WO2024060381A1

WO2024060381A1 - Incremental device fault diagnosis method

Info

Publication number: WO2024060381A1
Application number: PCT/CN2022/131657
Authority: WO
Inventors: 乔非; 关柳恩
Original assignee: 同济大学
Priority date: 2022-09-20
Filing date: 2022-11-14
Publication date: 2024-03-28
Also published as: CN115510963A

Abstract

An incremental device fault diagnosis method, using a trained fault diagnosis model to process data to be diagnosed having a new sample influx, to obtain a fault diagnosis result of a device. A construction method for the fault diagnosis model comprises: 1) constructing a complete sample set; 2) constructing an initial fault diagnosis model, and training the initial fault diagnosis model by using the complete sample set; 3) selectively retaining important sample subsets by using a sample retention method; 4) during the new sample influx, constructing an intermediate fault diagnosis model on the basis of the structure and parameters of the initial fault diagnosis model; and 5) on the basis of a knowledge distillation algorithm, jointly training the intermediate fault diagnosis model by using a new sample set and the important sample subsets, and obtaining and testing a final fault diagnosis model. The present invention realizes effective learning of new samples and retention of old samples, so that not only is the capability to determine new fault types achieved, but also good memory capability is kept for historical samples.

Description

An incremental equipment fault diagnosis method

Technical field

The invention belongs to the technical field of equipment fault diagnosis and relates to an incremental equipment fault diagnosis method.

Background technique

The complexity of modern systems has increased. Therefore, higher requirements are put forward for safety and stability. Equipment system faults need to be identified, diagnosed and quickly restored in a timely manner to avoid more serious economic losses and personal safety accidents. At present, the establishment of fault diagnosis models requires the strong assumption that training samples and test samples are equally distributed. However, as the equipment continues to operate, new fault characteristics and types may appear, and the original model is no longer applicable. Therefore, fault diagnosis needs to be The model is updated in a timely manner. The update process of fault diagnosis models usually requires combining new and old samples, that is, using a complete sample set in batch learning mode. However, in this process, the equipment has accumulated a large number of historical samples after running for a long time, and the storage cost is high. If the complete sample set is used to retrain the model, the time cost and calculation cost are relatively high; if only new samples are used to adjust the model parameters, the model is easy to Gradually forget old samples; too high an update frequency will cause unnecessary consumption, while too low an update frequency will make it difficult to ensure the performance of the model. In order to solve the above problems, people have proposed an incremental learning update method, which can continuously learn new knowledge from new samples that are continuously generated, and can also retain most of the old knowledge. It does not need to save all historical samples and reduces the occupation of storage space. Make full use of historical training results to improve model training efficiency.

In order to make the diagnostic model have the ability of incremental learning, people extend the functions at the algorithm level based on machine learning, deep learning and other methods. Among them, deep learning algorithms have more advantages than traditional machine learning algorithms in abstract feature representation and are widely used in the field of fault diagnosis. Most of the methods for achieving incremental learning in fault diagnosis models established based on deep learning adopt the structure and parameters of the model. Extension, but each update of the model requires careful design of additional network structures or allocation of weights.

For learning new samples, after searching the Chinese Publication No. CN112508192A, an incremental stacked width learning system with a deep structure was disclosed. This method incrementally updates the model by stacking multiple width learning systems, although only new parameters are added. Training is performed, but the direction of the new model structure is complicated and its feature mapping relationship is greatly restricted by the fixed old parameters, which is not conducive to improving the fitting ability of the model; for historical sample retention, in order to significantly reduce the data storage cost and model To reduce the training cost, people try to select and retain important samples from the complete sample set to replace the complete sample set. The common method of retaining important samples in the prior art is the sample retention method based on the nearest sample mean (NME). However, this method only considers the optimal sample of the current iteration rather than considering the whole. The selected sample is essentially the local optimal sample. Excellent sample. When there is a sample in the complete sample set that is very close to the mean center, there will be multiple samples of the same sample in the sample subset, thereby reducing the diversity of the sample set, causing serious information loss, and causing the model to be unable to effectively retain the knowledge of old samples. To sum up, the existing technology has the disadvantages of high model complexity and inability to effectively manage historical samples.

Contents of the invention

The purpose of the present invention is to provide an incremental equipment fault diagnosis method driven by a small number of labeled old samples to overcome the problems of high model complexity and poor historical sample management effects.

