Operation and maintenance data characterization method and related device

Abstract

The embodiment of the invention provides an operation and maintenance data characterization method and a related device, and is applied to the field of AI and ICT operation and maintenance. The method comprises the steps that each piece of source data in multi-source data is segmented into a plurality of data fragments from the time dimension, the multi-source data is operation and maintenance data generated in the product operation process, and the operation and maintenance data comprises data generated by a product and / or description information, generated by other devices, of the product; inputting the plurality of data fragments corresponding to each piece of source data into a data representation model for representation extraction to obtain a sub-sequence corresponding to each piece of source data; and fusing the plurality of periodic signals and the target sequence to obtain characterization information of the multi-source data. According to the embodiment of the invention, the characterization effect of the multi-source data fusion feature can be improved.

Description

Technical Field

[0001] The present application relates to fields such as artificial intelligence (AI), information and communication technology (ICT) operation and maintenance, and in particular to a method for representing operation and maintenance data and related devices. Background Art

[0002] As the scale of data that needs to be tracked by network operation and maintenance increases rapidly, the difficulty of data operation and maintenance increases dramatically. People expect to use artificial intelligence technology to improve the automation capability of operation and maintenance. Among them, data representation of operation and maintenance data is an important part of intelligent operation and maintenance. Through data representation, effective information in operation and maintenance data can be extracted, redundant information in data can be eliminated, and complex business data can be simplified and passed to subsequent algorithms. The quality of data representation may directly affect the effect of subsequent operation and maintenance models and algorithms.

[0003] However, since the operation and maintenance data includes various multi-source data with complex structures and the sampling intervals of multi-source data are different, some data may be missing in some time periods during the operation and maintenance process, which makes the accuracy of the representation fusion of operation and maintenance data low. Therefore, how to improve the representation effect of multi-source data fusion features is an urgent problem to be solved. Summary of the invention

[0004] The embodiments of the present application provide a method and related device for characterizing operation and maintenance data, which can effectively improve the characterization effect of multi-source data fusion features.

[0005] In a first aspect, an embodiment of the present application provides a method for characterizing operation and maintenance data, the method comprising:

[0006] Divide each source data in the multi-source data into multiple data segments from a time dimension, wherein the multi-source data is operation and maintenance data generated during the operation of the product, and the operation and maintenance data includes data generated by the product and / or description information of the product generated by other devices;

[0007] Inputting multiple data fragments corresponding to each source data into a data representation model for representation extraction to obtain a subsequence corresponding to each source data, wherein each subsequence includes multiple tokens corresponding to the multiple data fragments;

[0008] Multiple periodic signals and a target sequence are fused to obtain representation information of the multi-source data, wherein the target sequence includes multiple subsequences corresponding to multiple source data, a portion of the multiple periodic signals are association relationships between different Tokens of the same source data generated according to prior knowledge, and another portion of the multiple periodic signals are association relationships between different Tokens of different source data.

[0009] In the above method, by constructing a multi-source data sequence with a periodic pattern and adding a periodic signal in the representation learning process, the prior knowledge of operation and maintenance data is integrated into the training and reasoning process of the model, which can improve the representation effect of the multi-source data fusion feature.

[0010] In a possible implementation, the representation information of the multi-source data is used as an input to a downstream model, wherein the downstream model is used to predict or classify business.

[0011] In the above method, the method of integrating the prior knowledge of operation and maintenance data into the training and reasoning process of the model is applicable to all Transformer-based discriminative models or generative models, which enhances the representation and fusion effect of multi-source data and makes the business prediction or business classification through the training model more accurate.

[0012] In another possible implementation, the inputting of the multiple data segments corresponding to each source data into the data representation model for representation extraction to obtain the subsequence corresponding to each source data includes:

[0013] Inputting the multiple data segments corresponding to each source data into a large language model LLM to obtain data descriptions corresponding to the multiple data segments respectively;

[0014] Generate a semantic vector corresponding to each data description, wherein the semantic vector corresponding to the first data description is the Token corresponding to the first data segment, the first data description is the data description corresponding to the first data segment, and the first data segment is any one of the multiple data segments.

[0015] In the above method, the data representation model can be a large language model. After converting various forms of source data into natural language text form, different algorithms are used to calculate the semantic vector corresponding to each data description, and finally the semantic vector corresponding to the data description is used as the token corresponding to the data segment.

[0016] In another possible implementation, the data description includes one or more of the following:

[0017] objective data descriptions of the plurality of data segments;

[0018] a data summary obtained after analyzing the plurality of data segments;

[0019] The default data description.

[0020] In the above method, the data description can be a concrete and objective example description or an abstract and general description. In addition, if a source data is missing a data segment within a time slice, the data description corresponding to the missing data segment can be a preset data description (such as configuration, log, KPI and other data descriptions with business significance).

[0021] In another possible implementation, the step of inputting the multiple data segments corresponding to each source data into a data representation model for representation extraction to obtain a subsequence corresponding to each source data includes:

[0022] A feature extraction model is used to perform feature extraction on the multiple data fragments corresponding to each source data to obtain multiple feature vectors corresponding to the multiple data fragments, wherein the feature vector corresponding to the first data fragment is the token corresponding to the first data fragment, and the first data fragment is any one of the multiple data fragments.

[0023] In the above method, the data representation model can be a feature extraction model, which uses the existing data feature extraction model of each source data to map all data into a feature vector form with a uniform dimension, and then obtains multiple tokens of multiple data fragments corresponding to each source data.

[0024] In another possible implementation, fusing the multiple periodic signals with the target sequence to obtain the representation information of the multi-source data includes:

[0025] Converting the periodic signal into a bias matrix;

[0026] Add the bias matrix to the self-attention module that incorporates the target sequence, and calculate the association weights between different tokens in the target sequence;

[0027] The characterization information of the multi-source data is determined by the association weights.

[0028] In the above method, multiple periodic signals and target sequences are fused, which not only solves the problem of difficult alignment of multi-source data caused by different sampling granularities of source data, but also provides time information for the multi-source data fusion module, which is beneficial to the subsequent correlation analysis of multi-source data. The representation information of multi-source data is determined through correlation analysis, thereby improving the fusion representation effect of multi-source data.

[0029] In another possible implementation, the self-attention module includes a self-attention head, and the bias matrix corresponding to the self-attention head is a weighted accumulation result of the bias matrices of each periodic signal in the multiple periodic signals, wherein the bias matrix is ​​used to enhance the association relationship between the multiple tokens based on the association relationship calculated by the original self-attention head, or,

[0030] The self-attention module includes multiple self-attention heads, and the bias matrix corresponding to each of the multiple self-attention heads is the bias matrix corresponding to a periodic signal among the multiple periodic signals, or each self-attention head is the weighted accumulation result of the bias matrices of each periodic signal in the same source data, or each self-attention head is the original self-attention head without adding the periodic signal.

[0031] In the above method, the original part of the original self-attention head calculates the correlation between multiple tokens, and the newly added bias matrix can further enhance the correlation between multiple tokens based on the original calculated correlation, thereby improving the information extraction effect when multi-source data is fused, that is, the representation ability of the final multi-source data can be enhanced.

[0032] In another possible implementation, fusing the periodic signal and the target sequence to obtain the representation information of the multi-source data includes:

[0033] Converting the periodic signal into a coding matrix based on a cyclic coding matrix REM;

[0034] Adding the encoding matrix to a self-attention module that incorporates the target sequence, and calculating association weights between different tokens in the target sequence, wherein the self-attention module includes multiple attention heads;

[0035] The characterization information of the multi-source data is determined by the association weights.

[0036] In the above method, another optional method of adding periodic signals to the self-attention module in the form of a matrix is ​​to add the modeling advantage of the recurrent neural network (RNN) for periodic data to the multi-head attention of the self-attention module in the form of an encoding matrix based on the recurrence encoding matrix (REM). By configuring the periodic parameters of each attention mechanism, the ratio of the original self-attention matrix weights and the encoding matrix weights in the final result of the attention is adaptively adjusted in a data-driven manner, and the periodic signal preset based on prior knowledge is added to the representation learning of multi-source data, so as to improve the information extraction effect when multi-source data is fused.

[0037] In another possible implementation, dividing each source data in the multi-source data into a plurality of data segments from a time dimension includes:

[0038] Divide the first source data according to a preset time slice length from a time dimension to obtain a plurality of data segments corresponding to the first source data in the plurality of time slices, wherein if the first source data has first data in a first time slice, the first data is the data segment corresponding to the first time slice;

[0039] If the first source data does not exist in the first time slice, the data segment historically corresponding to the first time slice is the data segment currently corresponding to the first time slice, wherein the first time slice is any one of the multiple time slices, and the first source data is any one of the multiple source data.

[0040] In the above method, during the sequence generation process, if a source data is missing a data fragment within a time slice, the periodic characteristics of the source data can be used to select at least one data fragment of a time slice with a periodic association from the time slice before the missing time slice, and the average of the tokens corresponding to the data fragments of all selected time slices can be used as the token of the corresponding missing fragment. This solution can complete the data fragments of the missing time slices of the source data based on the periodic association characteristics, effectively solving the problems of heterogeneity and data missing faced by multi-source data.

[0041] In another possible implementation, fusing the multiple periodic signals with the target sequence to obtain the representation information of the multi-source data includes:

[0042] Comprehensively sorting multiple tokens of multiple data fragments corresponding to the multiple source data to obtain a sorted target sequence, wherein the token with the earlier corresponding time slice is sorted higher, and the token with the earlier corresponding source data in the same time slice is sorted higher;

[0043] The multiple periodic signals are fused with the sorted target sequence to obtain the representation information of the multi-source data.

[0044] In the above method, when the time slice lengths corresponding to different Tokens in multiple source data are the same, a comprehensive sorting is performed according to the sorting rule that the earlier the corresponding time slice, the earlier the Token is sorted, and the earlier the source data is sorted in the same time slice. The sorted target sequence is obtained. This scheme can make each source data have the same number of Tokens in the target sequence, and the Tokens corresponding to each source data appear in the form of periodic intervals.

[0045] In another possible implementation, comprehensively sorting the multiple Tokens of the multiple data fragments corresponding to the multiple source data to obtain a sorted target sequence includes:

[0046] Copy the first Token of the second source data in the second time slice to obtain multiple updated first Tokens of the second source data in the second time slice, wherein the time period of the second time slice is the same as the time period of the multiple time slices of the third source data, the contents of the multiple updated first Tokens are the same as the original first Tokens, the time slices of the multiple updated first Tokens are synchronized with the multiple time slices of the third source data in a one-to-one correspondence, the third source data is the source data with the shortest time slice length among the multiple source data, and the time slice length of the second source data is greater than the time slice length of the third source data;

[0047] The multiple Tokens of the multiple data fragments corresponding to the multiple source data are comprehensively sorted to obtain a sorted target sequence.

