A neural network model selection method for mechanical fault diagnosis based on sampling inspection and failure loss

Abstract

The application discloses a neural network model selection decision method for mechanical fault diagnosis based on sampling inspection and fault loss, constructs a plurality of neural network models for mechanical fault diagnosis; and takes the fault loss caused by the error of each neural network model in mechanical fault classification prediction as the basis for neural network model selection, so that the neural network model with the minimum fault loss can be selected from all the neural network models. According to the application, the actual production loss is taken as the basis for neural network model selection under the condition that the classification prediction accuracy of the neural network is basically the same, and the loss caused by the classification prediction error can be maximally reduced.

Description

TECHNICAL FIELD

[0001] The application relates to a neural network model selection decision method for mechanical fault diagnosis based on sampling inspection and fault loss, and belongs to the technical field of mechanical fault diagnosis. BACKGROUND

[0002] In recent years, neural network algorithms have developed rapidly. In 1943, McCulloch and Pitts proposed the first model of neurons, the MP model. Subsequently, in the late 1950s and early 1960s, Rosenblatt proposed a single-layer perception model, which put the concept of neural networks into practice. In 1986, Rumelhart et al. proposed the back propagation network (BP network). In 2006, Hinton et al. introduced the research of deep learning, which triggered a wave of artificial neural networks. At the same time, the application of neural networks in fault diagnosis is also developing. For example, in bearing fault diagnosis, Tang et al. optimized the support vector machine through sparrow search algorithm, and the bearing fault diagnosis accuracy was improved to 99.25%. Zhang et al. proposed a fault diagnosis method based on Gram angle field and particle swarm optimization convolutional neural network structure, which can obtain better bearing fault diagnosis accuracy than other mainstream convolutional neural network structures, and the average fault diagnosis accuracy can reach 99%. Zhang et al. proposed a bearing fault diagnosis model based on transfer learning, which improved the fault diagnosis accuracy. These literatures compare with other neural networks or optimization algorithms, and take the prediction accuracy as the basis to judge the performance of the model. The prediction accuracy of these neural network models has reached more than 99%, and it is difficult to significantly improve the algorithm.

[0003] With the development of neural networks, many excellent models have appeared, such as GA-BP, CNN, PSO-SVM, LSSVM, etc. Using these models to classify and predict mechanical faults is a hot research topic at present. In the selection of neural network models for mechanical fault diagnosis, the prediction accuracy is often taken as the criterion, and more attention is paid to the accuracy of neural network classification and prediction, while the loss caused by the error type of neural network classification and prediction is ignored. This selection of neural network models in mechanical fault diagnosis based on classification and prediction accuracy cannot be well applied to actual production, and the selected neural network model is not the optimal solution. SUMMARY

[0004] In view of the above-mentioned deficiencies in the prior art in using prediction accuracy as a criterion for selecting neural network models, the purpose of the present invention is to propose a neural network model selection decision method for mechanical fault diagnosis based on sampling inspection and fault loss. Under the condition that the neural network classification prediction accuracy is basically the same, the present invention uses actual production losses as the basis for selecting the neural network model, which can maximize the reduction of losses caused by classification prediction errors.

[0005] The technical solution of the present invention is achieved as follows:

[0006] A decision-making method for selecting a neural network model for mechanical fault diagnosis based on sampling inspection and failure loss is used to construct several neural network models for mechanical fault diagnosis. The failure loss caused by the error in the mechanical fault classification prediction of each neural network model is used as the basis for selecting the neural network model. The neural network model with the smallest failure loss is selected among all the neural network models.

[0007] Furthermore, when constructing the neural network model, the data set is divided into a training set and a test set, and the relationship between input and output features is constructed through the training set to obtain the neural network model; using the sampling inspection principle, the test results of the test set are used to infer the number of classification prediction errors of the constructed neural network model for the actual monitoring object in mechanical fault diagnosis. Different classification prediction errors and numbers correspond to different losses, and the sum of the losses of all classification prediction errors is the fault loss.