The object of the present invention can be achieved through the following technical solutions:

An incremental equipment fault diagnosis method, which uses a trained fault diagnosis model to process the data to be diagnosed with the influx of new samples to obtain fault diagnosis results of the equipment. The construction method of the fault diagnosis model includes:

Step S1: Obtain and process sensor data related to device status, and build a complete sample set;

Step S2: Construct an initial fault diagnosis model based on the deep neural network, and use the complete sample set to train the initial fault diagnosis model;

Step S3: Based on the sample retention method in the genetic algorithm, selectively retain important sample subsets from the complete sample set that are used to characterize the statistical characteristics of the complete sample set;

Step S4: When new samples pour in, based on the structure and parameters of the initial fault diagnosis model, an intermediate fault diagnosis model is constructed and its parameters are initialized;

Step S5: Adjust the objective function of the intermediate fault diagnosis model for parameter optimization. Based on the knowledge distillation algorithm, use the new sample set and the important sample subset to jointly train the intermediate fault diagnosis model, obtain and test the final fault diagnosis model, and end.

Further, the described construction of a complete sample set includes normalizing the sensor data and limiting the value to [0, 1]; aligning each sensor time and cutting it into several signal segments, each segment is used as a samples to construct a complete sample set.

Further, the selective retention of important sample subsets is achieved through a sample retention method based on genetic algorithms, including:

S310. Screen the sample set correctly classified by the initial fault diagnosis model;

S320. Binary encode the filtered sample set index to form a gene;

S330, randomly initialize the population;

S340. Calculate the fitness of each individual of the population;

S350. Perform roulette selection, two-point crossover and multi-point mutation operations on the population, and return to step S340 until the iteration stop condition is met and the final population is generated;

S360. Decode the optimal individuals in the final population to obtain important sample subsets.

Further, the fitness calculation method includes:

Step S341: Decode all individuals in the population generated by the current iteration to obtain the sample subset corresponding to each individual;

Step S342: Enter the complete sample set and the current sample subset into the fault diagnosis model respectively to obtain respective logits vector sets;

Step S343: Calculate the mean center of the logits vector set of the complete sample set and the current sample subset, and obtain

and μ, calculate the fitness of each individual.

Further, the calculation formula for calculating the fitness of each individual is:

Among them, F is the fitness of each individual.

Furthermore, the construction of the intermediate fault diagnosis model and initialization of its parameters include:

Step S410: Construct an intermediate fault diagnosis model with the same structure as the initial fault diagnosis model, and update the number of output neurons of the intermediate fault diagnosis model, where the number of output neurons is the same as the number of fault categories included in the sample set;

Step S420: Load the neuron weights and biases of the initial fault diagnosis model into the intermediate fault diagnosis model as its initial training weights and parameters, and initialize the excess neuron weights and biases to simulate zero output values.

Further, the final fault diagnosis model is implemented based on the knowledge distillation algorithm, including:

Step S510: Freeze the initial fault diagnosis model parameters so that they do not participate in the parameter optimization process, and merge the new sample set and important sample subsets into a training sample set;

Step S520: Input the training sample set into the initial fault diagnosis model and the intermediate fault diagnosis model at the same time. Under the adjustment of the temperature coefficient T, the soft label and soft prediction distribution of the old category are obtained respectively, and then the total distillation loss function is obtained, and the two are calculated. Distillation loss between

Step S530: Input the training sample set into the intermediate fault diagnosis model to obtain the prediction distribution of all categories, and calculate the cross-entropy loss between the prediction distribution of all categories and the real labels of the training sample set;

Step S540: Add the distillation loss and the cross-entropy loss to obtain the total loss. The total loss function is used as the objective function to reversely optimize the parameters of the intermediate fault diagnosis model to obtain the final fault diagnosis model.

Further, testing the final fault diagnosis model includes setting the temperature coefficient T to 1, inputting test samples into the model to obtain classification results, and performing performance evaluation.

Furthermore, the extra neuron weights and configurations are initialized to 1×10 ⁻⁶ .

Furthermore, the formula of the total distillation loss function is:

Among them, T represents the temperature coefficient, T is greater than 1; softmax is the normalized exponential function; cls _n and cls _o represent the number of new and old categories respectively;

and

Represents the i-th pixel of the feature map output by a certain layer of the old model and the new model respectively;

A soft label representing the output of the old model,

Represents the soft prediction distribution related to the old category output by the new model; θ represents the parameters of the deep neural network; ρ _l represents the constant coefficient of the l-th distillation network layer;

represents the distillation loss of the lth network layer, and L _kd represents the total distillation loss function.

