A big data-based communication machine room inspection method and system

Abstract

This invention discloses a communication equipment room inspection method and system based on big data, belonging to the field of big data technology. The method includes: processing equipment room inspection data to generate standard inspection data and dual-time-scaled response evidence envelopes; identifying operational disturbance events and matching response contract versions to generate an equipment room operational condition response contract diagram; correcting equipment response time and generating a response weaving tensor; inputting the response weaving tensor and the equipment room operational condition response contract diagram into an improved DAGMM model to generate response source accounting results and original default anti-chain; generating candidate fault causes and screening reversible perturbation actions to form verification instructions; executing perturbation verification to determine the target fault cause and generate inspection tasks; executing inspection tasks and updating inspection data to generate inspection results. This invention achieves accurate fault location in communication equipment rooms, improving inspection efficiency and result reliability.

Description

Technical Field

[0001] This invention relates to the field of big data technology, and in particular to a method and system for inspecting communication equipment rooms based on big data. Background Technology

[0002] As the scale of communication equipment rooms expands, communication equipment, power supply and distribution equipment, cooling equipment, environmental sensors, and data acquisition links continuously generate operational data, load data, configuration data, alarm data, and inspection feedback data. Existing communication equipment room inspection methods mostly employ fixed-period data collection, parameter threshold judgment, alarm correlation, historical statistical analysis, or anomaly detection models to identify equipment status and generate on-site inspection tasks based on anomaly results.

[0003] Existing methods often rely on deviations in parameters of a single device as the basis for judgment, lacking modeling of cross-device response processes caused by power supply switching, load changes, cooling start-up and shutdown, and control strategy adjustments. Different devices exhibit variations in response latency, duration, and recovery process. Furthermore, there are discrepancies between device time, gateway reception time, and platform data entry time. Fixed time windows can easily misinterpret data transmission delays as device response delays, making it difficult to accurately distinguish between upstream anomaly propagation, target device local faults, and data acquisition anomalies.

[0004] Some methods employ deep learning models to process multidimensional data center data, but the input data does not retain response direction, parent response relationship, propagation order, and operational condition contract version. The model output is mostly a single anomaly score, failing to decompose the effects of parent device inheritance, multi-device coupling, common operational conditions, insufficient local response, and insufficient data evidence. Inspection tasks are mostly generated directly based on abnormal equipment, lacking an active verification process for candidate fault causes, making it difficult to form a closed loop that connects response source location, fault cause confirmation, and inspection feedback updates.

[0005] Therefore, how to provide a communication equipment room inspection method and system based on big data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a communication equipment room inspection method and system based on big data. It adopts a dual-time-scaled response evidence envelope, a working condition response contract diagram, and an improved DAGMM model to identify cross-device response mismatches and verify the cause of faults through reversible perturbations. It has the advantages of accurate anomaly location, low misjudgment rate, strong inspection targeting, and complete feedback loop.

[0007] A communication equipment room inspection method based on big data according to an embodiment of the present invention includes: Collect data from the computer room inspection, perform unified identification, time correction, object association, and complete status marking, and generate standard computer room inspection data and dual time-scaled response evidence envelopes; Identify operational disturbance events from standard data center inspection data, match operational response contract versions, and generate a data center operational response contract diagram; Based on the dual-timescaled response evidence envelope correction device response time, response phase strips are constructed and arranged to generate a response weaving tensor; The response weaving tensor and the data center operating condition response contract graph are input into the improved DAGMM model. The improved DAGMM model includes a response compression and reconstruction module, a parent edge blocking combination inversion module, a duty Gaussian estimation module, and a default evidence output module to generate response source accounting results and original default anti-chain. Extract the original default response edge from the original default antichain, generate candidate failure causes by combining the response source accounting results, screen reversible perturbation actions based on the expected response differences of each candidate failure cause, and generate perturbation verification instructions. Execute the perturbation verification command, collect perturbation response data and re-input it into the improved DAGMM model, determine the cause of the target fault, and generate the inspection task; Perform inspection tasks and collect inspection feedback. Update the dual-timescaled response evidence envelope and equipment room inspection data based on the inspection feedback, and generate communication equipment room inspection results.

[0008] Optionally, the data center inspection data includes equipment operation data, data center operating condition data, configuration control data, data acquisition link data, operation and maintenance inspection data, operating condition response contract data, and safety operation data.

[0009] Optionally, the generation of standard data center inspection data and dual-time-scaled response evidence envelope includes: Extract device identifier, acquisition record identifier, device generation time, gateway reception time, platform entry time, clock source identifier, and clock calibration record from the data acquisition link data, and associate them with device operation data and data center operating status data; Based on the clock calibration record, the device generation time is corrected, the device physical response time is generated, and the data arrival time is generated based on the gateway reception time and the platform entry time. Two time ranges are determined based on equipment identification and clock source identification. The data integrity status is extracted from equipment operation data and computer room operating data, written into the acquisition record, and generated dual-timescale response evidence envelope and standard computer room inspection data.

[0010] Optionally, the generation of the data center operational condition response contract diagram includes: Extract equipment status, control commands, communication service load and environmental parameters from equipment operation data and computer room operating data; extract operating condition triggering conditions from operating condition response contract data; and identify operating condition disturbance events. Extract the contract version identifier, parent version identifier, validity period and applicable conditions from the operating condition response contract data, connect the contract versions to generate the operating condition response contract version grid, and determine the operating condition response contract version based on the current conditions extracted from configuration control data, equipment operation data, computer room operating condition data and operation and maintenance inspection data. Extract the source device, target device, response direction, response delay, response magnitude, response duration, response recovery process, parent response edge, and propagation order from the operating condition response contract version. Establish directed response edges with devices as nodes, and write the operating condition disturbance event and the operating condition response contract version into the directed response edges to generate the data center operating condition response contract graph.

[0011] Optionally, the arrangement generates a response weaving tensor, including: Based on the equipment identification and the equipment physical response time, the collected records in the dual-time-scaled response evidence envelope are associated with the operating condition disturbance events and response edges in the data center operating condition response contract diagram; Read the event occurrence time, response delay, response duration, response recovery process, response direction, parent response edge, and propagation sequence from the data center operating condition response contract diagram. Divide the response process into an early response zone, a response formation zone, a response stability zone, a response recovery zone, and a delayed response zone. Allocate equipment response data according to the overlap between the equipment's physical response time range and each response zone. Combine the response direction and data integrity status to form response phase stripes. Each bifurcation response phase strip is retained, and the parent response edge strip and target response edge strip that merge into the same target device are arranged separately. The response weaving tensor is generated by arranging the response edge, parent response edge, propagation order, and response region.

[0012] Optionally, the arrangement generates a response weaving tensor, including: The compression network and estimation network of the DAGMM model are replaced with the response compression reconstruction module and the responsibility Gaussian estimation module. A parent edge blocking combination inversion module is set between the two, and a mismatch evidence output module is set at the output end to construct an improved DAGMM model. The response compression and reconstruction module encodes the response phase strip, generates a low-dimensional response representation based on the combination encoding results of the parent response edge and propagation order, reconstructs the target device response, and returns the source device category, parent response edge combination, and propagation order. The parent edge occlusion combination inversion module extracts historical normal response records from the operation and maintenance inspection data and determines the main parent response edges. It then generates the combination inputs corresponding to all occlusion, single parent response edge retention, multi-parent response edge combination retention, and all retention and performs reconstruction. The parent edge independent inheritance response is generated by the difference between the single parent response edge retention result and the total occlusion result; the parent edge coupled response is generated by the remaining result after deducting the corresponding independent inheritance response from the multi-parent response edge combination retention result; and the common influence response is generated by the difference between the common parent response edge retention result and the total occlusion result. The parent edge independent inheritance response, parent edge coupled response, and common impact response are deducted from the actual response of the target device. The data evidence deficit is separated according to the data integrity status. The remaining part is determined as the target local response deficit. The back-pointing result is verified with the data center operating condition response contract diagram to generate the response source accounting result. The duty Gaussian estimation module generates basic duty keys based on the type of operating condition disturbance, the type of response relationship, the type of source and target equipment, the propagation sequence, and the version of the operating condition response contract. It organizes the duty Gaussian components according to the load interval and performs component splitting and merging of adjacent components to generate the conditional negative logarithmic density value of the operating condition sequence. The missing-redemption evidence output module determines the normal range of local response deficit and the normal range of condition density based on historical normal response records. Response edges whose results are both outside the corresponding normal range are identified as candidate missing-redemption response edges. Independent inherited responses and coupled responses passed in by abnormal ancestor response edges are deducted from the target local response deficit. Candidate missing-redemption response edges whose deduction results are within the normal range are deleted. The remaining candidate missing-redemption response edges that do not have ancestor-descendant relationships are combined into the primary missing-redemption antichain.

[0013] Optionally, the generation of perturbation verification instructions includes: Extract the original default response edge from the original default anti-chain, and generate candidate fault causes corresponding to source device action failure, target device execution failure, and data collection evidence failure based on the source device response, target local response deficit, and data evidence deficit. Replace the source device action field, target device fulfillment field and data acquisition field corresponding to the candidate fault causes respectively, and input the replaced response weaving tensor and the data center operating condition response contract diagram into the improved DAGMM model to generate the expected response, expected revenue sharing result and expected condition density change corresponding to the candidate fault causes. Extract executable actions, action boundaries, allowable load ranges, prohibited execution conditions, rollback instructions, and termination conditions from the safety operation data. Write each executable action into the expected response and update the response direction, response latency, and response magnitude according to the data center operating condition response contract diagram. By comparing the expected responses, expected revenue sharing results, and expected condition density changes under the same executable action, executable actions that do not meet the safe operation conditions are eliminated. Reversible perturbation actions are determined according to the degree of separability of candidate causes, and perturbation verification instructions are generated.

[0014] Optionally, the generation of the inspection task includes: Execute the perturbation verification command, extract perturbation response records from equipment operation data, data center operating data and data acquisition link data, correct the perturbation response time based on the dual time-scaled response evidence envelope, read the response relationship from the data center operating response contract diagram, and generate a perturbation response weaving tensor. Input the perturbation response weaving tensor and the data center operating condition response contract diagram into the improved DAGMM model to generate perturbation response source accounting results and perturbation operating condition order condition negative log density values. Compare the perturbation response records, perturbation response source accounting results, and negative logarithmic density values ​​of perturbation operating condition sequence conditions with the corresponding results of candidate fault causes. Candidate fault causes with consistent response direction, response delay, response amplitude, response source accounting category, and condition density change relationship are identified as target fault causes. If no consistent results are formed, an unknown fault cause label is generated. When the termination conditions are met, the action object is restored according to the rollback instruction, the restoration status is verified based on the safety operation data, the target device location is extracted from the configuration control data, the inspection items are extracted from the operation and maintenance inspection data, the inspection operation conditions are extracted from the safety operation data, and the inspection task is generated.

