A Two-Way Mining Method for Difficult Samples in Dynamic Equipment Fault Diagnosis Based on Joint Knowledge Distillation

Abstract

This application relates to the field of fault diagnosis technology, and in particular to a bidirectional mining method for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation. The method includes: acquiring difficult samples from the server and client based on a pre-built joint learning framework for a large-scale equipment fault diagnosis model; determining the difficult sample index based on the difficult samples from the server and client, and selecting the corresponding feature map based on the difficult sample index to obtain a difficult sample selection feature map; quantizing and compressing the difficult sample selection feature map, and transmitting the quantized and compressed difficult sample selection feature map between the server and client. This solves the problems of how to accurately select knowledge and perform difficult sample mining, achieving accurate knowledge localization, saving communication overhead, and improving model performance.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis technology, and in particular to a bidirectional mining method for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation. Background Technology

[0002] Due to harsh working environments and prolonged operation, major accidents caused by the failure of critical equipment in industrial production are increasingly drawing widespread public attention. This is particularly true in the petrochemical and energy sectors, where safety accidents due to equipment failure occur frequently. Rotating machinery, such as steam turbines, fans, and pumps, are critical equipment in industrial production. Equipment failures can lead to unexpected unit shutdowns, resulting in high operating and maintenance costs, and potentially causing significant economic losses and personnel accidents. The purpose of rotating machinery fault diagnosis is to detect, identify, and locate faults before they deteriorate, which is crucial for ensuring the safe and reliable operation of rotating machinery.

[0003] In recent years, with the popularization of online monitoring equipment, the industrial sector has accumulated a large amount of online monitoring data. However, the use of this data has only been limited to threshold alarms. How to make fuller use of monitoring data for real-time intelligent fault diagnosis has become an urgent and highly challenging problem. Currently, based on the collected vibration data, many rotating machinery fault diagnosis methods based on deep learning, knowledge graphs, and automated machine learning algorithms have been proposed, such as convolutional neural networks and long short-term memory networks.

[0004] However, most AI-based fault diagnosis methods for rotating machinery require training an independent model for each machine and suffer from poor generalization ability, failing to handle the varying data distributions caused by differences in structure and operating conditions between devices. To address this issue, transfer learning and domain adaptation are introduced into the field of fault diagnosis. However, most existing transfer learning methods can only handle variations in operating conditions such as speed and load. When applied to fault diagnosis of multiple machines, they still require a large amount of fault data and a complete range of fault types to support the training process, which is unsatisfactory in practical industrial applications.

[0005] Therefore, to address the issues of limited fault samples for single devices, poor local model performance, and the inability to diagnose multiple faults across multiple devices, joint learning is considered. Joint learning can learn fault diagnosis models from dispersed data while ensuring that private data remains locally accessible. However, in the scenario of rotating machinery fault diagnosis, traditional joint learning cannot meet the needs of heterogeneous equipment, data, and models. Training the same model for all participants does not yield ideal results. Furthermore, due to limited computing power among participants, training a single large model for each participant is impractical. Frequent uploading and downloading of model parameters incurs high communication overhead, affecting the efficiency of ordinary joint learning and potentially leading to the spread of "bad knowledge." Therefore, how to coordinate computing resources, better utilize server-side computing power and knowledge, and reduce communication overhead are urgent problems to be solved in rotating machinery fault diagnosis. Summary of the Invention

[0006] This application provides a bidirectional mining method for difficult samples in equipment fault diagnosis based on joint knowledge distillation, which solves the problems of how to accurately select knowledge and mine difficult samples, achieves accurate knowledge positioning, saves communication overhead, and improves model performance.

[0007] The first aspect of this application provides a bidirectional mining method for difficult samples in equipment fault diagnosis based on joint knowledge distillation, comprising the following steps: obtaining difficult samples from the server and client based on a pre-built joint learning framework for a large equipment fault diagnosis model; determining a difficult sample index based on the difficult samples from the server and client, and selecting a feature map corresponding to the difficult sample based on the difficult sample index to obtain a difficult sample selection feature map; quantizing and compressing the difficult sample selection feature map, and transmitting the quantized and compressed difficult sample selection feature map between the server and the client.

[0008] Optionally, before obtaining the hard samples of the server and client based on the pre-built joint learning framework for large-scale equipment fault diagnosis model, the method further includes: obtaining the public dataset of the server and training the server-side large-scale model using the public dataset; obtaining the private dataset of the client and training the client-side small-scale model using the private dataset; and jointly training the server-side large-scale model and the client-side small-scale model based on a preset bidirectional joint knowledge distillation strategy until a preset training termination condition is met, thereby obtaining the pre-built joint learning framework for large-scale equipment fault diagnosis model.

