenting the source domain historical operation data according to a fixed time window and performing amplitude normalization processing.

[0020] Furthermore, the fault diagnosis model includes a feature extractor. and classifier The feature extractor consists of two parts. Used for input vibration signals Feature extraction is performed to obtain sample features; the classifier It is used to map sample features to fault category probabilities, thereby obtaining the predicted fault category.

[0021] Furthermore, a source domain prototype alignment module is introduced during the source domain offline training process to impose constraints on the sample feature distribution, including intra-class clustering constraints, inter-class separation constraints, and same-class prototype alignment constraints, thus constructing the total loss function for the source domain offline training process:

[0022] Cross-entropy loss As a classification loss:

[0023] ;

[0024] in, For cross-entropy loss, This represents the number of samples in the current batch. This indicates that the fault diagnosis model will use vibration signals. Predicted as fault category The probability; softmax is the softmax function;

[0025] In each iteration of the source domain offline training, the global prototype of each fault category in the current batch is calculated. This belongs to the fault category. The mean of the sample features:

[0026] ;

[0027] in, Fault Category A global prototype under all operating conditions in the source domain; This indicates that it belongs to the fault category. The number of samples; To indicate the first Sample features obtained from each sample by the feature extractor; Indicates sample number Belongs to the fault category Values ​​are taken from the sample;

[0028] For a source domain with multiple operating conditions, calculate the fault category for each. Different operating conditions The following local prototype:

[0029] ;

[0030] in, Fault Category Operating conditions The local prototype below; Represents the source domain Operating conditions The sample features obtained by the feature extractor from the samples below Indicates the operating conditions in the source domain. The following are the fault categories The number of samples; Indicates sample number Subordinate to the source domain Operating conditions Values ​​are taken from the following samples;

[0031] Based on this, the source domain prototype alignment module constructs intra-class aggregation constraints, inter-class separation constraints, and same-class prototype alignment constraints.

[0032] 1) Intra-class aggregation constraint: Calculate the sample features of each sample in each batch of the source domain. According to its fault category global prototype The intra-class clustering loss is obtained by averaging the squared Euclidean distances between samples and the squared Euclidean distances of all samples. :

[0033] (4);

[0034] in, This is the intra-class aggregation loss;

[0035] 2) Inter-class separation constraint: Set a distance threshold. Global prototypes for each pair of different fault categories and Calculate the Euclidean distance between the two and compare it with the distance threshold. Comparison, when the Euclidean distance is less than the distance threshold At that time, calculate the inter-class separation loss. :

[0036] (5);

[0037] in, The number indicating the fault category, Fault Category With fault category The corresponding Euclidean distance between global prototypes;

[0038] 3) Prototype alignment constraint: For the case that there are multiple operating conditions in the source domain, the local prototypes of a fault class under any two different operating conditions are calculated, and the average of all such distances is obtained as the prototype alignment loss , :

[0039] (6);

[0040] wherein is the prototype alignment loss, , is the number of operating conditions, represents the number of samples in the source domain belonging to the fault class under the operating condition ;

[0041] The above three losses are weighted and summed according to the weight coefficients to obtain the total SPA loss:

[0042] (7);

[0043] wherein represents the SPA loss, represents the weight coefficient;

[0044] The final loss of the source domain pre-training is the sum of the classification loss and the SPA loss:

[0045] (8);

[0046] wherein represents the final loss;

[0047] The parameter learning of the fault diagnosis model is completed by minimizing the final loss.

[0048] Further, the method for obtaining the statistical representation of the source domain feature distribution is:

[0049] According to the set of sample features obtained by the feature extractor from all samples in the source domain, the statistical representation of the source domain feature distribution is constructed, and a Gaussian mixture model is selected to fit the features of each feature dimension of the sample features. Assuming that the sample feature dimension output by the feature extractor is , then for each feature dimension ​​​respectively, and the Gaussian mixture model parameters with are estimated by using the expectation-maximization algorithm, where is the pre-set number of Gaussian components;

[0050] (9);

[0051] wherein, denotes the log-likelihood function of the Gaussian mixture model parameter set , and the Gaussian mixture model parameters in the feature dimension are estimated by using the expectation-maximization algorithm; is a Gaussian mixture model parameter set in a feature dimension, is the number of samples of the source domain used for estimating the statistical representation of the source domain feature distribution, is the mean parameter set of the source domain feature distribution, is the variance parameter set of the source domain feature distribution, is the mixing coefficient of the source domain Gaussian mixture model, is the number of Gaussian components, and are elements in and , respectively, and represent the mean and standard deviation of the th Gaussian component, is an element in , and represents the mixing coefficient of the th Gaussian component; is the component of the sample feature of the th sample obtained by the feature extractor in the th feature dimension, denotes the value of the one-dimensional Gaussian probability density function with the mean and the standard deviation at the component ;

[0052] Then, the Gaussian mixture model parameter sets in all dimensions are saved as the statistical representation of the source domain feature distribution.

[0053] Further, the vibration monitoring data of the target domain are obtained, and the source domain model is adapted to the target domain online according to the statistical representation of the source domain feature distribution, to obtain a final fault diagnosis model, and specifically includes:

[0054] S1: obtaining the vibration monitoring data of the target domain;

[0055] Specifically, the vibration monitoring data continuously generated by the target domain form a set , which is first preprocessed and then input into the source domain model in a streaming form for processing, For a vibration monitoring data, This is an index for the vibration monitoring data in the current small batch. Indicates the target domain;

[0056] S2: Introduce a dynamic threshold diagnosis module to filter vibration monitoring data in the target domain and generate pseudo-labels. Then, use the filtered vibration monitoring data and its pseudo-labels to update the source domain model and obtain the final fault diagnosis model.

[0057] Furthermore, S2 specifically includes:

[0058] S2.1: Calculate the teacher distribution for each vibration monitoring data point;

[0059] For a single vibration monitoring data Activate the dropout mechanism of the source domain model, that is, randomly discard a certain proportion of neuron connections during inference, and repeat the process. The next random forward computation yields... Individual prediction probability vectors , will this The average of the predicted probability vectors is used as the source domain model for vibration monitoring data. The comprehensive predicted distribution, i.e., the teacher distribution:

[0060] (10);

[0061] in, For vibration monitoring data The comprehensive predicted distribution For the index of the number of random forward computations, Indicates the first The predicted probability vector calculated in the next random forward pass; The number of times the algorithm is computed in the random forward pass;

[0062] S2.2: Determine the uncertainty index for each vibration monitoring data point based on the teacher distribution for each vibration monitoring data point;

[0063] Based on teacher distribution Information entropy is used to quantify prediction uncertainty, resulting in an uncertainty index. The calculation formula is:

[0064] (11);

[0065] in, As an uncertainty index, yes In the target domain The probability of each fault category the number of fault classes in the target domain, the total number of known fault classes in the target domain, the total number of fault classes in the target domain containing unknown fault classes;

[0066] S2.3: According to the uncertainty indicator of each vibration monitoring data, the vibration monitoring data is divided into high confidence class, medium confidence class and low confidence class;

[0067] In the target domain, two data-updating entropies thresholds are set: low threshold and high threshold , satisfying ;

[0068] First, the uncertainty indicator of the vibration monitoring data in the current batch is normalized to interval, obtaining the normalized uncertainty indicator ; then the vibration monitoring data is graded according to the size of the normalized uncertainty indicator : when , the vibration monitoring data is classified as high confidence class; when , it is classified as low confidence class; when , it is classified as medium confidence class;

