A multi-depth network fusion migration distillation multi-working-condition rotating machinery fault diagnosis model

Abstract

The application discloses a kind of multi-depth network fusion transfer distillation multi-working condition rotating machinery fault diagnosis model, including feature map conversion module, multi-depth network fusion transfer module and knowledge distillation module, feature map conversion module is used to the original signal data and load working condition data, after normalization respectively, coding conversion is in the three channels of image;Multi-depth network fusion transfer module obtains the data according to feature map conversion module, using parmeter based transfer learning The way is simultaneously transferred and trained to multiple networks;Knowledge distillation module is used to compress the model of multi-depth network fusion transfer module.The multi-depth network fusion transfer learning multi-working condition rotating machinery fault diagnosis model of the application carries out integration to WideResNet module, ResNeSt module and ResNet152 module, reduces the time required for model training.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of rotating machinery fault diagnosis, and relates to a multi-depth network fusion transfer distillation multi-working-condition rotating machinery fault diagnosis model. BACKGROUND

[0002] With the development of steam engines and electric motors, rotating machinery replaces manual labor and is widely used in various fields of human production, life and work. Rotating machinery has a certain service life, and if the rotating machinery fault is not found in time, great losses may be suffered. Therefore, many researchers have successively studied the rotating machinery fault diagnosis method.

[0003] However, the existing researches have the following problems: (1) the single model fault diagnosis effect is poor; (2) the transfer learning training time is short; (3) the efficiency of the integrated model is low, and the required storage space is large, so each fault diagnosis model still has its own advantages and limitations, and the deep fault diagnosis model training still has the defect of long training time. SUMMARY

[0004] The technical problem to be solved by the application is to provide a multi-depth network fusion transfer distillation multi-working-condition rotating machinery fault diagnosis model to solve the technical problems in the prior art.

[0005] The technical scheme adopted by the application is: a multi-depth network fusion transfer distillation multi-working-condition rotating machinery fault diagnosis model, comprising a feature map conversion module, a multi-depth network fusion transfer module and a knowledge distillation module, the feature map conversion module is used to normalize and encode the original signal data (fan end and drive end data of the rotating machinery) and the load working condition data into three channels of an image; the multi-depth network fusion transfer module adopts a parameter based transfer learning mode to simultaneously transfer and train multiple networks according to the data obtained by the feature map conversion module; and the knowledge distillation module is used to compress the model of the multi-depth network fusion transfer module.

[0006] The multi-depth network fusion module includes a fused WideResNet module, ResNeSt module, and ResNet152 module. The ResNet152 module uses residual connections and employs BatchNorm for model regularization. The WideResNet module reduces the depth and increases the width based on the ResNet152 module, exploring model optimization ideas from the width direction. The WideResNet module uses the dropout module method for regularization. The ResNeSt module retains the idea of ​​residual connections in the ResNet152 module, divides the input feature map into G basis arrays along the channel dimension, and further splits the basis arrays into N modules. Thus, the average pooling combination of the g-th basis array is represented as shown in formula (1):

[0007]

[0008] Where N represents the number of splits in the base array, H and W represent the height and width of the image, respectively, and F k (i,j) represents the transformation operation performed on the (i,j) pixel in the feature map X at the k-th split. The attention calculation formula for the n-th split in the g-th basis array is shown in formula (2):

[0009]

[0010] In the formula, f j Let f represent the transformation function for the j-th input. n This represents the transformation function for the nth input, where the transformation function is a mapping trained by the neural network;

[0011] Therefore, the attention-weighted representation of the g-th basis array is shown in Equation (3):

[0012]

[0013] Among them, F N(g-1)+i (X) represents the transformation operation performed on the feature map X at the N(g-1)+ith split; the final expression after combining the basis arrays and adding the residual term is shown in formula (4):

[0014] Y = V b +Concat{V 1 V 2 ,...,V G} (4)

[0015] Among them, V b This represents the residual term.

[0016] The fusion formula for the fusion model:

[0017]

[0018] In the formula, h n’ (x) represents the output of the n'th model; n' represents the n'th model; M represents the total number of integrated models M; w n ' represents the weight of the nth model; H(x) represents the overall output.

