Device fault diagnosis method and device, computer device, and storage medium

Abstract

The embodiment of the application belongs to the field of equipment operation and maintenance, and relates to a device fault diagnosis method, comprising the following steps: extracting source domain features and target domain features in vibration data samples according to a preset deep residual network; determining attention weights corresponding to fault categories based on the source domain features; determining fault forward features according to the attention weights and the source domain features; and performing discrimination processing on the target domain features and the fault forward features through a domain discriminator to determine the device fault category. Thus, the multi-level features hidden in the vibration data can be better extracted. Due to the existence of the attention weights, the source domain features are decomposed into fault forward features related to the fault diagnosis task, blind alignment of the source domain features and the target domain features is avoided, the device fault diagnosis efficiency is higher, the accuracy of the device fault diagnosis is improved, the time cost and the effort of the equipment operation and maintenance personnel are greatly reduced, human intervention is not needed, and the degree of factory automation is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment operation and maintenance, and in particular to an equipment fault diagnosis method and device, a computer device and a storage medium. BACKGROUND

[0002] The operation and maintenance management of equipment is a crucial link for the stable operation of an enterprise factory. With the continuous development of the industrial revolution, the traditional equipment management mode gradually changes to automation. In the past, the operation and maintenance of equipment mostly relied on manual experience. Since the equipment fault diagnosis information cannot be obtained in time, the planned maintenance and after-maintenance methods are generally adopted. The consequences of this processing method will cause excessive maintenance, resource waste or safety accidents.

[0003] A large amount of data such as temperature, speed, vibration, etc. is generated in the daily operation of factory equipment, and these data are collected by sensors and sent to the front end of the equipment for analysis and use by workers. The temperature, speed, vibration and other data reflect the equipment state from various angles. Since the collected data have no labels, the operation and maintenance personnel need to monitor the equipment state according to these unlabeled data to detect in real time whether the equipment has a fault. When the equipment has a fault, the operation and maintenance personnel need to determine the fault type in time to facilitate the subsequent development of an effective maintenance plan, so as to realize real-time control of the equipment state and reduce the loss caused by the untimely response to equipment sudden failure. Since the manual diagnosis method has human subjective bias, high time cost, poor diagnosis accuracy and efficiency, it is not conducive to the scene of multiple high-speed running equipment.

[0004] Equipment fault diagnosis has always been the core part of equipment management. At present, the commonly used methods for fault diagnosis include a pure physical model according to the equipment mechanism, a frequency spectrum analysis on the collected signal data, a judgment on the equipment fault diagnosis according to the frequency or amplitude size combined with a threshold, and an automatic diagnosis on the equipment fault diagnosis type by using the popular convolutional neural network and adversarial transfer model. The traditional mechanism model needs to understand the equipment mechanism and greatly depends on the experience of experts, which is difficult for ordinary personnel. The frequency spectrum analysis method can roughly judge the general common fault type, but the effect will be greatly reduced for complex signal data. For the popular adversarial transfer neural network model, the model realizes automatic diagnosis and has good application effect for unlabeled data fault diagnosis, but this process focuses on how to reduce the distribution difference by aligning the features of source data and new data. However, for the alignment method of this type of method, there is no information to explicitly guide the features aligned to be helpful to the fault diagnosis task, which increases the processing time of the fault diagnosis task, which greatly reduces the prediction efficiency of the fault diagnosis. SUMMARY

[0005] The embodiment of the application aims to provide a device fault diagnosis method and device, computer equipment and storage medium, to solve the problems of low accuracy and low efficiency of device fault diagnosis.

[0006] To solve the above technical problems, the embodiment of the application provides a device fault diagnosis method, which adopts the technical scheme as follows:

[0007] According to the preset deep residual network, the source domain feature and the target domain feature in the vibration data sample are extracted;

[0008] Based on the source domain feature, the attention weight corresponding to the fault category is determined;

[0009] According to the attention weight and the source domain feature, the fault forward feature is determined;

[0010] The target domain feature and the fault forward feature are subjected to domain discriminator processing to determine the device fault category.

[0011] Further, the source domain feature includes a fault category label, and determining the attention weight corresponding to the fault category includes:

[0012] The fault category vector of the source domain feature and the prediction category label carried by the fault category vector are obtained;

[0013] When the prediction category label is the same as the fault category label, the fault category vector and the source domain feature are subjected to gradient calculation to obtain the attention weight.

[0014] Further, obtaining the fault category vector of the source domain feature and the prediction category label carried by the fault category vector includes:

[0015] The source domain feature is subjected to classification processing by the class discriminator to obtain the fault category vector;

[0016] The fault category probability of the fault category vector is calculated by a normalized exponential function;

[0017] According to the preset cross-entropy loss function and the fault category probability, the prediction category label carried by the fault category vector is determined.

[0018] Further, the target domain feature and the fault forward feature are subjected to discrimination processing to determine the device fault category, including:

[0019] The target domain feature and the fault forward feature are spliced to obtain spliced features;

[0020] The spliced features are subjected to gradient inversion processing to obtain fault features;

[0021] The device fault type is determined by discriminating the fault features.

[0022] Further, the extracting the source domain feature and the target domain feature in the vibration data sample according to the preset deep residual network comprises:

[0023] collecting the vibration data sample;

[0024] classifying the vibration data sample into the source domain sample carrying the fault category label and the target domain sample not carrying the fault category label;

[0025] extracting the source domain feature of the source domain sample and the target domain feature of the target domain sample respectively through the deep residual network.

