AI-Based Analysis Method and System for Maintenance Record Data

Abstract

The present invention relates to the field of data processing, and provides an analysis method and system for maintenance record data based on AI. By obtaining the first fault record data stream of a target terminal device and performing a segmentation operation, a first to-be-processed fault record data stream and a second to-be-processed fault record data stream are obtained; the first to-be-processed fault record data stream is loaded into a spatio-temporal feature fusion component in a fault forward reasoning network for multi-dimensional feature aggregation to obtain a spatio-temporal aggregation feature array; the second to-be-processed fault record data stream and the spatio-temporal aggregation feature array are loaded into a fault decision-making component in the fault forward reasoning network for fault mode decision-making to obtain a target fault prediction result corresponding to the target terminal device at a forward-looking time point. The present invention can enable the fault forward reasoning network to accurately process the first fault record data stream with a large time interval, and improve the prediction accuracy of the target fault prediction result for the target terminal device at the forward-looking time point.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and more particularly, to an AI-based analysis method and system for maintenance record data. Background Art

[0002] In many fields such as current industrial production, smart home, and communication, the wide application of various terminal devices has greatly promoted the development and progress of society. However, during the long-term operation of these devices, faults are inevitable. How to diagnose and predict device faults in a timely and accurate manner has become a key issue for ensuring the stable operation of devices and reducing maintenance costs.

[0003] Traditional device fault diagnosis methods mainly rely on manual experience and simple threshold judgment. Technicians rely on their professional knowledge and practical experience to judge whether a device has a fault by observing the operating state of the device and detecting key parameters. It has extremely high requirements for the experience of technicians, and there may be significant differences in the judgment results of different technicians; in addition, it is very difficult for humans to detect some potential and early faults in a timely manner.

[0004] With the development of data analysis technology, data-driven fault diagnosis methods have gradually emerged. Feasible data-driven methods include fault diagnosis algorithms based on machine learning, such as decision trees, support vector machines, etc. To a certain extent, they improve the accuracy and efficiency of fault diagnosis, but most of them can only diagnose existing faults and lack the ability to prospectively predict future faults. Especially for data streams with a large time interval, the perception of fault characteristics is insufficient, making the prediction accuracy not meet the requirements of practical applications. Summary of the Invention

[0005] In view of this, the present invention provides an AI-based analysis method and system for maintenance record data.

[0006] According to one aspect of the present invention, there is provided an AI-based analysis method for maintenance record data, the method comprising: obtaining a first fault record data stream corresponding to a target terminal device, the first fault record data stream representing the fault record data of the target terminal device in a first operation and maintenance monitoring period; performing a segmentation operation on the first fault record data stream to obtain a first to-be-processed fault record data stream and a second to-be-processed fault record data stream; loading the first to-be-processed fault record data stream into a spatio-temporal feature fusion component in a fault forward reasoning network for multi-dimensional feature aggregation to obtain a spatio-temporal aggregation feature array; loading the second to-be-processed fault record data stream and the spatio-temporal aggregation feature array into a fault decision-making component in the fault forward reasoning network for fault mode decision-making to obtain a target fault prediction result corresponding to the target terminal device at a forward-looking time point.

[0007] According to another aspect of the present invention, there is provided a computer system, including: one or more processors; and one or more memories, wherein computer-readable code is stored in the memories, and when the computer-readable code is run by the one or more processors, the one or more processors are caused to execute the method as described above.

[0008] The beneficial effects of the present invention at least include: based on the segmentation of the first fault record data stream of the target terminal device, the obtained first to-be-processed fault record data stream is loaded into the spatio-temporal feature fusion component in the fault forward reasoning network for multi-dimensional feature aggregation to obtain a spatio-temporal aggregation feature array, and the second to-be-processed fault record data stream and the spatio-temporal aggregation feature array are loaded into the fault decision component in the fault forward reasoning network for fault mode decision to obtain the target fault prediction result of the target terminal device at the forward-looking time point. Based on the spatio-temporal feature fusion component and the fault decision component, the fault forward reasoning network can process the first fault record data stream with a large time interval, increasing the prediction accuracy of the target fault prediction result of the target terminal device at the forward-looking time point, which is convenient for making forward-looking preparations for equipment maintenance.

[0009] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and do not limit the technical solution of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. The drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0011] Figure 1 is a schematic diagram of an application scenario provided by the present invention;

[0012] Figure 2 is a schematic flowchart of an analysis method for maintenance record data based on AI provided by the present invention;

[0013] Figure 3 is a schematic structural diagram of a computer system provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0015] To facilitate a clearer understanding of the present invention, first, a schematic diagram of the application scenario for implementing the AI-based maintenance record data application of the present invention is introduced. As Figure 1 shown, the application scenario includes a computer system 10 and a terminal cluster. The terminal cluster can include one or more terminals, and the number of terminals will not be limited here. As Figure 1 shown, the terminal cluster can specifically include Terminal 1, Terminal 2,..., Terminal n. It can be understood that Terminal 1, Terminal 2, Terminal 3,..., Terminal n can all be network-connected to the computer system 10 so that each terminal can perform data interaction with the computer system 10 through the network connection.

[0016] It can be understood that the computer system 10 can refer to the device that executes the method provided in the embodiments of the present invention, such as an operation and maintenance device, a server, a personal computer, etc. Among them, the server can be an independent physical server, or a server cluster or distributed system composed of at least two physical servers. It can also be 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, Content Delivery Network (CDN), and big data and artificial intelligence platforms. The terminal can specifically refer to a terminal device that needs to be maintained after being sold, such as an in-vehicle terminal, an industrial machine, a smart home appliance, a communication base station, etc., but is not limited thereto. Each terminal and the computer system 10 can be directly or indirectly connected through wired or wireless communication methods.

[0017] Further, please refer to Figure 2 , which is a flowchart of an analysis method for AI-based maintenance record data provided by an embodiment of the present invention. As Figure 2 shown, this method can be executed by the Figure 1 computer system 10 in. Among them, the analysis method for AI-based maintenance record data can include the following steps:

[0018] Step S100: Obtain a first failure record data stream corresponding to the target terminal device, and the first failure record data stream represents the failure record data of the target terminal device in the first operation and maintenance monitoring cycle.

[0019] The target terminal device is the object that needs to be remotely maintained and repaired, such as various devices that have been sold, such as machines on industrial production lines, smart home appliances, communication base stations, etc.

[0020] The first operation and maintenance monitoring period is a selected time period used to collect the fault record data of the target terminal device. This period can be set according to actual needs, such as one day, one week, one month, etc. Assuming that the first operation and maintenance monitoring period is set to one week, then all the fault record data of the target terminal device within this week will be collected.

[0021] The fault record data refers to the relevant information generated when the target terminal device fails during the first operation and maintenance monitoring period. This information can include the time of the fault occurrence, the type of the fault, the severity of the fault, etc.

[0022] The first fault record data stream is a data stream formed by arranging the fault record data of the target terminal device in a certain order during the first operation and maintenance monitoring period. This data stream can be time-series data, that is, arranged in the chronological order of the fault occurrence. For example, if it is collected that the target terminal device has three faults within a week, namely fault A at 10 am on Monday, fault B at 3 pm on Wednesday, and fault C at 8 pm on Friday, then the first fault record data stream can be expressed as [(10 am on Monday, fault A), (3 pm on Wednesday, fault B), (8 pm on Friday, fault C)].

[0023] For the target terminal device with communication function, a connection can be established with the target terminal device through the network communication protocol to receive the fault record data sent by the target terminal device in real time. For example, the machine on the industrial production line can send the fault record data to the computer system through the industrial Ethernet. Secondly, for the target terminal device without communication function, data can be collected through an intermediate device. For example, the smart home appliance can forward the fault record data to the computer system through the smart home gateway. In addition, the fault record data can also be read from the local storage device of the target terminal device. For example, the communication base station can store the fault record data in the local hard disk, and these data can be read through remote login.

[0024] Step S200: Perform a splitting operation on the first fault record data stream to obtain the first to-be-processed fault record data stream and the second to-be-processed fault record data stream.

[0025] The first fault record data stream is a data stream in which the fault record data of the target terminal device in the first operation and maintenance monitoring cycle are arranged in a certain order. In order to better predict faults, it is divided into two parts, namely the first fault record data stream to be processed and the second fault record data stream to be processed. This segmentation helps the subsequent feature extraction components (such as spatiotemporal feature fusion components) and decision components (such as fault decision components) to better process data in long time intervals, and capture the data failure modes of earlier time periods and later time periods respectively. By segmenting the data stream, the relative time intervals and sequential relationships between faults in different time periods can also be better captured, thereby improving the accuracy of fault prediction.

[0026] Specifically, a technical means for segmentation based on a time window length and a preset segmentation coefficient can be used. For example, the actual time window length of the first fault record data stream is obtained, and the actual time window length represents the data stream length of the first fault record data stream, which is the time difference between the initial fault record time and the latest fault record time in the data stream. Then, the actual time window length of the first fault record data stream is multiplied by the preset segmentation coefficient to obtain a segmentation product. Then, the segmentation product is subjected to a certain operation (which can be determined according to actual conditions, such as addition, etc.) with the initial fault time in the first fault record data stream to obtain the segmentation position corresponding to the first fault record data stream. Finally, according to the segmentation position, the fault record data streams located on both sides of the segmentation position in the first fault record data stream are respectively used as the first to-be-processed fault record data stream and the second to-be-processed fault record data stream, and the fault record data corresponding to the segmentation position can be determined according to specific rules to belong to the first to-be-processed fault record data stream or the second to-be-processed fault record data stream. The preset segmentation coefficient can be adjusted according to actual experience or experiments. For example, if the preset segmentation coefficient is set to 0.6, it means that the first fault record data stream is segmented according to a ratio of 60% and 40%. The most suitable splitting factor for the current target terminal device can be determined through historical experience or experiments.

[0027] In the subsequent steps, the first and second fault record data streams to be processed obtained by segmentation have different roles in the fault prospective reasoning network. The first fault record data stream to be processed will be loaded into the spatiotemporal feature fusion component in the fault prospective reasoning network for multi-dimensional feature aggregation to mine the spatiotemporal feature information in the fault record data. The second fault record data stream to be processed and the spatiotemporal aggregation feature array obtained by processing the spatiotemporal feature fusion component will be loaded into the fault decision component in the fault prospective reasoning network for fault mode decision, thereby obtaining the target fault prediction result corresponding to the target terminal device at the prospective time point.