[0015] Optionally, generating the communication equipment room inspection results includes: Perform inspection tasks, collect equipment verification results, data acquisition link verification results, configuration control verification results, maintenance and handling results, and inspection execution status, write them into operation and maintenance inspection data, and generate inspection feedback; Based on the inspection feedback, generate equipment fault markers, data acquisition anomaly markers, configuration control change markers, and normal response markers, update the dual time-stamped response evidence envelope, and write the normal response record and equipment fault record into the normal inspection record and abnormal inspection record, respectively. Candidate contract patches are generated based on configuration control change flags. The candidate contract patches are verified using normal inspection records and abnormal inspection records. When the verification is successful, a new operating condition response contract version is generated and written into the operating condition response contract version grid. The target fault cause, inspection execution status, maintenance and handling results and update records are collected to generate the communication equipment room inspection results.

[0016] A communication equipment room inspection system based on big data according to an embodiment of the present invention includes the following modules: The inspection data processing module is used to process the computer room inspection data and generate standard computer room inspection data and dual time-stamped response evidence envelopes. The working condition contract construction module is used to identify working condition disturbance events from standard data center inspection data and generate working condition response contract version format and data center working condition response contract diagram. The response phase weaving module is used to generate a response weaving tensor based on the dual-time-scaled response evidence envelope and the data center operating condition response contract diagram; The response default analysis module is used to input the response weaving tensor and the data center operating condition response contract diagram into the improved DAGMM model to generate the response source accounting results and the original default anti-chain. The perturbation instruction generation module is used to generate candidate fault causes, expected responses, and perturbation verification instructions based on the response source accounting results and the original default redemption backlink. The fault verification and inspection module is used to execute perturbation verification instructions, determine the target fault cause based on perturbation response data, candidate fault causes and expected response, and generate inspection tasks. The inspection feedback update module is used to execute inspection tasks and collect inspection feedback, update the dual-timescale response evidence envelope, data center inspection data and operating condition response contract version format, and generate communication data center inspection results.

[0017] The beneficial effects of this invention are: This invention generates standard data center inspection data and dual-timescaled response evidence envelopes by collecting equipment operation data, data center status data, configuration control data, data acquisition link data, operation and maintenance inspection data, status response contract data, and safety operation data. It separates the physical response time of the equipment from the data arrival time, reduces the interference of equipment clock offset, gateway transmission delay, and platform data entry delay on response timing judgment, improves the accuracy of cross-device response time positioning, and reduces the probability of data transmission anomalies being misjudged as equipment response anomalies.

[0018] This invention constructs a condition response contract version lattice and a data center condition response contract diagram, which associates equipment configuration, control strategy, load range, maintenance status with cross-device response direction, response delay, response amplitude, parent response edge, and propagation order. Based on the physical response time of the equipment, it generates response phase strips and response weaving tensors, so that asynchronous responses, bifurcation responses, and convergence responses caused by power supply switching, cooling regulation, and communication load changes can be expressed in a unified structure, thereby improving the completeness of the data center operating condition expression and the traceability of response relationships.

[0019] This invention improves the DAGMM model by performing response compression and reconstruction, parent-side blocking combination inversion, duty Gaussian estimation, and loss-of-conversion evidence output. It decomposes the target device response deviation into parent-side independent inherited response, parent-side coupled response, common influence response, target local response deficit, and data evidence deficit. It generates the primary loss-of-conversion anti-chain and filters reversible perturbation actions based on the expected response differences of candidate fault causes. This transforms the inspection task from being driven by a single anomaly score to being driven by fault source location and proactive evidence verification, thereby improving the accuracy of fault cause identification, the relevance of inspection tasks, and the closed-loop capability of inspection feedback. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They explain the invention together with the embodiments of the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a communication equipment room inspection method based on big data proposed in this invention; Figure 2 This is a schematic diagram of the improved DAGMM model structure of the communication equipment room inspection method based on big data proposed in this invention; Figure 3 This is a functional diagram of a communication equipment room inspection system based on big data proposed in this invention; Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1 and Figure 2 A big data-based method for inspecting communication equipment rooms includes: Collect data from the computer room inspection, perform unified identification, time correction, object association, and complete status marking, and generate standard computer room inspection data and dual time-scaled response evidence envelopes; Identify operational disturbance events from standard data center inspection data, match operational response contract versions, and generate a data center operational response contract diagram; Based on the dual-timescaled response evidence envelope correction device response time, response phase strips are constructed and arranged to generate a response weaving tensor; The response weaving tensor and the data center operating condition response contract graph are input into the improved DAGMM model. The improved DAGMM model includes a response compression and reconstruction module, a parent edge blocking combination inversion module, a duty Gaussian estimation module, and a default evidence output module to generate response source accounting results and original default anti-chain. Extract the original default response edge from the original default antichain, generate candidate failure causes by combining the response source accounting results, screen reversible perturbation actions based on the expected response differences of each candidate failure cause, and generate perturbation verification instructions. Execute the perturbation verification command, collect perturbation response data and re-input it into the improved DAGMM model, determine the cause of the target fault, and generate the inspection task; Perform inspection tasks and collect inspection feedback. Update the dual-timescaled response evidence envelope and equipment room inspection data based on the inspection feedback, and generate communication equipment room inspection results.

[0023] In this embodiment, the data room inspection data includes equipment operation data, data room operating status data, configuration control data, data acquisition link data, operation and maintenance inspection data, operating status response contract data, and safety operation data.

[0024] In this embodiment, the generation of standard data center inspection data and dual-timescaled response evidence envelope includes: Extract device identifier, acquisition record identifier, device generation time, gateway reception time, platform entry time, clock source identifier, and clock calibration record from the data acquisition link data, and associate them with device operation data and data center operating status data; Based on the clock calibration record, the device generation time is corrected, the device physical response time is generated, and the data arrival time is generated based on the gateway reception time and the platform entry time. Two time ranges are determined based on equipment identification and clock source identification. Data integrity status is extracted from equipment operation data and data center status data, written into the acquisition record, and dual-timescaled response evidence envelopes and standard data center inspection data are generated. Among these: Two time ranges are determined based on device identifier and clock source identifier, specifically: The collected records are grouped according to the device identifier and clock source identifier. Valid clock calibration records within 30 minutes before and after each collected record are read, centered on the device physical response time. The absolute value of the difference between the clock offset of each clock calibration record and the linear interpolation result of adjacent clock calibration records is calculated. The absolute values ​​of the differences are sorted in ascending order, and the 95th percentile value is read as the clock uncertainty boundary. The lower limit of the device physical response time range is obtained by subtracting the clock uncertainty boundary from the device physical response time. The upper limit of the device physical response time range is obtained by adding the clock uncertainty boundary to the device physical response time. The device-to-gateway transmission delay and the total device-to-platform arrival delay within the same time range are read. The device-to-gateway transmission delay is sorted in ascending order, and the 5th percentile value is read. The total device-to-platform arrival delay is sorted in ascending order, and the 95th percentile value is read. The lower limit of the data arrival time range is obtained by adding the device physical response time to the 5th percentile value of the device-to-gateway transmission delay. The upper limit of the data arrival time range is obtained by adding the device physical response time to the 95th percentile value of the total device-to-platform arrival delay. Thus, the device physical response time range and the data arrival time range are generated. Extract the data integrity status from equipment operation data and data center operating status data, and write it into the acquisition record, specifically: Read data sequence number, sampling time, sampling period, valid status, interpolation status, and hold status from equipment operation data and computer room operating data. Arrange data records according to equipment identification and parameter fields. When the valid status is valid, the interpolation status is 0, and the hold status is 0, write the measured status. When the interpolation status is 1, write the interpolation status. When the hold status is 1, write the hold status. When the valid status is invalid, write the invalid status. Calculate the difference between adjacent data sequence numbers. When the difference between data sequence numbers is greater than 1, create a missing record according to the missing data sequence number. Calculate the difference between adjacent sampling times. When the difference between sampling times is greater than 1.5 times the sampling period, create a missing record according to the sampling period. Write the generated complete data status into the corresponding acquisition record according to the acquisition record identifier. Generate a dual-time-stamped response evidence envelope and standard data center inspection data, specifically as follows: The system reads the device identifier, clock source identifier, device physical response time, device physical response time range, gateway reception time, platform data entry time, data arrival time range, device-to-gateway transmission latency, gateway-to-platform processing latency, device-to-platform total arrival latency, data integrity status, device operation record identifier, and data center operating condition record identifier according to the data collection record identifier. It then writes these read fields into the same envelope record, sorts the envelope records in ascending order by device identifier and device physical response time to generate a dual-timescaled response evidence envelope, and writes the device operation record and data center operating condition record into the corresponding data collection record according to the associated index. It also writes the device physical response time, data integrity status, and dual-timescaled response evidence envelope index into the data collection record, and sorts all data collection records in ascending order by device identifier, device physical response time, and data collection record identifier to generate standard data center inspection data.

