Multi-bearing time sequence data input model diagnosis method capable of explaining abnormal large vibration of steam turbine of thermal power plant

Through the comparative learning method, the timing data encoder and similarity encoder are constructed, which solves the insensitivity problem of the large model when processing multivariable timing data in the detection of multi-bearing vibration abnormality of steam turbines in thermal power plant, and realizes the vibration abnormality of multi-bearing vibration abnormality of multi-bearing Effective diagnosis and textual interpretation.

CN119004315APending Publication Date: 2024-11-22XI AN JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411055000.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-02
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The prior art relies on experience and manual inspection in the detection of multi-bearing vibration anomalies of steam turbines in thermal power plants. The machine learning results lack explanation, making it difficult to process multivariable timing data, and large models are insensitive in digital data processing, making it difficult to capture complex relationships.

Method used

The comparison learning method is used to build a time series data encoder and a similarity encoder. Through comparison learning, the encoder can learn the relationships and timing characteristics between multi-bearing timing data and the alignment of text descriptions, thereby enhancing the big model's understanding of timing data.

Benefits of technology

It realizes effective diagnosis and text interpretation of vibration abnormalities of multi-bearing turbines in thermal power plants, and improves the understanding and abnormal judgment ability of large models for multivariable timing data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119004315A_ABST
    Figure CN119004315A_ABST
Patent Text Reader

Abstract

The invention discloses a multi-bearing time series data input model diagnosis method capable of explaining abnormal vibration of a steam turbine of a thermal power plant, and the method employs two times of comparison learning to train encoders, the first encoder is used for processing the time series data relation between different bearings, and through learning the generality and difference of multi-bearing time series data, the difference of the time series data is calculated; the understanding ability and change sensing ability of the model to the complex relationship of the multi-feature time series data are enhanced; and the second encoder aligns the time sequence data with the text description in the corresponding time period so as to improve the text interpretation capability of the model on the time sequence data prediction result. And after pre-training, the coded time series data are input into the chatglm-6B large model, so that effective diagnosis of the abnormal vibration of the steam turbine bearing and text interpretation of the result are realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of anomaly detection and large models, and in particular to a large model diagnostic method for interpreting vibration anomalies of a thermal power plant steam turbine with multi-bearing time series data input, which can realize the diagnosis of multi-bearing vibration anomalies of a thermal power plant steam turbine and the corresponding textual interpretation of detection results. Background Art

[0002] Steam turbines are widely used in the power and industrial fields, but during operation, their high-load operation is easily affected by factors such as temperature, pressure, and material aging, which may cause abnormal bearing vibration or even failure. At present, the detection of abnormal vibration of steam turbines mainly relies on experience and manual inspection. Although sensor data analysis and machine learning algorithms have achieved certain results, they face challenges such as consuming a lot of manpower and material resources, lack of explanation of machine learning results, fault mode differentiation, real-time performance, and model generalization ability.

[0003] Large language models (LLMs) have important applications in fields such as natural language processing and computer vision. The advantages of large models are mainly reflected in their powerful natural language understanding and generation capabilities. By pre-training on large-scale corpora, large models can capture semantics, context, and complex language structures to generate coherent and meaningful text interpretations. It can automatically extract and integrate information, provide accurate and detailed interpretations, and reduce human intervention. Its adaptive learning ability enables it to handle diverse language inputs, improve the accuracy and consistency of interpretations, and meet the needs of text understanding in different scenarios. However, large models have shortcomings when processing numerical data. They usually show insensitivity and find it difficult to effectively capture the relationship between numerical changes. At the same time, large models find it difficult to support multi-variable data inputs, cannot learn key information between multiple features, and find it difficult to learn text information represented by time series data. This is mainly because large models are usually designed to be good at processing contextual semantics in natural language rather than precise relationships between numerical values. This design results in large models often having difficulty learning and understanding subtle differences and complex relationships between numbers in numerical-related tasks. Summary of the invention

[0004] In view of the shortcomings in the current prior art, the purpose of the present invention is to propose a large-scale diagnostic method for interpretable abnormal vibration of a steam turbine in a thermal power plant with multi-bearing time series data input. Using contrastive learning, two encoders are constructed. The first one is a time series data encoder. Different bearings belong to the same steam turbine. When an abnormality occurs, all bearing time series data will change. At the same time, because the bearings are located in different positions, the time series data of each bearing is different. Therefore, through contrastive learning, the encoder learns the relationship between the time series data of multiple bearings and learns the common characteristics of the changes in the time series data of different bearings. The second encoder is a similarity encoder, which imposes constraints. Through contrastive learning, the encoder learns the alignment of the time series features with the bearing vibration in the corresponding time period. In this way, it is expected that the large model will establish a connection between the time series data input and the text description of the vibration data.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] A large-scale diagnostic method for interpreting abnormal vibration of a steam turbine in a thermal power plant with multi-bearing time series data input includes the following steps:

[0007] Step 1: First, sample the bearing vibration time series data of the steam turbine in the thermal power plant at a certain frequency, and then perform time alignment on the data to ensure that each data corresponds to the same time base; for the aligned data, use the average value to fill the missing values, so as to obtain the processed complete time series data set X d , there are d bearing vibration data;

[0008] Step 2: Transform the complete time series dataset X d Cut and divide into different batches of data b1, b2, ..., b at the same time interval n , each batch is further divided into N time periods of data s 1 ,s 2 ,…,s N , and then for each time period time series data Perform manual annotation and provide a text description of the changes in vibration data during this period. 1 ,t 2 ,…,t N .

[0009] Step 3: Comparative learning of different bearing time series data: define an embedding function, Where T represents the number of samples in each time period, d represents the number of bearing vibration data, and u k It is the vector representation of the time series data after being transformed into M-dimensional space by function. First, all the subsequences of the bearing vibration data, that is, the time series data of each time period, are Embedded into the vector space through the embedding function, the embedding representation is u1 ,u 2 ,…,u k After obtaining the embedding representation of different subsequences, a neural network is constructed as a time series data encoder Among them, e a is the encoded representation of the time series data after passing through the encoder, the dimension remains unchanged, a is the serial number of the data time period, and the j-th bearing vibration data in the a-th time period is As the benchmark, j represents the number of columns, different feature columns represent different bearing vibration data, the turbine has d bearing vibration data, and the comparative learning enhancement data is performed to take the remaining bearing vibration data in the ath time period As positive samples for contrastive learning, the bearing vibration data in other time periods within the small batch is used as negative samples The encoder is trained using a loss function, which is defined as:

[0010]

[0011] Where: L is the training definition loss function of the time series data encoder, K is the batch size, represents the Euclidean distance between two column vectors, and Respectively and The kth component of is the time series data vector representation of other time periods in the small batch, r is the bearing number, and when calculating the loss function, the hyperparameter σ1 is 0.04, which is used to reduce the sensitivity to data distribution and improve the robustness of the loss function.

[0012] Step 4: Comparative learning of time series data and text data: Build another neural network as a similarity encoder Among them, e a is the encoded representation of the time series data after the time series encoder, and then the nonlinear projection x of the time series data is obtained through the similarity encoder a , the text data is obtained through the large model word embedding In this way, the dimensions of text data and time series data are converted to the same dimension, with a dimension size of Q. The large model cannot understand the meaning of the time series data. In order to enable the large model to learn the information contained in the time series features, the similarity is used to measure the relationship between the time series data and the corresponding text embedding information. Through comparative learning, the time series data x a The corresponding explanation text for the time period t a As positive samples, the time series data explanation text t corresponding to the remaining time periods in the mini-batch k(k≠a)As negative samples, through training, the time series data and the corresponding explanation text of the time series data have the maximum similarity to expect the big model to understand the internal information of the time series data. At the same time, the similarity with the explanation text of the time series data corresponding to other time periods is maximized, so that the big model can learn the differences and representative meanings of the time series data in different time periods. The training loss function of the similarity encoder is:

[0013]

[0014] r(x a ,t a )=(x a ) Τ t a

[0015] Where: σ2 is 0.07, which is used to reduce the sensitivity to data distribution and improve the robustness of the loss function.

[0016] Step 5: Select the reference time series data, i.e. the jth bearing vibration data in the ath time period The output of the pre-trained time series data encoder in step 3 and the pre-trained similarity encoder in step 4 are input into the chatglm-6B large model with frozen parameters. The text generation capability of the large model is used to finally obtain the abnormal diagnosis results of the time series data and the corresponding text explanation of the results.

[0017] Beneficial effects:

[0018] The present invention provides a large-scale diagnostic method for interpreting abnormal vibration of a steam turbine in a thermal power plant with multi-bearing time series data input, which has the following beneficial effects:

[0019] (1) Text data is usually unitary, while time series data is often multivariate. Therefore, it is difficult for traditional large models to input multivariate time series data and learn the complex relationship between multiple variables in time series data. The present invention adopts comparative learning. Because the bearing data of the same unit have commonalities, other bearing time series data are innovatively used for comparative learning, so that the model can learn the complex relationship between different bearings, and then the large model can learn multivariate time series data.

