A monitoring management method and system of an artificial intelligence development platform

Abstract

The application discloses a kind of monitoring management method and system of artificial intelligence development platform, specifically related to computer system monitoring management field, for solving the problem of unstable monitoring segment attribution and inconsistent alarm explanation in switching scene;By generating training task continuous identification based on submission record and establishing running instance identification mapping table, collecting monitoring data stream and forming monitoring segment when switching event arrives, calculating switching timing alignment index and training semantic continuation index and executing sequence inference to obtain segment attribution result, merging training task continuous identification to generate training task continuous record and generating switching instruction, performing consistency check on training task continuous record to form consistency check result and triggering correction instruction to update mapping table and rebuild training task continuous record, while revising boundary echo window construction rule and semantic anchor deduction rule, the stability of monitoring attribution and explanation caliber is realized.

Description

Technical Field

[0001] This invention relates to the field of computer system monitoring and management, and more specifically, to a monitoring and management method and system for an artificial intelligence development platform. Background Technology

[0002] In artificial intelligence development platforms, model training is typically run in a distributed manner across multiple computing devices. The platform continuously collects the running status, resource usage, and log information during the training process, and triggers alarms when anomalies occur, enabling timely handling by R&D and operations teams. Existing technologies, such as the "Method and Apparatus for Monitoring Model Training" (application number CN201910458041.8), propose that the platform receives monitoring information reported by each node, determines monitoring indicators and corresponding information accordingly, and executes alarms when alarm rules are triggered, thereby achieving monitoring and maintenance of the training task process. This solution enables visualized monitoring of the training process and facilitates alerts for common anomalies.

[0003] However, in real development platforms, training tasks often encounter interruptions and subsequent resumptions, switching between different runtime environments, or restarting in new runtime instances to adapt to resource scheduling, fault recovery, or changes in the operating environment. During these transitions, node identifiers, runtime instance identifiers, and log attribution paths change. If the monitoring system still uses nodes or runtime instances as the primary aggregation basis, it easily breaks down the same training task into multiple discontinuous monitoring records, sometimes even mixing residual information from previous segments with new information from later segments. The root cause lies in the fact that the identifiers on the scheduling side are inherently short-lived and reusable, while the monitoring side often aggregates these identifiers as stable identities. When switching occurs, monitoring data may overlap or be adjacent in time. Slightly improper merging rules can lead to incorrect alarm attribution, distorted trend judgments, and contaminated problem location clues, ultimately resulting in alarms that are difficult to interpret and the entire process that cannot be traced back.

[0004] Therefore, the key technical problem to be solved is: in the monitoring and management methods and systems of artificial intelligence development platforms, how to establish a stable identity for training tasks oriented towards the life cycle, and how to continuously merge and interpretably correlate monitoring data before and after task switching, so that alarms can accurately correspond to the actual operation process of the same training task and provide clear switching reasons and impact clues.

[0005] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a monitoring and management method and system for an artificial intelligence development platform. This method generates continuous training task identifiers based on submission records and establishes a mapping table of running instance identifiers. It collects monitoring data streams and forms monitoring segments when switching events occur. It calculates the switching sequence alignment index and the training semantic continuation index, performs sequence inference to obtain segment attribution results, merges training task continuous records according to the continuous training task identifiers, generates switching descriptions, performs consistency checks on the continuous training task records to form consistency check results, triggers correction instructions to update the mapping table and reconstruct the continuous training task records, and simultaneously revises the boundary echo window construction rules and semantic anchor point inference rules. This achieves stable and unified monitoring attribution and interpretation, thereby solving the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] S1: Generate continuous identification of training tasks based on the submission record of training tasks. After the scheduling side generates the running instance identification, establish a mapping table between the running instance identification and the continuous identification of training tasks.

[0009] S2: Collect the monitoring data stream of the training task and form a monitoring segment when the switching event arrives. Calculate the switching timing alignment index and training semantic continuation index for each monitoring segment and perform sequence inference to obtain the segment assignment result. Bind the monitoring segment to the continuous identifier of the training task according to the segment assignment result.

[0010] S3: Monitoring segments are merged according to the continuous identifier of training tasks to generate continuous training task records. The continuous training task records retain the segment order relationship and generate switching instructions. The switching instructions are used to explain the alarm attribution and changes in the operational phase.

[0011] S4: Perform consistency checks on the continuous records of the training task to obtain consistency check results. When the consistency check results indicate that the segment assignment is abnormal, a correction instruction is generated and the mapping table is updated accordingly and the continuous records of the training task are re-merged. At the same time, the boundary echo window construction rules and semantic anchor inference rules are revised using the consistency check results to improve the stability of subsequent segment assignment.

[0012] In a preferred embodiment, continuous identifier source information is extracted based on the training task submission record. The continuous identifier source information includes the task's spatial information, task definition information, and task construction information. The continuous identifier source information is then subjected to unified encoding, null value regularization, fixed key order, and delimiter escaping, and a continuous identifier input string is generated according to a fixed concatenation rule.

[0013] In a preferred embodiment, the continuous identifier input string is converted into a byte sequence and input into a one-way fingerprint function to generate a continuous identifier for the training task. When a conflict is found in the local registry, the continuous identifier for the training task is used to generate a conflict resolution factor by using the submission sequence information and submission timestamp information, and the continuous identifier for the training task is regenerated. After obtaining the running instance identifier, a mapping table is established. The mapping table maintains a one-to-one correspondence between the running instance identifier and the continuous identifier for the training task through atomic submission.

[0014] In a preferred embodiment, the indicator records, log records, tracking records, and event records in the monitoring data stream are normalized and encapsulated, and each monitoring record is appended with a running instance identifier, a collection channel identifier, and a collection timestamp to form a unified monitoring record; after a switch event arrives, the running instance identifier change information, switch reason information, and switch event occurrence timestamp are registered, and a monitoring segment is generated based on this information using the event center.

[0015] In a preferred embodiment, a boundary echo window is constructed for the monitoring segment, and log session reconstruction markers, indicator collector reconnection markers, tracing link reconstruction markers, and event queue replay markers are extracted to form a boundary signal set and the switching timing alignment index is calculated. Global step markers, recoverable point markers, data cursor markers, and model building signature markers are extracted for the monitoring segment to form a semantic anchor fingerprint and the training semantic continuation index is calculated. The segment attribution result is obtained based on sequence inference and the training task continuous identifier is bound through a mapping table.

[0016] In a preferred embodiment, the monitoring segments are grouped according to the continuous identifier of the training task to form a monitoring segment set. The segment start time stamp, segment end time stamp, switching event occurrence time stamp, and first unified monitoring record collection time stamp are read to perform stable sorting. The segment order relationship is generated according to the adjacent monitoring segments, and the connection type is determined according to the segment ownership result and the switching event reference.

[0017] In a preferred embodiment, training task continuous records are assembled along the segment order relationship, and the training task continuous records are written with original index and summary information. The original index includes acquisition channel identifier, earliest acquisition time stamp, latest acquisition time stamp, and storage partition identifier. The interpretation difference between adjacent monitoring segments is calculated to determine the interpretation center, and a switching description is generated based on a fixed field template and bound to the segment order relationship.

[0018] In a preferred embodiment, the training task is continuously recorded and a sequence of adjacent monitoring segments is constructed based on the segment order relationship. The segment start time stamp, segment end time stamp, boundary echo window, boundary signal set, semantic anchor fingerprint first anchor point, semantic anchor fingerprint tail anchor point, switching event reference, and alarm occurrence time stamp are extracted to form a consistency verification input and generate anomaly evidence reference.

[0019] In a preferred embodiment, the boundary consistency score, semantic anchor continuity score, alarm attribution anomaly ratio are calculated and the segment attribution anomaly score is synthesized. After the segment attribution anomaly score triggers the correction instruction, the range of running instance identifiers, the range of monitored segments, and the candidate set of continuous identifiers for training tasks are determined. The mapping table is atomically replaced and the continuous identifiers of training tasks are rebound to the monitored segments. The continuous records of training tasks are reconstructed and the boundary echo window construction rules and semantic anchor inference rules are revised.

[0020] A monitoring and management system for an artificial intelligence development platform, comprising:

[0021] Continuous Identification Module: Generates continuous identification of training tasks based on the submission record of training tasks. After the scheduling side generates the running instance identification, it establishes a mapping table between the running instance identification and the continuous identification of training tasks.

[0022] Segment Attribution Module: Collects the monitoring data stream of the training task and forms a monitoring segment when a switching event occurs. For each monitoring segment, it calculates the switching timing alignment index and the training semantic continuation index and performs sequence inference to obtain the segment attribution result. The monitoring segment is bound to the continuous identifier of the training task according to the segment attribution result.