[0008] Furthermore, the classification prediction error is divided into three error situations: the monitored object is in a fault state, and the classification prediction is in a healthy state; the monitored object is in a healthy state, and the classification prediction is in a fault state; the monitored object is in a fault state, and the classification prediction is in a fault state but the fault type is judged incorrectly; the former situation belongs to a safety failure, recorded as Type I misjudgment, and the corresponding loss is called a safety failure loss; the latter two situations belong to an efficiency failure, recorded as Type II misjudgment, and the corresponding loss is called an efficiency failure loss; using the sampling inspection principle, based on the occurrence frequency of Type I misjudgment and Type II misjudgment in the test set samples, the frequency of Type I misjudgment and Type II misjudgment of the corresponding neural network model in the actual monitoring process is inferred, and then the safety failure loss and efficiency failure loss of each neural network model are inferred, and the sum of the safety failure loss and efficiency failure loss is the failure loss.

[0009] Compared with the prior art, the present invention has the following beneficial effects:

[0010] The present invention combines the mechanical production process to obtain a new basis for selecting neural network models. While considering the accuracy of neural network predictions, it pays more attention to the types of errors in the classification and prediction of neural network models in mechanical production and the production losses caused. The present invention solves the problem that the neural network model is limited to theory. Based on the failure losses caused in actual production, it truly applies it to mechanical production, and maximizes the reduction of failure losses in the mechanical production process. The present invention draws an analogy between the principle of applying neural networks to mechanical fault diagnosis and the principle of sampling inspection, and makes a detailed division of the types of errors in neural network classification predictions, and based on this, proposes a failure loss function. Experimental results show that the smaller the value of the failure loss function, the smaller the loss caused by classification prediction errors of the neural network model in the actual mechanical production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is the principle diagram of sampling inspection.

[0012] Figure 2 This is the schematic diagram of mechanical fault diagnosis using a neural network model.

[0013] Figure 3 Schematic diagram of the neural network classification prediction error type of the present invention.

[0014] Figure 4 This is a sampling characteristic curve diagram of the present invention.

[0015] Figure 5 This is a schematic diagram of the bearing fault software operation interface I of Example 1 of the present invention.

[0016] Figure 6 This is a schematic diagram of the bearing fault software operation interface II of Example 1 of the present invention.

[0017] Figure 7 Schematic diagram of bearing fault software operation interface III according to embodiment 1 of the present invention.

[0018] Figure 8 Schematic diagram of the bearing fault software operation interface IV according to embodiment 1 of the present invention.

[0019] Figure 9 This is a schematic diagram of the bearing fault software operation interface V according to embodiment 1 of the present invention.

[0020] Figure 10 This is a schematic diagram of the equipment installation in Example 2 of the present invention.

[0021] Figure 11 This is a schematic diagram comparing the prediction accuracy of the BP neural network and the GA-BP neural network in Example 2 of the present invention.

[0022] Figure 12The confusion matrix comparison schematic diagram of the BP neural network and the GA-BP neural network in Embodiment 2 of the present application.

[0023] Figure 13 The prediction accuracy comparison schematic diagram of the convolutional neural network and the GA-BP neural network in Embodiment 2 of the present application.

[0024] Figure 14 The confusion matrix comparison schematic diagram of the convolutional neural network and the GA-BP neural network in Embodiment 2 of the present application. DETAILED DESCRIPTION

[0025] The specific embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0026] The present application is based on a neural network model selection decision method for mechanical fault diagnosis based on sampling inspection and fault loss, and constructs a plurality of neural network models for mechanical fault diagnosis; the fault loss caused by the classification prediction error of each neural network model for mechanical fault is used as the basis for neural network model selection, and the neural network model with the minimum fault loss is selected from all neural network models.

[0027] The present application is based on the fault loss caused in actual production, and is truly applied to mechanical production, thereby maximizing the reduction of fault loss in the mechanical production process. The present application analogizes the principle of neural network application in mechanical fault diagnosis and the principle of sampling inspection, and classifies the classification prediction error types of neural network in detail, and proposes a fault loss function based thereon. Specifically as follows.