[0020] (2) It is difficult for large models to effectively capture the relationship between the digital changes in time series data and they are insensitive. In order to enhance the perception of time series data by large models, an encoder is designed and contrastive learning is used to align time series data with text descriptions. This can also distinguish the differences in time series data in different time periods, thereby enhancing the large model’s ability to understand data and enabling effective anomaly judgment and result interpretation by large models. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a comparative learning diagram of different bearing data in the same time period in the present invention.

[0022] Figure 2 Learn graphs with maximum similarity between time series data and corresponding text descriptions. DETAILED DESCRIPTION

[0023] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0024] The present invention provides a large-scale diagnostic method for interpreting abnormal vibration of a steam turbine in a thermal power plant with multi-bearing time series data input, comprising the following steps:

[0025] Step 1: Data collection and preprocessing: First, obtain the vibration data of 8 bearings of steam turbine units 1, 2, 3, 4, 5, and 6 of a thermal power plant from 21 to 22. Resample the vibration data of these different bearings at 10s intervals and perform time alignment. The data is the average of the vibration data in each time interval to ensure that the corresponding time points of each data are the same. For the aligned data, the missing values ​​are filled with the average value of this column of data to obtain a processed complete time series data set. Some data are as follows:

[0026]

[0027] Step 2: Split the complete time series data set into sub-sequences, each containing 10 minutes of vibration data of 8 bearings. First, split the time series data into different batches b1, b2, ..., b in units of 2 hours. n Each batch of data is further divided into 12 subsequences s 1 ,s 2 ,…,s 12 , the length of each subsequence is 10 minutes.

[0028] Next, the vibration data of each batch of subsequence time periods are manually labeled and described to obtain t 1 ,t 2 ,…,t 12 Specifically, the changes in vibration data in each time period are described in the following aspects:

[0029] Vibration Amplitude: Describes the overall amplitude of the vibration data, including maximum, minimum and average values.

[0030] Vibration frequency: describes the frequency change of vibration data.

[0031] Vibration amplitude trend change: describes the change trend of the vibration amplitude during the time period, whether it increases or decreases.

[0032] Is there any abnormality: Determine whether there is any abnormal vibration during this time period and make a corresponding description. The text description of some subsequence data is as follows:

[0033]

[0034] Step 3: Comparative learning of different bearing time series data: define an embedding function See Figure 1 First, all bearing vibration data are divided into each subsequence The time series is embedded into a 512-dimensional vector space through the embedding function, and the embedding representation is u 1 ,u 2 ,…,u 12 After obtaining the embedding representation of different subsequences, a neural network is constructed as a time series data encoder Based on the vibration data of No. 4 bearing As a benchmark, different feature columns represent different bearing vibration data. Comparative learning is performed to enhance the data. The vibration data of other bearings in the same time period are taken. As positive samples for contrastive learning, the bearing vibration data in other time periods of the small batch is used as negative samples The encoder is trained using the contrastive loss function, which is defined as:

[0035]

[0036] Step 3: Comparative learning of time series data and text data: Another neural network was constructed as a similarity encoder e a The time series data is nonlinearly projected x by embedding the time series through the similarity encoder a , the text data is embedded in the large model word space text The large model cannot understand the meaning of the time series data. In order to enable the large model to learn the information contained in the time series features, the similarity is used to measure the relationship between the time series data and the corresponding text embedding information. Figure 2 , through contrastive learning, the time series data x a The explanatory text corresponding to the time period t a As positive samples, the time series data explanation text t corresponding to the remaining time periods in the mini-batch k(k≠a) As negative samples, through training, the time series data and the corresponding explanation text of the time series data have the maximum similarity to expect the big model to understand the internal information of the time series data. At the same time, the similarity with the explanation text of the time series data corresponding to other time periods is maximized, so that the big model can learn the differences and representative meanings of the time series data in different time periods. The contrast loss function of similar encoder training is:

[0037]

[0038] r(x a ,t a )=(x a ) Τ t a

[0039] Step 4: Convert the benchmark time series data Input, get the output through the two trained encoders above, input into the chatglm-6B large model with frozen parameters, and you can get the abnormal diagnosis results of the time series data in the corresponding time period and the corresponding result explanation.