[0025] In this embodiment, generating the data center operational condition response contract diagram includes: Extract equipment status, control commands, communication service load, and environmental parameters from equipment operation data and data center operating condition data; extract operating condition triggering conditions from operating condition response contract data; and identify operating condition disturbance events. Specifically, identifying operating condition disturbance events includes: Extract equipment identifier, equipment category, equipment status code, instruction execution status, instruction execution time and communication service load from equipment operation data; extract data center area identifier, sensor identifier, sampling time, data center temperature, relative humidity, smoke concentration, water immersion status and cabinet pressure difference from data center operating condition data; extract trigger condition identifier, applicable equipment category, trigger field, comparison method, trigger threshold, hold duration, debouncing duration and operating condition disturbance type from operating condition response contract data. The equipment operation data is arranged according to the equipment identification and the equipment physical response time. The computer room condition data is arranged according to the computer room area identification and the equipment physical response time. The trigger fields in the equipment operation data and the computer room condition data are compared with the corresponding condition trigger conditions. Records that meet the comparison method and trigger threshold and whose duration reaches the retention time are written into the pending confirmation event set. Read the time interval between adjacent trigger records in the set of events to be confirmed. If the time interval between adjacent trigger records does not exceed the debouncing duration, they are merged into one operating condition disturbance event. The time of the first trigger record is determined as the trigger time. The device status before triggering, the device status after triggering, the trigger field, the trigger value, the operating condition disturbance type, and the computer room area identifier are written into the operating condition disturbance event. The operating condition disturbance event identifier is generated according to the operating condition disturbance type, the triggering device identifier, and the trigger time. Extract the contract version identifier, parent version identifier, validity period, and applicable conditions from the operational condition response contract data. Concatenate the contract versions to generate an operational condition response contract version grid. Determine the operational condition response contract version based on the current conditions extracted from configuration control data, equipment operation data, data center operational condition data, and maintenance inspection data. Specifically, generating the operational condition response contract version grid by concatenating the contract versions involves: Read the contract version identifier, parent version identifier, applicable equipment category, applicable equipment identifier, configuration version identifier, control strategy identifier, load range, maintenance status, effective start time and effective end time from each working condition response contract data. Arrange the read fields in the order of applicable equipment range, configuration version, control strategy, load range, maintenance status and effective time to generate a contract application condition vector. Set the contract version with empty applicable equipment identifier, empty configuration version identifier, empty control strategy identifier, load range covering the complete operating range of the equipment, and empty maintenance status as the top node of the version grid, and connect the parent version identifier of each contract version with the corresponding contract version identifier to form the initial connection from the parent version to the lower version. For each initial connection, perform an applicable condition verification to determine whether the applicable device range of the lower-level version is included in the applicable device range of the upper-level version, whether the load range and effective time of the lower-level version are included in the corresponding range of the upper-level version, and whether the configuration version, control policy and maintenance status of the lower-level version are consistent with the upper-level version or refined from null values ​​to specific values. If all conditions are met, the corresponding connection is retained; if any conditions are not met, the corresponding connection is deleted and written into the version connection exception record. Perform topology sorting on the reserved connections, placing contract versions without parent connections at the top level and parent versions before the child versions. Read the parent version that is commonly connected to the two contract versions, and determine the common parent version with the largest connection depth as the minimum common parent version. Read the intersection of the applicable device range, load range, and effective time of the two contract versions, and combine the intersection fields with the consistent configuration version, control policy, and maintenance status to form the common child conditions. When there is a contract version in the working condition response contract data that has the same applicable conditions and common subordinate conditions, the corresponding contract version is determined as the common subordinate version. When there is no consistent contract version, a condition intersection node is generated and the common subordinate conditions are recorded. When there is no intersection of applicable conditions, no common subordinate connection is established. A working condition response contract version grid containing a top node, contract version node, condition intersection node and version partial order connection is generated. Extract the source device, target device, response direction, response latency, response amplitude, response duration, response recovery process, parent response edge, and propagation order from the operating condition response contract version. Establish directed response edges with devices as nodes, and write the operating condition disturbance event and the operating condition response contract version into the directed response edges to generate the data center operating condition response contract graph. Specifically, generating the data center operating condition response contract graph involves: Read the source device identifier, target device identifier, response relationship type, response direction, response delay lower limit, response delay upper limit, response amplitude lower limit, response amplitude upper limit, response duration, response recovery time, parent response edge identifier, and propagation order from the working condition response contract version one by one. Write the source device identifier and target device identifier into the device node table, sort the device nodes in ascending order of device identifier, and generate a node index. The response edge identifier is generated by concatenating the fields in the order of the operating condition disturbance event identifier, source device identifier, target device identifier, response relationship type and propagation order. A directed response edge is established from the source device node to the target device node. The response direction, response delay range, response amplitude range, response duration, response recovery time, parent response edge identifier, propagation order, operating condition disturbance event identifier and operating condition response contract version identifier are written into the corresponding directed response edge. Read the parent response edge identifier of each directed response edge. If the parent response edge identifier is not empty, retrieve the corresponding parent response edge. Verify whether the target device identifier of the parent response edge is consistent with the source device identifier of the current directed response edge. Verify whether the propagation order of the parent response edge is less than the propagation order of the current directed response edge. If both verifications pass, retain the current directed response edge. If there is an inconsistency, write the current directed response edge into the contract relationship exception record. An adjacency matrix is ​​built according to the device node index. When there is a valid directed response edge between the source device node and the target device node, the corresponding matrix position is written as 1, and when there is no valid directed response edge, it is written as 0. A parent edge index table is built according to the parent response edge identifier. The valid directed response edges are arranged in ascending order of propagation position. A data center operating condition response contract graph composed of the device node table, the directed response edge table, the adjacency matrix, and the parent edge index table is generated.

[0026] In this embodiment, the arrangement generates a response weaving tensor, including: Based on the equipment identification and the equipment physical response time, the collected records in the dual-time-scaled response evidence envelope are associated with the operating condition disturbance events and response edges in the data center operating condition response contract diagram; The event occurrence time, response delay, response duration, response recovery process, response direction, parent response edge, and propagation order are read from the data center operational condition response contract diagram. The response process is divided into an early response zone, a response formation zone, a response stabilization zone, a response recovery zone, and a delayed response zone. Equipment response data is allocated based on the overlap between the equipment's physical response time range and each response zone. Response phase stripes are formed by combining the response direction and data integrity status, where: Equipment response data is allocated based on the overlap between the equipment's physical response time range and each response zone, specifically as follows: Read the lower limit and upper limit of the physical response time range of the device for each collection record. For each response zone, compare the lower limit of the physical response time range of the device with the start of the response zone and take the later time as the overlap start. Compare the upper limit of the physical response time range of the device with the end of the response zone and take the earlier time as the overlap end. When the overlap end is later than the overlap start, calculate the time difference between the two and determine it as the overlap duration of the response zone. Calculate the sum of the overlap durations between the current acquisition record and all response zones. Divide the overlap duration of each response zone by the sum of the overlap durations to obtain the phase allocation weight of the corresponding response zone. The phase allocation weight is set according to the coverage ratio of the device's physical response time range in each response zone. It is used to retain the cross-zone time uncertainty caused by clock calibration error. The sum of all phase allocation weights is set to 1. When the lower limit of the physical response time range of the device is the same as the upper limit of the physical response time range of the device, the phase allocation weight corresponding to the response area into which the physical response time of the device falls is set to 1, and the phase allocation weight corresponding to the other response areas is set to 0. The device response data is copied to the response area with an overlap duration greater than 0 and written to the phase allocation weight. The response phase strip is formed by combining the response direction and the data integrity status, specifically as follows: For each response parameter, read the device response data in the early response zone where the complete state is the measured state, and determine the median value of the read data as the response reference value. When there is no measured data in the early response zone, read the measured data with the smallest time distance before the event occurrence time and determine it as the response reference value. Calculate the difference between the device response value in each response zone and the response reference value to obtain the device response change. Read the response direction and lower limit of response amplitude in the data center operating condition response contract diagram. When the response direction is upward and the change in the device response is not less than the lower limit of response amplitude, write a direction consistency mark. When the response direction is downward and the absolute value of the change in the device response is not less than the lower limit of response amplitude and the change in the device response is less than 0, write a direction consistency mark. When the absolute value of the change in the device response is less than the lower limit of response amplitude, write a response no mark. When the direction of device response change is opposite to the direction of contract response, write an opposite direction mark. The data integrity status is converted into a 5-dimensional integrity status code in the order of measured status, interpolated status, held status, missing status and invalid status. The corresponding status position is written as 1 and the non-corresponding status position is written as 0. The device response change of multiple acquisition records in the same response area is calculated as a weighted average according to the phase allocation weight. The 5-dimensional integrity status codes are summed according to the phase allocation weight to generate the data integrity status ratio of the corresponding response area. The source device response change, target device response change, direction consistency marker, response not formed marker, direction opposite marker, phase allocation weight, offset of device physical response time relative to event occurrence time, offset of data arrival time relative to device physical response time, and 5-dimensional complete state ratio are arranged in a fixed field order to generate phase units corresponding to each response zone. The five phase units are arranged in the order of early response zone, response formation zone, response stability zone, response recovery zone, and delayed response zone to form response phase strips. Each bifurcation response phase strip is retained, and the parent response edge strip and target response edge strip that converge to the same target device are arranged separately. These are then arranged according to response edge, parent response edge, propagation order, and response region to generate a response weaving tensor, where: The bifurcation response phase strips are retained separately, specifically as follows: Read response edges with the same source device identifier but different target device identifiers under the same operating condition disturbance event, set each response edge as a branch response group, copy the response phase strip corresponding to the source device to each branch response group, write the response phase strip corresponding to each target device to the corresponding branch response group, and do not perform summation or averaging on the response phase strips of different target devices. Generate bifurcation strip identifiers according to the operating condition disturbance event identifier, source device identifier, target device identifier and response edge identifier. Write the bifurcation response phase strips generated by the same source device into the same bifurcation group identifier, and retain the response direction, phase allocation weight and data integrity status ratio corresponding to different target devices. Arrange the parent response sidebands and target response sidebands that converge to the same target device separately, as follows: For each target response edge, read the parent response edge identifier, retrieve all parent response edges from the data center operating condition response contract diagram based on the parent response edge identifier, arrange the parent response edges in ascending order according to the propagation order, and arrange them in ascending order according to the parent response edge identifier when the propagation order is the same. Write the response phase strip corresponding to each parent response edge into an independent parent edge slot. Write the response phase strip corresponding to the target response edge into the target edge slot after all the parent edge slots. Do not merge multiple parent response edge strips. When the current target response edge does not have a parent response edge, directly write the target response edge strip into the target edge slot to generate an inflow response group composed of the parent response edge strip and the target response edge strip. The response weaving tensor is generated by arranging the response edges, parent response edges, propagation order, and response regions, as follows: The operating condition disturbance events are arranged in ascending order of their occurrence time. If the occurrence times are the same, they are arranged in ascending order of their operating condition disturbance event identifiers. For each operating condition disturbance event, the incoming response groups are arranged according to their propagation order and response edge identifier. For each incoming response group, the parent edge slots are arranged according to their parent response edge propagation order and parent response edge identifier. The target edge slots are arranged after all the parent edge slots. Calculate the maximum number of response edges and the maximum number of parent response edges in a single target response edge within the current batch of processing conditions. Construct a 5-dimensional tensor consisting of event dimension, response edge dimension, strip slot dimension, response region dimension, and state field dimension. Set the length of the response region dimension to 5, the length of the strip slot dimension to the maximum number of parent response edges plus 1, and the state field dimension to the number of fields in the response phase strip. Each response phase stripe is written into a 5D tensor according to the corresponding event position, response edge position, parent edge slot or target edge slot, response area position, and state field position. All response edge positions with fewer than the maximum number of response edges and parent edge slots with fewer than the maximum number of parent response edges are written to 0, and the stripe existence flag is written to 0. The position where a response phase stripe exists is written to 1, generating a response weaving tensor that preserves the bifurcation relationship, merging relationship, parent edge relationship, propagation order, and response stage.