[0048] In the above method, if the time slice lengths corresponding to different Tokens in multiple source data are different, the different time slice lengths of each source data can be copied into source data with the same time slice length according to the smallest time slice granularity, and the content of each copied source data remains the same as the original source data. Then, according to the sorting rule that the earlier the corresponding time slice, the earlier the Token is sorted, and the earlier the source data is sorted in the same time slice, the earlier the Token is sorted, and the sorted target sequence is obtained. This scheme can make each source data have the same number of Tokens in the target sequence, and the Tokens corresponding to each source data appear in the form of periodic intervals when the time slice lengths corresponding to different Tokens in multiple source data are different.

[0049] In yet another possible implementation, source data among the multiple source data that has a greater degree of influence on other source data among the multiple source data is ranked higher.

[0050] In the above method, when sorting the order of each source data within the shortest time slice, the operation and maintenance data that may affect other operation and maintenance data can be prioritized, for example, the configuration data that may have a global impact can be ranked in a front position.

[0051] In another possible implementation, if the prior knowledge indicates that the first Token in the first time slice and the second Token in the second time slice are in a preset periodic state and / or a similar state, the first periodic signal includes an association relationship between the first Token in the first time slice and the second Token in the second time slice, wherein the first Token and the second Token both belong to the first source data, and the first periodic signal is one of the multiple periodic signals.

[0052] In the above method, it is necessary to design a periodic signal based on prior knowledge, and convert the prior knowledge into a periodic signal that can be added to multiple tokens, that is, to use the periodic signal to establish additional connections between associated tokens. The way to convert prior knowledge into a periodic signal can be: when the prior knowledge is "the same source operation and maintenance data is in a similar state between different time slices, or in a periodic repetitive state", for example, some operation and maintenance data can be in a periodic repetitive state to varying degrees within different time spans, then the prior knowledge can be converted into: 1) connections between the same source data tokens between adjacent time slices; 2) connections between the same source data tokens across at least one time slice.

[0053] In another possible implementation, if the prior knowledge indicates that the time slice length of the second source data is greater than the time slice length of the third source data, and the interval between the third Token and the fourth Token spans multiple time slices, then the second periodic signal includes an association relationship between the third Token and the fourth Token, wherein the third Token and the fourth Token both belong to the second source data, and the second periodic signal is used to add prior knowledge with a longer period association.

[0054] In the above method, the way to convert prior knowledge into a periodic signal can also be: when the prior knowledge is "affected by the replication mechanism with different time slice lengths between multiple Tokens, and the time slice interval corresponding to the actual Token spans multiple time slices", then the prior knowledge can be converted into: the connection between Tokens of the same source data across at least one time slice.

[0055] In another possible implementation, the third periodic signal includes an association relationship between a fifth Token and a sixth Token, the fifth Token is any one Token in the fourth source data, the sixth Token subsequence is any one Token in the fifth source data, the fourth source data is the source data among the multiple source data that has the greatest impact on other source data, and the fifth source data is any one of the multiple source data except the fourth source data.

[0056] In the above method, the way to convert prior knowledge into a periodic signal can also be: when the prior knowledge is that "changes in some operation and maintenance data can take effect globally", such as changes in background configuration will affect other operation and maintenance data, then the prior knowledge can be converted into: the connection between tokens of different source data in the same time slice.

[0057] In another possible implementation, the bias matrix includes n rows and n columns, and the value of the element M(i, j) in the i-th row and j-th column of the bias matrix represents the degree of association between the i-th Token and the j-th Token in the target sequence. The larger the value of the element M(i, j), the higher the degree of association between the i-th Token and the j-th Token. n is the number of tokens in multiple tokens of multiple data fragments corresponding to each source data, i takes integers between 1 and n, and j takes integers between 1 and n.

[0058] In the above method, before adding the preset periodic signal to the self-attention module for fusion sequence, it is necessary to first convert the periodic signal into a bias matrix form, wherein each periodic signal has its own corresponding bias matrix, and the association relationship between the tokens of the same source data or different source data is determined according to the value of the element M(i, j) in the i-th row and j-th column of the bias matrix, and then the preset periodic signal is added to the self-attention module for fusion sequence in the form of a bias matrix, and the representation learning of multi-source data is realized by adding a preset inductive bias between multiple tokens.

[0059] In the second aspect, the present application also provides a device for characterizing operation and maintenance data. The characterization device can implement some or all of the functions performed by the electronic device described in the first aspect above. For example, the functions of the characterization device may have the functions of some or all of the embodiments of the electronic device described in the first aspect of the present application. The functions can be implemented by hardware or by hardware executing corresponding software. The hardware or software includes one or more units or modules corresponding to the above functions.

[0060] In an optional implementation, the characterization device includes a segmentation unit, an input unit, and a fusion unit, wherein:

[0061] The segmentation unit is used to segment each source data in the multi-source data into multiple data segments from a time dimension, wherein the multi-source data is operation and maintenance data generated during the operation of the product, and the operation and maintenance data includes data generated by the product and / or description information of the product generated by other devices;

[0062] The input unit is used to input the multiple data segments corresponding to each source data into the data representation model for representation extraction, so as to obtain a subsequence corresponding to each source data, wherein each subsequence includes a plurality of tokens corresponding to the multiple data segments;

[0063] The fusion unit is used to fuse multiple periodic signals and a target sequence to obtain representation information of the multi-source data, wherein the target sequence includes multiple subsequences corresponding to multiple source data, a part of the multiple periodic signals are association relationships between different Tokens of the same source data generated according to prior knowledge, and another part of the multiple periodic signals are association relationships between different Tokens of different source data.

[0064] In a third aspect, an embodiment of the present application provides an electronic device, which includes a processor; when the processor calls a computer program or instruction in a memory, it executes any method described in the first aspect.

[0065] Optionally, the electronic device further includes a communication interface, wherein the communication interface is used to receive and / or send data, and / or the communication interface is used to provide input and / or output for the processor.

[0066] It should be noted that the above embodiment is described by taking a processor (or general-purpose processor) that executes the method by calling a computer specification as an example. In the specific implementation process, the processor can also be a dedicated processor, in which case the computer instructions have been pre-loaded in the processor. Optionally, the processor can also include both a dedicated processor and a general-purpose processor.

[0067] Optionally, the communication device may further include a memory, which may be used to store computer programs or instructions. Furthermore, the memory may be located outside the processor, or integrated with the memory.

[0068] In a fourth aspect, the present application provides a computer-readable storage medium for storing instructions, which, when executed by a computer, implements the method described in any one of the first aspects above.

[0069] In a fifth aspect, the present application also provides a computer program product comprising instructions, which, when executed on a computer, implements the method described in any one of the first aspects above.

[0070] Optionally, the computer program product may be a software installation package or an image package. When the aforementioned method is required, the computer program product may be downloaded and executed on a computing device.

[0071] The beneficial effects of the technical solutions provided in the second to fifth aspects of the present application can refer to the beneficial effects of the technical solution of the first aspect, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The following is an introduction to the drawings used in the embodiments of the present application.

[0073] Figure 1 It is a schematic diagram of multi-source data feature fusion based on a data representation model provided in an embodiment of the present application;

[0074] Figure 2 is a schematic structural diagram of an electronic device 20 provided in an embodiment of the present application;

[0075] Figure 3 This is a business scenario diagram of a sequence prediction provided by an embodiment of the present application;

[0076] Figure 4 This is a business scenario diagram of a root cause analysis provided by an embodiment of the present application;

[0077] Figure 5 This is a business scenario diagram of anomaly detection provided by an embodiment of the present application;

[0078] Figure 6 It is a flow chart of a method for characterizing operation and maintenance data provided in an embodiment of the present application;

[0079] Figure 7 is a schematic diagram of constructing a target sequence of multi-source data provided by an embodiment of the present application;

[0080] Figure 8 is a schematic diagram of another method for constructing a target sequence of multi-source data provided by an embodiment of the present application;

[0081] Fig. 9 It is a schematic diagram of converting prior knowledge into a periodic signal provided by an embodiment of the present application;

[0082] Fig.10 This is a schematic diagram of a periodic signal of a Token of homologous data between adjacent time slices provided by an embodiment of the present application;

[0083] Fig.11 This is a schematic diagram of a periodic signal of a homologous data Token across time slices provided by an embodiment of the present application;

[0084] Fig.12 This is a schematic diagram of a periodic signal between Tokens of different source data in the same time slice provided by an embodiment of the present application;

[0085] Fig.13This is a flow chart of an operation and maintenance data fusion representation provided by an embodiment of the present application;

[0086] Fig.14 It is a schematic diagram of another process of operation and maintenance data fusion representation provided by an embodiment of the present application;

[0087] Fig.15 It is a structural diagram of a device 150 for representing operation and maintenance data provided in an embodiment of the present application. DETAILED DESCRIPTION

[0088] The technical solution in this application will be described below in conjunction with the accompanying drawings.

[0089] Due to the applications involved in the embodiments of the present application, in order to facilitate understanding, the relevant terms and other related concepts involved in the embodiments of the present application are first introduced below.

[0090] 1. Data Operation and Maintenance

[0091] Data operation and maintenance is the process of using data analysis technology to manage operation and maintenance data, that is, to ensure that the data is safe, reliable, effective, real-time, and available to users.

[0092] 2. Multi-source data

[0093] Multi-source data refers to data with multiple sources. Multi-source heterogeneous data comes from multiple data sources, including different database systems and / or data sets collected by different devices during work. Different data sources may have different operating systems and management systems (such as OA, CRM, HR, MES, SCM), different data storage modes and logical structures, and different data generation times, usage locations, code protocols, etc. This creates the "multi-source" feature of data. Currently, the commonly used feature fusion technology for multi-source data integrates all the information obtained through investigation and analysis, and uniformly evaluates the information to obtain unified information. The purpose of developing this technology is to integrate various different data information, absorb the characteristics of different data sources, and then extract unified information from them that is better and richer than single data.

[0094] 3. Prior Knowledge

[0095] Prior knowledge refers to knowledge that has been verified in advance. With the knowledge verified in advance, we can better infer the knowledge of the next stage. In conventional image detection, prior knowledge can be understood together with the fine-tuning / feature extractor in the training model. If we do not use prior knowledge to perform human image detection, it may take a long time and the effect may not be good. However, if we use prior knowledge to replace the weights in the model with the weight parameters, edges, contours, etc. of the known trained model, the effect will be much better.

[0096] 4. Multi-head attention mechanism

[0097] The purpose of the multi-head attention mechanism is to improve the expressiveness of the artificial intelligence (AI) model by capturing different attention information. The subspace of the multi-head matrix is ​​used to express the input data from different perspectives or dimensions. From the perspective of Bayesian neural networks, the multi-head attention mechanism is a sampling technique, and each attention head is a sample. The greater the discrimination of each head, the more perspectives, which enables the Transformer model to better fit the target.

[0098] The above description of related concepts can be applied to the following embodiments.