[0023] Continuous merging module: Monitoring segments are merged according to the continuous identifier of training tasks to generate continuous records of training tasks. The continuous records of training tasks retain the segment order relationship and generate switching instructions. The switching instructions are used to explain the alarm attribution and changes in the operation stage.

[0024] Consistency Correction Module: Performs consistency checks on continuous records of the training task to obtain consistency check results. When the consistency check result points to an abnormal segment assignment, a correction instruction is generated and the mapping table is updated accordingly, and continuous records of the training task are re-merged. At the same time, the consistency check result is used to revise the boundary echo window construction rules and semantic anchor inference rules to improve the stability of subsequent segment assignment.

[0025] The technical effects and advantages of the monitoring and management method and system for an artificial intelligence development platform of the present invention are as follows:

[0026] 1. The training task continuous identifier and mapping table unifies and merges the running instance identifiers of a training task after multiple migrations and restarts. The monitoring segment is bound to the training task continuous identifier according to the segment ownership result. The monitoring page and alarm page always revolve around the same training task continuous identifier. Operation and maintenance personnel do not need to switch repeatedly between multiple running instance identifiers to view indicators, logs, traces and events in a connected manner. When locating problems, they can directly trace back to the collection channel identifier and collection time scale range corresponding to the original index of the monitoring segment along the training task continuous record.

[0027] 2. The switching sequence alignment index and training semantic continuation index simultaneously incorporate the continuation evidence of the boundary signal set and semantic anchor fingerprint in the boundary echo window into sequence inference. The continuous recording of the training task solidifies the switching reason information and evidence into readable conclusions through the segment order relationship and switching description. After the consistency verification result triggers the correction instruction, the mapping table atomic replacement and the continuous recording reconstruction of the training task are completed. At the same time, the boundary echo window construction rules and semantic anchor inference rules are revised to ensure that the monitoring segment attribution and alarm interpretation remain consistent in the long term under the same switching scenario. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating a monitoring and management method for an artificial intelligence development platform according to the present invention.

[0029] Figure 2 This is a schematic diagram of the monitoring and management system of an artificial intelligence development platform according to the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example 1: Figure 1 The present invention provides a monitoring and management method for an artificial intelligence development platform, comprising:

[0032] S1: Generate continuous training task identifiers based on the submission records of training tasks. After the scheduling side generates the running instance identifiers, establish a mapping table between the running instance identifiers and the continuous training task identifiers.

[0033] S2: Collect the monitoring data stream of the training task and form a monitoring segment when the switching event arrives. Calculate the switching timing alignment index and training semantic continuation index for each monitoring segment and perform sequence inference to obtain the segment assignment result. Bind the monitoring segment to the continuous identifier of the training task according to the segment assignment result.

[0034] S3: Monitoring segments are merged according to the continuous identifier of the training task to generate a continuous record of the training task. The continuous record of the training task retains the segment order relationship and generates a switching description. The switching description is used to explain the alarm attribution and changes in the operation stage.

[0035] S4: Perform consistency checks on the continuous records of the training task to obtain consistency check results. When the consistency check results indicate that the segment assignment is abnormal, a correction instruction is generated and the mapping table is updated accordingly and the continuous records of the training task are re-merged. At the same time, the boundary echo window construction rules and semantic anchor inference rules are revised using the consistency check results to improve the stability of subsequent segment assignment.

[0036] This invention addresses the core issue of monitoring attribution drift caused by switching, creating a traceable, continuous chain from the submission of a training task to its execution after multiple migrations and restarts. First, a continuous identifier for the training task is generated using submission records. A mapping table is then used to uniformly associate the runtime instance identifiers generated by each scheduling with the same continuous identifier for the same training task, ensuring a stable primary key for monitoring attribution. Subsequently, when a switching event occurs, the monitoring data stream is segmented into monitoring fragments. Boundary evidence is extracted from the monitoring side, and semantic anchor evidence is extracted from the training side, forming a switching sequence alignment index and a training semantic continuation index. Sequence inference is then used to place adjacent fragments on the same timeline for comprehensive attribution determination, avoiding the use of a single record or a single switching event for fragment attribution. After attribution, the monitoring fragments are merged into continuous records for the training task according to the continuous identifier. These continuous records retain the fragment order relationship and organize the switching reason information, boundary evidence, and semantic evidence into a switching description, enabling alarm attribution and stage changes to be interpreted using the same set of criteria. Finally, when anomalies in segment attribution are detected through consistency verification, correction instructions are generated to drive the mapping table update and continuous record reconstruction of the training task. The confirmed boundary patterns and semantic connection patterns are written back into the boundary echo window construction rules and semantic anchor inference rules, making the segment attribution of subsequent similar switching more stable.

[0037] Training tasks on AI development platforms may migrate and restart during operation. Monitoring data streams often carry running instance identifiers rather than long-term stable training identities. Subsequent processing requires a stable attribution basis to accurately aggregate monitoring data streams under the same training task. However, running instance identifiers have short lifecycles and reusable characteristics. The semantic identity of the training task needs to be solidified at the submission stage and a searchable mapping established with the running instance identifier in order to support the attribution binding and sequence inference of subsequent monitoring segments.

[0038] Detailed implementation of step S1:

[0039] S101 Continuous Identifier Source Information Extraction and Continuous Identifier Input String Generation.

[0040] To ensure that the same training task maintains a consistent identity after migration and restart, after the training task submission request arrives, continuous identifier source information is extracted from the submission record. The continuous identifier source information is limited to the task's spatial information, task definition information, and task construction information.

[0041] Subsequently, normalized extraction is performed on the continuous identifier source information. Normalized extraction includes four types of actions: encoding standardization, null value normalization, key order fixing, and delimiter escaping. Encoding standardization is used to merge text encoding differences from different submission endpoints into a single encoding format; null value normalization is used to uniformly replace missing fields with fixed placeholder expressions; key order fixing is used to arrange field keys in a preset order, consistent with the order of the three types of fields in the continuous identifier source information; delimiter escaping is used to replace any delimiters that may appear in field values ​​with escaped expressions to avoid ambiguity after concatenation.

[0042] After the normalized extraction is completed, the key sequence is fixed, and the field keys and field values ​​are concatenated sequentially to form a continuous identifier input string. The continuous identifier input string is generated using deterministic concatenation rules to ensure that the same continuous identifier source information always generates the same continuous identifier input string.

[0043] S102 Training Task: Continuous Identifier Generation and Conflict Resolution.

[0044] To ensure that the continuous identifiers for the training task are stable, verifiable, and do not leak the source information of the continuous identifiers within the platform, after the continuous identifier input string is generated, it is first converted into a byte sequence according to a unified encoding rule. Then, the byte sequence is input into a one-way fingerprint function, which outputs a fingerprint byte sequence of fixed length. The fingerprint byte sequence is then converted into a displayable character encoding to form the continuous identifiers for the training task.

[0045] After training task consecutive identifiers are generated, they are immediately retrieved from the local registry. The local registry stores the correspondence between training task consecutive identifiers and consecutive identifier input strings, used to determine whether consecutive identifiers for the same training task correspond to consistent consecutive identifier input strings. If the search finds a match and the consecutive identifier input strings match, the training task consecutive identifier is directly confirmed as valid; if the search finds a match but the consecutive identifier input strings do not match, the conflict resolution process begins.

[0046] In the conflict resolution process, non-reusable information is extracted from the submission records to generate conflict resolution factors. The non-reusable information is selected from the submission sequence information and the submission timestamp information, and concatenated into conflict resolution factors according to a unified coding rule. Then, the conflict resolution factors are appended to the end of the continuous identifier input string to form a new continuous identifier input string. The one-way fingerprint function is executed again to generate new training task continuous identifiers, and the local registry search is repeated until the training task continuous identifier uniquely corresponds to a continuous identifier input string in the local registry.

[0047] In one embodiment, after a training task submission request enters the platform, the platform first reads the task's spatial information, task definition information, and task construction information from the submission record. After unifying the encoding and fixing the key order, it generates a continuous identifier input string. Then, it converts the continuous identifier input string into a byte sequence and inputs it into a one-way fingerprint function to obtain the continuous identifier of the training task. When the local registry search finds that the continuous identifier of the training task already exists and the continuous identifier input string is inconsistent, the platform reads the submission sequence information and submission timestamp information from the submission record to generate a conflict resolution factor. The conflict resolution factor is appended to the end of the continuous identifier input string, and the continuous identifier of the training task is generated again until the local registry search result satisfies the unique correspondence.