[0027] In this embodiment, the arrangement generates a response weaving tensor, including: An improved DAGMM model is constructed by replacing the compression and estimation networks of the DAGMM model with a response compression reconstruction module and a responsibility Gaussian estimation module, setting a parent edge occlusion combination inversion module between them, and setting a mismatch evidence output module at the output end. Specifically, the construction of the improved DAGMM model is as follows: The response compression and reconstruction module, the parent edge occlusion combination inversion module, the duty Gaussian estimation module, and the failure evidence output module are connected in sequence. The response weaving tensor is used as the numerical input, and the data center operating condition response contract graph is used as the structural index input. The data center operating condition response contract graph is used to provide the parent response edge, propagation order, operating condition response contract version, and ancestor path. The fully connected network is not directly input. The response weaving tensor slice corresponding to a single target response edge is set to batch size multiplied by 6, 5, and 14. The 6 strip slots correspond to 4 main parent response edges, 1 common parent response edge, and 1 target response edge, respectively. The 5 positions correspond to the early response area, response formation area, response stabilization area, response recovery area, and delayed response area. The 14 state fields include source device response change amount, target device response change amount, 3 response direction markers, phase allocation weight, device physical response time offset, data arrival time offset, 5 data integrity state ratios, and parent edge occlusion marker. Additionally, the stripes with batch size multiplied by 6 and 5 have a mask, which is not included in the state field dimension. The response compression and reconstruction module expands the 5x14 fields corresponding to each slot into a 70-dimensional strip vector. This vector is then processed through a 70-dimensional input layer, a 56-dimensional hidden layer, a 28-dimensional hidden layer, and a 14-dimensional strip encoding layer to generate a 14-dimensional strip encoding. The six strip encodings are concatenated into an 84-dimensional combined vector. This vector is then processed through a 64-dimensional combined compression layer and a 32-dimensional low-dimensional representation layer to generate a 32-dimensional low-dimensional response representation. Finally, the target device response changes in the five response regions are reconstructed through a 64-dimensional decoding hidden layer and a 5-dimensional decoding output layer. A 9-dimensional source device category back pointer, a 4-dimensional parent response edge combination back pointer, and a 7-dimensional propagation order back pointer are set. The duty Gaussian estimation module is configured with a 37-dimensional input layer, a 32-dimensional hidden layer, and a 4-dimensional duty probability output layer. The model contains a total of 12 trainable fully connected layers. The improved DAGMM model contains 18,111 trainable neural network parameters. Each 37-dimensional Gaussian component in the responsibility Gaussian estimation module contains 37 mean parameters, 703 independent covariance parameters, and 1 mixed weight parameter. When the total number of responsibility Gaussian components is denoted as K, the model parameter size is 18,111 plus 741 multiplied by K. The model output includes a 5-dimensional target device response reconstruction result, a 32-dimensional low-dimensional response representation, a 5x5 response source accounting result, a 1-dimensional working condition order condition negative log density value, and an M-row, 4-column primary default anti-chain record table, where M is the number of primary default response edges. Historical records verified as normal responses are extracted from operation and maintenance inspection data. These are then correlated with corresponding equipment operation data, data center status data, data acquisition link data, and status response contract data. A training response weaving tensor is generated according to rules consistent with online processing. The tensor is grouped by status disturbance event identifiers and divided into a 70% training set, a 20% validation set, and a 10% test set based on event occurrence time. Records corresponding to the same status disturbance event are written into the same dataset. The training set includes occlusion samples with different states retained by the parent response edges. The target response reconstruction error is used... The joint loss consists of source device category cross-entropy, parent response edge combination binary cross-entropy, propagation order cross-entropy, conditional negative log density loss, and covariance stability loss. The weights of the six components are set to 0.4, 0.1, 0.15, 0.1, 0.2, and 0.05 respectively. The weights are determined by searching the validation set with a step size of 0.05. The Adam optimizer is used for training with a learning rate of 0.001, a batch size of 64, and a maximum number of training epochs of 300. Training stops when the validation set loss decreases by no more than 0.0001 for 20 consecutive epochs. Compared with existing DAGMM models that compress flat data and output anomaly energy based on a uniform Gaussian mixture distribution, the improved DAGMM model retains the response stage, parent response edge, and propagation order. It adds parent edge blocking combination inversion to distinguish between upstream inherited responses and target device local responses, adds responsibility Gaussian estimation to adapt to different device responsibilities, operating condition versions, and load ranges, and adds failure evidence output to eliminate propagation anomalies that can be explained by anomaly ancestors. This enables the model to distinguish between upstream fault propagation, target device execution failure, and data acquisition evidence failure. The response compression and reconstruction module encodes the response phase strip, generates a low-dimensional response representation based on the combination encoding results of the parent response edge and propagation order, reconstructs the target device response, and returns the source device category, parent response edge combination, and propagation order. The parent edge occlusion combination inversion module extracts historical normal response records from the operation and maintenance inspection data and determines the main parent response edges. It then generates combined inputs for all occlusion, single parent response edge retention, multi-parent response edge combination retention, and all retention, and performs reconstruction. Specifically, generating and reconstructing the combined inputs for all occlusion, single parent response edge retention, multi-parent response edge combination retention, and all retention involves: According to the target response edge identifier, extract the historical response records from the operation and maintenance inspection data that have been confirmed to be normal by on-site inspection and have the same working condition response contract version. Read all the parent response edges corresponding to the target response edge, establish a single parent response edge retention input for each parent response edge, keep the current parent response edge slot with the original value, write the response value, response change amount and direction mark in the other parent response edge slots to 0, keep the stripe existence mark with the original value, input the trained response compression and reconstruction module, and generate the single parent response edge retention result; Write the response value, response change, and direction marker in all parent response edge slots to 0, generate all occlusion inputs and perform reconstruction. For each parent response edge, calculate the sum of the absolute values ​​of the differences between the single parent response edge retention result and the total occlusion result. Divide the sum of the absolute values ​​of the differences by the sum of the absolute values ​​of the differences of all parent response edges to generate the contribution ratio of the parent response edge. The median value of the contribution ratio of each parent response edge is calculated according to the historical normal response records. The parent response edges are arranged in descending order according to the median contribution ratio. The contribution ratio is accumulated starting from the first parent response edge. In this implementation, the parent response edges read before the cumulative contribution ratio reaches 0.85 are determined as the main parent response edges. When the number of main parent response edges exceeds 4, the first 4 parent response edges are retained. The cumulative contribution ratio of 0.85 is used to cover the main incoming contributions in the historical normal response of the target device. The upper limit of 4 is used to limit the number of parent response edge combinations to within 16 groups. Set the fusion weight for the response phase strips not included in the main parent response edge according to the median value of the contribution ratio. The fusion weight of each non-main parent response edge is the median value of the corresponding contribution ratio divided by the sum of the median values ​​of the contribution ratios of all non-main parent response edges. Multiply each state field by the fusion weight and add them together to generate a common parent response edge strip. If there is no non-main parent response edge, write all the common parent response edge strips to 0 and write the strip existence flag to 0. Establish binary retention codes according to the order of the main parent response edges. In the retention code, a value of 1 indicates that the corresponding parent response edge slot is retained, and a value of 0 indicates that the corresponding parent response edge slot is blocked. Traverse all binary retention codes to generate all blocking inputs, retention inputs for each single parent response edge, and retention inputs for each combination of multiple parent response edges. Establish a common parent response edge retention input that only retains the common parent response edge stripe. Establish all retention inputs that retain all main parent response edges and common parent response edges. All combined inputs are sequentially input into the same set of response compression and reconstruction module parameters. The encoder and decoder weights are not updated. The response reconstruction results of each target device are recorded according to the target response edge identifier, the parent response edge retention code, and the common parent response edge retention status. The parent edge independent inheritance response is generated by the difference between the retained result of a single parent response edge and the total occlusion results. The parent edge coupled response is generated by the remaining result after subtracting the corresponding independent inheritance response from the retained result of a multi-parent response edge combination. The common influence response is generated by the difference between the retained result of a common parent response edge and the total occlusion results. The parent edge independently inherits the response by using the difference between the retained result of the single-parent response edge and the total occlusion result, specifically as follows: Read all occlusion results and the single parent response edge retention results corresponding to each main parent response edge. Calculate the difference between the single parent response edge retention results and all occlusion results according to the response area and status field. Write the difference under the corresponding parent response edge identifier to generate the parent edge independent inheritance response corresponding to each main parent response edge. Preserve the index relationship between the parent response edge identifier, propagation order and target response edge identifier. The parent edge coupled response is generated by subtracting the corresponding independently inherited response from the retained result of the multi-parent response edge combination result, specifically: Read the combined retention results of multiple parent response edges in ascending order of the number of parent response edges retained. For combinations containing two parent response edges, deduct all occlusion results and the independent inherited responses of the parent edges corresponding to the two parent response edges from the combined retention results to generate the coupled response between the two parent response edges. For a combination containing 3 parent response edges, the combined retained result is subtracted from all occlusion results, the independent inherited responses of the corresponding parent edges of the 3 parent response edges, and the coupled responses of all 2 parent response edges in the combination, to generate the coupled response between the 3 parent response edges. For a combination containing 4 parent response edges, the combined retained result is subtracted from all occlusion results, all independently inherited responses of the parent edges, all coupled responses of the 2 parent response edges, and all coupled responses of the 3 parent response edges in the same order, to generate the coupled response between the 4 parent response edges. The coupled responses generated by combining each multi-parent response edge are added one by one according to the response area and state field to generate the parent edge coupled response corresponding to the target response edge. The parent response edge combination identifier corresponding to each coupled response is retained for the interpretation of abnormal ancestor responses. The common impact response is generated by the difference between the retained result of the common parent response edge and the total occlusion result, specifically as follows: Read the retained results of the common parent response edge and the complete occlusion results. Calculate the difference between the retained results of the common parent response edge and the complete occlusion results according to the response area and status field. Generate a common impact response. Write the original parent response edge identifier, fusion weight and propagation order range contained in the common parent response edge into the corresponding index of the common impact response. From the actual response of the target device, the parent-side independently inherited response, parent-side coupled response, and common impact response are subtracted. Based on the data integrity status, the data evidence deficit is separated, and the remaining portion is identified as the target local response deficit. The back-pointing results are verified against the data center operating condition response contract diagram to generate the response source accounting results, where: The parent edge's independently inherited response, parent edge's coupled response, and common influence response are subtracted from the target device's actual response, specifically: Read the target device response change from the response phase strip corresponding to the target response edge, and arrange them into the target device actual response vector according to the early response area, response formation area, response stabilization area, response recovery area and delayed response area. Arrange the parent edge independent inherited response, parent edge coupled response and common influence response according to the same response area and state field order. Add all the independent inherited responses of the parent edges one by one, and then subtract the sum of the independent inherited responses of the parent edges, the coupled responses of the parent edges, and the common influence responses from the actual response vector of the target device in sequence to generate the remaining vector of the target response. Based on the data integrity status, the data evidence deficit is separated, and the remaining part is identified as the target local response deficit, specifically: Extract the target device reconstruction error corresponding to the measured state, interpolated state, held state, missing state and invalid state from the historical normal response record, calculate the median value of the absolute value of the reconstruction error for each state, subtract the median value of the current state from the maximum median value and divide by the difference between the maximum median value and the minimum median value to generate the evidence reliability coefficient corresponding to the current state. The reliability coefficient of the measured state evidence is fixed at 1, and the reliability coefficient of the missing state and invalid state evidence is fixed at 0. When the maximum median value and the minimum median value are the same, the reliability coefficient of the imputed state evidence is set to 0.7, and the reliability coefficient of the held state evidence is set to 0.4. 0.7 is used to retain the data evidence inferred from the adjacent measured records, and 0.4 is used to retain the static numerical evidence corresponding to the held value and reduce the proportion of trend evidence. Read the proportion of complete 5-dimensional data states in each response zone, multiply each state proportion by the corresponding evidence reliability coefficient, and add them together to generate the evidence reliability of the response zone. Multiply each item of the target response remaining vector by 1 and subtract the evidence reliability of the response zone to generate the data evidence deficit. Multiply each item of the target response remaining vector by the evidence reliability of the response zone to generate the target local response deficit. Add the data evidence deficit and the target local response deficit together to equal the target response remaining vector. The return pointer results are verified against the data center operational condition response contract diagram to generate response source revenue sharing results, specifically: Read the source device category, parent response edge identifier set, and propagation order of the current target response edge from the data center operating condition response contract diagram. Compare the source device category back pointer result with the source device category, compare the parent response edge combination back pointer result with the parent response edge identifier set, and compare the propagation order back pointer result with the propagation order. When all three back-pointing results are consistent, write a back-pointing consistency flag; when there are inconsistent items, write a back-pointing inconsistency flag. Record the inconsistent source device category, parent response edge slot and propagation order. Write the parent edge independently inherited response, parent edge coupled response, common impact response, data evidence shortage, target local response shortage, index corresponding to each response and back-pointing verification result into the response source accounting result in a fixed field order. The duty Gaussian estimation module generates basic duty keys based on the type of operating condition disturbance, the type of response relationship, the type of source and target equipment, the propagation order, and the version of the operating condition response contract. It then organizes the duty Gaussian components according to the load interval and performs component splitting and merging of adjacent components to generate the operating condition order condition negative logarithmic density value. Specifically, the generation of the operating condition order condition negative logarithmic density value is as follows: The negative logarithmic density value of the operating condition sequence condition is generated as follows: read the operating condition disturbance type, response relationship type, source device category, target device category, propagation sequence, and operating condition response contract version identifier from the computer room operating condition response contract diagram, concatenate the fields according to the reading order to generate the basic responsibility key, and write the historical normal response records with consistent basic responsibility keys into the same responsibility sample set. The 32-dimensional low-dimensional response representation is concatenated with the response source accounting results. The mean absolute value after response amplitude range normalization is extracted from the parent edge independent inheritance response, parent edge coupled response, common impact response, data evidence deficiency and target local response deficiency, respectively, to generate a 5-dimensional accounting summary. The 32-dimensional low-dimensional response representation is concatenated with the 5-dimensional accounting summary to form a 37-dimensional responsibility feature vector. Read the load range from the corresponding working condition response contract version, write the 37-dimensional responsibility feature vector into the corresponding load range according to the communication service load when the historical normal response record occurs, calculate the mean, covariance matrix and sample number of the responsibility feature vector for each basic responsibility key and load range, divide the sample number by the total number of samples under the same basic responsibility key, and generate the responsibility Gaussian component mixture weight. Write 0.001 times the value of the diagonal elements of the covariance matrix into all diagonal elements to avoid singular results when inverting the covariance matrix. 0.001 is used to maintain the original covariance scale and provide a numerically stable term. For each responsibility Gaussian component, calculate the eigenvector corresponding to the largest eigenvalue of the covariance matrix, project all responsibility eigenvectors onto the corresponding eigenvectors and sort them in ascending order of projection values, read the segmentation position with the largest difference between adjacent projection values, and divide the samples on both sides of the segmentation position into two subsets; When the center distance between two subsets reaches twice the mean of the internal standard deviation of the subsets, and the sample size of each subset reaches 10 times the dimension of the responsibility feature plus 1, component splitting is performed. The mean, covariance matrix and mixing weight of the two new components are calculated respectively. The center distance of twice is used to reduce the erroneous splitting of overlapping sub-distributions, and the sample size condition is used to ensure the stability of the mean and covariance estimates. When the number of Gaussian component samples for the duty function is less than 10 times the duty function feature dimension plus 1, read the adjacent Gaussian components for the duty function under the same basic duty function key, calculate the Mahalanobis distance between the mean of the current component and the mean of each adjacent component, determine the adjacent component with the smallest Mahalanobis distance as the merging object, merge the samples of the two components and recalculate the mean, covariance matrix and mixing weight. When there are no adjacent load intervals, select the Gaussian component for the duty function with the smallest center distance of the load interval. Read the basic responsibility key and communication service load corresponding to the response edge to be detected, retrieve the responsibility Gaussian component with the same basic responsibility key and the load range covering the current communication service load, calculate the Gaussian probability density of the 37-dimensional responsibility feature vector in each responsibility Gaussian component, multiply the Gaussian probability density by the corresponding mixing weight and add them, take the natural logarithm of the addition result and change the sign to generate the negative logarithmic density value of the working condition sequence condition. The default evidence output module determines the normal range of local response deficit and the normal range of conditional density based on historical normal response records. Response edges whose results are both outside their corresponding normal ranges are identified as candidate default response edges. Independent inherited responses and coupled responses passed from abnormal ancestor response edges are deducted from the target local response deficit. Candidate default response edges whose deduction results are within the normal range are deleted. The remaining candidate default response edges that have no ancestor-descendant relationship are combined into a primary default antichain, where: The failure evidence output module determines the normal range of local response under-quantity and the normal range of condition density based on historical normal response records, specifically: Divide the absolute value of each response region of the target local response deficit by the corresponding response amplitude range width, and then calculate the weighted average value according to the phase allocation weight to generate the local response deficit value. When the response amplitude range width is 0, the median of the absolute value of the historical normal response amplitude of the same basic responsibility key is used as the divisor. Historical normal response records are filtered according to basic responsibility key, load range, and operating condition response contract version. Local response under-quantity values ​​and operating condition sequence condition negative logarithmic density values ​​are arranged separately. The median and median absolute deviation of each arranged object are calculated. The median value minus the product of 3 times the median absolute deviation and 1.4826 is determined as the lower limit of the normal range. The median value plus the product of 3 times the median absolute deviation and 1.4826 is determined as the upper limit of the normal range. When the median absolute deviation is 0, the standard deviation of the corresponding historical normal record is used as a substitute. When both the median absolute deviation and the standard deviation are 0, the lower limit of the normal range is set to the median value, and the upper limit of the normal range is set to the median value plus 0.000001. Response edges whose two results are both outside the corresponding normal range are identified as candidate unreacted response edges, specifically: Read the local response deficit value and the negative logarithmic density value of the operating condition sequence condition of the current response edge. When the local response deficit value is greater than the upper limit of the normal range of local response deficit and the negative logarithmic density value of the operating condition sequence condition is greater than the upper limit of the normal range of condition density, write the current response edge into the candidate failed response edge set. If neither value exceeds the upper limit of the corresponding normal range, do not write it into the candidate failed response edge set. From the target local response deficit, subtract the independent inherited responses and coupled responses passed in by the abnormal ancestor response edges, and delete the candidate mismatch response edges whose subtraction results are within the normal range. Specifically: For each candidate failed response edge, the ancestor response edge is retrieved level by level from the parent edge index table in the data center operating condition response contract graph, and the ancestor response edge that is simultaneously located in the candidate failed response edge set is written into the abnormal ancestor set. Based on the retained parent response edge identifier and parent response edge combination identifier, extract the parent edge independent inherited response and parent edge coupled response of the current candidate failed response edge passed to the propagation path corresponding to the abnormal ancestor set. Subtract the extracted results item by item from the target local response deficit of the current candidate failed response edge to generate the local response deficit after ancestor deduction. Calculate the local response deficit value after ancestor deduction according to the calculation method of the local response deficit value. When the local response deficit value after ancestor deduction is within the normal range of local response deficit, mark the current candidate unredeemed response edge as an ancestor interpretable response edge and delete it from the candidate unredeemed response edge set. If the local response deficit value after ancestor deduction is still greater than the upper limit of the normal range, retain the current candidate unredeemed response edge. The remaining candidate unredeliverable response edges that have no ancestor-descendant relationship with each other are combined into the original unredeliverable antichain, specifically: The remaining candidate missing response edges are arranged in ascending order of propagation order. If the propagation order is the same, they are arranged in descending order of the local response deficit value after ancestor deduction, and then in ascending order of response edge identifier. Read the candidate lost response edges sequentially. If the original lost response backchain is empty, write the current candidate lost response edge into the original lost response backchain. If the original lost response backchain is not empty, check the parent edge index table to see if there is an ancestor-descendant relationship between the current candidate lost response edge and the written response edge. If there is no ancestor-descendant relationship, write it into the original lost response backchain. If there is an ancestor-descendant relationship, retain the response edge with the smaller propagation order and delete the response edge with the larger propagation order. Generate an original lost response backchain composed of original lost response edges that do not have an ancestor-descendant relationship with each other.