[0026] Further, the extracting the source domain feature of the source domain sample and the target domain feature of the target domain sample respectively through the deep residual network comprises:

[0027] calculating the source domain sample and the target domain sample through the residual unit of the preset number of layers respectively to obtain a source domain output result and a target domain output result;

[0028] obtaining a residual weight corresponding to the residual unit;

[0029] adding the source domain output result and the residual weight to obtain the source domain feature;

[0030] adding the target domain output result and the residual weight to obtain the target domain feature.

[0031] To solve the above technical problems, the embodiment of the present application further provides a device fault diagnosis apparatus, comprising:

[0032] an extracting module configured to extract a source domain feature and a target domain feature in a vibration data sample according to a preset deep residual network;

[0033] an attention module configured to determine an attention weight corresponding to a fault category based on the source domain feature;

[0034] a forward feature module configured to determine a fault forward feature according to the attention weight and the source domain feature;

[0035] a discrimination module configured to perform discrimination processing on the target domain feature and the fault forward feature through a domain discriminator to determine a device fault category.

[0036] Further, the source domain feature comprises a fault category label, and the attention module comprises:

[0037] a predicted category unit configured to obtain a fault category vector of the source domain feature and a predicted category label carried by the fault category vector;

[0038] The gradient calculation unit is configured to perform gradient calculation on the fault category vector and the source domain feature to obtain an attention weight when the predicted category label is the same as the fault category label.

[0039] Further, the prediction category unit comprises:

[0040] The classification sub-unit is configured to perform classification processing on the source domain feature by the class discriminator to obtain a fault category vector.

[0041] The calculation sub-unit is configured to calculate a fault category probability of the fault category vector by a normalized exponential function.

[0042] The prediction sub-unit is configured to determine a predicted category label carried by the fault category vector according to a preset cross-entropy loss function and the fault category probability.

[0043] Further, the identification module comprises:

[0044] The splicing unit is configured to splice the target domain feature and the fault forward feature to obtain a spliced feature.

[0045] The gradient inversion unit is configured to perform gradient inversion processing on the spliced feature to obtain a fault feature.

[0046] The identification unit is configured to determine the device fault type by performing identification processing on the fault feature.

[0047] Further, the extraction module comprises:

[0048] The acquisition unit is configured to acquire vibration data samples.

[0049] The vibration data samples are classified into source domain samples carrying a fault category label and target domain samples not carrying a fault category label.

[0050] The extraction unit is configured to extract a source domain feature of the source domain sample and a target domain feature of the target domain sample by a deep residual network.

[0051] Further, the extraction unit comprises:

[0052] The residual calculation sub-unit is configured to calculate the source domain sample and the target domain sample by a residual unit of a preset number of layers to obtain a source domain output result and a target domain output result.

[0053] The first acquisition sub-unit is configured to acquire residual weights corresponding to the residual unit.

[0054] The first addition sub-unit is configured to add the source domain output result and the residual weights to obtain the source domain feature.

[0055] The second addition sub-unit is configured to add the target domain output result and the residual weights to obtain the target domain feature.

[0056] To solve the above technical problems, the embodiment of the present application also provides a computer device, comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps of the service management method described above when executing the computer program.

[0057] To solve the above technical problems, the embodiment of the present application also provides a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the steps of the service management method described above.

[0058] Compared with the prior art, the embodiment of the present application has the following beneficial effects:

[0059] By extracting the source domain features and the target domain features in the vibration data sample according to the preset deep residual network, determining the attention weight corresponding to the fault category, determining the fault positive feature according to the attention weight and the source domain feature, and performing discrimination processing on the target domain feature and the fault positive feature to determine the equipment fault category, the extraction of the source domain feature and the target domain feature is completed through the deep residual network, so that the multi-level features hidden in the vibration data can be better extracted. Due to the existence of the attention weight, the source domain feature is decomposed into the fault positive feature related to the fault diagnosis task, avoiding blind alignment of the source domain feature and the target domain feature, but starting from the final task of the model, i.e. the fault diagnosis classification, unnecessary irrelevant feature calculation is reduced, the influence of unnecessary irrelevant features on the fault diagnosis calculation accuracy is reduced, the fault diagnosis method is more intuitive than the previous fault diagnosis method, the equipment fault diagnosis efficiency is higher, reasonable, and the accuracy of the equipment fault diagnosis is improved. In addition, since the automatic diagnosis mode of the model network is used to replace artificial diagnosis, the time cost and effort of the equipment maintenance personnel are greatly reduced, and human intervention is not needed, thereby improving the degree of factory automation. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the schemes in the present application, the drawings needed in the description of the embodiments of the present application will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creating laborious work.

[0061] Figure 1 is an exemplary system architecture diagram to which the present application can be applied;

[0062] Figure 2 Flowchart of one embodiment of the equipment fault diagnosis method according to the present application;

[0063] Figure 3 is a schematic diagram of one residual unit in the embodiments of the present application;

[0064] Figure 4 is a flow diagram of one embodiment of the present application for extracting source domain features and target domain features;

[0065] Figure 5 is a flow diagram of one embodiment of the device fault diagnosis method in the embodiments of the present application;

[0066] Figure 6 is a structural schematic diagram of one embodiment of the device fault diagnosis apparatus according to the present application;

[0067] Figure 7 is a structural schematic diagram of one embodiment of the computer device according to the present application. DETAILED DESCRIPTION

[0068] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application; the description and the drawings are to be regarded as illustrative in nature and are not intended to limit the application; the terminology used in the description of the application herein including the abstract is not intended to be limiting of the application and is only used for the purpose of providing precise clear and concise description. The terms "comprising", "having", "including" and "containing" and any variations thereof used herein are intended to cover a non-exclusive inclusion. The terms "first", "second", and the like used in the description and in the claims of this application are used for distinguishing between similar objects talked about and do not necessarily have a particular order.