[0028] The basic principle of neural network and sampling inspection. The application principle of neural network model in mechanical fault diagnosis is similar to the principle of sampling inspection. Sampling inspection is to randomly extract a certain number of samples from a batch of products to be inspected, and to inspect the samples, and then to determine the quality of the products according to the inspection results of the samples, and the inspection principle is as shown in Figure 1 In the aspect of neural network, a limited data set is first extracted from the monitoring object and divided into a training set and a test set, the relationship between the input and output characteristics is constructed through the training set, and the number of classification prediction errors of the neural network model in mechanical fault diagnosis for the actual monitoring object is inferred through the test set, and the inspection principle is as shown in Figure 2 Both of them use the idea of "sample inference population".

[0029] Classification prediction error types of neural network model

[0030] In the actual mechanical production process, there are three kinds of errors in the neural network classification prediction: the monitoring object is in a fault state, but the classification prediction is a healthy state; the monitoring object is in a healthy state, but the classification prediction is a fault state; the monitoring object is in a fault state, but the classification prediction is a fault state but the fault type is wrong. The former situation belongs to safety failure, which is recorded as Class I misjudgment, and the latter two situations belong to efficiency failure, which are recorded as Class II misjudgment. The neural network classification prediction error types of the present invention are as follows: Figure 3 shown.

[0031] The confusion matrix is ​​a visualization tool in the field of machine learning. It displays the true value of the sample and the predicted value of the model in a confusion matrix, which can be used to intuitively determine the fault type and conduct diagnostic analysis. Take the following confusion matrix as an example:

[0032]

[0033] Each column of the confusion matrix represents the predicted category, and the total number of each column represents the number of data predicted to be of that category; each row represents the true category of the data, and the total number of data in each row represents the number of data instances of that category; the value in each column represents the number of true data predicted to be of that category.

[0034] When the true state of the sample is fault type 3, 5 samples are predicted by the model as healthy state, which is a type I misjudgment; when the true state of the sample is healthy state, 3 samples are predicted by the model as fault type 1, which is a type II misjudgment; when the true state of the sample is fault type 1, 2 samples are predicted by the model as fault type 2, which is a type II misjudgment.

[0035] The sampling characteristic curve uses the percentage of defective products p as the horizontal axis. The key is to find the number of defective products in the batch of products, obtain the acceptance probability L(p) as the vertical axis through probability distribution, and use this to judge the quality status of the entire batch of products. Figure 4 This is a sampling characteristic curve diagram of the present invention.

[0036] p=D / N×100% (1)

[0037]

[0038] In formula (1), p is the percentage of defective products; D is the number of defective products in the batch; and N is the total number of products in the batch. Formula (2) represents a Poisson distribution, where the sample size n and the acceptance number (i.e., the judgment criterion) Ac are the two most basic parameters in the sampling plan.

[0039] For example, using the sampling plan (30, 3), when L(p) = L(5%) = 0.943, it means that out of every 100 batches of products with this quality, approximately 94 batches will be judged as qualified products.

[0040] However, there are many factors that determine whether a product is unqualified, such as size, appearance, and performance. If the basis for each product's unqualified condition is used as the horizontal axis and the probability of acceptance L(p) is obtained through probability distribution, then 1-L(p) is the probability that defects such as size, appearance, and performance will appear in the entire batch of products. Similarly, in the neural network model, the percentage of Class I and Class II misjudgments can be obtained, and their calculation formula is:

[0041] P1=D1 / N×100% (3)

[0042] P2=D2 / N×100% (4)

[0043] P1 is the percentage of Class I misjudgments; P2 is the percentage of Class II misjudgments; D1 is the number of Class I misjudgments in the validation set; D2 is the number of Class II misjudgments in the validation set; N is the total number of validation sets.

[0044] By selecting a fixed sampling scheme (n, AC), we can obtain the acceptance probabilities L(P1) and L(P2). In this case, the acceptance probability L(P1) represents the probability that no Class I misjudgment will occur during the overall operation of the monitored object, and the acceptance probability L(P2) represents the probability that no Class II misjudgment will occur during the overall operation of the monitored object. Therefore, let α1 and α2 be the safety risk coefficient and efficiency risk coefficient, respectively, and their respective calculation formulas are:

[0045] α1=1-L(P1) (5)

[0046] α2=1-L(P2) (6)

[0047] In the formula, the safety risk coefficient α1 represents the probability of type I misjudgment in the overall operation of the monitored object, and α2 represents the probability of type II misjudgment in the overall operation of the monitored object.

