Parts life prediction method, device, equipment, storage medium and program product

Abstract

The present application provides a method, device, electronic device, computer-readable storage medium and computer program product for predicting the life of a part; the embodiment of the present application can be applied to vehicle-mounted scenarios, and the method includes: obtaining target sampling data of a part corresponding to a target time period in a target domain; when the target time period is completely in the degradation stage of the part, extracting multiple time domain features and multiple frequency domain features from the target sampling data; performing first attention processing on multiple time domain features and multiple frequency domain features to obtain time-frequency domain related features of the target domain; performing second attention processing on the time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain to obtain the life characteristics of the target domain; based on the life characteristics of the target domain, determining the remaining life of the part in the target domain starting from the target time period. Through the present application, the accuracy of the remaining life prediction of the part can be improved.

Description

Technical Field

[0001] The present application relates to artificial intelligence technology, and in particular to a method, device, electronic device, computer-readable storage medium and computer program product for predicting the life of a part. Background Art

[0002] Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0003] Machine parts are an important part of modern industrial equipment. Whether their operating status is normal directly affects the performance of the entire industrial equipment. For example, the operating status detection and remaining service life prediction of bearing parts are of great significance to the maintenance of industrial equipment and the safety of personnel in modern industry.

[0004] In the related art, the remaining life prediction method for parts is usually based on statistical reliability. Although the statistical reliability-based method is simple to model, it usually only considers life data and does not fully utilize the data of the working process, resulting in low accuracy of the remaining life prediction. Summary of the invention

[0005] The embodiments of the present application provide a method, device, electronic device, computer-readable storage medium and computer program product for predicting the life of a part, which can improve the accuracy of the remaining life prediction of a part.

[0006] The technical solution of the embodiment of the present application is implemented as follows:

[0007] The present application provides a method for predicting the life of a component, including:

[0008] Obtaining target sampling data of the part in a target domain corresponding to a target time period;

[0009] When the target time period is completely in the degradation stage of the part, extracting a plurality of time domain features and a plurality of frequency domain features from the target sampling data;

[0010] Performing a first attention process on the multiple time domain features and the multiple frequency domain features to obtain time-frequency domain related features of the target domain;

[0011] Performing a second attention process on the time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain to obtain the life span features of the target domain;

[0012] Based on the life characteristics of the target domain, a remaining life of the part in the target domain starting from the target time period is determined.

[0013] The present application embodiment provides a device for predicting the life of a part, comprising:

[0014] A sampling module, used to obtain target sampling data of the part in a target domain corresponding to a target time period;

[0015] An extraction module, used for extracting a plurality of time domain features and a plurality of frequency domain features from the target sampling data when the target time period is completely in the degradation stage of the part;

[0016] An attention module, configured to perform a first attention process on the multiple time domain features and the multiple frequency domain features to obtain time-frequency domain related features of the target domain;

[0017] The attention module is further used to perform a second attention process on the time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain to obtain the life span features of the target domain;

[0018] The life module is used to determine the remaining life of the part in the target domain starting from the target time period based on the life characteristics of the target domain.

[0019] In the above scheme, when the target time period is completely in the degradation stage of the part, before extracting multiple time domain features and multiple frequency domain features from the target sampling data, the sampling module is also used to: obtain multiple sampling data of the part from the start of work, wherein each of the sampling data includes multiple sampling points; perform multiple first sliding window processing based on the first sliding window size on the multiple sampling data sorted in ascending order based on the sampling time to obtain the sampling points in each first sliding window result; for each of the first sliding window results, determine the root mean square of the sampling points in the first sliding window result; perform multiple second sliding window processing based on the second sliding window size on the multiple root mean squares to obtain the root mean square in each second sliding window result; for each of the second sliding window results , based on the root mean square in the second sliding window result, determine the smoothed root mean square corresponding to the second sliding window result; perform multiple third sliding window processing based on the third sliding window size on multiple smoothed root mean squares to obtain the smoothed root mean square in each third sliding window result; for each third sliding window result, based on the smoothed root mean square in the third sliding window result, determine the smoothed mean value corresponding to the third sliding window result; based on the smoothed mean value corresponding to each third sliding window result, obtain the target third sliding window result that meets the degradation period condition among the multiple third sliding window results, and determine the earliest sampling time corresponding to the target third sliding window result as the starting time of the degradation stage; when the target time period is after the starting time, determine that the target time period is in the degradation stage.

[0020] In the above scheme, before performing multiple first sliding window processing based on the first sliding window size on the multiple sampling data to obtain the sampling points in each first sliding window result, the sampling module is also used to: take the inverse of the sampling frequency of the multiple sampling data to obtain the sampling interval of the multiple sampling points of the multiple sampling data; multiply the sampling interval by the number of multiple sampling points included in each of the sampling data to obtain the first sliding window size.

[0021] In the above scheme, the second sliding window size is larger than the first sliding window size, and the sampling module is further used to: perform multiple processing based on a first integer on the first sliding window size to obtain the second sliding window size; perform a first summation processing on the root mean square within the second sliding window result to obtain a first summation result; and determine the ratio of the first summation result to the second sliding window size as the smoothed root mean square corresponding to the second sliding window result.

[0022] In the above scheme, the sampling module is also used to: perform multiple processing based on a second integer on the second sliding window size to obtain the third sliding window size; perform a second summation processing on the smoothed root mean square within the third sliding window result to obtain a second summation result; and determine the ratio of the second summation result to the third sliding window size as the smoothed mean of the smoothed root mean square of the third sliding window result.

[0023] In the above scheme, the sampling module is also used to: perform the following processing for each of the third sliding window results: determine the absolute value between the smoothed mean and the stable mean of the third sliding window result; perform multiple processing of the stable standard deviation based on the third integer to obtain a mapping standard deviation; when the absolute value is greater than the mapping standard deviation, determine that the third sliding window result is the target third sliding result that meets the degradation period condition.

[0024] In the above scheme, the sampling module is also used to: based on the sampling time corresponding to each of the root mean squares, sort multiple corresponding smoothed root mean squares in ascending order, wherein each of the smoothed root mean squares corresponds to the root mean square of the earliest sampling time in the corresponding second sliding window result; use the multiple smoothed root mean squares ranked at the top in the ascending sorting result as the target smoothed root mean square; determine the mean of the multiple target smoothed root mean squares as the stable mean, and determine the standard deviation of the multiple target smoothed root mean squares as the stable standard deviation.

[0025] In the above scheme, the extraction module is also used to: perform feature processing on the time domain working signals of multiple sampling points included in the target sampling data to obtain the time-frequency domain features of the target sampling data; wherein the time-frequency domain features include multiple time domain features and multiple frequency domain features, and the feature processing is short-time Fourier transform processing or wavelet transform processing.

[0026] In the above scheme, the attention module is also used to: perform time domain self-attention processing on the multiple time domain features to obtain self-attention time domain features; perform frequency domain self-attention processing on the multiple frequency domain features to obtain self-attention frequency domain features; splice the self-attention time domain features with the self-attention frequency domain features to obtain fused features; perform fused attention processing on the target frequency domain features corresponding to the target moment in the target time period among the multiple frequency domain features and the fused features to obtain time-frequency domain related features of the target domain.

[0027] In the above scheme, the attention module is also used to: perform time domain embedding processing on the multiple time domain features to obtain multiple time domain embedding features that correspond one-to-one to the multiple time domain features; perform time domain position encoding processing on the multiple time domain embedding features to obtain multiple time domain position codes that correspond one-to-one to the multiple time domain features; perform time domain self-attention encoding processing on the multiple time domain position codes to obtain the multiple self-attention time domain features that correspond one-to-one to the multiple time domain features.

[0028] In the above scheme, the attention module is also used to: perform frequency domain embedding processing on the multiple frequency domain features to obtain multiple frequency domain embedded features that correspond one-to-one to the multiple frequency domain features; perform frequency domain position encoding processing on the multiple frequency domain embedded features to obtain multiple frequency domain position encodings that correspond one-to-one to the multiple frequency domain features; perform frequency domain self-attention encoding processing on the multiple frequency domain position encodings to obtain the multiple self-attention frequency domain features that correspond one-to-one to the multiple frequency domain features.

[0029] In the above scheme, the attention module is also used to: perform frequency domain embedding processing on the target frequency domain features to obtain target frequency domain embedded features corresponding to the target frequency domain features; perform frequency domain position encoding processing on the target frequency domain embedded features to obtain target frequency domain position encoding corresponding to the target frequency domain features; perform attention decoding processing on the target frequency domain position encoding and the fusion features to obtain time-frequency domain related features of the target domain.

[0030] In the above scheme, the attention module is also used to: perform dot product processing on the time-frequency domain related features of the target domain and multiple time-frequency domain related features of the source domain to obtain the dot product results corresponding to each time-frequency domain related feature of the source domain; perform maximum likelihood processing on the dot product results corresponding to each time-frequency domain related feature of the source domain to obtain the attention weight corresponding to each time-frequency domain related feature of the source domain; based on the attention weight, perform weighted sum processing on each time-frequency domain related feature of the source domain to obtain the source domain matching feature; fuse the source domain matching feature with the time-frequency domain related features of the target domain to obtain the life span feature of the target domain.

[0031] In the above scheme, the device also includes a training module, which is used to: obtain multiple source domain training samples corresponding to the source domain; forward propagate the multiple source domain training samples in the source domain life prediction model to obtain multiple first predicted remaining lifespans corresponding one-to-one to the multiple source domain training samples; determine a first loss based on the multiple first predicted remaining lifespans and multiple first pre-marked remaining lifespans, and update the parameters of the source domain life prediction model based on the first loss; use the parameters of the source domain life prediction model as initialization parameters of the target domain life prediction model; obtain multiple target domain training samples corresponding to the target domain; forward propagate the multiple target domain training samples in the target domain life prediction model to obtain multiple second predicted remaining lifespans corresponding one-to-one to the multiple target domain training samples; determine a second loss based on the multiple second predicted remaining lifespans and multiple second pre-marked remaining lifespans, and based on the second loss, update the parameters of the target domain life prediction model based on the initialization parameters.

[0032] An embodiment of the present application provides an electronic device, including:

[0033] A memory for storing executable instructions;

[0034] The processor is used to implement the life prediction method of parts provided in the embodiment of the present application when executing the executable instructions stored in the memory.