[0069] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive of one another. It is expressly understood that the embodiments described herein are merely examples from a whole class of comparable embodiments which are claimed by the application.

[0070] In order to make the technical personnel in the art better understand the scheme of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings.

[0071] As shown in Figure 1 , the system architecture 100 can include terminal devices 101, 102, 103, a network 104 and a server 105. The network 104 is used to provide a communication link medium between the terminal devices 101, 102, 103 and the server 105. The network 104 can include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0072] The user can use the terminal devices 101, 102, and 103 to interact with the server 105 through the network 104 to receive or send messages, etc. Various communication client applications can be installed on the terminal devices 101, 102, and 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0073] The terminal devices 101, 102, and 103 can be various electronic devices with display screens and supporting web browsing, including but not limited to smartphones, tablet computers, e-book readers, MP3 (Moving Picture Experts Group Audio Layer III) players, MP4 (Moving Picture Experts Group Audio Layer IV) players, laptop computers, desktop computers, etc.

[0074] The server 105 can be a server providing various services, such as a background server providing support for pages displayed on the terminal devices 101, 102, and 103.

[0075] It should be noted that the device fault diagnosis method provided in the embodiments of the present application is generally executed by a server / terminal device, and accordingly, the device fault diagnosis apparatus is generally arranged in a server / terminal device.

[0076] It should be understood that Figure 1 The number of terminal devices, networks, and servers in

[0077] With reference to Figure 2 , a flowchart of one embodiment of the device fault diagnosis method according to the present application is shown. The device fault diagnosis method includes the following steps:

[0078] S201, source domain features and target domain features in the vibration data sample are extracted according to a preset deep residual network.

[0079] In the present embodiment, the electronic device (for example, the terminal device 101) on which the device fault diagnosis method runs can be a terminal device, a server, or any other electronic device. Figure 1The server / terminal device shown) can receive the request for collecting the vibration data sample through a wired connection or a wireless connection. It should be noted that the wireless connection can include but is not limited to 3G / 4G / 5G connection, WiFi connection, Bluetooth connection, WiMA device fault diagnosis connection, Zigbee connection, UWB (ultra wideband) connection, and other now known or future developed wireless connection.

[0080] The deep residual network is used to extract different levels of features in the vibration data sample, that is, to extract source domain features and target domain features that affect the device fault diagnosis task, instead of the gradient dispersion problem caused by the previous multi-layer convolution layer for extracting features, and meanwhile, the influence of irrelevant features on subsequent fault diagnosis recognition is reduced. The vibration data sample is a sample data reflecting the state of the device, such as temperature, speed, vibration, etc.

[0081] In some embodiments, according to the preset deep residual network, the specific implementation of extracting the source domain features and the target domain features in the vibration data sample includes:

[0082] Collecting a vibration data sample;

[0083] Classifying the vibration data sample into a source domain sample carrying a fault category label and a target domain sample not carrying a fault category label;

[0084] Extracting the source domain features of the source domain sample and the target domain features of the target domain sample through the deep residual network, respectively.

[0085] The present application takes the fault diagnosis of a reduction gearbox as an example. Bearings are an extremely common assembly part in factory equipment. The fault diagnosis of bearings has always been a hot spot in the mechanical industry. Due to different types of equipment, the data distribution form of the bearing vibration signal will be different. In the embodiment of the present application, the health state diagnosis of the entire reduction gearbox uses the vibration signals of each important measuring point for fault diagnosis. Important parts of the reduction gearbox, such as input shaft, output shaft, and sensor-equipped measuring points of each transmission shaft, are used to collect vibration data, i.e., vibration signals, to monitor the running state of the entire reduction gearbox in real time according to the vibration signals collected by each measuring point.

[0086] Further, the collected vibration data is classified into source domain samples and target domain samples, and the source domain features of the source domain samples are pre-labeled with fault category labels. After being extracted by the deep residual network, the source domain features and the target domain features of the source domain samples and the target domain samples, respectively, share the same features and fault categories, but the feature distributions are different. The present application uses the information-rich source domain features to improve the performance of the target domain model for fault diagnosis.

[0087] Firstly, part of the high-quality fault data with labels is obtained according to the historical fault data accumulation and expert experience, that is, several common fault types of the device, and the research field of these fault data is called a source domain. The device unlabeled fault diagnosis problem researched in the application is a target domain. A fault diagnosis classifier is established for the source domain by using the labeled source domain samples. However, since the target domain samples are not labeled, the established source domain classifier is applied to the target domain by means of migration. However, in the whole migration process, in order to ensure the accuracy of the final fault diagnosis, the distribution difference between the source domain and the target domain should be reduced as much as possible. Unlike the previous method of aligning the source domain features and the target domain features, the application proposes a feature alignment oriented to the final fault diagnosis task. Such feature alignment can better serve the fault diagnosis classification task, and the method is more convincing than the previous method, and can further improve the accuracy of fault diagnosis on the basis of the original method.