[0099] See also Figure 1 , Figure 1 is a schematic diagram of a multi-source data feature fusion based on a data representation model provided in an embodiment of the present application, such as Figure 1 As shown in the figure, in one implementation, for the problem of feature fusion of multi-source data, the technology first divides each source data into multiple data segments according to the same time slice length, then uses different data representation models to serialize each source data segment into tokens of different lengths, then merges all tokens into a long token sequence in chronological order, and finally uses the self-attention module inside the Transformer model to perform feature fusion on the multi-source data in the long token sequence. However, this technology treats all tokens in the token sequence equally, ignores the strong prior of operation and maintenance data - periodicity, weakens the association between tokens of the same source data in different time periods, and thus reduces the fusion representation ability of operation and maintenance data; and due to the different sampling intervals and data scales of multi-source data, multi-source data may not be able to use the same time slice division granularity, and cannot process time slices of multi-source data of different lengths; and does not consider the possible data missing problems of multi-source data.

[0100] In response to the above-mentioned technical problems existing in the fusion representation of operation and maintenance data, the embodiment of the present application provides a new architecture for representing operation and maintenance data, and accordingly proposes a new method for representing operation and maintenance data based on the architecture, which can effectively improve the representation effect of multi-source data fusion features.

[0101] See also Figure 2 , Figure 2: is a structural schematic diagram of an electronic device 20 provided in an embodiment of the present application, the electronic device 20 is used to implement a method for characterizing operation and maintenance data, the electronic device 20 can be a device deployed locally, or a cloud server deployed in the cloud, the electronic device 20 can be a single device, or a distributed system composed of multiple devices. The electronic device 20 may include a processor 201, and may optionally include at least one memory 202. Further optionally, the electronic device 20 may also include a communication interface 203. Further optionally, it may also include a bus 204, wherein the processor 201, the memory 202 and the communication interface 203 are connected via the bus 204.

[0102] Among them, the processor 201 is a module that performs arithmetic operations and / or logical operations, and can specifically be a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MPU), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a complex programmable logic device (CPLD), a coprocessor (assisting the central processing unit to complete corresponding processing and applications), a microcontroller unit (MCU) and other processing modules or a combination of multiple thereof.

[0103] The memory 202 is used to provide a storage space, and the storage space can store data such as an operating system and a computer program. The memory 202 can be a random access memory (Random Access Memory, RAM), a read-only memory (Read-only Memory, ROM), an erasable programmable read-only memory (Erasable Programmable Read-only Memory, EPROM), or a portable read-only memory (Compact Disc Read-only Memory, CD-ROM) and the like, or a combination of multiple thereof.

[0104] The communication interface 203 may be used to provide information input or output for the at least one processor. And / or, the communication interface 203 may be used to receive data sent externally and / or send data externally, and may be a wired link interface including an Ethernet cable, etc., or a wireless link (Wi-Fi, Bluetooth, general wireless transmission, vehicle-mounted short-range communication technology, and other short-range wireless communication technologies, etc.) interface. Optionally, the communication interface 203 may also include a transmitter (such as a radio frequency transmitter, an antenna, etc.) coupled to the interface, or a receiver, etc.

[0105] The processor 201 in the electronic device 20 is used to execute Figure 6 The method described in the embodiment, for example, the processor 201 in the electronic device 20 calls a computer program, and specifically can perform the following operations:

[0106] The operation of fusion representation is performed on the operation and maintenance data, and the operation of fusion representation specifically includes the following steps.

[0107] First, each source data in the multi-source data is divided into multiple data segments from the time dimension, where the multi-source data is the operation and maintenance data generated during the operation of the product, and the operation and maintenance data includes data generated by the product and / or description information of the product generated by other devices (the data structure and mode of the above multi-source data may be different, but the same source data have similarity or periodicity in time series).

[0108] Then, through the periodic Token sequence construction module, multiple data fragments corresponding to each source data are input into the data representation model for representation extraction to obtain a subsequence corresponding to each source data, each subsequence including multiple Tokens corresponding to multiple data fragments.

[0109] Finally, multiple periodic signals and target sequences are fused through the prior knowledge conversion module and the multi-information fusion module to obtain the representation information of multi-source data, where the target sequence includes multiple subsequences corresponding to multiple source data, a part of the multiple periodic signals are the association relationships between different Tokens of the same source data generated according to prior knowledge, and another part of the multiple periodic signals are the association relationships between different Tokens of different source data.

[0110] By performing the above operations, the characterization effect of multi-source data fusion features can be effectively improved while retaining time information.

[0111] Optionally, the processor 201 may be a processor specifically used to execute these methods (referred to as a dedicated processor for convenience of distinction), or may be a processor that executes these methods by calling a computer program, such as a general-purpose processor. In this case, the memory 202 is required to store the computer program called by the processor. Optionally, the at least one processor may include both a dedicated processor and a general-purpose processor.

[0112] In addition, it should be noted that the method for representing operation and maintenance data proposed in the embodiment of the present application can be applied to the scenario of large model pre-training in the field of information and communication technology (ICT) operation and maintenance, and can also be widely used in various downstream tasks in the field of ICT operation and maintenance, such as sequence prediction (see Figure 3 , Figure 3 is a schematic diagram of a business scenario of sequence prediction provided by an embodiment of the present application, such as Figure 3 As shown, for example, multi-source data includes key performance indicators (KPI), alarms, signaling, logs, configurations and other operation and maintenance data. Among them, the sequence prediction for KPI can be KPI value, the sequence prediction for alarm can be: whether there is XX alarm, the sequence prediction for signaling can be: predict the business to be occurred, the sequence prediction for log can be: log content, the sequence prediction for configuration can be: whether to change the configuration, etc.), root cause analysis (see Figure 4 , Figure 4 is a business scenario diagram of a root cause analysis provided by an embodiment of the present application, such as Figure 4 As shown, for example, the root cause analysis for multi-source data can be: 1. Problem description XXX; 2. Problem analysis: including principle part, analysis part and root cause part. The analysis part can be XX log and XX alarm to indicate that the device has been powered off; the root cause part can be that account 1 is saved in the device configuration file and log at the same time, and the power failure during the upgrade causes XX alarm, and then the acquisition of XX information fails), anomaly detection (see Figure 5 , Figure 5 is a schematic diagram of a business scenario of anomaly detection provided by an embodiment of the present application, such as Figure 5 As shown, for example, Figure 5 The part marked in the middle is abnormal) etc.

[0113] In addition, since multi-source heterogeneity and periodicity are also typical characteristics of IT operation and maintenance data, this application, as a characterization method for multi-source data, can also be used in the IT operation and maintenance field to improve model pre-training and downstream task effects in this field.

[0114] See also Figure 6 , Figure 6 is a flow chart of a method for characterizing operation and maintenance data provided in an embodiment of the present application. Optionally, the characterization method can be applied to the aforementioned electronic device to implement, for example Figure 2 Electronic equipment shown.

[0115] The characterization method includes but is not limited to steps S601 to S603 as shown below. It should be understood that for the convenience of description, the order of steps S601 to S603 is described here, and it is not intended to limit the execution to the above order. The embodiment of the present application does not limit the execution order, execution time, execution number, etc. of the above one or more steps, and other steps can be interspersed between or before and after these steps as needed, where:

[0116] Step S601: the electronic device divides each source data in the multi-source data into multiple data segments from the time dimension.

[0117] Specifically, the multi-source data includes source data 1, source data 2, source data 3, ..., source data i, where, for example, D i (t s , t e ) represents the source data i from time t s At time t e The data fragments between.

[0118] Specifically, the electronic device may divide the source data 1 into data segments D1(t0, t1), data segment D1(t1, t2), data segment D1(t2, t3), ..., data segment D1(t s1 , t e1 ); divide the source data 2 into data segments D2(t0, t2), data segments D2(t2, t4), data segments D2(t4, t6), ..., data segments D2(t s2 , t e2 ); divide the source data 3 into data segments D3(t0, t3), data segment D3(t3, t6), data segment D3(t6, t9), ..., data segment D3(t s3 , t e3 );Split the source data i into data segments D i (t0, t4), data segment D2 (t4, t8), data segment D i (t8, t 12 ), ..., data fragment D i (t si , t ei ).

[0119] In an optional implementation, the time slice lengths of homologous data are the same.

[0120] Taking the above description as an example, when i=1, s=0, e=1, the data segment D1(t0, t1) is the data of the source data 1 from time t0 to time t1; when i=1, s=1, e=2, the data segment D1(t1, t2) is the data of the source data 1 from time t1 to time t2. The data segment D1(t0, t1) and the data segment D1(t1, t2) both belong to the data segments of the source data 1, and the time slice length of the two is 1 time unit, which can be hours, minutes, seconds, etc. The examples of time units will not be repeated below.

[0121] In an optional implementation, the lengths of time slices between different source data may be different.

[0122] Still taking the above description as an example, when i=1, s=1, e=2, the data segment D1(t1, t2) is the data of the source data 1 from time t1 to time t2; when i=2, s=2, e=4, the data segment D2(t2, t4) is the data of the source data 2 from time t2 to time t4, the data segment D1(t0, t2) and the data segment D2(t2, t4) are data segments of different source data, and the time slice length of the data segment D1(t0, t2) is 1 time unit, and the time slice length of the data segment D2(t2, t4) is 2 time units, and the time slice lengths of the two are different.

[0123] In an optional implementation, the lengths of time slices between different source data may be the same.

[0124] Still taking the above description as an example, when i=1, s=0, e=1, the data segment D1(t0, t1) is the data of source data 1 from time t0 to time t1; when i=2, s=1, e=2, the data segment D2(t1, t2) is the data of source data 2 from time t1 to time t2. The data segment D1(t0, t1) belongs to the data segment of source data 1, and the data segment D2(t1, t2) belongs to the data segment of source data 2. The two belong to different source data, but the time slice length is the same, which is 1 time unit.

[0125] Step S602: the electronic device inputs a plurality of data segments corresponding to each source data into a data representation model for representation extraction to obtain a subsequence corresponding to each source data.

[0126] Specifically, after obtaining the multiple data fragments corresponding to each source data after segmentation, the electronic device can further use the data representation model to process the multiple data fragments corresponding to each source data (the processing operation can be representation extraction or other operations on the multiple data fragments) to obtain a subsequence corresponding to each source data (for example, source data 1 corresponds to subsequence 1, source data 2 corresponds to subsequence 2,..., source data i corresponds to subsequence i), each subsequence includes multiple Tokens corresponding to the multiple data fragments (for example, subsequence 1 includes Token1, Token2 and Token3, subsequence 2 includes Token4, Token5 and Token6, and subsequence 3 includes Token7, Token8 and Token9).

[0127] In an optional implementation, step S602 may include the following content: the electronic device inputs multiple data fragments corresponding to each source data into a large language model LLM to obtain data descriptions corresponding to the multiple data fragments; then generates a semantic vector corresponding to each data description, and the semantic vector corresponding to each data description is a Token corresponding to each data fragment.

[0128] The semantic vector corresponding to the first data description is the token corresponding to the first data segment, the first data description is the data description corresponding to the first data segment, and the first data segment is any one of the multiple data segments. In other words, the description of the first data segment in this application document is only an example, and in fact, each of the multiple data segments has the same characteristics as the first data segment.

