Artificial intelligence-based terminal device lightweight operation and maintenance method

Abstract

The application discloses a terminal equipment light-weight operation and maintenance method based on artificial intelligence, relates to the technical field of terminal equipment light-weight operation and maintenance, and comprises the following steps: in the case that state continuation writing data is missing, extracting a state sequence after hot start from the running state record in the current period, constructing a hot start residual piece mutation mapping tensor, and determining a running state change piece through a field stability score and a boundary disturbance gradient model; inputting the running state change piece into a credibility mapping network, generating a credible mapping vector through calculation of a cross-field residual value smoothing factor and a local mutation mode alignment rate, and judging whether the operation and maintenance has credibility based on the credible mapping vector. The application solves the problem that operation and maintenance judgment is inaccurate due to missing state data after terminal equipment hot start, and realizes accurate quantification and dynamic regulation and control of operation and maintenance credibility under the condition that data is incomplete.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of terminal device lightweight operation and maintenance, and particularly relates to a terminal device lightweight operation and maintenance method based on artificial intelligence. BACKGROUND

[0002] The terminal device lightweight operation and maintenance based on artificial intelligence refers to realizing automatic perception, analysis, prediction and maintenance of the running state of distributed terminal devices (such as intelligent terminals, IoT devices, edge computing nodes, etc.) by using AI technology, and under the condition of limited resources, significantly reducing operation and maintenance costs and system resource consumption through model compression, edge deployment, autonomous decision-making and other ways. The existing lightweight operation and maintenance technology mainly relies on machine learning, deep learning and expert system technology in artificial intelligence, realizes fast response and partial autonomy by training operation and maintenance decision-making models (such as device anomaly identification, fault early warning, load prediction models) in the cloud, and then downlinking these models to the terminal after lightweight processing in the ways of quantization, pruning and distillation, and then running locally. The whole lightweight operation and maintenance process usually includes four key links: first, the running state data collection and preprocessing of terminal devices, which guarantees data quality and input effectiveness; second, training operation and maintenance models and generating strategies based on massive data in the cloud; third, deploying AI models to terminals through model optimization and compression technology to realize edge intelligent operation and maintenance; fourth, the terminal device makes local judgments based on the model, and supports the cloud to complete automatic decision-making and task execution, such as remote upgrade, load balancing, anomaly recovery and other operations. This technology improves the operation and maintenance efficiency and intelligent level of terminal devices, reduces the dependence on network resources and manual operation and maintenance, and is one of the key technical paths to realize large-scale terminal intelligent management.

[0003] The existing technology has the following deficiencies:

[0004] In the existing artificial intelligence-based terminal device lightweight operation and maintenance technology, in order to reduce resource overhead, a hot start and state continuation mechanism is usually adopted, that is, after the terminal device is hot restarted, the lightweight AI model compares the last cached operation and maintenance state with the current collected running state change segment to realize rapid abnormality judgment. However, during the device hot start process, due to factors such as RAM cache loss, file system not mounted or scheduling queue not completed recovery, the historical operation and maintenance state data originally used for inference reference may be partially missing, resulting in incomplete state continuation. In this case, in order to meet the lightweight design, the AI model only makes operation and maintenance judgment based on the current incomplete state change segment, and cannot reconstruct the complete behavior trajectory of the device running. Therefore, the existing artificial intelligence-based terminal device lightweight operation and maintenance technology cannot accurately judge whether the operation and maintenance judgment executed according to the running state change segment of the terminal device after the hot start state continuation data is missing has credibility. This problem will directly lead to the false judgment of the model output "normal" state, missing the fault warning or repair strategy that should be triggered, and then covering up problems such as memory leakage, I / O blocking, signal processing delay and the like which gradually worsen, and finally causing device downtime, communication interruption or service unavailability and other serious consequences during the continuous running of the system.

[0005] The above information disclosed in the background section is only used to enhance the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] The purpose of the present application is to provide an artificial intelligence-based terminal device lightweight operation and maintenance method to solve the problems in the background.

[0007] In order to achieve the above-mentioned purpose, the present application provides the following technical solution: an artificial intelligence-based terminal device lightweight operation and maintenance method, specifically comprising the following steps:

[0008] S1, collecting state information of the terminal device after completing hot start, comparing the operation and maintenance data fields before and after restart through field-level hash check factors, determining whether the operation and maintenance data fields are missing by combining the timestamp mapping diagram, and identifying whether the terminal device has a state continuation data missing condition;

[0009] S2, in the case of identifying the state continuation data missing condition, extracting the state sequence after hot start from the running state record in the current period, constructing a hot start residual piece mutation mapping tensor, and determining the running state change segment through a field stability score and a boundary disturbance gradient model;

[0010] S3, input the running state change segment into the credibility mapping network, generate a credibility mapping vector by calculating the cross-field residual value smoothing factor and the local mutation mode alignment rate, and judge whether the operation and maintenance has credibility based on the credibility mapping vector;

[0011] S4, perform a non-linear double threshold mapping operation according to the mapping position of the credibility mapping vector in the credibility space, divide the credibility judgment result into multiple levels, and correspond to different reasoning path control conditions;

[0012] S5, continuously record the deviation data between the credibility judgment result after the multi-period hot start and the actual operation and maintenance state, construct a time series residual inversion tensor, update the credibility mapping vector generation logic through a Bayesian inference model, and realize dynamic regulation and control of the credibility judgment process.

[0013] Preferably, S1 specifically comprises the following steps:

[0014] S101, collect the state information of the terminal device after completing the hot start, read the current values of the preset operation and maintenance data field set by executing the field extraction command, and generate a unique field identifier and a corresponding timestamp for each field to construct the current state information structure;

[0015] S102, match the operation and maintenance data field set saved before restart with the operation and maintenance data field set collected after hot start one by one through the field-level hash check factor, generate a field hash offset matrix, and calculate the hash consistency index of each field;

[0016] S103, combine the field hash consistency index and the time sequence correspondence of the field timestamp in the timestamp mapping diagram to construct a continuity search path, match the time adjacency of the field with the hash consistency threshold, identify the break position in the continuous field chain, and determine whether there is data field loss by judging the ratio of the number of broken fields to the total number of fields. When the ratio exceeds the preset loss threshold, it is determined that the terminal device has state continuation data loss.

[0017] Preferably, S102 specifically comprises:

[0018] Call the field hash constructor to perform hash processing on each field in the operation and maintenance data field set saved before restart and the operation and maintenance data field set collected after hot start, respectively, generate a field-level hash check factor based on the field content, collection order and data type;

[0019] Match the hash check factors of the corresponding fields in the field set before restart and the field set after hot start one by one, calculate the hash difference between each pair of fields, construct a field hash offset matrix, and use it to express the consistency change of the field value before and after hot start;

[0020] The field hash offset matrix is traversed, the normalized residual of the hash difference value of each field is calculated, and a hash consistency index is generated by combining the field data length and the collection interval, which is used to represent the similarity score result of the current field between two time periods.

[0021] Preferably, S2 specifically comprises the following steps:

[0022] S201, in the case of identifying the missing state continuation data, extracting the state sequence after hot start from the running state record in the current period, selecting a plurality of continuous operation and maintenance data samples adjacent to the hot start timestamp from the continuous state collection stream by locating the time node of hot start completion, and constructing a triple sequence containing field identification, field value and field timestamp for forming the state sequence after hot start;

[0023] S202, constructing a hot start residual piece mutation mapping tensor based on the state sequence after hot start, organizing the triple sequence into a three-dimensional tensor structure according to the field dimension, time dimension and value jump amplitude dimension, and calculating the value difference sequence and disturbance weight coefficient between adjacent time points for each field, which is used to describe the mutation trend and disturbance characteristics of the field in the early stage of hot start;

[0024] S203, performing field stability scoring operation on the hot start residual piece mutation mapping tensor, and calling the boundary disturbance gradient model to respectively fit and analyze the disturbance continuity and disturbance gradient fluctuation rate of the field in the time dimension, and locating the disturbance edge critical point according to the field stability scoring result and the boundary gradient change rate to determine the time boundary range of the running state change piece.

[0025] Preferably, S202 specifically comprises:

[0026] The triplets corresponding to each field in the state sequence after hot start are classified according to the field identification, the values of the same field at consecutive time stamps are constructed into time sequences, and all field corresponding time sequences are initially arranged according to the field dimension and the time dimension, which is used to construct the basic structure of the three-dimensional tensor;

[0027] The value difference between adjacent time points in the time sequence of each field is calculated to form a jump amplitude sequence, and each difference is normalized according to the field collection interval, unit change rate and historical disturbance fluctuation degree to generate the value jump amplitude dimension, which is used to complete the value jump amplitude dimension of the three-dimensional tensor;

[0028] Based on the combination characteristics of each field in the time dimension and the value jump amplitude dimension in the three-dimensional tensor, a disturbance weight coefficient sequence is constructed for each field, and the disturbance weight coefficient is generated according to the jump frequency, mutation persistence and jump amplitude standard deviation, which is used to quantify the mutation trend and disturbance characteristics of the field in the early stage of hot start.