[0035] An embodiment of the present application provides a computer-readable storage medium storing executable instructions for implementing the life prediction method of a part provided in the embodiment of the present application when executed by a processor.

[0036] An embodiment of the present application provides a computer program product, including a computer program or instructions, which, when executed by a processor, implements the method for predicting the life of a part provided in the embodiment of the present application.

[0037] The embodiments of the present application have the following beneficial effects:

[0038] The target sampling data of the part corresponding to the target time period in the target domain is obtained. When the target time period is completely in the degradation stage of the part, multiple time domain features and multiple frequency domain features are extracted from the target sampling data. Life prediction is performed only based on the data of the degradation period. In this way, inaccurate remaining life prediction caused by the healthy period data of the part can be avoided. The multiple time domain features and multiple frequency domain features are subjected to first attention processing to obtain the time-frequency domain related features of the target domain. Therefore, the correlation between the timing information and the frequency domain information of the part is considered. The time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain are subjected to second attention processing to obtain the life features of the target domain, thereby migrating the time-frequency domain related features learned in other domains to the target domain, thereby obtaining the life features of the target domain. Based on the life features of the target domain, the remaining life of the part in the target domain starting from the target time period is determined, which can effectively improve the accuracy of the remaining life prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1A It is a schematic diagram of the structure of the life prediction system of a part provided in an embodiment of the present application;

[0040] Figure 1B It is a structural diagram of a life prediction system based on a blockchain network provided in an embodiment of the present application;

[0041] Figure 2 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application;

[0042] Figure 3A-3C It is a flow chart of a method for predicting the life of a part provided in an embodiment of the present application;

[0043] Figure 4 It is an overall flow chart of the life prediction method of parts provided in the embodiment of the present application;

[0044] Figure 5 It is a schematic diagram of the structure of the life prediction model provided in the embodiment of the present application;

[0045] Figure 6 It is a schematic diagram of the structure of the time domain coding network provided in the embodiment of the present application;

[0046] Figure 7 is a schematic diagram of the structure of a decoding network provided in an embodiment of the present application;

[0047] Figure 8 It is a schematic diagram of the structure of the meta-learning framework provided in the embodiment of the present application;

[0048] Fig. 9 It is a query matching processing diagram provided in an embodiment of the present application. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0050] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it can be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0051] In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described here can be implemented in an order other than that illustrated or described here.

[0052] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0053] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0054] 1) Degradation: The performance of the engineering material of the parts degrades after long-term use. For example, the degradation phenomenon of common steel materials includes the aggregation of carbides (pearlite spheroidization) after long-term use at medium and high temperatures. As a result, the strength (or lasting strength) of the steel decreases, or the plasticity and toughness decrease, leading to unsafe factors in the industrial equipment where the parts are located.

[0055] 2) Working conditions refer to the working state of a part under conditions directly related to its movement, for example, the working state of parts being tightly connected, the working state of parts being overloaded, etc.

[0056] 3) Attention Mechanism: The attention mechanism originates from the study of human vision. In cognitive science, due to the bottleneck of information processing, humans selectively focus on a part of all information while ignoring other visible information. The above mechanism is usually called the attention mechanism. The attention mechanism has two main aspects: deciding which part of the input needs to be paid attention to; allocating limited information processing resources to the important part.

[0057] As an important part of modern industrial equipment, bearings and other parts have a direct impact on the performance of the entire equipment. Therefore, the health status detection and remaining life prediction of parts are of great significance to the maintenance of industrial equipment and the safety of personnel. The remaining life prediction of parts can be roughly divided into three categories: model-based prediction, statistical reliability-based prediction, and data-driven prediction. The model-based prediction method is to establish a failure model by analyzing the failure mechanism of parts such as fatigue cracks, wear, corrosion, etc. Since it is necessary to build a mathematical or physical model based on the failure mechanism, the modeling process is complicated and it is usually difficult to implement in actual engineering applications. The method of prediction based on statistical reliability uses failure data to establish a bearing reliability assessment model. It is a prediction method based on empirical data. It usually only considers life data and does not make full use of process data, resulting in low prediction accuracy. In the related art, the data-driven prediction method uses all data for model training without considering the impact of health stage data. Since part failure or failure generally occurs in the degradation period, if the data in the health stage is used for life prediction, it will cause interference. In addition, the related art uses a fixed threshold to determine whether it has entered the degradation period. Due to the complex working environment of the parts, the use of a fixed threshold is not applicable to all parts. At the same time, the data-driven prediction methods in related technologies are mostly modeled based on the working data of a certain target working condition of the part, without making full use of the working data of other working conditions of the part, and without considering the similarities and correlations between the target working condition and other working conditions.

[0058] Embodiments of the present application provide a method, device, electronic device, computer-readable storage medium, and computer program product for predicting the life of a part, which can improve the accuracy of the remaining life prediction of a part. The following describes an exemplary application of the electronic device provided in the embodiments of the present application.

[0059] The electronic device provided in the embodiments of the present application can be implemented as various types of user terminals such as a laptop computer, a tablet computer, a desktop computer, a set-top box, a mobile device (e.g., a mobile phone, a portable music player, a personal digital assistant, a dedicated messaging device, a portable gaming device), etc., and can also be implemented as a server. Below, an exemplary application when the device is implemented as a server will be described.

[0060] See also Figure 1A , Figure 1A It is a structural diagram of the life prediction system of parts provided in an embodiment of the present application. To implement a maintenance APP that supports an industrial equipment, a terminal 400 is connected to a server 200 via a network 300. The network 300 may be a wide area network or a local area network, or a combination of the two.

[0061] In some embodiments, the life prediction method of parts provided in the embodiments of the present application can be implemented collaboratively by a terminal and a server, the terminal 400 receives the target sampling data of the target time period sent by the sensor 500 of the industrial equipment, the terminal 400 sends a remaining life prediction request starting from the target time period to the server 200, the life prediction request includes the target sampling data of the target time period, detects whether the target time period belongs to the degradation stage of the part, and when the target time period is completely in the degradation stage of the part, extracts multiple time domain features and multiple frequency domain features from the target sampling data; performs first attention processing on the multiple time domain features and the multiple frequency domain features to obtain time-frequency domain related features of the target domain; performs second attention processing on the time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain to obtain the life features of the target domain; based on the life features of the target domain, determines the remaining life of the part in the target domain starting from the target time period, and returns the remaining life to the terminal 400 for presentation.

[0062] In some embodiments, the life prediction method of parts provided in the embodiments of the present application can also be implemented separately by a terminal or a server. Taking the implementation by a terminal alone as an example, the terminal 400 receives the target sampling data of the target time period sent by the sensor of the industrial equipment, and the terminal 400 detects whether the target time period belongs to the degradation stage of the part. When the target time period is completely in the degradation stage of the part, multiple time domain features and multiple frequency domain features are extracted from the target sampling data; the multiple time domain features and the multiple frequency domain features are subjected to a first attention processing to obtain time-frequency domain related features of the target domain; the time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain are subjected to a second attention processing to obtain the life features of the target domain; based on the life features of the target domain, the remaining life of the part in the target domain starting from the target time period is determined, and the remaining life is presented on the terminal 400.

[0063] In some embodiments, the embodiments of the present application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, assisted driving, etc. For example, the parts life prediction method provided by the embodiments of the present application can predict the remaining life of automobile parts in real time, thereby prompting users to replace parts in time to ensure safe driving of the vehicle.

[0064] In some embodiments, the server 200 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal 400 may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, an intelligent voice interaction device, a smart home appliance, a vehicle-mounted terminal, an aircraft, etc., but is not limited thereto. The terminal and the server may be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiments of the present application.

[0065] In some embodiments, the terminal or server can implement the life prediction method of the parts provided in the embodiments of the present application by running a computer program. For example, the computer program can be a native program or software module in the operating system; it can be a native application (APP, Application), that is, a program that needs to be installed in the operating system to run, such as an industrial equipment maintenance APP; it can also be a small program, that is, a program that can be run only by downloading it to a browser environment; it can also be a small program that can be embedded in any APP. In short, the above-mentioned computer program can be any form of application, module or plug-in.

[0066] In some embodiments, see Figure 1B , Figure 1B is a schematic diagram of the structure of the life prediction system based on the blockchain network provided by the embodiment of the present application. The following describes an exemplary application of the embodiment of the present application based on the blockchain network. Figure 1B , including a blockchain network 600 (exemplarily showing nodes 610 - 1 and node 610 - 2 included in the blockchain network 600 ), a server 200 , and a terminal 400 , which are described below respectively.

[0067] The server 200 (mapped as node 610-2) and the terminal 400 (mapped as node 610-1) can both join the blockchain network 600 and become nodes therein. Figure 1B The figure exemplarily shows that the terminal 400 is mapped as a node 610-1 of the blockchain network 600, and each node (such as node 610-1, node 610-2) has a consensus function and a bookkeeping function (i.e., maintaining a state database library, such as a key-value database).

[0068] The state database of each node (eg, node 610 - 1 ) records the target sampling data, target time period, and corresponding remaining life collected by terminal 400 , so that terminal 400 can query the target sampling data, target time period, and corresponding remaining life recorded in the state database.

[0069] In some embodiments, in response to receiving the target sampling data and the target time period, multiple servers 200 (each server is mapped to a node in the blockchain network) determine the remaining life starting from the target time period. For the remaining life of the target time period, when the number of nodes that have passed the consensus exceeds the node number threshold, it is determined that the consensus is passed, and the server 200 (mapped to node 610-2) sends the remaining life of the target time period that has passed the consensus to the terminal 400 (mapped to node 610-1), and presents it on the human-computer interaction interface of the terminal 400, and stores the target sampling data, target time period and the corresponding remaining life on the chain. Since the remaining life is obtained after consensus by multiple servers, the reliability of the remaining life can be effectively improved. Because the blockchain network is not easy to tamper with, the target sampling data, target time period and the corresponding remaining life stored on the chain will not be maliciously tampered with.