[0028] Step S300: Load the first data stream of fault records to be processed into the spatio-temporal feature fusion component in the fault forward reasoning network for multi-dimensional feature aggregation, and obtain a spatio-temporal aggregation feature array.

[0029] The first data stream of fault records to be processed is a part split from the first fault record data stream in step S200. The fault forward reasoning network is a neural network model for predicting the fault conditions of the target terminal device at the forward-looking time point. The spatio-temporal feature fusion component therein is a part of this network, for example, an encoder, such as a recurrent neural network or its variant, whose function is to perform multi-dimensional feature aggregation on the input fault record data and complete the encoding of the data.

[0030] The process of multi-dimensional feature aggregation is a feature extraction process, which extracts and integrates the features of the data stream in terms of time and space to obtain a feature extraction result suitable for analysis. For example, in the time dimension, faults occurring at different times may have different importance and relevance; in the space dimension, the distribution locations where different types of faults occur, etc. Through multi-dimensional feature aggregation, the potential rules and features in the fault record data can be captured more comprehensively.

[0031] The spatio-temporal aggregation feature array is the result obtained after being processed by the spatio-temporal feature fusion component, and is an array containing multi-dimensional feature information. This array will be used as an important input for subsequent fault mode decision-making to help the computer system more accurately predict the fault conditions of the target terminal device at the forward-looking time point.

[0032] In the implementation process, the computer system can index the corresponding temporal decay vector of each fault record data in the first data stream of fault records to be processed according to the target dynamic decay mapping matrix in the fault forward reasoning network. The target dynamic decay mapping matrix contains the corresponding relationships between different fault record data and the target temporal decay vectors. The temporal decay vectors are used to dynamically adjust the time decay weights of different fault records, enabling the neural network to distinguish the importance of fault types, capture long-term dependencies, and adapt to time sensitivity. For example, faults that occurred more recently may be more important for the current fault prediction, so the weights of their corresponding temporal decay vectors may be larger.

[0033] Then, according to the temporal distribution of each fault record data in the first fault record data stream to be processed in this data stream, determine the first target distribution positioning vector corresponding to the first fault record data stream to be processed. This vector can reflect the distribution characteristics of the fault record data in the time series. Next, perform a vector fusion operation on the first target temporal decay vector and the first target distribution positioning vector to obtain a spatio-temporal fusion input feature array. The vector fusion operation can adopt various methods, such as weighted summation, etc. Finally, load the spatio-temporal fusion input feature array into the spatio-temporal feature aggregation layer in the spatio-temporal feature fusion component for multi-dimensional feature aggregation to obtain a spatio-temporal aggregation feature array. The spatio-temporal feature aggregation layer can adopt deep learning models such as convolutional neural networks and recurrent neural networks to process the spatio-temporal fusion input feature array and extract the multi-dimensional feature information therein.

[0034] In practical applications, it is necessary to ensure the accuracy and rationality of the target dynamic decay mapping matrix. The design of the target dynamic decay mapping matrix needs to consider the characteristics of different fault types and the impact of time on fault prediction. The target dynamic decay mapping matrix can be optimized through a large number of experiments and data analyses to enable it to better reflect the actual situation of the fault record data.

[0035] The performance of the spatio-temporal feature fusion component will also affect the quality of the spatio-temporal aggregation feature array. Its performance can be optimized by adjusting the parameters of the spatio-temporal feature fusion component, selecting appropriate deep learning models, etc. For example, for different types of target terminal devices, spatio-temporal feature fusion components with different structures can be selected to better adapt to the characteristics of their fault record data.

[0036] The obtained spatio-temporal aggregation feature array will provide an important basis for subsequent fault mode decisions. By analyzing the feature information in the spatio-temporal aggregation feature array, it is possible to more accurately judge the possible fault types and probabilities of the target terminal device at the forward-looking time point.

[0037] Step S400: Load the second fault record data stream to be processed and the spatio-temporal aggregation feature array into the fault decision component in the fault forward-looking inference network for fault mode decision to obtain the target fault prediction result corresponding to the target terminal device at the forward-looking time point.

[0038] The second fault record data stream to be processed is another part split from the first fault record data stream in step S200, and together with the first fault record data stream to be processed, it constitutes the first fault record data stream. The spatio-temporal aggregation feature array is the result obtained in step S300 through multi-dimensional feature aggregation by loading the first fault record data stream to be processed into the spatio-temporal feature fusion component in the fault forward-looking inference network.

[0039] The fault decision-making component in the fault anticipation inference network is, for example, a decoder, such as the decoding structure in a recurrent neural network or its variants, or the decoding structure in a Transformer, or a fully connected network, which is used to make a fault mode decision based on the input second data stream of fault records to be processed and the spatio-temporal aggregation feature array. Fault mode decision-making refers to judging the possible fault types and probabilities of the target terminal device at the anticipation time point according to the input data.

[0040] The target fault prediction result is the result output after being processed by the fault decision-making component, which reflects the fault prediction situation corresponding to the target terminal device at the anticipation time point.

[0041] Taking the machine on an industrial production line as an example, the second data stream of fault records to be processed may contain information such as the fault occurrence time and fault type of the machine in the recent period, and the spatio-temporal aggregation feature array contains multi-dimensional feature information extracted from the previous fault record data. The fault decision-making component will comprehensively consider this information, judge the possible fault types of the machine at the anticipation time point (such as the next week), such as motor faults, sensor faults, etc., and give the corresponding fault occurrence probabilities.

[0042] When implementing step S400, for example, according to the target dynamic decay mapping matrix in the fault anticipation reasoning network, index the temporal decay vectors corresponding to each fault record data in the second to-be-processed fault record data stream. The target dynamic decay mapping matrix contains the correspondence between different fault record data and the target temporal decay vectors. The temporal decay vectors are used to dynamically adjust the time decay weights of different fault records, enabling the neural network to distinguish the importance of fault types, capture long-term dependencies, and adapt to time sensitivity. For example, newly occurred faults may be more important for current fault prediction, so the weights of their corresponding temporal decay vectors may be larger. Then, according to the temporal distribution of each fault record data in the second to-be-processed fault record data stream, determine the second target distribution positioning vector corresponding to the second to-be-processed fault record data stream. This vector can reflect the distribution characteristics of fault record data in the time series. Then, perform a vector fusion operation on the second target temporal decay vector and the second target distribution positioning vector to obtain a decision input fusion vector. The vector fusion operation can adopt various methods, such as weighted summation, etc. After that, load the decision input fusion vector into the decision feature extraction layer in the fault decision component to extract decision features and obtain a primary decision vector. The decision feature extraction layer can use deep learning models such as convolutional neural networks and recurrent neural networks to process the decision input fusion vector and extract the key feature information therein. Finally, load the primary decision vector and the spatio-temporal aggregation feature array into the cross-attention decision layer in the fault decision component for fault decision processing to obtain the target fault prediction result. The cross-attention decision layer can, through the attention mechanism, focus on the important information related to fault prediction in the primary decision vector and the spatio-temporal aggregation feature array, so as to make a more accurate fault mode decision.

[0043] As an implementation manner, for the above step S200, perform a splitting operation on the first fault record data stream to obtain a first to-be-processed fault record data stream and a second to-be-processed fault record data stream, which may specifically include:

[0044] Step S210: Obtain the actual time window length of the first fault record data stream. The actual time window length of the first fault record data stream represents the data stream length of the first fault record data stream, and the actual time window length of the first fault record data stream is the time difference between the initial fault record moment and the latest fault record moment in the data stream;

[0045] Step S220: Multiply the actual time window length of the first fault record data stream by a preset splitting coefficient to obtain a splitting product;

[0046] Step S230: Multiply the splitting product by the initial fault moment in the first fault record data stream to obtain the splitting position corresponding to the first fault record data stream;

[0047] Step S240: According to the splitting position, the fault record data streams on both sides of the splitting position in the first fault record data stream are respectively used as the first fault record data stream to be processed and the second fault record data stream to be processed, where the fault record data corresponding to the splitting position belongs to the first fault record data stream to be processed or belongs to the second fault record data stream to be processed.

[0048] In step S210, the actual time window length represents the data stream length of the first fault record data stream, which is the time difference between the earliest fault record time and the latest fault record time in the data stream. The actual time window length reflects the time range covered by the first fault record data stream and is an important basis for subsequent splitting operations. For example, for the first fault record data stream of an intelligent air conditioner, assume that the earliest fault record time is 10:00 am on January 1, 2024, and the latest fault record time is 3:00 pm on January 10, 2024. Then, by calculating the time difference between these two times, the computer system obtains an actual time window length of 9 days and 5 hours. The computer system can query the timestamps of each fault record in the first fault record data stream, find the earliest and latest timestamps, and then calculate the time difference to obtain the actual time window length.

[0049] In step S220, the actual time window length of the first fault record data stream is multiplied by a preset splitting coefficient to obtain a splitting product. The preset splitting coefficient is a preset value, and its value range is usually between 0 and 1, which is used to determine the splitting ratio of the first fault record data stream. For example, the preset splitting coefficient is set to 0.6. Combining the previous example of the intelligent air conditioner, its actual time window length is 9 days and 5 hours, which is approximately 9×24 + 5 = 221 hours when converted to hours. Then, the splitting product is 221×0.6 = 132.6 hours. The computer system can directly use multiplication operation to implement this step. Let the actual time window length be T and the preset splitting coefficient be k, then the splitting product P = k×T.

[0050] In step S230, the computer system performs a certain operation on the splitting product and the earliest fault time in the first fault record data stream to obtain the splitting position corresponding to the first fault record data stream. Here, for example, the time length corresponding to the splitting product is added to the earliest fault time to determine the splitting position. Continuing with the example of the intelligent air conditioner, the earliest fault time is 10:00 am on January 1, 2024, and the splitting product is 132.6 hours. Converting 132.6 hours into days and hours, 132.6÷24 = 5 days with a remainder of 12.6 hours. Then, the time corresponding to the splitting position is 10:36 pm on January 6, 2024 (0.6 hours is 36 minutes). The date and time calculation function can be used to accumulate the time length corresponding to the splitting product to the earliest fault time to obtain the time corresponding to the splitting position. If the earliest fault time is tmin , if the split product is P, then the time t corresponding to the split position split =t min +P.