[0029] Preferably, S3 specifically comprises the following steps:

[0030] S301, arrange the running state change segment in field order and time order into a multi-channel input structure, vectorize the field value change trajectory using a feature encoder, and maintain the sequence information of the field in the time dimension through a position encoding mechanism, input the encoding result into the credibility mapping network for initializing the credibility calculation process;

[0031] S302, in the credibility mapping network, calculate the cross-field residual value smoothing factor based on the variation trend of the residual value between different fields, the residual value smoothing factor reflects the fluctuation consistency of the field residual in the continuous time slice, and the disturbance pattern of the mutation section in the change segment is extracted, and compared with the typical abnormal pattern marked in the historical training set, the similarity degree of the current disturbance pattern and the abnormal pattern is evaluated using the alignment rate, to generate a credibility mapping vector;

[0032] S303, based on the vector distribution position of the credibility mapping vector in the pre-defined credibility space, use the distance function and multi-class classifier to calculate the distance relationship between the center point of each credibility level, and judge whether the operation and maintenance judgment corresponding to the running state change segment has credibility. When the credibility mapping vector deviates from the high credibility area and falls into the low credibility distribution cluster, it is determined that the operation and maintenance judgment does not have credibility.

[0033] Preferably, S302 specifically comprises

[0034] Extract the numerical value change sequence of each field in the running state change segment in the continuous time slice, calculate the residual sequence of each field at adjacent time points, and construct a residual trend graph based on the residual mean value variation trend of the same time slice between fields. By analyzing the synchronization degree of the residual change of each field in the trend graph, a cross-field residual value smoothing factor reflecting the fluctuation consistency of the residual between fields is generated;

[0035] After generating the cross-field residual value smoothing factor, the high-amplitude mutation position in the running state change segment is labeled, the disturbance intensity, duration and fluctuation frequency of the mutation section are extracted, and a disturbance pattern description vector is constructed to express the disturbance structure characteristics of the mutation segment;

[0036] Compare the disturbance pattern description vector with the standard abnormal disturbance vector in the pre-trained abnormal pattern feature library through multi-dimensional feature comparison, calculate the alignment rate of the disturbance pattern and the historical typical abnormal pattern, and the alignment rate is used to evaluate the consistency degree of the current mutation structure and the known abnormal behavior. Combine the cross-field residual value smoothing factor to construct a credibility mapping vector as the basis for credibility judgment.

[0037] Preferably, S4 specifically comprises:

[0038] According to the mapping position of the trusted mapping vector in the trusted space, a nonlinear double-threshold mapping operation is performed, preset trusted upper and lower thresholds are constructed, a segmented activation function is constructed to map the position of the trusted mapping vector, and the response sensitivity of the mapping result to the boundary region change is enhanced through nonlinear distance compression between the vector and the double-threshold boundary.

[0039] According to the result of the nonlinear double-threshold mapping operation, the grade interval to which the trustworthiness judgment result belongs is determined, the mapped trusted mapping vector is divided into a high-trustworthiness region, a medium-trustworthiness region and a low-trustworthiness region, and an independent trustworthiness grade label is configured for each grade interval, which is used to support the judgment basis of a multi-grade operation and maintenance response strategy.

[0040] Based on the divided trustworthiness judgment result grade, different inference path control conditions are correspondingly matched, the inference path control conditions include field weight adjustment parameters, abnormal inference model calling identification and decision output confidence lower threshold, so as to ensure that the operation and maintenance judgment is completed through an adaptive operation mechanism to complete a differentiated decision control process under different trustworthiness grades.

[0041] Preferably, S5 specifically includes:

[0042] The offset data between the trustworthiness judgment result after the hot start of multiple periods and the actual operation and maintenance state is continuously recorded, the trusted mapping vector and the corresponding actual operation and maintenance state label in each period are obtained, the trustworthiness offset residual of each period is calculated based on the field dimension and the time dimension, and a multi-period trustworthiness offset record set is formed.

[0043] According to the multi-period trustworthiness offset record set, a time series residual inversion tensor is constructed, the trustworthiness offset residual is arranged in a period order, and field identification dimension, period index dimension and residual amplitude dimension are introduced to generate a three-dimensional tensor structure, which is used to express the error distribution characteristics of the trustworthiness judgment result after the hot start with the period change.

[0044] A Bayesian inference model is used to dynamically probabilistically model the time series residual inversion tensor, the prior probability distribution is updated based on the trustworthiness offset trend of each field in different periods in the tensor, and the residual value smoothing factor and abnormal alignment rate mapping parameter in the trusted mapping vector generation logic are corrected in combination with the posterior inference logic, so as to realize dynamic regulation and control of the trustworthiness judgment process.

[0045] In the above technical solution, the technical effects and advantages provided by the present application are as follows:

[0046] 1、The application can still realize deep modeling and credibility determination of the current running state change segment under the condition of state continuation data missing after terminal equipment hot start, thereby effectively avoiding the misjudgment problem caused by data defects. Through field-level hash check and timestamp mapping mechanism, the missing paragraph in the state data chain can be accurately identified, and a multi-dimensional mutation mapping tensor is constructed based on the state sequence after hot start, and field stability score and boundary disturbance gradient analysis are further introduced to identify the key running state change segment, thereby effectively improving the utilization efficiency and sensitivity of the operation and maintenance data under the incomplete state. At the same time, the system introduces a credibility mapping network, generates a credibility mapping vector through the alignment rate calculation of the cross-field residual value smoothing factor and the local disturbance mode, not only realizes the quantitative expression of the credibility of the current state judgment result of the equipment, but also lays a foundation for the differential reasoning control path.

[0047] 2、The application grades the credibility result through a nonlinear double threshold mapping mechanism, and flexibly adapts the reasoning strategy of different credibility levels by combining field weight adjustment, abnormal model calling and decision confidence threshold, so as to realize differential response control of complex scenes such as "highly credible normal", "medium credible and need to review", "low credible high risk" and the like. In addition, the system constructs a time series residual inversion tensor, combines Bayesian reasoning logic, dynamically tracks the deviation trend between the multi-period credibility judgment and the actual state, so as to realize dynamic correction and adaptive evolution of the credibility mapping logic, and greatly improves the robustness and continuous effectiveness of the model under long-term running conditions. On the basis of ensuring the lightweight of the model, the reliability and risk perception ability of the operation and maintenance judgment are significantly enhanced, and a feasible and highly robust technical path is provided for intelligent operation and maintenance in the resource-constrained terminal scene. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0049] Figure 1 The flowchart of the terminal equipment lightweight operation and maintenance method based on artificial intelligence of the present application. DETAILED DESCRIPTION

[0050] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the inventive aspects to those skilled in the art.

[0051] The application provides a terminal device lightweight operation and maintenance method based on artificial intelligence as shown in the formula Figure 1 The application provides a terminal device lightweight operation and maintenance method based on artificial intelligence as shown in the formula

[0052] S1, collecting state information of the terminal device after completing hot start, comparing the operation and maintenance data fields before and after restart through a field-level hash check factor, determining whether the operation and maintenance data fields are missing in combination with a timestamp mapping diagram, and identifying whether the terminal device has a state continuation data missing condition;

[0053] In this embodiment, S1 specifically includes the following steps:

[0054] S101, collecting state information of the terminal device after completing hot start, reading current values of a preset operation and maintenance data field set by executing a field extraction command, and generating a unique field identifier and a corresponding timestamp for each field, which are used to construct a current state information structure;

[0055] After the terminal device completes hot start, the key to collecting current state information lies in the directional extraction of data associated with operation and maintenance in the device running environment. The field extraction command is an operation and maintenance data calling action triggered by a software instruction, which is used to obtain data fields with explicit structured identifiers from an operating system, a driver cache, a kernel log, a task scheduling queue, a hardware sensor interface or a data collection interface embedded in the device. The preset operation and maintenance data field set is a set of high-correlation indexes determined through expert experience, historical data analysis and model training process, and usually includes CPU usage, memory occupation, thread state, IO read-write rate, cache hit rate, network packet loss rate, temperature sensor reading, voltage stability identifier and other key running dimensions. In the process of executing the field extraction command, a data pulling operation is performed on each target field, and a unique field identifier is assigned to each field immediately after the reading is completed. The identifier usually includes a field name, a data type mark, a collection node number and other information. At the same time, a high-precision timestamp is generated for each field, which is used to represent the collection time sequence position of the field data. The above collection results are uniformly constructed into a current state information structure, that is, a data collection body with field identifier as index, timestamp as sequence and field value as content is formed. This method can maximize the timeliness and structural consistency of data collection, facilitate subsequent accurate comparison with the state before restart, effectively avoid unnecessary system resource consumption by controlling the range of the field set, and improve the lightweight operation and maintenance efficiency.