[0088] Among them, the source domain feature represents a different field from the target domain feature (also called a test sample), and has rich supervised label information; the target domain feature represents the field of the test sample, which is unlabeled or has only a small amount of labels. The source domain feature and the target domain feature often belong to the same type of task, but have different distributions.

[0089] In the whole process of building the fault diagnosis model, for the diagnosis of the bearing health state, the bearing vibration signal is generally analyzed. The collected bearing vibration signal is generally complex and contains time domain information and frequency domain information. In addition, the influence of external noise or other factors makes the information contained in it more complex, and the use of ordinary signal decomposition to extract features cannot obviously achieve more accurate and detailed information. The application uses a deep residual network (ResNet) in the field of image processing to extract deeper features from the source domain samples and the target domain samples. This network is more stable than the previous convolutional network and is more conducive to extracting deeper features. Since it is the fault diagnosis of the whole reduction gearbox, and the reduction gearbox has multiple important part measurement points, the vibration data collected from each part measurement point is spliced and used as the input of the deep residual network, and the feature extraction work is completed through the deep residual network.

[0090] Among them, the source domain feature and the target domain feature of the source domain sample and the target domain sample are extracted by the deep residual network, including:

[0091] The source domain sample and the target domain sample are calculated by the residual unit with a preset number of layers to obtain the source domain output result and the target domain output result;

[0092] The residual weight corresponding to the residual unit is obtained;

[0093] The source domain output result and the residual weight are added to obtain a source domain feature;

[0094] The target domain output result and the residual weight are added to obtain a target domain feature.

[0095] The preset number of layers is determined by actual network training, and is not limited here. The entire network is composed of multiple residual units, and a single residual unit is as shown in Figure 3 The vibration data x is input into the residual unit, and the vibration data is added to the calculated result to output the extraction result of the residual unit by calculating through the weight layer of the set number of layers. The residual calculation process is as shown in formula (1) and formula (2).

[0096] y l =h(x l )+F(x l ,W l ) Formula (1)

[0097] x l+1 =f(y l ) Formula (2)

[0098] Wherein, x l and x l+1 respectively represent the input (i.e. the source domain sample and the target domain sample) and the output (i.e. the source domain output result and the target domain output result) of the lth residual unit, y l represents the mapping in the lth residual unit, W l is a weight matrix, F is a residual function, i.e. a learned residual, and h(x l ) = x l represents an identity mapping, and f is a ReLU activation function.

[0099] Based on the above formula, the learning feature x L from the shallow layer l to the deep layer L can be obtained according to formula (3) as follows:

[0100]

[0101] Wherein, is the residual weight obtained by calculating and accumulating through the residual function and the weight matrix, and the obtained residual weight is the residual weight corresponding to the residual unit. In the embodiment of the present application, L = 3, i.e. the feature is extracted through 3 residual units, and the residual weight corresponding to the residual unit is i.e. the residual weight corresponding to two residual units, and finally the source domain feature is represented as fs and the target domain feature is represented as ft.

[0102] In the embodiment of the present application, as shown in Figure 4The flowchart of an embodiment of source domain features and target domain features shown is illustrated. Three residual units are used for both source domain feature extraction and target domain feature extraction. The entire deep residual network consists of three residual units connected in series, and the size of the convolution kernel is 3×3. Finally, features in the two domains are obtained respectively.

[0103] By using a deep residual network to extract source domain features from source domain samples and target domain features from target domain samples, instead of using a simple convolutional neural network, the hidden multi-level features in the vibration signal can be extracted more effectively, laying the foundation for subsequent fault category diagnosis and thus improving the accuracy of fault diagnosis.

[0104] S202, determine the attention weights corresponding to the fault categories based on source domain features.

[0105] When a fault category is related to certain source domain features, that is, when the fault category is affected by these source domain features, a certain weight value is assigned to the source domain features so that the corresponding fault category can be inferred based on the weight value of the source domain features.

[0106] In some implementations, the source domain features include fault category labels, and S202, which determines the attention weights corresponding to the fault categories, includes the following specific implementations:

[0107] Obtain the fault category vector of the source domain features and the predicted category label carried by the fault category vector;

[0108] When the predicted category label is the same as the fault category label, the gradient of the fault category vector and the source domain features is calculated to obtain the attention weight.

[0109] like Figure 5 The flowchart of one embodiment of the equipment fault diagnosis method shown in the figure shows that the source domain features obtained are processed by a class discriminator to finally obtain a fault category vector. Then, combined with the known fault category scores provided by the prior knowledge of the class discriminator, attention weights are calculated. The class discriminator is used to determine the fault category based on the source domain features. The fault category vector is a predicted category score vector.