[0051] In step S240, based on the split position, the computer system uses the fault record data streams on both sides of the split position in the first fault record data stream as the first to-be-processed fault record data stream and the second to-be-processed fault record data stream respectively. Among them, the fault record data corresponding to the split position belongs to either the first to-be-processed fault record data stream or the second to-be-processed fault record data stream, and the specific attribution can be determined according to actual requirements and rules. Still taking the smart air conditioner as an example, taking 10:36 PM on January 6, 2024 as the split position, the fault record data before this time (including this moment, if it is stipulated that the split position data belongs to the first to-be-processed fault record data stream) constitutes the first to-be-processed fault record data stream, and the fault record data after this time constitutes the second to-be-processed fault record data stream. Each fault record in the first fault record data stream can be traversed, and according to the comparison between its timestamp and the time corresponding to the split position, the fault record is classified into the corresponding to-be-processed fault record data stream.

[0052] Different preset split coefficients will lead to different split results, thereby affecting the accuracy of subsequent fault prediction. If the preset split coefficient is too large, the first to-be-processed fault record data stream contains too much fault record data, which may cause the data volume processed by the spatio-temporal feature fusion component to be too large, increasing the computational burden. At the same time, it may introduce too many early fault record data with little impact on the current fault prediction, reducing the accuracy of fault prediction; if the preset split coefficient is too small, the first to-be-processed fault record data stream contains too little fault record data, which may not provide enough information for the spatio-temporal feature fusion component to perform effective feature aggregation, also affecting the accuracy of fault prediction. The most suitable preset split coefficient for the current target terminal device can be determined through historical experience or multiple experiments.

[0053] When dividing the first to-be-processed fault record data stream and the second to-be-processed fault record data stream, the computer system needs to determine the attribution of the fault record data corresponding to the split position according to the preset rules. If the rule stipulates that the fault record data corresponding to the split position belongs to the first to-be-processed fault record data stream, then the computer system should add this data to the first to-be-processed fault record data stream during division; otherwise, add it to the second to-be-processed fault record data stream. At the same time, the computer system also needs to ensure the data integrity of the two divided to-be-processed fault record data streams to avoid data loss.

[0054] Different types of target terminal devices may have different failure occurrence rules and characteristics. Therefore, when performing the segmentation operation, it is necessary to select an appropriate preset segmentation coefficient according to the type and characteristics of the target terminal device. For some target terminal devices with relatively frequent failures and complex failure rules, a relatively small preset segmentation coefficient may need to be selected to analyze the recent failure record data more carefully; for some target terminal devices with relatively stable failures and simple failure rules, a relatively large preset segmentation coefficient can be selected to reduce the computational amount. The computer system can establish a preset segmentation coefficient database, store the corresponding recommended preset segmentation coefficients according to different types of target terminal devices, and automatically select an appropriate preset segmentation coefficient according to the type of the target terminal device when performing the segmentation operation.

[0055] As an implementation manner, in step S300, the first to-be-processed failure record data stream is loaded into the spatio-temporal feature fusion component in the failure forward reasoning network for multi-dimensional feature aggregation to obtain a spatio-temporal aggregation feature array, including:

[0056] Step S310: According to the target dynamic attenuation mapping matrix in the failure forward reasoning network, index the corresponding temporal attenuation vector of each failure record data in the first to-be-processed failure record data stream to obtain the first target temporal attenuation vector corresponding to the first to-be-processed failure record data stream. The target dynamic attenuation mapping matrix contains the corresponding situations of different failure record data and the target temporal attenuation vector;

[0057] Step S320: Determine the first target distribution positioning vector corresponding to the first to-be-processed failure record data stream according to the temporal distribution situation of each failure record data in the first to-be-processed failure record data stream;

[0058] Step S330: Perform a vector fusion operation on the first target temporal attenuation vector and the first target distribution positioning vector to obtain a spatio-temporal fusion input feature array;

[0059] Step S340: Load the spatio-temporal fusion input feature array into the spatio-temporal feature aggregation layer in the spatio-temporal feature fusion component for multi-dimensional feature aggregation to obtain a spatio-temporal aggregation feature array.

[0060] In step S310, the computer system indexes the time-series decay vector corresponding to each fault record data in the first fault record data stream to be processed based on the target dynamic decay mapping matrix in the fault foresight inference network, and obtains the first target time-series decay vector corresponding to the first fault record data stream to be processed. The target dynamic decay mapping matrix is a matrix that stores the correspondence between different fault record data and the target time-series decay vector. Its purpose is to dynamically adjust the time decay weights of different fault records, enabling the neural network to distinguish the importance of fault types, capture long-term dependencies, and adapt to time sensitivity. The time-series decay vector is a vector with a preset dimension, and each dimension in the vector represents different meanings, such as the importance of different fault types, the impact degree of the distance of the fault occurrence time on the current fault prediction, etc. For example, for the first fault record data stream of an industrial robot, it contains fault records of different types such as motor faults and sensor faults. The target dynamic decay mapping matrix may stipulate that the time-series decay vector corresponding to the motor fault is [0.8, 0.2, 0.1], and the time-series decay vector corresponding to the sensor fault is [0.3, 0.6, 0.1]. The computer system finds the corresponding time-series decay vector for each fault record data in the first fault record data stream to be processed by querying the target dynamic decay mapping matrix, and then combines these vectors to obtain the first target time-series decay vector. If there are n fault record data in the first fault record data stream to be processed, and the time-series decay vector corresponding to each fault record data is , then the first target time-series decay vector . Data structures such as hash tables can be used to store the target dynamic decay mapping matrix, and the corresponding time-series decay vector can be quickly found by using the identifier of the fault record data as the key. The time-series decay vector corresponding to each fault record data can be generated by performing feature learning on historical fault record data and analyzing the time relationship between the occurrence time of each fault record in the historical fault record data and subsequent faults using a neural network. During the training process of the neural network, it is also iteratively adjusted along with the network parameters.

[0061] In step S320, the computer system determines a first target distribution positioning vector corresponding to the first data stream of to-be-processed fault records according to the temporal distribution of each fault record data in the first data stream of to-be-processed fault records. The first target distribution positioning vector is used to reflect the distribution characteristics of the fault record data in the time series, such as information on the frequency of fault occurrence, time interval, etc. Taking an industrial robot as an example, assume that in the first data stream of to-be-processed fault records, the fault occurrence frequency is relatively low in the first half of the time period and relatively high in the second half of the time period. Then the first target distribution positioning vector can reflect this distribution difference. The computer system can determine the first target distribution positioning vector by counting the number of fault record data in each time interval, calculating the time interval between adjacent fault record data, etc. Suppose the first data stream of to-be-processed fault records is divided into m time intervals, and the number of fault record data in each time interval is , and the average time interval between adjacent fault record data is , then the first target distribution positioning vector , where f is a custom function used to convert the fault quantity and time interval information into elements in the vector. The computer system can use the sliding window technique to count the fault information in different time intervals.

[0062] In step S330, the computer system performs a vector fusion operation on the first target temporal attenuation vector and the first target distribution positioning vector to obtain a spatio-temporal fusion input feature array. The purpose of the vector fusion operation is to combine the attenuation information in the time dimension and the positioning information of the fault distribution to form an array containing multi-dimensional features. Feasible vector fusion methods include weighted summation, concatenation, etc. The fusion weights can be adjusted according to the actual situation to highlight the importance of different features.

[0063] In step S340, the computer system loads the spatio-temporal fusion input feature array into the spatio-temporal feature aggregation layer in the spatio-temporal feature fusion component for multi-dimensional feature aggregation to obtain a spatio-temporal aggregation feature array. The spatio-temporal feature aggregation layer is a component in the fault forward reasoning network, and its role is to further process the spatio-temporal fusion input feature array, mine the potential feature information therein, and complete data encoding. The spatio-temporal feature aggregation layer can adopt deep learning models such as convolutional neural network (CNN) and recurrent neural network (RNN). Taking CNN as an example, the convolutional layer in CNN can extract local features in the spatio-temporal fusion input feature array, and the pooling layer can perform dimensionality reduction processing on the features to reduce the amount of calculation. Through multiple convolution and pooling operations, the spatio-temporal feature aggregation layer can obtain a more advanced feature representation, that is, the spatio-temporal aggregation feature array. Assuming that the dimension of the spatio-temporal fusion input feature array is d, after being processed by the spatio-temporal feature aggregation layer, the dimension of the spatio-temporal aggregation feature array may become d'. The computer system can optimize the effect of feature aggregation by adjusting the parameters of the spatio-temporal feature aggregation layer, such as the size of the convolution kernel, the number of convolution kernels, the type of pooling layer, etc.

[0064] For the target dynamic attenuation mapping matrix, different types of target terminal devices may have different fault characteristics and time dependencies. Therefore, it is necessary to design a suitable target dynamic attenuation mapping matrix for different target terminal devices. For some devices that are more sensitive to time, such as communication base stations, the faults that occurred recently may have a greater impact on the current fault prediction. Then, the weight of the temporal attenuation vector corresponding to the recent faults in the target dynamic attenuation mapping matrix should be larger; for some devices with relatively stable fault occurrences, such as ordinary office computers, the long-term fault record data may all have certain reference value. Then, the difference in the weights of the temporal attenuation vectors corresponding to the faults at different times in the target dynamic attenuation mapping matrix can be relatively small. The computer system can optimize the target dynamic attenuation mapping matrix through a large number of experiments and data analysis to make it better reflect the actual fault situation of the target terminal device.

[0065] As an implementation manner, the fault decision component includes a decision feature extraction layer and a cross-attention decision layer. In step S400, the second to-be-processed fault record data stream and the spatio-temporal aggregation feature array are loaded into the fault decision component in the fault forward reasoning network for fault mode decision to obtain the target fault prediction result corresponding to the target terminal device at the forward-looking time point, including:

[0066] Step S410: According to the target dynamic attenuation mapping matrix in the fault forward reasoning network, index the temporal attenuation vector corresponding to each fault record data in the second to-be-processed fault record data stream to obtain the second target temporal attenuation vector corresponding to the second to-be-processed fault record data stream. The target dynamic attenuation mapping matrix contains the corresponding situations of different fault record data and the target temporal attenuation vector;

[0067] Step S420: Determine the second target distribution positioning vector corresponding to the second to-be-processed fault record data stream according to the temporal distribution of each fault record data in the second to-be-processed fault record data stream.