Compared with the existing technology, the present invention has the following characteristics:

1. This invention is based on the knowledge distillation algorithm, uses the new sample set and important sample subsets to obtain the total loss function as the objective function and jointly trains the constructed intermediate fault diagnosis model to obtain the final training model, realizing the migration of old samples and effective learning of new samples, so that It not only has the ability to identify new fault characteristics and new fault types, but also has the ability to maintain good memory for historical samples.

2. The present invention builds an intermediate fault diagnosis model based on the structure and parameters of the initial fault diagnosis model, which can mine the potential correlation between new and old samples from the initial fault diagnosis model, adaptively constrain the direction of model parameter optimization, and further Reduce memory consumption.

3. The present invention establishes a fault diagnosis model based on deep learning, which has excellent nonlinear feature representation and fitting capabilities, and can extract key features from a large number of samples and accurately identify fault categories.

Description of drawings

Figure 1 is a flow chart of the incremental equipment fault diagnosis method of the present invention;

Figure 2 is a schematic diagram of chromosome coding according to the present invention;

Figure 3 is a flow chart of the GA-based sample retention method of the present invention;

Figure 4 is a schematic diagram of the update model structure of the present invention;

Figure 5 is an application flow chart of the knowledge distillation method of the present invention, wherein (5a): a flow chart of the knowledge distillation method of the present invention applied to the training process; (5b): a flow chart of the knowledge distillation method of the present invention applied to the testing process;

Figure 6 shows the experimental results of the incremental equipment fault diagnosis method of the present invention, where (6a): the experimental result when the number of sample retention is 5; (6b): the experimental result when the number of sample retention is 10; (6c): sample Experimental results when the retention number is 20; (6d): Experimental results when the sample retention number is 30.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention and provides detailed implementation modes and specific operating procedures. However, the protection scope of the present invention is not limited to the following embodiments.

Example:

This embodiment proposes an incremental equipment fault diagnosis method. This method uses a trained fault diagnosis model to process the data to be diagnosed with the influx of new samples, and obtains the fault diagnosis results of the equipment. The construction method of the fault diagnosis model includes:

Figure 1 shows a schematic flow chart of the incremental equipment fault diagnosis method, including the following steps:

Step S2: Construct an initial fault diagnosis model based on the deep neural network, train the initial fault diagnosis model using a complete sample set, and optimize its parameters;

Step S3, selectively retaining an important sample subset from the complete sample set for characterizing the statistical characteristics of the complete sample set;

Step S4: When new samples pour in, build an intermediate fault diagnosis model based on the initial fault diagnosis model and initialize parameters;

Step S5, adjust the objective function of the intermediate fault diagnosis model parameter optimization, based on the knowledge distillation algorithm, use the new sample set and the important sample subset to jointly train the intermediate fault diagnosis model, obtain and test the final fault diagnosis model, and end.

Among them, step S1 includes:

Step S110: Select data representing various faults and health states of the equipment from a variety of sensor data for monitoring equipment status, where the status data must be continuously measured and recorded.

Step S120: Perform normalization calculation on the sensor data and limit the value to [0, 1]. For each sensor time, it is cut into several signal segments, and each segment is used as a sample to construct a complete sample set.

Among them, step S2 includes:

Step S210: Build a deep neural network model based on CNN, and use fully connected layers and softmax layers to implement multi-classification tasks. The size of the input feature map and the number of channels are determined based on the sampling size and the number of data sources. The final output is the probability vector that the input sample belongs to each fault category.

Step S220: Preset all hyperparameters related to model training, such as learning rate, maximum number of iterations, etc. Use a complete sample set to train the fault diagnosis model, and use the Adam optimizer to optimize the parameters of the model.