[0048] Mechanical fault diagnosis technology, as an emerging discipline, plays a vital role in monitoring, diagnosing, and predicting the operating status of mechanical equipment. Understanding the manifestations of faults is fundamental to mechanical fault diagnosis. Faults in mechanical production often occur in two forms: sudden failures caused by component failure and short-term failures caused by repair or maintenance of mechanical equipment. The former focuses on analyzing the damage caused to the mechanical equipment itself and the safety of operators, known as safety failure losses; the latter focuses on analyzing the production window losses caused by repair or maintenance of mechanical equipment, known as efficiency failure losses.

[0049] The ratio of safety failure loss to efficiency failure loss can be used to calculate the failure loss magnitude t. For example, if the failure loss magnitude t = 1000, a safety failure results in 1000 unit losses, while an efficiency failure results in 1 unit loss.

[0050] When different levels of attention are paid to safety failures and efficiency failures, the magnitude of the failure loss t will change. In actual production processes, it can be divided into three situations:

[0051] ① When the loss from safety failure is significantly greater than the loss from efficiency failure or when it is hoped to avoid casualties due to safety accidents, more attention should be paid to the loss caused by safety failure. Therefore, the value of t should be as large as possible (t tends to infinity) to increase the proportion of safety failure loss in failure loss.

[0052] ② When the loss from safety failure is significantly smaller than the loss from efficiency failure or the possible safety accidents in the production process cause less harm to personnel, more attention should be paid to the failure loss caused by efficiency problems. Therefore, the t value should be as small as possible (t tends to 0) to increase the proportion of efficiency loss in failure loss.

[0053] ③ When the failure losses caused by safety issues and efficiency issues are approximately the same, the two are given equal weight, and the t value is 1. The two losses account for approximately the same proportion of the failure losses. The following table provides a direct comparison of the three scenarios.

[0054] The scale between safety and efficiency Failure loss ratio Focus on safety t tends to positive infinity Safety failure losses account for a large proportion Focus on efficiency t tends to 0 Efficiency failure losses account for a large proportion Both are given equal attention t＝1 The loss ratio is approximately the same

[0055] In the process of mechanical production, once mechanical equipment fails, it often causes huge economic losses and casualties. At this time, the loss due to safety failure is significantly greater than the loss due to efficiency failure, and the magnitude of the failure loss t should be as large as possible.

[0056] When the monitoring object runs m times, mα1 and mα2 represent the number of Class I and Class II misjudgments of the monitoring object, respectively. At this time, the safety failure loss and efficiency failure loss are expressed as tmα1 and mα2, respectively. Let η1 and η2 be the safety failure loss function and efficiency failure loss function, respectively. η1 = safety failure loss / batch number of products; η2 = efficiency failure loss / batch number of products, so the calculation formula is as follows:

[0057] η1=t*a1 (7)

[0058] η2=a2 (8)

[0059] The safety failure loss function and the efficiency failure loss function represent the loss caused by each sample. t is the magnitude of safety and efficiency (i.e., the magnitude of failure loss); α1 is the safety risk coefficient; and α2 is the efficiency risk coefficient.

[0060] Combined with the above analysis, the failure loss function η is proposed, which is the sum of the safety failure loss function and the efficiency failure loss function. Its calculation formula is as follows:

[0061] η=η1+η2(9)

[0062] Therefore, according to formulas (2)(3)(4)(5)(6)(7)(8)(9), from the perspective of safety and efficiency, according to the degree of failure loss of safety and efficiency, we can flexibly select the magnitude between safety and efficiency, and obtain the failure loss function η as the judgment standard for selecting the neural network model. Generally speaking, the smaller the failure loss function η, the smaller the failure loss caused. However, in the actual mechanical production process, in order to avoid major safety accidents, efficiency is often sacrificed to ensure safety. The calculation principle can be expressed as follows: artificially increase the failure loss magnitude t so that the efficiency loss function η2 can be ignored. At this time, the failure loss function is approximately equal to the safety loss function, that is, η=η1=t*α1. Therefore, the selection of the neural network is determined by the safety risk coefficient α1.

[0063] The present invention is described in detail below through two embodiments.