[0068] Step S430: Perform a vector fusion operation on the second target temporal attenuation vector and the second target distribution positioning vector to obtain a decision input fusion vector.

[0069] Step S440: Load the decision input fusion vector into the decision feature extraction layer for decision feature extraction to obtain a primary decision vector.

[0070] Step S450: Load the primary decision vector and the spatio-temporal aggregation feature array into the cross-attention decision layer for fault decision processing to obtain the target fault prediction result.

[0071] In step S410, the computer system indexes the temporal attenuation vector corresponding to each fault record data in the second to-be-processed fault record data stream according to the target dynamic attenuation mapping matrix in the fault forward reasoning network, and obtains the second target temporal attenuation vector corresponding to the second to-be-processed fault record data stream. The target dynamic attenuation mapping matrix stores the corresponding relationship between different fault record data and the target temporal attenuation vector. Its function is to dynamically adjust the time attenuation weights of different fault records, so that the neural network can distinguish the importance of fault types, capture long-term dependencies, and adapt to time sensitivity. The temporal attenuation vector is a vector of a preset dimension, and each dimension of the vector can represent meanings such as the attenuation degree of different fault characteristics in the time dimension. For example, for a production device in an intelligent factory, its second to-be-processed fault record data stream contains different types of fault records such as electrical faults and mechanical faults. The target dynamic attenuation mapping matrix may stipulate that the temporal attenuation vector corresponding to the electrical fault is [0.7, 0.2, 0.1], and the temporal attenuation vector corresponding to the mechanical fault is [0.4, 0.5, 0.1]. The computer system queries this matrix to match the corresponding temporal attenuation vector for each fault record data in the second to-be-processed fault record data stream, and then combines them to obtain the second target temporal attenuation vector. If there are n fault record data in the second to-be-processed fault record data stream, and the temporal attenuation vector corresponding to each fault record data is , then the second target temporal attenuation vector . The computer system can use data structures such as hash tables to store the target dynamic attenuation mapping matrix, and use the identifier of the fault record data as the key to achieve fast search for the corresponding temporal attenuation vector.

[0072] In step S420, based on the temporal distribution of each fault record data in the second data stream of fault records to be processed in the computer system, the computer system determines the second target distribution positioning vector corresponding to the second data stream of fault records to be processed. The second target distribution positioning vector is used to characterize the distribution characteristics of fault record data in the time series, such as the frequency of fault occurrence, the regularity of time intervals, etc. Taking the production equipment in an intelligent factory as an example, if in the second data stream of fault records to be processed, faults occur concentratedly in certain time periods and rarely occur in other time periods, then the second target distribution positioning vector can reflect this distribution feature. The computer system can determine this vector by counting the number of fault record data in different time intervals, calculating the time interval between adjacent fault record data, etc. Suppose the second data stream of fault records to be processed is divided into m time intervals, and the number of fault record data in each time interval is , and the average time interval between adjacent fault record data is t j , then the second target distribution positioning vector , where f is a user-defined function for converting the fault quantity and time interval information into elements in the vector. The computer system can adopt the sliding window technique to count the fault information in different time intervals.

[0073] In step S430, the computer system performs a vector fusion operation on the second target temporal attenuation vector and the second target distribution positioning vector to obtain a decision input fusion vector. The purpose of the vector fusion operation is to integrate the attenuation information in the time dimension and the positioning information of the fault distribution to form a vector containing multi-dimensional features, providing more comprehensive information for subsequent fault decision-making. Feasible vector fusion methods include weighted summation, concatenation, etc. For example, the weighted summation method is adopted.

[0074] In step S440, the computer system loads the decision input fusion vector into the decision feature extraction layer for decision feature extraction to obtain a primary decision vector. The decision feature extraction layer is part of the fault decision component, and its role is to extract key features related to fault decision from the decision input fusion vector. The decision feature extraction layer can adopt deep learning models such as convolutional neural network (CNN) and multi-layer perceptron (MLP). Taking CNN as an example, the convolutional layer in CNN can extract local features in the decision input fusion vector. Through the sliding operation of the convolutional kernel, it captures feature information at different positions; the pooling layer can perform dimensionality reduction on the features to reduce the computational amount. After multiple convolutional and pooling operations, the decision feature extraction layer can obtain a more advanced feature representation, that is, the primary decision vector. Assuming the dimension of the decision input fusion vector is d, after being processed by the decision feature extraction layer, the dimension of the primary decision vector may become d'. The computer system can optimize the effect of feature extraction by adjusting the parameters of the decision feature extraction layer, such as the size of the convolutional kernel, the number of convolutional kernels, the type of pooling layer, etc.

[0075] In step S450, the computer system loads the primary decision vector and the spatio-temporal aggregation feature array into the cross-attention decision layer for fault decision processing to obtain the target fault prediction result. The cross-attention decision layer uses the attention mechanism to focus on the important information related to fault prediction in the primary decision vector and the spatio-temporal aggregation feature array, so as to make a more accurate fault mode decision. The attention mechanism can assign different weights to different feature elements, highlighting important feature information and suppressing irrelevant feature information. For example, when predicting the faults of production equipment in an intelligent factory, the cross-attention decision layer may pay more attention to the fault features that occurred recently and the fault features related to the core components of the equipment. By calculating the attention weights, the primary decision vector and the spatio-temporal aggregation feature array are weighted and combined to obtain the final decision result, that is, the target fault prediction result. Let the primary decision vector be , the spatio-temporal aggregation feature array be , the attention weight matrix be , then the target fault prediction result (where represents concatenating the primary decision vector and the spatio-temporal aggregation feature array). The attention weight matrix can be learned through training so that the cross-attention decision layer can adaptively adjust the attention distribution according to different input data.

[0076] As an implementation manner, the method provided by the present invention further includes the training process of the fault forward reasoning network, specifically including:

[0077] Step S10: Obtain an initialized fault forward reasoning network;

[0078] Step S20: The historical fault record data stream corresponding to the target terminal device, where the historical fault record data stream represents the fault record data of the target terminal device during the past operation and maintenance monitoring cycles, and the past operation and maintenance monitoring cycles include the first operation and maintenance monitoring cycle;

[0079] Step S30: Perform a splitting operation on the historical fault record data stream to obtain a first fault record data stream and a second fault record data stream;

[0080] Step S40: Iteratively adjust the parameters of the initialized fault forward reasoning network according to the second fault record data stream to obtain a converged fault forward reasoning network.

[0081] In step S10, the initialized fault forward reasoning network is a pre-constructed neural network model that needs to be trained, which includes a preset network structure and initial parameters. This network structure consists of a preset spatio-temporal feature fusion component and a preset fault decision-making component. The preset spatio-temporal feature fusion component is used to perform multi-dimensional feature aggregation on the input fault record data, and the preset fault decision-making component is used to make fault mode decisions based on the aggregated features. For example, for a fault forward reasoning network used to predict the faults of an industrial robot, during initialization, the preset spatio-temporal feature fusion component may adopt a structure that combines a convolutional neural network (CNN) and a recurrent neural network (RNN), and the preset fault decision-making component may adopt a structure of a multi-layer perceptron (MLP). The computer system can use deep learning frameworks such as TensorFlow and PyTorch to build and initialize this network. In these frameworks, the computer system can define parameters such as the number of layers of the network, the number of neurons in each layer, and the activation function, so as to obtain the initialized fault forward reasoning network.

[0082] In step S20, the computer system obtains the historical fault record data stream corresponding to the target terminal device. The historical fault record data stream represents the fault record data of the target terminal device during the past operation and maintenance monitoring cycles, and the past operation and maintenance monitoring cycles include the first operation and maintenance monitoring cycle. The historical fault record data stream is an important data source for training the fault forward reasoning network, which contains various fault information of the target terminal device at different time points. Taking an intelligent air conditioner as an example, the historical fault record data stream may include information such as the fault occurrence time, fault type (such as refrigeration fault, fan fault, etc.), and fault severity of the air conditioner in the past year. The computer system can establish a communication connection with the target terminal device to read the fault record data from the local storage of the device; or receive the fault record data uploaded by the device in real time through the network interface, and organize these data into a historical fault record data stream in chronological order.

[0083] In step S30, the computer system performs a splitting operation on the historical fault record data stream to obtain a first fault record data stream and a second fault record data stream. This splitting operation is similar to the splitting operation in step S200, and its purpose is to divide the historical fault record data stream into two parts, one part for training the preset spatio-temporal feature fusion component and the other part for training the preset fault decision-making component. The computer system can split the historical fault record data stream into two parts in chronological order according to a preset splitting coefficient. For example, if the splitting coefficient is set to 0.7, the computer system takes the first 70% of the fault record data in the historical fault record data stream as the first fault record data stream, and the last 30% of the fault record data as the second fault record data stream. The specific splitting method can be to first calculate the actual time window length of the historical fault record data stream, then multiply the actual time window length by the preset splitting coefficient to obtain the splitting position, and finally split the data stream into two parts according to the splitting position. Let the actual time window length of the historical fault record data stream be T, and the preset splitting coefficient be k, then the time corresponding to the splitting position is t split = k×T, and the historical fault record data stream is split into a first fault record data stream and a second fault record data stream according to this time.

[0084] In step S40, the computer system iteratively adjusts the parameters of the initialized fault forward reasoning network based on the second fault record data stream to obtain a converged fault forward reasoning network. This is the core step of the training process. By continuously adjusting the parameters of the network, the network can better fit the fault patterns in the second fault record data stream, thereby improving the accuracy of fault prediction. The computer system first determines one or more fault record training data stream sets and the corresponding prior fault labels based on the second fault record data stream. The fault record training data stream sets include a first fault record training data stream and a second fault record training data stream, and the prior fault label refers to the known fault type or the true situation of the fault occurrence. For example, for the second fault record data stream of an intelligent air conditioner, the computer system can divide it according to a certain time window to obtain multiple fault record training data stream sets. The first fault record training data stream in each set is used to input the preset spatio-temporal feature fusion component, and the second fault record training data stream is used to be input into the preset fault decision-making component together with the training spatio-temporal aggregation feature array output by the preset spatio-temporal feature fusion component. At the same time, according to the historical maintenance records or other reliable information, the corresponding prior fault label is determined for each fault record training data stream set.