[0048] A one-way fingerprint function is a computational function that maps a variable-length input, such as a continuous identifier input string, into a fixed-length fingerprint through hashing. The hashing operation satisfies the condition that the fingerprint changes significantly when the input changes slightly, and it is computationally infeasible to deduce the original input from the fingerprint. Platforms usually use cryptographic hashing algorithms to implement this function. Common implementations include secure hashing series or equivalent hashing implementations, which are used to stably compress the task's spatial information, task definition information, and task construction information into continuous identifiers for the training task.

[0049] The conflict resolution factor is a non-reusable marker appended to the end of the continuous identifier input string to distinguish between two inputs when the same continuous identifier for a training task is found to correspond to different continuous identifier input strings in the local registry. The conflict resolution factor is usually obtained by concatenating the submission pipeline information and the submission timestamp information according to a fixed code, and then re-entering the continuous identifier input string into the same hash operation process to generate a new continuous identifier for the training task, thereby eliminating collisions and retaining traceable associations while maintaining the consistency of the generation rules.

[0050] S103 Mapping Table Creation and Atom Submission.

[0051] To ensure that the monitoring system can reliably retrieve the continuous training task identifier from the running instance identifier, after the scheduling side creates a running instance, the platform obtains the running instance identifier and establishes a mapping table record to correspond the running instance identifier with the continuous training task identifier.

[0052] The mapping table records at least the running instance identifier, training task continuity identifier, running instance creation time, running instance current status, and reference to the most recent switch event. The running instance creation time is taken from the creation information returned by the scheduler and is used to align with the time stamp of subsequent switch events; the running instance current status indicates whether the running instance is in the creation, running, or terminated state; the reference to the most recent switch event is initialized to a null placeholder value for subsequent writing of switch event references and to maintain field consistency.

[0053] The mapping table update adopts the atomic commit method. The atomic commit uses the running instance identifier as the unique positioning key to execute a single record write or a single record replacement. After the commit action is completed, it is ensured that the running instance identifier corresponds to only one training task consecutive identifier, avoiding concurrent writes that cause the running instance identifier to point to multiple training task consecutive identifiers.

[0054] In one embodiment, the platform's job console receives a training task submission request. The submission record includes the project space name, training task name, training script entry point, mirror address, dependency list, and configuration file path. The platform arranges these fields in a fixed order, standardizes the character format, and fills in placeholders for missing fields to form a continuous identifier input string. This continuous identifier input string enters the fingerprint generation process to obtain the training task continuous identifier. The scheduling side then creates a running instance and returns the running instance identifier. The platform writes the running instance identifier and the training task continuous identifier into a mapping table, along with the running instance creation time and current status. When the job console page is refreshed, the training task card displays the training task continuous identifier, and the running instance identifier is displayed in the running information area of ​​the same card. Subsequent monitoring pages consistently display the training task continuous identifier as the entry point.

[0055] After step S1 is completed, the continuous identifier of the training task is used as a stable primary key to express the semantic identity of the training task, and the mapping table is used to express the unique correspondence between the running instance identifier and the continuous identifier of the training task. After the monitoring data stream forms a monitoring segment, it can retrieve the continuous identifier of the training task according to the mapping table and complete the attribution binding, so that the monitoring attribution in the migration and restart scenarios has a consistent entry point and a consistent judgment basis.

[0056] Step S1 has generated continuous training task identifiers and formed a mapping table, which establishes a one-to-one correspondence between running instance identifiers and continuous training task identifiers. When migration and restart occur, the monitoring data stream first carries the running instance identifier, and the switching event carries the running instance identifier change information and switching reason information. Monitoring records may be delayed or rearranged in the acquisition link. However, the monitoring segments must complete the verifiable segment attribution determination before binding the continuous training task identifiers in order to obtain continuous and reliable continuous training task records during the merging phase.

[0057] Detailed implementation of step S2:

[0058] S201 Unified Monitoring Record Construction and Switching Event Registration.

[0059] The runtime instance identifier is used for mapping table retrieval, the acquisition channel identifier is used to distinguish the source of boundary evidence, and the acquisition timestamp is used for slicing and alignment determination. After receiving the monitoring data stream, the platform performs normalization encapsulation on indicator-type records, log-type records, tracking-type records, and event-type records. The normalization encapsulation adds a runtime instance identifier, acquisition channel identifier, and acquisition timestamp to each monitoring record, forming a unified monitoring record and writing it to the monitoring record cache queue.

[0060] After a switching event occurs, the platform extracts the runtime instance identifier change information, switching reason information, and switching event occurrence timestamp from the switching event. The switching event occurrence timestamp is written into the switching event registration table, the switching reason information is written into the switching event registration table, and the runtime instance identifier change information is written into the switching event registration table. The switching event registration table is used to drive the transition constraints of subsequent slicing and sequence inference.

[0061] Once completed, the unified monitoring records will have alignable timestamps and traceable acquisition channel identifiers, and the switching events will have alignable switching event occurrence timestamps and explainable switching reason information. The monitoring segment generation will then enter the executable state.

[0062] S202 monitoring segment generation and initial determination of slice boundaries.

[0063] The switching event time stamp provides the slice center. Late arrivals and reordering of monitoring records require the slice boundaries to allow for echo coverage. After a switching event is registered, the platform triggers the slicing process, extracting a unified set of monitoring records from the monitoring record cache queue centered on the switching event time stamp, generating a monitoring segment. The monitoring segment contains the set of records within the segment and the switching event reference associated with the segment.

[0064] The initial segment boundary is determined using the event-center truncation rule: a range is truncated forward and backward from the switching event's time stamp. The truncation range is first formed using the minimum truncation span preset in the acquisition link to create the initial segment start and end time stamps, which are then written into the monitoring segment. After the boundary echo window is calculated in step S203, it is overwritten into the monitoring segment to correct the initial segment start and end time stamps to verifiable ones, ensuring that the monitoring segment can cover late and rearranged records near the switching event.

[0065] After completing step S202, the monitoring segment has formed a basic data carrier, the monitoring segment is associated with switching events, and subsequent index calculations have a clear input object.

[0066] During platform migrations, restarts, and other switching scenarios, the monitoring data stream simultaneously exhibits two types of uncertainty: one stems from the transmission characteristics of the acquisition link, where monitoring records may be delayed, rearranged, or exhibit asynchronous boundary signals on different acquisition channels; the other arises from the business semantics of the training process, where training tasks may continue along the same training semantic link, or semantic breaks may occur due to rollbacks, replays, or reinitializations. Attributing data solely based on runtime instance identifiers or switching event references can easily lead to incorrect segmentation of monitoring records near boundaries, or misjudging segments after semantic breaks as continuations of the same training task, resulting in conflicts between alarm attribution and operational phase interpretation. Therefore, it is necessary to establish separate criteria for the credibility of the monitoring side boundary and the semantic continuity of the training side: the switching timing alignment index is used to measure whether the switching event occurrence time stamp and the multi-channel boundary signal form a consistent arrival interval and consistent sequence relationship on the monitoring side, thereby determining whether the slice boundary is reliable and avoiding boundary drift caused by delays and rearrangements; the training semantic continuation index is used to measure the interpretable connection relationship of semantic anchors before and after the switch in terms of global step, recoverable point, data cursor and model construction signature, thereby determining whether the segment belongs to the continuation of the same training semantic link and avoiding misclassifying semantic restarts or reconstructions into the original link. These two parameters cover physical timing consistency and business semantic consistency, respectively, and together serve as the observation basis for sequence inference, so as to obtain verifiable segment attribution results in the real environment of dense switching and asynchronous acquisition. Therefore, the analysis of subsequent steps S203 at least S205 is carried out.

[0067] S203 Switching Timing Alignment Index Calculation.

[0068] The handover event registry records the timestamps of handover events, the monitoring record cache queue records the arrival rhythm of unified monitoring records, and the boundary echo window is used to cover boundary drift caused by delays and reordering. The platform first constructs the boundary echo window and then calculates the handover timing alignment index.

[0069] The boundary echo window construction involves two stages of computation.

[0070] The first calculation process calculates the water level advance lag time span: The platform reads and collects the water level advance information of the monitoring and collecting pipeline. The water level advance information of the monitoring and collecting pipeline includes the water level timestamp, which indicates the time position of the event to which the monitoring and collecting pipeline has stably advanced. The water level advance lag time span is defined as the non-negative time difference between the switching event occurrence timestamp and the water level timestamp. The non-negative time difference is the maximum of the difference between the two and the zero time span, ensuring that the water level advance lag time span is always a non-negative time span.