[0069] The self-adaptive estimation is performed on the first , small batches in the initial stage of online adaptation. For a small batch, the normalized uncertainty indicators of all vibration monitoring data in the batch are collected and sorted in ascending order. Under the given target rejection ratio , the values of the uncertainty indicators at the and positions in the sorted sequence are taken respectively, denoted as the temporary entropy threshold and of the batch; subsequently, the temporary entropy thresholds of the first small batches are accumulated and the average value is calculated to obtain the final and , which are then fixedly used in the entire online adaptation process;

[0070] S2.4: For high confidence class, medium confidence class and low confidence vibration monitoring data of the target domain are set pseudo labels;

[0071] (12);

[0072] wherein, a pseudo label of the vibration monitoring data of the target domain is obtained, a hard pseudo label of the vibration monitoring data belonging to the high-confidence class is obtained, specifically, the vibration monitoring data is set as a hard pseudo label of the vibration monitoring data corresponding to the fault class with the maximum probability in the teacher distribution of the vibration monitoring data as the hard pseudo label of the sample; a soft pseudo label of the vibration monitoring data belonging to the medium-confidence class is obtained, specifically, the probability distribution after temperature scaling of the self-teacher distribution ; ;

[0073] S2.5: the source domain model is updated by using the vibration monitoring data of the high-confidence class and the medium-confidence class and the labels thereof, to obtain a final fault diagnosis model;

[0074] for the vibration monitoring data belonging to the high-confidence class, the vibration monitoring data and the hard pseudo label thereof are used as labeled samples to perform target domain online adaptation on the source domain model, and the forward supervision loss is:

[0075] (13);

[0076] wherein, the forward supervision loss is denoted as, the standard forward maximum likelihood of the student distribution is denoted as, and the prediction probability of the vibration monitoring data on the hard pseudo label is denoted as;

[0077] the soft cross-entropy loss between the standard forward maximum likelihood of the current student distribution of the source domain model and the soft pseudo label is calculated as :

[0078] (14);

[0079] (15);

[0080] wherein,​​ For soft cross-entropy loss, This is the weighting factor for the uncertainty of vibration monitoring data, with a value range of [value range missing]. Uncertainty index after normalization In the interval The inner linear decrease; Let be the standard forward maximum likelihood of the student distribution, representing the student distribution's likelihood of vibration monitoring data. In soft fake tags The predicted probability on the value; ReLU(·) is the activation function;

[0081] For those belonging to low confidence Vibration monitoring data of this type will not be used for gradient training, i.e., its pseudo-labels will be discarded.

[0082] Calculate high confidence Class and medium confidence The loss of vibration monitoring data can be used to obtain the current total self-supervised classification loss;

[0083] (16);

[0084] in, Total self-monitored classification loss;

[0085] Using the target domain feature set obtained from the current batch of vibration monitoring data through a feature extractor, Gaussian mixture model parameter estimation is performed to obtain a statistical representation of the target domain feature distribution. ;

[0086] In the case of Gaussian mixture, there is no simple closed-form solution. A weighted approximation is used to calculate the Gaussian mixture distance alignment loss:

[0087] (17);

[0088] in, The Gaussian mixture distance alignment loss is used. The distance is a Gaussian mixture. and These are the statistical representations of the source domain feature distribution and the target domain feature distribution, respectively. For the same feature dimension Above, source domain With the target domain The Middle The mean of the one-dimensional Gaussian corresponding to each Gaussian component; For the same feature dimension Above, source domain With the target domain The Middle The variance of a one-dimensional Gaussian corresponding to each Gaussian component; for the same feature dimension source domain and target domain the mixing coefficient of the first Gaussian component;

[0089] The total loss is:

[0090] (18);

[0091] wherein, is the total loss, is a weighting coefficient;

[0092] Finally, the total loss is used to perform gradient descent update on the source domain model parameters, completing the adaptive adjustment of the current batch. Then the next batch of vibration monitoring data is processed, and the cycle is repeated until the end of an epoch target domain data stream.

[0093] In another aspect, the present application also provides an online industrial bearing streaming data fault diagnosis system for implementing an online industrial bearing streaming data fault diagnosis method, comprising:

[0094] A data acquisition module is configured to acquire source domain historical operation data.

[0095] A preprocessing module is configured to preprocess the source domain historical operation data to obtain preprocessed source domain historical operation data.

[0096] A model construction and source domain offline training module is configured to construct a fault diagnosis model and train the model offline using the preprocessed source domain historical operation data to obtain a source domain model, and save the statistical representation of the source domain feature distribution obtained during the source domain offline training process.

[0097] A target domain online adaptation module is configured to acquire target domain vibration monitoring data and adapt the source domain model to the target domain according to the statistical representation of the source domain feature distribution to obtain a final fault diagnosis model.

[0098] A fault diagnosis module is configured to acquire real-time vibration monitoring data and perform online fault diagnosis using the final fault diagnosis model.

[0099] In a third aspect, the present application provides an electronic device comprising one or more processors and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the online industrial bearing streaming data fault diagnosis method.

[0100] In a fourth aspect, the present application provides a computer readable storage medium storing executable instructions, which when executed cause a processor to perform the online industrial bearing streaming data fault diagnosis method.

[0101] In a fifth aspect, the present application provides a computer program product comprising a computer program or instructions, which when executed by a processor implement the online industrial bearing streaming data fault diagnosis method.

[0102] Compared with the prior art, the present application has the following beneficial effects:

[0103] (1) Source data is independent, safe and efficient: the method of the present application does not need to access the original data of the source domain in the target domain adaptation process, and can realize knowledge transfer only by relying on source model parameters and a small amount of statistical features, solving the pain point of difficulty in obtaining source data in industrial field, meeting the requirements of data privacy and safety. At the same time, since long-term offline retraining is not required, only a single round of online update is needed to complete the adaptation, so the adaptation speed is fast, the computational overhead is low, and it is suitable for real-time deployment in the field.

[0104] (2) Make full use of unlabeled streaming data: the present application uses a dynamic threshold multi-level screening strategy to finely divide and utilize the unlabeled data coming in the stream. Compared with the fixed high threshold strategy, the amount of discarded data is greatly reduced, the number and diversity of pseudo-label samples are significantly improved, and the model can benefit from more target domain information. At the same time, different training strategies are adopted for different confidence samples, which ensures the reliable use of high confidence samples, and also obtains additional beneficial signals from medium confidence samples, improving the efficiency and effect of online adaptation.

[0105] (3) Prevent model drift and improve discrimination ability: with the help of the discriminative features learned by the source domain prototype alignment module, the model of the present application has strong intra-class cohesion and inter-class separation degree when starting to adapt, which lays a foundation for target domain alignment. In the adaptation process, the Gaussian mixture distance alignment (GMD) loss is introduced to continuously constrain the consistency of the target feature distribution and the source distribution, preventing the model from shifting the discriminant boundary or collapsing the features due to learning noise pseudo labels. Experimental results show that the present method can effectively avoid the degradation of model performance under long-term continuous training, and maintain high recognition rate for each known fault class.

[0106] (4) Support the rejection of unknown faults: Although the present application is aimed at closed set fault diagnosis (i.e. the target domain fault type is consistent with the source domain), the proposed dynamic threshold mechanism allows the model to "not guess recklessly" when the uncertainty is high, thereby detecting and rejecting potential unknown faults. When a fault beyond the training set range occurs, the model tends to give a high entropy value and low confidence, at which point the sample is rejected, avoiding the misjudgment of unknown faults as incorrect known categories, and improving the safety and reliability of the diagnosis results.