[0085] Next, the computer system loads the first fault record training data stream into a preset spatio-temporal feature fusion component for multi-dimensional feature aggregation to obtain a training spatio-temporal aggregation feature array. This process is similar to the operation in step S300. The computer system indexes the temporal decay vector corresponding to each fault record data in the first fault record training data stream according to the preset dynamic decay mapping matrix in the initialized fault forward reasoning network to obtain a training temporal decay vector; determines a training distribution positioning vector according to the temporal distribution of each fault record data in the first fault record training data stream; performs a vector fusion operation on the training temporal decay vector and the training distribution positioning vector to obtain a training spatio-temporal fusion input feature array; finally, loads the training spatio-temporal fusion input feature array into the spatio-temporal feature aggregation layer of the preset spatio-temporal feature fusion component for multi-dimensional feature aggregation to obtain a training spatio-temporal aggregation feature array.

[0086] Then, the computer system loads the second fault record training data stream and the training spatio-temporal aggregation feature array into a preset fault decision component for fault mode decision to obtain a deduced fault decision result. The preset fault decision component will calculate through the internal neural network structure based on the input data and output a prediction result of the fault type or the probability of fault occurrence, that is, the deduced fault decision result.

[0087] After that, the computer system obtains a fault deduction error based on the deduced fault decision result and the prior fault label. The fault deduction error is used to measure the difference between the deduced fault decision result and the actual fault situation. Feasible error calculation methods include mean square error (MSE), cross-entropy loss, etc. Taking the mean square error as an example, let the deduced fault decision result be , and the prior fault label be y, then the fault deduction error , where n is the number of samples.

[0088] Finally, the computer system iteratively adjusts the parameters of the initialized fault forward reasoning network according to the fault deduction error. A feasible parameter adjustment method is to use the backpropagation algorithm. By calculating the gradient of the fault deduction error with respect to the network parameters, and then updating the network parameters according to the direction and magnitude of the gradient. The computer system continuously repeats the above process of determining the fault record training data stream set and the prior fault label, performing feature aggregation, fault mode decision, calculating the error, and adjusting the parameters until the preset iteration stop requirements are met, such as the fault deduction error is less than a certain threshold, reaching the maximum number of iterations, etc. At this time, a converged fault forward reasoning network is obtained.

[0089] In practical applications, the selection of the structure and parameters for initializing the fault anticipation inference network will affect the training effect and efficiency. Different network structures are suitable for different types of target terminal devices and fault record data. For example, for target terminal devices with obvious time series characteristics in the fault record data, using RNN or its variants (such as LSTM, GRU) as the preset spatio-temporal feature fusion component may achieve better results; for target terminal devices with spatial characteristics in the fault record data, CNN may be more suitable. The computer system can compare the training effects under different network structures and parameter combinations through experiments and select the optimal initialization configuration.

[0090] The computer system also needs to monitor and evaluate the training process. The validation set can be used to evaluate the performance of the fault anticipation inference network, calculate indicators such as fault deduction error, accuracy, and recall rate on the validation set, and observe the changes in these indicators during the training process. If it is found that the performance on the validation set deteriorates, it may indicate that the network has overfitting problems. The computer system can adopt some regularization methods, such as L1 and L2 regularization, Dropout, etc., to alleviate the overfitting problem.

[0091] As an implementation method, the initialization of the fault anticipation inference network includes a preset spatio-temporal feature fusion component and a preset fault decision component. Step S40: Iteratively adjust the parameters of the initialization fault anticipation inference network according to the second fault record data stream to obtain a converged fault anticipation inference network, including:

[0092] Step S41: Determine one or more fault record training data stream sets and the prior fault labels corresponding to one or more fault record training data stream sets according to the second fault record data stream. Each fault record training data stream set includes a first fault record training data stream and a second fault record training data stream;

[0093] Step S42: Load the first fault record training data stream into the preset spatio-temporal feature fusion component for multi-dimensional feature aggregation to obtain a training spatio-temporal aggregation feature array;

[0094] Step S43: Load the second fault record training data stream and the training spatio-temporal aggregation feature array into the preset fault decision component for fault mode decision to obtain a deduced fault decision result;

[0095] Step S44: Obtain a fault deduction error according to the deduced fault decision result and the prior fault label;

[0096] Step S45: Iteratively adjust the parameters of the initialization fault anticipation inference network according to the fault deduction error to obtain a fault anticipation inference network.

[0097] In the implementation of step S40, steps S41 - S45 detail the specific process by which the computer system iteratively adjusts the parameters of the initialized fault forward reasoning network based on the second fault record data stream, thereby obtaining the converged fault forward reasoning network. This series of steps is the core of the fault forward reasoning network training process. By continuously optimizing the network parameters, the network can more accurately predict the faults of the target terminal device.

[0098] In step S41, the computer system determines one or more fault record training data stream sets and the corresponding prior fault labels based on the second fault record data stream. Each fault record training data stream set includes a first fault record training data stream and a second fault record training data stream. The prior fault label is known real fault information used to measure the prediction accuracy of the network. For example, for an automated device on an industrial production line, its second fault record data stream contains the fault occurrence situations in the past period. The computer system can divide the second fault record data stream according to a certain time window to obtain multiple fault record training data stream sets. Suppose the time window is set to one week, then the fault record data of each week constitutes a fault record training data stream set. In each set, the data is further divided into a first fault record training data stream and a second fault record training data stream according to a certain ratio. For example, the first 70% of the data can be used as the first fault record training data stream, and the last 30% of the data can be used as the second fault record training data stream. At the same time, according to the device's maintenance records or the judgment of professionals, the corresponding prior fault labels, such as motor faults, sensor faults, etc., are determined for each fault record training data stream set. The computer system can use the sliding window technique to implement the division of the second fault record data stream. By setting appropriate window sizes and sliding steps, multiple non - overlapping or partially overlapping fault record training data stream sets can be obtained.

[0099] In step S42, the computer system loads the first fault record training data stream into a preset spatio-temporal feature fusion component for multi-dimensional feature aggregation to obtain a training spatio-temporal aggregation feature array. The preset spatio-temporal feature fusion component is an important part of the initialized fault forward reasoning network, and its role is to extract multi-dimensional spatio-temporal feature information from the first fault record training data stream. The computer system first indexes the time series attenuation vector corresponding to each fault record data in the first fault record training data stream according to the preset dynamic attenuation mapping matrix in the initialized fault forward reasoning network to obtain a training time series attenuation vector. The preset dynamic attenuation mapping matrix stores the corresponding relationship between different fault record data and the preset time series attenuation vector. The time series attenuation vector is used to dynamically adjust the time attenuation weight of different fault records to reflect the influence degree of faults at different times on the current fault prediction. For example, for the above industrial automation equipment, faults that occurred recently may be more important for the current fault prediction, so the weight of the corresponding time series attenuation vector will be relatively large. Then, the computer system determines a training distribution positioning vector according to the time series distribution of each fault record data in the first fault record training data stream, and this vector reflects the distribution characteristics of the fault record data in the time series. Then, a vector fusion operation is performed on the training time series attenuation vector and the training distribution positioning vector to obtain a training spatio-temporal fusion input feature array. Feasible fusion methods include weighted summation, etc. Finally, the training spatio-temporal fusion input feature array is loaded into the spatio-temporal feature aggregation layer of the preset spatio-temporal feature fusion component for multi-dimensional feature aggregation. The spatio-temporal feature aggregation layer can adopt deep learning models such as convolutional neural network (CNN) and recurrent neural network (RNN). After processing, a training spatio-temporal aggregation feature array is obtained.

[0100] In step S43, the computer system loads the second fault record training data stream and the training spatio-temporal aggregation feature array into a preset fault decision component for fault mode decision to obtain a deduced fault decision result. The preset fault decision component is another important part of the initialized fault forward reasoning network, and its task is to judge the fault mode of the target terminal device according to the input second fault record training data stream and the training spatio-temporal aggregation feature array. The second fault record training data stream contains fault record information in the subsequent time period, and the training spatio-temporal aggregation feature array contains multi-dimensional spatio-temporal feature information extracted from the first fault record training data stream. The preset fault decision component will comprehensively consider this information, perform calculations and judgments through the internal neural network structure, and output the prediction result of the fault type or the probability of fault occurrence, that is, the deduced fault decision result. For example, for industrial automation equipment, the preset fault decision component may judge according to the input data that the probability of the equipment having a motor fault in the next period of time is 80%, and the probability of having a sensor fault is 20%.

[0101] In step S44, the computer system obtains a fault deduction error based on the deduced fault decision result and the prior fault mark. The fault deduction error is used to measure the difference between the deduced fault decision result and the actual fault situation, and is an important basis for evaluating network performance and adjusting parameters. Feasible error calculation methods include mean square error (MSE), cross-entropy loss, etc. Taking the mean square error as an example, assume that the deduced fault decision result is , the prior fault mark is y, and the number of samples is n. Then the fault deduction error . In the example of industrial automation equipment, if the prior fault mark indicates that the equipment actually has a motor fault, and the deduced fault decision result predicts that the probability of the motor fault is 80% and the probability of other faults is 20%, then the computer system can calculate the current fault deduction error according to the mean square error formula.

[0102] In step S45, the computer system iteratively adjusts the parameters of the initialized fault forward reasoning network based on the fault deduction error to obtain the fault forward reasoning network. This is a process of continuously optimizing network parameters, aiming to make the output result of the network as close as possible to the actual fault situation, that is, to reduce the fault deduction error. The computer system first iteratively adjusts the network parameters in the initialized fault forward reasoning network and the preset time-series attenuation vector in the preset dynamic attenuation mapping matrix according to the fault deduction error. A feasible parameter adjustment method is to use the backpropagation algorithm, which calculates the gradient of the fault deduction error with respect to the network parameters and then updates the network parameters according to the direction and magnitude of the gradient. At the same time, the preset time-series attenuation vector in the preset dynamic attenuation mapping matrix is also adjusted to better reflect the time attenuation weights of different fault records. Then, the computer system iteratively executes steps S41 - S45 based on the iterated initialized fault forward reasoning network, that is, determines the fault record training data stream set and the prior fault mark again, performs feature aggregation, fault mode decision, error calculation and parameter adjustment until the preset iteration stop requirements are met, such as the fault deduction error is less than a certain threshold, the maximum number of iterations is reached, etc. At this time, the converged fault forward reasoning network is obtained.