[0129] In this solution, the data representation model can be a large language model (optionally, the large language model can be a unified language model that supports processing various source data. After the data segments corresponding to each source data are input into the large language model, the data description corresponding to each data segment can be obtained. For example, the data type of source data 1 is KPI, and the data type of source data 2 is alarm. The large language model can process both KPI business data and alarm business data. Source data 1 and source data 2 are input into the large language model to obtain the corresponding data description, that is, the source data and the large language model have a many-to-one correspondence. The large language model can also be a language model composed of a plurality of language models that process single source data. Each source data segment is input into the corresponding processing unit. For example, source data 1 is input into sub-language model 1 to obtain the data description corresponding to source data 1, and source data 2 is input into sub-language model 2 to obtain the data description corresponding to source data 2, that is, the source data and the large language model are in a many-to-many correspondence). For any source data in text form or that can be described by text (such as common alarm business data, KPI business data, and log business data in operation and maintenance data), the source data can be described in text first. After converting the source data into natural language text form, different algorithms are used to calculate the semantic vector corresponding to each data description, and finally the semantic vector corresponding to the data description is used as the Token corresponding to the data segment.

[0130] In an optional implementation, the data description includes one or more of the following: an objective data description of multiple data segments, a data overview obtained after analyzing multiple data segments, and a preset data description.

[0131] It should be noted that the preset data description is a historical data description with business significance.

[0132] Specifically, the above data description includes but is not limited to the example description provided in Table 1. As shown in Table 1, if the data type of source data 1 in the multi-source data is KPI business data, the data type of source data 2 is alarm business data, the data type of source data 3 is log business data, the data type of source data 4 is signaling business data, and the data type of source data 5 is configuration business data, wherein, taking KPI business data as an example, the data description of KPI business data is illustrated as follows:

[0133] For KPI cell switching service data, if there is no data missing in the KPI cell switching service data, its objective data description can be that the success rate of the KPI cell switching service data in the operation and maintenance process drops sharply from 90% to 75%; if an abstract data description is required, the summary data overview obtained after analyzing the KPI cell switching service data can be that the success rate of the KPI cell switching service data drops by more than 20%. If there is data missing in the KPI cell switching service data, it means that the specific service data cannot be obtained at present, and further, the specific service data cannot be described. At this time, the corresponding data description with business meaning can be selected as the data description of the current missing service data. For example, the preset data overview for the KPI cell switching service data can be that there is no KPI cell switching service data in the current period.

[0134] For the data description of other business data, please refer to the aforementioned KPI business data, which will not be repeated here.

[0135] Table 1

[0136]

[0137]

[0138] In an optional implementation, step S602 may include the following content: the electronic device uses a feature extraction model to extract features from multiple data segments corresponding to each source data, and obtains multiple feature vectors corresponding to the multiple data segments, where the feature vector corresponding to each data segment is a Token corresponding to each data segment.

[0139] The feature vector corresponding to the first data segment is a Token corresponding to the first data segment, and the first data segment is any one of the multiple data segments.

[0140] In this scheme, the data representation model can be a feature extraction model, which can use the existing data feature extraction model of each source data to map all data into a feature vector form with a uniform dimension (optionally, the feature vector in the embodiment of the present application includes but is not limited to the following situations: it can be an artificial feature vector designed for each source data, such as a state vector after anomaly detection; it can be a simple feature vector processed by a traditional machine learning model, such as a feature vector processed by a decision tree; it can also be a complex feature vector processed by a deep learning network, such as a feature vector of an image processed by a convolutional neural network), and then obtain multiple Tokens for multiple data fragments corresponding to each source data.

[0141] During the characterization extraction process, if a source data is missing a data fragment within a time slice, in order to fill the missing data fragment, the electronic device can use the periodic characteristics of the source data to select at least one data fragment of a time slice with a periodic association from the time slice before the missing time slice, and can use the average of the tokens corresponding to the data fragments of all selected time slices as the token of the corresponding missing fragment.

[0142] Specifically, the electronic device may first divide the first source data according to a preset time slice length from a time dimension to obtain a plurality of data segments corresponding to the first source data in a plurality of time slices, wherein if the first source data has first data in a first time slice, the first data is the data segment corresponding to the first time slice;

[0143] If the first source data does not exist in the first time slice, the data segment historically corresponding to the first time slice is the data segment currently corresponding to the first time slice, wherein the first time slice is any one of multiple time slices, and the first source data is any one of the multiple source data.

[0144] For example, when the length of the time slice is 1 hour, if the data with a short period of 24 hours, a long period of 7x24 hours, and an ultra-long period of 365 days is missing between 10 and 11 am on Saturday, April 1, 2023, then three historical time segments with periodic associations with the missing time slices can be selected: 10-11 am on March 31, 2023, 10-11 am on Saturday, March 25, 2023, and 10-11 am on April 1, 2022, and the token corresponding to any of the three time segments is used as the token of the corresponding data segment of the missing time slice. Optionally, the average value of all tokens on the source data can also be used as the token of the corresponding data segment of the missing time slice.

[0145] Step S603: The electronic device fuses the multiple periodic signals and the target sequence to obtain representation information of the multi-source data.

[0146] Among them, the target sequence includes multiple subsequences corresponding to multiple source data (specifically, one source data corresponds to one subsequence. For example, the target sequence includes subsequence 1, subsequence 2 and subsequence 3, source data 1 corresponds to subsequence 1, source data 2 corresponds to subsequence 2, and source data 3 corresponds to subsequence 3), and a part of the periodic signals in the multiple periodic signals are the association relationship between different Tokens of the same source data generated according to prior knowledge (for example, there are 6 periodic signals, and a part of the 6 periodic signals are periodic signal 1, periodic signal 2 and periodic signal 3, periodic signal 1 is the association relationship between Token 1 and Token 2 of source data 1 generated according to prior knowledge, and periodic signal 2 is the association relationship between Token 3 and Token 4 of source data 2 generated according to prior knowledge). , and periodic signal 3 is the association relationship between Token5 and Token6 of source data 3 generated according to prior knowledge), another part of the periodic signals in the multiple periodic signals are the association relationship between different Tokens of different source data (for example, another part of the 6 periodic signals are periodic signal 4, periodic signal 5 and periodic signal 6, periodic signal 4 is the association relationship between Token7 of source data 1 and Token8 of source data 2 generated according to prior knowledge, periodic signal 5 is the association relationship between Token9 of source data 1 and Token10 of source data 3 generated according to prior knowledge, and periodic signal 6 is the association relationship between Token11 of source data 1 and Token12 of source data 4 generated according to prior knowledge).

[0147] Optionally, the electronic device finally obtains the representation information of the multi-source data after performing the following steps in sequence, specifically:

[0148] In step 1, the electronic device needs to comprehensively sort multiple Tokens of multiple data fragments corresponding to multiple source data to obtain a sorted target sequence.

[0149] Among them, in case 1, if the time slice lengths of multiple source data are the same, the electronic device can directly comprehensively sort the multiple Tokens of the multiple data fragments corresponding to the multiple source data. The sorting rule can be: the Token with the earlier time slice corresponds to the earlier the sorting, and the Token with the same time slice has the earlier source data to which it belongs to the earlier the sorting.

[0150] See also Figure 7 , Figure 7 is a schematic diagram of constructing a target sequence of multi-source data provided by an embodiment of the present application, such as Figure 7 As shown, from step S601, it can be seen that the multi-source data includes source data 1, source data 2, source data 3, ..., source data i, where, for example, D i (ts , t e ) represents the source data i from time t s At time t e The data between, further, X i s , e is the data segment D i (t s , t e ) The corresponding token, the X marked in the dotted box i s , e is the preset token for source data i to replace the missing source data. For example, if the multi-source data includes source data 1, source data 2 and source data 3, and the time slice lengths of source data 1, source data 2 and source data 3 are the same, source data 1 is divided into data segments D1 (t0, t2), data segments D1 (t2, t4) and data segments D1 (t4, t6), and then the above data segments D1 are tokenized, and the tokens corresponding to the data segments D1 (t0, t2) are obtained as X. 1 0，2 、The Token corresponding to the data fragment D1(t2, t4) is X 1 2，4 And the Token corresponding to the data fragment D1(t4, t6) is X 1 4，6 .

[0151] Similarly, source data 2 is divided into data segments D2(t0, t2), data segment D2(t2, t4) and data segment D2(t4, t6). After tokenizing the data segment D2, the tokens corresponding to the data segment D2(t0, t2) are X and 2 0，2 、The Token corresponding to the data fragment D2(t2, t4) is X 2 2，4 And the Token corresponding to the data segment D2 (t4, t6) is X 2 4, 6, and divide the source data 3 into data fragment D3 (t0, t2) and data fragment D3 (t4, t6), and then tokenize the above data fragment D3, and obtain the tokens corresponding to the data fragment D3 (t0, t2) as X. 3 Token corresponding to 0, 2 and data segment D3 (t4, t6) is X 3 4, 6.

[0152] It can be seen that since there is data missing in the source data 3, the missing data can be filled first (for details, please refer to the processing process for data missing described in step S601, which will not be repeated here), and the missing data segment is obtained as data segment D3 (t2, t4), and its corresponding Token is X 3 2, 4. After all data are complete and aligned, the multiple tokens of the multiple data fragments corresponding to the multiple source data are comprehensively sorted according to the sorting rules, that is, the tokens corresponding to the data fragments from time t0 to time t2 in source data 1, source data 2 and source data 3 are first arranged in the front, and then the tokens corresponding to the data fragments from time t2 to time t4 are arranged in the second place, and the tokens corresponding to the data fragments from time t4 to time t6 are arranged last. After the tokens corresponding to the data fragments of each source data are sorted in the first round according to the order of the above time slices, if the order of the time slices is the same, the tokens of the multiple tokens that belong to the earlier source data are sorted in the front (optionally, the source data of the multiple source data that has a greater impact on other source data in the multiple source data can be ranked higher). For example, among source data 1, source data 2 and source data 3, if source data 1 has the greatest impact on source data 2 and source data 3, such as the data type of source data 1 is configuration business data, the data type of source data 2 is log business data, and the data type of source data 3 is alarm business data, if the configuration business data is locally changed, it may cause the global effect of log business data and alarm business data, so source data 1 can be placed before source data 2 and source data 3). According to the above steps, after comprehensively sorting multiple tag Tokens of multiple data fragments corresponding to multiple source data (that is, constructing sequences of the obtained multiple Tokens), the target sequence (that is, a flattened multi-source data Token sequence with the same time slice length) can be obtained, so that the user's visualization effect is stronger.

[0153] Case 2: If the time slice lengths of multiple source data are different, the electronic device needs to first copy the first Token of the second source data in the second time slice to obtain multiple updated first Tokens of the second source data in the second time slice. After the multiple first Tokens are updated and the time slice lengths of the multiple source data are the same, the tokens corresponding to the updated multiple source data are comprehensively sorted according to the sorting rules to obtain the sorted target sequence.

[0154] Among them, the time period of the second time slice is the same as the time period of multiple time slices of the third source data, the content of the updated multiple first Tokens is the same as the original first Tokens, the time slices of the updated multiple first Tokens are synchronized with the multiple time slices of the third source data one by one, the third source data is the source data with the shortest time slice length among the multiple source data, and the time slice length of the second source data is greater than the time slice length of the third source data.

