Time-varying rotating speed small sample fault diagnosis method based on depth residual error alignment

Abstract

The invention discloses a time-varying rotating speed small sample fault diagnosis method based on depth residual error alignment. The method comprises the following steps: acquiring first vibration data and corresponding first rotating speed data in a healthy state, and second vibration data and second rotating speed data in a fault state; processing the first vibration data to obtain a corresponding enhanced sample set, the enhanced sample set including a plurality of positive example enhanced sample data, and the plurality of positive example enhanced sample data being negative example enhanced sample data; inputting the positive example enhanced sample data into an extractor for self-supervised contrast learning training to obtain a trained extractor; training a regression device through the positive example enhanced sample data and the first rotating speed data to obtain a trained regression device; updating the trained regression device and classifier through the second vibration data and the second rotating speed data; and acquiring real-time vibration data and real-time rotating speed data, and inputting the real-time vibration data and the real-time rotating speed data into the trained extractor, the regression device and the classifier to obtain a fault diagnosis result. According to the time-varying rotating speed small sample fault diagnosis method based on depth residual error alignment, adaptive alignment can be carried out on characteristics of characteristic signals under each time-varying rotating speed working condition, and the consistency capability of a model for processing vibration signals under continuous change working conditions is improved; and the specific health state of the rotating machine can be effectively discriminated under the condition that small sample data and the rotating speed working condition change along with time.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a time-varying speed small sample fault diagnosis method based on deep residual alignment. Background Art

[0002] The problem of mechanical equipment fault diagnosis has long been constrained by the sparse and low-quality fault data in actual industrial application scenarios and has not been able to achieve substantial development and application. In addition, complex and changeable operating conditions are another major problem facing fault diagnosis technology at this stage. Large rotating machinery is often accompanied by continuous changes in speed during actual operation. Large changes in speed will cause the amplitude of the vibration signal and the energy generation level to jump, thus making many fault diagnosis models invalid. Therefore, how to achieve fault diagnosis under time-varying speed conditions and small sample conditions is a scientific and application problem that needs to be solved urgently.

[0003] As a major branch of data-driven fault diagnosis technology, deep learning-based methods and self-supervised learning are deep learning models of deep learning, which have powerful fitting and induction capabilities for massive measured data. The deep learning-based model can learn the characteristic distribution of various faults from uniformly sampled training data, and perform reasoning and health judgment on the newly generated vibration data based on the learned model. However, at present, the fault diagnosis problem under time-varying speed is difficult to adapt to various traditional deep learning models due to the interference coupling between the continuous and variable working conditions and the fault categories, resulting in a sharp decline in effectiveness and robustness. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a time-varying speed small sample fault diagnosis method based on deep residual alignment, which can adaptively align the characteristic signal features under time-varying speed conditions, improve the consistency of the model's processing of vibration signals under continuously changing conditions, and effectively identify the specific health status of rotating machinery under small sample data and conditions where the speed conditions change over time.

[0005] The present invention provides a time-varying speed small sample fault diagnosis method based on deep residual alignment, comprising the following steps:

[0006] Acquire first vibration data and corresponding first rotation speed data in a healthy state, and second vibration data and second rotation speed data in a fault state;

[0007] Processing the first vibration data to obtain a corresponding enhanced sample set, wherein the enhanced sample set includes a plurality of positive enhanced sample data, and the plurality of positive enhanced sample data are mutually negative enhanced samples;

[0008] Inputting the positive enhanced sample data into the extractor for self-supervised contrastive learning training to obtain a trained extractor;

[0009] Training a regressor using the positive enhanced sample data and the first speed data to obtain a trained regressor;

[0010] Updating the trained regressor and classifier by using the second vibration data and the second speed data;

[0011] Real-time vibration data and real-time rotation speed data are acquired and input into the trained extractor, the regressor and the classifier to obtain a fault diagnosis result.

[0012] In one embodiment, the processing of the first vibration data to obtain a corresponding enhanced sample set, wherein the enhanced sample set includes a plurality of positive enhanced sample data, and the plurality of positive enhanced sample data are mutually negative enhanced sample data, further includes:

[0013] Segmenting the first vibration data to obtain a one-dimensional vibration signal sample set;

[0014] Each sample in the one-dimensional vibration signal sample set is subjected to a negation operation and a flipping operation to obtain an enhanced sample set.