[0103] When the computer system executes steps S41 - S45, it comprehensively considers various factors, including the design of the preset dynamic attenuation mapping matrix, the selection of the time window, the method and weight selection of the vector fusion operation, the models and parameter settings of the preset spatio-temporal feature fusion component and the preset fault decision-making component, the adjustment of the learning rate, the monitoring and evaluation of the training process, and the recording of operation information. Through reasonable operations and optimizations, a converged fault foresight inference network is obtained, providing accurate and reliable support for the fault prediction of the target terminal device, improving the reliability and availability of the device, and reducing the maintenance cost. At the same time, the computer system can also continuously learn and optimize the entire training process to adapt to different types of target terminal devices and complex application scenarios, further improving the effect of fault prediction.

[0104] As an implementation, in step S42, the first fault record training data stream is loaded into the preset spatio-temporal feature fusion component for multi-dimensional feature aggregation to obtain a training spatio-temporal aggregation feature array, including:

[0105] Step S421: According to the preset dynamic attenuation mapping matrix in the initialized fault foresight inference network, index the temporal attenuation vectors corresponding to each fault record data in the first fault record training data stream to obtain the training temporal attenuation vector corresponding to the first fault record training data stream. The preset dynamic attenuation mapping matrix contains the corresponding situations of different fault record data and the preset temporal attenuation vectors;

[0106] Step S422: Determine the training distribution positioning vector corresponding to the first fault record training data stream according to the temporal distribution situation of each fault record data in the first fault record training data stream;

[0107] Step S423: Perform a vector fusion operation on the training temporal attenuation vector and the training distribution positioning vector to obtain a training spatio-temporal fusion input feature array;

[0108] Step S424: Load the training spatio-temporal fusion input feature array into the spatio-temporal feature aggregation layer of the preset spatio-temporal feature fusion component for multi-dimensional feature aggregation to obtain a training spatio-temporal aggregation feature array.

[0109] In step S421, the computer system indexes the time-series decay vector corresponding to each fault record data in the first fault record training data stream according to the preset dynamic decay mapping matrix in the initialized fault forward reasoning network, and obtains the training time-series decay vector corresponding to the first fault record training data stream. The preset dynamic decay mapping matrix is an important part of the initialized fault forward reasoning network, which stores the corresponding relationship between different fault record data and the preset time-series decay vector. Its function is to dynamically adjust the time decay weights of different fault records, enabling the neural network to distinguish the importance of fault types, capture long-term dependencies, and adapt to time sensitivity. The time-series decay vector is a vector with a preset dimension, and each dimension in the vector can represent different meanings, such as the importance of different fault types, the influence degree of the distance of the fault occurrence time on the current fault prediction, etc.

[0110] Taking the transformer in the smart grid as an example, the first fault record training data stream may contain fault records such as overheating of the oil temperature, overheating of the winding, and insulation aging of the transformer over a period of time. The preset dynamic decay mapping matrix may stipulate that the preset time-series decay vector corresponding to the overheating of the oil temperature fault is [0.8, 0.1, 0.1], the preset time-series decay vector corresponding to the overheating of the winding fault is [0.6, 0.3, 0.1], and the preset time-series decay vector corresponding to the insulation aging fault is [0.3, 0.5, 0.2]. The computer system finds the corresponding time-series decay vector for each fault record data in the first fault record training data stream by querying the preset dynamic decay mapping matrix.

[0111] In step S422, the computer system determines the training distribution positioning vector corresponding to the first fault record training data stream according to the time-series distribution of each fault record data in the first fault record training data stream. The training distribution positioning vector is used to reflect the distribution characteristics of the fault record data in the time series, such as the frequency of fault occurrence, time interval, etc. Continuing with the example of the transformer in the smart grid, assume that in the first fault record training data stream, the frequency of fault occurrence is relatively low in the first half of the time period and relatively high in the second half of the time period, then the training distribution positioning vector can reflect this distribution difference.

[0112] The computer system can determine the training distribution positioning vector by counting the number of fault record data in each time interval, calculating the time interval between adjacent fault record data, etc. Suppose the first fault record training data stream is divided into m time intervals, and the number of fault record data in each time interval is , and the average time interval between adjacent fault record data is t j , then the training distribution positioning vector , where f is a custom function used to convert the fault quantity and time interval information into elements in a vector. The sliding window technology can be used to count the fault information in different time intervals.

[0113] In step S423, the computer system performs a vector fusion operation on the training time series attenuation vector and the training distribution location vector to obtain a training spatiotemporal fusion input feature array. The purpose of the vector fusion operation is to combine the attenuation information in the time dimension and the location information of the fault distribution to form an array containing multi-dimensional features, providing more comprehensive information for subsequent feature aggregation. Feasible vector fusion methods include weighted summation and splicing. The fusion weight can be adjusted according to the actual situation to highlight the importance of different features.

[0114] In step S424, the computer system loads the training spatiotemporal fusion input feature array into the spatiotemporal feature aggregation layer of the preset spatiotemporal feature fusion component for multi-dimensional feature aggregation to obtain the training spatiotemporal aggregation feature array. The spatiotemporal feature aggregation layer is the core part of the preset spatiotemporal feature fusion component, and its function is to further process the training spatiotemporal fusion input feature array and mine the potential feature information therein. The spatiotemporal feature aggregation layer can adopt deep learning models such as convolutional neural network (CNN) and recurrent neural network (RNN).

[0115] Taking CNN as an example, the convolution layer in CNN can extract local features from the training spatiotemporal fusion input feature array, and capture feature information at different positions through the sliding operation of the convolution kernel; the pooling layer can reduce the dimension of the features to reduce the amount of calculation. After multiple convolution and pooling operations, the spatiotemporal feature aggregation layer can obtain a more advanced feature representation, that is, the training spatiotemporal aggregation feature array. Assuming that the dimension of the training spatiotemporal fusion input feature array is d, after being processed by the spatiotemporal feature aggregation layer, the dimension of the training spatiotemporal aggregation feature array may become d'. The computer system can optimize the effect of feature aggregation by adjusting the parameters of the spatiotemporal feature aggregation layer, such as the size of the convolution kernel, the number of convolution kernels, the type of pooling layer, etc.

[0116] In practical applications, the design of the preset dynamic attenuation mapping matrix needs to fully consider the characteristics of the target terminal device. For different types of devices, the patterns and time dependencies of fault occurrences may vary greatly. For some devices with high real-time requirements, such as avionics equipment, the impact of recent faults on future fault prediction may be greater. Therefore, the weight of the preset time-series attenuation vector corresponding to recent faults in the preset dynamic attenuation mapping matrix should be relatively large. For some relatively stable devices, such as large generator sets, long-term fault record data may all have certain reference value. At this time, the difference in the weights of the preset time-series attenuation vectors corresponding to faults at different times can be relatively small. The computer system can optimize the preset dynamic attenuation mapping matrix through a large number of experiments and data analyses to enable it to better reflect the actual fault conditions of the target terminal device.

[0117] As an implementation manner, step S45, iteratively adjusting the parameters of the initialized fault forward reasoning network according to the fault deduction error to obtain the fault forward reasoning network, includes:

[0118] Step S451: According to the fault deduction error, iteratively adjust the network parameters in the initialized fault forward reasoning network and the preset time-series attenuation vector in the preset dynamic attenuation mapping matrix to obtain the iterated initialized fault forward reasoning network;

[0119] Step S452: According to the iterated initialized fault forward reasoning network, iteratively execute determining one or more fault record training data stream sets and the prior fault marks corresponding to one or more fault record training data stream sets based on the second fault record data stream to according to the fault deduction error, iteratively adjusting the network parameters in the initialized fault forward reasoning network and the preset time-series attenuation vector in the preset dynamic attenuation mapping matrix to obtain the iterated initialized fault forward reasoning network until the preset iteration stop requirement is met to obtain the fault forward reasoning network.

[0120] In step S451, the computer system iteratively adjusts the network parameters in the initialized fault forward reasoning network and the preset time-series attenuation vector in the preset dynamic attenuation mapping matrix according to the fault deduction error to obtain the iterated initialized fault forward reasoning network. The fault deduction error is calculated in step S44 based on the deduced fault decision result and the prior fault mark, and it measures the difference between the current prediction result of the network and the actual fault situation. The network parameters are the connection weights and bias values between the neurons in the initialized fault forward reasoning network, and the preset time-series attenuation vector in the preset dynamic attenuation mapping matrix is used to adjust the time attenuation weights of different fault records.

[0121] Taking an initialized fault forward reasoning network for predicting industrial robot faults as an example, assume that the fault deduction error shows that the current network has a low prediction accuracy for robot motor faults. The computer system will use the backpropagation algorithm to calculate the gradients of the fault deduction error with respect to the network parameters and the preset time decay vector. The gradient represents the rate of change of the error function in the parameter space. By updating the parameters in the opposite direction of the gradient, the error can be reduced. For example, for a connection weight w in the network, the update formula can be expressed as , where is the learning rate,[[]]END]] is the partial derivative of the error E with respect to the weight w, and w new is the updated connection weight, and w old is the connection weight before update. For the preset time decay vector in the preset dynamic decay mapping matrix, a similar method will also be used for update to adjust the time decay weights of different fault records, so that the network can better capture the time dependence of faults.

[0122] In actual operation, the computer system can use optimizers provided by deep learning frameworks, such as Stochastic Gradient Descent (SGD), Adagrad, Adadelta, Adam, etc., to automatically complete the parameter update. These optimizers will adjust the learning rate according to different algorithm rules to improve the efficiency and stability of training. For example, the Adam optimizer combines the ideas of momentum and adaptive learning rate, and can adaptively adjust the learning rate of each parameter during training, making the parameter update more stable and efficient.