As shown in FIG2 , step S3 selects and stores important sample subsets from the complete sample set to characterize the statistical characteristics of the complete sample set, specifically including:

Step S310: Screen the samples that are correctly classified by the initial fault diagnosis model from the complete sample set;

Step S320: Chromosome binary encoding. The index of the filtered sample set is binary encoded, and the length is equal to the size of the even number of sample subsets that need to be constructed. See Figure 3. "11" represents a sample with the index subscript 11 in the complete sample set, which is encoded into genes in binary "1011", and so on;

Step S330, randomly initialize the population;

Step S340: Calculate the fitness of each individual in the population;

Step S350, use roulette selection, two-point crossover and multi-point mutation as the selection operator, crossover operator and mutation operator to generate a new population, return to step S340 until the iteration stop condition is met and the final population is generated;

Step S360: Decode the optimal individuals in the final population to obtain all sample indices, use corresponding samples to construct sample subsets, and generate important sample subsets.

Among them, the calculation method of fitness in step S340 includes:

Step S342: Input the complete sample set and the current sample subset into the fault diagnosis model respectively to obtain the logits vector set of the complete sample set and the logits vector set of the current sample subset. The logits vector is the output of the last fully connected layer. vector.

and μ, and then calculate the fitness

Obtain the fitness of each individual in the current population.

In a specific implementation, constructing an intermediate fault diagnosis model in step S4 and optimizing its parameters include:

Step S410: Construct an intermediate fault diagnosis model with the same structure as the initial fault diagnosis model. Update the number of output neurons of the intermediate fault diagnosis model. The number of output neurons is the same as the number of fault categories included in the sample set, see Figure 4;

Step S420: Load the weights and biases of the initial fault diagnosis model into this round of intermediate fault diagnosis models as their initial training weights and parameters. Since the number of output layer neurons of the initial fault diagnosis model is less than that of the intermediate fault diagnosis model, The number of output layer neurons, so the extra neuron weights and biases are initialized to 1×10 ^-6 to simulate zero output values.

Among them, the final fault diagnosis model obtained in step S5 includes:

Step S520: Input the training sample set into the initial fault diagnosis model and the intermediate fault diagnosis model at the same time to obtain the soft label and soft prediction distribution of the old category under the adjustment of the temperature coefficient T, and calculate the two based on the ratio of the number of old and new categories. Distillation loss between

Step S530: input the training sample set into the intermediate fault diagnosis model to obtain the predicted distribution of all categories, and calculate the cross entropy loss between the predicted distribution of all categories and the true label of the training sample set;

Among them, testing the final fault diagnosis model includes setting the temperature coefficient T to 1, inputting the test sample into the model to obtain the classification results, and conducting performance evaluation.

Figure (5a) is a schematic flowchart of the application of the knowledge distillation method in the training process of the present invention, and Figure (5b) is a schematic flowchart of the application and testing process of the knowledge distillation method of the present invention.

In a specific implementation, when the training sample set is input into the initial fault diagnosis model and the intermediate fault diagnosis model respectively in step S520, the distillation loss is calculated for the feature maps output by multiple intermediate layers, and the loss is rigidly adjusted according to the change in sample category proportion. coefficient. The specific calculation is as follows:

The calculation formula of the distillation loss function of a single network layer is:

and

A soft label representing the output of the old model,

For deep neural networks, knowledge distillation needs to be performed on each network layer with downsampling function. Since such network layers are usually evenly distributed, the number of distilled network layers can be dynamically adjusted based on the network depth.

According to the changes in the number of new and old categories, the adaptability coefficient is set. When the number of samples in each category is equivalent, the ratio of the number of old categories cls _o to the number of new categories cls _n can be adjusted. Secondly, considering the influence of temperature coefficient T on the amplitude of distillation loss, T2 is used for compensation.

The calculation formula of the total distillation loss function is:

Among them, L _kd is the total distillation loss function, ρ _l represents the constant coefficient of the l-th distillation network layer,

Represents the distillation loss and constant coefficient of the l-th distillation network layer, cls _n and cls _o represent the number of new and old categories respectively.

In a specific implementation, in order to verify the performance of the embodiment, the Case Western Reserve University (CWRU) bearing data set in the United States is used for case study and analysis. The information of the CWRU data set is as follows: There are 3 accelerometers collecting vibration data at different ends, namely the drive end accelerometer data (DE), the fan end accelerometer data (FE) and the basic acceleration data (BA); there are 4 different speeds. The operating states are 1730, 1750, 1772 and 1797rpm respectively; there are 3 different fault diameters, 0.007, 0.014 and 0.021 respectively; there are 3 fault states, namely inner ring fault (IRF) and rolling element fault (BF) And outer ring fault (ORF), in which the outer ring fault also includes 3 measuring points, namely the 6 o'clock position directly located in the load area, the 3 o'clock position orthogonal to the load area, and the 3 o'clock position opposite to the load area. 12 o'clock position.