[0070] See also Figure 2 , Figure 2 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application, Figure 2 The terminal 400 shown includes: at least one processor 410, a memory 450, at least one network interface 420 and a user interface 430. The various components in the terminal 400 are coupled together via a bus system 440. It is understood that the bus system 440 is used to achieve connection and communication between these components. In addition to the data bus, the bus system 440 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 440 is not shown in FIG. Figure 2 Various buses are labeled as bus system 440 .

[0071] Processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0072] The user interface 430 includes one or more output devices 431 that enable presentation of media content, including one or more speakers and / or one or more visual display screens. The user interface 430 also includes one or more input devices 432, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0073] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drives, optical disk drives, etc. The memory 450 may optionally include one or more storage devices that are physically remote from the processor 410.

[0074] The memory 450 includes a volatile memory or a non-volatile memory, and may also include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 450 described in the embodiments of the present application is intended to include any suitable type of memory.

[0075] In some embodiments, memory 450 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplarily described below.

[0076] Operating system 451, including system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0077] A network communication module 452, for reaching other computing devices via one or more (wired or wireless) network interfaces 420, exemplary network interfaces 420 include: Bluetooth, Wireless Compatibility Authentication (WiFi), and Universal Serial Bus (USB);

[0078] a presentation module 453 for enabling presentation of information via one or more output devices 431 (e.g., display screen, speaker, etc.) associated with the user interface 430 (e.g., a user interface for operating peripherals and displaying content and information);

[0079] The input processing module 454 is used to detect one or more user inputs or interactions from one of the one or more input devices 432 and translate the detected inputs or interactions.

[0080] In some embodiments, the device provided in the embodiments of the present application can be implemented in software. Figure 2 The life prediction device 455 of the parts stored in the memory 450 is shown, which can be software in the form of a program and a plug-in, etc., including the following software modules: a sampling module 4551, an extraction module 4552, an attention module 4553, a life module 4554 and a training module 4555. These modules are logical, so they can be arbitrarily combined or further split according to the functions implemented. The functions of each module will be explained below.

[0081] In other embodiments, the life prediction device of the parts provided in the embodiments of the present application can be implemented in hardware. As an example, the life prediction device of the parts provided in the embodiments of the present application can be a processor in the form of a hardware decoding processor, which is programmed to execute the method provided in the embodiments of the present application. For example, the processor in the form of a hardware decoding processor can adopt one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field programmable gate arrays (FPGAs) or other electronic components.

[0082] The life prediction method of parts provided in the embodiment of the present application will be explained in combination with the exemplary application and implementation of the terminal provided in the embodiment of the present application.

[0083] See also Figure 3A , Figure 3A is a flow chart of the method for predicting the life of a part provided in the embodiment of the present application, which will be combined with Figure 3A Steps 101 to 105 are shown for explanation.

[0084] In step 101, target sampling data of a part in a target domain corresponding to a target time period is obtained.

[0085] As an example, from the time when the part starts to work normally, the working data of the part is sampled to obtain multiple sampling points, and the set number of sampling points are combined into a certain sampling data. For example, every time 5 sampling points are sampled, a sampling data is generated, and each sampling data corresponds to a time period, that is, the time period from sampling to the first sampling point to sampling to the fifth sampling point. The target time period is the time period for calculating the remaining life from this starting point. For example, it can be the time period at the current moment, which is determined according to the needs of the user. For example, the remaining life corresponding to the first time period will obviously be greater than the remaining life corresponding to the 600th time period.

[0086] In some embodiments, see Figure 3B , Figure 3B It is a flow chart of the method for predicting the life of a part provided in an embodiment of the present application. After step 101 and before step 102, the following steps 106 to 114 are performed.

[0087] In step 106 , a plurality of sampling data of the part from the beginning of working is acquired, wherein each sampling data includes a plurality of sampling points.

[0088] As an example, it usually refers to multiple sampling data starting from normal operation. The time when normal operation starts can be determined based on historical experience time. The time when normal operation starts can be judged based on the actual working status, such as the vibration signal has stabilized, such as the amplitude of the vibration signal is stable within a set range.

[0089] In step 107, a plurality of sampling data sorted in ascending order based on the sampling time are subjected to a plurality of first sliding window processes based on the first sliding window size to obtain sampling points within each first sliding window result.

[0090] In some embodiments, in step 107, multiple sampling data are subjected to multiple first sliding window processing based on the first sliding window size, and before obtaining the sampling points in each first sliding window result, the sampling frequencies of the multiple sampling data are reciprocally processed to obtain the sampling intervals of the multiple sampling points of the multiple sampling data; the sampling interval is multiplied by the number of the multiple sampling points included in each sampling data to obtain the first sliding window size.

[0091] As an example, the size of the first sliding window is determined according to the time period corresponding to the sampling data. The size of the first sliding window can be determined by formula (1):

[0092]

[0093] Wherein, the sampling frequency is f, n sampling points are taken as one sampling data, and the size of the first sliding window can be expressed as w1.

[0094] In step 108, for each first sliding window result, the root mean square of the sampling points in the first sliding window result is determined.

[0095] As an example, if 0.1 seconds of data are sampled at each sampling point, the RMS (Root Mean Square) in the first sliding window w1 is calculated with a step size of 0.1 seconds. Square), for example, sampling is performed from the beginning of the stable operation of the part to obtain 100 sampling points, each sampling point samples 0.1 second of data, f is 10, that is, sampling is performed 10 times per second, 5 sampling points are used as a training sample, the size of the first sliding window is 0.5 seconds, and 0.1 seconds is used as the moving distance of each sliding window to calculate the RMS in the first sliding window obtained by each movement. For example, the root mean square is calculated based on the data values ​​(such as vibration signal values) of the first 5 sampling points (S1 to S5) of the 100 sampling points as the RMS in the first sliding window result, which can be represented by r1, and then the time period corresponding to the first sliding window result is moved backward by 0.1 seconds, that is, the root mean square is calculated based on the data values ​​(such as vibration signal values) of the 2nd to 6th sampling points (S2 to S6) of the 100 sampling points as the RMS in the second first sliding window result, which can be represented by r2, and so on, and the above calculation process is repeated until all 100 sampling points slide to.

[0096] In step 109, multiple RMSs are processed multiple times based on the second sliding window size to obtain the RMS in each second sliding window result.

[0097] In step 110, for each second sliding window result, a smoothed root mean square of the corresponding second sliding window result is determined based on the root mean square within the second sliding window result.

[0098] In some embodiments, the second sliding window size is larger than the first sliding window size. In step 110, the smoothed root mean square corresponding to the second sliding window result is determined based on the root mean square in the second sliding window result. This can be achieved by the following technical solution: the first sliding window size is processed based on multiples of the first integer to obtain the second sliding window size; the root mean square in the second sliding window result is processed by a first summation to obtain a first summation result; and the ratio of the first summation result to the second sliding window size is determined as the smoothed root mean square in the second sliding window result. Smoothing the RMS in the calculated multiple first sliding window results by means of sliding average can effectively reduce noise interference.

[0099] As an example, since the data values ​​(e.g., vibration signals) of the sampling points obtained by the data collector (e.g., sensor) all contain noise components, in order to reduce the interference of the noise, the RMS in the calculated multiple first sliding window results are smoothed by a sliding average method. The sliding average method adopts a second sliding window. The size of the second sliding window is represented by w2. w2 can be multiple times of w1, for example, 20 times of w1. The RMS after smoothing can be expressed as The smoothed RMS of any first sliding window result can be calculated using formula (2):

[0100]

[0101] Among them, r i represents the RMS calculated within the i-th first sliding window result, Represents r i The value after smoothing, w2 represents the size of the second sliding window, r i Characterized by r i Taking a second sliding window result as a starting point, all RMS in the second sliding window result are summed.

[0102] If the timestamps of all sampling points involved in a certain RMS are within the second sliding window result, the RMS is considered to be the RMS within the second sliding window result. The second sliding window can be understood as a number of timestamps of all sampling points involved in a certain RMS. i Slide on the constructed sequence, with each sliding step being a root mean square, and the number of root mean squares in the second sliding window result is a multiple of w2 and w1.

[0103] For example, continuing with the above example, if the size of the second sliding window is 1 second, which is equivalent to twice the size of the first sliding window, the RMS number in the second sliding window result is 2, for example, Yes 1 and r 2 The ratio of the sum of and w2, Yes 2 and r 3 The ratio of the sum of and w2, Yes 3 and r 4 The ratio of the sum of and w2.

[0104] In step 111, a plurality of smoothed RMSs are subjected to a plurality of third sliding window processes based on a third sliding window size to obtain a smoothed RMS in each third sliding window result.

[0105] In step 112, for each third sliding window result, a smoothed mean value of the corresponding third sliding window result is determined based on the smoothed root mean square within the third sliding window result.

[0106] In some embodiments, in step 112, determining the smoothed mean of the corresponding third sliding window result based on the smoothed root mean square in the third sliding window result can be achieved by the following technical solution: performing multiple processing based on the second integer on the second sliding window size to obtain the third sliding window size; performing a second summation processing on the smoothed root mean square in the third sliding window result to obtain a second summation result; and determining the ratio of the second summation result to the third sliding window size as the smoothed mean of the smoothed root mean square of the third sliding window result.

[0107] As an example, the energy stability of the working data is determined by using the third sliding window. The size of the third sliding window is represented by w3. w3 can be a multiple of w2, for example, 200 times of w2. The mean m2 of multiple smoothed RMS in a third sliding result is calculated. The mean m2 is calculated by formula (3):

[0108]

[0109] Where m2 is the mean of multiple smoothed RMS values ​​within a third sliding result, Represents r i The value after smoothing, w3 represents the size of the third sliding window, Characterized by Taking a third sliding result as a starting point, all smoothed RMS values ​​in the third sliding result are summed.

[0110] If the timestamps of all sampling points involved in a smoothed RMS are within the third sliding result, the smoothed RMS is considered to be the RMS within the third sliding result. The third sliding window can be understood as a window in multiple Slide on the constructed sequence, with each sliding step being a smoothed root mean square, and the number of smoothed root mean squares in the third sliding result is a multiple of w3 and w2.

[0111] For example, continuing with the above example, if the size of the third sliding window is 2 seconds, which is equivalent to twice the size of the second sliding window, the smoothed RMS number in the third sliding result is 2, for example, yes and The summation result is yes and The summation result is yes and The summation result of .