[0064] Example 1:

[0065] Using Matlab R2024a, we designed software for bearing failure detection and applied it to a company's machine tool production line. An investigation revealed that safety incidents caused by bearing failures in this production line cost approximately RMB 2 million, and the production downtime caused by the need to replace faulty bearings cost approximately RMB 10,000. The magnitude of the failure loss was t = 200.

[0066] (1) Input safety failure loss and efficiency failure loss, such as Figure 5 shown.

[0067] (2) Click 'Click' to jump to the interface for selecting the neural network model, such as Figure 6 The neural network models selected in this embodiment include BP neural network, GA-BP neural network and CNN neural network.

[0068] (3) Import the bearing data set, such as Figure 7 shown.

[0069] (4) Select a neural network model, such as Figure 8 shown.

[0070] (5) Calculation results, such as Figure 9 shown.

[0071] The analysis found that the BP neural network model's failure loss value was 44,930 yuan, significantly lower than the GA-BP neural network's failure loss value of 32,390 yuan and the CNN neural network's failure loss value of 79,880 yuan. Therefore, the BP neural network model is more suitable as a neural network model for fault diagnosis of this company's machine tool production line.

[0072] Example 2:

[0073] Example 2 uses the University of Ottawa, Canada bearing dataset for model training and testing to verify the effectiveness of the method and idea. The dataset contains vibration signals collected from bearings in five health conditions under time-varying speed, which are healthy (H), inner race defect fault (IF), outer race defect fault (OF), ball defect fault (BF), and combined defect fault (CF) of inner race, outer race and ball. Each dataset contains two channels: "Channel_1" is the vibration data measured by the accelerometer, and "Channel_2" is the speed data measured by the encoder. All these data are sampled at a frequency of 200,000 Hz, and the sampling time lasts for 10 s. The CPR (Cycles Per Revolution) of the encoder is 1024. The equipment installation of Example 2 is shown in Figure 10 .

[0074] This experiment takes vibration data and speed data as input features and health condition as output features. To analyze the health status of the bearing during the starting stage, the signal values of vibration data under the condition of increasing speed in five health conditions are selected. In the initial stage of testing, the speed is small, and the vibration difference is not obvious, so the vibration data when the speed increases to the maximum and gradually stabilizes are selected. Through observation of the data, it is found that the outer race defect fault (OF) and the ball defect fault (BF) data are similar and are not suitable for analysis together, so the following four categories are mainly studied: healthy (H), inner race defect fault (IF), outer race defect fault (OF), and combined defect fault (CF) of inner race, outer race and ball. Each health condition is selected in units of 10 cycles, a total of 284 groups of data (including vibration data and speed data), and each health condition is numbered as 1, 2, 3, 4 in turn. Among them, 199 groups are selected as the training set to construct the GA-BP neural network, and the remaining 85 groups are used as the test set for performance verification, the fault classification test chart is output, and the confusion matrix is used to show the classification result statistical information of the model to the fault class samples.

[0075] In MATLAB, first, the collected data is divided into training set and test set, and the mapminmax function is used to normalize the data in it, mapping the data to the [0, 1] interval. The newff function is used to establish the BP neural network, with the number of hidden layers set to 10, forming a network structure of 2-10-1. Then the training parameters are determined: the maximum number of iterations is 1000, the target training error is 1x10^(-6), and the learning rate is 0.01. Finally, simulation testing and data processing are performed to obtain the test set prediction results. After optimizing the weight and threshold of the BP neural network by genetic algorithm, the above steps are repeated to obtain the test set prediction result comparison, as shown in Figure 11 and Figure 12 .

[0076] By comparing the BP neural network and the GA-BP neural network, it was found that the prediction accuracy of the two is the same, both 92.9412%, and they have strong similarities in terms of confusion matrix. The only difference is the number of error types they make. In the case of 85 total samples, the BP neural network makes 1 type I misjudgment and 5 type II misjudgments, while the GA-BP neural network makes 0 type I misjudgments and 6 type II misjudgments. The type I misjudgment percentage P1 and type II misjudgment percentage P2 of each can be obtained by formula (3) and (4), and the acceptance probability L(p) is then obtained by formula (2). The fault loss function η is obtained by formulas (5), (6), (7), (8), and (9) to judge the performance of the model.