[0009] Optionally, the public dataset includes multiple types of faults, including shaft faults, bearing faults, and gearbox faults.

[0010] Optionally, the public dataset may also include currently publicly available fault simulation experimental data.

[0011] Optionally, determining the hard sample index based on the hard samples of the server and the client includes: obtaining the predicted value and the true label value of the server and the client on a public validation set, wherein the public validation set is constructed by multiple clients; traversing all samples in the public validation set, obtaining a hard sample index set based on the indexes of all samples whose difference between the predicted value and the true label value is greater than a preset threshold, and using the hard sample index set as the hard sample index.

[0012] Optionally, the above-mentioned bidirectional mining method for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation further includes: determining whether the predicted value is transmitted between the server and the client; if the predicted value is transmitted between the server and the client, then the predicted value is transmitted directly.

[0013] Optionally, the step of selecting the feature map corresponding to the difficult sample based on the difficult sample index to obtain the difficult sample selection feature map includes: retaining the feature map corresponding to the difficult sample index in the difficult sample index, and removing the feature map corresponding to the difficult sample index that is not in the difficult sample index, so as to obtain the difficult sample selection feature map.

[0014] A second aspect of this application provides a bidirectional mining device for difficult samples in equipment fault diagnosis based on joint knowledge distillation, comprising: a first acquisition module, used to acquire difficult samples from a server and a client based on a pre-built joint learning framework for a large-scale equipment fault diagnosis model; a second acquisition module, used to determine a difficult sample index based on the difficult samples from the server and the client, and select a feature map corresponding to the difficult sample based on the difficult sample index to obtain a difficult sample selection feature map; and a compression module, used to quantize and compress the difficult sample selection feature map, and transmit the quantized and compressed difficult sample selection feature map between the server and the client.

[0015] Optionally, before obtaining the hard samples of the server and client based on the pre-built joint learning framework for large-scale equipment fault diagnosis model, the first acquisition module is further configured to: acquire the public dataset of the server and train the server-side large-scale model using the public dataset; acquire the private dataset of the client and train the client-side small-scale model using the private dataset; and jointly train the server-side large-scale model and the client-side small-scale model based on a preset bidirectional joint knowledge distillation strategy until a preset training termination condition is met, thereby obtaining the pre-built joint learning framework for large-scale equipment fault diagnosis model.

[0016] Optionally, the public dataset includes multiple types of faults, including shaft faults, bearing faults, and gearbox faults.

[0017] Optionally, the public dataset may also include currently publicly available fault simulation experimental data.

[0018] Optionally, the second acquisition module is further configured to: acquire the predicted values ​​and true label values ​​of the server and the client on the public validation set, wherein the public validation set is constructed by multiple clients; traverse all samples of the public validation set, obtain a set of hard sample indices based on the indices of all samples whose difference between the predicted value and the true label value is greater than a preset threshold, and use the set of hard sample indices as the hard sample index.

[0019] Optionally, the above-mentioned bidirectional mining device for difficult samples of dynamic equipment fault diagnosis based on joint knowledge distillation further includes: a judgment module, used to determine whether the predicted value is transmitted between the server and the client; if the predicted value is transmitted between the server and the client, then the predicted value is transmitted directly.

[0020] Optionally, the second acquisition module is further configured to: retain the feature maps corresponding to the difficult samples indexed in the difficult sample index and remove the feature maps corresponding to the difficult samples not indexed in the difficult sample index, so as to obtain the difficult sample selection feature map.

[0021] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the program to implement the bidirectional mining method for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation as described in the above embodiments.

[0022] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the bidirectional mining method for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation as described in the above embodiments.

[0023] Therefore, this application, based on a pre-built joint learning framework for a large-scale equipment fault diagnosis model, acquires hard samples from both the server and client sides. It determines hard sample indices based on these indices, selects feature maps corresponding to the hard samples, and obtains hard sample selection feature maps. These selected feature maps are then quantized and compressed, and transmitted between the server and client sides. This solves the problems of accurately selecting knowledge and mining hard samples, achieving precise knowledge localization, saving communication overhead, and improving model performance.

[0024] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0025] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0026] Figure 1 This is a flowchart of a bidirectional mining method for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation, according to an embodiment of this application.