[0110] In the embodiments of the present application, the obtained source domain feature fs, m is the number of fault categories in the source domain sample, and the predicted category score vector y, each component of y is the score of predicting as one of the m fault categories. For example, if the number of fault categories is 3, that is, there are three types of faults, represented by (0, 1, 2), the obtained predicted category score vector y = (y1, y2, y3) at this time, y1 represents the score of predicting the category label as the "0" fault type, y2 represents the score of predicting the category label as the "1" fault type, and y3 represents the score of predicting the category label as the "2" fault type. Assuming that the predicted category label k is the correct category (ground-truth class), that is, the predicted category label carried by the fault category vector is the same as the fault category label carried by the corresponding source domain feature, the score y k and the source domain feature fs to obtain the gradient w cls That is, the obtained is the attention weight, and the attention weight calculation formula is as formula (4):

[0111]

[0112] Through the above formula (4), the attention weight w cls is obtained, and then the domain discriminator is constructed through the attention weight. The domain discriminator is the most core part in the adversarial transfer network. In the past, the adversarial transfer network fully mixes the source domain feature and the target domain feature, and respectively constructs the domain discriminator and the class discriminator, which easily considers irrelevant features and increases the loss error of model recognition, and reduces the accuracy of fault diagnosis. The calculation formula of w cls in the present application can be known that the predicted category score vector y in the class discriminator is used in the calculation process, the attention weight is calculated under the condition that y is known, that is, the calculation process is completed based on some prior knowledge of the class discriminator, which is equivalent to having the "guidance" (Classification meta-knowledge) of the class discriminator, so that the domain discriminator can better distinguish the source domain sample and the target domain sample based on the final fault diagnosis classification task.

[0113] In some embodiments, the fault category vector of the source domain feature and the predicted category label carried by the fault category vector are specifically implemented as follows:

[0114] The source domain feature is classified by the class discriminator to obtain the fault category vector;

[0115] The fault category probability of the fault category vector is calculated by the normalized exponential function;

[0116] According to the preset cross-entropy loss function and the fault category probability, the predicted category label carried by the fault category vector is determined.

[0117] In the embodiment of the present application, the source domain features are classified by the class discriminator to obtain a fault class vector, that is, a class score vector. Since the class discriminator is used to determine which fault type the input source domain sample belongs to, it is essentially a multi-classification problem. The present embodiment uses the commonly used softmax (normalized exponential function) in deep learning to complete the classification task. Through the softmax function, the probability p i i.e. the fault class probability, and the cross-entropy loss function L C is used to determine the predicted class label, where L C is defined as formula (5):

[0118]

[0119] where m is the number of fault classes, y i is the predicted class label, that is, if the predicted class label is i, then y i = 1, otherwise equal to 0, p j is the softmax output fault class probability, p i i.e. the probability that the fault class is i. Since the cross-entropy loss function represents the distribution of the true label, it can be used to measure the similarity between the predicted class label and the original fault class label carried, thereby improving the accuracy of the predicted class label.

[0120] S203, determining the fault positive feature according to the attention weight and the source domain feature.

[0121] The source domain features obtained in S201 are weighted. The source domain sample is fault data with class label. In the process of fault class diagnosis, the correct class (ground-truth class) is recorded as fps, which is called fault positive feature. Correspondingly, the feature fns unrelated to the discrimination task can be obtained at the same time, which is called fault negative feature. This idea is based on Grad-CAM, which is used to locate the most important classification features in the convolutional neural network model. The gradient of the final prediction score corresponding to the ground-truth class transmits the task discrimination information, and the relevant features are identified, so that the correct class is correctly identified. Therefore, the gradient of the prediction score corresponding to the ground-truth class is used as the attention weight to obtain the fault positive feature related to the discrimination task.

[0122] After the source domain feature fs is acted on by the attention weight w cls , the fault positive feature related to the discrimination task fps is obtained. At this time, the fault positive feature fps is calculated as shown in formula (6):

[0123] fps = fs ⊙ wcls Formula (6)

[0124] The previous feature extraction process is to fully mix the source domain features and the target domain features, and align all the features, while performing domain discriminator training and class determiner training, that is, two tasks in parallel. Since the overall features obtained from the source domain samples or the target domain samples usually contain information related to classification discrimination information and task-irrelevant information, intuitively, aligning the task-irrelevant features cannot effectively reduce the gap between the two domains, and has no obvious benefit to the classification task. Aligning the target characteristics with the source characteristics irrelevant to the task will impair the recognition ability of the target characteristics. Therefore, in the present application, the method of aligning all the features is no longer used, but the overall features of each source domain sample are decomposed into fault positive features related to the fault discrimination task and task-irrelevant features, so as to align the task-oriented fault positive features with the target features, so as to improve the recognition accuracy of fault diagnosis.

[0125] S204, the target domain features and the fault positive features are subjected to discrimination processing to determine the equipment fault category.

[0126] The discrimination processing refers to further judging whether the classified fault category is accurate through the domain discriminator, wherein the discriminator outputs a confidence degree of [0, 1] given an input classification information, the closer to 1, the higher the confidence, and the closer to 0, the lower the confidence, and then judging whether the classification is true.

[0127] In some optional implementation manners of the embodiment, S204, the target domain features and the fault positive features are subjected to discrimination processing to determine the equipment fault category, including:

[0128] The target domain features and the fault positive features are spliced to obtain spliced features;

[0129] The spliced features are subjected to gradient inversion processing to obtain fault features;

[0130] The fault features are subjected to discrimination processing to determine the equipment fault type.