[0040] The present invention adopts two contrast learning training encoders. The first encoder is used to process the time series data relationship between different bearings. By learning the commonalities and differences of multi-bearing time series data, the model's ability to understand the complex relationship of multi-feature time series data and perceive changes is enhanced; the second encoder aligns the time series data with the text description within the corresponding time period to improve the model's ability to interpret the text of the time series data prediction results.

Claims

1. A large-scale diagnostic method for interpreting abnormal vibration of a steam turbine in a thermal power plant with multi-bearing time series data input, characterized in that: The steps include: Step 1: First, sample the bearing vibration time series data of the steam turbine in the thermal power plant at a certain frequency, and then perform time alignment on the data to ensure that each data corresponds to the same time base; for the aligned data, use the average value to fill the missing values, so as to obtain the processed complete time series data set X d , there are d bearing vibration data; Step 2: Transform the complete time series dataset X d Cut and divide into different batches of data b1, b2, ..., b at the same time interval n , each batch is further divided into N time periods of data s 1 ,s 2 ,…,s N , and then for each time period time series data Perform manual annotation and provide a text description of the changes in vibration data during this period. 1 ,t 2 ,…,t N ; Step 3: Comparative learning of different bearing time series data: define an embedding function, Where T represents the number of samples in each time period, d represents the number of bearing vibration data, and u k It is the vector representation of the time series data after being transformed into M-dimensional space by function. First, all the subsequences of the bearing vibration data, that is, the time series data of each time period, are Embedded into the vector space through the embedding function, the embedding representation is u 1 ,u 2 ,…,u k After obtaining the embedding representations of different subsequences, a neural network is constructed as a time series data encoder Among them, e a is the encoded representation of the time series data after passing through the encoder, the dimension remains unchanged, a is the serial number of the data time period, and the j-th bearing vibration data in the a-th time period is As the benchmark, j represents the number of columns, different feature columns represent different bearing vibration data, the turbine has d bearing vibration data, and the comparative learning enhancement data is performed to take the remaining bearing vibration data in the ath time period As positive samples for contrastive learning, the bearing vibration data in other time periods within the small batch is used as negative samples The encoder is trained using a loss function, which is defined as: Where: L is the training loss function of the time series data encoder, K is the batch size, represents the Euclidean distance between two column vectors, and Respectively and The kth component of is the time series data vector representation of other time periods in the small batch, r is the bearing number, σ1 is the hyperparameter size of 0.04, which is used to reduce the sensitivity to data distribution and improve the robustness of the loss function; Step 4: Comparative learning of time series data and text data: Build another neural network as a similarity encoder Among them, e a is the encoded representation of the time series data after the time series encoder, and then the nonlinear projection x of the time series data is obtained through the similarity encoder a , the text data is obtained through the large model word embedding In this way, the dimensions of text data and time series data are converted to the same dimension, with a dimension size of Q. The large model cannot understand the meaning of the time series data. In order to enable the large model to learn the information contained in the time series features, the similarity is used to measure the relationship between the time series data and the corresponding text embedding information. Through comparative learning, the time series data x a The corresponding explanation text for the time period t a As positive samples, the time series data explanation text t corresponding to the remaining time periods in the mini-batch k(k≠a) As negative samples, through training, the time series data and the corresponding explanation text of the time series data have the maximum similarity to expect the big model to understand the internal information of the time series data. At the same time, the similarity with the explanation text of the time series data corresponding to other time periods is maximized, so that the big model can learn the differences and representative meanings of the time series data in different time periods. The training loss function of the similarity encoder is: r(x a ,t a )=(x a ) Τ t a Where: L′ is the training loss function of the similarity encoder, σ2 is a hyperparameter used to reduce the sensitivity to data distribution and improve the robustness of the loss function. Step 5: Select the reference time series data, i.e. the jth bearing vibration data in the ath time period The output of the pre-trained time series data encoder in step 3 and the pre-trained similarity encoder in step 4 are input and then input into the large model with frozen parameters. The text generation capability of the large model is used to finally obtain the abnormal diagnosis results of the time series data and the corresponding text explanation of the results.

2. The diagnostic method for abnormal vibration of a steam turbine in a thermal power plant with interpretable large model based on multi-bearing time series data input according to claim 1 is characterized by: The hyperparameter σ1 is 0.04 and the hyperparameter σ2 is 0.

07.

3. The diagnostic method for abnormal vibration of a steam turbine in a thermal power plant with interpretable large model based on multi-bearing time series data input according to claim 1 is characterized by: The large model described in step 5 is the chatglm-6B large model.