[0015] In one embodiment, the step of inputting the positive enhanced sample data into an extractor for self-supervised contrastive learning training to obtain a trained extractor further comprises:

[0016] Inputting the positive enhanced sample data and the corresponding negative enhanced sample data into an extractor to obtain a first characterization feature;

[0017] Inputting the first characterization feature into a projector to map the first characterization feature to a latent space by nonlinear mapping, thereby obtaining a positive latent feature vector and a negative latent feature vector;

[0018] The representation similarity between the positive example latent feature vector and the negative example latent feature vector is calculated by ratio learning loss, wherein the representation similarity includes the negative example latent feature similarity and the positive example latent feature similarity, and the training of the extractor and the projector is guided in a mode of maximizing the negative example latent feature similarity and minimizing the positive example latent feature similarity to obtain the trained extractor and the projector.

[0019] The time-varying speed small sample fault diagnosis method based on deep residual alignment provided by the present invention can adaptively align the characteristic signal features under time-varying speed conditions, improve the consistency of the model's processing of vibration signals under continuously changing conditions, and effectively identify the specific health status of rotating machinery under small sample data and conditions where the speed conditions change over time. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0021] Figure 1 A schematic flow chart of the time-varying speed small sample fault diagnosis method based on deep residual alignment provided by the present invention.

[0022] Figure 2 Schematic diagram of self-supervised comparative training of an extractor and a projector according to an embodiment of the present invention.

[0023] Figure 3 4 is a schematic diagram of nonlinear mapping training of a regressor according to an embodiment of the present invention.

[0024] Figure 4 It is a schematic diagram of small sample fault classification training based on residual vector according to an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The specific embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the description of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] In the description of the present invention, unless otherwise clearly specified and limited, the terms "set", "install", "connection" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0027] The directions or positional relationships indicated by terms such as “upper”, “lower”, “left”, “right”, “front”, “back”, “top”, “bottom”, “inside” and “outside” are based on the directions or positional relationships shown in the accompanying drawings, or are the directions or positional relationships in which the inventive product is usually placed when used. They are only for the convenience of description and simplified description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as a limitation on the present invention.

[0028] The terms "first", "second", "third", etc. are merely used to distinguish elements of similar nature, and do not indicate or imply relative importance or a particular order.

[0029] The terms "comprises," "comprising," or any other variations thereof, are intended to cover a non-exclusive inclusion of the elements listed and may also include additional elements not expressly listed.

[0030] See also Figure 1 The time-varying speed small sample fault diagnosis method based on deep residual alignment provided by the present invention comprises the following steps:

[0031] S1, obtaining first vibration data and corresponding first rotation speed data in a healthy state, and second vibration data and corresponding second rotation speed data in a fault state.

[0032] It can be understood that the healthy state refers to the state where the mechanical equipment is fault-free, and correspondingly, the fault state refers to the state where the mechanical equipment is faulty. The second vibration data and the second speed data can be data under the historical fault state of the mechanical equipment.

[0033] S2, processing the first vibration data to obtain a corresponding enhanced sample set, wherein the enhanced sample set includes a number of positive enhanced sample data, and the number of positive enhanced sample data are each other's negative enhanced sample data.

[0034] It is understandable that the above steps may further include:

[0035] S201 , segmenting first vibration data to obtain a one-dimensional vibration signal sample set.

[0036] It can be known that segmenting the first vibration data can better facilitate subsequent analysis of the fault data.

[0037] S202 , performing a negation operation and a flipping operation on each sample monomer in the one-dimensional vibration signal sample set to obtain an enhanced sample set.

[0038] It can be known that the formula for the flip operation can be expressed as:

[0039]

[0040] in, is the first vibration data after flipping, is a data point of the first vibration data, and 1 to n are the corresponding vibration signal lengths respectively.

[0041] The formula for negation operation can be expressed as:

[0042]

[0043] in, is the first vibration data after negation, is a data point of the first vibration data, and 1 to n are the corresponding vibration signal lengths respectively.

[0044] Mutually negative enhanced samples can be understood as one of them being positive enhanced sample data, and the rest being negative enhanced sample data of the positive enhanced sample data.

[0045] See also Figure 2 ,S3, input the positive enhanced sample data into the extractor for self-supervised contrastive learning training to obtain a trained extractor.