[0131] In the embodiment of the present application, as Figure 5As shown, the spliced features obtained after splicing the target domain features and the fault positive features are taken as the input of the domain discriminator. In the embodiment of the present application, the spliced features will not be directly input to the domain discriminator, but there is a gradient reversal layer between the domain discriminator and the spliced features, so that the spliced features obtain the fault features after gradient reversal processing. The training target of the domain discriminator is to try to divide the input information into the correct domain category, that is, whether the source domain sample is the known fault category or the target domain sample is the unknown fault category. In the process of network back propagation parameter update, due to the existence of the gradient reversal layer, the gradient direction will gradually change to the opposite direction. For the final fault category judgment, the back propagation makes the spliced features gradually iterate towards the direction of confusing the discrimination result, so that the domain discriminator cannot correctly judge which domain the information comes from, thus forming an antagonistic relationship. At the same time, the existence of the class discriminator enables the fault to be classified. Therefore, based on the antagonistic relationship and the class discriminator, the fault category can be well distinguished when distinguishing the source of the vibration data sample, a better classifier can be trained, and the fault category of the target sample can be judged.

[0132] Further, the domain discriminator loss function L D to determine the equipment fault type:

[0133]

[0134] wherein G represents the feature extractor, L C represents the class discriminator loss function, X s and X t represent the set of source domain samples and target domain samples, and are the samples in the corresponding set, N s and N t are the number of source domain samples and target domain samples respectively.

[0135] Based on this situation, the application proposes a fault diagnosis task-oriented unsupervised adaptive (ToAlign-UDA) equipment fault diagnosis method. The method performs fault diagnosis on the real-time acquired equipment state data without labels. First, according to the historical fault data accumulation and expert experience, part of the high-quality fault data with labels is obtained. The data with labels are taken as a source domain (source), and the equipment unlabeled fault diagnosis problem researched by the application is taken as a target domain (target). Then, the labeled source domain data is used to establish a fault diagnosis classifier for the source domain. However, since the target domain data itself has no label, the established source domain classifier is applied to the target domain by means of the transfer idea. However, in the whole transfer process, in order to ensure the accuracy of the final fault diagnosis, the distribution difference between the source domain and the target domain should be reduced as much as possible. Unlike the previous method of aligning the features of the source domain and the target domain, the application proposes a feature alignment oriented to the final fault diagnosis task. Such feature alignment can better serve the fault diagnosis classification task and further improve the accuracy of fault diagnosis on the basis of the original. Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When the program is executed, the processes of the above-mentioned embodiments can be included. The storage medium can be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0136] It should be understood that, although each step in the flowchart of the accompanying drawings is displayed in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and they can be executed in other sequences. Moreover, at least part of the steps in the flowchart of the accompanying drawings can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence is not necessarily sequential, but can be executed alternately or alternately with at least part of other steps or other steps, sub-steps or stages.

[0137] Further referring to Figure 6 , as an implementation of the method shown in Figure 2 , the application provides an embodiment of an equipment fault diagnosis device. The device embodiment corresponds to the method embodiment shown in Figure 2 . The device can be applied to various electronic equipment.

[0138] AsFigure 6 As shown, the device fault diagnosis apparatus described in this embodiment includes an extraction module 61, an attention module 62, a forward feature module 63, and a discrimination module 64. Among them:

[0139] The extraction module 61 is configured to extract source domain features and target domain features in the vibration data sample according to a preset deep residual network;

[0140] The attention module 62 is configured to determine an attention weight corresponding to a fault category based on the source domain features;

[0141] The forward feature module 63 is configured to determine a fault forward feature according to the attention weight and the source domain features;

[0142] The discrimination module 64 is configured to perform discrimination processing on the target domain features and the fault forward feature through a domain discriminator to determine a device fault category.

[0143] In some optional implementation manners of this embodiment, the source domain features include a fault category label, and the attention module 62 includes:

[0144] A prediction category unit is configured to obtain a fault category vector of the source domain features and a prediction category label carried by the fault category vector;

[0145] A gradient calculation unit is configured to perform gradient calculation on the fault category vector and the source domain features to obtain the attention weight when the prediction category label is the same as the fault category label.

[0146] In some optional implementation manners of this embodiment, the prediction category unit includes:

[0147] A classification subunit is configured to perform classification processing on the source domain features through a category discriminator to obtain the fault category vector;

[0148] A calculation subunit is configured to calculate a fault category probability of the fault category vector through a normalized exponential function;

[0149] A prediction subunit is configured to determine the prediction category label carried by the fault category vector according to a preset cross-entropy loss function and the fault category probability.

[0150] In some optional implementation manners of this embodiment, the discrimination module 64 includes:

[0151] A splicing unit is configured to splice the target domain features and the fault forward feature to obtain spliced features;

[0152] A gradient inversion unit is configured to perform gradient inversion processing on the spliced features to obtain fault features;

[0153] A discrimination unit is configured to perform discrimination processing on the fault features to determine a device fault type.

[0154] In some optional implementations of the embodiment, the extraction module 61 comprises:

[0155] a collection unit, configured to collect vibration data samples;

[0156] classify the vibration data samples into source domain samples carrying fault category labels and target domain samples without carrying fault category labels;

[0157] an extraction unit, configured to extract source domain features of the source domain samples and target domain features of the target domain samples respectively through a deep residual network.

[0158] Further, the extraction of the source domain features of the source domain samples and the target domain features of the target domain samples through the deep residual network comprises:

[0159] calculating the source domain samples and the target domain samples through residual units of a preset number of layers respectively to obtain source domain output results and target domain output results;

[0160] obtaining residual weights corresponding to the residual units;

[0161] adding the source domain output results and the residual weights to obtain the source domain features;

[0162] adding the target domain output results and the residual weights to obtain the target domain features.

