CN-122020872-A - Method and system for constructing universal base model of rotating equipment and electronic equipment

CN122020872ACN 122020872 ACN122020872 ACN 122020872ACN-122020872-A

Abstract

The invention relates to the technical field of fault diagnosis of rotating equipment, and discloses a method, a system and electronic equipment for constructing a universal base model of the rotating equipment, wherein the method comprises the steps of obtaining operation data of multi-source rotating equipment; the method comprises the steps of sequentially carrying out acquisition frequency alignment, length alignment, amplitude alignment, channel alignment and data enhancement on data to obtain a standardized training data set, constructing a base model comprising a shared encoder, a mask reconstruction branch, a domain countermeasure branch, a reconstruction loss, a spectrum sparsity loss and a domain classification loss, adopting a progressive pre-training strategy based on the training data set, and completing model training by adjusting weight coefficients of three loss functions in stages to obtain a general base model of the rotating equipment. The method breaks through the bottleneck of cross-equipment and cross-working condition diagnosis, so that the model can realize high-precision fault diagnosis, the repeated training cost of the model in industrial operation and maintenance is greatly reduced, and the fault diagnosis adaptability and efficiency are improved.

Inventors

Ai Zhenpeng
XIE KANG
XIAO JINGHUA
Luo Tingyu
WANG YUEMIAO
LIU JIA

Assignees

广州佳都智通科技有限公司
佳都科技集团股份有限公司
广东华之源信息工程有限公司
广州华佳软件有限公司
广州佳都城轨智慧运维服务有限公司

Dates

Publication Date: 20260512
Application Date: 20251216

Claims (10)

1. A method for constructing a universal base model of a rotating device, comprising: acquiring multi-source rotating equipment operation data under different equipment types and different working conditions; Sequentially carrying out acquisition frequency alignment, length alignment, amplitude alignment, channel alignment and data enhancement processing on the operation data of the multi-source rotating equipment to obtain a processed training data set; Constructing a base model comprising a shared encoder, a mask reconstruction branch, a domain countermeasure branch and three loss functions, wherein the three loss functions are respectively reconstruction loss, spectrum sparsity loss and domain classification loss; Based on the training data set, a progressive pre-training strategy is adopted, model training is completed by adjusting weight coefficients of three loss functions in stages, and a universal base model of the rotating equipment is obtained.
2. The method of claim 1, wherein the multi-source rotating equipment operational data comprises a public data set, a test stand collected data set, and a real production collected data set, including vibration data for one channel or multiple channels.
3. The method according to claim 1 or 2, wherein the processing procedure of the acquisition frequency alignment comprises presetting a reference acquisition frequency, when the acquisition frequency is higher than the reference acquisition frequency, setting a cut-off frequency as a nyquist frequency of the reference acquisition frequency through low-pass filtering, and then downsampling the filtered data to the reference acquisition frequency; Selecting the shortest data length in the multi-source rotating equipment operation data as a reference length, and dividing other data into one or more fragments with the length equal to the reference length through a sliding window; the processing process of the amplitude alignment comprises the steps of carrying out standardized processing on each piece of data to ensure that the data accords with the distribution with the mean value of 0 and the standard deviation of 1; The channel alignment processing process comprises the steps of keeping the data unchanged if the data is single-channel data, and calculating window length if the data is multi-channel data And remainder R=L% C, wherein C is the number of channels, L is the data length after length alignment, then each channel is segmented into a plurality of subsequences with length W, the subsequences are spliced into C new sequences, and finally R0 are filled at the tail end of each new sequence, so that the length of the new sequence is reduced to L; the data enhancement processing comprises the steps of adding Gaussian noise to the data or carrying out random scaling on the data.
4. The method of claim 1, wherein the shared encoder is configured to extract timing characteristics of vibration data using a TimesNet, TCN, or fransformer network architecture; the mask code reconstruction branch is used for learning the time sequence dependency relationship of data and calculating reconstruction loss and spectrum sparsity loss; The domain countermeasure branch comprises a classifier and a gradient inversion layer, wherein the classifier is used for distinguishing which data source the data comes from, and the gradient inversion layer is used for inverting the gradient of domain classification loss in the training process so as to force the shared encoder to learn domain independent general characteristics.
5. The method of claim 4, wherein the three loss functions are calculated by the following formulas: Reconstruction loss: Where N is the batch size at training, L is the sequence length, Is a mask matrix, where an element representation of 1 is masked, Is the value of the input data and, Is a reconstructed data value; Spectral sparsity loss: Wherein, the Is the mean of the ith sample; domain classification loss: where D is the number of data sources, i.e., the number of fields, Is the one-hot encoding of the real domain label, Is the predicted domain distribution probability; the total model loss can be expressed as: Wherein, the 、 And The weight coefficients of the three losses are respectively.
6. The method of claim 5, wherein the model training is accomplished by adjusting weight coefficients of three loss functions in stages using a progressive pre-training strategy, comprising: The first stage is to train only the reconstruction loss and the spectrum sparsity loss, and weight coefficient , And the training rounds at this stage account for 10% of the total training rounds The second stage is to gradually introduce domain countermeasures to loss, and its weight coefficient With the nonlinear increase of training rounds, the adjustment is performed according to the following formula: Wherein epoch is the current training round number, warmup _epoch is the round number used in the first stage, train_epoch is the round total number of the whole training process, and the training round in the stage accounts for 50% of the total training round; third stage, fixed weight coefficient of reconstruction loss Weight coefficient of sum domain countermeasures against loss Weight coefficient for increasing loss of spectrum sparsity And maintaining the weight coefficient combination until the training is finished, wherein the training round at the stage accounts for 40% of the total training rounds.
7. A rotating equipment universal base model building system, comprising: The data acquisition module is used for acquiring the operation data of the multi-source rotating equipment under different equipment types and different working conditions; the data processing module is used for sequentially carrying out acquisition frequency alignment, length alignment, amplitude alignment, channel alignment and data enhancement processing on the operation data of the multi-source rotating equipment to obtain a processed training data set; the model construction module is used for constructing a base model comprising a shared encoder, a mask reconstruction branch, a domain countermeasure branch and three loss functions, wherein the three loss functions are reconstruction loss, spectrum sparsity loss and domain classification loss respectively; And the model training module is used for completing model training by adjusting weight coefficients of three loss functions in stages by adopting a progressive pre-training strategy based on the training data set to obtain a universal base model of the rotating equipment.
8. An electronic device, comprising: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory having stored therein computer instructions, the processor executing the computer instructions to perform the rotational device universal base model building method of any one of claims 1 to 6.
9. A computer-readable storage medium having stored thereon computer instructions for causing a computer to execute the rotating apparatus universal base model construction method according to any one of claims 1 to 6.
10. A computer program product comprising computer instructions for causing a computer to perform the rotating apparatus universal base model construction method of any one of claims 1 to 6.

