Neutron chopper fault diagnosis method and system based on multi-sensor data fusion

Abstract

The invention relates to a neutron chopper fault diagnosis method based on multi-sensor data fusion, and the method comprises the following steps: a comparison coding model is designed, the comparison coding model comprises a feature extractor Genc and an autoregression model Gar, the feature extractor Genc extracts features from a plurality of one-dimensional data samples, and fuses the features into a feature vector zt; the feature extraction operation can be expressed by a formula: zt = Gent (xt); the autoregressive model Gar takes the feature representation z < = t of the previous time step as the input, generates the potential feature representation ct of the context through ct = Gar (z < = t), and then predicts the feature representation zgt of the subsequent time step by using the output ct of the autoregressive model; t; the method comprises the following steps: acquiring original data of a neutron chopper through multiple sensors, and converting the original data into a frequency domain signal by using fast Fourier transform (FFT); and inputting the converted frequency signal into a comparison coding model, and predicting the fault state of the neutron chopper through the coding model.

Description

Technical Field

[0001] The present invention relates to the field of neutron chopper fault diagnosis, and in particular to a multi-sensor data fusion neutron chopper fault diagnosis method and system. Background Art

[0002] The neutron chopper is an indispensable key equipment in the China Spallation Neutron Source time-of-flight neutron scattering spectrometer. The neutron chopper belongs to the type of rotating mechanical equipment. Its main function is to remove the fast neutron background, select neutrons in a certain band range, and select monochromatic neutrons in the time-of-flight neutron scattering spectrometer, so it is also called a neutron energy selector.

[0003] In actual operation, it is difficult to establish a fault diagnosis model for monitoring the operating status of the T0 neutron chopper, mainly because its monitoring data has the problems of single source, sparse labels and diverse working conditions: 1) The neutron chopper can only obtain bearing vibration signals on the shielding shell. Since the vibration of the rolling bearing is attenuated when it propagates on the shielding shell, and the operation of the vacuum and water cooling system causes a large amount of vibration noise on the shielding shell, it is difficult to obtain effective fault characteristics from it when using a single vibration sensor for monitoring; 2) Since adding labels to the monitoring data requires professional knowledge and a lot of manpower, due to cost constraints, only a small amount of data can be labeled. Therefore, the distribution of the collected data labels is very sparse, that is, only a few samples have labels, and sometimes even no labels, which makes it difficult to use traditional supervised learning methods to build a fault diagnosis model; 3) The T0 neutron chopper has several different working conditions in actual operation, among which the monitoring data collected under common working conditions (source domain) is large, while the monitoring data under uncommon working conditions (target domain) is small, and it is impossible to directly establish a fault diagnosis model under uncommon working conditions.

[0004] In recent years, the application of supervised deep learning methods in the field of mechanical fault diagnosis has brought new breakthroughs. This method can learn accurate mapping relationships from monitoring data and provide accurate, efficient and reliable solutions for fault diagnosis. In actual industrial applications, the original vibration signals require a large number of experts in professional fields to label, which consumes a lot of manpower and material resources. How to use the large amount of unlabeled data collected has become a difficult problem. In complex industrial systems, it is difficult to extract effective fault features from a single signal source data. Therefore, it is necessary to study how to extract fault features from multi-source data to improve the performance of rotating machinery fault diagnosis. Summary of the invention

[0005] Based on contrastive learning and multi-sensor data fusion, the present invention proposes a self-supervised fault diagnosis method based on contrastive predictive coding multi-sensor data fusion and applies it to the fault diagnosis research of neutron chopper, aiming to solve the problem of limited fault diagnosis performance caused by a large number of unlabeled data samples and single signal source data feature extraction.