[0028] In this embodiment, the generation of perturbation verification instructions includes: Extract the original default response edge from the original default anti-chain, and generate candidate fault causes corresponding to source device action failure, target device execution failure, and data collection evidence failure based on the source device response, target local response deficit, and data evidence deficit. Replace the source device action field, target device fulfillment field, and data acquisition field corresponding to the candidate fault causes, respectively. Input the replaced response weaving tensor and the data center operating condition response contract diagram into the improved DAGMM model to generate the expected response, expected revenue sharing result, and expected condition density change corresponding to the candidate fault causes. Specifically, the generation of the expected response, expected revenue sharing result, and expected condition density change corresponding to the candidate fault causes is as follows: Each candidate tensor copy and the data center operating condition response contract diagram are input into the improved DAGMM model. The target device reconstruction response is read from the response compression and reconstruction module. The target device reconstruction response is arranged according to the early response zone, response formation zone, response stability zone, response recovery zone and delayed response zone to generate the expected response corresponding to the candidate fault cause. Read the parent edge independent inherited response, parent edge coupled response, common impact response, target local response deficit and data evidence deficit from the parent edge blocking combination inversion module, and generate the expected revenue sharing results corresponding to the candidate fault causes according to the fixed field order; Read the negative logarithmic density value of the working condition sequence condition corresponding to the original response weaving tensor, read the negative logarithmic density value of the working condition sequence condition corresponding to the candidate tensor copy, calculate the difference between the value corresponding to the candidate tensor copy and the value corresponding to the original response weaving tensor, generate the expected condition density change corresponding to the candidate fault cause, and write the candidate fault cause identifier, expected response, expected accounting result and expected condition density change into the candidate cause expected record. Extract executable actions, action boundaries, allowable load ranges, prohibited execution conditions, rollback instructions, and termination conditions from the safety operation data. Write each executable action into the expected response and update the response direction, response latency, and response magnitude according to the data center operational condition response contract diagram. Specifically, writing each executable action into the expected response and updating the response direction, response latency, and response magnitude according to the data center operational condition response contract diagram involves: Read the action object, control field, action direction, action amplitude, action duration and corresponding operating condition disturbance type from each executable action. Retrieve the starting response edge consistent with the source device from the data center operating condition response contract diagram according to the action object and the corresponding operating condition disturbance type. Read the starting response edge and downstream response edge in ascending order of propagation sequence. For each executable action and each candidate fault cause, copy the expected record of the candidate cause, write the action direction and action amplitude into the action field of the source device corresponding to the action object, write the action execution time into the start point of the response formation area, write the action duration into the duration of the response stabilization area, and generate an action candidate tensor. When the candidate fault cause is that the source device fails to perform its action, the response of the source device corresponding to the action object is kept at the baseline value of the advance response zone, the change in the source device response is written to 0, and the writing of the action response to the downstream response edge is stopped. When the candidate fault cause is that the target device fails to perform its action, the action response is propagated to the target device according to the data center operating condition response contract diagram, the response of the target device is kept at the baseline value of the advance response zone, and the writing of the action response to the downstream response edge of the target device is stopped. When the candidate fault cause is the invalidation of data acquisition evidence, the physical response fields of the source device and the downstream target device are updated according to the response direction, response delay and response amplitude recorded in the data center operating condition response contract diagram, and the data integrity status code and data arrival time offset are kept as the data acquisition field replacement results. For response edges that are allowed to propagate, the formation time of the previous level response is added to the response delay range of the current response edge to generate the expected response delay of the current response edge. The response direction recorded by the current response edge is written into the corresponding response direction field, and the response amplitude range recorded by the current response edge is written into the corresponding response amplitude field. All downstream response edges are updated according to the propagation order. The action candidate tensor and the data center operating condition response contract diagram are input into the improved DAGMM model again to update the expected response, expected revenue sharing result and expected condition density change under each executable action, and generate the action identifier, candidate fault cause identifier and the expected record of the three types of expected results. Compare the expected responses, expected revenue sharing results, and expected condition density changes under the same executable action; eliminate executable actions that do not meet the safe operation conditions; determine reversible perturbation actions based on the separability of candidate causes; and generate perturbation verification instructions, wherein: The following compares the changes in expected responses, expected revenue sharing results, and expected condition density under the same executable action: Read all expected records of action causes corresponding to the same executable action according to the action identifier, select candidate fault causes one by one, compare the expected response directions corresponding to the two candidate fault causes, generate a direction difference value of 0 when the directions are consistent, and generate a direction difference value of 1 when the directions are inconsistent. Calculate the absolute value of the difference between the midpoints of the two expected response delay ranges and divide it by the width of the response delay range in the corresponding operating condition response contract version to obtain the delay difference value. Calculate the absolute value of the difference between the midpoints of the two expected response amplitude ranges and divide it by the width of the response amplitude range in the corresponding operating condition response contract version to obtain the amplitude difference value. If the delay difference value or amplitude difference value is greater than 1, write the corresponding difference value as 1. Compare the two expected revenue sharing results in the order of parent edge independent inheritance response, parent edge coupled response, common impact response, target local response deficit, and data evidence deficit. Divide the absolute value of the difference of each field by the normal range width of the corresponding field in the historical normal response record, calculate the average value of the normalized field difference, and obtain the revenue sharing difference value. If the revenue sharing difference value is greater than 1, write the revenue sharing difference value as 1. Calculate the absolute value of the difference between two expected condition density changes, and divide it by the normal range width of the negative logarithmic density value of the condition sequence in the historical normal response record to obtain the condition density difference value. If the condition density difference value is greater than 1, write the condition density difference value as 1. Based on the separability of candidate causes, reversible perturbation actions are determined, and perturbation verification instructions are generated, specifically as follows: The weights for the direction difference value, time delay difference value, amplitude difference value, revenue sharing difference value and condition density difference value are all set to 0.2, and the sum of the five weights is 1. The basis for setting the same weight is that all five types of evidence have been normalized and respectively represent the response direction, response time, response amplitude, response source and condition density, so as to avoid a single piece of evidence from dominating the action screening results. Multiply the 5 difference values ​​by their corresponding weights and sum them to generate a pair of candidate fault causes under the current executable action. Calculate the cause separation values ​​for all candidate fault causes pair by pair, and extract the cause separation value with the smallest value as the degree of separation of the candidate cause for the current executable action. When the candidate fault cause contains only 1 item, use the original response with the unexecuted field replacement as the reference for the normal response and calculate the degree of separation of the candidate cause. Read the current device load, action parameters and safety status, delete executable actions that are outside the allowable load range, have action parameters outside the action boundary, meet the prohibition conditions, lack rollback instructions or lack termination conditions, arrange the remaining executable actions in descending order of the degree of separability of candidate causes, arrange them in ascending order of action duration when the degree of separability of candidate causes is the same, arrange them in ascending order of action duration when the action duration is the same, and determine the first executable action as the reversible perturbation action. Write the original lost-response edge identifier, candidate fault cause identifier, reversible perturbation action identifier, action object, control field, action direction, action amplitude, action duration, expected response corresponding to each candidate fault cause, expected accounting result, expected condition density change, rollback instruction and termination condition into the verification record to generate perturbation verification instruction.