[0107] (5) Strong real-time adaptive ability: The method of the present application is designed for sensor data stream, which can realize continuous online learning, and the model can continuously run and update parameters in the target domain in a streaming manner, timely adapting to the gradual changes of working conditions.

[0108] In summary, the present application realizes an efficient and reliable online fault diagnosis method under closed set conditions through source domain local prototype alignment, target domain dynamic threshold screening and distribution alignment, etc., and has important industrial application value. BRIEF DESCRIPTION OF DRAWINGS

[0109] Figure 1 A flowchart of an online industrial bearing streaming data fault diagnosis method in an embodiment of the present application;

[0110] Figure 2 A general structure framework schematic diagram of an online industrial bearing streaming data fault diagnosis method in an embodiment of the present application;

[0111] Figure 3 A structure schematic diagram of a fault diagnosis model in an embodiment of the present application;

[0112] Figure 4 A structure schematic diagram of a source domain prototype alignment module in an embodiment of the present application;

[0113] Figure 5 A working flow schematic diagram of a dynamic threshold diagnosis module in an embodiment of the present application;

[0114] Figure 6 A source domain and target domain feature distribution alignment schematic diagram in an embodiment of the present application;

[0115] Figure 7 A different model classification feature visualization result graph when the adaptation type is close-set (closed set domain adaptation) in an embodiment of the present application;

[0116] Figure 8 A different model classification feature visualization result graph when the adaptation type is PDA (partial set domain adaptation) in an embodiment of the present application;

[0117] Figure 9A visualization result chart of different model classification features when the adaptation type is ODA (open set domain adaptation) in the embodiment of the application;

[0118] Figure 10 A visualization result chart of different model classification features when the adaptation type is OPDA (open partial set domain adaptation) in the embodiment of the application. DETAILED DESCRIPTION

[0119] The application will be described in detail below with reference to the drawings and embodiments.

[0120] Embodiment 1

[0121] In view of the prior art, a new type of diagnosis mechanism is needed, which does not depend on source domain data and can fully utilize target domain unlabeled streaming data for fast online adaptation. A balance should be struck between effectively mining data value and preventing the model from diverging due to noisy labels, and adaptation lag should be avoided due to excessive sample rejection. At the same time, the model must maintain the discriminability of known fault patterns during continuous learning and have sufficient robustness to changes in operating conditions and potential new patterns, so as to ultimately realize a high-precision, high-stability, industrial field-applicable online intelligent diagnosis system for rotating machinery bearings. To this end, the application proposes an online industrial bearing streaming data fault diagnosis method facing streaming data, without the participation of source domain data, which combines dynamic sample screening and stable model updating mechanism, as shown in Figure 1 and Figure 2 , comprising the following steps:

[0122] Step 1: Obtain source domain historical operation data;

[0123] The source domain historical operation data is represented as , which includes samples of several bearing fault states and corresponding fault class labels and a source domain identifier. Each sample of a bearing fault state is a vibration signal collected by a vibration sensor installed on the bearing , and the source domain identifier is the operating condition in the source domain , represents the source domain, represents the number of samples of the bearing fault state, represents the number of fault classes, represents the total number of fault classes, represents the number of operating conditions, represents the total number of operating conditions;

[0124] Step 2: Preprocess the source domain historical operation data to obtain preprocessed source domain historical operation data;

[0125] The preprocessing includes segmenting the source domain historical operation data into fixed time windows and performing amplitude normalization to obtain preprocessed source domain historical operation data for subsequent feature extraction and model training.

[0126] Step 3: Construct a fault diagnosis model and use the preprocessed historical operating data of the source domain to train it offline in the source domain to obtain the source domain model. At the same time, save the statistical representation of the source domain feature distribution obtained during the offline training process.

[0127] The fault diagnosis model includes a feature extractor. and classifier Two parts, such as Figure 3 As shown, the feature extractor Used for input vibration signals Feature extraction is performed to obtain sample features; the classifier It is used to map sample features to fault category probabilities, thereby obtaining the predicted fault category;

[0128] The feature extractor The deep neural network model adopts a one-dimensional convolutional neural network structure, specifically including a first convolutional layer (C1 layer), a second convolutional layer (C2 layer), and a flatten layer; each convolutional layer includes 1-dimensional convolution (Conv1d), an activation function (RELU), and average pooling (AvgPool).

[0129] The classifier It includes a first fully connected layer (F1 layer), a second fully connected layer (F2 layer), a Dropout layer (p=0.5), and a fourth fully connected layer (F4 layer) connected in sequence.

[0130] This invention, based on conventional training, introduces a Source Prototype Alignment (SPA) module to impose constraints on the sample feature distribution, such as... Figure 4As shown, the discriminant ability of the fault diagnosis model is enhanced in three aspects: first, the global prototype of each fault category (i.e., the mean of the features of all samples of the fault category) is calculated, and an intra-class aggregation constraint is imposed to make each sample feature close to the global prototype of the fault category to which it belongs, thereby reducing the intra-class difference; second, the distance between the global prototypes of different fault categories is calculated, and an inter-class separation constraint is imposed to require the distance between the global prototypes of different fault categories to be not less than a preset distance threshold, thereby widening the inter-class interval and avoiding a large number of difficult-to-classify samples near the decision boundary; and third, for the case where the source domain contains multiple operating conditions (subdomains), the local prototypes of the same fault category under different operating conditions are calculated, and a same-class prototype alignment constraint is imposed to make the local prototypes of the same fault category under different operating conditions close to each other, so as to eliminate the offset introduced by the operating condition difference. Through the above SPA strategy, the fault diagnosis model can learn a feature representation that has higher discrimination between fault categories and stronger robustness to operating condition changes.

[0131] During the source domain offline training process, different source domains can access each other, and high accuracy of the source domain fault categories is obtained through supervised learning;

[0132] The construction process of the total loss function of the source domain offline training process is as follows:

[0133] The cross-entropy loss is taken as the classification loss:

[0134] ;

[0135] wherein, is the cross-entropy loss, is the number of samples in the current batch, represents the probability that the fault diagnosis model predicts the vibration signal to be the fault category , represents the real fault category (the correct fault category corresponds to a component of 1, and the others are 0); and softmax is a softmax function;

[0136] In each source domain offline training iteration, the global prototype of each fault category in the current batch is calculated, i.e., the mean of the features of the samples belonging to the fault category :

[0137] ;

[0138] wherein, is the global prototype of the fault category under all operating conditions of the source domain; represents the global prototype of the fault category the number of samples in the source domain; represents the sample feature of the i-th sample in the source domain; represents the sample number from the samples in the source domain;

[0139] For the source domain with multiple operating conditions, the local prototypes of the fault category under different operating conditions are calculated respectively:

[0140] ;

[0141] wherein, is the local prototype of the fault category under the operating condition ; represents the sample feature of the sample in the source domain under the operating condition , represents the number of samples belonging to the fault category under the operating condition of the source domain ; represents the sample number from the samples in the source domain under the operating condition ;

[0142] The source domain prototype alignment module constructs the intra-class aggregation constraint, the inter-class separation constraint and the same-class prototype alignment constraint according to the above;

[0143] 1) Intra-class aggregation constraint: the Euclidean distance square between the sample feature of each sample in each batch of the source domain and the global prototype of the fault category to which the sample belongs is calculated, and the Euclidean distance squares of all samples are averaged to obtain the intra-class aggregation loss :

[0144] (4);

[0145] wherein, is the intra-class aggregation loss, which encourages the sample feature distribution of each fault category to shrink and gather, reduces the intra-class variance, and makes the samples of the same fault category form a compact cluster distribution in the feature space, thereby improving the robustness of the fault diagnosis model to unseen operating conditions. In practice, the sample feature can be approximated to the global prototype of the fault category to which it belongs by minimizing , and batch normalization technology can be combined to stabilize the training;​​​​