[0155] The following is a detailed description of the update process of multiple first tokens:

[0156] See also Figure 8 , Figure 8 is a schematic diagram of another method for constructing a target sequence of multi-source data provided by an embodiment of the present application, such as Figure 8 As shown, for example, if the multi-source data includes source data 1, source data 2 and source data 3, and the time slice lengths of source data 1, source data 2 and source data 3 are different, source data 1 is first divided into data segments D1 (t0, t2), data segments D1 (t2, t4) and data segments D1 (t4, t6), where the Token corresponding to data segment D1 (t0, t2) is X 1 0，2 、The Token corresponding to the data fragment D1(t2, t4) is X 1 2，4 And the Token corresponding to the data fragment D1(t4, t6) is X 1 4，6 . Then divide the source data 2 into data segments D2 (t0, t6), and its corresponding Token is X 2 0，6 . Then divide the source data 3 into data segments D3(t0, t1), data segment D3(t1, t2), data segment D3(t3, t4) and data segment D3(t4, t5), where the Token corresponding to data segment D3(t0, t1) is X 3 0，1 、The Token corresponding to the data fragment D3(t1, t2) is X 3 1，2 、The Token corresponding to the data segment D3(t3, t4) is X 3 3，4 The Token corresponding to the data segment D3 (t4, t5) is X 3 4，5 .

[0157] In the above process, since there are missing data segments in source data 3, it is necessary to first fill in the missing data in source data 3. The tokens marked with dotted lines are the filled tokens, which include the tokens corresponding to the data segment D3 (t2, t3). 3 2，3 TokenX corresponding to data fragment D3(t5, t6) 3 5，6 .

[0158] As can be seen from the above, the time slice granularity of source data 3 is the smallest, and source data 2 with the longest time slice length needs to have its corresponding TokenX 2 0，6 Copy 5 times, the source data 1 with the second longest slice length needs to change the X in the corresponding Token 1 0，2 , X 1 2, 4 and X 1 4，6 They are all copied once, so that the time slice where the token of the updated source data 2 is located is synchronized one-to-one with the time slice where the token of source data 3 is located, and the time slice where the token of the updated source data 1 is located is synchronized one-to-one with the time slice where the token of source data 3 is located, thereby ensuring that the number of tokens of each source data is the same, solving the heterogeneity and alignment problems faced by multi-source data.

[0159] Step 2: fuse multiple periodic signals with the sorted target sequence to obtain the representation information of multi-source data.

[0160] It should be clear that with the knowledge verified in advance, the knowledge of the next stage can be better inferred. Therefore, using prior knowledge to replace the weights in the training model with the weight parameters, edges, contours, etc. of the known trained model will have a much better effect than not using prior knowledge. Therefore, this application designs a periodic signal based on prior knowledge, converts the prior knowledge into a periodic signal that can be added to multiple Tokens, that is, uses the periodic signal to establish additional connections between multiple associated Tokens. Among them, including but not limited to the following prior knowledge conversion methods:

[0161] Method 1: If prior knowledge indicates that the first Token in the first time slice and the second Token in the second time slice are in a preset periodic state and / or similar state, the first periodic signal includes the association relationship between the first Token in the first time slice and the second Token in the second time slice.

[0162] The first Token and the second Token both belong to the first source data, and the first time slice and the second time slice are adjacent time slices.

[0163] For example, see Fig. 9 , Fig. 9 is a schematic diagram of converting prior knowledge into a periodic signal provided by an embodiment of the present application, such as Fig. 9 As shown in (a) ①, the first periodic signal can be X 1 0，2 With X 1 2，4 Between or X 1 2，4 With X 1 4，6 Between or X 3 0，2 With X 3 2，4 Between or X 3 2，4 With X 3 4，6 The relationship between them is that both of the above two tokens belong to the tokens of the same source data between adjacent time slices, or the above X 1 0，2 With X 1 2，4 It can also be a token of homologous data separated by at least one time slice.

[0164] For example, Fig. 9 As shown in ② of (a), the first periodic signal can also be X 1 0，2 With X 1 4，6 The relationship between X 1 0，2 With X 1 4，6 A token belonging to the same source data across at least one time slice, or X 1 0，2 With X 1 4，6 It can also be a token that belongs to the same source data between adjacent time slices.

[0165] Method 2: If prior knowledge indicates that the time slice length of the second source data is greater than the time slice length of the third source data, and the interval between the third Token and the fourth Token spans multiple time slices, then the second periodic signal includes the association relationship between the third Token and the fourth Token.

[0166] Among them, the third Token and the fourth Token both belong to the second source data, and the second periodic signal is used to add longer prior knowledge with periodic association.

[0167] For example, if Fig. 9 As shown in (b) ②, due to X 2 0，4 With X 2 4，8 The time slice length of source data 1 is the longest, which is greater than the time slice length of source data 1 and source data 3. Therefore, the time slice length of source data 2 needs to be updated to be the same as that of source data 1 and source data 3 (the content of source data 2 does not change). In the case where the prior knowledge is that the time slice interval of the actual Token spans multiple time slices due to the influence of the Token replication mechanism, the second periodic signal can be X 2 0，4 With X 2 4，8 The relationship between X 2 0，4 With X 2 4，8 A token that belongs to homologous data that spans at least one time slice.

[0168] Method three, the third periodic signal includes the association relationship between the fifth Token and the sixth Token, the fifth Token is any Token in the fourth source data, the sixth Token is any Token in the fifth source data, the fourth source data is the source data among multiple source data that has the greatest impact on other source data, and the fifth source data is any source data among multiple source data except the fourth source data.

[0169] If the prior knowledge is that changes to some operation and maintenance data can take effect globally, such as changes to background configurations that affect other operation and maintenance data, for example, Fig. 9 As shown in ③ of (c), if source data 1 has the greatest impact on source data 2 and source data 3, for example, the data type of source data 1 is configuration business data, the data type of source data 2 is log business data, and the data type of source data 3 is alarm business data, if the configuration business data is locally changed, it may cause the global effect of log business data and alarm business data. In this case, the third period signal can be X 1 0，2 With X 2 0，4 The relationship between them, or X 1 0，2 With X 3 0，2 The relationship between X 1 0，2 With X 20，4， and X 1 0，2 With X 3 0，2 All of them belong to the tokens of multi-source data in the same time slice.

[0170] Furthermore, the periodic signal preset in step 2 is added to the self-attention module for fusion sequence in the form of a bias matrix. By adding a preset inductive bias between multiple tokens, the operation and maintenance prior knowledge is introduced into the representation learning of multi-source data. The specific steps are as follows:

[0171] Step 1: Convert the periodic signal into a bias matrix.

[0172] It should be noted that the conversion of periodic signals includes but is not limited to the following conversion methods:

[0173] The bias matrix consists of n rows and n columns. The value of the element M(i, j) in the i-th row and j-th column of the bias matrix represents the degree of association between the i-th Token and the j-th Token in the target sequence. The larger the value of the element M(i, j), the higher the degree of association between the i-th Token and the j-th Token. n is the number of tokens in the target sequence, i is an integer between 1 and n, and j is an integer between 1 and n. For example, if the element M(i, j)∈(0,1], please refer to Fig.10 , Fig.10 is a schematic diagram of a periodic signal of a Token of homologous data between adjacent time slices provided in an embodiment of the present application, such as Fig.10 As shown, when the element M(i, j) = 1, it means that the correlation between the i-th Token and the j-th Token is the highest, that is, Fig.10 The black area shown; when the element M(i, j) = 0, it means that the correlation between the i-th Token and the j-th Token is the lowest, that is, Fig.10 The white area shown.

[0174] In an optional implementation, the value of the element M(i, j) includes but is not limited to the following methods: 1. Fixed value form: M(i, j+τ*ρ)=M(i, j); 2. Linear decreasing form: M(i, j+τ*ρ)=(1-α*τ)*M(i, j); 3. Exponential decreasing form: M(i, j+τ*ρ)=α τ *M(i,j).

[0175] Among them, ρ is the period length of the corresponding periodic signal, τ is the number of periods between two Tokens, and α is the preset attenuation coefficient.

[0176] The following is an example of the process of converting different periodic signals into corresponding bias matrices, combining the methods 1, 2, and 3 in the above step 2 respectively:

[0177] For the first periodic signal, when the conversion is the periodic signal of the token of the same source data connected between adjacent time slices, it is still the same as Fig.10 As shown, source data 1 and source data 3 use periodic signal 1 ( Fig.10 The solid line part) and periodic signal 2 ( Fig.10 The dashed line in the middle connects the corresponding tokens between adjacent time slices. For example, X 1 0，2 With X 1 2，4 Connect, X 3 0，2 With X 3 2，4 connection, where, due to X 1 0，2 With X 1 2，4 Connect, and X 3 0，2 With X 3 2，4 The connection process spans 3 tokens, so ρ of periodic signal 1 and periodic signal 2 are both 3, and the number of periods τ between two tokens is 1. The attenuation coefficient α can be set by the user according to actual needs. The embodiment of the present application does not limit the specific value of the attenuation coefficient.

[0178] For another possible situation of the first periodic signal and the second periodic signal (i.e., the periodic signal of the Token subsequence of the same source data across at least one time slice), see Fig.11 , Fig.11 is a schematic diagram of a periodic signal of a homologous data Token across time slices provided by an embodiment of the present application, such as Fig.11 As shown, in order to add prior knowledge of a period longer than the adjacent event slices, source data 1 uses periodic signal 3 ( Fig.11 The solid line in the middle) connects the tokens of the same source data across time slices. As can be seen from the second method, due to the influence of the token replication mechanism, the time slice interval of the actual token spans multiple time slices, so the source data 2 uses periodic signal 4 ( Fig.11 The solid line in the middle connects the tokens of the same source data across time slices, that is, across multiple replicated tokens. For example, X 1 0, 2 and X 1 4，6 Connect, X 2 0，4With X 2 4，8 connection, where, due to X 1 0，2 With X 1 4，6 The connection spans 6 tokens, so the ρ of periodic signal 3 is 6, and the number of interval periods τ is 2; and because X 2 0，4 With X 2 4，8 The connection process also spans 6 tokens, so the ρ of periodic signal 4 is 6 and the number of interval periods τ is 2.

[0179] For the third periodic signal (i.e., the periodic signal between tokens of different source data in the same time slice), see Fig.12 , Fig.12 This is a schematic diagram of a periodic signal between Tokens of different source data in the same time slice provided by an embodiment of the present application, such as Fig.12 As shown in the third method, source data 2 and source data 3 in the same time slice will be affected by source data 1, so the periodic signal 5 ( Fig.12 The solid line in the middle) connects the tokens of source data 1 and source data 2 in the same time slice, using periodic signal 6 ( Fig.12 The dashed line in the middle connects the tokens of source data 1 and source data 3 in the same time slice. For example, X 1 0，2 With X 2 0，4 Connect, X 1 0，2 With X 3 0，2 connection, where, due to X 1 0，2 With X 2 0，4 The connection process spans 1 Token, so the ρ of the periodic signal 5 is 1, and the number of interval cycles τ is 1; and because X 1 0，2 With X 3 0，2 The connection process spans 2 tokens, so the ρ of periodic signal 6 is 2 and the number of interval periods τ is 1.