[0112] In step 113, based on the smoothed mean value corresponding to each third sliding window result, a target third sliding window result that meets the degradation period condition is obtained from multiple third sliding window results, and the earliest sampling time corresponding to the target third sliding window result is determined as the starting time of the degradation stage.

[0113] In some embodiments, in step 113, based on the smoothed mean value corresponding to each third sliding window result, obtaining the target third sliding window result that meets the degradation period condition from among the multiple third sliding window results can be implemented by the following technical solution: performing the following processing for each third sliding window result: determining the absolute value between the smoothed mean value and the stable mean value of the third sliding window result; performing multiple processing based on the third integer on the stable standard deviation to obtain the mapping standard deviation; when the absolute value is greater than the mapping standard deviation, determining that the third sliding window result is the target third sliding result that meets the degradation period condition. The number of target third sliding window results is at least one, and the earliest sampling time corresponding to at least one target third sliding window result is determined as the starting time. The target third sliding window result is obtained based on multiple smoothed root mean squares, each smoothed root mean square is obtained based on multiple root mean squares, and each root mean square is obtained based on multiple sampling points, so that the earliest sampling time from among the multiple sampling points corresponding to the target third sliding window result can be determined as the starting time, and by comparing the smoothed mean value of the third sliding window result with the stable mean value, the target third sliding result that meets the degradation period condition can be accurately determined.

[0114] In some embodiments, based on the sampling time corresponding to each root mean square, multiple corresponding smoothed root mean squares are sorted in ascending order, wherein each smoothed root mean square corresponds to the root mean square of the earliest sampling time in the corresponding second sliding window result; multiple smoothed root mean squares ranked first in the ascending sorting result are used as target smoothed root mean squares; the mean of multiple target smoothed root mean squares is determined as a stable mean, and the standard deviation of multiple target smoothed root mean squares is determined as a stable standard deviation. By determining the stable mean and the stable standard deviation, reference data of the healthy period of the part can be obtained, thereby accurately determining the degradation stage.

[0115] As an example, calculate the mean and standard deviation of the first m consecutive smoothed RMS, starting from the time the part starts to run steadily. The mean m1 and standard deviation σ of the system are as follows: the specific value of m depends on the actual working conditions. The mean m1 and standard deviation σ are taken as the mean and standard deviation of the healthy stage, that is, the stable mean and stable standard deviation.

[0116] If the absolute value of m2 minus m1 is greater than three times the standard deviation σ, it can be considered that the part enters the degradation period at the last moment of the third sliding window, where three times the standard deviation σ is the mapping standard deviation, see formula (4):

[0117] |m2-m1|>3σ (4);

[0118] In step 114 , when the target time period is after the starting time, it is determined that the target time period is in the degradation stage.

[0119] In step 102, when the target time period is completely in the degradation stage of the part, a plurality of time domain features and a plurality of frequency domain features are extracted from the target sampling data.

[0120] In some embodiments, extracting multiple time domain features and multiple frequency domain features from the target sampling data in step 102 can be achieved through the following technical solution: performing feature processing on the time domain working signals of multiple sampling points included in the target sampling data to obtain the time-frequency domain features of the target sampling data; wherein the time-frequency domain features include multiple time domain features and multiple frequency domain features, and the feature processing is short-time Fourier transform processing or wavelet transform processing. By utilizing the time domain features and the frequency domain features at the same time, the modeling effect can be effectively improved.

[0121] As an example, the sampled data includes multiple sampling points, each sampling point includes a sampling time point and a time domain working signal (such as a vibration signal of a part). Since the vibration signal is usually interfered by noise, the modeling effect of directly using the time domain working signal is not good. The time domain working signal is converted into time-frequency domain features using short-time Fourier transform or wavelet transform. The time-frequency domain features include time domain features and frequency domain features. Since the sampled data includes multiple sampling points, for example, the sampled data includes n sampling points, and the time span corresponding to these n sampling points is T, firstly, the time-frequency domain features of the time domain working signal are extracted using short-time Fourier transform or wavelet transform. After the time-frequency domain conversion, the sampled data changes from T dimension to T*F dimension, where T represents the time dimension and F represents the frequency dimension.

[0122] In step 103, a first attention process is performed on a plurality of time domain features and a plurality of frequency domain features to obtain time-frequency domain related features of the target domain.

[0123] See also Figure 3C , Figure 3C is a flow chart of the method for predicting the life of a part provided in an embodiment of the present application. In step 103, a plurality of time domain features and a plurality of frequency domain features are first processed to obtain the time-frequency domain related features of the target domain, which can be obtained by Figure 3C Steps 1031 to 1034 are implemented.

[0124] In step 1031, time domain self-attention processing is performed on multiple time domain features to obtain self-attention time domain features.

[0125] In some embodiments, in step 1031, time domain self-attention processing is performed on multiple time domain features to obtain self-attention time domain features, which can be achieved through the following technical solutions: time domain embedding processing is performed on multiple time domain features to obtain multiple time domain embedded features corresponding one-to-one to the multiple time domain features; time domain position coding processing is performed on the multiple time domain embedded features to obtain multiple time domain position codes corresponding one-to-one to the multiple time domain features; time domain self-attention coding processing is performed on the multiple time domain position codes to obtain multiple self-attention time domain features corresponding one-to-one to the multiple time domain features.

[0126] As an example, the time-frequency domain features are used as the input data of the life prediction model. In the time domain direction, the time domain features of a certain input training sample are embedded through the time domain embedding network to obtain the time domain embedding features of the historical moment. This process is equivalent to compressing the discrete data into a dense vector, and then position-encoding the time domain embedding features of the historical moment through the time domain position coding network to obtain the time domain position coding of the historical moment, and then encoding the time domain position coding of the historical moment through the time domain coding network to obtain the self-attention time domain features. The encoding process is to extract the autocorrelation of the time domain dimension through the multi-head self-attention mechanism. Therefore, the autocorrelation of the time domain dimension can be extracted, effectively improving the representation ability of subsequent life characteristics.

[0127] In step 1032, frequency domain self-attention processing is performed on multiple frequency domain features to obtain self-attention frequency domain features.

[0128] In some embodiments, in step 1032, frequency domain self-attention processing is performed on multiple frequency domain features to obtain self-attention frequency domain features, which can be achieved through the following technical solutions: frequency domain embedding processing is performed on multiple frequency domain features to obtain multiple frequency domain embedded features corresponding one-to-one to the multiple frequency domain features; frequency domain position encoding processing is performed on the multiple frequency domain embedded features to obtain multiple frequency domain position codes corresponding one-to-one to the multiple frequency domain features; frequency domain self-attention encoding processing is performed on the multiple frequency domain position codes to obtain multiple self-attention frequency domain features corresponding one-to-one to the multiple frequency domain features.

[0129] As an example, in the frequency domain direction, the frequency domain features of a certain input training sample are embedded through the frequency domain embedding network to obtain the frequency domain embedded features of the historical moment. This process is equivalent to compressing the discrete data into a dense vector, and then performing position encoding on the frequency domain embedded features of the historical moment through the frequency domain position encoding network to obtain the frequency domain position coding of the historical moment, and then encoding the frequency domain position coding of the historical moment through the frequency domain coding network to obtain the self-attention frequency domain features. The encoding process is to extract the autocorrelation of the frequency domain dimension through the multi-head self-attention mechanism. Therefore, the autocorrelation of the frequency domain dimension can be extracted, thereby effectively improving the characterization ability of subsequent life characteristics.

[0130] In step 1033, the self-attention time domain features and the self-attention frequency domain features are concatenated to obtain fused features.

[0131] In step 1034, the target frequency domain features corresponding to the target time period in the target time period and the fusion features in the multiple frequency domain features are subjected to fusion attention processing to obtain the time-frequency domain related features of the target domain.

[0132] In some embodiments, in step 1034, the target frequency domain features corresponding to the target moment in the target time period among multiple frequency domain features are fused with the fused features for attention processing to obtain the time-frequency domain related features of the target domain, which can be achieved by the following technical solutions: frequency domain embedding processing is performed on the target frequency domain features to obtain the target frequency domain embedded features corresponding to the target frequency domain features; frequency domain position encoding processing is performed on the target frequency domain embedded features to obtain the target frequency domain position coding corresponding to the target frequency domain features; attention decoding processing is performed on the target frequency domain position coding and the fused features to obtain the time-frequency domain related features of the target domain.

[0133] As an example, the frequency domain features of the target time are embedded through the frequency domain embedding network to obtain the frequency domain embedded features of the target time. For example, the target time is the current time, and the current time is the last sampling time point in the target sampling data. The embedding process is equivalent to compressing the discrete data into a dense vector, and then position encoding the frequency domain embedded features of the target time through the frequency domain position encoding network to obtain the frequency domain position encoding of the target time. The frequency domain position encoding and fusion features of the target time are used as the input of the decoding network, see Figure 7 , Figure 7 It is a structural diagram of the decoding network provided in an embodiment of the present application. The structure of the decoding network here adopts the structure of the decoder in the Transformer model. The decoding includes network layers such as a masked multi-head self-attention layer, a decoding residual connection layer, and a normalization layer. The decoding network uses a multi-head self-attention network model to extract the correlation between the frequency domain features of the target moment and the time-frequency domain features of the historical moment, and obtains decoding features containing comprehensive time-frequency domain correlation, that is, the time-frequency domain correlation features of the target domain.

[0134] In step 104, the time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain are subjected to a second attention process to obtain the life span features of the target domain.

[0135] In some embodiments, in step 104, the time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain are subjected to a second attention process to obtain the life characteristics of the target domain, which can be achieved by the following technical scheme: the time-frequency domain related features of the target domain are subjected to a dot product process with multiple time-frequency domain related features of the source domain to obtain the dot product result of each time-frequency domain related feature of the corresponding source domain; the dot product result of each time-frequency domain related feature of the corresponding source domain is subjected to maximum likelihood processing to obtain the attention weight corresponding to each time-frequency domain related feature of the source domain; based on the attention weight, each time-frequency domain related feature of the source domain is subjected to weighted summation processing to obtain the source domain matching feature; the source domain matching feature is subjected to fusion processing with the time-frequency domain related features of the target domain to obtain the life characteristics of the target domain. Through the embodiment of the present application, the time-frequency domain related features learned in other domains are migrated to the target domain, so as to improve the learning accuracy and the generalization degree of the life characteristics, thereby improving the accuracy of the subsequent use of the life characteristics to predict the remaining life.