[0146] 2) Inter-class separation constraint: set a distance threshold (safety margin) , for each pair of different fault classes, compute the Euclidean distance between the global prototypes and , and compare it with the distance threshold , when the Euclidean distance is less than the distance threshold , calculate the inter-class separation loss :

[0147] (5) ;

[0148] where denotes the number of fault classes, is the Euclidean distance between the global prototypes corresponding to fault class and fault class ; , i.e. the distance threshold is not enough, when the Euclidean distance is greater than the distance threshold , it is not included in the loss; this forces the fault diagnosis model to pull apart the global prototypes of different fault classes, at least to the distance threshold , to increase the discriminant margin between fault classes; this measure effectively reduces the risk of sample confusion near the decision boundary, and provides a certain "buffer zone" when there are unknown fault classes, which is beneficial for subsequent open set rejection;

[0149] 3) Same-class prototype alignment constraint: for data with multiple operating conditions in the source domain, a fault class appears in the source domain operating condition set , calculate the distance between the local prototypes of this fault class under any two different operating conditions , , and , and take the average of all such distances to get the prototype alignment loss :

[0150] (6) ;

[0151] where is the prototype alignment loss, , is the number of operating conditions, denotes the number of samples in the source domain that belong to fault class under operating condition ;

[0152] Minimize the prototype alignment loss The local prototypes of the same fault category in different operating condition environments can be made close to each other, so that feature deviation caused by differences in sensor positions, operating conditions and the like is weakened. By aligning the local prototypes of each operating condition, the fault diagnosis model can learn fault category features that are independent of the operating condition type, and the cross-operating-condition migratability is improved.

[0153] The three losses are weighted and summed according to weight coefficients to obtain a total SPA loss:

[0154] (7);

[0155] wherein, represents the SPA loss, and the weight coefficient;

[0156] The final loss of the source domain pre-training is the sum of the classification loss and the SPA loss:

[0157] (8);

[0158] wherein, represents the final loss;

[0159] After the parameter learning of the fault diagnosis model is completed by minimizing the loss, the source domain model not only has a high classification accuracy on the training data, but also can identify the essential features of different fault categories in different operating conditions, which lays a good foundation for the adaptation of the source domain model to the target domain. After the training is completed, the source domain model parameters and the statistical representation of the source domain feature distribution will be used for the next step of target domain self-adaptation.

[0160] The method for obtaining the statistical representation of the source domain feature distribution is as follows:

[0161] According to the analysis of the above technical solution, this embodiment selects a Gaussian mixture model (GMM) to fit the features of each feature dimension of the sample features, assuming that the sample feature dimension output by the feature extractor is Then, a one-dimensional Gaussian mixture model (GMM) is established for each feature dimension The expectation maximization (EM) algorithm is used to estimate the Gaussian mixture model parameters with Gaussian components, wherein is a pre-set number of Gaussian components, which is only related to the fitting accuracy of the Gaussian mixture model (GMM) and is irrelevant to the feature dimension ;

[0162] (9);

[0163] wherein, denotes the log-likelihood function of the Gaussian mixture model parameter set is estimated by the Expectation Maximization (EM) algorithm; is a Gaussian mixture model parameter set for a feature dimension, is the number of samples of the source domain used to estimate the statistical representation of the source domain feature distribution, is the mean parameter set of the source domain feature distribution, is the variance parameter set of the source domain feature distribution, is the mixing coefficient of the source domain Gaussian mixture model, is the number of Gaussian components, and are and respectively denote the mean and standard deviation of the th Gaussian component, is the element in denotes the mixing coefficient of the th Gaussian component; is the component of the sample feature of the th sample in the th feature dimension after feature extraction, denotes the value of the one-dimensional Gaussian probability density function with mean and standard deviation at component ;

[0164] Then the Gaussian mixture model parameter set of all dimensions is saved as the statistical representation of the source domain feature distribution, which condenses the distribution information of the source domain feature space and is used to guide the target domain alignment. It should be noted that the parameters only reflect the aggregated statistical features and do not contain specific sample data, so they do not infringe the confidentiality and privacy of the source data.

[0165] Step 4: Obtain the vibration monitoring data of the target domain, and perform online adaptation of the source domain model according to the statistical representation of the source domain feature distribution to obtain the final fault diagnosis model;

[0166] The target domain adaptation stage includes two core parts: a dynamic threshold diagnosis module (DTD) and a Gaussian mixture distance alignment (GMD) loss calculation, which work together to realize the continuous updating and accurate diagnosis of the source domain model, including the following steps:

[0167] Step 4.1: Obtain the vibration monitoring data of the target domain;

[0168] Specifically: the source domain model and the statistical representation of the source domain features obtained by offline training of the source domain are deployed to the target device on site to start online fault diagnosis on the target domain data. A set of vibration monitoring data continuously generated by the target domain , first preprocessed (such as denoising, normalization, etc.), and then input into the source domain model in a stream form (batchsize) for processing, is a vibration monitoring data, is the index of the vibration monitoring data in the current mini-batch, represents the target domain;

[0169] Step 4.2: Introduce a dynamic threshold diagnosis module (DTD) to filter and generate pseudo labels for the vibration monitoring data of the target domain, and use the filtered vibration monitoring data and its pseudo labels to update the source domain model to obtain the final fault diagnosis model;

[0170] Inspired by teacher-student network training, the vibration monitoring data continuously generated by the target domain is divided into several batches. For each arriving batch, first enable Monte Carlo random dropout (MC-Dropout) to perform times of random forward calculation under the premise of keeping the network structure of the source domain model unchanged, obtain prediction probability vectors for each vibration monitoring data and take the average to form the teacher distribution of the vibration monitoring data, each prediction probability vector including the probability of the vibration monitoring data belonging to each fault category, and then the normal forward propagation of the source domain model outputs the student distribution of the vibration monitoring data;

[0171] Since the vibration monitoring data continuously generated by the target domain has no label, direct training needs to rely on a pseudo label strategy. Therefore, the dynamic threshold diagnosis module (DTD) takes the teacher distribution of each vibration monitoring data as the basis, on the one hand, uses its prediction probability vector to calculate the prediction entropy as an uncertainty indicator, and on the other hand, selects the fault category with the highest probability from the teacher distribution as the pseudo label candidate of the vibration monitoring data, to realize uncertainty evaluation and screening of the vibration monitoring data;

[0172] The specific working process of the dynamic threshold diagnosis module (DTD) is as follows:

[0173] Step 4.2.1: Calculate the teacher distribution of each vibration monitoring data;

[0174] For a single vibration monitoring data , activate the dropout mechanism of the model, that is, also randomly discard a certain proportion of neuron connections during inference, as shown in Figure 5 , repeat times of random forward calculation to obtain independent prediction probability vectors , will this The average of the predicted probability vectors is used as the source domain model for vibration monitoring data. The comprehensive predicted distribution, i.e., the teacher distribution:

[0175] (10);

[0176] in, For vibration monitoring data The comprehensive predicted distribution For the index of the number of random forward computations, Indicates the first The predicted probability vector calculated in the next random forward pass; The number of times the algorithm is computed in the random forward pass;

[0177] Step 4.2.2: Determine the uncertainty index for each vibration monitoring data point based on the teacher distribution for each vibration monitoring data point;

[0178] Based on teacher distribution (comprehensive predicted distribution) Information entropy is used to quantify prediction uncertainty, resulting in an uncertainty index. The calculation formula is:

[0179] (11);

[0180] in, As an uncertainty index, yes In the target domain The probability of each fault category This is the fault category number for the target domain. The total number of known fault categories in the target domain. The total number of fault categories containing unknown fault types in the target domain; if the uncertainty index A value close to 0 indicates a comprehensive predicted distribution. The source domain model is quite certain that it holds an absolute advantage in a certain fault category. It belongs to this fault category; otherwise, if the uncertainty index If the value is large, then the overall predicted distribution will be large. The source domain model, with its approximately uniform distribution across multiple fault categories, supports vibration monitoring data. The classification is rather uncertain;

[0181] Step 4.2.3: Based on the uncertainty index of each vibration monitoring data point, classify the vibration monitoring data into high-confidence categories. Class, medium confidence level Class and low confidence kind;

[0182] To facilitate adaptive filtering of vibration monitoring data with different confidence levels, this invention sets two entropy thresholds in the target domain that are updated with the data: a low threshold and a high threshold. and high threshold ,satisfy ;

[0183] First, the vibration monitoring data from the current batch Uncertainty index Normalization to The interval is used to obtain the normalized uncertainty index. Then, according to the normalized uncertainty index... The magnitude of the vibration monitoring data is used to classify the data: when At that time, vibration monitoring data was classified as high confidence level. Class; when At that time, it was classified as low confidence. Class; when At that time, it is classified as a medium confidence level. kind;

[0184] In order to enable the entropy threshold to automatically adjust with the overall uncertainty level, this invention adapts in the initial stage of online adaptation. On a small batch , Adaptive estimation is performed by collecting the normalized uncertainty index of all vibration monitoring data for a small batch. And sort them from smallest to largest, given the rejection rate of the target Under the given conditions, take the sorted sequence respectively and The value of the uncertainty index at location is denoted as the temporary entropy threshold for this batch. and Subsequently, regarding the former The temporary entropy thresholds of each small batch are accumulated and averaged to obtain the final value. and , , The value is automatically determined by the overall uncertainty level of the vibration monitoring data at that stage, rather than being pre-specified. This set of entropy thresholds is then used consistently throughout the online adaptation process to achieve high confidence levels for vibration monitoring data in subsequent batches. Class, medium confidence level Class and low confidence The classification of classes, since this invention was completed before... After adaptive verification of a small batch and Fixed no longer update, thus avoid excessive or too little vibration monitoring data is rejected due to the overall confidence fluctuations, make model screening strategy and the current target domain data characteristics match;

[0185] Step 4.2.4: For high confidence Class, medium confidence Class and low confidence Class vibration monitoring data set pseudo label;

[0186] (12);

[0187] Wherein, The vibration monitoring data of the target domain Pseudo label, The hard pseudo label of the vibration monitoring data Belonging to high confidence Class, specifically, the maximum probability Corresponding fault class in the teacher distribution Of vibration monitoring data As the hard pseudo label of the sample; The soft pseudo label of the vibration monitoring data Belonging to medium confidence Class, the probability distribution after temperature Scaling, Self-teacher distribution, ;

[0188] Step 4.2.5: using high confidence Class, medium confidence Class vibration monitoring data and its label, update the source domain model to obtain the final fault diagnosis model;

[0189] For vibration monitoring data Belonging to high confidence Class, vibration monitoring data and its hard pseudo label as labeled samples, online adaptation of the source domain model to the target domain, forward supervision loss Is:

[0190] (13);

[0191] Wherein, Indicates the forward supervision loss, The standard forward maximum likelihood of student distribution, indicates the prediction probability of vibration monitoring data On the hard pseudo label (fault class) Of student distribution; since the source domain model is more confident about these vibration monitoring data, the reliability of its pseudo label is high, and normal weight can be given during training;

[0192] For those belonging to the medium confidence level For vibration monitoring data of this class, instead of directly using a single class label, soft pseudo-labels are used. Then, the soft cross-entropy loss between the standard forward maximum likelihood of the current student distribution in the source domain model and the soft pseudo-labels is calculated. This encourages the source domain model output to converge towards its own average prediction.

[0193] (14);

[0194] (15);

[0195] in, For soft cross-entropy loss, This is the weighting factor for the uncertainty of vibration monitoring data, with a value range of [value range missing]. Uncertainty index after normalization In the interval The inner linear decreases when Approaching low threshold hour, The vibration monitoring data has a relatively large weight in the soft cross-entropy loss; when Approaching a high threshold hour, Therefore, its impact on the loss is weakened; Let be the standard forward maximum likelihood of the student distribution, representing the student distribution's likelihood of vibration monitoring data. In soft fake tags The predicted probability on; ReLU(·) is the activation function, guaranteeing that when At the same time, it dynamically suppresses the interference of samples near the low confidence boundary on the source domain model update. This method still assigns the relevant parameters to the model. The vibration monitoring data of this class is softly pseudo-labeled, but its loss weight is reduced, thus decreasing its contribution to gradient updates. Ensure that it belongs to... The vibration monitoring data of this type is partially utilized without causing drastic disturbances to the model's decision-making;

[0196] For those belonging to low confidence Vibration monitoring data of a certain type are not used in gradient training, meaning their pseudo-labels are discarded to avoid misleading the model. Most high-confidence samples are used for model updates, while only a small number of high-uncertainty samples are rejected. This achieves a balance between fully mining data stream information and ensuring the reliability of pseudo-labels, overcoming the shortcomings of traditional single-threshold methods that waste data due to overly strict screening or introduce noise due to overly lenient screening. Simultaneously, the source domain model can identify low-confidence samples... Vibration monitoring data of this type are considered abnormal inputs and are marked as unknown type in the final diagnostic output. This helps avoid misdiagnosis.

[0197] Calculate high confidence Class and medium confidence The loss of vibration monitoring data can be used to obtain the current total self-supervised classification loss;

[0198] (16);

[0199] in, Total self-monitored classification loss;

[0200] While implementing the aforementioned gating supervision, to suppress discriminative structure drift and catastrophic forgetting that may result from relying solely on pseudo-label learning, such as... Figure 6 As shown, using the target domain feature set obtained from the current batch of vibration monitoring data through a feature extractor, GMM parameter estimation (similar to the source domain offline step) is performed to obtain a statistical representation of the target domain feature distribution. ;

[0201] In the case of Gaussian mixtures, there is no simple closed-form solution. Referring to existing literature methods, a weighted approximation is used to calculate the Gaussian mixture distance alignment (GMD) loss:

[0202] (17);

[0203] in, The loss is Gaussian mixture distance alignment (GMD). The distance is a Gaussian mixture. and These are the statistical representations of the source domain feature distribution and the target domain feature distribution, respectively. For the same feature dimension Above, source domain With the target domain The Middle The mean of the one-dimensional Gaussian corresponding to each Gaussian component; For the same feature dimension Above, source domain With the target domain The Middle The variance of a one-dimensional Gaussian corresponding to each Gaussian component; For the same feature dimension Above, source domain With the target domain The Middle The mixing coefficients of the Gaussian components; the formula explicitly measures the difference in distribution between the source and the target on each feature dimension and averages it using the mixing coefficients; For the corresponding Gaussian component pairs, i.e., the source domain... The Gaussian component and the target domain Bhattacharyya distance of Gaussian components;

[0204] Calculated Subsequently, the total loss is formed by combining it with the total self-supervised classification loss. During the update process, the source domain model not only needs to fit the target pseudo-label but is also driven to maintain the discriminative structure of the source domain features, avoiding severe feature drift. The overall GMD alignment mechanism provides cross-domain anchoring when source data is unavailable. It utilizes the statistical representation of the source domain feature distribution to guide the evolution of target features, ensuring that the source domain model does not lose its original fault category discrimination ability while learning new domain knowledge. The total loss is:

[0205] (18);

[0206] in, For the total loss, As a tradeoff coefficient, it can be preset to a fixed value based on experience to balance the relative contributions of the loss from pseudo-label learning and distribution alignment.