[0180] Step 2: Add the bias matrix to the self-attention module that integrates the target sequence, calculate the association weights between different tokens in the target sequence, and determine the representation information of the multi-source data through the association weights.

[0181] Specifically, after getting N pThe bias matrix P corresponding to the periodic signal is [P1, P2, ..., P Np ], the electronic device needs to add the bias matrix P to the self-attention module for fusing the target sequence, where:

[0182] If the self-attention module includes a self-attention head, the bias matrix corresponding to the self-attention head is the weighted accumulation result of the bias matrices of each periodic signal in the multiple periodic signals (for example, it can be the weighted accumulation result of the bias matrices of the above-mentioned periodic signal 1, periodic signal 2, periodic signal 3, periodic signal 4, periodic signal 5 and periodic signal 6), wherein the bias matrix is ​​used to enhance the association relationship between multiple tokens on the basis of calculating the association relationship through the original self-attention head. The final representation fusion calculation model of multi-source data is:

[0183]

[0184] Where Q = K = V = X

[0185] Where X is the target sequence, d k is the vector dimension of the semantic vector or feature vector corresponding to each Token in the target sequence. The vector dimension of all Tokens in the target sequence is the same. μ is the preset periodic signal weight. α e is the weight coefficient corresponding to each periodic signal (for example, α1 is the weight coefficient corresponding to periodic signal 1, α2 is the weight coefficient corresponding to periodic signal 2, and α3 is the weight coefficient corresponding to periodic signal 3).

[0186] If the self-attention module includes multiple self-attention heads, the bias matrix corresponding to each of the multiple self-attention heads is the bias matrix corresponding to a periodic signal among the multiple periodic signals (for example, the bias matrix corresponding to self-attention head 1 can be the bias matrix P1 corresponding to periodic signal 1, and the bias matrix corresponding to self-attention head 2 can be the bias matrix P2 corresponding to periodic signal 2) (corresponding to (P in formula (2) e ), or each self-attention head is the weighted accumulation result of the bias matrices of each periodic signal in the same source data (for example, the bias matrix corresponding to self-attention head 1 can be the weighted accumulation result of the bias matrix P1 corresponding to the periodic signal 1 in the above source data 1 and the bias matrix P3 corresponding to the periodic signal 3) (corresponding to formula (2) Or each self-attention head is the original self-attention head without adding periodic signals (corresponding to 0 in formula (2)), then the final representation fusion calculation model of multi-source data is:

[0187] MultiHead(Q,K,V,P)=Concat(head1,...,head h )W O

[0188] head e =Attention(QW e Q ,KW e K ,VW e V ,P e )

[0189]

[0190]

[0191]

[0192] Among them, d k =d v =d m / h, each W i is the learnable parameter matrix, h is the number of self-attention heads, This is an explanation of the three variables Q, K, and V. It is also a common mathematical expression used to illustrate the vector dimensions of the three variables Q, K, and V in each dimension. m is the dimension corresponding to the fusion representation result finally output by the self-attention module, n s is the number of periodic signals on the source data s, and n is the number of tokens in the target sequence.

[0193] In this step, the self-attention head Some will calculate the correlation between multiple tokens, and the bias matrix converted from the newly added periodic signal can further enhance the correlation between multiple tokens based on the original calculated correlation, thereby improving the information extraction effect when the token sequence features of multi-source data are fused, that is, improving the final multi-source data representation capability.

[0194] In an optional embodiment, if the self-attention module includes multiple attention heads, the electronic device can also convert the periodic signal into a coding matrix based on the cyclic coding matrix REM; add the coding matrix to the self-attention module of multiple tokens that fuse multiple data fragments corresponding to each source data, calculate the association weights between different tokens in the target sequence, and finally determine the representation information of the multi-source data by adjusting the ratio of the original self-attention matrix weights and the coding matrix weights.

[0195] Finally, the result of fusion processing of multiple tokens of multi-source data by the self-attention module is regarded as the fusion representation of multi-source data and used as the model input of the downstream model. The downstream model can be a pre-trained model of multi-source data or a dedicated model for processing downstream tasks of multi-source data, including but not limited to the existing Transformer-based discriminative model and generative model.

[0196] In an embodiment of the present application, by constructing a multi-source data sequence with a periodic pattern, the heterogeneity and alignment problems faced by multi-source data are solved while retaining time information. In addition, by adding periodic signals in the representation learning process, the prior knowledge of operation and maintenance data is integrated into the training and reasoning process of the model, thereby improving the representation effect of the multi-source data fusion features.

[0197] Figure 6 The embodiment shown in the figure explains the fusion characterization process of operation and maintenance data in detail. Fig.13 , Fig.14 Let us take two specific examples.

[0198] See also Fig.13 , Fig.13 is a flow chart of an operation and maintenance data fusion representation provided by an embodiment of the present application, such as Fig.13 As shown, the details of Case 1 are as follows:

[0199] Step 11: The electronic device divides each source data in the multi-source data into data segments according to time.

[0200] Step 12: The electronic device generates a corresponding data description for each data segment using the LLM, and uses the semantic vector corresponding to the description as the Token of the corresponding data segment.

[0201] Step 13: The electronic device combines all the tokens obtained above into a flattened token sequence with periodic properties, so that the tokens of each source data are presented in a periodic interval form in the sequence.

[0202] Step 14: The electronic device designs the periodic signal based on the prior knowledge.

[0203] Step 15: The electronic device adds the preset periodic signal in the form of a bias matrix to the self-attention module for fusing the Token sequence.

[0204] Step 16: The electronic device uses the result of the fusion processing of the multi-source Token sequence by the self-attention module as the fusion representation of the multi-source data, which is used as the model input of the downstream model.

[0205] See also Fig.14 , Fig.14is another flow chart of operation and maintenance data fusion representation provided by an embodiment of the present application, such as Fig.14 As shown, the details of Case 2 are as follows:

[0206] Step 21: The electronic device divides each source data in the multi-source data into data segments according to time.

[0207] Step 22: The electronic device uses the feature extraction model of each source data to extract the feature vector of each data segment, and uses each feature vector as a token of the corresponding data segment.

[0208] Step 23: The electronic device combines all the tokens obtained above into a flattened token sequence with periodic properties, so that the tokens of each source data are presented in a periodic interval form within the sequence.

[0209] Step 24: The electronic device designs the periodic signal based on the prior knowledge.

[0210] Step 25: The electronic device adds the preset periodic signal in the form of a bias matrix to the self-attention module for fusing the Token sequence.

[0211] Step 26: The electronic device uses the result of the fusion processing of the multi-source Token sequence by the self-attention module as the fusion representation of the multi-source data, which is used as the model input of the downstream model.

[0212] It should be noted that the detailed explanations of the above steps 11 to 16 and steps 21 to 26 can be found in Figure 6 The embodiments described will not be described in detail here.

[0213] The method of the embodiment of the present application is described in detail above, and the device of the embodiment of the present application is provided below.

[0214] It can be understood that the multiple devices provided in the embodiments of the present application, such as the characterization device, include hardware structures, software modules, or a combination of hardware structures and software structures corresponding to executing each function in order to implement the functions in the above method embodiments.

[0215] Those skilled in the art should easily realize that, in combination with the units and algorithm steps of each example described in the embodiments disclosed herein, the embodiments of the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different device implementations to implement the aforementioned method embodiments in different usage scenarios, and different implementations of the device should not be considered to exceed the scope of the embodiments of the present application.

[0216] The embodiment of the present application can divide the device into functional modules. For example, each functional module can be divided according to each function, or two or more functions can be integrated into one functional module. The above integrated modules can be implemented in the form of hardware or in the form of software functional modules. It should be noted that the division of modules in the embodiment of the present application is schematic and is only a logical functional division. There may be other division methods in actual implementation.

[0217] For example, in the case of dividing the functional modules of the device in an integrated manner, the present application cites several possible processing devices.

[0218] See also Fig.15 , Fig.15 is a schematic diagram of a structure of a representation device 150 for operation and maintenance data provided in an embodiment of the present application. The representation device 150 may be Figure 2 The electronic device shown may be a device in the electronic device, such as a chip, a software module, an integrated circuit, etc. The characterization device 150 has the function of implementing the electronic device described in the embodiment of the present application. For example, the characterization device 150 includes a module or unit or means (means) corresponding to the steps involved in the electronic device described in the embodiment of the present application for the electronic device to execute, and the function or unit or means (means) can be implemented by software, or by hardware, or by hardware executing the corresponding software implementation, or by a combination of software and hardware. For details, please refer to the corresponding description in the aforementioned corresponding method embodiment.

[0219] In a possible design, the characterization device 150 includes a segmentation unit 1501, an input unit 1502, and a fusion unit 1503, wherein:

[0220] The segmentation unit 1501 is used to segment each source data in the multi-source data into multiple data segments from a time dimension, wherein the multi-source data is operation and maintenance data generated during the operation of the product, and the operation and maintenance data includes data generated by the product and / or description information of the product generated by other devices;

[0221] The input unit 1502 is used to input the multiple data segments corresponding to each source data into the data representation model for representation extraction, so as to obtain a subsequence corresponding to each source data, wherein each subsequence includes a plurality of tokens corresponding to the multiple data segments;

[0222] The fusion unit 1503 is used to fuse multiple periodic signals and a target sequence to obtain representation information of the multi-source data, wherein the target sequence includes multiple subsequences corresponding to multiple source data, a part of the multiple periodic signals are association relationships between different Tokens of the same source data generated according to prior knowledge, and another part of the multiple periodic signals are association relationships between different Tokens of different source data.

[0223] In an optional implementation, the representation information of the multi-source data is used as an input to a downstream model, wherein the downstream model is used to predict or classify the business.

[0224] In another optional implementation, in the aspect of inputting the multiple data segments corresponding to each source data into the data representation model for representation extraction to obtain the subsequence corresponding to each source data, the input unit 1502 is specifically used to:

[0225] Inputting the multiple data segments corresponding to each source data into a large language model LLM to obtain data descriptions corresponding to the multiple data segments respectively;

[0226] Generate a semantic vector corresponding to each data description, wherein the semantic vector corresponding to the first data description is the Token corresponding to the first data segment, the first data description is the data description corresponding to the first data segment, and the first data segment is any one of the multiple data segments.

[0227] In another optional implementation manner, the data description includes one or more of the following:

[0228] objective data descriptions of the plurality of data segments;

[0229] a data summary obtained after analyzing the plurality of data segments;

[0230] The default data description.

[0231] In another optional implementation, in the aspect of inputting the multiple data segments corresponding to each source data into the data representation model for representation extraction to obtain the subsequence corresponding to each source data, the input unit 1502 is specifically used to:

[0232] A feature extraction model is used to perform feature extraction on the multiple data fragments corresponding to each source data to obtain multiple feature vectors corresponding to the multiple data fragments, wherein the feature vector corresponding to the first data fragment is a Token subsequence corresponding to the first data fragment, and the first data fragment is any one of the multiple data fragments.