[0019] The knowledge distillation module uses ResNet18 as the student model. Knowledge distillation is performed on the model of the multi-deep network fusion module at a temperature of T=7. The softmax function after introducing temperature T is shown in formula (5).

[0020]

[0021] Where T represents temperature, y is the dependent variable (the output of the current layer), x is the independent variable (the output of the previous layer of the neuron (the input of the current layer)), l represents the l-th input variable, and exp represents the exponential function.

[0022] The beneficial effects of this invention are as follows: Compared with the prior art, the multi-deep network fusion transfer learning multi-condition rotating machinery fault diagnosis model of this invention integrates the WideResNet, ResNeSt, and ResNet152 modules, reducing the training time required for the model. Transfer learning of the multi-deep network fusion model significantly improves the model iteration efficiency, enabling rapid model convergence. Simultaneously, the transfer and fine-tuning of the source model parameters allows the model to effectively improve fault diagnosis accuracy, generalization ability, and fault diagnosis stability under various operating conditions with only a small dataset. The method of integrating operating load, drive end, and fan end vibration signals into the three channels of the fault feature map respectively enables the model to comprehensively utilize various information to identify faults under different operating conditions. Furthermore, it transforms the one-dimensional time series classification problem into a two-dimensional image classification problem, facilitating automatic feature extraction using two-dimensional convolutional kernels. The knowledge distillation algorithm is used to compress the model, reducing the storage space required by the TLMDN model and improving model inference efficiency. Attached Figure Description

[0023] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0024] The present invention will be further described below with reference to specific embodiments.

[0025] Example 1: As Figure 1As shown, a multi-deep network fusion transfer distillation multi-condition rotating machinery fault diagnosis model mainly includes three modules: a feature map transformation module, a multi-deep network fusion transfer module, and a knowledge distillation module. The following sections will provide a detailed explanation of each module of the TLMDN model.

[0026] (1) Feature Map Conversion Module

[0027] Feature map conversion module, such as Figure 1 The left half shows the main function of normalizing the raw signal data (fan and drive data from the Western Reserve University fault diagnosis dataset) and load condition data, then encoding and converting them into the three channels of the image. This forms a feature map containing the complete information needed for fault diagnosis, transforming the one-dimensional time series classification problem into a two-dimensional image classification problem. Each type of fault feature map has its own distinct texture features, providing the foundation for using deep convolutional neural networks for fault diagnosis.

[0028] (2) Multi-depth network fusion migration module

[0029] Multi-depth network fusion modules such as Figure 1 As shown in the middle section, it mainly consists of three parts: the WideResNet module, the ResNeSt module, and the ResNet152 module. Each module converts the vibration signal (one-dimensional sequence data) into a two-dimensional matrix (i.e., an image) in a row-major manner, and obtains a feature map through a convolutional layer. Transfer learning involves using the parameters of models trained on other data to initialize its own model. The transfer learning capability effectively solves the problem of insufficient labeled data and great difficulty in data labeling, which is more common in rotating machinery fault diagnosis than in other fields. The WideResNet module, the ResNeSt module, and the ResNet152 module will be introduced below.

[0030] The fusion formula for the fusion model of the three modules:

[0031]

[0032] In the formula, h n’ (x) represents the output of the n'th model; n' represents the n'th model; M represents the total number of integrated models M; w n ' represents the weight of the nth model; H(x) represents the overall output.

[0033] The ResNet152 module primarily uses residual connections to improve the model's fitting and expressive capabilities from a depth perspective, and mainly employs BatchNorm for model regularization.

[0034] The WideResNet module primarily reduces the depth and increases the width of ResNet, exploring model optimization strategies from the width perspective. WideResNet improves the model's training performance. Since the number of parameters increases rapidly with increasing width, dropout is used for regularization.