[0123] In step S452, the computer system iteratively executes the process from determining one or more fault record training data stream sets and their corresponding prior fault labels based on the second fault record data stream to iteratively adjusting the network parameters in the initialized fault forward reasoning network and the preset time decay vector in the preset dynamic decay mapping matrix according to the fault deduction error until the preset iteration stop requirement is met, and a fault forward reasoning network is obtained. The preset iteration stop requirement can be that the fault deduction error is less than a certain threshold, reaching the maximum number of iterations, etc.

[0124] Continuing with the example of industrial robot fault prediction, the computer system will again divide the second fault record data stream into multiple fault record training data stream sets according to a certain time window, and determine the corresponding prior fault labels for each set. Then, load the first fault record training data stream in each set into the preset spatio-temporal feature fusion component of the iterated initialized fault forward reasoning network for multi-dimensional feature aggregation to obtain a new training spatio-temporal aggregation feature array; load the second fault record training data stream and the new training spatio-temporal aggregation feature array into the preset fault decision-making component for fault mode decision-making to obtain a new deduced fault decision result; calculate the new fault deduction error according to the new deduced fault decision result and the prior fault label; finally, iteratively adjust the network parameters and the preset time series attenuation vector again according to the new fault deduction error.

[0125] During this iterative process, the computer system needs to continuously monitor the change of the fault deduction error. If the fault deduction error no longer decreases significantly after multiple iterations, or begins to show an upward trend, it may indicate that the network has an overfitting phenomenon. Overfitting refers to the situation where the network performs well on the training data but poorly on the unseen data. To avoid overfitting, the computer system can adopt some regularization methods, such as L1 and L2 regularization, Dropout, etc. L1 and L2 regularization limit the size of the network parameters by adding regularization terms to the error function to prevent overfitting caused by too large parameters; Dropout randomly ignores some neurons during the training process to make the network more robust.

[0126] During the entire iterative process, the computer system records the relevant information of each iteration, such as the fault deduction error, the update of the network parameters, the change of the preset dynamic attenuation mapping matrix, etc. These information can be used to analyze the training process and performance of the network to help the computer system adjust the training strategy. For example, if it is found that a certain preset time series attenuation vector changes little after multiple iterations, it may indicate that the initial setting of this vector is relatively reasonable, or the impact of this fault record on fault prediction is relatively small, and it can be further analyzed whether it is necessary to adjust the structure of the preset dynamic attenuation mapping matrix.

[0127] When the preset iteration stop requirement is met, what the computer system obtains is the converged fault forward reasoning network. This network has undergone multiple iterative trainings and can better fit the fault patterns in the second fault record data stream, thereby improving the prediction accuracy of the fault situation of the target terminal device at the forward-looking time point. For example, the trained industrial robot fault prediction network can more accurately predict the occurrence probability of robot motor faults, joint faults, etc., providing strong support for equipment maintenance and management.

[0128] As an implementation method, the method further includes:

[0129] Step S1: Obtain a preset maximum time window length and the actual time window length of the second fault record data stream;

[0130] In step S452, based on the second fault record data stream, determine one or more fault record training data stream sets and the prior fault markers corresponding to one or more fault record training data stream sets, including:

[0131] Step S4521: According to the actual time window length and the preset maximum time window length of the second fault record data stream, perform a splitting operation on the second fault record data stream to obtain one or more third fault record data streams;

[0132] Step S4522: Generate the prior fault marker corresponding to each third fault record data stream according to the target fault record data in each third fault record data stream, where the target fault record data is the fault record data with the latest recording time corresponding to each third fault record data stream;

[0133] Step S4523: Extract the fault record data stream other than the target fault record data in each third fault record data stream to obtain the fourth fault record data stream corresponding to each third fault record data stream;

[0134] Step S4524: Determine the splitting position corresponding to each fourth fault record data stream according to the actual time window length of each fourth fault record data stream;

[0135] Step S4525: Perform a splitting operation on each fourth fault record data stream according to the splitting position to obtain one or more fault record training data stream sets.

[0136] In step S1, the computer system obtains the preset maximum time window length and the actual time window length of the second fault record data stream. The preset maximum time window length is a preset time length standard for subsequent segmentation operations on the second fault record data stream. The actual time window length of the second fault record data stream refers to the time difference between the earliest fault moment and the latest fault moment in the data stream, which reflects the actual time range covered by the second fault record data stream. For example, for the second fault record data stream of an intelligent air conditioner, the earliest fault record time is 10:00 am on January 1, 2024, and the latest fault record time is 3:00 pm on January 15, 2024. Then its actual time window length is 14 days and 5 hours. The computer system can obtain the actual time window length by querying the timestamps of each fault record in the second fault record data stream, finding the earliest and latest timestamps, and then calculating the time difference. The preset maximum time window length can be determined based on experience or analysis of the fault pattern of the target terminal device. For example, for this intelligent air conditioner, based on past fault data and maintenance experience, the preset maximum time window length is set to 7 days.

[0137] In step S4521 of the implementation manner of step S452, the computer system segments the second fault record data stream based on the actual time window length and the preset maximum time window length of the second fault record data stream, obtaining one or more third fault record data streams. The purpose of this step is to reasonably divide the longer second fault record data stream according to the preset maximum time window length for more detailed analysis and processing in the future. Continuing with the example of the intelligent air conditioner, the actual time window length of the second fault record data stream is 14 days and 5 hours, and the preset maximum time window length is 7 days. Then the computer system will segment the second fault record data stream into two third fault record data streams. The first one covers the fault records from 10:00 am on January 1, 2024 to around 10:00 am on January 8, 2024, and the second one covers the fault records from 10:00 am on January 8, 2024 to 3:00 pm on January 15, 2024. The computer system can complete the segmentation operation by traversing the timestamps of the second fault record data stream and determining the segmentation points according to the preset maximum time window length.

[0138] In step S4522, based on the target fault record data in each third fault record data stream, the computer system generates a prior fault label corresponding to each third fault record data stream. The target fault record data is the fault record data with the latest recording time in each third fault record data stream. The prior fault label is known fault information used for subsequent training and evaluating network performance. For the first third fault record data stream of the above intelligent air conditioner, assuming that the fault record data with the latest recording time is a compressor fault, then the prior fault label corresponding to this third fault record data stream is a compressor fault. The computer system can find the latest fault record data by comparing the timestamps of each fault record in each third fault record data stream, and generate a prior fault label based on information such as the fault type of this data.

[0139] In step S4523, the computer system extracts the fault record data stream other than the target fault record data from each third fault record data stream, obtaining a fourth fault record data stream corresponding to each third fault record data stream. This step is to separate the target fault record data used to generate the prior fault label from other fault record data for subsequent separate processing. For the first third fault record data stream of the intelligent air conditioner, after excluding the target fault record data of the compressor fault, the remaining other fault record data constitutes the corresponding fourth fault record data stream. The computer system can obtain the fourth fault record data stream by simply screening out the target fault record data from the third fault record data stream.

[0140] In step S4524, the computer system determines the segmentation position corresponding to each fourth fault record data stream according to the actual time window length of each fourth fault record data stream. The method for determining the segmentation position is similar to that for determining the segmentation position of the first fault record data stream in step S200. For example, it is obtained by multiplying the actual time window length of the fourth fault record data stream by a preset segmentation coefficient and then combining it with the initial fault time. Assuming that the preset segmentation coefficient is 0.6, for the first fourth fault record data stream of the intelligent air conditioner, its actual time window length is 6 days and 23 hours, which is approximately 6×24 + 23 = 167 hours when converted to hours. The segmentation product is 167×0.6 = 100.2 hours. If the initial fault time of this fourth fault record data stream is 10 am on January 1, 2024, converting 100.2 hours to days and hours, 100.2÷24 = 4 days with a remainder of 4.2 hours, then the time corresponding to the segmentation position is 2 pm and 12 minutes on January 5, 2024 (0.2 hours is 12 minutes). The computer system can obtain the segmentation position by calculating the actual time window length of the fourth fault record data stream and then calculating according to the above rules.

[0141] In step S4525, the computer system performs a splitting operation on each fourth fault record data stream according to the splitting position, obtaining one or more fault record training data stream sets. Each fault record training data stream set includes a first fault record training data stream and a second fault record training data stream, which will be used for training the subsequent preset spatio-temporal feature fusion component and preset fault decision component respectively. For the first fourth fault record data stream of the intelligent air conditioner, taking 2:12 PM on January 5, 2024 as the splitting position, the fault record data before this time (including this moment, if it is stipulated that the splitting position data belongs to the first fault record training data stream) constitutes the first fault record training data stream, and the fault record data after this time constitutes the second fault record training data stream. These two data streams form a fault record training data stream set. The computer system traverses each fault record in the fourth fault record data stream, compares its timestamp with the time corresponding to the splitting position, and classifies the fault record into the corresponding training data stream, thereby obtaining the fault record training data stream set.

[0142] Please refer to Figure 3 , which is a schematic structural diagram of a computer system provided by an embodiment of the present invention. As Figure 3 shown, the above computer system 10 may include: a processor 1001, a network interface 1004, and a memory 1005. In addition, the above computer system 10 may further include: a user interface 1003 and at least one communication bus 1002. Among them, the communication bus 1002 is used to realize the connection and communication between these components. Among them, the user interface 1003 may include a display screen (Display) and a keyboard (Keyboard). Optionally, the user interface 1003 may further include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. Optionally, the memory 1005 may further be at least one storage device far from the aforementioned processor 1001. As Figure 3 shown, the memory 1005, as a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a device control application program.

[0143] In Figure 3 the computer system 10 shown, the network interface 1004 can provide network communication functions; while the user interface 1003 is mainly used to provide an input interface; and the processor 1001 can be used to call the device control application program stored in the memory 1005 to implement the method provided in the above embodiments.

[0144] In the description, claims and drawings of the embodiments of the present invention, terms such as "first" and "second" are used to distinguish the content in different media, rather than to describe a specific order. In addition, the term "comprising" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, device, product or equipment that includes a series of steps or units is not limited to the listed steps or modules, but may optionally further include steps or modules not listed, or may optionally further include other step units inherent to these processes, methods, devices, products or equipment.

[0145] The embodiments of the present invention further provide a computer program product, including a computer program / instructions, and when the computer program / instructions are executed by a processor, they implement the Figure 2 description of the above-mentioned method for analyzing maintenance record data based on AI in the corresponding embodiments. Therefore, it will not be elaborated here. In addition, the description of the beneficial effects of using the same method will not be elaborated either. For the technical details not disclosed in the embodiments of the computer program product involved in the present invention, please refer to the description of the method embodiments of the present invention.