[0071] The second stage of the calculation process calculates the reporting jitter span: The platform traces back one sampling range from the time stamp of the switching event, and the sampling range is determined by the configuration of the monitoring acquisition pipeline; within the sampling range, the platform takes the acquisition time stamp intervals of adjacent unified monitoring records in the order of acquisition time stamps to obtain multiple acquisition time stamp intervals; the reporting jitter span is defined as the maximum value among the acquisition time stamp intervals minus the minimum value among the acquisition time stamp intervals, ensuring that the reporting jitter span has a time dimension and reflects the dispersion of the acquisition rhythm.

[0072] The half-width of the boundary echo window is defined as the sum of the water level advance lag time span and the reported jitter span. The left boundary of the boundary echo window is defined as the switching event occurrence time stamp minus the half-width of the boundary echo window. The right boundary of the boundary echo window is defined as the switching event occurrence time stamp plus the half-width of the boundary echo window. The boundary echo window is written into the monitoring segment and used for subsequent boundary signal consistency verification.

[0073] The multi-channel boundary consistency check employs executable boundary signal extraction rules. The platform extracts a set of boundary signals from the monitoring segments, which consists of four types of boundary signals.

[0074] The boundary signal corresponding to the log session reconstruction marker comes from the session identifier field of the log collection component. The log collection component generates the boundary signal time stamp when it detects a change in the session identifier.

[0075] The boundary signal corresponding to the reconnection marker of the indicator acquisition unit comes from the connection identifier field of the indicator acquisition component. The indicator acquisition component generates the boundary signal time stamp when it detects a change in the connection identifier.

[0076] The boundary signal corresponding to the link reconstruction marker originates from the link root identifier field of the tracking acquisition component. The tracking acquisition component generates the boundary signal time stamp when it detects a change in the link root identifier.

[0077] The boundary signal corresponding to the event queue replay marker comes from the consumption position field of the event consumption component. The event consumption component generates the boundary signal time stamp when it detects that the consumption position has been rolled back or reset.

[0078] After the boundary signal set is extracted, the platform performs two counts. The first count is the window hit count, which represents the number of boundary signals whose timestamps fall into the boundary echo window. The second count is the old-new sequence satisfaction count, which represents the number of pairs where the timestamps of the old running instance boundary signals are earlier than those of the new running instance boundary signals. The old and new running instance boundary signals are distinguished based on the running instance identifier change information in the switch event registration table.

[0079] The switching timing alignment index is defined as the sum of the window hit count and the old-new order satisfaction count, divided by the sum of the total number of boundary signals and the total number of old-new pairings. The switching timing alignment index remains dimensionless. The closer the switching timing alignment index is to one, the more concentrated the boundary evidence and the more consistent the old-new order. The closer the switching timing alignment index is to zero, the more dispersed the boundary evidence or the less consistent the order relationship.

[0080] After completing step S203, the boundary echo window and the switching timing alignment index both fall within the monitoring segment range, and the switching boundary judgment has traceable evidence.

[0081] S204 Training Semantic Continuation Index Calculation.

[0082] At the switching boundary of the monitored segment, it is necessary to simultaneously verify whether the training semantics continue. The training semantic anchor fingerprint provides the training semantic progression trajectory, and the training semantic continuation index provides a verifiable continuation determination. The platform first generates the semantic anchor fingerprint, and then performs continuation consistency deduction to obtain the training semantic continuation index.

[0083] Semantic anchor fingerprint generation employs structured extraction rules. The platform extracts training semantic anchors from monitoring segments, which are derived from the training framework's reporting interface and structured fields in the training logs.

[0084] The global step marker is extracted from the step field reported by the training framework. When the step field is missing, it is extracted from the step field in the training log. When both step fields are missing, a missing placeholder expression is written.

[0085] The recoverable point marker is extracted from the recoverable point field reported by the training framework. When the recoverable point field is missing, it is extracted from the recoverable point field in the training log. When both recoverable point fields are missing, a missing placeholder expression is written.

[0086] The data cursor marker is extracted from the data cursor field reported by the training framework. When the data cursor field is missing, it is extracted from the data cursor field in the training log. When both data cursor fields are missing, a missing placeholder expression is written.

[0087] The model build signature is extracted from the build signature field reported by the training framework. If the build signature field is missing, it is extracted from the build signature field in the training log. If both the build signature fields are missing, the build signature is derived from the task build information in the training task submission record and written into the model build signature.

[0088] The platform assembles training semantic anchors into semantic anchor fingerprints according to their order of appearance within the monitoring segment. The semantic anchor fingerprints are written into the monitoring segment, and the order relationship between anchor type and anchor value is preserved.

[0089] The continuity consistency inference uses four consistency checks and is summarized into a training semantic continuity index.

[0090] The step consistency check determines whether the global step marker after the switch meets the condition to continue advancing, or meets the condition to retreat from the recoverable point marker and then advance again. If the condition is met, it is marked as passed; if the condition is not met, it is marked as failed.

[0091] The cursor consistency check determines whether the data cursor markings after the switch meet the requirements for continued reading, or meet the mode of replaying subsequent readings from the recoverable point mark. If the conditions are met, it is marked as passed; if the conditions are not met, it is marked as failed.

[0092] The signature consistency check determines whether the model construction signature tag is consistent with that before the switch. When the switch reason information indicates that the model construction signature change is allowed, the signature consistency check is recorded as passed if it is within the allowed range, and as failed if it is outside the allowed range.

[0093] The recoverable point consistency check determines whether the recoverable point marker after the switch is consistent with that before the switch, or whether the recoverable point marker after the switch is in a missing placeholder expression. If the condition is met, it is marked as passed; if the condition is not met, it is marked as failed.

[0094] The training semantic continuation index is defined as the number of passed items divided by the total number of checked items. The training semantic continuation index remains dimensionless. The closer the training semantic continuation index is to one, the stronger the evidence of training semantic continuity. The closer the training semantic continuation index is to zero, the stronger the evidence of training semantic breakage.

[0095] In one embodiment, a migration occurs during the training task execution. The switching event registry records the switching event time stamp and switching reason information. The monitoring record cache queue shows changes in the session identifier of the log collection component and the connection identifier of the indicator collection component near the switching event time stamp. The boundary signal time stamps are concentrated in the boundary echo window, and the switching timing alignment index is close to one. The structured fields of the training log show that the recoverable point markers are consistent, the global step markers continue to advance from near the recoverable point markers, the training semantic continuation index is close to one, and the monitoring segment is determined to be in a continuous state in sequence inference.

[0096] After completing step S204, the monitored segment simultaneously possesses switching boundary evidence and training semantic evidence, and the sequence inference has verifiable input.

[0097] S205 sequence inference generates fragment attribution results and binds them to continuous identifiers for the training task.

[0098] The monitoring segments adjacent to the switching event may be delayed or rearranged. The judgment of a single segment is easily affected by local noise. Sequence inference incorporates the consistency of adjacent monitoring segments into the segment ownership status selection. The segment ownership result is used to fall to the unique running instance identifier and retrieve the mapping table to complete the binding.

[0099] The platform takes a continuous sequence of monitored segments as input, and the segment ownership status adopts three states: continuation state, transition state, and restart state. The platform constructs a emission score for each monitored segment and each segment ownership status. The emission score is based on the switching time sequence alignment index and the training semantic continuation index and adopts a logarithmic transformation. The emission score uses a positive lower bound to avoid the logarithm being undefined.

[0100] The emission score for the continuation state is the minimum of the switching timing alignment index and the training semantic continuation index, plus a positive lower bound and then the logarithm.

[0101] The transmit score in the transition state is calculated by subtracting the switching timing alignment exponent from one, adding a positive lower bound, and then taking the logarithm.

[0102] The launch score for the restart state is the minimum of the switching timing alignment index and the training semantic continuation index minus one, plus a positive lower bound and then the logarithm.

[0103] The platform constructs state transition constraints based on the switching reason information. State transitions that are not permitted by the switching reason information are given an unreachable penalty, which uses a fixed negative penalty value and is written into the transition score. The platform uses dynamic programming to calculate the optimal state path. Dynamic programming uses the sequence of monitored segments as the time axis, calculates the optimal cumulative score for each state for each segment, and records the predecessor state. The state with the highest cumulative score at the end of the monitored segment sequence is selected as the termination state, and the segment's state sequence is obtained by backtracking along the predecessor state.