[0163] The application extracts source domain features and target domain features in the vibration data sample through the extraction module, determines the attention weight corresponding to the fault category by using the attention module, determines the fault forward feature through the forward feature module, and then determines the equipment fault category through the discrimination module. Unlike the past, in the aspect of feature alignment, the source domain features and the target domain features are not blindly aligned in total, but from the perspective of the final task of the model, i.e. the fault classification diagnosis, the source domain features are decomposed into forward features related to the classification task and irrelevant features. Based on the idea of Grad-CAM, the "key part" directly affecting the correct classification of the subsequent category discriminator is calculated, i.e. the gradient is calculated as the key attention weight, then the source domain features and the attention weight interact to obtain the fault forward feature, at this time the fault forward feature and the target domain feature are aligned to serve as the input of the domain discriminator to complete the judgment of whether the sample comes from the source domain or the target domain. In this process, the fault classification task and the domain discriminator task are not parallel, but the domain discriminator is actually completed under the guidance of the classification discriminator due to the existence of the attention weight, and ultimately serves the fault diagnosis task. Intuitively, it is more in line with the purpose of the model, i.e. it can further improve the accuracy of fault diagnosis. Secondly, in the process of feature extractor, a simple convolutional neural network is not used, but a deep residual network is selected to extract the source domain features and the target domain features, so that the hidden multi-level features in the signal data can be better extracted, laying a foundation for the subsequent model. Compared with the past model method, it is more intuitive, efficient and reasonable, and can further improve the accuracy of fault diagnosis. In addition, due to the automation of the model, the pressure on the equipment maintenance personnel is reduced, and human intervention is not needed, improving the degree of factory automation.

[0164] To solve the above technical problems, the embodiment of the application further provides a computer device. For details, please refer to Figure 7 , Figure 7 The basic structure block diagram of the computer device of the embodiment is shown in the figure.

[0165] The computer device 7 comprises a memory 71, a processor 72, and a network interface 73 which are communicatively connected by a system bus. It should be noted that only the computer device 7 with components 71-73 is shown in the figure, but it should be understood that not all of the shown components are required to be implemented, and more or fewer components can be alternatively implemented. Among them, those skilled in the art can understand that the computer device herein is a device capable of automatically performing numerical calculation and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0166] The computer device can be a desktop computer, a notebook computer, a palm computer, a cloud server, and the like. The computer device can interact with the user through a keyboard, a mouse, a remote controller, a touchpad, a voice control device, and the like.

[0167] The memory 71 comprises at least one type of readable storage medium, including a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., an SD or D device fault diagnosis memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, and the like. In some embodiments, the memory 71 can be an internal storage unit of the computer device 7, such as a hard disk or a memory of the computer device 7. In other embodiments, the memory 71 can also be an external storage device of the computer device 7, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, and the like. Of course, the memory 71 can also include both the internal storage unit and the external storage device of the computer device 7. In the present embodiment, the memory 71 is generally used to store an operating system and various application software installed in the computer device 7, such as program codes of a device fault diagnosis method, and the like. In addition, the memory 71 can also be used to temporarily store various data that have been output or will be output.

[0168] The processor 72 may, in some embodiments, be a Central Processing Unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor 72 is generally used to control the overall operation of the computer device 7. In the present embodiment, the processor 72 is configured to run program code or process data stored in the memory 71, such as program code of the device fault diagnosis method.

[0169] The network interface 73 may include a wireless network interface or a wired network interface, and is generally used to establish a communication connection between the computer device 7 and other electronic devices.

[0170] The present application also provides another embodiment, i.e., a computer readable storage medium storing a device fault diagnosis program, which can be executed by at least one processor to enable the at least one processor to perform the steps of the device fault diagnosis method as described above.

[0171] From the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be realized by means of software and necessary general hardware platforms, and of course, can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as a ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to perform the methods described in the various embodiments of the present application.

[0172] Obviously, the above-described embodiments are only some embodiments of the present application, rather than all the embodiments, and the drawings show the preferred embodiments of the present application, but do not limit the patent scope of the present application. The present application can be realized in many different forms, and conversely, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing specific embodiments, or make equivalent replacements to some technical features. Any equivalent structure made by referring to the contents of the present application specification and drawings, directly or indirectly applied to other related technical fields, is also within the scope of the patent protection of the present application.