[0136] As an example, see Fig. 9 , Fig. 9 : This is a query matching processing schematic diagram provided in an embodiment of the present application, wherein the decoding features (vectors) output by the decoding network of the life prediction model of the working data of the target domain working condition are used as the query vector, and the multiple original time-frequency domain related features of the source domain stored in the external storage structure are clustered to obtain the clustering processing result (multiple time-frequency domain related features of the source domain, i.e., cluster vectors of clustering processing) as the key vector, wherein the multiple original time-frequency domain related features of the source domain are also obtained by a processing method similar to steps 101 to 102, the only difference being that the training data of the source domain is different from the training data of the target domain, resulting in different parameters in the processing process, and the attention mechanism is used to calculate the correlation feature representation h2 (source domain matching feature) of the query vector h1 (time-frequency domain related features of the target domain) and the key vector, specifically, by a dot product method, the attention weight of each source domain time-frequency domain related feature is determined, and then the multiple source domain time-frequency domain related features are weightedly summed based on the attention weight to obtain the source domain matching feature, and then h1 and h2 are concatenated as the final life prediction feature, and finally the feature is used as the input of the fully connected layer in the life prediction model.

[0137] In step 105 , the remaining life of the part in the target domain starting from the target time period is determined based on the life characteristics of the target domain.

[0138] As an example, the remaining life from the target time period can be the remaining life from the starting moment of the target time period, or the remaining life from the end point of the target time period, or the remaining life from any moment in the target time period, which depends on the specific moment of the target moment in step 1034, that is, the remaining life from the target time period is essentially the remaining life calculated from the target moment.

[0139] In some embodiments, before executing step 102, multiple source domain training samples corresponding to the source domain are obtained; the multiple source domain training samples are forward propagated in the source domain life prediction model to obtain multiple first predicted remaining lifespans corresponding to the multiple source domain training samples; based on the multiple first predicted remaining lifespans and the multiple first pre-labeled remaining lifespans, a first loss is determined, and the parameters of the source domain life prediction model are updated based on the first loss; the parameters of the source domain life prediction model are used as initialization parameters of the target domain life prediction model; multiple target domain training samples corresponding to the target domain are obtained; the multiple target domain training samples are forward propagated in the target domain life prediction model to obtain multiple second predicted remaining lifespans corresponding to the multiple target domain training samples; based on the multiple second predicted remaining lifespans and the multiple second pre-labeled remaining lifespans, a second loss is determined, and based on the second loss, the parameters of the target domain life prediction model are updated based on the initialization parameters.

[0140] As an example, considering that the amount of data for the same working condition as the actual working condition of the part (part working environment) is limited, while the data for other working conditions is more, and each working condition has both similarities and specificities, the working data of the parts in other working conditions is used as the source domain data during the training process, and the working data of the parts in the working condition to be tested is used as the target domain data. The network framework of meta-learning is used to quickly learn the training samples, and the information learned from the source domain is transferred to the target working condition with a relatively small amount of data to be tested, so as to obtain the part life prediction model for the target working condition. The specific migration process is to use the data of other working conditions to train the life prediction model of other working conditions (source domain life prediction model), and use the parameters of the life prediction model of other working conditions obtained by training as the initialization parameters of the part life prediction model of the target working condition (target domain life prediction model).

[0141] The target sampling data of the part corresponding to the target time period in the target domain is obtained. When the target time period is completely in the degradation stage of the part, multiple time domain features and multiple frequency domain features are extracted from the target sampling data. Life prediction is performed only based on the data of the degradation period. In this way, inaccurate remaining life prediction caused by the healthy period data of the part can be avoided. The multiple time domain features and multiple frequency domain features are subjected to first attention processing to obtain the time-frequency domain related features of the target domain. Therefore, the correlation between the timing information and the frequency domain information of the part is considered. The time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain are subjected to second attention processing to obtain the life features of the target domain, thereby migrating the time-frequency domain related features learned in other domains to the target domain, thereby obtaining the life features of the target domain. Based on the life features of the target domain, the remaining life of the part in the target domain starting from the target time period is determined, which can effectively improve the accuracy of the remaining life prediction.

[0142] The following is an explanation of an exemplary application of the embodiments of the present application in a practical application scenario.

[0143] In some embodiments, a terminal receives target sampling data of a target time period sent by a sensor of industrial equipment, and the terminal sends a remaining life prediction request starting from the target time period to a server, wherein the life prediction request includes the target sampling data of the target time period, and detects whether the target time period belongs to the degradation stage of the part. When the target time period is completely in the degradation stage of the part, multiple time domain features and multiple frequency domain features are extracted from the target sampling data; a first attention processing is performed on the multiple time domain features and the multiple frequency domain features to obtain time-frequency domain related features of the target domain; a second attention processing is performed on the time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain to obtain the life features of the target domain; based on the life features of the target domain, the remaining life of the part in the target domain starting from the target time period is determined, and the remaining life is returned to the terminal for presentation.

[0144] The embodiment of the present application proposes a method for detecting the time when the bearing enters the degradation period based on the root mean square calculation of the working data (such as vibration signal) of the sliding window and combining the 3σ criterion to adaptively detect the time when the bearing enters the degradation period, and only uses the data of the degradation period for model training. Since the time domain working data is often interfered by noise, the embodiment of the present application first uses the signal processing method of time-frequency domain conversion to convert the time domain signal into time-frequency domain features. In order to fully explore the deeper time-frequency domain features and make full use of the time domain correlation, frequency domain correlation and time-frequency domain correlation, the embodiment of the present application proposes a network structure based on the time-frequency domain bidirectional self-attention mechanism as a life prediction model. In addition, since the working data of the target working condition that can be collected is limited, and different working conditions have both specificity and similarity, the life prediction model proposed in the embodiment of the present application is trained based on the meta-learning framework, and the working data of other working conditions can be used as the source domain data, and the working data of the target working condition can be used as the target domain data. The information learned from the source domain is transferred to the target working condition with a relatively small amount of working data, and the life prediction model of the parts for the target working condition is obtained.

[0145] The parts life prediction method provided in the embodiment of the present application is applicable to all scenarios involving sensor signal analysis and fault prediction. The collected sampling data can be converted from the time domain to the time-frequency domain through a signal processing method. The sampling data can be the status detection data of aircraft parts, the status detection data of transportation scene parts, etc.

[0146] In some embodiments, the life prediction method of parts provided in the embodiments of the present application can be applied to the scene of part degradation period detection or fault warning in specific working conditions, as well as the scene where the data volume of the target working condition is limited and the data volume of other working conditions is relatively abundant. The life prediction method of parts provided in the embodiments of the present application can detect the time when the parts enter the degradation period in real time and adaptively based on the calculation of the root mean square of the sliding window and the 3σ criterion, and fully extract the correlation of the time-frequency domain characteristics of the degradation period data to predict the remaining life of the parts based on the correlation. Moreover, the life prediction method of parts provided in the embodiments of the present application uses the model parameters obtained by data training based on multiple working conditions as the initialization parameters of the model training under the target working condition based on the meta-learning framework, thereby migrating the information learned from the source domain (other working conditions) to the target working condition with relatively less data to be measured, thereby improving the stability of the model migration.

[0147] See also Figure 4 , Figure 4 is an overall flow chart of the life prediction method of a part provided in an embodiment of the present application, Figure 4 The training process of the target domain life prediction model and the prediction process of the target domain life prediction model are shown in FIG. In the training process of the target domain life prediction model, the training data is the working data of the part after the degradation period detection algorithm determines that the part data has entered the degradation period. The working data of the part entering the degradation period is subjected to feature processing to obtain time-frequency domain features. The target domain life prediction model is trained using the time-frequency domain features to obtain a trained target domain life prediction model M. In the prediction process of the target domain life prediction model, the parts are subjected to real-time degradation period detection. Once the parts enter the degradation period, the working data of the parts entering the degradation period can be subjected to feature processing, and the time-frequency domain features obtained after feature processing are input into the trained target domain life prediction model M to obtain the predicted remaining life.

[0148] During the training process of the life prediction method for parts provided in the embodiment of the present application, it is first necessary to detect whether the part is in the degradation period, calculate the root mean square (RMS) of the working data (for example, vibration signal) based on the sliding window, and combine the 3σ criterion to detect the time when the part enters the degradation period in real time and adaptively, so as to avoid the working data of the part in a healthy state interfering with the training of the target domain life prediction model. In the prediction process, it is also necessary to detect whether the part is in the degradation period, calculate the root mean square (RMS) of the working data (for example, vibration signal) based on the sliding window, and combine the 3σ criterion to detect the time when the part enters the degradation period in real time and adaptively, so as to effectively obtain the working data for life prediction, and avoid using the working data of the part in a healthy state, which leads to the inability to accurately realize life prediction.

[0149] The processing flow of adaptive degradation period detection based on the RMS of the sliding window and the 3σ criterion is as follows:

[0150] During the training process, it is necessary to collect training data through adaptive degradation period detection. The size of the first sliding window is determined according to the size of the training samples in the training data. The size of the first sliding window can be determined by formula (5):

[0151]

[0152] Wherein, the sampling frequency is f, n sampling points are taken as a training sample, and the size of the first sliding window can be expressed as w1.

[0153] If 0.1 seconds of data are sampled at each sampling point, the RMS in the first sliding window w1 is calculated with a step size of 0.1s. For example, sampling is performed from the beginning of the stable operation of the part to obtain 100 sampling points, and 0.1 seconds of data are sampled at each sampling point. f is 10, that is, 10 samples are taken per second. 5 sampling points are used as a training sample. The size of the first sliding window is 0.5 seconds. 0.1 seconds is used as the moving distance of each sliding window to calculate the RMS in the first sliding window obtained by each movement. The timestamp of the first sampling point is used as the starting point of the first sliding window to determine the first sliding window. The data value of each sampling point corresponding to the time period of a sliding window, that is, the root mean square is calculated based on the data values ​​of the first 5 sampling points (S1 to S5) of the 100 sampling points, as the RMS in the first first sliding window, which can be represented by r1, and then the time period corresponding to the first first sliding window is moved backward by 0.1 seconds, that is, the root mean square is calculated based on the data values ​​of the 2nd to 6th sampling points (S2 to S6) based on the 100 sampling points, as the RMS in the second first sliding window, which can be represented by r2, and so on, the above calculation process is completed until all 100 sampling points are slid.