This embodiment uses the status data of bearings with fault diameters of 0.007, 0.014 and 0.021 operating at a rotation speed of 1797 rpm, covering three faults: IRF, BF and ORF (loaded position 6:00). Bearing status data includes DE, FE and BA data, with a sampling frequency of 12kHz. The fault status of the data set is divided into 9 categories. The specific fault type numbers are shown in Table 1, and the category number range is 0-8; the amount of data contained in each fault type is shown in Table 2. The model structure parameters are shown in Table 3, where cls represents the current number of fault categories, and the output of the convolutional layer with a step size of 2 will be used to calculate the distillation loss.

Table 1 Fault type numbers in the CWRU data set

Table 2 Data volume of each fault type in the CWRU data set

Table 3 Network model parameters

This embodiment is divided into four stages, namely one initial learning stage and three incremental learning stages, which are respectively referred to as "initial stage" and "incremental stage i" (i=1, 2, 3). In the initial stage, there are only samples of fault categories 0-2, and in each subsequent incremental stage, two new categories of samples will appear for model update learning. The experimental results of incremental equipment fault diagnosis are shown in Figure 6. This experiment compared the effects of using only the sample retention method and combining the knowledge distillation and sample retention methods for incremental learning, where "sample retention N" means only using the GA-based sample retention method and retaining N key samples. "Knowledge distillation + sample retention N" means using methods based on knowledge distillation and sample retention for incremental learning and retaining N key samples.

It can be seen from Figure (6a) that "sample retention 5" has the worst incremental learning effect during the entire experiment. By the incremental stage 3, the model has a better performance on the test sets of categories 0-2, 3-4, and 5-6. The diagnostic accuracy of the samples has dropped to 42.16%, 56.39% and 51.55% respectively. Due to the regularization effect of "Knowledge Distillation + Sample Retention 5", by the incremental stage 3, the model's diagnostic accuracy for category 0-2, 3-4, and 5-6 test set samples are 63.17%, 90.59%, and 90.59% respectively. 55.05%, which shows a relatively large accuracy improvement overall.

Similarly, it can be seen from Figure (6b) that the diagnostic accuracy of the test set samples of categories 0-2, 3-4, and 5-6 from "knowledge distillation + sample retention 10" to incremental stage 3 is 83.73% and 97.44% respectively. and 70.79%, which is generally higher than the incremental learning effect of "sample retention 10" (61.37%, 79.71% and 76.97%).

In Figure (6c) and Figure (6d), when the number of retained samples reaches 20 and 30, under the joint action of knowledge distillation and sample retention, the effect of model incremental learning remains basically stable, and it has relatively good fault category discrimination performance both in new categories and old categories. Among them, the total accuracy difference between "knowledge distillation + sample retention 30" and "knowledge distillation + sample retention 20" is 2.29%; if only the incremental stage 3 is considered, the total difference in diagnostic accuracy is 1.35%. This means that the number of retained samples of each old category has increased by 10, but the total increase in diagnostic accuracy is relatively small. The total diagnostic accuracy of "knowledge distillation + sample retention 20" is 1.01% higher than that of "sample retention 30". This means that the accuracy of the model is close to saturation at this time, and the use of knowledge distillation can further reduce the number of retained samples while improving the model's diagnostic accuracy for old category samples.

If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention essentially or the part that contributes to the existing technology or the part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program code. .

The preferred embodiments of the present invention are described in detail above. It should be understood that those skilled in the art can make many modifications and changes based on the concept of the present invention without creative efforts. Therefore, any technical solutions that can be obtained by those skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention and on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

An incremental equipment fault diagnosis method, which is characterized in that the method uses a trained fault diagnosis model to process the data to be diagnosed with the influx of new samples, and obtains fault diagnosis results of the equipment. The construction method of the fault diagnosis model include:

Step S1: Obtain and process sensor data related to device status, and build a complete sample set;

Step S2: Construct an initial fault diagnosis model based on the deep neural network, and use the complete sample set to train the initial fault diagnosis model;

Step S3: Based on the sample retention method in the genetic algorithm, selectively retain important sample subsets from the complete sample set that are used to characterize the statistical characteristics of the complete sample set;

Step S4: When new samples pour in, based on the structure and parameters of the initial fault diagnosis model, an intermediate fault diagnosis model is constructed and its parameters are initialized;