[0056] S102, one-to-one matching the operation and maintenance data field set saved before restart with the operation and maintenance data field set collected after hot start through a field-level hash check factor, generating a field hash offset matrix, and calculating a hash consistency index of each field;

[0057] S103, in combination with the field hash consistency index and the time sequence correspondence of the field timestamp in the timestamp mapping diagram, a continuity search path is constructed, the time adjacency of the fields is matched with the hash consistency threshold, the breaking position in the continuous field chain is identified, and whether data field loss occurs is judged according to the ratio of the number of broken fields to the total number of fields. When the ratio exceeds a preset loss threshold, it is determined that the terminal device has a state continuation data loss condition.

[0058] After the terminal device completes the hot start, in order to identify whether there is a state continuation data loss condition, the integrity of the field chain can be analyzed by constructing a continuity search path. The specific implementation includes: first, using the timestamp information of each operation and maintenance data field, a timestamp mapping diagram is constructed to indicate the distribution position of each field on the time axis to form a time sequence structure atlas; then, in combination with the hash consistency index of each field, the fields are sorted according to time adjacency, and whether the hash consistency index of each field in the field sequence formed after sorting is higher than the set consistency threshold is matched. If a plurality of adjacent fields in succession have low hash consistency index but are adjacent in time, it means that there is an obvious fault in the state continuity of these fields. At this time, the sequence can be marked as a "breaking position". After marking all the breaking positions, the number of corresponding fields is counted, and a ratio is calculated with the total number of complete field sets to obtain a state data continuity loss ratio. When the ratio exceeds a system-set preset loss threshold, it can be judged that the terminal device has a state continuation data loss problem after hot start, for example, if 35 fields are identified as broken fields in 100 fields and the threshold is 30%, it is confirmed that there is a serious state breaking risk.

[0059] The hash consistency index is a measure of the consistency of the field content, reflecting the preservation degree of the field in the semantic and numerical levels; the timestamp mapping diagram strengthens the time sequence relationship between the fields through the time axis dimension, so that the adjacency relationship between the fields can be explicitly expressed. The continuity search path is a comprehensive expression of time and content continuity, which is used to find invisible data breaks. The preset loss threshold is a tolerance upper limit determined by historical training samples or expert experience, which is used to distinguish between slight data jitter and serious state breaking, and is usually set according to the type of terminal device, operation and maintenance scene and abnormal recovery tolerance. The threshold cannot be too high to miss the real breaking, nor too low to misjudge the normal state fluctuation, which is one of the core parameters to realize the credibility evaluation of operation and maintenance judgment. The whole process combines time information and hash structure information to quantitatively judge the data integrity after hot start in an interpretable way, which guarantees the reliability and accuracy of the artificial intelligence model in the lightweight operation and maintenance scene.

[0060] In this embodiment, S102 specifically includes:

[0061] The field hash constructor is called to perform hash processing on each field in the set of operation and maintenance data saved before restart and the set of operation and maintenance data collected after hot start respectively, and a field-level hash check factor is generated based on the field content, collection sequence and data type;

[0062] In the data consistency check process of terminal device lightweight operation and maintenance, hash processing needs to be performed on the set of operation and maintenance data saved before restart and the set of operation and maintenance data collected after hot start, so as to perform field-level comparison subsequently. To achieve this purpose, the field hash constructor can be called to perform weighted mapping processing on each field. The constructor is a kind of algorithm component specially used for structured data fast consistency check, and the input parameters include the original numerical value of the field, the content structure of the field (such as single value, array, nested object, etc.), the collection sequence index, and the data type code of the field. In the execution process, the constructor will first perform standardization operation on the field content, and convert different data types into unified format comparison units, such as processing Boolean, enumeration, integer, floating point type data through normalization, encoding mapping, format correction and other methods. Then, combined with the collection sequence information and the position index of the field in the set, the structural features are extracted through convolution mapping method to further improve the distinguishability of the field content. In order to enhance the uniqueness and anti-collision ability of the hash factor, the data type factor also needs to be introduced to participate in the disturbance calculation, so that the same field generates different hash results even if the values are the same. Finally, the generated field-level hash check factor can be used to judge the real difference of the field before and after restart, effectively avoiding the situation that the field name is consistent but the value is changed and is misjudged as consistent. The method has the advantages of high speed, low resource occupation and scalability, and is especially suitable for fast state synchronization scene in embedded or edge device, which helps to improve the accuracy of terminal device operation and maintenance judgment under the driving of artificial intelligence.

[0063] The hash check factors of the corresponding fields in the set of fields before restart and the set of fields after hot start are matched one by one, the hash difference between each pair of fields is calculated, and a field hash offset matrix is constructed to express the consistency change of the field value before and after hot start;

[0064] In the consistency analysis process of the operation and maintenance state after the terminal device hot start, in order to evaluate whether the field value changes critically before and after the hot start, the field set in the two time periods needs to be matched with the hash factor and the difference is calculated. Specifically, first, according to the unique identifier of the field, each field in the field set before the restart and the field set after the hot start is matched one by one to ensure that the fields with the same function semantics are correctly associated. For example, if the field name is "cpu_usage_1s", it needs to ensure that the fields that record "CPU usage 1 second average" before and after the restart are matched. Then, for each pair of fields, the hash check factor is extracted and the numerical difference is calculated, usually using unsigned integer or vector difference to represent the difference result, to quantify the structural and content variation amplitude of the same field in the two periods. The hash difference results of all fields will be summarized into a two-dimensional matrix structure, namely the field hash offset matrix, where the row represents the field name and the list represents the hash difference dimension. This matrix not only reveals the consistency change degree between fields, but also can be used to mine potential broken fragments or state decay areas. For example, if the hash value of a field after the hot start deviates from the hash value before the restart by more than a certain threshold, it can be inferred that the operation and maintenance semantics of the field may be damaged or the data source is not restored, and then marked as inconsistent field. This method uses structured hash offset mapping to quickly complete field-level difference evaluation without restoring the original field value, greatly improving the efficiency of judging the state continuity and data integrity in the terminal device operation and maintenance process driven by artificial intelligence.

[0065] Traverse the field hash offset matrix, calculate the normalized residual of the hash difference value of each field, and generate a hash consistency index combining the field data length and the collection interval to represent the similarity score result of the current field between the two periods.

[0066] In the terminal device hot start state consistency evaluation process, in order to accurately measure the consistency degree of each field before and after hot start, the field hash offset matrix needs to be analyzed item by item, and the normalized residual of the hash difference between each field in the two periods is calculated. The core of this operation is to unify the scale of different hash differences, so that the difference results of fields of different lengths and different types have comparability. Specifically, for each field, first calculate the proportional relationship between its hash offset and the maximum possible offset range of the field in the historical record, and obtain the normalized residual value. This value reflects the position of the current offset within the acceptable range. For example, for the "memory_utilization" field, if the hash offset value is in the upper quartile of the historical offset fluctuation range, the normalized residual value will be close to 1, indicating high abnormality. Next, combined with the data length of the field in the two collections (i.e. the number of data points covered by the sampling window) and the collection interval (i.e. the refresh period of the field), introduce the data volume factor and time density factor as correction parameters. The reason is that the more dense the field data and the more abundant the samples, the stronger the statistical stability of the hash difference, and the higher the reliability. Finally, the normalized residual value and the data length correction term and the collection interval weight term jointly generate the hash consistency index. This index is a real number score, which is used to express whether the field has retained sufficient information continuity and structural integrity before and after restart. For example, the field with a hash consistency index close to 1 indicates that its operation and maintenance state has good inheritability and high reliability; while the field with an index lower than the preset threshold will be marked as a state fracture high-risk field. In this way, quantitative basis can be provided for subsequent reliability evaluation and field elimination, which is a key link to realize accurate operation and maintenance judgment of artificial intelligence in lightweight scenarios.

[0067] S2, in the case of identifying the missing state continuation data, extracting the state sequence after hot start from the running state record in the current period, constructing a hot start residual piece mutation mapping tensor, and determining the running state change piece through the field stability score and the boundary disturbance gradient model;

[0068] In this embodiment, S2 specifically includes the following steps:

[0069] S201, in the case of identifying the missing state continuation data, extracting the state sequence after hot start from the running state record in the current period, locating the time node of hot start completion, selecting a plurality of continuous operation and maintenance data samples adjacent to the hot start timestamp from the continuous state collection stream, and constructing a triple sequence containing field identifier, field value and field timestamp, to form the state sequence after hot start;

[0070] After the terminal device performs a hot start, in order to accurately evaluate the running state of the device in the early stage of restart, it is necessary to first accurately locate the time node of the completion of the hot start. The time point can be determined by listening to the key event signal in the startup process, the state identification code in the kernel log or the CPU scheduling recovery mark. After the time node is located, a continuous operation and maintenance data sample in a time range is extracted according to the time stamp. The sample is derived from a lightweight state collection stream inside the device, which records running data at a fixed time interval, including but not limited to CPU usage, memory fluctuation, power state, thread scheduling, abnormal interruption count and other key dimensions. The extraction operation is usually completed through a lightweight data interface or embedded middleware, and is structured and packaged according to the three elements of field identification, field value and time stamp, forming a plurality of three-tuple structures, and is sorted in chronological order, thereby building a complete and usable tensor model for the post-hot-start state sequence.