[0154] Since the vibration signal obtained by the data collector (such as a sensor) contains noise components, in order to reduce the interference of noise, the RMS in the first sliding windows is smoothed by sliding average. The sliding average method adopts a second sliding window. The size of the second sliding window is represented by w2. w2 can be multiple times of w1, for example, 20 times of w1. The RMS after smoothing can be expressed as The smoothed RMS of any RMS in the first sliding window can be calculated using formula (6):

[0155]

[0156] Among them, r i represents the calculated RMS in the i-th first sliding window, Represents ri The smoothed value, w2 represents the time size of the second sliding window, r i Characterized by r i As the starting point of a second sliding window, all RMS in the second sliding window are summed.

[0157] If the timestamps of all sampling points involved in a certain RMS are within the second sliding window, the RMS is considered to be the RMS within the second sliding window. The second sliding window can be understood as the number of sampling points within a plurality of r i Slide on the constructed sequence, with each sliding step being a root mean square, and the number of root mean squares in the second sliding window is a multiple of w2 and w1.

[0158] For example, continuing with the above example, if the size of the second sliding window is 1 second, which is equivalent to twice the size of the first sliding window, the RMS number in the second sliding window is 2, for example, Yes 1 and r 2 The ratio of the sum of and w2, Yes 2 and r 3 The ratio of the sum of and w2, Yes 3 and r 4 The ratio of the sum of and w2.

[0159] Calculate the mean and standard deviation of the RMS after smoothing for the previous several consecutive times, starting from the time when the part starts to run stably. The mean m1 and standard deviation σ of the system are as follows: the specific value of m depends on the actual working conditions, and the mean m1 and standard deviation σ are taken as the mean and standard deviation of the healthy stage.

[0160] Finally, the third sliding window is used to judge the energy stability of the working data. The size of the third sliding window is represented by w3, which can be a multiple of w2, for example, 200 times of w2. The mean m2 of multiple smoothed RMS values ​​in a third sliding window is calculated, and the mean m2 is calculated by formula (7):

[0161]

[0162] Where m2 is the mean of multiple smoothed RMS values ​​in a third sliding window, Represents r i The smoothed value, w3 represents the time size of the third sliding window, Characterized by As the starting point of a third sliding window, all the smoothed RMS values ​​in the third sliding window are summed.

[0163] If the timestamps of all sampling points involved in a smoothed RMS are within the third sliding window, the smoothed RMS is considered to be the RMS within the third sliding window. The third sliding window can be understood as the number of sampling points in a plurality of sliding windows. Slide on the constructed sequence, with each sliding step being a smoothed RMS, and the number of smoothed RMS in the third sliding window is a multiple of w3 and w2.

[0164] For example, continuing with the above example, if the size of the third sliding window is 2 seconds, which is equivalent to twice the size of the second sliding window, the smoothed RMS number in the third sliding window is 2, for example, yes and The summation result is yes and The summation result is yes and The summation result of .

[0165] If the absolute value of m2 minus m1 is greater than three times the standard deviation σ, it can be considered that the part enters the degradation period at the last moment of the third sliding window, see formula (8):

[0166] |m2-m1|>3σ (8);

[0167] The above degradation period detection process can be understood as judging the change of the part operation trend by comparing the stability of the part vibration signal energy in the short term with the stability of the part vibration signal energy in the healthy stage. After the starting timestamp of the degradation stage of the part is detected, the training samples after the starting timestamp are used for model training.

[0168] In the prediction process, the same method is used to detect the time when the part enters the degradation period, so as to determine whether the current moment is in the degradation period of the part. When it is determined that the training sample at the current moment is in the degradation period of the part, the training sample at the current moment is used as the input of the target domain life prediction model.

[0169] In some embodiments, the target domain life prediction model provided by the embodiments of the present application is a time-frequency domain bidirectional self-attention deep learning model, which is trained based on a meta-learning framework.

[0170] See also Figure 5 , Figure 5It is a structural diagram of the target domain life prediction model provided in an embodiment of the present application. Since the vibration signal as working data is usually interfered by noise, the direct use of time domain signals for modeling is not effective. First, the time domain signal is converted into time-frequency domain features (as the input of the target domain life prediction model) by short-time Fourier transform or wavelet transform. The time-frequency domain features of the parts are processed by the target domain life prediction model. The target domain life prediction model includes an encoder (time domain encoding network) and a decoder (decoding network). In the encoder, multi-head self-attention is used to perform encoding operations in the time domain direction and frequency domain direction respectively, and then the attention mechanism is used again by the decoder to extract the correlation between the current sampling data and the historical sampling data.

[0171] The training samples from the time when the detection enters the degradation period to the time when the operation fails are used as the data for the training model. Every n sampling points are still used as a training sample, and the time span corresponding to these n sampling points is T. Then, short-time Fourier transform or wavelet transform is first used to extract the time-frequency domain features of the time domain vibration signal (working data). After the time-frequency domain conversion, the single training sample changes from T dimension to T*F, where T represents the time dimension and F represents the frequency dimension.

[0172] The training process of the lifespan prediction model is as follows:

[0173] Considering that the amount of data for the same working condition as the actual working condition of the part (part working environment) is limited, while the data for other working conditions is more, and each working condition has both similarities and specificities, the working data of the parts in other working conditions is used as the source domain data, and the working data of the parts in the working condition to be tested is used as the target data during the training process. The network framework of meta-learning is used to quickly learn the training samples, and the information learned from the source domain is transferred to the target working condition with a relatively small amount of data to be tested, so as to obtain the part life prediction model for the target working condition. The specific migration process is to use the data of other working conditions to train the source domain life prediction model of other working conditions, and use the parameters of the source domain life prediction model of other working conditions obtained by training as the initialization parameters of the target domain life prediction model of the parts in the target working condition.

[0174] The time-frequency domain features are used as the input data of the life prediction model. In the time domain direction, the time domain features of a certain training sample are embedded through the time domain embedding network to obtain the time domain embedding features of the historical moment. This process is equivalent to compressing the discrete data into a dense vector, and then the time domain embedding features of the historical moment are position-encoded by the time domain position coding network to obtain the time domain position coding of the historical moment, and then the time domain position coding of the historical moment is encoded by the time domain coding network. The encoding process is to extract the autocorrelation of the time domain dimension through the multi-head self-attention mechanism; similarly, in the frequency domain direction, the frequency domain features of a certain training sample are embedded through the frequency domain embedding network to obtain the frequency domain embedding features of the historical moment. This process is equivalent to compressing the discrete data into a dense vector, and then the frequency domain embedding features of the historical moment are position-encoded by the frequency domain position coding network to obtain the frequency domain position coding of the historical moment, and then the frequency domain position coding of the historical moment is encoded by the frequency domain coding network. The encoding process is to extract the autocorrelation of the frequency domain dimension through the multi-head self-attention mechanism.

[0175] See also Figure 6 , Figure 6 It is a structural diagram of the time domain coding network provided in an embodiment of the present application. The structures of the frequency domain coding network and the time domain coding network here adopt the structure of the encoder in the Transformer model. The time domain coding network includes network layers such as time domain multi-head self-attention layer, time domain residual connection layer, and normalization layer. The structural diagram of the frequency domain coding network is similar to the structural diagram of the time domain coding network, and the only difference is that the parameters are different.

[0176] Continue to see Figure 5 , through the feature fusion network of the time domain features with time domain correlation, the time domain features with time domain correlation output by the time domain coding network and the frequency domain features with frequency domain correlation output by the frequency domain coding network are connected in series to form a fusion feature.

[0177] While executing the above processing, the frequency domain features of the current moment are embedded through the frequency domain embedding network to obtain the frequency domain embedded features of the current moment. This process is equivalent to compressing discrete data into a dense vector, and then position encoding the frequency domain embedded features of the current moment through the frequency domain position encoding network to obtain the frequency domain position coding of the current moment. The frequency domain position coding and fusion features of the current moment are used as the input of the decoding network.

[0178] See also Figure 7The structure of the decoding network here adopts the structure of the decoder in the Transformer model. The decoding includes network layers such as the masked multi-head self-attention layer, the decoding residual connection layer, and the normalization layer. The decoding network uses the multi-head self-attention network model to extract the correlation between the frequency domain features of the current moment and the time-frequency domain features of the historical moment, and obtains the decoding features containing the comprehensive time-frequency domain correlation (that is, the time-frequency domain correlation features of the target domain).

[0179] The embodiment of the present application utilizes a meta-learning framework that combines an attention mechanism and external storage to enable the life prediction model to learn information from multiple other working conditions, see Figure 8 , Figure 8 It is a structural diagram of the meta-learning framework provided by the embodiment of the present application. Since the essence of meta-learning is to learn the similarities behind different tasks, the embodiment of the present application uses the meta-learning framework to learn the common features of the working data of parts in different working conditions, and uses an external storage structure to save the common features; and since the working data of the target working condition is limited, and the time-frequency domain features of the parts will change with the changes in working conditions and time, in order to improve the stability of migration, it is necessary to learn the common features containing comprehensive time-frequency domain correlation from the working data of multiple working conditions. In order to reduce the memory pressure of the external storage structure, the common features of the working data of other working conditions are clustered, and the external storage structure only stores the cluster vectors after clustering. The decoded features of the target working condition (i.e., the time-frequency domain related features of the target domain) are queried and matched based on the external storage structure to obtain life prediction features.

[0180] See also Fig. 9 , the decoding features (i.e., the vector representation of the time-frequency domain related features of the target domain) output by the decoding network of the life prediction model of the working data of the target domain working condition are used as the query vector, and the cluster vector of the external storage structure is used as the key vector. The attention mechanism is used to calculate the correlation feature representation h2 between the query vector and the key vector, and then h1 and h2 are concatenated as the final life prediction feature, and then this feature is used as the input of the fully connected layer in the life prediction model.