[0104] The segment assignment result is determined by the segment assignment status and the change information of the running instance identifier. The continuation status corresponds to the running instance identifier before the switchover, the restart status corresponds to the running instance identifier after the switchover, and the transition status corresponds to the transition set of the running instance identifiers before and after the switchover. The unique running instance identifier of the transition status is selected from the assignment side whose boundary signal timestamp is closest to the timestamp of the switchover event. The assignment side is determined by the change direction of the running instance identifier in the switchover event registration table.

[0105] The platform obtains a unique running instance identifier based on the fragment attribution result, retrieves the training task continuity identifier from the mapping table using the unique running instance identifier, and binds the monitoring fragment to the training task continuity identifier. When binding the monitoring fragment, the switching sequence alignment index, training semantic continuation index, semantic anchor fingerprint, boundary echo window, fragment attribution result, and switching event reference are written synchronously to ensure that subsequent merging and consistency verification have a complete chain of evidence input.

[0106] The emission score is a numerical expression of the degree of matching of observed evidence in a certain state in a sequence inference model. It is used to quantify the consistency between a single monitoring segment and the observed evidence when the segment belongs to a given state. The emission score is constructed from observation features such as switching time alignment index, training semantic continuation index, semantic anchor fingerprint consistency conclusion, and boundary echo window hit situation into log-likelihood or equivalent scoring terms. These are used as inputs for dynamic programming or Viterbi decoding, so that the optimal state path is simultaneously affected by the strength of evidence of each segment and the state transition constraints of adjacent segments. In engineering implementation, emission probability modeling of Hidden Markov Model, state scoring modeling of Conditional Random Field, or logistic regression is used to map the observation features into state scores. The output participates in the optimal path search in a logarithmic domain accumulation manner.

[0107] In one embodiment, after the training task starts running, the metric acquisition component continuously reports changes in memory usage and throughput, the log acquisition component continuously reports training log lines, the tracking acquisition component continuously reports operator link information, and the event consumption component continuously reports queue consumption positions. The platform adds a running instance identifier, acquisition channel identifier, and acquisition timestamp to each monitoring record and writes it to a unified monitoring record cache queue. During training, a migration occurs, and a switch event arrives carrying switch reason information and running instance identifier change information. The platform extracts monitoring records centered on the timestamp of the switch event to generate monitoring segments and writes the switch event reference into the monitoring segments. Subsequently, the platform identifies log session reconstruction markers and metric acquisition device reconnection markers within the monitoring segments, and simultaneously identifies recoverable point markers and global step markers in the structured fields of the training logs. Based on this, the platform calculates the switch sequence alignment index and training semantic continuation index, and performs sequence inference on adjacent monitoring segments to obtain the segment attribution result. After the monitoring page is refreshed, the monitoring curves before and after the migration are strung together under the same continuous training task identifier, and the attribution column of the alarm list maintains the same continuous training task identifier, no longer splitting into two tasks due to changes in the running instance identifier.

[0108] After step S2 is completed, the monitoring data stream forms a monitoring segment under the switching event. The monitoring segment extracts boundary evidence through the boundary echo window and calculates the switching timing alignment index. It extracts training semantic evidence through semantic anchor fingerprint and calculates the training semantic continuation index. Sequence inference outputs the segment attribution result and binds the monitoring segment to the corresponding training task continuous identifier in the mapping table. The monitoring attribution in the migration and restart scenarios enters a stable state that is verifiable and traceable.

[0109] Step S2 has already output the monitoring segment, and the monitoring segment has been bound to the training task continuity identifier, while also carrying the segment attribution result, switching event reference, switching time sequence alignment index, training semantic continuation index, semantic anchor fingerprint, and boundary echo window. Migration and restart will split the same training task into multiple monitoring segments. The monitoring segments need to be organized into a continuous time chain and given a switching explanation. However, if the switching explanation lacks a unified generation rule, the alarm attribution and changes in the running stage can easily lead to inconsistencies in the interpretation of different pages.

[0110] Detailed implementation of step S3:

[0111] S301 Monitoring Segment Grouping and Segment Order Relationship Construction.

[0112] The continuous identifier of the training task serves as the primary key for monitoring attribution. Monitoring segments are first grouped according to the continuous identifier of the training task. The grouping results form an index table of the monitoring segment set corresponding to the continuous identifier of the training task. The index table records the list of monitoring segment identifiers associated with each continuous identifier of the training task.

[0113] When constructing the segment order relationship within the monitoring segment set, the start and end timestamps of each monitoring segment are first read. The segment order relationship is generated using a stable sorting rule, with the sorting keys sequentially being the segment start timestamp, segment end timestamp, switching event occurrence timestamp, and the acquisition timestamp of the first unified monitoring record within the segment. During sorting, only when the preceding sorting key is the same will the subsequent sorting key be compared, ensuring the sorting result is unique and verifiable. After sorting, adjacent monitoring segments are grouped into a sequential connection record, which is written into the segment order relationship. The segment order relationship is stored as an order field in the continuous records of the training task.

[0114] The order of the segments is also written into the connection type field. The connection type is determined by the segment affiliation result of two adjacent monitoring segments and the reference of the switching event: if the running instance identifiers corresponding to the two adjacent monitoring segments are the same, it is recorded as natural continuation; if the running instance identifiers corresponding to the two adjacent monitoring segments are different, it is recorded as switching transition; when the switching event reference is missing, the connection type is recorded as unmarked and written into the missing placeholder expression. The missing placeholder expression is used for subsequent consistency verification to locate evidence gaps.

[0115] S302 Training Task: Continuous Record Assembly and Original Index Summary Generation.

[0116] The training task continuously records data using the segment order as the main thread, writing monitoring segment information sequentially along this order. Each monitoring segment contains an original index and a summary; the original index is used for tracing and locating, while the summary is used for reading and display.

[0117] The original index consists of a unified set of monitoring record storage locations, which are grouped according to the acquisition channel identifier. Each acquisition channel identifier corresponds to one storage location record, which includes the acquisition channel identifier, the earliest acquisition time stamp, the latest acquisition time stamp, and the storage partition identifier. The earliest acquisition time stamp is taken from the smallest acquisition time stamp of the unified monitoring records under the same acquisition channel identifier within the monitoring segment, and the latest acquisition time stamp is taken from the largest acquisition time stamp of the unified monitoring records under the same acquisition channel identifier within the monitoring segment. The storage partition identifier is taken from the partition information returned when the monitoring acquisition link is written; if the storage partition identifier is missing, a placeholder expression is written. The list of storage location records is written into the original index field of the monitoring segment to ensure that any monitoring segment has a locatable entry point on the storage side.

[0118] The summary information is extracted from semantic anchor fingerprints, switching temporal alignment indices, and training semantic continuation indices. Semantic anchor fingerprint summaries use first and last anchors; the first anchor is the first anchor record in the semantic anchor fingerprint, and the last anchor is the last anchor record. When a semantic anchor fingerprint is empty, the first and last anchors are written into a placeholder representation. When the switching temporal alignment indices and training semantic continuation indices are written into the summary information, their original values ​​are retained, and the generated timestamp is recorded. The generated timestamp is taken from the end timestamp of the monitored segment, ensuring that the summary information and the monitored segment have consistent time boundaries.

[0119] In one embodiment, the training task undergoes a migration, switching event references to enter the monitoring segment, and the monitoring segment is bound to the training task continuity identifier. The platform aggregates the monitoring segment set according to the training task continuity identifier, completes stable sorting according to the segment start time index, and generates segment order relationships; the training task continuity record is written to the storage location record list of each monitoring segment along the segment order relationship, and the first and last anchor points of the semantic anchor fingerprint are written. The training semantic trajectory before and after the migration is continuously presented in the same training task continuity record.

[0120] S303 Switching Instructions: The elements for generating and explaining the switchover instructions are fixed.

[0121] Switching instructions are generated between two adjacent monitoring segments, and each switching instruction corresponds one-to-one with the sequential connection record in the segment sequence relationship. Before generating the switching instructions, the interpretation focus is determined, which is used to decide whether the switching instructions mainly cite boundary evidence or semantic evidence.

[0122] The interpretation focus is determined using a textual calculation process. First, the handover timing alignment index is read from two adjacent monitoring segments, and the smaller value of the two handover timing alignment indices is selected. Then, the training semantic continuation index is read from two adjacent monitoring segments, and the smaller value of the two training semantic continuation indices is selected. The smaller value of the handover timing alignment index is subtracted from the smaller value of the training semantic continuation index to obtain the interpretation difference. When the interpretation difference is less than zero, the interpretation focus is selected as boundary evidence; when the interpretation difference is greater than zero, the interpretation focus is selected as semantic evidence; when the interpretation difference is equal to zero, the interpretation focus is selected as handover reason information.