[0071] The core features involved in this process include three key points: first, the field identification is used to mark the source and physical meaning of each independent parameter in the operation and maintenance data, for example, to distinguish whether it is the CPU core temperature or the memory paging number; second, the field value represents the actual measured value of the field at the corresponding time stamp, which needs to be standardized through the pre-defined collection accuracy and filtering logic; third, the field time stamp records the time point of data collection, usually using a high-precision clock source to ensure the timing accuracy between consecutive samples. Through the ordered combination of field identification, field value and field time stamp, the original data in the continuous state record stream can be converted into sequence information with clear structure, which can be further processed by the model, providing a data basis for subsequent tensor construction and anomaly reasoning. This design ensures that in the case of missing state continuation data, the key behavior trajectory can still be recovered through the post-hot-start state sequence, meeting the data integrity requirements of artificial intelligence in lightweight operation and maintenance.

[0072] S202, based on the post-hot-start state sequence, a hot-start residual piece mutation mapping tensor is constructed, the three-tuple sequence is organized into a three-dimensional tensor structure according to the field dimension, time dimension and value jump amplitude dimension, and the value difference sequence and disturbance weight coefficient between adjacent time points are calculated for each field, which are used to describe the mutation trend and disturbance characteristics of the field in the early stage of hot start;

[0073] S203, performing a field stability scoring operation on the hot-start residual piece mutation mapping tensor, and calling a boundary disturbance gradient model to respectively fit and analyze the disturbance continuity and disturbance gradient fluctuation rate of the field in the time dimension, and locating the disturbance edge critical point according to the field stability scoring result and the boundary gradient change rate to determine the time boundary range of the running state change piece.

[0074] After the tensor of hot-start residual fragment mutation mapping is constructed, in order to extract the fragments that truly reflect the change of the running state of the device from the tensor, the field stability score of the fields in the tensor needs to be calculated first. The score is evaluated by counting the change trend, change frequency and fluctuation direction consistency of the disturbance weight coefficient of the field in a short time window, so as to measure the stability performance of each field in the whole hot-start period. Then, the boundary disturbance gradient model is called for the fields with low scores, and the disturbance continuity and disturbance gradient fluctuation rate of these fields in the time dimension are fitted and analyzed. The disturbance continuity reflects whether the mutation event of the field is in a continuous state, and the disturbance gradient fluctuation rate describes the steepness of the change of the mutation intensity over time. The fitting results of the two are jointly analyzed with the field stability score to identify the critical turning points with significant gradient fluctuations and scores lower than the set threshold, and finally the disturbance edge critical points are located. These critical points represent the key nodes of the change of the running state of the device, and by marking the time stamps of these nodes, the boundary range of the running state change fragment in time can be determined, so as to realize the accurate extraction of the device abnormal behavior window.

[0075] The field stability score is used to quantify the fluctuation consistency of the field in the hot-start sequence, which is calculated by fusing the stability degree, jump direction continuity and fluctuation frequency of the disturbance weight coefficient in multiple consecutive time segments. The lower the score, the more dramatic and unpredictable the change of the field in the hot-start process, and the higher the abnormal reference value. The boundary disturbance gradient model is used to model the change rate of the disturbance amplitude of the field over time, which is to extract the gradient change trajectory of the disturbance curve of each field and perform sliding window convolution to calculate the fluctuation rate distribution. The disturbance continuity is used to evaluate whether the jump behavior presents a time aggregation trend, that is, whether it frequently mutates in a short time. The boundary gradient change rate describes the rising or falling rate of the mutation intensity, which is used to determine the boundaries of the entry and end of the field disturbance behavior. By aligning the extreme points of the gradient fluctuation rate curve of the field with low stability score with the time dimension, a set of disturbance edge critical points can be constructed, and the complete running state change fragment can be extracted, providing accurate input data for subsequent operation and maintenance credibility evaluation.

[0076] In this embodiment, S202 specifically comprises:

[0077] The triplets corresponding to each field in the state sequence after hot-start are classified according to the field identifier, the values of the same field at consecutive time stamps are constructed into a time sequence, and all the time sequences corresponding to the fields are initially arranged according to the field dimension and the time dimension, which is used to construct the basic structure of the three-dimensional tensor;

[0078] In order to construct the hot-start residual fragment mutation mapping tensor, the triplets in the state sequence after the hot start need to be structured and sorted. In specific implementation, first, all triplets are classified according to field identifiers, and triplets with the same field identifier are classified into the same field category. Each field category contains multiple triplets with different timestamps but the same field identifier, representing the state changes of the field at different time points after the hot start. Then, for each field category, the field values at consecutive timestamps are extracted and arranged in chronological order to construct the time series of the field. The time series of all fields form a two-dimensional structure, with one dimension being the field dimension and the other dimension being the time dimension. On this basis, according to the correspondence between the field dimension and the time dimension, the time series of each field is mapped to the two-dimensional slice surface of the tensor structure, and the third dimension is reserved for subsequent filling of the numerical jump amplitude. This initial arrangement operation aims to provide a stable basic framework for the three-dimensional tensor, so that each position in the tensor can accurately represent the state information of a field at a certain time point. The joint arrangement of the field dimension and the time dimension realized in this way not only ensures the order of the data structure, but also provides a high-resolution local state basis for mutation analysis, enhancing the effectiveness of the tensor in capturing short-period abnormal behavior. The field identifier ensures the accuracy of the classification, the timestamp ensures the continuity of the sequence, and the field value provides the original basic data for behavior modeling, all of which constitute the key inputs for the construction of the three-dimensional tensor.

[0079] For each field, the numerical difference between adjacent time points in its time series is calculated to form a jump amplitude sequence, and each difference is normalized according to the field acquisition interval, unit change rate, and historical disturbance fluctuation to generate a numerical jump amplitude dimension, which is used to complete the numerical jump amplitude dimension of the three-dimensional tensor.

[0080] In constructing the hot start residual fragment mutation mapping tensor, fine-grained jump analysis needs to be performed on the time series data of each field. In the specific implementation process, first, for the time series of each field, the difference value of the field value between two consecutive time points is calculated to generate the original jump amplitude sequence. This sequence reflects the numerical fluctuation of the field in a short time window and is an important basis for identifying sudden state changes. In order to unify the differences in collection frequency, unit scale and fluctuation sensitivity of different fields, the original difference value needs to be normalized. In the normalization calculation process, each difference value is divided by the collection interval of the corresponding field to measure the change speed in unit time; then it is multiplied by the unit change rate of the field, which represents the standard response rate of the field in the physical sense; finally, combined with the historical disturbance fluctuation degree, the current change is compared with the normal fluctuation range of the field in the known state before and after the hot start to reflect the degree of deviation from normal behavior. Through the above normalization steps, the jump amplitude value under the unified dimension system is generated and filled into the three-dimensional tensor as the numerical jump amplitude dimension, thereby realizing the completion of the third dimension of the tensor. This processing not only improves the tensor expression ability, making it have both time continuity and numerical fluctuation, but also provides high-dimensional semantic support for subsequent mutation trend identification and disturbance modeling. The field collection interval ensures the time dimension accuracy, the unit change rate reflects the index reaction characteristics, and the historical disturbance fluctuation degree enhances the consistency of the context semantics of abnormal identification. For example, if a field is CPU usage, it jumps from 45% to 88% between two sampling points, and its unit change rate is 10% per second, and the historical fluctuation range is ±15%, then the jump will be represented as a high mutation weight after normalization, directly affecting the expression strength in the disturbance tensor.

[0081] Based on the combination characteristics of each field in the three-dimensional tensor in the time dimension and the numerical jump amplitude dimension, a disturbance weight coefficient sequence is constructed for each field. The disturbance weight coefficient is generated according to the jump frequency, mutation persistence and jump amplitude standard deviation, and is used to quantify the mutation trend and disturbance characteristics of the field in the early stage of hot start.