[0006] In order to achieve the above purpose, the technical solution adopted by the present invention is: a neutron chopper fault diagnosis method based on multi-sensor data fusion, comprising the following steps:

[0007] Step 1: Design a contrastive coding model. The contrastive coding model includes a feature extractor Genc and an autoregressive model Gar. The feature extractor Genc extracts features from multiple one-dimensional data samples and fuses them into a feature vector z t , the feature extraction operation can be expressed as: t =Genc(x t ); the autoregressive model Gar represents z with the features of the previous time step ≤t as input and pass c t =Gar(z ≤t ) to generate the contextual latent feature representation c t , then, using the output c of the autoregressive model t To predict the feature representation z of the next time step >t ; Input data x t , t is the current input time step, Z t is the feature vector of this step;

[0008] Step 2: Collect the raw data of the neutron chopper through multiple sensors and convert it into a frequency domain signal using fast Fourier transform (FFT);

[0009] Step 3: Input the converted frequency signal into a comparison coding model, and predict the fault state of the neutron chopper through the coding model.

[0010] Preferably, the model encodes the input data x (each sample is 1024 data points) through a series of nonlinear transformations, learns the latent semantics c of the context, and converts it into a vector form, which is defined as:

[0011]

[0012] I(x,c) is the mutual information between the original signal x and the contextual latent semantics c;

[0013] The goal of the model is to maximize the mutual information between the original signal and the contextual latent semantics c. The model adopts the method of constructing a density function f, which retains the model input sample xt+k (Each sample has 1024 data points, 8 samples form a group, a total of 8192 data points) and contextual latent semantics c t The mutual information between them is expressed as:

[0014]

[0015] Where k represents the kth moment, ∝ represents proportional to, and a concise log-linear model equation is used to represent f:

[0016]

[0017] Among them It is used to express that for each step k, a linear converter W is used k To make predictions; by calculating the density ratio f k (x t+k ,c t ), the model can evaluate the prediction results Accuracy in high-dimensional distributions, and distinguishing the relationship between its own model and input data to achieve effective separation.

[0018] Preferably, the feature extractor Genc is divided into three modules: feature extraction, feature concatenation and feature fusion. First, the preprocessed one-dimensional vibration signal is input into the designed convolutional neural network model for feature extraction. After feature extraction, multiple features are obtained respectively, and then the features are concatenated into a feature vector through the feature concatenation layer, and sent to the fully connected layer for fusion to obtain the z in CPC. t =Genc(x t ); mathematically summarized as follows:

[0019] Assume that there are m different sensors collecting data at different locations, and the expression is:

[0020] y(t)=[y 1 (t),y 2 (t)…y m (t)]

[0021] Where y(t) is the input data set;

[0022] The length of the signal data is reconstructed into a matrix of the input of the feature extractor Genc, ​​as follows:

[0023]

[0024] The extracted features are concatenated as shown in the equation:

[0025]

[0026] The obtained features are concatenated and input into the feature fusion layer to obtain the z required in CPC t =Genc(x t ).

[0027] Preferably, in the contrastive coding model, the infoNCE function is used to train the feature extractor Genc and the autoregressive model Gar, which evaluates the performance of the model by calculating the cross entropy between positive samples and negative samples. The positive samples are obtained from P(x t+k |c t ), and the negative sample is a sample randomly selected from P(x t+k ) randomly selects N-1 samples; the expression of its loss function is as follows:

[0028]

[0029] The above equation is used to calculate the model prediction value f k / ∑ X f k The corresponding cross entropy, in this equation, f k represents the predicted value of the model, and ∑ X f k It represents the sum of the predicted values ​​of all samples, and further introduces the optimal loss probability p(d=i|X,c t ), where the sample x i The positive sample in can be expressed as d = i, from the conditional distribution p(x t+k |c t ) is a sample x extracted from i The probability can be expressed by the following equation:

[0030]

[0031] f k (x t+k ,c t ) is the optimal value as p(x t+k |c t ) / p(x t+k ) increases, that is, it is a proportional relationship, and the selection of the number of negative samples N-1 will not affect the proportional relationship between the two;

[0032] x t+k With variable c t The mutual information between can be evaluated by the following expression:

[0033] I(x t+k ,c t )≥log(N)-L N

[0034] As the loss L is minimizedN At the same time, the lower bound of the common information can be maximized, and the mutual information will become more solid.

[0035] The technical effect of the present invention is as follows: a study is conducted when a large number of neutron chopper data sets are unlabeled, and in view of the fact that a single sensor is difficult to fully reflect the true state of the neutron chopper, a fault diagnosis method of multi-sensor data fusion based on comparative predictive coding is proposed. The method is elaborated in detail from five aspects, including data preprocessing steps, construction of CNN model, structural design of comparative coding model, loss function, construction of feature fusion layer and the overall execution flow of the method; the problem of limited fault diagnosis performance caused by a large number of unlabeled data samples and feature extraction of single signal source data is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is the structural diagram of the contrast coding (CPC) model.