[0181] The embodiment of the present application detects the time when a part enters the degradation period in real time and adaptively, avoiding the inaccuracy of life prediction using data generated in a healthy state. Since the amount of information contained in the time domain part data is limited, the embodiment of the present application fully exploits the correlation in the time domain, the correlation in the frequency domain, and the correlation between the current moment and the historical time and frequency domain characteristics of the part time and frequency domain features. Since the working data of the same working condition that can be collected is limited, the embodiment of the present application fully utilizes the working data of multiple working conditions, improving the generalization of the model while taking into account the specificity of specific working conditions.

[0182] The following is a description of an exemplary structure of a component life prediction device 455 provided in an embodiment of the present application implemented as a software module. In some embodiments, Figure 2 As shown, the software modules in the life prediction device 455 of the part stored in the memory 450 may include: a sampling module 4551, which is used to obtain target sampling data of the part corresponding to the target time period in the target domain; an extraction module 4552, which is used to extract multiple time domain features and multiple frequency domain features from the target sampling data when the target time period is completely in the degradation stage of the part; an attention module 4553, which is used to perform a first attention processing on the multiple time domain features and the multiple frequency domain features to obtain time-frequency domain related features of the target domain; the attention module 4553 is also used to perform a second attention processing on the time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain to obtain the life characteristics of the target domain; a life module 4554, which is used to determine the remaining life of the part in the target domain starting from the target time period based on the life characteristics of the target domain.

[0183] In some embodiments, when the target time period is completely in the degradation stage of the part, before extracting multiple time domain features and multiple frequency domain features from the target sampling data, the sampling module 4551 is also used to: obtain multiple sampling data of the part from the beginning of work, wherein each sampling data includes multiple sampling points; perform multiple first sliding window processing based on the first sliding window size on the multiple sampling data sorted in ascending order based on the sampling time, to obtain the sampling points in each first sliding window result; for each first sliding window result, determine the root mean square of the sampling points in the first sliding window result; perform multiple second sliding window processing based on the second sliding window size on the multiple root mean squares, to obtain the root mean square in each second sliding window result; for each first sliding window result, The method comprises the steps of: determining a smoothed root mean square of the second sliding window result and performing a third sliding window processing based on a third sliding window size on the multiple smoothed root mean squares to obtain a smoothed root mean square in each third sliding window result; determining a smoothed mean value of the corresponding third sliding window result based on the smoothed root mean square in the third sliding window result for each third sliding window result; obtaining a target third sliding window result that meets the degradation period condition among the multiple third sliding window results based on the smoothed mean value corresponding to each third sliding window result, and determining the earliest sampling time corresponding to the target third sliding window result as the starting time of the degradation stage; and determining that the target time period is in the degradation stage when the target time period is after the starting time.

[0184] In some embodiments, before performing multiple first sliding window processing based on the first sliding window size on multiple sampling data to obtain the sampling points in each first sliding window result, the sampling module 4551 is also used to: take the inverse of the sampling frequency of the multiple sampling data to obtain the sampling interval of the multiple sampling points of the multiple sampling data; multiply the sampling interval by the number of multiple sampling points included in each sampling data to obtain the first sliding window size.

[0185] In some embodiments, the second sliding window size is larger than the first sliding window size, and the sampling module 4551 is further used to: perform multiple processing based on the first integer on the first sliding window size to obtain the second sliding window size; perform a first summation processing on the root mean square within the second sliding window result to obtain a first summation result; and determine the ratio of the first summation result to the second sliding window size as the smoothed root mean square within the second sliding window result.

[0186] In some embodiments, the sampling module 4551 is also used to: perform multiple processing based on the second integer on the second sliding window size to obtain a third sliding window size; perform a second summation processing on the smoothed root mean square within the third sliding window result to obtain a second summation result; and determine the ratio of the second summation result to the third sliding window size as the smoothed mean of the smoothed root mean square of the third sliding window result.

[0187] In some embodiments, the sampling module 4551 is also used to: perform the following processing for each third sliding window result: determine the absolute value between the smoothed mean and the stable mean of the third sliding window result; perform multiple processing of the stable standard deviation based on the third integer to obtain a mapping standard deviation; when the absolute value is greater than the mapping standard deviation, determine that the third sliding window result is a target third sliding result that meets the degradation period condition.

[0188] In some embodiments, the sampling module 4551 is also used to: sort multiple corresponding smoothed root mean squares in ascending order based on the sampling time corresponding to each root mean square, wherein each smoothed root mean square corresponds to the root mean square of the earliest sampling time in the corresponding second sliding window result; use the multiple smoothed root mean squares with the highest order in the ascending sorting result as the target smoothed root mean square; determine the mean of the multiple target smoothed root mean squares as the stable mean, and determine the standard deviation of the multiple target smoothed root mean squares as the stable standard deviation.

[0189] In some embodiments, the extraction module 4552 is also used to: perform feature processing on the time domain working signals of multiple sampling points included in the target sampling data to obtain the time-frequency domain features of the target sampling data; wherein the time-frequency domain features include multiple time domain features and multiple frequency domain features, and the feature processing is short-time Fourier transform processing or wavelet transform processing.

[0190] In some embodiments, the attention module 4553 is also used to: perform time domain self-attention processing on multiple time domain features to obtain self-attention time domain features; perform frequency domain self-attention processing on multiple frequency domain features to obtain self-attention frequency domain features; splice the self-attention time domain features with the self-attention frequency domain features to obtain fused features; perform fused attention processing on the target frequency domain features corresponding to the target time period in the multiple frequency domain features and the fused features to obtain the time-frequency domain related features of the target domain.

[0191] In some embodiments, the attention module 4553 is also used to: perform time domain embedding processing on multiple time domain features to obtain multiple time domain embedded features corresponding one-to-one to the multiple time domain features; perform time domain position coding processing on multiple time domain embedded features to obtain multiple time domain position codes corresponding one-to-one to the multiple time domain features; perform time domain self-attention coding processing on multiple time domain position codes to obtain multiple self-attention time domain features corresponding one-to-one to the multiple time domain features.

[0192] In some embodiments, the attention module 4553 is also used to: perform frequency domain embedding processing on multiple frequency domain features to obtain multiple frequency domain embedded features corresponding one-to-one to the multiple frequency domain features; perform frequency domain position encoding processing on the multiple frequency domain embedded features to obtain multiple frequency domain position codes corresponding one-to-one to the multiple frequency domain features; perform frequency domain self-attention encoding processing on the multiple frequency domain position codes to obtain multiple self-attention frequency domain features corresponding one-to-one to the multiple frequency domain features.

[0193] In some embodiments, the attention module 4553 is also used to: perform frequency domain embedding processing on the target frequency domain features to obtain target frequency domain embedded features corresponding to the target frequency domain features; perform frequency domain position coding processing on the target frequency domain embedded features to obtain target frequency domain position coding corresponding to the target frequency domain features; perform attention decoding processing on the target frequency domain position coding and fusion features to obtain time-frequency domain related features of the target domain.

[0194] In some embodiments, the attention module 4553 is also used to: perform dot product processing on the time-frequency domain related features of the target domain and multiple time-frequency domain related features of the source domain to obtain the dot product results of each time-frequency domain related feature of the corresponding source domain; perform maximum likelihood processing on the dot product results of each time-frequency domain related feature of the corresponding source domain to obtain the attention weight corresponding to each time-frequency domain related feature of the source domain; based on the attention weight, perform weighted sum processing on each time-frequency domain related feature of the source domain to obtain the source domain matching feature; fuse the source domain matching feature with the time-frequency domain related features of the target domain to obtain the life span feature of the target domain.

[0195] In some embodiments, the device also includes a training module 4555, which is used to: obtain multiple source domain training samples corresponding to the source domain; forward propagate the multiple source domain training samples in the source domain life prediction model to obtain multiple first predicted remaining lifespans corresponding to the multiple source domain training samples; determine a first loss based on the multiple first predicted remaining lifespans and the multiple first pre-labeled remaining lifespans, and update the parameters of the source domain life prediction model based on the first loss; use the parameters of the source domain life prediction model as initialization parameters of the target domain life prediction model; obtain multiple target domain training samples corresponding to the target domain; forward propagate the multiple target domain training samples in the target domain life prediction model to obtain multiple second predicted remaining lifespans corresponding to the multiple target domain training samples; determine a second loss based on the multiple second predicted remaining lifespans and the multiple second pre-labeled remaining lifespans, and based on the second loss, update the parameters of the target domain life prediction model based on the initialization parameters.

[0196] The embodiment of the present application provides a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the above-mentioned method for predicting the life of a part in the embodiment of the present application.

[0197] The present application embodiment provides a computer-readable storage medium storing executable instructions, wherein the executable instructions are stored. When the executable instructions are executed by a processor, the processor will be caused to execute the life prediction method of the parts provided in the present application embodiment, for example, Figures 3A-3C A method for predicting the life of a component is shown.

[0198] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface storage, optical disk, or CD-ROM; or it may be various devices including one or any combination of the above memories.

[0199] In some embodiments, executable instructions may be in the form of a program, software, software module, script or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine or other unit suitable for use in a computing environment.

[0200] As an example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files storing one or more modules, subroutines, or code portions).

[0201] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices located at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0202] In summary, through the embodiment of the present application, the target sampling data of the part corresponding to the target time period in the target domain is obtained. When the target time period is completely in the degradation stage of the part, multiple time domain features and multiple frequency domain features are extracted from the target sampling data. Life prediction is performed only based on the data of the degradation period. This can avoid the inaccurate remaining life prediction caused by the healthy period data of the part. The multiple time domain features and the multiple frequency domain features are subjected to first attention processing to obtain the time-frequency domain related features of the target domain. Therefore, the correlation between the timing information and the frequency domain information of the part is considered. The time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain are subjected to second attention processing to obtain the life features of the target domain, thereby migrating the time-frequency domain related features learned in other domains to the target domain, thereby obtaining the life features of the target domain. Based on the life features of the target domain, the remaining life of the part in the target domain starting from the target time period is determined, which can effectively improve the accuracy of the remaining life prediction.