[0207] Finally, using Gradient descent is performed to update the parameters of the source domain model, completing the adaptive adjustment for the current batch. Then, the next batch of vibration monitoring data is processed, and this cycle continues until the target domain data stream of one epoch is completed. Throughout the entire process, the original source domain data is never used; the update of the source domain model relies entirely on the inherent characteristics and statistical constraints of the unlabeled target domain data, thus achieving true source data isolation and online adaptation.

[0208] It should be noted that the target domain adaptation of this invention is a single-pass online mode, meaning that the target domain data is assumed to be processed only once, and the source domain will not be trained offline for multiple rounds. This strategy meets the needs of real-time monitoring of industrial processes and also ensures the efficiency of the method. Nevertheless, if the target domain operates for a long time, the source domain model can also continuously learn by applying the aforementioned DTD and GMD mechanisms to the continuously flowing new data, thereby tracking long-term distribution changes.

[0209] Step 5: Obtain real-time vibration monitoring data and perform online fault diagnosis using the final fault diagnosis model;

[0210] After the above online adaptation stage, the fault diagnosis model has been optimized and adjusted for the target domain environment. At this time, for the real-time data input of the target equipment, the fault diagnosis model can directly output the corresponding fault category diagnosis result. Overall, after the model is adapted by the method of the application, the known bearing fault types of the target domain can be identified with high accuracy, the state of the rotating mechanical equipment is accurately diagnosed and alarmed, and the unknown fault detection capability is also possessed. This provides a reliable technical means for equipment health management in industrial field.

[0211] Examples and performance verification:

[0212] The effectiveness of the method of the application is verified by the bearing fault data set CWRU in the streaming target domain scenario. In order to verify the alignment of the model to the known classes and the "rejection" effect to the unknown classes, use to represent the source domain label fault category set, and to represent the target domain label fault category set, the CWRU data set task is divided into four categories: close-set (closed set domain adaptation) represents that the source domain label and the target domain label are the same (i.e. the source domain and the target domain are the same) ), PDA (partial set domain adaptation) represents that the source domain has private classes (i.e. the source domain and the target domain are different) ), ODA (open set domain adaptation) represents that the target domain has private classes (i.e. the source domain and the target domain are different) ), and OPDA (open partial set domain adaptation) represents that the source domain and the target domain both have private classes, and the overall task constitutes as shown in Table 1.

[0213] Table 1: CWRU data set task composition

[0214] ;

[0215] The application is compared and evaluated with the existing representative advanced methods CoWA-JMDS, LEAD and GMM scheme based on Gaussian mixture modeling; in order to ensure fairness, all methods are offline initialized on the same feature extractor and source domain training data, and no target label or source domain original sample is used in the target domain stage, and small batch data stream arriving in time sequence is uniformly used for one epoch online processing. The evaluation index covers both known class recognition and unknown class rejection: for the close-set and PDA tasks which do not contain unknown classes, the known class recognition accuracy is used to measure the classification performance after cross-condition alignment:

[0216] (19);

[0217] wherein, is a true negative sample, is a true positive sample, is a false negative sample, is a false positive sample;

[0218] For the tasks ODA and OPDA (A3, A4) that the target domain contains unknown classes, the harmonic mean of the known class accuracy and the unknown class accuracy is given, and the H-score is used to evaluate the comprehensive performance of the classification model, which combines the classification accuracy and the misclassification rate, and the calculation formula is:

[0219] (20);

[0220] wherein, is the classification accuracy of the known fault class sample, indicating the classification effect of the model on the known fault class; is the classification accuracy of the unseen fault class sample, indicating the classification effect of the model on the unseen class;

[0221] The source domain model only represents that the pre-training model is not modified at all, and is directly deployed online on the target domain, so the source domain model only can not perform fault detection on the PDA, ODA and OPDA tasks. The experimental results are shown in Table 2, and compared with the source domain model without adaptation, the fault recognition accuracy of the present method is improved by about 20 percentage points. In addition, compared with the adaptation methods of CoWA-JMDS (method 1), LEAD (method 2) and GMM (method 3), the present method converges faster and has higher final accuracy due to the use of more medium confidence samples; it is shown that after introducing the GMD distribution alignment, the model can still maintain stable discrimination ability under long-time continuous adaptation, and there is no phenomenon of sharp decline in accuracy or model collapse. These results fully prove the effectiveness and robustness of the present application in practical application, and the visualization results of different model feature distributions are shown in Figures 7-10 .

[0222] Table 2 Experimental accuracy (%) of different models on CWRU data;

[0223] ;

[0224] In summary, the present application provides an innovative rotating machinery equipment bearing fault diagnosis method, which realizes fast and efficient self-adaptation of the model in the online scene without source data and labels through source domain prototype alignment pre-training, target domain dynamic threshold grading self-training and source-target distribution alignment, and improves the intelligent level and practical value of industrial equipment fault diagnosis.

[0225] Embodiment 2:

[0226] An online industrial bearing streaming data fault diagnosis system is used to realize an online industrial bearing streaming data fault diagnosis method, comprising:

[0227] The data acquisition module is configured to acquire source domain historical operation data.

[0228] The preprocessing module is configured to preprocess the source domain historical operation data to obtain preprocessed source domain historical operation data.

[0229] The model construction and source domain offline training module is configured to construct a fault diagnosis model, and perform source domain offline training on the fault diagnosis model by using the preprocessed source domain historical operation data to obtain a source domain model, and save a statistical representation of a source domain feature distribution obtained in a source domain offline training process.

[0230] The target domain online adaptation module is configured to acquire vibration monitoring data of a target domain, and perform target domain online adaptation on the source domain model according to the statistical representation of the source domain feature distribution to obtain a final fault diagnosis model.

[0231] The fault diagnosis module is configured to acquire real-time vibration monitoring data, and perform online fault diagnosis by using the final fault diagnosis model.

[0232] Embodiment 3

[0233] The embodiment provides an electronic device, which includes one or more processors and a memory. The memory is configured to store instructions, and when the instructions are executed by the one or more processors, the one or more processors perform the online industrial bearing streaming data fault diagnosis method.

[0234] The electronic device can be a mobile phone, a computer, a tablet computer, or the like, and includes a memory and a processor. The memory stores a computer program, and the computer program is executed by the processor to implement the online industrial bearing streaming data fault diagnosis method as described in the embodiments. It can be understood that the electronic device can further include an input / output (I / O) interface and a communication component.

[0235] The processor is configured to execute all or part of the steps of the online industrial bearing streaming data fault diagnosis method as described in the above embodiments. The memory is configured to store various types of data, which can include, for example, instructions of any application program or method in the electronic device, and application program related data.

[0236] The processor can be an Application Specific Integrated Cricuit (ASIC), a Digital Signal Processor (DSP), a Programmable Logic Device (PLD), a Field Programmable Gate Array (FPGA), a controller, a microcontroller, a microprocessor, or other electronic elements, which are used to execute the online industrial bearing streaming data fault diagnosis method described in the above embodiments.

[0237] Embodiment 4:

[0238] The embodiment provides a computer readable storage medium storing executable instructions, which, when executed, can be stored in one computer readable storage medium if implemented in the form of a software function unit and sold or used as an independent product.