Description

Method and system for constructing universal base model of rotating equipment and electronic equipment Technical Field The invention relates to the technical field of fault diagnosis of rotating equipment, in particular to a method and a system for constructing a universal base model of the rotating equipment and electronic equipment. Background In the field of industrial production, rotating equipment (such as a bearing, a gear box, a motor and the like) is used as a core component for energy conversion and transmission, and the running state of the rotating equipment directly determines the production efficiency and the safety, so that the fault diagnosis technology is always a key research direction of industrial operation and maintenance. At present, fault diagnosis of rotating equipment mainly depends on two types of traditional technical schemes, but has obvious limitations, and is difficult to meet the diagnosis requirements under complex working conditions. The first type is a fault diagnosis method based on expert rules. The method takes the experience knowledge of field experts as a core, and judges the state of equipment through preset fault characteristics and diagnosis rules (such as vibration amplitude threshold values, frequency spectrum peak positions and the like). However, the diagnosis precision is completely dependent on the integrity of expert experience, and when facing dynamic working conditions such as variable rotation speed, variable load and the like or composite faults of bearing stripping and gear tooth breakage coexistence, the preset rule is easy to misjudge or miss judge, and the adaptability is very poor. The second category is fault diagnosis methods based on conventional machine learning. The method is to manually extract the time domain (such as peak value and root mean square) or frequency domain (such as harmonic component) characteristics of the vibration signal, and then train a diagnosis model by using a support vector machine, a random forest and other models. In actual production, the failure rate of the rotating equipment is generally low, so that sample data marked with the health state and failure type of the equipment is scarce, and effective training of a supporting model is difficult, and the trained model is only suitable for specific equipment types and fixed working conditions, when the equipment types are replaced or the operation parameters are changed, the model is lowered, the data are required to be collected again, the new model is trained, the operation and maintenance cost is increased, and the diagnosis lag risk exists. Disclosure of Invention In the existing rotating equipment fault diagnosis, the method based on expert rules has limited precision and is difficult to cope with complex working conditions, the method based on traditional machine learning depends on a large amount of labeling data and has poor model generalization capability (can not adapt to cross-equipment and cross-working condition scenes), so that the problem needs to be solved to realize fault diagnosis with low labeling dependence and high suitability under the complex scenes. In a first aspect, the present invention provides a method for constructing a universal base model of a rotating device, including: acquiring multi-source rotating equipment operation data under different equipment types and different working conditions; Sequentially carrying out acquisition frequency alignment, length alignment, amplitude alignment, channel alignment and data enhancement processing on the operation data of the multi-source rotating equipment to obtain a processed training data set; Constructing a base model comprising a shared encoder, a mask reconstruction branch, a domain countermeasure branch and three loss functions, wherein the three loss functions are respectively reconstruction loss, spectrum sparsity loss and domain classification loss; Based on the training data set, a progressive pre-training strategy is adopted, model training is completed by adjusting weight coefficients of three loss functions in stages, and a universal base model of the rotating equipment is obtained. The embodiment of the invention forms a complete universal base model construction system of rotating equipment by integrating multisource data acquisition, full-flow data preprocessing, multi-component base model construction and progressive pre-training strategies. Compared with the traditional fault diagnosis method, the method breaks through the defects that expert rules depend on experience and the generalization of traditional machine learning is poor, multi-source data cover different equipment and working conditions, a foundation is laid for general feature learning, full-flow data preprocessing guarantees input standardization, model training is prevented from being interfered by data differences, a multi-component model is combined with progressive training, model learning time sequence dependence