[0037] Figure 2 This is the structural diagram of the feature extractor Genc.

[0038] Figure 3 is a two-dimensional projection embedded using a T-distributed random neighborhood.

[0039] Figure 4 Line graph of neutron chopper diagnostic results.

[0040] Figure 5 An experimental platform was built.

[0041] Figure 6 This is a framework diagram of a neutron chopper state monitoring system based on multi-sensor data fusion.

[0042] Figure 7 This is the interactive interface diagram of the neutron chopper monitoring and diagnostic system. DETAILED DESCRIPTION

[0043] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings. The description in this part is only exemplary and explanatory and should not have any limiting effect on the protection scope of the present invention.

[0044] 1.1. Basic Algorithm Theory

[0045] The redesigned contrast coding (CPC) model structure is as follows Figure 1 As shown in Figure 2, the model contains a feature extractor Genc, ​​which can extract features from multiple one-dimensional data samples and fuse them into a feature vector z t , the feature extraction operation can be expressed as: t =Genc(xt ). In the model structure, an autoregressive model Gar is also included. This model represents z with the features of the previous time step. ≤t as input and pass c t =Gar(z ≤t ) to generate the contextual latent feature representation c t , then, using the output c of the autoregressive model t To predict the feature representation z of the next time step >t That is, the model predicts the feature representation of the future time step by learning the feature representation of the previous time step. In the T0 neutron chopper fault diagnosis scenario, the input continuous multi-sensor data x t-3:t+4 After the feature extractor, we can get the feature representation z t-3:t+4 , the autoregressive model Gar can represent z from the feature t-3:t The context feature representation c is learned t , and use c t To predict the subsequent feature representation z t+1:t+4 If the autoregressive model Gar can accurately predict the feature representation of subsequent data, it can be said that the feature extractor Genc has successfully extracted representative features.

[0046] The above model can use existing information to complete the prediction of future information, and the model encodes the input data x through a series of nonlinear transformations, learns the latent semantics c of the context, and converts it into a vector form. This representation method helps to retain the mutual information between the original signal x and the latent semantics c of the context in the vector space. Its definition equation is shown in (1-1):

[0047]

[0048] Where x is the original signal and c is the contextual latent semantics.

[0049] The goal of the model is to maximize the mutual information between the original signal and the contextual latent semantics c. To achieve this goal, the model adopts a method of constructing a density function f that can retain the model input sample x. t+k (Each sensor data sample is 1024 data points, 8 samples are grouped together, a total of 8192 data points) and contextual latent semantics c t The mutual information between them is expressed as (1-2):

[0050]

[0051] Here ∝ represents proportional to, and the density ratio f does not need to be integrated to 1. This paper uses a concise log-linear model equation (1-3) to represent f:

[0052]

[0053] Among them, W k c t It is used to express that for each step k, a linear converter W is used k By calculating the density ratio f k (x t+k ,c t ), the model can evaluate the prediction results Accuracy in high-dimensional distributions and distinguishing the relationship between its own model and the input data, thereby achieving effective separation.

[0054] 1.2. Feature Extractor

[0055] Figure 2 The basic structure of the feature extractor Genc can be divided into three modules: feature extraction, feature concatenation and feature fusion. First, the preprocessed one-dimensional vibration signal is input into the designed convolutional neural network model for feature extraction. After feature extraction, multiple features are obtained respectively, and then the features are concatenated into a feature vector through the feature concatenation layer and sent to the fully connected layer for fusion to obtain the z in CPC. t =Genc(x t ). Mathematically it can be summarized as follows.

[0056] (1) Assume that there are m different sensors collecting data at different locations. The expression is shown in (1-4):

[0057] y(t)=[y 1 (t),y 2 (t)…y m (t)] (1-4)

[0058] Where y(t) is the input data set

[0059] (2) Reconstruct the length of the signal data into a matrix for the input of the feature extractor Genc, ​​as shown in (1-5):

[0060]

[0061] (3) Concatenate the extracted features, as shown in equation (1-6):

[0062]

[0063] Where x t The features after splicing

[0064] (4) Concatenate the obtained features and input them into the feature fusion layer to obtain the z required in CPCt =Genc(x t ).