[0027] Figure 2 This is a schematic diagram of a joint knowledge distillation framework for fault diagnosis of moving equipment according to an embodiment of this application;

[0028] Figure 3 This is a flowchart of a bidirectional mining method for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation, according to an embodiment of this application;

[0029] Figure 4 This is an example diagram of a bidirectional mining device for difficult samples of dynamic equipment fault diagnosis based on joint knowledge distillation, according to an embodiment of this application.

[0030] Figure 5 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0031] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0032] The following describes, with reference to the accompanying drawings, a bidirectional mining method for difficult samples in moving equipment fault diagnosis based on joint knowledge distillation, according to embodiments of this application. Addressing the problem mentioned in the background section regarding how to accurately select knowledge for difficult sample mining, this application provides a bidirectional mining method for difficult samples in moving equipment fault diagnosis based on joint knowledge distillation. In this method, based on a pre-built joint learning framework for a large-scale equipment fault diagnosis model, difficult samples from the server and client are obtained. Difficult sample indices are determined based on these indices, and feature maps corresponding to the difficult samples are selected to obtain difficult sample selection feature maps. These selected feature maps are then quantized and compressed, and transmitted between the server and client. This solves the problem of how to accurately select knowledge for difficult sample mining, achieving accurate knowledge localization, saving communication overhead, and improving model performance.

[0033] Specifically, Figure 1 This is a flowchart illustrating a bidirectional mining method for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation, provided in an embodiment of this application.

[0034] like Figure 1 As shown, the bidirectional mining method for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation includes the following steps:

[0035] In step S101, based on the pre-built joint learning framework of the large-scale equipment fault diagnosis model, difficult samples from the server and the client are obtained.

[0036] Optionally, in some embodiments, before obtaining hard samples from the server and client based on the pre-built joint learning framework for large-scale equipment fault diagnosis model, the method further includes: obtaining the public dataset of the server and training the server-side large-scale model using the public dataset; obtaining the private dataset of the client and training the client-side small-scale model using the private dataset; and jointly training the server-side large-scale model and the client-side small-scale model based on a preset bidirectional joint knowledge distillation strategy until a preset training termination condition is met, thereby obtaining the pre-built joint learning framework for large-scale equipment fault diagnosis model.

[0037] Optionally, in some embodiments, the public dataset includes multiple types of faults, including shaft faults, bearing faults, and gearbox faults.

[0038] In some embodiments, the public dataset also includes currently publicly available fault simulation experimental data.

[0039] Specifically, in this embodiment of the application, the factory is the client and a platform operator is the joint learning server to build a joint learning framework for a large-scale equipment fault diagnosis model.

[0040] The server-side architecture comprises a large model and a publicly available dataset consisting of multiple fault types, extensive online fault simulation data from universities and research institutions, and a publicly available validation set jointly constructed by all client-side factories. The validation set is a dataset compiled from data collected from multiple clients. The server-side model is pre-trained using this publicly available dataset. The fault types include shaft faults, bearing faults, and gearbox faults. Each client-side architecture has a smaller model, with a dataset consisting of a local private dataset and the publicly available validation set. Heterogeneous modeling is supported between clients, enabling the diagnosis of one to two types of equipment faults. Clients train their models on their local private datasets.

[0041] Furthermore, the server and client transmit predicted values ​​and feature maps, and train the model using a bidirectional joint knowledge distillation method to obtain a pre-built joint learning framework for the large-scale equipment fault diagnosis model, until the joint learning model converges.

[0042] Specifically, the pre-built large-scale equipment fault diagnosis model joint learning framework, such as... Figure 2 As shown, in terms of data, the factory in the heterogeneous FL (Federated Learning) framework collects data from local monitoring devices to form a local private dataset X. i (i = 1, 2, ..., n). The cloud-based management of heterogeneous FL frameworks utilizes a publicly available dataset X. public This includes a large amount of publicly available data in the relevant field. A public validation set X is constructed by collecting data from each factory. s This is used for training KD (Knowledge Distillation). To obtain the data distribution for each participant, the proportion of local data contributed by each factory can be 1 / 5 or less.

[0043] Secondly, regarding the model, the cloud maintains a model using X. public Pre-trained large-scale DL (Deep Learning) models. The DL models among the factories are heterogeneous in terms of model parameters and structure. Therefore, each factory can choose a local DL model of different sizes based on local data and computing resource constraints.