[0082] To quantify the possible drastic change behavior of each field at the initial stage of hot start, a disturbance weight coefficient sequence is constructed for each field based on the constructed three-dimensional tensor, extracting the combined features of the time dimension and the numerical jump amplitude dimension. In specific implementation, first, the time sequence and the jump amplitude sequence corresponding to each field in the tensor structure are traversed, and a joint analysis is performed thereon. The jump frequency is obtained by counting the number of times the jump amplitude exceeds the set sensitivity threshold within a unit time, which is used to reflect the mutation activity of the field; the mutation persistence is measured by analyzing the time span covered by consecutive jump events, and if a field has large fluctuations at multiple consecutive time points, its mutation persistence score is higher; the jump amplitude standard deviation is used to capture the unevenness of field fluctuations, and high standard deviation means that the field has drastic but unstable changes after hot start. These three indicators are integrated into the disturbance weight coefficient, forming the disturbance weight coefficient sequence corresponding to each field, which is used to represent the mutation trend and disturbance intensity of the field at the initial stage of hot start. For example, if a field has 5 consecutive numerical jumps with an amplitude greater than the upper limit of regular fluctuations within the first 30 seconds, and these jump events are distributed in adjacent time windows, and the fluctuation intensity is inconsistent, then the disturbance weight coefficient of the field will be evaluated as high, which means that it has strong influence on the judgment of the running state change segment. This processing method not only realizes the structured modeling of the mutation behavior of the field, but also provides a quantifiable disturbance basis for subsequent identification of the state change boundary by the gradient model.

[0083] S3, input the running state change segment into the credibility mapping network, generate a credibility mapping vector by calculating the cross-field residual value smoothing factor and the local mutation pattern alignment rate, and judge whether the operation and maintenance judgment has credibility based on the credibility mapping vector;

[0084] In this embodiment, S3 specifically includes the following steps:

[0085] S301, arrange the running state change segment into a multi-channel input structure according to the field order and the time order, use a feature encoder to vectorize the field value change trajectory, and use a position encoding mechanism to maintain the sequence information of the field in the time dimension, input the encoding result into the credibility mapping network, and use it to initialize the credibility calculation process;

[0086] In actual implementation, the process of arranging the running state change segment in field order and time order into a multi-channel input structure can be implemented by constructing a two-dimensional matrix, each row of the matrix corresponding to an independent field, and each column representing a time slice, thereby forming the value change trajectory of the field at different times. The feature encoder can be a lightweight transformer structure or a one-dimensional convolutional neural network, which is used to extract the local dynamic features of the field value change. By scanning the time series of each field in a sliding window manner, the change trend, fluctuation frequency, amplitude feature and other information of the field are extracted and vectorized. In order to ensure that the time sequence information is not lost, a position encoding mechanism needs to be introduced in the encoding process, for example, using a sine-cosine position function to generate a unique position vector for each time point, and fusing it with the field feature vector to enhance the model's perception of sequence order. The final encoding result is organized in a multi-channel format, each channel representing the dynamic features of a field at multiple time points, serving as the input of the credibility mapping network, facilitating subsequent credibility reasoning calculation.

[0087] The field order is used to define the arrangement relationship between channels in the multi-channel input structure, ensuring that the same field is always in a fixed channel position, avoiding confusion of field meaning. The time order dimension controls the evolution trajectory of the field value, so that the subsequent calculation can capture the mutation point and trend fluctuation. The feature encoder is used to model the time variation law of the field value, and outputs a set of fixed-dimension vectors representing the time sequence features of each field. The position encoding mechanism provides the positioning function of the time point, preventing the loss of time sequence information due to inter-layer mapping in the deep neural network. In this process, the time resolution, field sampling interval, encoding dimension and other parameters can be customized to adapt to different types of terminal device running data. The credibility mapping network accepts the multi-channel data processed by feature extraction and position fusion, and can establish a nonlinear association between the running state change segment and the credibility, improving the model's ability to identify the credibility of abnormal judgment. Through this structured data preprocessing and vectorized input, the credibility reasoning process is robust and generalizable.

[0088] S302, in the credibility mapping network, a cross-field residual value smoothing factor is calculated based on the fluctuation trend of the residual values between different fields, the residual value smoothing factor reflects the fluctuation consistency of the field residual in consecutive time slices, and the disturbance pattern of the mutation section in the change segment is extracted, and compared with the typical abnormal pattern marked in the historical training set, the similarity between the current disturbance pattern and the abnormal pattern is evaluated by using the alignment rate, to generate a credibility mapping vector;

[0089] S303, based on the vector distribution position of the trusted mapping vector in the pre-defined trusted space, the distance relationship with each trusted level center point is calculated by using the distance function and the multi-class classifier, whether the operation and maintenance judgment corresponding to the running state change segment has trustworthiness is judged, when the trusted mapping vector deviates from the high trust region and falls into the low trust distribution cluster, it is judged that the operation and maintenance judgment does not have trustworthiness.

[0090] In order to judge whether the operation and maintenance judgment corresponding to the running state change segment has trustworthiness, a multi-dimensional vector space can be constructed as a pre-defined trusted space, which is divided into multiple trusted level regions by unsupervised clustering or supervised learning according to the high trust, medium trust and low trust state instances in the historical labeled data. A trusted level center point is set in each level region to represent the trusted mapping vector aggregation characteristics of the typical operation and maintenance state of the level. In actual operation, the generated trusted mapping vector is projected into the pre-defined trusted space, and the distance between the vector and each level center point is calculated by using the Euclidean distance, Manhattan distance or Gaussian kernel function. Then, by introducing support vector machine, multilayer perception or other multi-class classifier, the trusted mapping vector is classified, and the distance relationship with each level center point is combined to judge the closest trusted level. If the vector deviates from the high trust region and falls into the low trust level cluster, it is judged that the operation and maintenance judgment does not have trustworthiness, and the subsequent abnormal processing process is triggered.

[0091] In this process, the trusted mapping vector as the key input carries the comprehensive characteristics of the cross-field residual value smoothing factor and the disturbance mode alignment rate, and has high distinguishability. The construction of the pre-defined trusted space not only requires that the training samples are representative, but also needs to keep the distribution of the trusted level boundary clear in the high-dimensional feature space. The multi-class classifier is used to strengthen the boundary discrimination ability and avoid overlapping misjudgment between different levels. Each trusted level center point needs to be dynamically updated by iterative clustering algorithm to ensure synchronization with the change of actual operation state. Through the double decision mechanism of distance function and classifier, not only the judgment accuracy can be improved, but also the gradual processing of edge samples can be realized, which effectively avoids the operation and maintenance response error caused by misclassification of critical state. The whole process constitutes a closed loop structure from feature extraction to trustworthiness classification, which provides a stable and reliable intelligent judgment basis for the lightweight operation and maintenance system in the data missing scene.

[0092] In this embodiment, S302 is specifically

[0093] The value change sequence of each field in the running state change segment in the continuous time slice is extracted, the residual sequence of each field at adjacent time points is calculated, and the residual trend graph is constructed based on the residual mean change trend of the same time slice between fields. The cross-field residual value smoothing factor reflecting the consistency of residual fluctuation between fields is generated by analyzing the synchronization degree of residual change of each field in the trend graph.

[0094] In the implementation process, the numerical change sequence of each field on the continuous time slice can be first extracted from the running state change segment to form a plurality of one-dimensional time sequences, and each sequence represents the state trajectory of a field evolving over time. For each field sequence, the numerical difference between adjacent two time points is calculated to obtain the residual sequence of the field, which is used to express the short-term fluctuation characteristics. After aligning all field residual sequences according to the time slice, the residual trend graph is constructed based on the mean value change trend of all field residual values on each time slice, which is used to reveal whether the fluctuation directions of different fields are consistent at the same time. Further analysis of the synchronization degree of the residual change of each field in the trend graph can evaluate the change consistency between fields by using standard deviation normalization, dynamic time warping or Pearson correlation. If a plurality of fields present similar residual change direction and amplitude at most time points, it can be determined that they have high coupling in short-term behavior. By accumulating the residual synchronization at a plurality of time points, a comprehensive index is calculated to represent the fluctuation coordination degree between fields, which is the cross-field residual value smoothing factor. The higher the smoothing factor, the more consistent the fluctuations of each field, indicating that there may be a systematic abnormal trend rather than a single-point anomaly in this state change segment; otherwise, it indicates that the abnormal behavior is more likely to be an isolated event, which is helpful for accurate judgment of the credibility level. The field residual, residual trend graph, residual synchronization degree and final cross-field residual value smoothing factor in the calculation process can be quantitatively realized, providing reliable feature input basis for the credibility mapping vector.

[0095] After generating the cross-field residual value smoothing factor, the high-amplitude mutation position in the running state change segment is labeled, the disturbance intensity, duration and fluctuation frequency of the mutation segment are extracted, and the disturbance mode description vector is constructed to express the disturbance structure characteristics of the mutation segment.

[0096] In the actual implementation process, based on the time sequence of each field in the running state change segment, first, the position point of high amplitude jump is detected, and such mutation position usually shows that the field value changes sharply in a short time, and the jump amplitude exceeds the normal fluctuation threshold. For each identified jump point, the adjacent time window can be traversed forward and backward, the continuous fluctuation segment of the mutation segment is extracted, and three core disturbance features are calculated from it: disturbance intensity, duration and disturbance fluctuation frequency. The disturbance intensity represents the amplitude peak of the value change in the segment, reflecting the degree of mutation behavior; the duration defines the time length of the abnormal fluctuation, showing whether the disturbance is a transient peak or a slow evolution; the disturbance fluctuation frequency is measured by analyzing the switching frequency of the jump direction in the continuous fluctuation, and is used to reveal the internal rhythm in the disturbance process. The three indexes can be quantified to form a vector, and after unified normalization processing, they are spliced into a disturbance mode description vector. The vector can fully express the disturbance structure characteristics of a mutation segment in multiple dimensions, provide high-dimensional semantic support for subsequent pattern alignment comparison with known abnormal patterns, and help to judge whether the mutation belongs to the abnormal type that has appeared in history, thereby improving the resolution ability and accuracy of the overall credibility evaluation.