[0146] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described according to functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0147] The method and related device provided by the embodiments of the present invention are described with reference to the method flowcharts and / or structure diagrams provided by the embodiments of the present invention. Specifically, each process and / or block of the method flowchart and / or structure diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable network-connected devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable network-connected devices generate a device for implementing the Figure 1 one process or multiple processes and / or structure diagrams Figure 1 function specified in one block or multiple blocks. These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable network-connected devices to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured product including an instruction device, and the instruction device implements the processFigure 1 One process or multiple processes and / or structural schematic Figure 1 The functions specified in one box or multiple boxes. These computer program instructions can also be loaded onto a computer or other programmable network-connected device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one process Figure 1 One process or multiple processes and / or structural schematic steps for the functions specified in one box or multiple boxes. The above-disclosed are only the preferred embodiments of the present invention. Of course, the scope of the rights of the present invention cannot be limited thereby. Therefore, equivalent changes made according to the claims of the present invention still fall within the scope covered by the present invention.

Claims

1. An AI-based analysis method for maintenance record data, characterized in that, The method includes: Obtaining a first fault record data stream corresponding to a target terminal device, where the first fault record data stream represents the fault record data of the target terminal device in a first operation and maintenance monitoring period, and is a data stream formed by arranging the fault record data of the target terminal device in the first operation and maintenance monitoring period in sequence; Performing a segmentation operation on the first fault record data stream to obtain a first to-be-processed fault record data stream and a second to-be-processed fault record data stream; Loading the first to-be-processed fault record data stream into a spatio-temporal feature fusion component in a fault forward reasoning network for multi-dimensional feature aggregation to obtain a spatio-temporal aggregation feature array; Loading the second to-be-processed fault record data stream and the spatio-temporal aggregation feature array into a fault decision-making component in the fault forward reasoning network for fault mode decision-making to obtain a target fault prediction result corresponding to the target terminal device at a forward-looking time point; the fault decision-making component is a decoder for performing fault mode decision-making according to the input second to-be-processed fault record data stream and spatio-temporal aggregation feature array; Wherein, the performing a segmentation operation on the first fault record data stream to obtain a first to-be-processed fault record data stream and a second to-be-processed fault record data stream includes: Obtaining the actual time window length of the first fault record data stream, where the actual time window length of the first fault record data stream represents the data stream length of the first fault record data stream, and the actual time window length of the first fault record data stream is the time difference between the earliest fault record moment and the latest fault record moment in the data stream; Multiplying the actual time window length of the first fault record data stream by a preset segmentation coefficient to obtain a segmentation product; Multiplying the segmentation product by the earliest fault moment in the first fault record data stream to obtain the segmentation position corresponding to the first fault record data stream; According to the segmentation position, taking the fault record data streams on both sides of the segmentation position in the first fault record data stream as the first to-be-processed fault record data stream and the second to-be-processed fault record data stream respectively, where the fault record data corresponding to the segmentation position belongs to the first to-be-processed fault record data stream or the second to-be-processed fault record data stream; The loading the first to-be-processed fault record data stream into a spatio-temporal feature fusion component in a fault forward reasoning network for multi-dimensional feature aggregation to obtain a spatio-temporal aggregation feature array includes: Indexing the time series attenuation vector corresponding to each fault record data in the first to-be-processed fault record data stream according to the target dynamic attenuation mapping matrix in the fault forward reasoning network to obtain a first target time series attenuation vector corresponding to the first to-be-processed fault record data stream, where the target dynamic attenuation mapping matrix includes the corresponding situations of different fault record data and target time series attenuation vectors; Determining a first target distribution positioning vector corresponding to the first to-be-processed fault record data stream according to the time series distribution situation of each fault record data in the first to-be-processed fault record data stream; Perform a vector fusion operation on the first target temporal decay vector and the first target distribution localization vector to obtain a spatio-temporal fusion input feature array; Load the spatio-temporal fusion input feature array into the spatio-temporal feature aggregation layer in the spatio-temporal feature fusion component for multi-dimensional feature aggregation to obtain the spatio-temporal aggregation feature array. Among them, the spatio-temporal feature fusion component is an encoder used to perform multi-dimensional feature aggregation on the input fault record data to complete the encoding of the data.

2. The method according to claim 1, wherein The fault decision component includes a decision feature extraction layer and a cross-attention decision layer. Loading the second to-be-processed fault record data stream and the spatio-temporal aggregation feature array into the fault decision component in the fault forward reasoning network for fault mode decision to obtain the target fault prediction result corresponding to the target terminal device at the forward-looking time point includes: According to the target dynamic decay mapping matrix in the fault forward reasoning network, index the temporal decay vector corresponding to each fault record data in the second to-be-processed fault record data stream to obtain the second target temporal decay vector corresponding to the second to-be-processed fault record data stream. The target dynamic decay mapping matrix contains the corresponding relationship between different fault record data and the target temporal decay vector; Determine the second target distribution localization vector corresponding to the second to-be-processed fault record data stream according to the temporal distribution of each fault record data in the second to-be-processed fault record data stream; Perform a vector fusion operation on the second target temporal decay vector and the second target distribution localization vector to obtain a decision input fusion vector; Load the decision input fusion vector into the decision feature extraction layer for decision feature extraction to obtain a primary decision vector; Load the primary decision vector and the spatio-temporal aggregation feature array into the cross-attention decision layer for fault decision processing to obtain the target fault prediction result.

3. The method according to claim 1, characterized in that The method further includes: Obtain an initialized fault forward reasoning network; Before obtaining the first fault record data stream corresponding to the target terminal device, the method further includes: Obtain the past fault record data stream corresponding to the target terminal device. The past fault record data stream represents the fault record data of the target terminal device in the past operation and maintenance monitoring cycle, and the past operation and maintenance monitoring cycle includes the first operation and maintenance monitoring cycle; Perform a segmentation operation on the past fault record data stream to obtain the first fault record data stream and the second fault record data stream; Iteratively adjust the parameters of the initialized fault forward reasoning network according to the second fault record data stream to obtain the converged fault forward reasoning network.

4. The method according to claim 3, wherein The initialized fault forward reasoning network includes a preset spatio-temporal feature fusion component and a preset fault decision component. Iteratively adjusting the parameters of the initialized fault forward reasoning network according to the second fault record data stream to obtain the converged fault forward reasoning network includes: Determine one or more sets of fault record training data streams and the corresponding prior fault labels for the one or more sets of fault record training data streams according to the second fault record data stream. Each set of fault record training data streams includes a first fault record training data stream and a second fault record training data stream; Load the first fault record training data stream into the preset spatio-temporal feature fusion component for multi-dimensional feature aggregation to obtain a training spatio-temporal aggregation feature array; Load the second fault record training data stream and the training spatio-temporal aggregation feature array into the preset fault decision component for fault mode decision to obtain a deduced fault decision result; Obtain a fault deduction error according to the deduced fault decision result and the prior fault label; Iteratively adjust the parameters of the initialized fault forward reasoning network according to the fault deduction error to obtain the fault forward reasoning network.

5. The method according to claim 4, wherein The step of loading the first fault record training data stream into the preset spatio-temporal feature fusion component for multi-dimensional feature aggregation to obtain a training spatio-temporal aggregation feature array includes: Index the corresponding temporal decay vector of each fault record data in the first fault record training data stream according to the preset dynamic decay mapping matrix in the initialized fault forward reasoning network to obtain the training temporal decay vector corresponding to the first fault record training data stream. The preset dynamic decay mapping matrix contains the corresponding relationship between different fault record data and the preset temporal decay vector; Determine the training distribution positioning vector corresponding to the first fault record training data stream according to the temporal distribution of each fault record data in the first fault record training data stream; Perform a vector fusion operation on the training temporal decay vector and the training distribution positioning vector to obtain a training spatio-temporal fusion input feature array; Load the training spatio-temporal fusion input feature array into the spatio-temporal feature aggregation layer of the preset spatio-temporal feature fusion component for multi-dimensional feature aggregation to obtain the training spatio-temporal aggregation feature array.

6. The method according to claim 5, characterized in that, The step of iteratively adjusting the parameters of the initialized fault forward reasoning network according to the fault deduction error to obtain the fault forward reasoning network includes: Iteratively adjust the network parameters in the initialized fault forward reasoning network and the preset temporal decay vector in the preset dynamic decay mapping matrix according to the fault deduction error to obtain the iterated initialized fault forward reasoning network; According to the iterated initialized fault forward reasoning network, iteratively execute the steps from determining one or more sets of fault record training data streams and the corresponding prior fault labels for the one or more sets of fault record training data streams according to the second fault record data stream to iteratively adjusting the network parameters in the initialized fault forward reasoning network and the preset temporal decay vector in the preset dynamic decay mapping matrix according to the fault deduction error to obtain the iterated initialized fault forward reasoning network until the preset iteration stop requirement is met to obtain the fault forward reasoning network.

7. The method according to claim 4, wherein The method further includes: Obtain a preset maximum time window length and the actual time window length of the second fault record data stream; The determining one or more fault record training data stream sets and the prior fault markers corresponding to the one or more fault record training data stream sets according to the second fault record data stream includes: Performing a segmentation operation on the second fault record data stream according to the actual time window length of the second fault record data stream and the preset maximum time window length to obtain one or more third fault record data streams; Generating a prior fault marker corresponding to each third fault record data stream according to the target fault record data in each third fault record data stream, where the target fault record data is the fault record data with the latest record time corresponding to each third fault record data stream; Extracting the fault record data stream other than the target fault record data in each third fault record data stream to obtain a fourth fault record data stream corresponding to each third fault record data stream; Determining a segmentation position corresponding to each fourth fault record data stream according to the actual time window length of each fourth fault record data stream; Performing a segmentation operation on each fourth fault record data stream according to the segmentation position to obtain the one or more fault record training data stream sets.

8. A computer system, characterized in that, Including: One or more processors; And one or more memories, wherein computer-readable code is stored in the memories, and when the computer-readable code is run by the one or more processors, the one or more processors execute the method according to any one of claims 1 to 7.

Images