[0046] It is understandable that the above steps may further include:

[0047] S301, inputting positive enhanced sample data and corresponding negative enhanced samples into an extractor to obtain a first representation feature;

[0048] It can be understood that the positive enhanced sample data and the corresponding negative enhanced samples are input into the extractor, and the first representation features of the positive enhanced sample data and the corresponding negative enhanced samples in the feature space can be obtained.

[0049] S302, inputting the first characterization feature into a projector to map the first characterization feature to a latent space by nonlinear mapping, thereby obtaining a positive example latent feature vector and a negative example latent feature vector.

[0050] S303, calculating the representation similarity between the positive example latent feature vector and the negative example latent feature vector by ratio learning loss, wherein the representation similarity includes the negative example latent feature similarity and the positive example latent feature similarity, and guiding the training of the extractor and the projector in a mode of maximizing the negative example latent feature similarity and minimizing the positive example latent feature similarity to obtain the trained extractor and projector.

[0051] It can be understood that the representation similarity between the positive potential feature vector and the negative potential feature vector is the positive potential feature similarity, and the positive potential feature vector has the corresponding positive enhanced sample data, and the feature similarity between the negative enhanced sample data corresponding to the positive enhanced sample data is the negative potential feature similarity. The calculation formula of the contrastive learning loss can be expressed as:

[0052] .

[0053] in and They represent the potential feature vectors of the two enhanced samples, which are obtained by stepwise calculation by the extractor and the projector. is an indicator function if and only if The value is 0, A temperature coefficient is built in for self-supervised contrastive learning, usually 0.07. By clustering similar vibration samples and pushing dissimilar vibration samples apart, the automatic induction and extraction of the characteristic manifold of vibration signals under time-varying speeds without any labels can be achieved. The extractor itself can be a network of any structure, usually including but not limited to Convolutional Neural Network (CNN), Fully Connected Neural Network (FCNN), Long Short-term Memory Network (LSTM) and Transformer Network (Transformer), etc.

[0054] See also Figure 3 , S4, train the regressor through the positive example enhanced sample data and the first speed data to obtain a trained regressor.

[0055] It is understandable that the above steps may further include:

[0056] S401, inputting the positive enhanced sample data into the trained extractor to obtain a second representation feature;

[0057] S402, inputting the first speed data into the trained extractor to obtain a first regression feature;

[0058] S403, guiding the regressor training by the mean square error between the second representation feature and the first regression feature to obtain a trained regressor.

[0059] It can be understood that the regressor can be a nonlinear mapper between the first speed data and the second characterization feature, and when a new speed value is generated, it can quickly calculate the core position of the current speed value in the characterization space; the loss function of the regressor can be given by the following formula. The regressor is a fully connected neural network FCNN with a single hidden layer.

[0060]

[0061] in, is the kth vibration signal in a single training batch, is the kth speed signal in a single training batch, and are extractors and regressors respectively, and N is the number of samples in a single training batch.

[0062] See also Figure 4 , S5, update the trained regressor and classifier through the second vibration data and the second rotation speed data.

[0063] It is understandable that the above steps may further include:

[0064] Based on the extractor trained in the above steps and the trained regressor, the vibration data of all fault types (i.e., the second vibration data) are input into the extractor to obtain the feature of the fault table, and the corresponding speed value is input into the regressor to obtain the fault regression feature. Then the residual vector between the fault characterization feature and the regression feature is calculated. The residual vector is input into the classifier to obtain the corresponding fault type. The classification loss between the classifier output and the true small sample label is calculated to generate the training gradient. The gradient is passed to the regressor through the classifier, and the parameters of the classifier and the regressor are updated at the same time to minimize the classification loss under each speed condition. During the entire classifier training process, the extractor remains frozen. The loss function of the classifier training stage is shown in the following formula. The classifier is also a fully connected neural network FCNN with a single hidden layer.

[0065]

[0066] in, is the category label, is the classifier, is the characteristic vector, is the regression feature vector, is the prediction vector, is the number of all health states.

[0067] S6, obtain real-time vibration data and real-time speed data, and input them into the trained extractor, regressor and classifier to obtain fault diagnosis results.

[0068] The method provided by the present invention achieved higher fault diagnosis accuracy than other comparison methods under 20 samples, 15 samples, 10 samples and 5 samples, and the average diagnosis accuracy reached 89.4708%. The selected data are small sample data under extremely harsh conditions, which shows that the method of the present invention is still effective and robust.