[0233] In another optional implementation, in fusing the multiple periodic signals and the target sequence to obtain the representation information of the multi-source data, the fusion unit 1503 is specifically used to:

[0234] Converting the periodic signal into a bias matrix;

[0235] Add the bias matrix to the self-attention module that incorporates the target sequence, and calculate the association weights between different tokens in the target sequence;

[0236] The characterization information of the multi-source data is determined by the association weights.

[0237] In another optional implementation, the self-attention module includes a self-attention head, and the bias matrix corresponding to the self-attention head is a weighted accumulation result of the bias matrices of each periodic signal in the multiple periodic signals, wherein the bias matrix is ​​used to enhance the association relationship between the multiple tokens based on the association relationship calculated by the original self-attention head, or,

[0238] The self-attention module includes multiple self-attention heads, and the bias matrix corresponding to each of the multiple self-attention heads is the bias matrix corresponding to a periodic signal among the multiple periodic signals, or each self-attention head is the weighted accumulation result of the bias matrices of each periodic signal in the same source data, or each self-attention head is the original self-attention head without adding the periodic signal.

[0239] In another optional implementation, in fusing the multiple periodic signals and the target sequence to obtain the representation information of the multi-source data, the fusion unit 1503 is specifically used to:

[0240] Converting the periodic signal into a coding matrix based on a cyclic coding matrix REM;

[0241] Adding the encoding matrix to a self-attention module that incorporates the target sequence, and calculating association weights between different tokens in the target sequence, wherein the self-attention module includes multiple attention heads;

[0242] The characterization information of the multi-source data is determined by the association weights.

[0243] In another optional implementation manner, in the aspect of dividing each source data in the multi-source data into a plurality of data segments from the time dimension, the dividing unit 1501 is specifically used to:

[0244] Divide the first source data according to a preset time slice length from a time dimension to obtain a plurality of data segments corresponding to the first source data in the plurality of time slices, wherein if the first source data has first data in a first time slice, the first data is the data segment corresponding to the first time slice;

[0245] If the first source data does not exist in the first time slice, the data segment historically corresponding to the first time slice is the data segment currently corresponding to the first time slice, wherein the first time slice is any one of the multiple time slices, and the first source data is any one of the multiple source data.

[0246] In another optional implementation, in fusing the multiple periodic signals and the target sequence to obtain the representation information of the multi-source data, the fusion unit 1503 is specifically used to:

[0247] Comprehensively sorting multiple tokens of multiple data fragments corresponding to the multiple source data to obtain a sorted target sequence, wherein the token with the earlier corresponding time slice is sorted higher, and the token with the earlier corresponding source data in the same time slice is sorted higher;

[0248] The multiple periodic signals are fused with the sorted target sequence to obtain the representation information of the multi-source data.

[0249] In another optional implementation, in the aspect of comprehensively sorting the multiple Tokens of the multiple data fragments corresponding to the multiple source data to obtain the sorted target sequence, the sorting unit is specifically used to:

[0250] Copy the first Token of the second source data in the second time slice to obtain multiple updated first Tokens of the second source data in the second time slice, wherein the time period of the second time slice is the same as the time period of the multiple time slices of the third source data, the contents of the multiple updated first Tokens are the same as the original first Tokens, the time slices of the multiple updated first Tokens are synchronized with the multiple time slices of the third source data in a one-to-one correspondence, the third source data is the source data with the shortest time slice length among the multiple source data, and the time slice length of the second source data is greater than the time slice length of the third source data;

[0251] The multiple Tokens of the multiple data fragments corresponding to the multiple source data are comprehensively sorted to obtain a sorted target sequence.

[0252] In yet another optional implementation, source data among the multiple source data that has a greater degree of influence on other source data among the multiple source data is ranked higher.

[0253] In another optional implementation, if the prior knowledge indicates that the first Token in the first time slice and the second Token in the second time slice are in a preset periodic state and / or a similar state, the first periodic signal includes an association relationship between the first Token in the first time slice and the second Token in the second time slice, wherein the first Token and the second Token both belong to the first source data, and the first periodic signal is one of the multiple periodic signals.

[0254] In another optional implementation, if the prior knowledge indicates that the time slice length of the second source data is greater than the time slice length of the third source data, and the interval between the third Token and the fourth Token spans multiple time slices, then the second periodic signal includes an association relationship between the third Token and the fourth Token, wherein the third Token and the fourth Token both belong to the second source data, and the second periodic signal is used to add prior knowledge with a longer period association.

[0255] In another optional embodiment, the third periodic signal includes an association relationship between a fifth Token and a sixth Token, the fifth Token is any one Token in the fourth source data, the sixth Token subsequence is any one Token in the fifth source data, the fourth source data is the source data among the multiple source data that has the greatest impact on other source data, and the fifth source data is any one of the multiple source data except the fourth source data.

[0256] In another optional implementation, the bias matrix includes n rows and n columns, and the value of the element M(i, j) in the i-th row and j-th column of the bias matrix represents the degree of association between the i-th Token and the j-th Token in the target sequence. The larger the value of the element M(i, j), the higher the degree of association between the i-th Token and the j-th Token. n is the number of tokens in multiple tokens of multiple data fragments corresponding to each source data, i takes integers between 1 and n, and j takes integers between 1 and n.

[0257] The embodiments of the present application and the method embodiments shown above are based on the same concept, and the technical effects they bring are also the same. For the specific principles, please refer to the description of the embodiments shown above, and no further details will be given.

[0258] In the various embodiments of the present application, unless otherwise specified or logically conflicting, the terms and / or descriptions between the various embodiments are consistent and may be referenced to each other, and the technical features in different embodiments may be combined to form new embodiments according to their inherent logical relationships.

[0259] The present application also provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, and when the instructions are executed on at least one processor, the aforementioned method for representing operation and maintenance data is implemented, for example Figure 6 The method described.

[0260] The present application also provides a computer program product, which includes computer instructions, and when executed by a computing device, implements the above-mentioned method for representing operation and maintenance data, for example Figure 6 The method described.

[0261] In the embodiments of the present application, the words "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

[0262] The "at least one" mentioned in the embodiments of the present application refers to one or more, and "plurality" refers to two or more. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can be represented by: a, b, c, (a and b), (a and c), (b and c), or (a and b and c), where a, b, c can be single or multiple. "And / or" describes the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can be represented by: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. The character " / " generally indicates that the associated objects before and after are in an "or" relationship.

[0263] Furthermore, unless otherwise specified, the ordinal numbers such as "first" and "second" used in the embodiments of the present application are used to distinguish multiple objects, and are not used to limit the order, timing, priority or importance of multiple objects. For example, the first device and the second device are only for the convenience of description, and do not represent the difference in structure, importance, etc. between the first device and the second device. In some embodiments, the first device and the second device can also be the same device.

[0264] In the above embodiments, the term "when..." can be interpreted as meaning "if..." or "after..." or "in response to determining..." or "in response to detecting...", depending on the context. The above description is only an optional embodiment of the present application and is not intended to limit the present application. Any modification, equivalent substitution, improvement, etc. made within the concept and principle of the present application shall be included in the protection scope of the present application.

[0265] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0266] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for representing operation and maintenance data, characterized in that: The method comprises: Divide each source data in the multi-source data into multiple data segments from a time dimension, wherein the multi-source data is operation and maintenance data generated during the operation of the product, and the operation and maintenance data includes data generated by the product and / or description information of the product generated by other devices; Inputting multiple data fragments corresponding to each source data into a data representation model for representation extraction to obtain a subsequence corresponding to each source data, wherein each subsequence includes multiple tokens corresponding to the multiple data fragments; Multiple periodic signals and a target sequence are fused to obtain representation information of the multi-source data, wherein the target sequence includes multiple subsequences corresponding to multiple source data, a portion of the multiple periodic signals are association relationships between different Tokens of the same source data generated according to prior knowledge, and another portion of the multiple periodic signals are association relationships between different Tokens of different source data.

2. The method according to claim 1, characterized in that: The representation information of the multi-source data is used as an input to a downstream model, wherein the downstream model is used to predict or classify business.

3. The method according to claim 1 or 2, characterized in that: The step of inputting the multiple data segments corresponding to each source data into the data representation model for representation extraction to obtain the subsequence corresponding to each source data includes: Inputting the multiple data segments corresponding to each source data into a large language model LLM to obtain data descriptions corresponding to the multiple data segments respectively; Generate a semantic vector corresponding to each data description, wherein the semantic vector corresponding to the first data description is the Token corresponding to the first data segment, the first data description is the data description corresponding to the first data segment, and the first data segment is any one of the multiple data segments.

4. The method according to claim 3, characterized in that The data description includes one or more of the following: objective data descriptions of the plurality of data segments; a data summary obtained after analyzing the plurality of data segments; The default data description.

5. The method according to claim 1 or 2, characterized in that: The step of inputting the multiple data segments corresponding to each source data into the data representation model for representation extraction to obtain the subsequence corresponding to each source data includes: A feature extraction model is used to perform feature extraction on the multiple data fragments corresponding to each source data to obtain multiple feature vectors corresponding to the multiple data fragments, wherein the feature vector corresponding to the first data fragment is the Token corresponding to the first data fragment, and the first data fragment is any one of the multiple data fragments.

6. The method according to any one of claims 1 to 5, characterized in that: The fusing of the multiple periodic signals and the target sequence to obtain the representation information of the multi-source data includes: Converting the periodic signal into a bias matrix; Add the bias matrix to the self-attention module that incorporates the target sequence, and calculate the association weights between different tokens in the target sequence; The characterization information of the multi-source data is determined by the association weights.

7. The method according to claim 6, characterized in that: The self-attention module includes a self-attention head, and the bias matrix corresponding to the self-attention head is a weighted accumulation result of the bias matrices of each periodic signal in the multiple periodic signals, wherein the bias matrix is ​​used to enhance the association relationship between the multiple tokens based on the association relationship calculated by the original self-attention head, or, The self-attention module includes multiple self-attention heads, and the bias matrix corresponding to each of the multiple self-attention heads is the bias matrix corresponding to a periodic signal among the multiple periodic signals, or each self-attention head is the weighted accumulation result of the bias matrices of each periodic signal in the same source data, or each self-attention head is the original self-attention head without adding the periodic signal.

8. The method according to any one of claims 1 to 5, characterized in that: The step of fusing the periodic signal and the target sequence to obtain the representation information of the multi-source data includes: Converting the periodic signal into a coding matrix based on a cyclic coding matrix REM; Adding the encoding matrix to a self-attention module that incorporates the target sequence, and calculating association weights between different tokens in the target sequence, wherein the self-attention module includes multiple attention heads; The characterization information of the multi-source data is determined by the association weights.

9. The method according to claim 5, characterized in that The method of dividing each source data in the multi-source data into multiple data segments from the time dimension includes: Divide the first source data according to a preset time slice length from a time dimension to obtain a plurality of data segments corresponding to the first source data in the plurality of time slices, wherein if the first source data has first data in a first time slice, the first data is the data segment corresponding to the first time slice; If the first source data does not exist in the first time slice, the data segment historically corresponding to the first time slice is the data segment currently corresponding to the first time slice, wherein the first time slice is any one of the multiple time slices, and the first source data is any one of the multiple source data.