[0044] The steps for jointly training the server-side large model and the client-side small model are as follows:

[0045] In step 1), according to X public Build large DL models using publicly available data and tasks, and use X public Pre-training large models in the cloud. Model pre-training can improve the generality of large cloud models in relevant fields and facilitate the cold start of the entire joint KD framework;

[0046] In step 2), all factories use their local private dataset X. i Train the local deep learning model locally;

[0047] In step 3), based on the public verification set X s Bidirectional KD training is performed between the factory and cloud servers, including selective knowledge fusion from the factory to the cloud (see details). Figure 3 (as shown) and KD, as well as KD from the cloud server to each factory;

[0048] In step 4), repeat steps 2) through 3) until X. sand X i The training reached convergence.

[0049] In step S102, the hard sample index is determined based on the hard samples of the server and the client, and the feature map corresponding to the hard sample is selected based on the hard sample index to obtain the hard sample selection feature map.

[0050] Specifically, in terms of communication, since transmitting feature maps requires a lot of communication overhead, feature maps are selected for uploading based on hard samples, and the selected feature maps are further quantized and compressed to reduce communication costs by 7 to 10 times while ensuring performance.

[0051] Optionally, in some embodiments, determining the hard sample index based on the hard samples of the server and the client includes: obtaining the predicted values ​​and true label values ​​of the server and the client on a public validation set, wherein the public validation set is constructed by multiple clients; traversing all samples in the public validation set, obtaining a hard sample index set based on the indexes of all samples whose difference between the predicted value and the true label value is greater than a preset threshold, and using the hard sample index set as the hard sample index.

[0052] It should be understood that the process of obtaining the hard sample index required by the client / server includes obtaining the predicted value and true label of the client / server on the public validation set, traversing all samples of the public validation set, and if the difference between the predicted value and the true label is greater than a preset threshold, then the index of all samples is added to the hard sample index set, the sample traversal is completed, and the resulting hard sample index set is used as the hard sample index required by the client / server.

[0053] The selective knowledge fusion-hard sample bidirectional mining method is described as follows: Figure 3 As shown, adding the feature map KD improves the accuracy and stability of both the client and server models, but it also increases communication overhead due to the transmission of large-scale feature maps. Furthermore, the size of the feature map is directly related to the level of feature map extraction and the number of samples in the dataset, introducing uncertainty into network transmission. Additionally, in the later stages of KD training, X... s Most of the samples in the database have been stored on the server (f) server ) and client-side model (f i Correct identification leads to the transfer of a large amount of useless knowledge. On the other hand, repeated training on correctly identified samples will lead to overfitting of the DL model.

[0054] Therefore, to reduce communication overhead and prevent overfitting, the pre-built large-scale device fault diagnosis model joint learning framework uses hard sample selection, which only uses the feature maps of hard samples. i This needs to be transmitted between the client and the server. The specific method for determining hard samples is as follows: when the predicted value (l) i) and the true label value (y i The absolute value of the difference between y and y exceeds a preset threshold, i.e., the judgment condition |y| is met. i -l i |>σ, are considered hard samples, and are used as the hard sample index set to obtain the corresponding hard sample index.

[0055] Furthermore, in some embodiments, the feature map corresponding to the difficult sample is selected based on the difficult sample index to obtain the difficult sample selection feature map, including: retaining the feature map corresponding to the difficult sample whose index is in the difficult sample index, and removing the feature map corresponding to the difficult sample whose index is not in the difficult sample index, so as to obtain the difficult sample selection feature map.

[0056] It is understandable that the selection of feature maps means: retaining the feature maps corresponding to samples whose indices are in the hard sample index, and discarding the feature maps corresponding to samples whose indices are not in the hard sample index, thus obtaining the hard sample selection feature map.

[0057] In step S103, the hard sample selection feature map is quantized and compressed, and the quantized and compressed hard sample selection feature map is transmitted between the server and the client.

[0058] Optionally, in some embodiments, the above-described bidirectional mining method for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation further includes: determining whether a predicted value is transmitted between the server and the client; if a predicted value is transmitted between the server and the client, then the predicted value is transmitted directly.

[0059] It is understandable that if the server transmits predicted values ​​to each client, the transmission is performed directly. If the server transmits feature maps to each client, the transmission is performed using the feature maps corresponding to the samples in the hard sample index.

[0060] During feature map transmission, a list of hard samples is sent to the other end (server or client) of a pre-built joint learning framework for a large-scale device fault diagnosis model. When the other party receives the list of hard samples, only the map of the hard sample list is generated after local KD training. need The feature maps of the samples are transmitted back, while the predicted values ​​for all samples are transmitted. Specifically, the cloud server receives the feature maps of hard samples from all clients, and since the clients share the same list of hard samples, the feature maps are aggregated on an average basis.