[0069] Table 1 Comparison of fault diagnosis results obtained by various methods

[0070] Model 20 samples 15 samples 10 samples 5 samples average Comparison method 1t 70.7207 68.4685 65.4279 47.0721 62.9223 Comparison Method 2 52.5901 45.3829 41.5541 38.1757 44.4257 Comparison method 3 54.6171 52.7027 34.1216 30.5180 42.9899 Comparison method 4 76.5766 61.3739 58.1081 26.6892 55.6870 Comparison method 5 71.8468 63.2883 47.8604 25.0000 51.9989 Method of the present invention 93.5811 89.7523 87.2748 87.2748 89.4708

[0071] From the above description, it can be known that the time-varying speed small sample fault diagnosis method based on deep residual alignment provided by the present invention can adaptively align the characteristic signal features under time-varying speed conditions, improve the consistency of the model's ability to process vibration signals under continuously changing conditions, and can effectively identify the specific health status of rotating machinery under small sample data and conditions where the speed conditions change over time.

[0072] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A time-varying speed small sample fault diagnosis method based on deep residual alignment, characterized in that: The steps include: Acquire first vibration data and corresponding first rotation speed data in a healthy state, and second vibration data and second rotation speed data in a fault state; Processing the first vibration data to obtain a corresponding enhanced sample set, wherein the enhanced sample set includes a plurality of positive enhanced sample data, and the plurality of positive enhanced sample data are each other's negative enhanced sample data; Inputting the positive enhanced sample data into the extractor for self-supervised contrastive learning training to obtain a trained extractor; Training a regressor using the positive enhanced sample data and the first speed data to obtain a trained regressor; Updating the trained regressor and classifier by using the second vibration data and the second speed data; Real-time vibration data and real-time rotation speed data are acquired and input into the trained extractor, the regressor and the classifier to obtain a fault diagnosis result.

2. The time-varying speed small sample fault diagnosis method based on deep residual alignment according to claim 1 is characterized in that: The first vibration data is processed to obtain a corresponding enhanced sample set, wherein the enhanced sample set includes a plurality of positive enhanced sample data, and the plurality of positive enhanced sample data are each other's negative enhanced sample data, further comprising: Segmenting the first vibration data to obtain a one-dimensional vibration signal sample set; Each sample in the one-dimensional vibration signal sample set is subjected to a negation operation and a flipping operation to obtain an enhanced sample set.

3. The time-varying speed small sample fault diagnosis method based on deep residual alignment according to claim 1 is characterized in that: The step of inputting the positive enhanced sample data into the extractor for self-supervised contrastive learning training to obtain a trained extractor further includes: Inputting the positive enhanced sample data and the corresponding negative enhanced sample data into an extractor to obtain a first characterization feature; Inputting the first characterization feature into a projector to map the first characterization feature to a latent space by nonlinear mapping, thereby obtaining a positive latent feature vector and a negative latent feature vector; The representation similarity between the positive example latent feature vector and the negative example latent feature vector is calculated by ratio learning loss, wherein the representation similarity includes the negative example latent feature similarity and the positive example latent feature similarity, and the training of the extractor and the projector is guided in a mode of maximizing the negative example latent feature similarity and minimizing the positive example latent feature similarity to obtain the trained extractor and the projector.

4. The time-varying speed small sample fault diagnosis method based on deep residual alignment according to claim 1 is characterized in that: The step of training the regressor by using the positive enhanced sample data and the first speed data to obtain a trained regressor further includes: Inputting the positive enhanced sample data into the trained extractor to obtain a second characterization feature; Inputting the first speed data into the trained extractor to obtain a first regression feature; The regressor training is guided by the mean square error between the second characterization feature and the first regression feature to obtain the trained regressor.

5. The time-varying speed small sample fault diagnosis method based on deep residual alignment according to claim 1 is characterized in that: The updating of the trained regressor and classifier by the second vibration data and the second speed data further includes: Inputting the second vibration data into the trained extractor to obtain a third characterization feature; Inputting the second speed data into the trained regressor to obtain a second regression feature; Calculating a residual vector between the third characterization feature and the second regression feature; Inputting the residual vector into the classifier to obtain the fault type; Calculating a loss function between the fault type and the true label; Get the training gradient according to the loss function; Passing the training gradient to the regressor through the classifier; The classifier and the regressor are updated according to the training gradient.

Images