[0065] 1.3. Loss Function

[0066] In the contrastive coding model, the infoNCE function is used to train the feature extractor Genc and the autoregressive model Gar. It evaluates the performance of the model by calculating the cross entropy between positive and negative samples. The positive samples are obtained from P(x t+k |c t ), and the negative sample is a sample randomly selected from P(x t+k ) N-1 samples are randomly selected from . The expression (1-7) of its loss function is as follows.

[0067]

[0068] Where f k / ∑ X f k is the predicted value of the model.

[0069] The above equation is used to calculate the model prediction value f k / Σ X f k The corresponding cross entropy, in this equation, f k represents the predicted value of the model, and Σ X f k It represents the sum of the predicted values ​​of all samples, and further introduces the optimal loss probability p(d=i|X,c t ), where the sample x i The positive samples in can be expressed as d = i. From the conditional distribution p(x t+k |c t ) is a sample x extracted from i The probability can be expressed by equation (1-8).

[0070]

[0071] Therefore, from equation (1-8) summarized in this article, it can be seen that f k (x t+k ,c t ) is the optimal value as p(x t+k |c t ) / p(x t+k ), which is a proportional relationship, and the selection of the number of negative samples N-1 will not affect the proportional relationship between the two.

[0072] In addition, x t+k With variable c tThe mutual information between can be evaluated by expression (1-9):

[0073] I(x t+k ,c t )≥log(N)-L N (1-9)

[0074] According to the above formula, we can make corresponding judgments, and minimize the loss L N At the same time, the lower bound of the common information can be maximized, and the mutual information will become more solid.

[0075] 1.4. Neutron chopper fault diagnosis experiment

[0076] 1.4.1 Model design

[0077] The feature extractor in this algorithm is built by a convolutional neural network, which consists of 5 convolutional layers and two fully connected layers (Dense) to reduce the feature dimension to 16 dimensions.

[0078] Table 1.1 Comparison of encoding feature extractor parameters

[0079]

[0080]

[0081] Note: F-single; K-convolution kernel size; S-step length 1.4.2. Construction of feature fusion layer

[0082] The feature fusion layer consists of two fully connected layers. The number of neurons in the first fully connected layer is 32, and the number of neurons in the second fully connected layer is 16. As shown in Table 1.2 and Table 1.3:

[0083] Table 1.2 Feature fusion layer

[0084]

[0085] Table 1.3 Comparison of coded autoregressive model parameters

[0086]

[0087] 1.4.3 Experiments

[0088] A fault simulation experimental system for a neutron chopper was built, including a motor drive system, a rotor system, a vacuum control system and a data acquisition system.

[0089] Rolling bearing failure of T0 Neutron Chopper will cause the bearing to emit a pulse signal. The waveform of this signal is short, the frequency range is wide, and the energy decays rapidly. In addition, this fault signal will propagate around the bearing in a spherical form, which means that the characteristics of the bearing signal must be fully utilized during data acquisition and analysis. In order to obtain more accurate vibration data, the sensor should be as close to the bearing as possible. When collecting data, select the horizontal (x), vertical (y) and axial (z) directions. The reasons for selecting these directions for data collection are as follows:

[0090] Horizontal (x) direction: This direction is perpendicular to the rolling direction of the bearing and can capture the lateral vibration signal caused by the bearing rolling. The data in this direction mainly detects the relative movement between the inner and outer ring rolling elements of the bearing.

[0091] Vertical (y) direction: This direction is perpendicular to the bearing and can capture the vibration signal of the bearing in the vertical direction. The data in this direction is mainly used to detect the overall vibration condition of the bearing and the rigidity characteristics of the bearing seat.

[0092] Axial (z) direction: This direction is consistent with the bearing axis and can capture the vibration signal of the bearing in the axial direction. The data in this direction is used to detect the axial movement of the bearing, the bearing preload and the vibration information of the bearing outer ring.