[0061] Compared with the prior art, the technical solution proposed in this application embodiment, in the process of joint bidirectional knowledge distillation for fault diagnosis of moving equipment, can greatly save communication overhead, avoid the spread of "bad knowledge", and improve the overall performance of the model by first selecting difficult samples on the server / client side, determining the index of difficult samples, and then compressing and uploading the corresponding feature maps.

[0062] The bidirectional mining method for difficult samples in equipment fault diagnosis based on joint knowledge distillation proposed in this application involves acquiring difficult samples from both the server and client sides using a pre-built joint learning framework for a large-scale equipment fault diagnosis model. Difficult sample indices are determined based on these indices, and feature maps corresponding to the difficult samples are selected to obtain the difficult sample selection feature map. This selected feature map is then quantized and compressed, and transmitted between the server and client sides. This solves the problems of accurately selecting knowledge and mining difficult samples, achieving precise knowledge localization, saving communication overhead, and improving model performance.

[0063] Secondly, referring to the accompanying drawings, this application proposes a bidirectional mining device for difficult samples of dynamic equipment fault diagnosis based on joint knowledge distillation, according to an embodiment of the present application.

[0064] Figure 4 This is a block diagram of a bidirectional mining device for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation, according to an embodiment of this application.

[0065] like Figure 4 As shown, the bidirectional mining device 10 for difficult samples of dynamic equipment fault diagnosis based on joint knowledge distillation includes: a first acquisition module 100, a second acquisition module 200, and a compression module 300.

[0066] The first acquisition module 100 is used to acquire hard samples from the server and client based on a pre-built joint learning framework for a large-scale equipment fault diagnosis model; the second acquisition module 200 is used to determine the hard sample index based on the hard samples from the server and client, and select the feature map corresponding to the hard sample based on the hard sample index to obtain the hard sample selection feature map; the compression module 300 is used to quantize and compress the hard sample selection feature map, and transmit the quantized and compressed hard sample selection feature map between the server and client.

[0067] Optionally, in some embodiments, before obtaining hard samples from the server and client based on the pre-built joint learning framework for large-scale equipment fault diagnosis model, the first acquisition module 100 is further configured to: acquire the public dataset of the server and train the server-side large-scale model using the public dataset; acquire the private dataset of the client and train the client-side small-scale model using the private dataset; and jointly train the server-side large-scale model and the client-side small-scale model based on a preset bidirectional joint knowledge distillation strategy until the preset training termination condition is met, thereby obtaining the pre-built joint learning framework for large-scale equipment fault diagnosis model.

[0068] Optionally, in some embodiments, the public dataset includes multiple types of faults, including shaft faults, bearing faults, and gearbox faults.

[0069] Optionally, in some embodiments, the public dataset also includes currently publicly available fault simulation experimental data.

[0070] Optionally, in some embodiments, the second acquisition module 200 is further configured to: acquire the predicted values ​​and true label values ​​of the server and the client on the public validation set, wherein the public validation set is constructed by multiple clients; traverse all samples of the public validation set, obtain a set of hard sample indexes based on the indexes of all samples whose difference between the predicted value and the true label value is greater than a preset threshold, and use the set of hard sample indexes as the hard sample index.

[0071] Optionally, in some embodiments, the above-described bidirectional mining device 10 for difficult samples of dynamic equipment fault diagnosis based on joint knowledge distillation further includes: a judgment module, used to determine whether the server and the client transmit the predicted value; if the server and the client transmit the predicted value, then the predicted value is transmitted directly.

[0072] Optionally, the second acquisition module 200 is further configured to: retain the feature maps corresponding to the difficult samples whose indices are in the difficult sample index, and remove the feature maps corresponding to the difficult samples whose indices are not in the difficult sample index, so as to obtain the difficult sample selection feature map.

[0073] It should be noted that the foregoing explanation of the embodiment of the bidirectional mining method for difficult samples of moving equipment fault diagnosis based on joint knowledge distillation also applies to the bidirectional mining device for difficult samples of moving equipment fault diagnosis based on joint knowledge distillation in this embodiment, and will not be repeated here.

[0074] The bidirectional mining device for difficult samples in equipment fault diagnosis based on joint knowledge distillation proposed in this application obtains difficult samples from both the server and client based on a pre-built joint learning framework for a large-scale equipment fault diagnosis model. It determines the difficult sample index based on these samples and selects the corresponding feature map based on the index, resulting in a difficult sample selection feature map. This selected feature map is then quantized and compressed, and transmitted between the server and client. This solves the problems of accurately selecting knowledge and mining difficult samples, achieving precise knowledge localization, saving communication overhead, and improving model performance.