[0029] In this embodiment, generating the inspection task includes: The perturbation verification command is executed to extract perturbation response records from equipment operation data, data center operating condition data, and data acquisition link data. The perturbation response time is corrected based on the dual-timescaled response evidence envelope. The response relationship is read from the data center operating condition response contract diagram to generate a perturbation response weaving tensor. Specifically, the process of reading the response relationship from the data center operating condition response contract diagram and generating the perturbation response weaving tensor is as follows: According to the reversible perturbation action corresponding to the operating condition perturbation type and action object, the starting response edge that is consistent with the source device and the action object is retrieved from the computer room operating condition response contract diagram. All downstream response edges are read along the parent response edge index and directed adjacency relationship of the starting response edge. The response direction, response delay, response duration, response recovery process, parent response edge and propagation order of each response edge are extracted. The action execution time is taken as the event occurrence time. The response delay, response duration and response recovery process are divided into the early response area, response formation area, response stability area, response recovery area and delayed response area. Read the physical response time range of the perturbation device from each perturbation response record, calculate the overlap time between the physical response time range of the perturbation device and each response zone, divide the overlap time of each response zone by the sum of the overlap times of all response zones, generate phase allocation weights, and write the perturbation response value, the response change amount relative to the reference value before the action is executed, the response direction mark, the phase allocation weight, the data arrival time offset, and the data integrity status code into the corresponding response zone to generate perturbation response phase strips; An empty perturbation tensor matrix is ​​constructed based on the operating condition perturbation event dimension, response edge dimension, strip slot dimension, response region dimension, and state field dimension used in the original response weaving tensor. Parent edge stripes are arranged according to the propagation order of the parent response edges. The target response edge stripes are arranged after the parent edge stripes. Each perturbation response phase stripe is written to the corresponding tensor position. The position where there is no stripe is written with 0 and the stripe existence mark is written with 0. The position where there is a stripe is written with 1. A perturbation response weaving tensor with the same field order, dimension definition, and slot rules as the original response weaving tensor is generated. The perturbation response weaving tensor and the data center operating condition response contract diagram are input into the improved DAGMM model to generate perturbation response source allocation results and perturbation operating condition order-conditional negative log density values. Specifically, the generation of perturbation response source allocation results and perturbation operating condition order-conditional negative log density values ​​is as follows: The trained improved DAGMM model and corresponding data normalization parameters are called, without updating the model weights. The perturbation response weaving tensor is normalized according to the field mean and field standard deviation used in the training phase. The normalization result is input into the response compression and reconstruction module. Each perturbation response phase strip is encoded according to the parent response edge and propagation order to generate the perturbation low-dimensional response representation and the perturbation target response reconstruction result. The perturbation response weaving tensor is input into the parent edge occlusion combination inversion module. According to the S main parent response edge set, the combination inputs corresponding to full occlusion, single parent response edge retention, multi-parent response edge combination retention and full retention are generated. Shared parameter reconstruction is performed respectively. The difference between the single parent response edge retention result and the full occlusion result is used to generate the perturbation parent edge independent inheritance response. The remaining result after deducting the corresponding perturbation parent edge independent inheritance response from the multi-parent response edge combination retention result is used to generate the perturbation parent edge coupling response. The difference between the common parent response edge retention result and the full occlusion result is used to generate the perturbation common impact response. The above responses are deducted from the actual response of the perturbation target device and the perturbation data evidence deficiency is separated according to the data integrity status. The remaining part is determined as the perturbation target local response deficiency, and the perturbation response source accounting result is generated. Based on the disturbance type, response relationship type, source device category, target device category, propagation sequence, disturbance response contract version, and load interval at the time of disturbance in the data center operation condition response contract diagram, a basic responsibility key for the disturbance is generated. The responsibility Gaussian component with the same basic responsibility key and corresponding load interval is retrieved from the responsibility Gaussian estimation module. The low-dimensional response representation of the disturbance and the source accounting result of the disturbance response are written into the responsibility Gaussian component. The negative log density of the corresponding sample in the responsibility Gaussian component is calculated, and the conditional negative log density value of the disturbance operation condition sequence is generated. By comparing the perturbation response records, perturbation response source allocation results, and the negative logarithmic density values ​​of the perturbation condition sequence conditions with the corresponding results of candidate fault causes, candidate fault causes with consistent response direction, response delay, response amplitude, response source allocation category, and condition density change relationships are identified as target fault causes. If no consistent results are found, an unknown fault cause label is generated. Specifically, the comparison of the perturbation response records, perturbation response source allocation results, and the negative logarithmic density values ​​of the perturbation condition sequence conditions with the corresponding results of candidate fault causes is as follows: Read the expected record of action cause according to the reversible perturbation action identifier, and form a candidate comparison set by combining the candidate fault cause identifier, expected response, expected accounting result and expected condition density change corresponding to the same reversible perturbation action. Read the collection time of the first response direction after the action is executed that conforms to the data center working condition response contract diagram and the response change reaches the lower limit of the response amplitude from the perturbation response record. Subtract the action execution time from the collection time to obtain the actual response delay. Read the maximum value of the absolute value of the response change in the response stability region as the actual response amplitude. Determine the actual response direction from the response change direction in the response formation region and the response stability region. For each candidate fault cause, a direction consistency degree of 1 is generated when the actual response direction is consistent with the expected response direction, and a direction consistency degree of 0 is generated when they are inconsistent. The absolute value of the difference between the midpoint of the actual response delay and the expected response delay interval is calculated. The difference is divided by the width of the expected response delay interval, and the result that is not less than 1 is written as 1. Then, the processing result is subtracted from 1 to obtain the delay consistency degree. The absolute value of the difference between the midpoint of the actual response amplitude and the expected response amplitude interval is calculated. The difference is divided by the width of the expected response amplitude interval, and the result that is not less than 1 is written as 1. Then, the processing result is subtracted from 1 to obtain the amplitude consistency degree. The order of edge independent inheritance response, parent edge coupled response, common impact response, target local response deficit and data evidence deficit is used to calculate the absolute value of the difference between the corresponding field of the perturbation response source accounting result and the expected accounting result. Each field difference is divided by the normal range width of the corresponding field in the historical normal response record. The average value of all normalized field differences is calculated. The average value that is not less than 1 is written as 1. Then, the accounting consistency is obtained by subtracting the processing result from 1. When the normal range width is 0, the median of the absolute value of the corresponding field's historical normal sample is used as the divisor. When the median of the absolute value is still 0, 1 is used as the divisor. The actual condition density change is obtained by subtracting the negative logarithmic density value of the sequence condition before the perturbation from the negative logarithmic density value of the sequence condition before the perturbation. When the positive and negative directions of the actual condition density change and the expected condition density change are consistent, the absolute value of the difference between the two is calculated and divided by the normal range width of the condition density change in the historical normal response record. The result that is not less than 1 is written as 1. Then, the condition density consistency is obtained by subtracting the processing result from 1. When the positive and negative directions are inconsistent, the condition density consistency is written as 0. When the termination conditions are met, the action object is restored according to the rollback command. The restoration status is verified based on the safety operation data. The target device location is extracted from the configuration control data, the inspection items are extracted from the operation and maintenance inspection data, and the inspection operation conditions are extracted from the safety operation data to generate the inspection task. When the termination conditions are met, the action object is restored according to the rollback command, and the restoration status is verified based on the safety operation data, specifically as follows: The termination condition is read from the perturbation verification command. The termination condition includes any one of the following conditions: the duration of the action reaches the action duration limit, the operating parameters of the action object reach the safety boundary, the target response evidence collection is completed, or the continuous loss of the control link reaches the number of times the safe operation data record is lost. When the termination condition is detected, the reversible perturbation action is stopped and a rollback command is sent to the action object. The rollback command sending time, device acknowledgment time, and rollback execution status are recorded. Extract the discrete state target value, continuous operating parameter rollback target value, rollback allowable deviation, and rollback verification time corresponding to the action object from the safety operation data. Write a discrete state recovery flag when the discrete state is consistent with the discrete state before the perturbation. Write a continuous parameter recovery flag when the absolute value of the difference between the continuous operating parameters and the continuous operating parameters before the perturbation does not exceed the rollback allowable deviation. Write a recovery success status when both the discrete state recovery flag and the continuous parameter recovery flag are formed within the rollback verification time. If a recovery success status is not formed, generate a rollback anomaly flag and write the action object into the emergency inspection object set. The inspection task is generated as follows: Read the original failure response edge identifier, source device identifier, target device identifier and candidate failure cause category from the target failure cause; extract the location code, cabinet number, port number and device connection relationship of the source device and target device from the configuration control data; extract the inspection items, inspection tools, operation sequence and result filling fields corresponding to the candidate failure cause category from the operation and maintenance inspection data; and extract the power failure conditions, bypass conditions, personnel protection conditions and prohibited operation conditions from the safety operation data. When the target fault is caused by the source device failing to perform its action, the source device is identified as the primary inspection object and the target device is identified as the associated verification object. When the target fault is caused by the target device failing to perform its action, the target device is identified as the primary inspection object and the source device is identified as the associated verification object. When the target fault is caused by the data acquisition evidence failing, the corresponding sensor, acquisition interface, gateway port, and data link of the target device are identified as the primary inspection objects. When there is an unknown fault cause marker, all associated devices are extracted along the abnormal ancestor path and downstream response path of the original failed response edge and written into the extended inspection object set. Write the following fields into the task record: inspection task identifier, target fault cause, main inspection object, associated verification object, extended inspection object set, equipment location, inspection items, operation sequence, inspection tools, safe operation conditions, disturbance response evidence, disturbance response source accounting result, rollback status, and result entry field. Generate inspection tasks in the order of main inspection object first, associated verification object second, and extended inspection objects arranged in ascending order of propagation sequence.