[0093] By collecting data in these three directions, we can fully understand the vibration characteristics of the bearing and improve the accuracy and reliability of bearing fault diagnosis. Therefore, when collecting vibration data, it is necessary and effective to approach the bearing and collect data in the horizontal, vertical and axial directions, which can provide a more reliable data basis for fault detection and prediction.

[0094] Experimental platform such as Figure 5 As shown, the T0 neutron chopper data set collected in the experiment has a sampling rate of 20 kHz, a sampling time of 10 s, and a number of sensors of 4.

[0095] Overall experimental steps:

[0096] Step 1: Build an experimental system to collect the required raw data;

[0097] Step 2: Cut the acquired raw data samples and convert them into frequency domain signals using Fast Fourier Transform (FFT).

[0098] Step 3: Construct a feature extractor and apply the feature extractor to the contrastive encoding model.

[0099] Step 4: Use single-source sensor, dual-source sensor, triple-source sensor, and quad-source sensor data for feature extraction respectively.

[0100] Step 5: Save the feature extractor parameters, and draw T-SNE graphs of the data features of single-source sensors, dual-source sensors, three-source sensors, and four-source sensors, respectively, to perform qualitative evaluation based on feature visualization; and perform quantitative evaluation based on intra-class distance and inter-class distance.

[0101] Step 6: Perform comparative evaluation based on the above results.

[0102] The experiment of this application first uses the vibration signal data of 200r / min without load under vacuum state. In order to verify and evaluate the diagnostic performance of the proposed method, the results of single sensor data, dual sensor data, three sensor data and four sensor data are introduced for comparison. In the training process of comparative coding, this application uses Adam optimizer, which has the characteristics of adaptive learning rate and can efficiently update parameters. Its learning rate is set to 1e-3, that is, 0.001. A smaller learning rate can make the parameter update more stable, and the number of iterations (Epoch) is set to 10, that is, the data set is iterated 10 times during the training process. In order to ensure the fairness of the experimental process, all experiments are carried out on a computer with Intel (R) Xeon (R) Gold 5218R*2 processor, NVIDIA GeForce RTX 3090*2 graphics card and 64G memory, and the environment is tensorflow2.6. The ratio of training set, test set and validation set is 6:2:2 to ensure the reasonable division of data.

[0103] 1.5 Results Analysis

[0104] In the results, this application adopts two evaluation methods, namely, qualitative evaluation based on visualization and quantitative evaluation based on intra-class distance and inter-class distance. The qualitative evaluation based on visualization can directly determine whether the model can accurately separate samples of different categories, while the quantitative evaluation based on intra-class distance and inter-class distance can help understand the degree of separation between different categories and the degree of aggregation within the same category. These two methods complement each other and provide a comprehensive evaluation of model performance.

[0105] 1.5.1 Qualitative evaluation of data features based on dimensionality reduction visualization

[0106] After successfully completing the pre-training of the comparative predictive coding model, the extracted features are first reduced in dimension using the T-SNE method. Figure 3 As shown, the visual feature results can be better compared.

[0107] After the CPC training is completed using single sensor, dual sensor, triple sensor, and quad sensor data, the trained feature extractor Genc is used to extract features from the data samples. In order to better observe the extracted feature data, this paper uses the T-SNE algorithm to visualize the extracted 16-dimensional feature data. The results are as follows: Figure 3 As shown. Observing the graphical results, we can see that: Figure 3 (a) to Figure 3 (d) The trend of change, where 0, 1, 2, and 3 represent the bearing conditions of normal, outer ring pitting, outer ring spalling, and rolling element spalling, respectively. Figure 3 The clustering effect of (d) is significantly better than Figure 3 (a) Figure 3 (b) and Figure 3 (c) This shows that with the increase of sensor data, the effect of feature clustering has been significantly improved, and the clustering between features has become tighter, and samples of different categories are easier to distinguish. This shows that the multi-sensor information fusion method can provide richer feature expressions in self-supervised training, thereby improving the effect of feature extraction. In contrast, a single sensor can only provide limited information, resulting in poor feature clustering effect.