[0097] The disturbance mode description vector is compared with the standard abnormal disturbance vector in the pre-trained abnormal mode feature library, the alignment rate of the disturbance mode and the historical typical abnormal mode is calculated, the alignment rate is used to evaluate the consistency of the current mutation structure and the known abnormal behavior, and the cross-field residual value smoothing factor is combined to construct a credibility mapping vector, which is used as the basis for credibility judgment.

[0098] To realize the high-precision matching process between the disturbance pattern description vector and the standard abnormal disturbance vector, a multi-dimensional feature comparison technology can be used. In the established abnormal pattern feature library, the standard abnormal disturbance vectors corresponding to multiple known fault types are constructed in advance by combining unsupervised clustering and manual labeling. Each vector covers high-dimensional feature indicators such as disturbance intensity, duration, and fluctuation frequency. In actual operation, the current extracted disturbance pattern description vector is first standardized, and then cosine similarity, Euclidean distance, or Mahalanobis distance similarity measurement functions are used to match with the standard abnormal disturbance vectors one by one. The similarity score between each pair of vectors is calculated, and the highest score is selected as the alignment rate. The alignment rate is used to measure the structural consistency between the current disturbance structure and the historical typical abnormality. The higher the value, the closer the current behavior is to a known abnormal feature. Subsequently, the alignment rate and the previously generated cross-field residual smoothing factor are combined to construct a trusted mapping vector. The alignment rate is used to provide an abnormal correlation index in the semantic matching dimension, and the residual smoothing factor is used to supplement the global fluctuation consistency information. This fusion vector will reflect the disturbance commonality between fields and the similarity of abnormal behavior, providing a discriminative multi-dimensional feature expression for the subsequent credibility classification judgment process.

[0099] S4, performing a nonlinear double-threshold mapping operation according to the mapping position of the trusted mapping vector in the trusted space, dividing the credibility judgment result into multiple levels, respectively corresponding to different reasoning path control conditions;

[0100] In this embodiment, S4 is specifically:

[0101] According to the mapping position of the trusted mapping vector in the trusted space, a nonlinear double-threshold mapping operation is performed. A preset trusted upper threshold and a trusted lower threshold are constructed to build a segmented activation function for position mapping of the trusted mapping vector. By nonlinear distance compression between the vector and the double-threshold boundary, the response sensitivity of the mapping result to the boundary region change is enhanced.

[0102] To implement the nonlinear double-threshold mapping operation based on the trusted mapping vector, first, two static threshold boundaries are defined in the trusted space, which constitute the trusted upper threshold and the trusted lower threshold, respectively. These two thresholds are usually set according to the clustering centers of trusted and untrusted samples in the spatial distribution of the training data, ensuring the formation of high, medium, and low confidence transition areas in the space. When performing position mapping, the trusted mapping vector is input into a segmented activation function structure, which can be designed using a nonlinear function with a saturation region and a sensitive region (such as a double sigmoid combination or a softplus function superposition form). By nonlinearly compressing the Euclidean distance or cosine similarity between the vector and the upper and lower thresholds, vectors in the trusted boundary region have higher response strength to slight perturbations, thereby improving the model's discrimination accuracy for the edge state of confidence. This mapping mechanism also supports the rapid transition of vectors near the confidence level critical point, enhancing the system's ability to identify ambiguous states.

[0103] The trusted mapping vector represents the position characteristics of the running state change segment in the multi-dimensional abnormality discrimination space, and is a high-dimensional vector result generated by aligning the cross-field residual value smoothing factor and the perturbation mode after the pre-sequence network calculation. The trusted space provides an embedding domain with a structured confidence level distribution for the vector, and the nonlinear double-threshold mapping operation is applied to the continuous distribution structure of this space to compress the ambiguous transition zone between the high-confidence region and the low-confidence region. The trusted upper threshold is used to distinguish the boundary between high-confidence judgment and medium-confidence judgment, while the trusted lower threshold is used to identify the trigger boundary of low-confidence judgment. The design of the segmented activation function makes the output stable in the high-confidence and low-confidence regions, while the output in the critical region has a higher slope, thereby improving the judgment accuracy. Nonlinear distance compression makes it easier for the system to detect misjudgment risks in critical perturbation situations, which is an important foundation for improving the sensitivity and accuracy of confidence evaluation.

[0104] According to the results of the nonlinear double-threshold mapping operation, the confidence level interval to which the confidence judgment result belongs is determined, the mapped trusted mapping vector is divided into high-confidence, medium-confidence, and low-confidence regions, and each level interval is configured with an independent confidence level label to support the judgment basis for multi-level operation and maintenance response strategies.

[0105] In order to realize the level interval division of the credibility judgment result, interval positioning processing needs to be performed on the mapping output after performing the nonlinear double threshold mapping operation. The processing first compares the response value of the credible mapping vector with the preset upper threshold and lower threshold according to the continuous numerical value output by the nonlinear mapping function, judges the credible space area it falls into. If the response value is higher than the upper threshold of credibility, it is divided into the high credibility area; if the response value is lower than the lower threshold of credibility, it is divided into the low credibility area; the value between the two is classified into the medium credibility area. After completing the interval division, a unique credibility level label is assigned to each interval, which is used as the trigger basis of the inference control condition in the system, for selecting different levels of operation and maintenance response paths in the subsequent. For example, the results in the high credibility area can directly drive the rapid decision-making process, while the results in the low credibility area need to be guided to the redundant verification mechanism or manual review link to enhance the robustness and security of the system.

[0106] The division of the credibility level interval relies on the distribution state of the mapping result in the credible space, and the mapping value of the credible mapping vector is mapped to the standardized dimension and compared with the threshold. The high credibility area represents the judgment result with very high credibility, which is usually highly consistent with the stable running state or reproducible historical behavior pattern; the medium credibility area represents the state that the judgment result has partial deviation but is still within the tolerable range; the low credibility area represents the low confidence result caused by significant abnormality or structural deficiency. The credibility level label is a symbolic label used to identify the response level of the current judgment, and these labels are used as control signals in the inference engine to match the operation and maintenance strategy templates, realizing the multi-branch response mechanism of the inference path. Through this division mechanism, the system can execute differentiated action strategies according to the reliability of the judgment, avoiding the occurrence of misjudgment, missed judgment or decision delay due to unclear judgment.

[0107] Based on the divided credibility judgment result level, different inference path control conditions are correspondingly selected, including field weight adjustment parameters, abnormal inference model calling identification and decision output confidence lower threshold, to ensure that the operation and maintenance judgment completes the differentiated decision control process through the adaptive operation mechanism under different credibility levels.

[0108] After completing the credibility judgment result level division, an independent inference path control condition needs to be configured for each credibility level to ensure that the operation and maintenance judgment under different credibility levels can adopt the most suitable operation mechanism. In specific implementation, a set of control parameters can be preset for each level, including field weight adjustment parameters, abnormal inference model calling identifier, and decision output confidence lower threshold. The field weight adjustment parameters are used to dynamically adjust the influence weight of each operation and maintenance field in the model in the inference process, thereby enhancing the attention to key fields; the abnormal inference model calling identifier is used to decide whether to call an abnormal analysis model with higher complexity or with a historical comparison mechanism to improve the judgment rigor in the case of low credibility; and the confidence lower threshold is used to limit the output condition of the final inference result, ensuring that in the case of insufficient credibility, the system does not directly intervene, but enters the manual review or redundancy verification channel. These control conditions participate in the calculation process as input parameters in the model execution stage, and according to the credibility level corresponding to the mapping vector, the system will automatically load the matching control parameter group to form a complete path branch logic.

[0109] The core technical features of the inference path control condition lie in its dynamic and configurability. The system takes the credibility level mapping result as the decision condition input, loads the corresponding parameter template through the condition mapping relationship, and realizes the fine management of the subsequent inference process. The field weight adjustment parameters usually appear as a learnable vector, which participates in the weight product calculation during the forward propagation of the model, and determines the feature expression strength of different fields; the abnormal inference model calling identifier controls whether to switch to a deep model with historical trajectory reconstruction or causal chain analysis capability through a Boolean variable; and the confidence lower threshold is used as an output filtering threshold, which allows direct entry into the execution stage only when the confidence of the model prediction result is higher than the threshold. Through the combined setting of the three key control variables, the judgment under different credibility levels will trigger completely different operation and maintenance processing flows, so that the entire system has the ability of elastic response, high fault tolerance and self-adaptive discrimination, thereby significantly improving the practicality and safety of intelligent operation and maintenance.