[0077] When the sampling plan is (85, 30, 3) and t=100, the data obtained by the two models are shown in the following table:

[0078]

[0079] When the sampling scheme is (85, 30, 3) and t=1000, the data obtained by the two models are shown in the following table:

[0080]

[0081] From the table, we can see that when the failure loss magnitude t = 100 (i.e., a safety failure loss of 100 units and an efficiency failure loss of 1 unit), the BP neural network is selected because its failure loss function η = 0.153 is smaller than that of the GA-BP neural network (η = 0.1648). Similarly, when the failure loss magnitude t = 1000 (i.e., a safety failure loss of 1000 units and an efficiency failure loss of 1 unit), the GA-BP neural network is selected because its failure loss function η = 0.603 is larger than that of the GA-BP neural network (η = 0.1648). This shows that as the failure loss magnitude t increases, the choice of neural network changes, and the advantage of neural networks with a lower percentage of Type I misclassifications becomes increasingly apparent.

[0082] According to the above experimental method, the experimental results of CNN neural network are obtained. By comparing CNN neural network and GA-BP neural network, it is found that Figure 13 and Figure 14As shown in the figure, the prediction accuracy of the CNN neural network is significantly higher than that of the GA-BP neural network. However, the confusion matrix shows that, when the total sample size is 85, the CNN neural network makes 2 type I errors and 3 type II errors, while the GA-BP neural network makes 0 type I errors and 6 type II errors. The type I error percentage P1 and type II error percentage P2 of the two can be obtained by formulas (3) and (4), and the acceptance probability L(p) can be obtained by formula (2). The fault loss function η can be obtained by formulas (5), (6), (7), (8), and (9) to judge the performance of the model.

[0083] When the sampling scheme is (85, 30, 3) and t=100, the data obtained by the two models are shown in the following table:

[0084]

[0085] When the sampling scheme is (85, 30, 3) and t=1000, the data obtained by the two models are shown in the following table:

[0086]

[0087] The table shows that regardless of the fault loss magnitude t = 100 or 1000, the GA-BP neural network's fault loss function η is always smaller than that of the CNN, so the GA-BP neural network is selected. This also directly indicates that even if the CNN's prediction accuracy is higher than that of the GA-BP, the selection of the neural network model will change after introducing the potential fault losses in actual production.

[0088] The above experiments show that although the prediction accuracy of the CNN neural network is higher than that of the BP neural network and the GA-BP neural network, by analyzing the prediction error types and introducing the concept of fault loss, the selection of the neural network will change according to the change of the fault loss magnitude t and is determined by the fault loss function.

[0089] From the perspective of safety and efficiency, this invention calculates the failure loss magnitude t based on the losses caused by safety and efficiency failures during mechanical production. Using a confusion matrix, the neural network model's erroneous predictions are divided into safety and efficiency failures, denoted as Class I and Class II misjudgments, respectively. Using the principle of sampling inspection, the frequency of Class I and Class II misjudgments in the test set samples is used to infer the frequency of Class I and Class II misjudgments in the actual monitoring process for each neural network model. This frequency is then multiplied by the failure loss magnitude t to obtain the failure loss function η. Using this failure loss function, the optimal neural network for fault diagnosis is determined. In general, the choice of neural network model for fault diagnosis varies depending on the failure loss magnitude t.

[0090] Under the condition that the classification prediction accuracy of the neural network model reaches more than 90%, the present invention distinguishes the classification prediction error types. At the same time, taking into account the need to avoid losses caused by safety accidents in the mechanical production process as much as possible, the basis for selecting the neural network model is no longer based on the classification prediction accuracy, but the losses caused by failures in the mechanical production process are used as the basis for model selection. This is more in line with actual selection considerations and can also reduce losses caused by failures.

[0091] Finally, it should be noted that the above examples of the present invention are merely illustrative of the present invention and are not intended to limit the embodiments of the present invention. Although the applicant has described the present invention in detail with reference to preferred embodiments, those skilled in the art will appreciate that other variations and modifications can be made based on the above description. It is not possible to enumerate all embodiments here. Any obvious variations or modifications derived from the technical solution of the present invention remain within the scope of protection of the present invention.