[0108] 1.5.2 Quantitative evaluation based on intra-class distance and inter-class distance

[0109] Table 1.4 lists in detail the mean and standard deviation of the intra-class distances between the four categories calculated, as well as the mean and standard deviation of the inter-class distances of the other three methods.

[0110] Table 1.4 Intra-class distance and inter-class distance

[0111]

[0112] In Table 1.4, the present application provides the intra-class distances of single sensor, dual sensor, triple sensor and quad sensor. By observing these values, some conclusions can be drawn: first, the average value of quad sensor is 22.880, while the average values ​​of intra-class distances of single sensor, dual sensor and triple sensor are 59.133, 23.809 and 39.704 respectively. This means that after the four-sensor data fusion training, the feature extractor can extract more compact features from the one-dimensional vibration signal. In contrast, the average value of intra-class distance of single sensor is higher, indicating that in the case of a single sensor, the degree of aggregation between features is lower. The average intra-class distance of dual sensor and triple sensor is between that of single sensor and quad sensor, indicating that the degree of aggregation of features increases with the increase in the number of sensors. In addition, the average values ​​of inter-class distances of single sensor, dual sensor, triple sensor and quad sensor are 60.567, 99.032, 101.165 and 113.205 respectively. It can be seen that the average inter-class distance of the four sensors is significantly higher than that of the other three methods, which indicates that the difference between features is further enhanced by fusing the information of the four sensors. In contrast, the average inter-class distance of the single sensor is lower, which means that the degree of separation between different categories is lower. These results show that in fault diagnosis tasks, the multi-sensor data fusion method can provide more reliable and effective feature representation, thereby improving the performance and accuracy of the model.

[0113] 1.5.3 Fault Diagnosis Experiment with Few Labels

[0114] The above results all show the effectiveness of four-sensor data fusion. Therefore, four-sensor data is used in subsequent few-label experiments. Compared with the above experimental steps, a step of using a large amount of unlabeled data to train the contrastive coding model is added. The purpose of this step is to obtain a feature extractor with representative features through the training of the contrastive coding model, and freeze its parameters during the training process. By maximizing the similarity between samples of the same class and minimizing the similarity between samples of different classes, a feature representation with discriminability is learned. The parameters of the feature extractor are fixed, and the classification layer is trained using a small amount of labeled data. Finally, a large number of samples are used to verify the accuracy of the method and comprehensively evaluate the performance of the model. By using a contrastive predictive coding model trained with a large amount of unlabeled data, the feature extractor can learn more representative feature representations, and freezing the parameters of the feature extractor can ensure that the learned features can perform best in the classification task. Such a training strategy makes full use of a small amount of labeled data, so that the model can show higher performance in the case of few labels, and can determine the effectiveness of the proposed method in the case of few labels, providing empirical support for its promotion in practical applications. The results of this step have important guiding significance for further improving and optimizing the model based on contrastive predictive coding.

[0115] Table 1.5 shows the diagnostic results of the neutron chopper using the contrast prediction coding pre-training model and the non-contrast coding pre-training model on different loads. For each load type, this paper uses 5 to 200 labeled data for supervised training of the model. When the model has a pre-training operation, the feature extractor is frozen and only the classification layer participates in the training. For the model that has not been pre-trained, the entire model participates in the training. In the table, this paper has bolded the results with higher accuracy. In order to better display the diagnostic results, Figure 4 Presented as line graphs, these results provide an intuitive understanding of the diagnostic performance under different training strategies.

[0116] Table 1.5 Neutron chopper diagnosis results (%)

[0117]

[0118] Note: L1 has no load; L2 has one weight; L3 has two weights; L4 has three weights. 5, 10, and 20 represent the amount of labeled data.

[0119] According to the results, there is a rule under various working conditions: the diagnostic accuracy of the model pre-trained with contrastive predictive coding is always significantly higher than that of the model without pre-training when there is a small amount of labeled data. This situation always exists when the data labels are less than 100, but when the number of data labels exceeds 100, the accuracy of the model without pre-training begins to approach or even surpass the model with pre-training. This is because the model with pre-training can extract excellent feature representations from the unlabeled data, which enables the model to use these pre-trained features for diagnosis when the labeled data is limited, thereby achieving better results. However, when the labels are sufficient, the model without training has no restrictions on optimizing the parameters of the feature extractor, thereby achieving better diagnostic accuracy. This phenomenon shows that the performance of pre-trained and non-pre-trained models will vary under different amounts of labeled data, and the appropriate model strategy should be selected according to the actual situation.