Step S5: Adjust the objective function of the intermediate fault diagnosis model for parameter optimization. Based on the knowledge distillation algorithm, use the new sample set and the important sample subset to jointly train the intermediate fault diagnosis model, obtain and test the final fault diagnosis model, and end.
An incremental equipment fault diagnosis method according to claim 1, characterized in that said constructing a complete sample set includes normalizing the sensor data and limiting the value to [0, 1]; Align the time of each sensor and cut it into several signal fragments. Each fragment is used as a sample to build a complete sample set.
An incremental equipment fault diagnosis method according to claim 1, characterized in that the selective retention of important sample subsets is implemented by a sample retention method based on genetic algorithms, including:

S310. Screen the sample set correctly classified by the initial fault diagnosis model;

S320. Binary encode the filtered sample set index to form a gene;

S330, randomly initialize the population;

S340. Calculate the fitness of each individual of the population;

S350. Perform roulette selection, two-point crossover and multi-point mutation operations on the population, and return to step S340 until the iteration stop condition is met and the final population is generated;

S360. Decode the optimal individuals in the final population to obtain important sample subsets.
An incremental equipment fault diagnosis method according to claim 3, characterized in that the fitness calculation method includes:

Step S341: Decode all individuals in the population generated by the current iteration to obtain the sample subset corresponding to each individual;

Step S342: Enter the complete sample set and the current sample subset into the fault diagnosis model respectively to obtain respective logits vector sets;

Step S343: Calculate the mean center of the logits vector set of the complete sample set and the current sample subset, and obtain
and μ, calculate the fitness of each individual.
An incremental equipment fault diagnosis method according to claim 4, characterized in that the calculation formula for calculating the fitness of each individual is:

Among them, F is the fitness of each individual.
An incremental equipment fault diagnosis method according to claim 1, characterized in that said constructing an intermediate fault diagnosis model and initializing its parameters includes:

Step S410: Construct an intermediate fault diagnosis model with the same structure as the initial fault diagnosis model, and update the number of output neurons of the intermediate fault diagnosis model, where the number of output neurons is the same as the number of fault categories included in the sample set;

Step S420: Load the neuron weights and biases of the initial fault diagnosis model into the intermediate fault diagnosis model as its initial training weights and parameters, and initialize the excess neuron weights and biases to simulate zero output values.
An incremental equipment fault diagnosis method according to claim 1, characterized in that the final fault diagnosis model is implemented based on a knowledge distillation algorithm, including:

Step S510: Freeze the initial fault diagnosis model parameters so that they do not participate in the parameter optimization process, and merge the new sample set and important sample subsets into a training sample set;

Step S520: Input the training sample set into the initial fault diagnosis model and the intermediate fault diagnosis model at the same time. Under the adjustment of the temperature coefficient T, the soft label and soft prediction distribution of the old category are obtained respectively, and then the total distillation loss function is obtained, and the two are calculated. Distillation loss between

Step S530: Input the training sample set into the intermediate fault diagnosis model to obtain the prediction distribution of all categories, and calculate the cross-entropy loss between the prediction distribution of all categories and the real labels of the training sample set;

Step S540: Add the distillation loss and the cross-entropy loss to obtain the total loss. The total loss function is used as the objective function to reversely optimize the parameters of the intermediate fault diagnosis model to obtain the final fault diagnosis model.
An incremental equipment fault diagnosis method according to claim 1, characterized in that said testing the final fault diagnosis model includes setting the temperature coefficient T to 1, inputting the test sample into the model to obtain the classification result, and performing performance testing. evaluate.
An incremental equipment fault diagnosis method according to claim 6, characterized in that the weight and configuration of the extra neurons are initialized to 1×10 -6 .
An incremental equipment fault diagnosis method according to claim 7, characterized in that the formula of the total distillation loss function is:

Among them, T represents the temperature coefficient, T is greater than 1; softmax is the normalized exponential function; cls n and cls o represent the number of new and old categories respectively;
and
Represents the i-th pixel of the feature map output by a certain layer of the old model and the new model respectively;
A soft label representing the output of the old model,
Represents the soft prediction distribution related to the old category output by the new model; θ represents the parameters of the deep neural network; ρ l represents the constant coefficient of the l-th distillation network layer;
represents the distillation loss of the first network layer, and L kd represents the total distillation loss function.