[0123] The switching description text is generated using a fixed field template. This template includes switching reason information, switching event timestamp, boundary echo window, boundary signal set summary, semantic anchor fingerprint summary, switching time sequence alignment index summary, training semantic continuation index summary, and segment attribution result summary. The boundary signal set summary extracts window hit information from log session reconstruction markers, indicator collector reconnection markers, tracking link reconstruction markers, and event queue replay markers. Window hit information is expressed as a hit type list plus a hit timestamp range. The semantic anchor fingerprint summary directly references the first and last anchors of two adjacent monitoring segments. When the interpretation focus is on boundary evidence, the switching description prioritizes filling the boundary echo window and boundary signal set summary; when the interpretation focus is on semantic evidence, the switching description prioritizes filling the semantic anchor fingerprint summary and marking semantic discontinuities; when the interpretation focus is on switching reason information, the switching description prioritizes filling the switching reason information and writes the boundary evidence and semantic evidence in summary form. The switching description is written to the continuous records of the training task and bound to the sequential connection records, ensuring that the switching description references the same record on both the display and verification ends.

[0124] In one embodiment, the training task undergoes one migration and one restart, and the platform has obtained multiple monitoring segments and completed the binding of continuous training task identifiers. The platform aggregates monitoring segments with the same continuous training task identifier into a monitoring segment set, and performs stable sorting according to the segment start time and the switching event occurrence time, generating a segment order relationship. The continuous training task record is written to the original index of each monitoring segment, which is displayed as the acquisition time range and storage partition identifier corresponding to the acquisition channel identifier; the continuous training task record is also written with the first and last anchor points of the semantic anchor fingerprint, and the page can see that the training semantics enters the segment from a certain step position and ends the segment at another step position. The platform generates switching descriptions between adjacent monitoring segments, which directly reference the switching reason information, the boundary echo window range, the boundary signal hit status, and the semantic anchor change conclusions; when the operation and maintenance personnel open any alarm, the alarm details page can display which monitoring segment the alarm belongs to and can see the switching description text of the adjacent segment.

[0125] After step S3 is completed, the set of monitoring segments corresponding to the continuous identifiers of the training task is stably sorted and a segment order relationship is generated. The continuous records of the training task are written into the original index and summary information according to the segment order relationship. The switching instructions are generated according to the adjacent monitoring segments and the interpretation elements are fixed. The monitoring segment splits caused by migration and restart are organized into traceable, readable and verifiable continuous records.

[0126] Step S3 has generated a continuous training task record, which includes the segment order relationship, switching instructions, original index of the monitoring segment, and summary information. The monitoring segments have been bound to training task continuous identifiers in step S2 and carry boundary echo windows, semantic anchor fingerprints, switching sequence alignment indices, and training semantic continuation indices. Delays and reorderings caused by migration and restarts may still assign a small number of monitoring segments to incorrect training task continuous identifiers. Changes in alarm attribution and runtime stages lead to interpretation conflicts. However, consistency verification requires pinpointing the conflict to the specific monitoring segment and switching event reference, and providing executable correction instructions.

[0127] Detailed implementation of step S4:

[0128] S401 Consistency Check Input Organization and Check Pair Construction.

[0129] Migration and restart split the same training task into multiple monitoring segments, and the segment order relationship can provide fixed pairs of adjacent monitoring segments. After reading the continuous records of the training task, adjacent monitoring segment pairs are extracted sequentially according to the segment order relationship to form a check pair sequence.

[0130] For each adjacent monitoring segment pair, the following fields are extracted and written into a checksum: segment start time stamp, segment end time stamp; switching event reference corresponding to the switching event occurrence time stamp and switching reason information; left and right boundaries of the boundary echo window; boundary signal set and the acquisition channel identifier corresponding to the boundary signal; semantic anchor fingerprint first and last anchor points; alarm occurrence time stamp and alarm belonging segment identifier in the alarm summary. When a field is missing, a placeholder expression is written, and this placeholder expression is added to the missing field list referenced by the abnormal evidence.

[0131] After the checksum sequence is formed, consistency verification uses the checksum sequence as the sole input to avoid inconsistencies caused by repeatedly taking data from multiple locations. The checksum sequence is written into the check object list of the consistency verification results, ensuring that each piece of abnormal evidence references can point back to the original index of the specific monitoring segment.

[0132] S402 Fragment Boundary Consistency Verification and Boundary Anomaly Evidence Citation.

[0133] Switching boundaries requires consistent evidence on the monitoring side. The degree of overlap of boundary echo windows and the concentration of boundary signals within the overlapping interval directly reflect boundary consistency. For each check pair, the overlapping interval of the boundary echo windows is calculated first, followed by the boundary consistency score.

[0134] The overlap interval of the boundary echo window is calculated by taking the larger of the left boundaries of the two windows as the left end of the overlap interval and the smaller of the right boundaries of the two windows as the right end of the overlap interval. When the right end of the overlap interval is earlier than the left end, the length of the overlap interval is recorded as zero. The overlap ratio of the boundary echo window is calculated by dividing the length of the overlap interval by the length of the union of the two windows. The union length is calculated by subtracting the smaller of the left boundaries of the two windows from the larger of the right boundaries of the two windows. When the union length is zero, it is replaced by a positive lower bound to ensure that the overlap ratio is defined.

[0135] The hit ratio of boundary signals is calculated around the overlapping interval. The boundary signal sets of the preceding and subsequent monitoring segments are merged into a single boundary signal time stamp list. Each boundary signal time stamp is checked to see if it falls into the overlapping interval. If it does, it is counted as a hit. The hit ratio is calculated by dividing the hit ratio by the total number of times in the boundary signal time stamp list. If the total number is zero, the total number is replaced by a positive lower bound to ensure that the hit ratio is defined.

[0136] The boundary consistency score is the smaller of the boundary echo window overlap ratio and the boundary signal hit ratio. When the boundary consistency score is lower than the consistency judgment threshold, a boundary anomaly evidence reference is written. The boundary anomaly evidence reference includes the overlapping interval boundary; a list of missed boundary signal timestamps; a list of acquisition channel identifiers corresponding to the missed boundary signals; a list of storage location records in the original index of the corresponding monitoring segment; and a reference to the corresponding switching event. The boundary anomaly evidence reference is written into the consistency verification result and the relevant monitoring segment identifier is marked to ensure that the correction command can directly locate the monitoring segment range that needs to be adjusted.

[0137] S403 Semantic Anchor Continuity Verification and Citation of Semantic Anomaly Evidence.

[0138] Semantic breaks in training can lead to discontinuous progress trajectories in continuous training task recordings. Semantic anchor fingerprints can pinpoint the breakpoints down to the anchor point level. For each verification pair, the tail anchor point of the semantic anchor fingerprint of the preceding monitoring segment and the first anchor point of the semantic anchor fingerprint of the following monitoring segment are read. The global step marker, recoverable point marker, data cursor marker, and model construction signature marker are extracted respectively. Four consistency checks are performed, and the semantic anchor continuity score is summarized.

[0139] The global step consistency determination uses two paths: if the subsequent global step mark is not less than the previous global step mark, it is determined to pass; if the subsequent recoverable point mark is consistent with the previous recoverable point mark, it is determined to pass.

[0140] The data cursor consistency determination uses two paths: if the subsequent data cursor mark is not less than the previous data cursor mark, it is considered to pass; if the subsequent recoverable point mark is consistent with the previous recoverable point mark, it is considered to pass.

[0141] The consistency determination of model construction signatures uses two paths: if the signature marker of the subsequent model construction is consistent with the signature marker of the previous model construction, it is considered to pass; if the switching reason information indicates that the model construction signature change is allowed and the change type falls within the allowed range, it is considered to pass.

[0142] The consistency determination of recoverable points adopts two paths: if the subsequent recoverable point marker is consistent with the preceding recoverable point marker, it is considered to pass; if the subsequent recoverable point marker is a missing placeholder expression, it is considered to pass.

[0143] Each of the four criteria outputs either "pass" or "fail". The semantic anchor continuity score is calculated by dividing the number of passed items by the total number of criteria, which is fixed at four. When the semantic anchor continuity score is lower than the consistency threshold, a semantic anomaly evidence reference is written. The semantic anomaly evidence reference includes the preceding tail anchor and the following first anchor; a list of failed criteria; a list of storage location records in the original index of the corresponding monitoring segment; the corresponding switch reason information segment; and the corresponding switch event reference.