[0075] Figure 5 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include:

[0076] The memory 501, the processor 502, and the computer program stored on the memory 501 and capable of running on the processor 502.

[0077] When the processor 502 executes the program, it implements the bidirectional mining method for difficult samples of dynamic equipment fault diagnosis based on joint knowledge distillation provided in the above embodiments.

[0078] Furthermore, electronic devices also include:

[0079] Communication interface 503 is used for communication between memory 501 and processor 502.

[0080] The memory 501 is used to store computer programs that can run on the processor 502.

[0081] The memory 501 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0082] If the memory 501, processor 502, and communication interface 503 are implemented independently, then the communication interface 503, memory 501, and processor 502 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 5 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0083] Optionally, in a specific implementation, if the memory 501, processor 502, and communication interface 503 are integrated on a single chip, then the memory 501, processor 502, and communication interface 503 can communicate with each other through an internal interface.

[0084] Processor 502 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0085] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described bidirectional mining method for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation.

[0086] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0087] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0088] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0089] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0090] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0091] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0092] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0093] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A bidirectional mining method for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation, characterized in that, Includes the following steps: Based on a pre-built large-scale equipment fault diagnosis model joint learning framework, difficult samples from both the server and client sides are obtained. The hard sample index is determined based on the hard samples from the server and the client, and the feature map corresponding to the hard sample is selected based on the hard sample index to obtain the hard sample selection feature map. as well as The hard sample selection feature map is quantized and compressed, and the quantized and compressed hard sample selection feature map is transmitted between the server and the client; before obtaining the hard samples from the server and the client based on the pre-built equipment fault diagnosis large model joint learning framework, the method further includes: Obtain the public dataset of the server and use the public dataset to train a large model on the server. Obtain the client's private dataset and use the private dataset to train a small client model; Based on a preset bidirectional joint knowledge distillation strategy, the server-side large model and the client-side small model are jointly trained until the preset training termination condition is met, thus obtaining the pre-built joint learning framework of the device fault diagnosis large model. The step of determining the hard sample index based on the hard samples from the server and the client includes: Obtain the predicted values ​​and true label values ​​of the server and client on a public validation set, wherein the public validation set is constructed by multiple clients; wherein the method for determining the hard sample is as follows: when the absolute value of the difference between the predicted value and the true label value exceeds a preset threshold, it is considered a hard sample; Traverse all samples in the public validation set, obtain a set of hard sample indices based on the indices of all samples whose difference between the predicted value and the true label value is greater than a preset threshold, and use the set of hard sample indices as the hard sample index.

2. The method according to claim 1, characterized in that, The publicly available dataset includes multiple types of faults, including shaft faults, bearing faults, and gearbox faults.

3. The method according to claim 2, characterized in that, The publicly available dataset also includes currently available fault simulation experimental data.

4. The method according to claim 1, characterized in that, Also includes: Determine whether the predicted value is transmitted between the server and the client; If the server and the client are transmitting the predicted value, then the predicted value is transmitted directly.

5. The method according to claim 1, characterized in that, The step of selecting the feature map corresponding to the difficult sample based on the difficult sample index to obtain the difficult sample selection feature map includes: The feature maps corresponding to the difficult samples indexed in the difficult sample index are retained, and the feature maps corresponding to the difficult samples not indexed in the difficult sample index are removed to obtain the difficult sample selection feature map.

6. A bidirectional mining device for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation, characterized in that, The apparatus for implementing the bidirectional mining method for difficult samples in dynamic equipment fault diagnosis based on joint knowledge distillation as described in any one of claims 1-5 includes: The first acquisition module is used to acquire difficult samples from the server and client based on the pre-built joint learning framework of the large-scale equipment fault diagnosis model. The second acquisition module is used to determine the hard sample index based on the hard samples from the server and the client, and select the feature map corresponding to the hard sample based on the hard sample index to obtain the hard sample selection feature map; and The compression module is used to quantize and compress the hard sample selection feature map and transmit the quantized and compressed hard sample selection feature map between the server and the client.

7. An electronic device, characterized in that, include: The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the bidirectional mining method for difficult samples of dynamic equipment fault diagnosis based on joint knowledge distillation as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the bidirectional mining method for difficult samples of dynamic equipment fault diagnosis based on joint knowledge distillation as described in any one of claims 1-5.

Images