[0120] 1.5.4 Monitoring system programming framework

[0121] To achieve the above-mentioned object of the present invention, according to another aspect of the present invention, the present invention provides a neutron chopper state monitoring system based on multi-sensor data fusion, which can be secondary developed for any other related equipment. Figure 6 Design a framework diagram for the program.

[0122] The monitoring system has four modes to choose from, namely monitoring mode, saving mode, diagnosis mode and history mode. When the program is running, first set the sampling frequency of the vibration signal, and judge whether the machine is in operation by whether there is data input. After the program is running, it is in monitoring mode by default. At the same time, you can choose the saving mode, diagnosis mode and history mode by yourself. The saving mode can save the collected vibration signal to the hard disk, the diagnosis mode loads the current vibration signal input by the Python model as a sample for diagnosis, and the history mode allows users to select saved data for observation.

[0123] Figure 7 The interactive interface of the neutron chopper monitoring and diagnosis system is displayed, in which window ① displays the real-time kurtosis of the vibration signal, window ② displays the real-time skewness, window ③ displays the Fourier transform of the real-time signal, window ④ displays the real-time vibration signal, and window ⑤ allows users to select the current mode. When the diagnosis mode is selected, the current fault type is displayed.

[0124] 1.6 Summary

[0125] This application proposes a neutron chopper fault diagnosis algorithm based on contrast prediction coding multi-sensor data fusion. First, the theoretical basis of the contrast coding model is introduced, and the multi-sensor data fusion algorithm model is described in depth. Then, the algorithm is applied to the neutron chopper data set, and an experimental comparison is conducted under the condition of unlabeled data. The results show that in the visual qualitative evaluation, it can be clearly seen that the feature clustering effect of the four-sensor data is more obvious than that of the single-sensor data. In the quantitative evaluation of the inter-class distance and intra-class distance, the average value of the feature intra-class distance extracted when training with the four-sensor data is 36.253 lower than that of the single sensor, while the average value of the inter-class distance of the four-sensor data is 52.629 higher than that of the single sensor data. From the results, it can be seen that the multi-sensor feature fusion method can make full use of the complementarity between different sensors and can obtain better feature extraction effects, which is of great significance for fault diagnosis tasks, because better feature extraction can improve the model's ability to identify faults or abnormal conditions. Therefore, combining the multi-sensor feature fusion method helps to improve the feature effect in self-supervised training and further improve the accuracy and reliability of fault diagnosis.

[0126] Secondly, a classification experiment of fault diagnosis is designed for the case of sparse data labels. The experimental results show that the diagnostic accuracy is higher on the basis of contrastive predictive coding pre-training. It is worth noting that when the number of data labels is less than 100, the diagnostic accuracy of the contrastive coding pre-training is always higher than that of the model without pre-training. From the results, it can be seen that the fewer the number of data labels, the more obvious the advantage of the proposed method in diagnostic accuracy over the method without pre-training. Therefore, the proposed method has a high application value in the case of sparse data labels of neutron choppers, and provides an effective solution for the field of neutron chopper fault diagnosis. It also provides new ideas and methods for improving the fault diagnosis capability of neutron choppers, which helps to improve the performance and reliability of the equipment.

[0127] It should be noted that, in this article, the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or apparatus.

[0128] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. The above is only a preferred implementation of the present invention. It should be pointed out that due to the limitations of textual expression and the objective existence of infinite specific structures, ordinary technicians in this technical field can make several improvements, modifications or changes without departing from the principles of the present invention, and can also combine the above technical features in an appropriate manner; these improvements, modifications, changes or combinations, or the direct application of the inventive concept and technical solution to other occasions without improvement, should be regarded as the protection scope of the present invention.