[0110] In the divided confidence judgment result level, each level corresponds to a group of predefined reasoning path control conditions, which is used to adapt the operation and maintenance decision accuracy and fault tolerance requirement under different confidence levels. Among them, the field weight adjustment parameter is a multi-dimensional vector structure, which is used to dynamically adjust the importance distribution of the input field in the reasoning process. The weight parameter under the high confidence level is more inclined to maintain the original distribution of the model. The bias coefficient is introduced under the medium confidence level to strengthen the response to part of the fluctuating field. The attention weight of the historical high-risk field is increased under the low confidence level. The abnormal reasoning model calling identifier is a control signal, which is used to indicate whether to enable the model branch with more complex reasoning ability, such as the module with graph neural network structure or time sequence attention mechanism, which is usually activated under medium or low confidence level. The decision output confidence lower threshold is a numerical boundary setting, which is used to limit the minimum confidence level of the final reasoning result. The high confidence level corresponds to a lower threshold, which allows direct output. The medium confidence level sets a medium threshold, which needs to meet certain conditions before output. The low confidence level sets a high threshold limit, which needs to trigger the subsequent process through multiple model confirmations or manual review mechanisms, so as to realize the hierarchical control and risk release of the reasoning strategy.

[0111] S5, continuously record the offset data between the confidence judgment result after multiple periodic hot starts and the actual operation and maintenance state, construct a time sequence residual inversion tensor, update the confidence mapping vector generation logic through a Bayesian inference model, and realize dynamic regulation and control of the confidence judgment process.

[0112] In this embodiment, S5 is specifically:

[0113] The offset data between the confidence judgment result after multiple periodic hot starts and the actual operation and maintenance state is continuously recorded, the confidence mapping vector and the corresponding actual operation and maintenance state label in each period are obtained, the confidence offset residual of each period is calculated based on the field dimension and the time dimension, and a multi-period confidence offset record set is formed;

[0114] In the multi-period hot start scene, in order to realize dynamic monitoring and optimization of the confidence judgment process, the offset data between the confidence judgment result after each hot start in multiple continuous periods and the actual operation and maintenance state needs to be recorded. Specifically, the confidence mapping vector and the actual feedback operation and maintenance state label in each period can be recorded, and the two are corresponded on the same field dimension and time dimension, and the difference between the two is calculated to form a confidence offset residual. For example, in a certain hot start period, if the confidence mapping vector evaluates the confidence of a certain field state as high level, and the actual operation and maintenance state feedbacks that the field is in abnormal state, then the confidence offset residual of the field in the current period will be negative high amplitude. By archiving the offset residuals of all fields in multiple periods, a multi-period confidence offset record set containing time sequence and field classification can be formed, which will be used for subsequent dynamic regulation modeling.

[0115] The trusted mapping vector is a multi-dimensional vector output from the trust mapping network, each dimension corresponding to a trust judgment score of an operation and maintenance field; the actual operation and maintenance state label is marked by an artificial or system feedback mechanism, indicating the real state performance of each field in the current period; the field dimension is used to distinguish different operation and maintenance parameters such as temperature, current and voltage, and the time dimension is used to represent the time sequence in multiple periods; the trust offset residual is generated by calculating the deviation between the trust judgment value and the actual state of the same field in the same period, used to quantify the deviation degree of the judgment accuracy, and then reflect the trust stability problem existing in the long-term operation of the system.

[0116] According to the multi-period trust offset record set, a time sequence residual inversion tensor is constructed, the trust offset residual is arranged in period order, and a field identification dimension, a period index dimension and a residual amplitude dimension are introduced to generate a three-dimensional tensor structure, which is used to express the error distribution characteristics of the trust judgment result changing with the period after the hot start;

[0117] In order to comprehensively depict the evolution law of the deviation between the trust judgment result and the actual operation and maintenance state in the multi-period hot start process, it is necessary to construct a time sequence residual inversion tensor based on the collected multi-period trust offset record set. In the specific implementation process, first, the trust offset residual of each field in each period is classified and arranged in time sequence, ensuring that the record set has complete time sequence structure. On this basis, by setting the field identification dimension, the period index dimension and the residual amplitude dimension, the offset residual values of each field in different periods are mapped to the element values in the tensor, thereby constructing a three-dimensional tensor structure. The tensor can intuitively reflect the long-term offset trend and periodic error accumulation mode of the trust judgment of each field in the hot start operation, providing continuous input for the subsequent trust adjustment strategy based on the Bayesian inference model.

[0118] The multi-period trust offset record set is a time sequence set of trust errors of each field in multiple operation periods, the field identification dimension is used to distinguish different operation and maintenance indicators in the tensor, the period index dimension is used to identify each specific hot start period, ensuring the time sequence continuity of the data, and the residual amplitude dimension represents the error size of each field in the corresponding period. The three-dimensional tensor structure is used as the core form of data organization and storage in this process, which not only improves the compactness and structure of data representation, but also supports the subsequent model to fit and invert the error trend in the time and field dimensions, thereby realizing accurate modeling of the trust evolution process.

[0119] The time series residual inversion tensor is dynamically probabilistically modeled by a Bayesian inference model, prior probability distribution is updated based on the reliable offset trend of each field in different periods, and the residual value smoothing factor and abnormal alignment rate mapping parameter in the reliable mapping vector generation logic are corrected combined with the posterior inference logic, to realize dynamic regulation of the reliability judgment process.

[0120] In order to realize the dynamic regulation of the reliability judgment logic, it is necessary to dynamically probabilistically model the time series residual inversion tensor based on the Bayesian inference model. In specific implementation, first, the constructed three-dimensional tensor is taken as input, the reliability offset trend of each field in different periods is analyzed, and the error evolution feature sequence at the field level is extracted. Then, the prior probability distribution in the Bayesian model is updated using these trend information, so as to reflect the error fluctuation in the real environment after hot start. Then, combined with the observed current offset data, the posterior inference is carried out to generate new parameter estimation results. The posterior result will be used to correct the core parameters in the reliable mapping vector generation logic, including the dynamic adjustment factor of the cross-field residual value smoothing factor and the alignment rate mapping weight of the abnormal disturbance pattern, so that the model still maintains stability and adaptability when facing long-term offset and periodic disturbance, effectively improving the sensitive identification ability to abnormal changes.

[0121] The Bayesian inference model is an updating mechanism based on probability theory, which continuously corrects the distribution parameters in the belief model by introducing historical offset trends. The prior probability distribution represents the initial cognition of the model on the reliable change of the field before observing the current period data, and the posterior inference logic dynamically outputs the reliability update suggestion by integrating the current tensor observation results and prior information. The residual value smoothing factor is a weight factor to measure the consistency of field value fluctuation, and the abnormal alignment rate mapping parameter is used to describe the similarity between the current disturbance structure and the historical abnormal model. These key parameters will be automatically adjusted after being reevaluated by the Bayesian model, to realize the adaptive evolution of the reliable mapping vector generation mechanism in the dynamic environment. Through this mechanism, the system can continuously optimize the accuracy of reliability evaluation and maintain high sensitivity response to complex operation and maintenance state.

[0122] The above-described embodiments can be implemented in part or in whole through software, hardware, firmware or any combination thereof. When implemented in software, the above-described embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When loaded and executed by a computer, the computer instructions or computer programs cause the computer to perform all or part of the processes or functions described in the embodiments of the present application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired or wireless (for example, infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0123] It should be understood that the size of the sequence number of each process described above in various embodiments of the present application does not mean the order of execution, and the execution order of each process should be determined by its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0124] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be realized in electronic hardware, or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0125] In several embodiments provided in the present application, it should be understood that the disclosed system and method can be implemented in other ways. For example, the above-described embodiments are only illustrative, for example, the division of units is only a logical function division, and actual implementation can have another division manner, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0126] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0127] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit.