Claims

1. A method for diagnosing equipment faults, characterized in that, include: Based on the preset deep residual network, the source domain features and target domain features are extracted from the vibration data samples; The attention weights corresponding to the fault categories are determined based on the source domain features. Based on the attention weights and the source domain features, positive fault features are determined; The target domain features and the positive fault features are processed for identification to determine the equipment fault category; The source domain features include fault category labels, and the determination of attention weights corresponding to fault categories based on the source domain features includes: Obtain the fault category vector of the source domain features and the predicted category label carried by the fault category vector; When the predicted category label is the same as the fault category label, the gradient of the fault category vector and the source domain feature is calculated to obtain the attention weight; Where the predicted category label carried by the fault category vector is the same as the fault category label carried by its corresponding source domain feature, the score corresponding to the fault category vector is... Calculate the gradient of fs with source domain features That is, what we get is the attention weight, and the formula for calculating the attention weight is as shown in formula (4): ; Specifically, determining the positive fault features based on the attention weights and the source domain features includes: The source domain features are weighted. The source domain samples are fault data with category labels. During fault category diagnosis, those correctly classified are denoted as FPS, which are called positive fault features. After attention weight After the process, positive fault features related to the discrimination task are obtained. At this time, the positive characteristics of the fault The calculation is shown in formula (6): ; The discrimination process refers to further judging whether the classified fault categories are accurate by using a domain discriminator. The discriminator is given an input classification information and outputs a confidence level of [0,1]. The closer to 1, the higher the confidence level, and the closer to 0, the lower the confidence level, thereby judging whether the classification is true. The concatenated features obtained by concatenating the target domain features and the positive fault features are used as input to the domain discriminator. The concatenated features are not directly input to the domain discriminator. Instead, there is a gradient inversion layer between the domain discriminator and the concatenated features. This gradient inversion process results in fault features. The training objective of the domain discriminator is to classify the input information into the correct domain category, i.e., whether it is the known fault category of the source domain sample or the unknown fault category of the target domain sample. During the backpropagation of the network to update the parameters, due to the presence of the gradient inversion layer, the gradient direction will gradually change in the opposite direction. For the final fault category judgment, backpropagation causes the concatenated features to iterate in the direction of confusing the discrimination result, making it impossible for the domain discriminator to correctly determine which domain the information comes from. This creates an adversarial relationship. At the same time, the presence of the class discriminator enables the classification of the fault.

2. The equipment fault diagnosis method according to claim 1, characterized in that, The process of obtaining the fault category vector of the source domain features and the predicted category label carried by the fault category vector includes: The source domain features are classified using a class discriminator to obtain a fault category vector; The fault category probability of the fault category vector is calculated using a normalized exponential function; Based on the preset cross-entropy loss function and the fault category probability, the predicted category label carried by the fault category vector is determined.

3. The equipment fault diagnosis method according to claim 1, characterized in that, The step of processing the target domain features and the positive fault features to determine the equipment fault category includes: The target domain features and the positive fault features are concatenated to obtain concatenated features; The spliced ​​features are then subjected to gradient inversion processing to obtain the fault features. The fault characteristics are identified to determine the type of equipment fault.

4. The equipment fault diagnosis method according to any one of claims 1-3, characterized in that, The step of extracting source domain features and target domain features from vibration data samples based on a preset deep residual network includes: Collect vibration data samples; The vibration data samples are classified into source domain samples carrying fault category labels and target domain samples without fault category labels. The source domain features of the source domain samples and the target domain features of the target domain samples are extracted using the deep residual network.

5. The equipment fault diagnosis method according to claim 4, characterized in that, The step of extracting the source domain features of the source domain samples and the target domain features of the target domain samples through the deep residual network includes: The source domain samples and the target domain samples are respectively calculated through residual units with a preset number of layers to obtain the source domain output result and the target domain output result; Obtain the residual weights corresponding to the residual units; The source domain output result and the residual weight are added together to obtain the source domain features; The target domain features are obtained by adding the target domain output and the residual weights.

6. A device for diagnosing equipment faults, characterized in that, include: The extraction module is used to extract source domain features and target domain features from vibration data samples based on a preset deep residual network. The attention module is used to determine the attention weights corresponding to the fault categories based on the source domain features; A positive feature module is used to determine fault positive features based on the attention weights and the source domain features; The identification module is used to process the target domain features and the positive fault features to determine the equipment fault category. The source domain features include fault category labels, and the attention module includes: A prediction category unit is used to obtain the fault category vector of the source domain features and the prediction category label carried by the fault category vector; The gradient calculation unit is used to calculate the gradient of the fault category vector and the source domain features when the predicted category label is the same as the fault category label, so as to obtain the attention weight; Where the predicted category label carried by the fault category vector is the same as the fault category label carried by its corresponding source domain feature, the score corresponding to the fault category vector is... Calculate the gradient of fs with source domain features That is, what we get is the attention weight, and the formula for calculating the attention weight is as shown in formula (4): ; Specifically, the positive feature module is used to weight the source domain features. The source domain samples are fault data with category labels. During the fault category diagnosis process, those correctly classified are denoted as fps, which are called fault positive features. After attention weight After the process, positive fault features related to the discrimination task are obtained. At this time, the positive characteristics of the fault The calculation is shown in formula (6): ; The discrimination process refers to further judging whether the classified fault categories are accurate by using a domain discriminator. The discriminator is given an input classification information and outputs a confidence level of [0,1]. The closer to 1, the higher the confidence level, and the closer to 0, the lower the confidence level, thereby judging whether the classification is true. The concatenated features obtained by concatenating the target domain features and the positive fault features are used as input to the domain discriminator. The concatenated features are not directly input to the domain discriminator. Instead, there is a gradient inversion layer between the domain discriminator and the concatenated features. This gradient inversion process results in fault features. The training objective of the domain discriminator is to classify the input information into the correct domain category, i.e., whether it is the known fault category of the source domain sample or the unknown fault category of the target domain sample. During the backpropagation of the network to update the parameters, due to the presence of the gradient inversion layer, the gradient direction will gradually change in the opposite direction. For the final fault category judgment, backpropagation causes the concatenated features to iterate in the direction of confusing the discrimination result, making it impossible for the domain discriminator to correctly determine which domain the information comes from. This creates an adversarial relationship. At the same time, the presence of the class discriminator enables the classification of the fault.

7. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the device fault diagnosis method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the device fault diagnosis method as described in any one of claims 1 to 5.

Images