[0144] In one embodiment, the training task is restarted after migration, a pair of adjacent monitoring segments appear in the verification sequence, the overlap interval of the boundary echo window is very short, and the boundary signal time stamps are concentrated outside the overlap interval; the semantic anchor fingerprint shows that the recoverable point markers are consistent, and the global step markers continue to advance from the vicinity of the recoverable point markers; the consistency verification result is written to the boundary abnormal evidence reference but not to the semantic abnormal evidence reference, and the correction instruction focuses on the consistency of the slice boundary and the mapping table switching.

[0145] S404 alarm attribution consistency verification, correction command generation, mapping table update, record reconstruction, and rule revision.

[0146] Alarm attribution must fall within the time interval of the assigned segment. Alarms falling outside this interval will directly cause conflicts in the interpretation of continuous records in the training task. After reading the alarm summary from the continuous records of the training task, the alarm occurrence time stamp and alarm attribution segment identifier are extracted for each alarm. The corresponding monitoring segment is located based on the alarm attribution segment identifier, and the segment start time stamp and segment end time stamp are read. Alarm occurrence times stamps falling within the interval between the segment start time stamp and segment end time stamp are considered consistent; those falling outside the interval are considered inconsistent. The number of inconsistent alarms divided by the total number of alarms yields the alarm attribution anomaly ratio. When the total number of alarms is zero, the alarm attribution anomaly ratio is recorded as zero.

[0147] The correction trigger determination uses the maximum value of the three anomaly severity levels. Boundary anomaly severity is determined by subtracting the boundary consistency score from the boundary score; semantic anomaly severity is determined by subtracting the semantic anchor continuity score from the semantic anchor continuity score; and alarm attribution anomaly severity is determined by the alarm attribution anomaly ratio. The maximum value of the three anomaly severity levels is taken as the fragment attribution anomaly score. A correction instruction is generated when the fragment attribution anomaly score exceeds the anomaly trigger threshold.

[0148] The correction instructions include the range of running instance identifiers, the range of monitoring segments, and the candidate set of continuous identifiers for training tasks. All three parts are derived from the reference of abnormal evidence.

[0149] The scope of the running instance identifier is derived from the switching event reference in the abnormal evidence reference, which extracts the switching event occurrence time stamp and the running instance identifier change information. The running instance identifier before and after the switching in the running instance identifier change information together constitute the scope of the running instance identifier.

[0150] The monitoring segment range is obtained by directly extracting the relevant monitoring segment identifiers from the abnormal evidence references, and incorporating the monitoring segment identifiers that are adjacent to the relevant monitoring segments in the segment order relationship into the monitoring segment range to ensure that the rebinding covers the boundary neighborhood.

[0151] The candidate set of continuous training task identifiers prioritizes consistency of model construction signature tags: Model construction signature tags are extracted from the monitored segment range; task construction information corresponding to the continuous training task identifiers is read from the local registry and model construction signature tags are derived; continuous training task identifiers with consistent model construction signature tags are added to the candidate set. When the candidate set is empty, a new candidate set is generated based on the consistency of the task's spatial information.

[0152] The mapping table update performs atomic replacement according to the correction instruction. For each running instance identifier within the range of running instance identifiers, an atomic replacement of the mapping table is performed once, replacing the training task contiguous identifier corresponding to the running instance identifier with a selected item from the candidate set of training task contiguous identifiers. The selection of the item adopts the time proximity rule: read the running instance creation time and the reference to the most recent switch event in the mapping table record, and select the training task contiguous identifier whose running instance creation time is closer to the time stamp of the switch event, ensuring that the training task contiguous identifier pointed back to by the running instance identifier and the switch event reference maintain the same temporal context.

[0153] After the mapping table is updated, the monitoring segment range is rebounded. Rebinding retrieves the training task continuity identifier from the mapping table using the instance identifier and overwrites it in the monitoring segment binding result. After rebinding, the training task continuity record is regenerated according to the merging rules in step S3 and overwritten in the training task continuity record storage location. The switching descriptions within the training task continuity record are reconstructed during the merging process to avoid old interpretations remaining.

[0154] The boundary echo window construction rule revision extracts the boundary signal superwindow offset from boundary anomaly evidence references. For each boundary anomaly evidence reference, the time stamp of the switching event and the time stamp of the missed boundary signal are read; the absolute value of the time difference between the boundary signal time stamp and the switching event time stamp is calculated for each reference; the maximum absolute value of the time difference is taken as the coverage span of this revision; the coverage span is written into the minimum half-width field of the boundary echo window construction rule, and the lower half-width is written according to the acquisition channel identifier to ensure that the late arrival and rearrangement characteristics of the same acquisition channel are continuously absorbed.

[0155] The semantic anchor inference rule revision extracts passed and failed patterns from semantic anomaly evidence citations. Passed patterns are written into the rule base, which stores patterns of recoverable point marker changes, global step marker changes, data cursor marker changes, and model construction signature marker changes. Failed patterns are written into the prohibited list in the rule base and associated with switching reason information to ensure that subsequent continuity inferences can provide stable judgments when faced with similar switching reason information.

[0156] In one embodiment, continuous training task records have been generated and displayed on the monitoring page. Operations personnel discover that an alarm occurred after migration, but its affiliation falls within a monitoring segment from before the migration. The platform reads the continuous training task records, extracts adjacent monitoring segment pairs according to the segment sequence, first checks the overlap of boundary echo windows, then checks whether boundary signals are concentrated in the overlapping interval, and simultaneously performs continuity determination on the first and last anchor points of the semantic anchor fingerprint. It also performs interval inclusion determination on the alarm occurrence timestamp and the segment start and end timestamps of the alarm's affiliation segment. The inspection results indicate that the boundary signals are concentrated outside the overlapping interval, and the alarm occurrence timestamp falls outside the affiliation segment interval. The abnormal evidence references the storage location record and switching event reference corresponding to the log collection component. Based on this, the platform generates a correction instruction, listing the involved running instance identifier range and monitoring segment range, and filters out a candidate set of continuous training task identifiers from the consistency of model construction signature markers. After the mapping table completes atomic replacement, the platform rebinds the monitoring segments to continuous training task identifiers and reconstructs the continuous training task records. When the monitoring page is reopened, the alarm for the previously misplaced segment is returned to the monitoring segment after migration, the switching description text is updated synchronously, and the changes before and after migration are explained consistently in the continuous records of the same training task.

[0157] Step S4 outputs the consistency verification result and the corrected training task continuity record. The consistency verification result includes the boundary consistency score, semantic anchor continuity score, alarm attribution anomaly ratio, and anomaly evidence reference; the correction instruction clarifies the scope of the running instance identifier, the scope of the monitoring segment, and the candidate set of training task continuity identifiers; after the mapping table is atomically replaced, the monitoring segment is rebound and the training task continuity record is rebuilt; the boundary echo window construction rules and semantic anchor inference rules are updated veribly based on the anomaly evidence reference.

[0158] Example 2: Figure 2 The present invention provides a monitoring and management system for an artificial intelligence development platform, comprising: a continuous identification module, a segment attribution module, a continuous merging module, and a consistency correction module;

[0159] Continuous Identification Module: Generates continuous identification of training tasks based on the submission record of training tasks. After the scheduling side generates the running instance identification, it establishes a mapping table between the running instance identification and the continuous identification of training tasks.

[0160] Segment Attribution Module: Collects the monitoring data stream of the training task and forms a monitoring segment when a switching event occurs. For each monitoring segment, it calculates the switching timing alignment index and the training semantic continuation index and performs sequence inference to obtain the segment attribution result. The monitoring segment is bound to the continuous identifier of the training task according to the segment attribution result.

[0161] Continuous merging module: Monitoring segments are merged according to the continuous identifier of training tasks to generate continuous records of training tasks. The continuous records of training tasks retain the segment order relationship and generate switching instructions. The switching instructions are used to explain the alarm attribution and changes in the operation stage.

[0162] Consistency Correction Module: Performs consistency checks on continuous records of the training task to obtain consistency check results. When the consistency check result points to an abnormal segment assignment, a correction instruction is generated and the mapping table is updated accordingly, and continuous records of the training task are re-merged. At the same time, the consistency check result is used to revise the boundary echo window construction rules and semantic anchor inference rules to improve the stability of subsequent segment assignment.

[0163] Specifically, the above description is only a preferred embodiment of this application and is not intended to limit this application.

[0164] The consistency determination threshold and the anomaly trigger threshold can be pre-calibrated through offline simulation testing, or set to fixed values ​​according to the on-site operating procedures.