[0203] The above is only an embodiment of the present application and is not intended to limit the protection scope of the present application. Any modifications, equivalent substitutions and improvements made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. A method for predicting the life of a component, characterized in that: The method comprises: Obtaining target sampling data of the part in a target domain corresponding to a target time period; Performing multiple first sliding window processing based on the first sliding window size on a plurality of sampling data of the part sorted in ascending order according to the sampling time since the start of work, and determining the root mean square of the sampling points in each first sliding window result, wherein each of the sampling data includes a plurality of sampling points; Performing multiple second sliding window processing based on the second sliding window size on the multiple RMSs, and determining a smoothed RMS of the RMSs in each second sliding window result; Performing a plurality of third sliding window processes based on a third sliding window size on the plurality of smoothed RMSs, and determining a smoothed mean value of the smoothed RMSs in each third sliding window result; Determine a target third sliding window result that meets the degradation period condition among the plurality of third sliding window results based on the smoothed mean, and determine the earliest sampling time corresponding to the target third sliding window result as the starting time of the degradation stage; When the target time period is after the starting time, determining that the target time period is in the degradation stage; When the target time period is completely in the degradation stage of the part, extracting a plurality of time domain features and a plurality of frequency domain features from the target sampling data; Performing a first attention process on the multiple time domain features and the multiple frequency domain features to obtain time-frequency domain related features of the target domain; Performing a second attention process on the time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain to obtain the life span features of the target domain; Based on the life characteristics of the target domain, a remaining life of the part in the target domain starting from the target time period is determined.

2. The method according to claim 1, characterized in that: Before performing a plurality of first sliding window processes based on a first sliding window size on the plurality of sampling data sorted in ascending order according to the sampling time since the start of the work of the part, the method further comprises: Performing reciprocal processing on the sampling frequencies of the plurality of sampling data to obtain sampling intervals of a plurality of sampling points of the plurality of sampling data; The sampling interval is multiplied by the number of sampling points included in each of the sampling data to obtain the first sliding window size.

3. The method according to claim 1, characterized in that The second sliding window size is larger than the first sliding window size, and determining the smoothed root mean square of the root mean square in each second sliding window result includes: Performing multiple processing based on the first integer on the first sliding window size to obtain the second sliding window size; Performing a first summation process on the root mean square in the second sliding window result to obtain a first summation result; The ratio of the first summation result to the second sliding window size is determined as a smoothed root mean square corresponding to the second sliding window result.

4. The method according to claim 1, characterized in that: The third sliding window size is larger than the second sliding window size, and determining the smoothed mean of the smoothed root mean square in each third sliding window result includes: Performing multiple processing based on the second integer on the second sliding window size to obtain the third sliding window size; Performing a second summation process on the smoothed RMS in the third sliding window result to obtain a second summation result; The ratio of the second summation result to the third sliding window size is determined as a smoothed mean of the smoothed root mean square of the third sliding window result.

5. The method according to claim 1, characterized in that The determining, based on the smoothed mean, a target third sliding window result that satisfies the degradation period condition among the plurality of third sliding window results comprises: The stable standard deviation is processed based on multiples of the third integer to obtain a mapping standard deviation; The following processing is performed for each of the third sliding window results: Determine an absolute value between a smoothed mean and a stable mean of the third sliding window result; When the absolute value is greater than the mapping standard deviation, it is determined that the third sliding window result is a target third sliding result that satisfies the degradation period condition.

6. The method according to claim 5, characterized in that The method further comprises: Based on the sampling time corresponding to each of the root mean squares, a plurality of corresponding smoothed root mean squares are sorted in ascending order, wherein each of the smoothed root mean squares corresponds to the root mean square of the earliest sampling time in the corresponding second sliding window result; The top smoothed RMS in the ascending sorting results are used as the target smoothed RMS; The mean of the plurality of target smoothed root mean squares is determined as the stable mean, and the standard deviation of the plurality of target smoothed root mean squares is determined as the stable standard deviation.

7. The method according to claim 1, characterized in that The extracting a plurality of time domain features and a plurality of frequency domain features from the target sampled data comprises: Performing feature processing on time-domain working signals of a plurality of sampling points included in the target sampling data to obtain time-frequency domain features of the target sampling data; The time-frequency domain features include the multiple time-domain features and the multiple frequency-domain features, and the feature processing is short-time Fourier transform processing or wavelet transform processing.

8. The method according to claim 1, characterized in that: The performing first attention processing on the multiple time domain features and the multiple frequency domain features to obtain the time-frequency domain related features of the target domain includes: Performing time-domain self-attention processing on the multiple time-domain features to obtain self-attention time-domain features; Performing frequency domain self-attention processing on the multiple frequency domain features to obtain self-attention frequency domain features; Concatenate the self-attention time domain feature and the self-attention frequency domain feature to obtain a fusion feature; The target frequency domain features corresponding to the target moment in the target time period among the multiple frequency domain features are fused with the fusion features to obtain the time-frequency domain related features of the target domain.

9. The method according to claim 8, characterized in that The performing time-domain self-attention processing on the multiple time-domain features to obtain self-attention time-domain features includes: Performing time domain embedding processing on the multiple time domain features to obtain multiple time domain embedding features corresponding to the multiple time domain features one by one; Performing time domain position coding processing on the multiple time domain embedding features to obtain multiple time domain position codes corresponding to the multiple time domain features one by one; The multiple time-domain position codes are subjected to time-domain self-attention coding processing to obtain multiple self-attention time-domain features corresponding one-to-one to the multiple time-domain features.

10. The method according to claim 8, characterized in that The performing frequency domain self-attention processing on the multiple frequency domain features to obtain self-attention frequency domain features includes: Performing frequency domain embedding processing on the multiple frequency domain features to obtain multiple frequency domain embedding features corresponding one to one to the multiple frequency domain features; Performing frequency domain position coding processing on the multiple frequency domain embedded features to obtain multiple frequency domain position codes corresponding to the multiple frequency domain features one by one; The multiple frequency domain position codes are subjected to frequency domain self-attention coding processing to obtain multiple self-attention frequency domain features corresponding one-to-one to the multiple frequency domain features.

11. The method according to claim 8, characterized in that The step of fusing the target frequency domain features corresponding to the target time in the target time period with the fusion features among the multiple frequency domain features to obtain the time-frequency domain related features of the target domain includes: Performing frequency domain embedding processing on the target frequency domain feature to obtain a target frequency domain embedding feature corresponding to the target frequency domain feature; Performing frequency domain position coding processing on the target frequency domain embedding feature to obtain a target frequency domain position coding corresponding to the target frequency domain feature; The target frequency domain position encoding and the fusion feature are subjected to attention decoding processing to obtain the time-frequency domain related features of the target domain.

12. The method according to claim 1, characterized in that The step of performing a second attention process on the time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain to obtain the life span features of the target domain includes: Performing a dot product process on the time-frequency domain related features of the target domain and the multiple time-frequency domain related features of the source domain to obtain a dot product result corresponding to each time-frequency domain related feature of the source domain; Performing maximum likelihood processing on the dot product result of each time-frequency domain related feature corresponding to the source domain to obtain an attention weight corresponding to each time-frequency domain related feature of the source domain; Based on the attention weight, weighted sum processing is performed on each time-frequency domain related feature of the source domain to obtain a source domain matching feature; The source domain matching features are fused with the time-frequency domain related features of the target domain to obtain the life span features of the target domain.

13. The method according to claim 1, characterized in that The first attention processing, the second attention processing, and the processing of determining the remaining life are implemented by a target domain life prediction model, and the method further includes: Acquire a plurality of source domain training samples corresponding to the source domain; Forward propagating the plurality of source domain training samples in a source domain life prediction model to obtain a plurality of first predicted remaining lifespans corresponding one-to-one to the plurality of source domain training samples; Determining a first loss based on the plurality of first predicted remaining lifetimes and the plurality of first pre-marked remaining lifetimes, and updating a parameter of the source domain lifetime prediction model based on the first loss; Using the parameters of the source domain life prediction model as initialization parameters of the target domain life prediction model; Acquire a plurality of target domain training samples corresponding to the target domain; Forward propagating the plurality of target domain training samples in a target domain life prediction model to obtain a plurality of second predicted remaining lifespans corresponding one-to-one to the plurality of target domain training samples; Based on the multiple second predicted remaining lifetimes and the multiple second pre-marked remaining lifetimes, a second loss is determined, and based on the second loss, the parameters of the target domain lifetime prediction model are updated on the basis of the initialization parameters.

14. A device for predicting the life of a part, characterized in that: The device comprises: A sampling module, used to obtain target sampling data of the part in a target domain corresponding to a target time period; An extraction module is used to perform multiple first sliding window processings based on a first sliding window size on a plurality of sampling data of the part sorted in ascending order according to sampling time since the start of work, and determine the root mean square of the sampling points in each first sliding window result, wherein each of the sampling data includes multiple sampling points; perform multiple second sliding window processings based on a second sliding window size on the multiple root mean squares, and determine the smoothed root mean square of the root mean square in each second sliding window result; perform multiple third sliding window processings based on a third sliding window size on the multiple smoothed root mean squares, and determine the smoothed mean of the smoothed root mean square in each third sliding window result; determine a target third sliding window result that meets the degradation period condition among the multiple third sliding window results based on the smoothed mean, and determine the earliest sampling time corresponding to the target third sliding window result as the starting time of the degradation stage; when the target time period is after the starting time, determine that the target time period is in the degradation stage; when the target time period is completely in the degradation stage of the part, extract multiple time domain features and multiple frequency domain features from the target sampling data; An attention module, configured to perform a first attention process on the multiple time domain features and the multiple frequency domain features to obtain time-frequency domain related features of the target domain; The attention module is further used to perform a second attention process on the time-frequency domain related features of the target domain and the time-frequency domain related features of the source domain to obtain the life span features of the target domain; The life module is used to determine the remaining life of the part in the target domain starting from the target time period based on the life characteristics of the target domain.

15. An electronic device, characterized in that: The electronic device comprises: A memory for storing executable instructions; A processor, configured to implement the method for predicting the life of a part as described in any one of claims 1 to 13 when executing the executable instructions stored in the memory.

16. A computer-readable storage medium storing executable instructions, characterized in that: When the executable instructions are executed by a processor, the life prediction method of a component according to any one of claims 1 to 13 is implemented.

17. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the life prediction method of a component according to any one of claims 1 to 13 is implemented.

Images