Claims

1. A neutron chopper fault diagnosis method based on multi-sensor data fusion, characterized in that: The steps include: Step 1: Design a contrastive coding model. The contrastive coding model includes a feature extractor Genc and an autoregressive model Gar. The feature extractor Genc extracts features from multiple one-dimensional data samples and fuses them into a feature vector z t , the feature extraction operation can be expressed as: t =Genc(x t ); the autoregressive model Gar represents z with the features of the previous time step ≤t as input and pass c t =Gar(z ≤t ) to generate the contextual latent feature representation c t , then, using the output c of the autoregressive model t To predict the feature representation z of the next time step >t ; Input data x t , t is the current input time step, Z t is the feature vector of this step Step 2: Collect the raw data of the neutron chopper through multiple sensors and convert it into a frequency domain signal using Fast Fourier Transform (FFT); Step 3: Input the converted frequency signal into a comparison coding model, and predict the fault state of the neutron chopper through the coding model.

2. A neutron chopper fault diagnosis method based on multi-sensor data fusion according to claim 1, characterized in that: The model encodes the input data x through a series of nonlinear transformations, learns the latent semantics c of the context, and converts it into a vector form, which is defined as: I(x,c) is the mutual information between the original signal x and the contextual latent semantics c; The goal of the model is to maximize the mutual information between the original signal and the contextual latent semantics c. The model adopts the method of constructing a density function f, which retains the model input sample x t+k and contextual latent semantics c t The mutual information between them is expressed as: Where k represents the kth moment, ∝ represents proportional to, and a concise log-linear model equation is used to represent f: Among them, W k c t It is used to express that for each step k, a linear converter W is used k To make predictions; by calculating the density ratio f k (x t+k ,c t ), the model can evaluate the prediction results Accuracy in high-dimensional distributions, and distinguishing the relationship between its own model and input data to achieve effective separation.

3. The neutron chopper fault diagnosis method based on multi-sensor data fusion according to claim 1, characterized in that: The feature extractor Genc is divided into three modules: feature extraction, feature concatenation and feature fusion. First, the preprocessed one-dimensional vibration signal is input into the designed convolutional neural network model for feature extraction. After feature extraction, multiple features are obtained respectively. Then, the features are concatenated into a feature vector through the feature concatenation layer and sent to the fully connected layer for fusion to obtain the z in CPC. t =Genc(x t ); mathematically summarized as follows: Assume that there are m different sensors collecting data at different locations, and the expression is: y(t)=[y1(t),y2(t)…y m (t)] Where y(t) is the input data set; The length of the signal data is reconstructed into a matrix of the input of the feature extractor Genc, ​​as follows: The extracted features are concatenated as shown in the equation: The obtained features are concatenated and input into the feature fusion layer to obtain the z required in CPC t =Genc(x t ).

4. The neutron chopper fault diagnosis method based on multi-sensor data fusion according to claim 2 is characterized in that: In the contrastive coding model, the infoNCE function is used to train the feature extractor Genc and the autoregressive model Gar. It evaluates the performance of the model by calculating the cross entropy between positive and negative samples. The positive samples are obtained from P(x t+k |c t ), and the negative sample is a sample randomly selected from P(x t+k ) randomly selects N-1 samples; the expression of its loss function is as follows: Where X={x1,x2,…,x N } is a set of samples, (x t+k ,c t ) is a positive sample pair, (x j ,c t ) is a negative sample pair, that is, the positive sample comes from the current context c t The samples are separated by k steps, and the negative samples come from samples randomly selected from the sequence. The above equation is used to calculate the model prediction value f k / ∑ X f k The corresponding cross entropy, in this equation, f k represents the predicted value of the model, and ∑ X f k It represents the sum of the predicted values ​​of all samples, and further introduces the optimal loss probability p(d=i|X,c t ), where the sample x i The positive sample in can be expressed as d = i, from the conditional distribution p(x t+k |c t ) is a sample x extracted from i The probability can be expressed by the following equation: f k (x t+k ,c t ) is the optimal value as p(x t+k |c t ) / p(x t+k ) increases, that is, it is a proportional relationship, and the selection of the number of negative samples N-1 will not affect the proportional relationship between the two; x t+k With variable c t The mutual information between can be evaluated by the following expression: I(x t+k ,c t )≥log(N)-L N As the loss L is minimized N At the same time, the lower bound of the common information can be maximized, and the mutual information will become more solid.

Images