10. The method according to any one of claims 1 to 9, characterized in that: The fusing of the multiple periodic signals and the target sequence to obtain the representation information of the multi-source data includes: Comprehensively sorting multiple tokens of multiple data fragments corresponding to the multiple source data to obtain a sorted target sequence, wherein the token with the earlier corresponding time slice is sorted higher, and the token with the earlier corresponding source data in the same time slice is sorted higher; The multiple periodic signals are fused with the sorted target sequence to obtain the representation information of the multi-source data.

11. The method according to claim 10, characterized in that The step of comprehensively sorting the multiple Tokens of the multiple data fragments corresponding to the multiple source data to obtain a sorted target sequence includes: Copy the first Token of the second source data in the second time slice to obtain multiple updated first Tokens of the second source data in the second time slice, wherein the time period of the second time slice is the same as the time period of the multiple time slices of the third source data, the contents of the multiple updated first Tokens are the same as the original first Tokens, the time slices of the multiple updated first Tokens are synchronized with the multiple time slices of the third source data in a one-to-one correspondence, the third source data is the source data with the shortest time slice length among the multiple source data, and the time slice length of the second source data is greater than the time slice length of the third source data; The multiple Tokens of the multiple data fragments corresponding to the multiple source data are comprehensively sorted to obtain a sorted target sequence.

12. The method according to claim 10 or 11, characterized in that: The source data among the multiple source data that has a greater influence on other source data among the multiple source data is ranked higher.

13. The method according to claim 12, characterized in that: If the prior knowledge indicates that the first Token in the first time slice and the second Token in the second time slice are in a preset periodic state and / or a similar state, the first periodic signal includes the association relationship between the first Token in the first time slice and the second Token in the second time slice, wherein the first Token and the second Token both belong to the first source data, and the first periodic signal is one of the multiple periodic signals.

14. The method according to claim 12 or 13, characterized in that: If the prior knowledge indicates that the time slice length of the second source data is greater than the time slice length of the third source data, and the interval between the third Token and the fourth Token spans multiple time slices, then the second periodic signal includes the association relationship between the third Token and the fourth Token, wherein the third Token and the fourth Token both belong to the second source data, and the second periodic signal is used to add prior knowledge with a longer period association.

15. The method according to any one of claims 12 to 14, characterized in that: The third periodic signal includes an association relationship between a fifth Token and a sixth Token, the fifth Token is any one of the fourth source data, the sixth Token is any one of the fifth source data, the fourth source data is the source data that has the greatest impact on other source data among the multiple source data, and the fifth source data is any one of the multiple source data except the fourth source data.

16. The method according to claim 6, characterized in that: The bias matrix includes n rows and n columns. The value of the element M(i, j) in the i-th row and j-th column of the bias matrix represents the degree of association between the i-th Token and the j-th Token in the target sequence. The larger the value of the element M(i, j), the higher the degree of association between the i-th Token and the j-th Token. n is the number of tokens in the target sequence, i is an integer between 1 and n, and j is an integer between 1 and n.

17. A device for representing operation and maintenance data, characterized in that: The characterization device includes a segmentation unit, an input unit and a fusion unit, wherein: The segmentation unit is used to segment each source data in the multi-source data into multiple data segments from a time dimension, wherein the multi-source data is operation and maintenance data generated during the operation of the product, and the operation and maintenance data includes data generated by the product and / or description information of the product generated by other devices; The input unit is used to input the multiple data segments corresponding to each source data into the data representation model for representation extraction, so as to obtain a subsequence corresponding to each source data, wherein each subsequence includes a plurality of tokens corresponding to the multiple data segments; The fusion unit is used to fuse multiple periodic signals and a target sequence to obtain representation information of the multi-source data, wherein the target sequence includes multiple subsequences corresponding to multiple source data, a part of the multiple periodic signals are association relationships between different Tokens of the same source data generated according to prior knowledge, and another part of the multiple periodic signals are association relationships between different Tokens of different source data.

18. The device according to claim 17, characterized in that: The representation information of the multi-source data is used as an input to a downstream model, wherein the downstream model is used to predict or classify business.

19. The device according to claim 17 or 18, characterized in that In the aspect of inputting the multiple data segments corresponding to each source data into the data representation model for representation extraction to obtain the subsequence corresponding to each source data, the input unit is specifically used to: Inputting the multiple data segments corresponding to each source data into a large language model LLM to obtain data descriptions corresponding to the multiple data segments respectively; Generate a semantic vector corresponding to each data description, wherein the semantic vector corresponding to the first data description is the Token corresponding to the first data segment, the first data description is the data description corresponding to the first data segment, and the first data segment is any one of the multiple data segments.

20. The device according to claim 19, characterized in that The data description includes one or more of the following: objective data descriptions of the plurality of data segments; a data summary obtained after analyzing the plurality of data segments; The default data description.

21. The device according to claim 17 or 18, characterized in that In the aspect of inputting the multiple data segments corresponding to each source data into the data representation model for representation extraction to obtain the subsequence corresponding to each source data, the input unit is specifically used to: A feature extraction model is used to perform feature extraction on the multiple data fragments corresponding to each source data to obtain multiple feature vectors corresponding to the multiple data fragments, wherein the feature vector corresponding to the first data fragment is the Token corresponding to the first data fragment, and the first data fragment is any one of the multiple data fragments.

22. The device according to any one of claims 17 to 21, characterized in that In the aspect of fusing the multiple periodic signals and the target sequence to obtain the representation information of the multi-source data, the fusion unit is specifically used for: Converting the periodic signal into a bias matrix; Add the bias matrix to the self-attention module that incorporates the target sequence, and calculate the association weights between different tokens in the target sequence; The characterization information of the multi-source data is determined by the association weights.

23. The device according to claim 22, characterized in that: The self-attention module includes a self-attention head, and the bias matrix corresponding to the self-attention head is a weighted accumulation result of the bias matrices of each periodic signal in the multiple periodic signals, wherein the bias matrix is ​​used to enhance the association relationship between the multiple tokens based on the association relationship calculated by the original self-attention head, or, The self-attention module includes multiple self-attention heads, and the bias matrix corresponding to each of the multiple self-attention heads is the bias matrix corresponding to a periodic signal among the multiple periodic signals, or each self-attention head is the weighted accumulation result of the bias matrices of each periodic signal in the same source data, or each self-attention head is the original self-attention head without adding the periodic signal.

24. The device according to any one of claims 17 to 21, characterized in that In the aspect of fusing the multiple periodic signals and the target sequence to obtain the representation information of the multi-source data, the fusion unit is specifically used for: Converting the periodic signal into a coding matrix based on a cyclic coding matrix REM; Adding the encoding matrix to a self-attention module that incorporates the target sequence, and calculating association weights between different tokens in the target sequence, wherein the self-attention module includes multiple attention heads; The characterization information of the multi-source data is determined by the association weights.

25. The device according to claim 21, characterized in that In the aspect of dividing each source data in the multi-source data into a plurality of data segments from the time dimension, the dividing unit is specifically used for: Divide the first source data according to a preset time slice length from a time dimension to obtain a plurality of data segments corresponding to the first source data in the plurality of time slices, wherein if the first source data has first data in a first time slice, the first data is the data segment corresponding to the first time slice; If the first source data does not exist in the first time slice, the data segment historically corresponding to the first time slice is the data segment currently corresponding to the first time slice, wherein the first time slice is any one of the multiple time slices, and the first source data is any one of the multiple source data.

26. The device according to any one of claims 17 to 25, characterized in that In the aspect of fusing the multiple periodic signals and the target sequence to obtain the representation information of the multi-source data, the fusion unit is specifically used for: Comprehensively sorting multiple tokens of multiple data fragments corresponding to the multiple source data to obtain a sorted target sequence, wherein the token with the earlier corresponding time slice is sorted higher, and the token with the earlier corresponding source data in the same time slice is sorted higher; The multiple periodic signals are fused with the sorted target sequence to obtain the representation information of the multi-source data.

27. The device according to claim 26, characterized in that In the aspect of comprehensively sorting the multiple Tokens of the multiple data fragments corresponding to the multiple source data to obtain the sorted target sequence, the sorting unit is specifically used to: Copy the first Token of the second source data in the second time slice to obtain multiple updated first Tokens of the second source data in the second time slice, wherein the time period of the second time slice is the same as the time period of the multiple time slices of the third source data, the contents of the multiple updated first Tokens are the same as the original first Tokens, the time slices of the multiple updated first Tokens are synchronized with the multiple time slices of the third source data in a one-to-one correspondence, the third source data is the source data with the shortest time slice length among the multiple source data, and the time slice length of the second source data is greater than the time slice length of the third source data; The multiple Tokens of the multiple data fragments corresponding to the multiple source data are comprehensively sorted to obtain a sorted target sequence.

28. The device according to claim 26 or 27, characterized in that The source data among the multiple source data that has a greater influence on other source data among the multiple source data is ranked higher.

29. The device according to claim 28, characterized in that: If the prior knowledge indicates that the first Token in the first time slice and the second Token in the second time slice are in a preset periodic state and / or a similar state, the first periodic signal includes the association relationship between the first Token in the first time slice and the second Token in the second time slice, wherein the first Token and the second Token both belong to the first source data, and the first periodic signal is one of the multiple periodic signals.

30. The device according to claim 28 or 29, characterized in that: If the prior knowledge indicates that the time slice length of the second source data is greater than the time slice length of the third source data, and the interval between the third Token and the fourth Token spans multiple time slices, then the second periodic signal includes the association relationship between the third Token and the fourth Token, wherein the third Token and the fourth Token both belong to the second source data, and the second periodic signal is used to add prior knowledge with a longer period association.

31. The device according to any one of claims 28 to 30, characterized in that: The third periodic signal includes an association relationship between a fifth Token and a sixth Token, the fifth Token is any one Token in the fourth source data, the sixth Token subsequence is any one Token in the fifth source data, the fourth source data is the source data among the multiple source data that has the greatest impact on other source data, and the fifth source data is any one source data among the multiple source data except the fourth source data.

32. The device according to claim 22, characterized in that: The bias matrix includes n rows and n columns. The value of the element M(i, j) in the i-th row and j-th column of the bias matrix represents the degree of association between the i-th Token and the j-th Token in the target sequence. The larger the value of the element M(i, j), the higher the degree of association between the i-th Token and the j-th Token. n is the number of tokens in the multiple tokens of the multiple data fragments corresponding to each source data. i takes integers between 1 and n in turn, and j takes integers between 1 and n in turn.

33. An electronic device, characterized in that: The electronic device comprises at least one processor and a communication interface; The communication interface is used to receive and / or send data, and / or the communication interface is used to provide input and / or output for the processor; The at least one processor is configured to implement the method according to any one of claims 1 to 16.

34. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed on at least one processor, the method according to any one of claims 1 to 16 is implemented.

Images