[0030] In this embodiment, generating the communication equipment room inspection results includes: Perform inspection tasks, collect equipment verification results, data acquisition link verification results, configuration control verification results, maintenance and handling results, and inspection execution status, write them into operation and maintenance inspection data, and generate inspection feedback; Based on the inspection feedback, generate equipment fault markers, data acquisition anomaly markers, configuration control change markers, and normal response markers, update the dual time-stamped response evidence envelope, and write the normal response record and equipment fault record into the normal inspection record and abnormal inspection record, respectively. Candidate contract patches are generated based on configuration control change flags. The candidate contract patches are verified using normal inspection records and abnormal inspection records. When the verification is successful, a new operating condition response contract version is generated and written into the operating condition response contract version grid. The target fault cause, inspection execution status, maintenance and handling results and update records are collected to generate the communication equipment room inspection results.

[0031] refer to Figure 3 A communication equipment room inspection system based on big data includes the following modules: The inspection data processing module is used to process the computer room inspection data and generate standard computer room inspection data and dual time-stamped response evidence envelopes. The working condition contract construction module is used to identify working condition disturbance events from standard data center inspection data and generate working condition response contract version format and data center working condition response contract diagram. The response phase weaving module is used to generate a response weaving tensor based on the dual-time-scaled response evidence envelope and the data center operating condition response contract diagram; The response default analysis module is used to input the response weaving tensor and the data center operating condition response contract diagram into the improved DAGMM model to generate the response source accounting results and the original default anti-chain. The perturbation instruction generation module is used to generate candidate fault causes, expected responses, and perturbation verification instructions based on the response source accounting results and the original default redemption backlink. The fault verification and inspection module is used to execute perturbation verification instructions, determine the target fault cause based on perturbation response data, candidate fault causes and expected response, and generate inspection tasks. The inspection feedback update module is used to execute inspection tasks and collect inspection feedback, update the dual-timescale response evidence envelope, data center inspection data and operating condition response contract version format, and generate communication data center inspection results.

[0032] Example 1: In a communication room configured with power, cooling, environmental monitoring, and data acquisition links, 84 communication devices, 2 AC power distribution cabinets, 2 uninterruptible power supplies, 4 precision air conditioners, 2 sets of batteries, 48 ​​environmental sensors, and 3 data acquisition gateways are deployed. The devices upload data at 1-second intervals. During operation, the room experiences device clock skew, gateway buffering, platform data entry delays, and cross-device response transmission. Traditional fixed-time-window methods easily misidentify link delays as device delays and classify alarms caused by upstream anomalies as independent faults, leading to an increase in the number of inspection targets and a longer location time.

[0033] The system extracted 528 normal operating condition disturbance events from the inspection records that had been verified on-site, forming a total of 18,432 normal response records. Of these, 12,914 were used as training data, 3,699 as validation data, and 1,819 as normal test data. Each event record was included in only one dataset. The training samples included load changes, power supply switching, cooling start / stop, and data acquisition link delays. In one sample, the communication service load increased from 61.3% to 78.5%, the uninterruptible power supply output current increased by 22.6% after 1.7s, the percentage of the precision air conditioner fan's rated speed increased by 15.4 percentage points after 6.2s, and the device generation to gateway reception delay was 0.38s. In another sample, the device physical response delay was 2.1s, and the device generation to gateway reception delay was 4.8s.

[0034] Each response record is converted into a 6×5×14 response weaving tensor slice. The 6 slots correspond to 4 primary parent response edges, 1 common parent response edge, and 1 target response edge. The improved DAGMM model generates a 32-dimensional low-dimensional response representation and a 37-dimensional responsibility feature vector. Parent edge occlusion samples are added to the training data. The Adam optimizer is used for training with a learning rate of 0.001 and a batch size of 64. Training stops at the 187th round, and the parameters from the 167th round are saved.

[0035] During the simulated inspection, the communication service load increased from 54.8% to 76.9% within 15 seconds, an increase of 22.1 percentage points. The operational condition response contract stipulates that the uninterruptible power supply (UPS) output current should increase by 18%–30% within 0.8–2.5 seconds, the percentage of the precision air conditioner fan's rated speed should increase by 12–20 percentage points within 4–9 seconds, and the return air temperature should decrease by 0.4–1.1℃ within 35–90 seconds. The data collection results show that the UPS output current increased by 23.4% after 1.6 seconds, the precision air conditioner controller returned a successful execution response after 5.2 seconds, the percentage of the fan's rated speed only increased by 1.8 percentage points, the motor current only increased by 0.1A, and the return air temperature only decreased by 0.1℃ at the end of the 90-second observation window. The device generation time for temperature sampling records was offset by 0.9 seconds relative to the occurrence time of the disturbance event, resulting in a 7.4-second link delay from device generation to gateway reception. Traditional methods generated four types of alarms.

[0036] This invention establishes a dual-timescale response evidence envelope based on device generation time, gateway reception time, and platform data entry time. After correction, the physical response delay of the uninterruptible power supply is 1.6s, the air conditioning control response delay is 5.2s, and the additional delay of the temperature sampling link is 7.4s. The system matches the current operating condition response contract version, establishes corresponding inter-device response edges, writes the device responses into five response regions, and generates a response weaving tensor. The improved DAGMM model normalizes the absolute response quantities of the five response sources, obtaining the dimensionless response source proportions. Among them, the parent edge independent inheritance response is 0.31, the parent edge coupled response is 0.12, the common influence is 0.08, the target local response deficiency is 0.42, the data evidence deficiency is 0.07, and the operating condition sequence condition negative logarithmic density value is 8.63, exceeding the upper limit of the normal range of 5.21. After deducting the influence of abnormal ancestor transmission, the dimensionless local response deficit of the precision air conditioner fan response edge is 0.39, which is still higher than the upper limit of the normal range of 0.14. The physical response deficit of the temperature sensor response edge falls back to the normal range. The system determines the precision air conditioner fan response edge as the original out-of-range response edge.

[0037] The system generates two candidate fault causes: target equipment execution failure and data acquisition evidence failure. Safety operation data limits the allowable range of the fan's rated speed percentage to 40%–85%, with a single adjustment not exceeding 5 percentage points, an action duration not exceeding 20 seconds, and a motor current not exceeding 7A. The system selects increasing the fan setpoint by 5 percentage points as a reversible perturbation action. After the command is executed, both the controller setpoint and output feedback increase by 5 percentage points, the fan rated speed percentage increases by only 0.7 percentage points, the motor current increases by 0.05A, the conditional negative logarithmic density value increases from 8.63 to 9.11, the target equipment execution failure matching value is 0.92, and the data acquisition evidence failure matching value is 0.36. The system determines the target equipment execution failure as the target fault cause.

[0038] The system restored the fan setpoints and generated inspection tasks for the fan actuator circuit contactor, fan actuator, and feedback interface. On-site verification revealed that the resistance of the fan actuator circuit contactor had increased. After maintenance, the same amplitude perturbation was executed again, and the fan rated speed ratio increased by 5.1 percentage points, the motor current increased by 0.82A, and the return air temperature decreased by 0.7℃ within 61s. All responses were restored to the contract range.