[0165] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0166] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A monitoring and management method for an artificial intelligence development platform, characterized in that, Including the following steps: S1: Generate continuous identification of training tasks based on the submission record of training tasks. After the scheduling side generates the running instance identification, establish a mapping table between the running instance identification and the continuous identification of training tasks. S2: Collect the monitoring data stream of the training task and form a monitoring segment when the switching event arrives; construct a boundary echo window for the monitoring segment. The boundary echo window construction rule is as follows: add the non-negative time difference between the switching event occurrence time marker and the water level time marker to the difference between the maximum and minimum values ​​in the time interval of adjacent unified monitoring records within the sampling range to obtain the half-width of the boundary echo window. Subtract and add the half-width of the boundary echo window from the switching event occurrence time marker to obtain the left and right boundaries of the boundary echo window, respectively; extract the number of boundary signals falling into the boundary echo window as the window hit count, extract the number of pairs where the time marker of the boundary signal of the old running instance is earlier than the time marker of the boundary signal of the new running instance as the old-new order satisfaction count, and then construct the window. The hit count and the old-new order satisfy the condition that the sum of the counts is divided by the sum of the total number of boundary signals and the total number of old-new pairs to obtain a dimensionless switching timing alignment index between 0 and 1. Global step markers, recoverable point markers, data cursor markers, and model construction signature markers are extracted from the monitoring segment. The semantic anchor inference rule is as follows: based on the global step markers, recoverable point markers, data cursor markers, and model construction signature markers, step consistency checks, cursor consistency checks, construction signature consistency checks, and recoverable point consistency checks are performed, and a training semantic continuation index is generated based on the number of passed items and the total number of checked items. The training semantic continuation index is calculated according to the number of passed items / the total number of checked items, and the training semantic continuation index is a dimensionless value between 0 and 1. Based on the switching timing alignment index and the training semantic continuation index, the execution sequence inference is performed to obtain the segment attribution result of the running instance identifier before the switch, the running instance identifier after the switch, or the transition set consisting of the running instance identifier before the switch and the running instance identifier after the switch corresponding to the monitoring segment. The unique running instance identifier is determined according to the segment attribution result, and the corresponding training task continuous identifier is retrieved and bound through the mapping table. S3: Monitoring segments are merged according to the continuous identifier of training tasks to generate continuous training task records. The continuous training task records retain the segment order relationship and generate switching instructions. The switching instructions are used to explain the alarm attribution and changes in the operational phase. S4: Perform consistency checks on continuous training task records to obtain consistency check results. When the consistency check result points to an abnormal segment assignment, a correction instruction is generated. The correction instruction determines the range of running instance identifiers, the range of monitored segments, and the candidate set of continuous training task identifiers. It then performs atomic replacement of the mapping table on the running instance identifiers within the range of running instance identifiers to replace their corresponding continuous training task identifiers with the selected items in the candidate set. Based on this, the continuous training task identifiers are rebound for the monitored segments and the continuous training task records are re-merged. At the same time, the consistency check results are used to revise the boundary echo window construction rules and semantic anchor inference rules to improve the stability of subsequent segment assignments.

2. The monitoring and management method for an artificial intelligence development platform according to claim 1, characterized in that, Step S1 includes the following: Continuous identifier source information is extracted based on training task submission records. The continuous identifier source information includes the spatial information of the task, the task definition information, and the task construction information. The continuous identifier source information is encoded uniformly, null values ​​are regularized, key order is fixed, and delimiters are escaped. A continuous identifier input string is generated according to a fixed concatenation rule.

3. The monitoring and management method for an artificial intelligence development platform according to claim 2, characterized in that, Step S1 also includes the following: The continuous identifier input string is converted into a byte sequence and input into a one-way fingerprint function to generate continuous identifiers for training tasks. When conflicts are found in the local registry, the continuous identifiers for training tasks are used to generate conflict resolution factors by using submission serial information and submission timestamp information, and then the continuous identifiers for training tasks are regenerated. After obtaining the runtime instance identifier, a mapping table is established. The mapping table maintains a one-to-one correspondence between the runtime instance identifier and the continuous identifier of the training task through atomic commits.

4. The monitoring and management method for an artificial intelligence development platform according to claim 3, characterized in that, Step S2 includes the following: Normalization and encapsulation are performed on indicator-type records, log-type records, tracking-type records, and event-type records in the monitoring data stream. Each monitoring record is appended with a running instance identifier, a collection channel identifier, and a collection timestamp to form a unified monitoring record. When a switch event arrives, the change information of the running instance identifier, the switch reason information, and the switch event occurrence timestamp are registered, and monitoring segments are generated based on this information using the event center.

5. The monitoring and management method for an artificial intelligence development platform according to claim 4, characterized in that, Step S2 also includes the following: For each monitoring segment, a boundary echo window is constructed and log session reconstruction markers, indicator collector reconnection markers, tracing link reconstruction markers, and event queue replay markers are extracted to form a boundary signal set for calculating the switching timing alignment index. For each monitoring segment, global step markers, recoverable point markers, data cursor markers, and model building signature markers are extracted to form a semantic anchor fingerprint for calculating the training semantic continuation index. Based on sequence inference, the segment attribution result is obtained and the training task continuous identifier is bound through a mapping table.

6. The monitoring and management method for an artificial intelligence development platform according to claim 5, characterized in that, Step S3 includes the following: The monitoring segments are grouped into a monitoring segment set according to the continuous identifier of the training task. The start time stamp, end time stamp, switching event occurrence time stamp, and first unified monitoring record collection time stamp of the segment are read to perform stable sorting. The segment order relationship is generated according to the adjacent monitoring segments, and the connection type is determined according to the segment ownership result and switching event reference.

7. The monitoring and management method for an artificial intelligence development platform according to claim 6, characterized in that, Step S3 also Includes the following: The training task continuous records are assembled according to the segment order relationship. The training task continuous records are written into the original index and summary information. The original index includes the acquisition channel identifier, the earliest acquisition time stamp, the latest acquisition time stamp, and the storage partition identifier. The interpretation difference between adjacent monitoring segments is calculated to determine the interpretation center. Switching descriptions are generated based on fixed field templates and bound to the segment order relationship.

8. The monitoring and management method for an artificial intelligence development platform according to claim 7, characterized in that, Step S4 includes the following: Read the continuous records of the training task and construct a sequence of adjacent monitoring segments based on the segment order relationship. Extract the segment start time stamp, segment end time stamp, boundary echo window, boundary signal set, semantic anchor fingerprint first anchor point, semantic anchor fingerprint tail anchor point, switching event reference, alarm occurrence time stamp, form consistency verification input and generate abnormal evidence reference.

9. The monitoring and management method for an artificial intelligence development platform according to claim 8, characterized in that, Step S4 also includes the following: Calculate the boundary consistency score, semantic anchor continuity score, alarm attribution anomaly ratio, and synthesize the segment attribution anomaly score. After the segment attribution anomaly score triggers the correction instruction, determine the running instance identifier range, the monitoring segment range, and the training task continuous identifier candidate set. Perform atomic replacement of the mapping table and rebind the monitoring segment to the training task continuous identifier. Reconstruct the training task continuous record and revise the boundary echo window construction rules and semantic anchor inference rules.

10. A monitoring and management system for an artificial intelligence development platform, used to implement the monitoring and management method for an artificial intelligence development platform as described in any one of claims 1-9, characterized in that, include: Continuous Identification Module: Generates continuous identification of training tasks based on the submission record of training tasks. After the scheduling side generates the running instance identification, it establishes a mapping table between the running instance identification and the continuous identification of training tasks. Segment Attribution Module: Collects the monitoring data stream of the training task and forms a monitoring segment when the switching event arrives. For each monitoring segment, it calculates the switching timing alignment index and the training semantic continuation index and performs sequence inference to obtain the target running instance identifier corresponding to the monitoring segment as the segment attribution result. Based on the mapping table, it binds the monitoring segment to the training task continuation identifier corresponding to the target running instance identifier. Continuous merging module: Monitoring segments are merged according to the continuous identifier of training tasks to generate continuous records of training tasks. The continuous records of training tasks retain the segment order relationship and generate switching instructions. The switching instructions are used to explain the alarm attribution and changes in the operation stage. Consistency Correction Module: Performs consistency checks on continuous training task records to obtain consistency check results. When the consistency check result points to an abnormal segment assignment, a correction instruction is generated. Based on this, the mapping relationship between the running instance identifier and the continuous training task identifier within the abnormal range is replaced and updated. The relevant monitoring segments are rebound to the continuous training task identifier and the continuous training task records are re-merged. At the same time, the consistency check result is used to revise the boundary echo window construction rules and semantic anchor inference rules to improve the stability of subsequent segment assignment.

Images