[0128] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A lightweight operation and maintenance method for terminal devices based on artificial intelligence, characterized in that, Specifically, the following steps are included: S1. Collect the status information of the terminal device after it has completed a hot start. Compare the operation and maintenance data fields before and after the restart using the field-level hash verification factor. Combine the timestamp mapping to determine whether the operation and maintenance data fields are missing and identify whether the terminal device has missing status continuation data. S2. When the missing state continuation data is identified, extract the state sequence after hot start from the running state record in the current time period, construct the hot start fragment mutation mapping tensor, and determine the running state change fragments through field stability score and boundary perturbation gradient model. S2 specifically includes the following steps: S201. When the status continuation data is found to be missing, extract the status sequence after hot start from the operation status record in the current time period. By locating the time node when the hot start is completed, select multiple continuous operation and maintenance data samples adjacent to the hot start timestamp from the continuous status acquisition stream, and construct a triplet sequence containing field identifier, field value and field timestamp to form the status sequence after hot start. S202. Construct a hot-start fragment mutation mapping tensor based on the state sequence after hot start. Organize the triplet sequence into a three-dimensional tensor structure according to the field dimension, time dimension and numerical jump amplitude dimension. Calculate the numerical difference sequence and perturbation weight coefficient between adjacent time points for each field to describe the mutation trend and perturbation characteristics of the field in the early stage of hot start. S202 specifically refers to: After a hot start, the triplet corresponding to each field in the state sequence is classified according to the field identifier. The values ​​of the same field at consecutive timestamps are used to construct a time series. The time series corresponding to all fields are initially arranged according to the field dimension and the time dimension to construct the basic structure of the three-dimensional tensor. For each field, the numerical difference between adjacent time points is calculated in its time series to form a jump amplitude sequence. Each difference is normalized according to the field acquisition interval, unit rate of change, and historical disturbance fluctuation to generate a numerical jump amplitude dimension, which is used to complete the numerical jump amplitude dimension of the three-dimensional tensor. Based on the combined features of each field in the three-dimensional tensor in the time dimension and the numerical jump amplitude dimension, a perturbation weight coefficient sequence is constructed for each field. The perturbation weight coefficient is generated according to the jump frequency, mutation persistence and jump amplitude standard deviation, and is used to quantify the mutation trend and perturbation characteristics of the field in the early stage of hot start. S203. Perform field stability scoring on the hot-start fragment mutation mapping tensor and call the boundary perturbation gradient model to perform fitting analysis on the perturbation continuity and perturbation gradient volatility of the field in the time dimension. Based on the field stability scoring results and the boundary gradient change rate, locate the perturbation edge critical point and determine the time boundary range of the running state change segment. S3. Input the segment of the change in the running status into the credibility mapping network, generate a credibility mapping vector by calculating the alignment rate between the cross-field residual smoothing factor and the local mutation mode, and determine whether the operation and maintenance judgment has credibility based on the credibility mapping vector. S4. Perform a nonlinear double threshold mapping operation based on the mapping position of the trustworthy mapping vector in the trustworthy space to divide the trustworthiness judgment result into multiple levels, each corresponding to different inference path control conditions. S5. Continuously record the offset data between the credibility judgment results and the actual operation and maintenance status after multiple cycles of hot start, construct the time-series residual inversion tensor, update the credibility mapping vector generation logic through the Bayesian inference model, and realize the dynamic control of the credibility judgment process.

2. The lightweight operation and maintenance method for terminal devices based on artificial intelligence according to claim 1, characterized in that, S1 specifically includes the following steps: S101. Collect the status information of the terminal device after it has completed a hot start. Read the current value of the preset operation and maintenance data field set by executing the field extraction command, and generate a unique field identifier and corresponding timestamp for each field to construct the current status information structure. S102. By using field-level hash verification factors, the set of operation and maintenance data fields saved before the restart and the set of operation and maintenance data fields collected after the hot start are matched one by one to generate a field hash offset matrix and calculate the hash consistency index of each field. S103. Combining the field hash consistency index with the temporal correspondence of the field timestamp in the timestamp mapping graph, a continuous retrieval path is constructed. By matching the temporal adjacency and hash consistency threshold of the fields, the break position in the continuous field chain is identified. The ratio of the number of broken fields to the total number of fields is used to determine whether there is a missing data field. When the ratio exceeds the preset loss threshold, it is determined that the terminal device has a situation where the state continuation data is missing.

3. The lightweight operation and maintenance method for terminal devices based on artificial intelligence according to claim 2, characterized in that, S102 specifically refers to: The field hash constructor is called to perform hash processing on each field in the set of operation and maintenance data fields saved before the restart and the set of operation and maintenance data fields collected after the hot start, respectively, and generate a field-level hash verification factor based on the field content, collection order and data type; The hash verification factors of the corresponding fields in the field set before restart and the field set after hot start are matched one by one. The hash difference between each pair of fields is calculated, and a field hash offset matrix is ​​constructed to express the consistency change of field values ​​before and after hot start. Traverse the field hash offset matrix, calculate the normalized residual of the hash difference for each field, and generate a hash consistency index by combining the field data length and the collection interval, which is used to represent the similarity score of the current field between two time periods.

4. The lightweight operation and maintenance method for terminal devices based on artificial intelligence according to claim 1, characterized in that, S3 specifically includes the following steps: S301. Arrange the running status change segments into a multi-channel input structure according to the field order and time order. Use the feature encoder to vectorize the field value change trajectory and maintain the order information of the fields in the time dimension through the position encoding mechanism. Input the encoding result into the credibility mapping network to initialize the credibility calculation process. S302. In the credibility mapping network, the cross-field residual smoothing factor is calculated based on the changing trend of residual values ​​between different fields. The residual smoothing factor reflects the consistency of the fluctuation of field residuals in continuous time slices. At the same time, the perturbation pattern of the abrupt segment in the change segment is extracted and compared with the typical abnormal patterns labeled in the historical training set. The alignment rate is used to evaluate the similarity between the current perturbation pattern and the abnormal pattern in order to generate a credibility mapping vector. S303. Based on the vector distribution position of the trusted mapping vector in the predefined trusted space, use the distance function and multi-class classifier to calculate the distance relationship between it and the center point of each trusted level, and determine whether the operation and maintenance judgment corresponding to the segment of the change in the running state has credibility. When the trusted mapping vector deviates from the high-credibility region and falls into the low-credibility distribution cluster, it is determined that the operation and maintenance judgment does not have credibility.

5. The lightweight operation and maintenance method for terminal devices based on artificial intelligence according to claim 4, characterized in that, S302 specifically refers to Extract the numerical change sequence of each field in the segment of running status change on continuous time slices, calculate the residual sequence of each field at adjacent time points, and construct a residual trend map based on the trend of the mean residual change of the fields in the same time slice. By analyzing the degree of synchronization of the residual changes of each field in the trend map, a cross-field residual smoothing factor that reflects the consistency of residual fluctuations between fields is generated. After generating the cross-field residual smoothing factor, the high-amplitude mutation locations in the running state change segments are marked, and the disturbance intensity, duration and disturbance fluctuation frequency of the mutation segments are extracted to construct a disturbance mode description vector to express the disturbance structure characteristics of the mutation segments. The perturbation pattern description vector is compared with the standard anomalous perturbation vector in the pre-trained anomalous pattern feature library using multidimensional features. The alignment rate between the perturbation pattern and the historical typical anomalous patterns is calculated. The alignment rate is used to evaluate the consistency between the current mutation structure and known anomalous behavior. A reliable mapping vector is constructed together with the cross-field residual smoothing factor, which serves as the basis for the reliability judgment.

6. The lightweight operation and maintenance method for terminal devices based on artificial intelligence according to claim 1, characterized in that, S4 specifically refers to: Based on the mapping position of the trusted mapping vector in the trusted space, a nonlinear double-threshold mapping operation is performed. A trusted upper threshold and a trusted lower threshold are preset, and a piecewise activation function is constructed to map the position of the trusted mapping vector. By compressing the nonlinear distance between the vector and the double-threshold boundary, the sensitivity of the mapping result to changes in the boundary region is enhanced. Based on the results of the nonlinear double threshold mapping operation, the level interval to which the credibility judgment result belongs is determined. The mapped credibility mapping vector is divided into high credibility zone, medium credibility zone and low credibility zone, and an independent credibility level label is configured for each level interval to support the judgment basis of multi-level operation and maintenance response strategy. Based on the defined confidence level, different inference path control conditions are assigned. These conditions include field weight adjustment parameters, abnormal inference model call identifiers, and decision output confidence thresholds. This ensures that maintenance judgments can complete differentiated decision control processes through appropriate computational mechanisms at different confidence levels.

7. The lightweight operation and maintenance method for terminal devices based on artificial intelligence according to claim 1, characterized in that, S5 specifically refers to: The offset data between the credibility judgment result and the actual operation and maintenance status after multiple cycles of hot start is continuously recorded. The credibility mapping vector and the corresponding actual operation and maintenance status label in each cycle are obtained. The credibility offset residual of each cycle is calculated based on the field dimension and the time dimension to form a multi-cycle credibility offset record set. Based on the multi-period reliable offset record set, a temporal residual inversion tensor is constructed. The reliability offset residuals are arranged in period order, and the field identifier dimension, period index dimension and residual amplitude dimension are introduced to generate a three-dimensional tensor structure, which is used to express the error distribution characteristics of the reliability judgment result after hot start as the period changes. Dynamic probability modeling of the temporal residual inversion tensor is performed using a Bayesian inference model. The prior probability distribution is updated based on the credibility offset trend of each field in the tensor under different periods. The residual smoothing factor and abnormal alignment rate mapping parameters in the credibility mapping vector generation logic are corrected by combining the posterior inference logic, so as to achieve dynamic control of the credibility judgment process.

Images