[0035] The ResNeSt module, while retaining the idea of ​​ResNet residual connections, divides the input feature map into G basis arrays along the channel dimension, and further splits the basis arrays into N modules. Thus, the average pooling combination of the g-th basis array is represented as shown in Equation (1):

[0036]

[0037] Where N represents the number of splits in the base array, H and W represent the height and width of the image, respectively, and F k (i,j) represents the transformation operation performed on the (i,j) pixel in the feature map X at the k-th split. The attention calculation formula for the n-th split in the g-th base array is shown in formula (2):

[0038]

[0039] In the formula, f j Let f represent the transformation function for the j-th input. n This represents the transformation function for the nth input, where the transformation function is a mapping trained by the neural network;

[0040] Therefore, the attention-weighted representation of the g-th basis array is shown in Equation (3):

[0041]

[0042] Among them, F N(g-1)+i (X) represents the transformation operation performed on the feature map X at the N(g-1)+ith split. The final expression after combining the basis arrays and adding the residual term is shown in Equation (4):

[0043] Y = V b +Concat{V 1 V 2 ,...,V G} (4)

[0044] Among them, V b This represents the residual term.

[0045] Multi-depth network migration modules such as Figure 1As shown in the right half, since general object recognition can extract image features, and the textures of each fault type have their own unique characteristics, this paper uses parameter-based transfer learning to simultaneously transfer and train multiple networks. This paper refines the network, eliminating the need for retraining from scratch and saving time; simultaneously, the source models for transfer are pre-trained on large datasets, resulting in stronger robustness and generalization ability.

[0046] (3) Knowledge Distillation Module

[0047] Knowledge distillation module, such as Figure 1 The lower half of the figure shows that ResNet18 is used as the student model. Knowledge distillation is performed on the TLMDN model at a temperature of T=7. The main function of temperature T is to make the softmax function output smoother. The softmax function after introducing temperature T is shown in formula (5).

[0048]

[0049] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the claims.

Claims

1. A multi-depth network fusion migration distillation multi-condition rotating machinery fault diagnosis device, characterized in that: The system includes a feature map transformation module, a multi-deep network fusion transfer module, and a knowledge distillation module. The feature map transformation module normalizes the original signal data and load condition data. The original signal data includes vibration signals from the fan end and drive end of rotating machinery. The normalized fan end vibration signal data, drive end vibration signal data, and load condition data are encoded and transformed into three channels of the image, respectively. The multi-deep network fusion transfer module uses parameter-based transfer learning to simultaneously transfer and train multiple networks based on the data obtained from the feature map transformation module. The knowledge distillation module compresses the model from the multi-deep network fusion transfer module. The multi-depth network fusion module includes a fused WideResNet module, ResNeSt module, and ResNet152 module. The ResNet152 module uses residual connections and employs BatchNorm for model regularization. The WideResNet module reduces the depth and increases the width based on the ResNet152 module, exploring model optimization ideas from the width direction. The WideResNet module uses the dropout module method for regularization. The ResNeSt module retains the idea of ​​residual connections in the ResNet152 module, divides the input feature map into G basis arrays along the channel dimension, and further splits the basis arrays into N modules. Thus, the average pooling combination of the g-th basis array is represented as shown in formula (1): （1）， Where N represents the number of splits in the base array, H and W represent the height and width of the image, respectively, and F k (i, j) represents the transformation operation performed on the (i, j) pixel in the feature map at the k-th split, and the attention calculation formula for the n-th split in the g-th base array is shown in formula (2): （2）， In the formula, Indicates the first Transformation function for each input, f n This represents the transformation function for the nth input, where the transformation function is a mapping trained by the neural network; Therefore, the attention-weighted representation of the g-th basis array is shown in Equation (3): （3）， in, The feature map X represents the first time. The transformation operations performed on each split; the final expression after combining the base arrays and adding the residual term is shown in formula (4): （4）， in, This represents the residual term.

2. The multi-depth network fusion migration distillation multi-condition rotating machinery fault diagnosis device according to claim 1, characterized in that: Model fusion formula: ， In the formula, h n’ (x) represents the output of the n'th model; n' represents the n'th model; M represents the total number of integrated models M; w n’ H(x) represents the weight of the n'th model; H(x) represents the overall output.

3. A multi-depth network fusion migration distillation multi-condition rotating machinery fault diagnosis device according to any one of claims 1-2, characterized in that: The knowledge distillation module uses ResNet18 as the student model. When performing knowledge distillation on the model of the multi-deep network fusion module, the softmax function after introducing temperature T is shown in formula (5): （5）， Where T represents temperature, y is the dependent variable, x is the independent variable, l represents the l-th input variable, and exp represents the exponential function.

Images