Claims

1. A decision-making method for selecting a neural network model for mechanical fault diagnosis based on sampling inspection and failure loss, characterized by: Construct several neural network models for mechanical fault diagnosis; use the failure loss caused by each neural network model's incorrect prediction of mechanical fault classification as the basis for selecting the neural network model, and select the neural network model with the smallest failure loss among all the neural network models; When constructing a neural network model, the data set is divided into a training set and a test set. The relationship between input and output features is constructed through the training set to obtain the neural network model. Using the sampling inspection principle, the test results of the test set are used to infer the number of classification prediction errors of the constructed neural network model for the actual monitored objects in mechanical fault diagnosis. Different classification prediction errors and numbers correspond to different losses. The sum of the losses of all classification prediction errors is the fault loss. The classification prediction error is divided into three error situations: the monitored object is in a fault state, but the classification prediction is in a healthy state; the monitored object is in a healthy state, but the classification prediction is in a fault state; the monitored object is in a fault state, but the classification prediction is in a fault state but the fault type is misjudged; the former situation belongs to a safety fault, recorded as Class I misjudgment, and the corresponding loss is called safety fault loss; the latter two situations belong to efficiency faults, recorded as Class II misjudgment, and the corresponding losses are called efficiency fault losses; Using the sampling inspection principle, based on the occurrence frequency of Class I and Class II misjudgments in the test set samples, the frequency of Class I and Class II misjudgments of the corresponding neural network model in the actual monitoring process is inferred, and then the safety failure loss and efficiency failure loss of each neural network model are inferred. The sum of the safety failure loss and efficiency failure loss is the failure loss.

2. The neural network model selection and decision-making method for mechanical fault diagnosis based on sampling inspection and failure loss according to claim 1 is characterized in that: The percentage of type I misjudgment and the percentage of type II misjudgment are calculated as follows: P1=D1 / N×100% (3) P2=D2 / N×100% (4) Where P1 is the percentage of Class I misjudgments; P2 is the percentage of Class II misjudgments; D1 is the number of Class I misjudgments in the validation set; D2 is the number of Class II misjudgments in the validation set; N is the total number of validation sets; Assume that the acceptance probability L(P1) represents the probability that no type I misjudgment will occur during the overall operation of the monitored object, and the acceptance probability L(P2) represents the probability that no type II misjudgment will occur during the overall operation of the monitored object; let α1 and α2 be the safety risk coefficient and efficiency risk coefficient, respectively. The calculation formulas are: α1=1-L(P1) (5) α2=1-L(P2) (6) The safety risk coefficient α1 represents the probability of type I misjudgment in the overall operation of the monitored object, and α2 represents the probability of type II misjudgment in the overall operation of the monitored object; The fault loss magnitude t is defined as the ratio of safety fault loss to efficiency fault loss. Let mα1 and mα2 represent the number of type I and type II misjudgments of the monitored object when the monitored object is run m times, respectively. At this time, the safety fault loss and efficiency fault loss are expressed as tmα1 and mα2, respectively. Let η1 and η2 be the safety fault loss function and efficiency fault loss function, respectively, representing the safety fault loss and efficiency fault loss caused by each sample. η1 is defined as safety fault loss / number of batch products; η2 is efficiency fault loss / number of batch products. The simplified calculation formula is as follows: η1=t*α1 (7) η2=α2 (8) Let the failure loss function η be the sum of the safety failure loss function η1 and the efficiency failure loss function η2, that is: η=η1+η2(9) Calculate the fault loss function η of all neural network models, and select the neural network model with the smallest fault loss function η among all neural network models.

3. The neural network model selection and decision-making method for mechanical fault diagnosis based on sampling inspection and failure loss according to claim 2 is characterized by: In the mechanical production process, the fault loss magnitude t is increased so that the efficiency fault loss function η2 is ignored, that is, η = η1 = t*α1. At this time, the selection of the neural network model is determined by the safety risk coefficient α1. Among all the neural network models, the neural network model with the smallest safety risk coefficient α1 is selected.

4. The neural network model selection and decision-making method for mechanical fault diagnosis based on sampling inspection and failure loss according to claim 1 is characterized in that: The classification and prediction accuracy of all neural network models reached over 90%.

Images