[0239] The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the online industrial bearing streaming data fault diagnosis method described in various embodiments of the present application.

[0240] The aforementioned storage medium includes a flash memory, a hard disk, a multimedia card, a card-type memory (for example, an SD (Secure Digital Memory Card) or a DX (Memory Data Register, MDR) memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, a server, an APP (Application) application store, and various media capable of storing a program check code, on which a computer program is stored, which can implement the steps of the online industrial bearing streaming data fault diagnosis method described above when executed by a processor.

[0241] Embodiment 5:

[0242] The embodiment provides a computer program product comprising computer programs or instructions, which, when executed by a processor, implement the online industrial bearing streaming data fault diagnosis method.

[0243] Based on such understanding, the technical solution of the present application or the part of the technical solution which essentially contributes to the prior art or the part of the technical solution can be embodied in the form of a computer program product.

[0244] The embodiments in the present application are described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment mainly describes the difference from other embodiments.

[0245] The protection scope of the present application is not limited to the above-mentioned embodiments. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the scope and spirit of the present application. If these modifications and changes belong to the scope of the present application and equivalent technologies thereof, the present application also includes these modifications and changes.

Claims

1. An online industrial bearing streaming data fault diagnosis method, characterized in that, The method comprises the following steps: obtaining source domain historical running data, including samples of bearing fault states and corresponding fault category labels and source domain identifiers; preprocessing the source domain historical running data to obtain preprocessed source domain historical running data; constructing a fault diagnosis model and performing source domain offline training on the fault diagnosis model by using the preprocessed source domain historical running data to obtain a source domain model, and saving a statistical representation of source domain feature distribution obtained in the source domain offline training process; the fault diagnosis model is used for fault diagnosis on input samples of bearing fault states to obtain fault categories; in the source domain offline training process, a source domain prototype alignment module is introduced to impose constraints on sample feature distribution, including intra-class aggregation constraints, inter-class separation constraints and same-class prototype alignment constraints; obtaining target domain vibration monitoring data and performing target domain online adaptation on the source domain model according to the statistical representation of the source domain feature distribution to obtain a final fault diagnosis model, specifically, introducing a dynamic threshold diagnosis module to screen and generate pseudo labels for the target domain vibration monitoring data, and updating the source domain model by using the screened vibration monitoring data and the pseudo labels to obtain the final fault diagnosis model; obtaining real-time vibration monitoring data and performing online fault diagnosis by using the final fault diagnosis model.

2. The online industrial bearing streaming data fault diagnosis method according to claim 1, wherein, The source domain historical operation data is represented as Each sample of the bearing fault state is a vibration signal collected by a vibration sensor installed on the bearing The source domain is identified as an operating condition in the source domain , is a fault category label, represents the source domain, represents the number of samples of the bearing fault state, represents the number of fault categories, represents the total number of fault categories, represents the number of operating conditions, represents the total number of operating conditions.

3. The online industrial bearing streaming data fault diagnosis method of claim 1, wherein, The preprocessing comprises segmenting the source domain historical running data according to a fixed time window and performing amplitude normalization processing.

4. The online industrial bearing streaming data fault diagnosis method of claim 1, wherein, The fault diagnosis model comprises a feature extractor and a classifier Two parts, the feature extractor is used for feature extraction on the input vibration signal to obtain sample features; the classifier is used for mapping the sample features to fault category probabilities, and further obtaining predicted fault categories.

5. The online industrial bearing streaming data fault diagnosis method of claim 1, wherein, a total loss function of the source domain offline training process is constructed: cross-entropy loss as a classification loss: ； wherein, is a cross-entropy loss, is the number of samples in the current batch, denotes the probability that the fault diagnosis model predicts the vibration signal as a fault class ; softmax is a softmax function; In each source domain offline training iteration, compute global prototypes for each failure class in the current batch i.e. the sample feature mean belonging to the failure class : ； in, Fault Category A global prototype under all operating conditions in the source domain; This indicates that it belongs to the fault category. The number of samples; To indicate the first Sample features obtained from each sample by the feature extractor; Indicates sample number Belongs to the fault category Values ​​are taken from the sample; For the source domain with multiple operating conditions, the fault categories are calculated respectively under different operating conditions of the local prototype: ； in, Fault Category Operating conditions The local prototype below; Represents the source domain Operating conditions The sample features obtained by the feature extractor from the samples below Indicates the operating conditions in the source domain. The following are the fault categories The number of samples; Indicates sample number Subordinate to the source domain Operating conditions Values ​​are taken from the following samples; the source domain prototype alignment module constructs the intra-class aggregation constraints, the inter-class separation constraints and the same-class prototype alignment constraints accordingly; 1) Intra-class clustering constraint: compute the squared Euclidean distance between the sample feature of each sample in each batch of the source domain and the global prototype of the fault class it belongs to :​​​ （4）； wherein, is the intra-class clustering loss; 2) Inter-class separation constraint: set a distance threshold , for each pair of different fault classes and , calculate the Euclidean distance between them and compare it with the distance threshold , when the Euclidean distance is less than the distance threshold , calculate the inter-class separation loss : （5）； wherein, denotes a number of a failure class, is a failure class with a failure class the Euclidean distance between the corresponding global prototypes; 3) Alignment of prototypes of the same class: for the case where there are multiple operating conditions in the source domain, the local prototype of a fault class under any two different operating conditions is calculated, and the average of all such distances is obtained as the prototype alignment loss : ​​​​​ （6）； wherein, is a prototype alignment loss, , is a number of operating conditions, denotes a number of samples in the source domain that belong to the failure class under the operating condition . The above three losses are summed up according to weight coefficients The total SPA loss is obtained by weighted summation: （7）； wherein, represents the SPA loss, represents the weight coefficient; the final loss of the source domain pre-training is the sum of a classification loss and an SPA loss: （8）； wherein, represents the final loss; the parameter learning of the fault diagnosis model is completed by minimizing the final loss.

6. The online industrial bearing streaming data fault diagnosis method of claim 1, wherein, The statistical representation of the source domain feature distribution is obtained by: According to a set of sample features obtained by the feature extractor from all samples in the source domain, a statistical representation of the source domain feature distribution is constructed, a Gaussian mixture model is selected to fit the features of each feature dimension of the sample features, and the sample feature dimension output by the feature extractor is set as Then, for each feature dimension A one-dimensional Gaussian mixture model is established respectively, and an expectation maximization algorithm is used to estimate the Gaussian mixture model parameters with Gaussian components, wherein is a pre-set number of Gaussian components; （9）； in, Represents the parameter set of the Gaussian mixture model The log-likelihood function is used to estimate the parameters of the Gaussian mixture model on this feature dimension using the expectation-maximization algorithm; A Gaussian mixture model parameter set with one feature dimension. The number of samples in the source domain used to estimate the statistical representation of the source domain's characteristic distribution. This is the set of mean parameters of the source domain feature distribution. This is the set of variance parameters for the characteristic distribution of the source domain. The mixing coefficients of the source domain Gaussian mixture model are given. For the Gaussian components, and for and The elements in the array represent the first, second, third, fourth, fifth, sixth, seventh, eighth, and eleventh elements respectively. The mean and standard deviation of each Gaussian component. for The element in represents the first element. The mixing coefficient of the Gaussian components; For the first The sample features obtained by the feature extractor from the nth sample are in the nth sample. Components in each feature dimension The mean is Standard deviation is The one-dimensional Gaussian probability density function in components The value at; then, a set of Gaussian mixture model parameters of all dimensions is saved as the statistical representation of the source domain feature distribution.