[0039] The comparative experiment used the same 12,914 training records, 3,699 verification records, 1,819 normal test records, and 240 abnormal test records. The traditional method uses ordinary DAGMM to calculate the abnormal energy, and then combines fixed time windows, equipment thresholds, and alarm association rules to generate fault causes and inspection objects. The traditional method correctly identified 175 records, with an accuracy rate of 72.9%, while the present invention correctly identified 220 records, with an accuracy rate of 91.7%. Among 61 data link delay anomaly records, the traditional method misclassified 15, with a misclassification rate of 24.6%, while the present invention misclassified 4, with a misclassification rate of 6.6%. The traditional method generated 76 repeated alarms for propagation anomalies, with a repeat alarm rate of 31.7%, while the present invention generated 20, with a repeat alarm rate of 8.3%. The average number of inspection objects corresponding to each anomaly decreased from 3.2 to 1.4, and the average fault location time was shortened from 43.8 minutes to 18.6 minutes. The number of consistent records verified after maintenance increased from 188 to 227, and the consistency rate improved from 78.3% to 94.6%. This embodiment demonstrates that the present invention can separate the physical response of the equipment from the data link delay, identify the original lost response edge, and verify the cause of the fault through reversible perturbation, thereby improving the accuracy of fault location and inspection efficiency.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A communication equipment room inspection method based on big data, characterized in that, include: Collect data from the computer room inspection, perform unified identification, time correction, object association, and complete status marking, and generate standard computer room inspection data and dual time-scaled response evidence envelopes; Identify operational disturbance events from standard data center inspection data, match operational response contract versions, and generate a data center operational response contract diagram; Based on the dual-timescaled response evidence envelope correction device response time, response phase strips are constructed and arranged to generate a response weaving tensor; The response weaving tensor and the data center operating condition response contract graph are input into the improved DAGMM model. The improved DAGMM model includes a response compression and reconstruction module, a parent edge blocking combination inversion module, a duty Gaussian estimation module, and a default evidence output module to generate response source accounting results and original default anti-chain. Extract the original default response edge from the original default antichain, generate candidate failure causes by combining the response source accounting results, screen reversible perturbation actions based on the expected response differences of each candidate failure cause, and generate perturbation verification instructions. Execute the perturbation verification command, collect perturbation response data and re-input it into the improved DAGMM model, determine the cause of the target fault, and generate the inspection task; Perform inspection tasks and collect inspection feedback. Update the dual-timescaled response evidence envelope and equipment room inspection data based on the inspection feedback, and generate communication equipment room inspection results.

2. The communication equipment room inspection method based on big data according to claim 1, characterized in that, The data center inspection data includes equipment operation data, data center operating status data, configuration control data, data acquisition link data, operation and maintenance inspection data, operating status response contract data, and safety operation data.

3. The communication equipment room inspection method based on big data according to claim 1, characterized in that, The generated standard data center inspection data and dual-timescaled response evidence envelope include: Extract device identifier, acquisition record identifier, device generation time, gateway reception time, platform entry time, clock source identifier, and clock calibration record from the data acquisition link data, and associate them with device operation data and data center operating status data; Based on the clock calibration record, the device generation time is corrected, the device physical response time is generated, and the data arrival time is generated based on the gateway reception time and the platform entry time. Two time ranges are determined based on equipment identification and clock source identification. The data integrity status is extracted from equipment operation data and computer room operating data, written into the acquisition record, and generated dual-timescale response evidence envelope and standard computer room inspection data.

4. The communication equipment room inspection method based on big data according to claim 1, characterized in that, The generated data center operational condition response contract diagram includes: Extract equipment status, control commands, communication service load and environmental parameters from equipment operation data and computer room operating data; extract operating condition triggering conditions from operating condition response contract data; and identify operating condition disturbance events. Extract the contract version identifier, parent version identifier, validity period and applicable conditions from the operating condition response contract data, connect the contract versions to generate the operating condition response contract version grid, and determine the operating condition response contract version based on the current conditions extracted from configuration control data, equipment operation data, computer room operating condition data and operation and maintenance inspection data. Extract the source device, target device, response direction, response delay, response magnitude, response duration, response recovery process, parent response edge, and propagation order from the operating condition response contract version. Establish directed response edges with devices as nodes, and write the operating condition disturbance event and the operating condition response contract version into the directed response edges to generate the data center operating condition response contract graph.

5. The communication equipment room inspection method based on big data according to claim 1, characterized in that, The arrangement generates a response weaving tensor, including: Based on the equipment identification and the equipment physical response time, the collected records in the dual-time-scaled response evidence envelope are associated with the operating condition disturbance events and response edges in the data center operating condition response contract diagram; Read the event occurrence time, response delay, response duration, response recovery process, response direction, parent response edge, and propagation sequence from the data center operating condition response contract diagram. Divide the response process into an early response zone, a response formation zone, a response stability zone, a response recovery zone, and a delayed response zone. Allocate equipment response data according to the overlap between the equipment's physical response time range and each response zone. Combine the response direction and data integrity status to form response phase stripes. Each bifurcation response phase strip is retained, and the parent response edge strip and target response edge strip that merge into the same target device are arranged separately. The response weaving tensor is generated by arranging the response edge, parent response edge, propagation order, and response region.

6. The communication equipment room inspection method based on big data according to claim 1, characterized in that, The arrangement generates a response weaving tensor, including: The compression network and estimation network of the DAGMM model are replaced with the response compression reconstruction module and the responsibility Gaussian estimation module. A parent edge blocking combination inversion module is set between the two, and a mismatch evidence output module is set at the output end to construct an improved DAGMM model. The response compression and reconstruction module encodes the response phase strip, generates a low-dimensional response representation based on the combination encoding results of the parent response edge and propagation order, reconstructs the target device response, and returns the source device category, parent response edge combination, and propagation order. The parent edge occlusion combination inversion module extracts historical normal response records from the operation and maintenance inspection data and determines the main parent response edges. It then generates the combination inputs corresponding to all occlusion, single parent response edge retention, multi-parent response edge combination retention, and all retention and performs reconstruction. The parent edge independent inheritance response is generated by the difference between the single parent response edge retention result and the total occlusion result; the parent edge coupled response is generated by the remaining result after deducting the corresponding independent inheritance response from the multi-parent response edge combination retention result; and the common influence response is generated by the difference between the common parent response edge retention result and the total occlusion result. The parent edge independent inheritance response, parent edge coupled response, and common impact response are deducted from the actual response of the target device. The data evidence deficit is separated according to the data integrity status. The remaining part is determined as the target local response deficit. The back-pointing result is verified with the data center operating condition response contract diagram to generate the response source accounting result. The duty Gaussian estimation module generates basic duty keys based on the type of operating condition disturbance, the type of response relationship, the type of source and target equipment, the propagation sequence, and the version of the operating condition response contract. It organizes the duty Gaussian components according to the load interval and performs component splitting and merging of adjacent components to generate the conditional negative logarithmic density value of the operating condition sequence. The missing-redemption evidence output module determines the normal range of local response deficit and the normal range of condition density based on historical normal response records. Response edges whose results are both outside the corresponding normal range are identified as candidate missing-redemption response edges. Independent inherited responses and coupled responses passed in by abnormal ancestor response edges are deducted from the target local response deficit. Candidate missing-redemption response edges whose deduction results are within the normal range are deleted. The remaining candidate missing-redemption response edges that do not have ancestor-descendant relationships are combined into the primary missing-redemption antichain.

7. The communication equipment room inspection method based on big data according to claim 1, characterized in that, The generation of perturbation verification instructions includes: Extract the original default response edge from the original default anti-chain, and generate candidate fault causes corresponding to source device action failure, target device execution failure, and data collection evidence failure based on the source device response, target local response deficit, and data evidence deficit. Replace the source device action field, target device fulfillment field and data acquisition field corresponding to the candidate fault causes respectively, and input the replaced response weaving tensor and the data center operating condition response contract diagram into the improved DAGMM model to generate the expected response, expected revenue sharing result and expected condition density change corresponding to the candidate fault causes. Extract executable actions, action boundaries, allowable load ranges, prohibited execution conditions, rollback instructions, and termination conditions from the safety operation data. Write each executable action into the expected response and update the response direction, response latency, and response magnitude according to the data center operating condition response contract diagram. By comparing the expected responses, expected revenue sharing results, and expected condition density changes under the same executable action, executable actions that do not meet the safe operation conditions are eliminated. Reversible perturbation actions are determined according to the degree of separability of candidate causes, and perturbation verification instructions are generated.

8. The communication equipment room inspection method based on big data according to claim 1, characterized in that, The generation of inspection tasks includes: Execute the perturbation verification command, extract perturbation response records from equipment operation data, data center operating data and data acquisition link data, correct the perturbation response time based on the dual time-scaled response evidence envelope, read the response relationship from the data center operating response contract diagram, and generate a perturbation response weaving tensor. Input the perturbation response weaving tensor and the data center operating condition response contract diagram into the improved DAGMM model to generate perturbation response source accounting results and perturbation operating condition order condition negative log density values. Compare the perturbation response records, perturbation response source accounting results, and negative logarithmic density values ​​of perturbation operating condition sequence conditions with the corresponding results of candidate fault causes. Candidate fault causes with consistent response direction, response delay, response amplitude, response source accounting category, and condition density change relationship are identified as target fault causes. If no consistent results are formed, an unknown fault cause label is generated. When the termination conditions are met, the action object is restored according to the rollback instruction, the restoration status is verified based on the safety operation data, the target device location is extracted from the configuration control data, the inspection items are extracted from the operation and maintenance inspection data, the inspection operation conditions are extracted from the safety operation data, and the inspection task is generated.

9. The communication equipment room inspection method based on big data according to claim 1, characterized in that, The generated communication equipment room inspection results include: Perform inspection tasks, collect equipment verification results, data acquisition link verification results, configuration control verification results, maintenance and handling results, and inspection execution status, write them into operation and maintenance inspection data, and generate inspection feedback; Based on the inspection feedback, generate equipment fault markers, data acquisition anomaly markers, configuration control change markers, and normal response markers, update the dual time-stamped response evidence envelope, and write the normal response record and equipment fault record into the normal inspection record and abnormal inspection record, respectively. Candidate contract patches are generated based on configuration control change flags. The candidate contract patches are verified using normal inspection records and abnormal inspection records. When the verification is successful, a new operating condition response contract version is generated and written into the operating condition response contract version grid. The target fault cause, inspection execution status, maintenance and handling results and update records are collected to generate the communication equipment room inspection results.

10. A communication equipment room inspection system based on big data, comprising executing the communication equipment room inspection method based on big data as described in any one of claims 1 to 9, characterized in that, Includes the following modules: The inspection data processing module is used to process the computer room inspection data and generate standard computer room inspection data and dual time-stamped response evidence envelopes. The working condition contract construction module is used to identify working condition disturbance events from standard data center inspection data and generate working condition response contract version format and data center working condition response contract diagram. The response phase weaving module is used to generate a response weaving tensor based on the dual-time-scaled response evidence envelope and the data center operating condition response contract diagram; The response default analysis module is used to input the response weaving tensor and the data center operating condition response contract diagram into the improved DAGMM model to generate the response source accounting results and the original default anti-chain. The perturbation instruction generation module is used to generate candidate fault causes, expected responses, and perturbation verification instructions based on the response source accounting results and the original default redemption backlink. The fault verification and inspection module is used to execute perturbation verification instructions, determine the target fault cause based on perturbation response data, candidate fault causes and expected response, and generate inspection tasks. The inspection feedback update module is used to execute inspection tasks and collect inspection feedback, update the dual-timescale response evidence envelope, data center inspection data and operating condition response contract version format, and generate communication data center inspection results.

Images