7. The online industrial bearing streaming data fault diagnosis method of claim 1, wherein, The vibration monitoring data of the target domain is specifically a set composed of a plurality of vibration monitoring data continuously generated by the target domain , first pre-processed, and then input into the source domain model in a stream form for processing, is a vibration monitoring data, is an index of the vibration monitoring data in the current mini-batch, represents the target domain.

8. The online industrial bearing streaming data fault diagnosis method of claim 7, wherein, The dynamic threshold diagnosis module is introduced to screen and generate pseudo labels for the target domain vibration monitoring data, and the source domain model is updated by using the screened vibration monitoring data and the pseudo labels to obtain the final fault diagnosis model, specifically comprising: S1: calculating a teacher distribution of each vibration monitoring data; S2: determining an uncertainty index of each vibration monitoring data according to the teacher distribution of each vibration monitoring data; For single vibration monitoring data , the dropout mechanism of the source domain model is activated, that is, a certain proportion of neuron connections are randomly discarded during inference, and the random forward calculation is repeated times to obtain independent prediction probability vectors , and the prediction probability vectors are averaged as the comprehensive prediction distribution of the source domain model for the vibration monitoring data , that is, the teacher distribution: （10）； wherein, is the vibration monitoring data is the synthetic predictive distribution, is an index of the number of stochastic forward calculations, is the predictive probability vector of the stochastic forward calculation; is the number of stochastic forward calculations; in the Gaussian mixture case, there is no simple closed-form solution, and a weighted approximation calculation is adopted to calculate the Gaussian mixture distance alignment loss: The prediction uncertainty is quantified by the information entropy of the teacher distribution to obtain an uncertainty index , and the calculation formula is: （11）； wherein, is an uncertainty indicator, is is the probability of the target domain being in the th fault class, is the number of fault classes in the target domain, is the total number of known fault classes in the target domain, is the total number of fault classes in the target domain that contain unknown fault classes; S3: dividing the vibration monitoring data into high confidence class, medium confidence class, and low confidence class; Two entropy thresholds that are updated with the data in the target domain are set: a low threshold and a high threshold , satisfying ; First, the vibration monitoring data from the current batch Uncertainty index Normalization to The interval is used to obtain the normalized uncertainty index. Then, according to the normalized uncertainty index... The magnitude of the vibration monitoring data is used to classify the data: when At that time, vibration monitoring data was classified as high confidence level. Class; when At that time, it was classified as low confidence. Class; when At that time, it is classified as a medium confidence level. kind; In the early stages of online adaptation On a small batch , Adaptive estimation is performed by collecting the normalized uncertainty index of all vibration monitoring data for a small batch. And sort them from smallest to largest, given the rejection rate of the target Under the given conditions, take the sorted sequence respectively and The value of the uncertainty index at location is denoted as the temporary entropy threshold for this batch. and Subsequently, regarding the former The temporary entropy thresholds of each small batch are accumulated and averaged to obtain the final value. and This set of entropy thresholds was then used consistently throughout the online adaptation process; S4: high confidence class, medium confidence class and low confidence vibration monitoring data of the class sets a pseudo label; （12）； wherein, vibration monitoring data of the target domain a pseudo label, vibration monitoring data belonging to a high-confidence class a hard pseudo label of the vibration monitoring data, specifically, the vibration monitoring data a teacher distribution a maximum probability corresponding fault class as a hard pseudo label of the sample; vibration monitoring data belonging to a medium-confidence class a soft pseudo label of the vibration monitoring data, a self-teacher distribution a probability distribution after temperature scaling, ; S5: Utilize high confidence Class, medium confidence The vibration monitoring data of the class and the label thereof are used to update the source domain model to obtain a final fault diagnosis model. For the vibration monitoring data belonging to the high confidence class , the source domain model is adaptively targeted online with the vibration monitoring data and its hard pseudo labels as labeled samples, and the forward supervision loss is: （13）； wherein, represents the forward supervised loss, is the standard forward maximum likelihood of the student distribution, representing the probability of the vibration monitoring data on hard pseudo-labels predicted probability; computing a soft cross-entropy loss between the standard forward maximum likelihood of the current student distribution of the source domain model and the soft pseudo-labels : （14）； （15）； wherein, is a soft cross-entropy loss, is a weighting coefficient of the uncertainty of the vibration monitoring data, and the value range is ; the normalized uncertainty index is linearly reduced in the interval is a standard forward maximum likelihood of the student distribution, indicating the prediction probability of the vibration monitoring data on the soft pseudo label ; ReLU(·) is an activation function.​ For the vibration monitoring data belonging to the low-confidence class, it does not participate in gradient training, that is, its pseudo label is discarded. Computing high confidence Class and medium confidence Loss of vibration monitoring data for the class, the current total self-supervised classification loss can be obtained; （16）； wherein, is the total self-supervised classification loss; The target domain feature set obtained by the feature extractor using the current batch of vibration monitoring data is used to perform Gaussian mixture model parameter estimation to obtain a statistical representation of the target domain feature distribution ; the total loss is: （17）； wherein, is a Gaussian Mixture Distance Alignment loss, is a Gaussian Mixture Distance, and are the statistical representation of the source domain feature distribution and the statistical representation of the target domain feature distribution, respectively; is the mean of the one-dimensional Gaussian corresponding to the th Gaussian component in the source domain and the target domain , respectively; is the mean of the one-dimensional Gaussian corresponding to the th Gaussian component in the source domain and the target domain , respectively; is the variance of the one-dimensional Gaussian corresponding to the th Gaussian component in the source domain and the target domain , respectively; is the mixing coefficient of the th Gaussian component in the source domain and the target domain comprising: （18）； wherein is the total loss, is the trade-off coefficient; Finally, the total loss The gradient descent update is performed on the source domain model parameters, completing the adaptive adjustment of the current batch. Then the next batch of vibration monitoring data is processed, and the cycle continues until the end of an epoch of target domain data stream.

9. An online industrial bearing streaming data fault diagnosis system for implementing the online industrial bearing streaming data fault diagnosis method of any one of claims 1-8, characterized in that, a data acquisition module configured to obtain source domain historical running data; a preprocessing module configured to preprocess the source domain historical running data to obtain preprocessed source domain historical running data; a model construction and source domain offline training module configured to construct a fault diagnosis model and perform source domain offline training on the fault diagnosis model by using the preprocessed source domain historical running data to obtain a source domain model, and save a statistical representation of source domain feature distribution obtained in the source domain offline training process; a target domain online adaptation module configured to obtain target domain vibration monitoring data and perform target domain online adaptation on the source domain model according to the statistical representation of the source domain feature distribution to obtain a final fault diagnosis model; and a data acquisition module configured to obtain source domain historical running data; a preprocessing module configured to preprocess the source domain historical running data to obtain preprocessed source domain historical running data; a model construction and source domain offline training module configured to construct a fault diagnosis model and perform source domain offline training on the fault diagnosis model by using the preprocessed source domain historical running data to obtain a source domain model, and save a statistical representation of source domain feature distribution obtained in the source domain offline training process; a target domain online adaptation module configured to obtain target domain vibration monitoring data and perform target domain online adaptation on the source domain model according to the statistical representation of the source domain feature distribution to obtain a final fault diagnosis model; and A fault diagnosis module is configured to acquire real-time vibration monitoring data and perform online fault diagnosis by using the final fault diagnosis model.

10. A computer program product, characterised in that, The computer program or instructions are executed by a processor to implement the online industrial bearing streaming data fault diagnosis method of any one of claims 1-8.

Images