A Deep Learning-Based Intelligent Monitoring System for Chip Operation Status

By constructing a deep learning-based intelligent monitoring system for chip operating status, generating state sequences and offset sequences, and combining closed-loop transmission topology, the system solves the problem of monitoring the linked changes in chip operating status, and achieves high-precision anomaly identification and location.

CN123114889APending Publication Date: 2026-11-17XIONGAN MIG TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611278717.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing chip operating status monitoring technologies are unable to uniformly depict the interconnected changes between multiple operating states, and lack the ability to analyze the monitoring stability and anomaly propagation process under complex load scenarios.

Method used

A deep learning-based intelligent monitoring system for chip operation status is constructed. By generating operation status sequences, state evolution process sequences, steady-state reference surfaces, and state offset sequences, and combining functionally coupled closed-loop conduction topology, a process offset CfC model is built to achieve continuous analysis of chip operation status.

Benefits of technology

It improves the accuracy of chip operating status classification, enhances the continuity of abnormal offset intensity analysis and the completeness of propagation relationship identification, can locate the starting position of abnormality, and improves the interpretability of monitoring results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114889A_ABST
    Figure CN123114889A_ABST
Patent Text Reader

Abstract

This invention discloses a deep learning-based intelligent monitoring method for chip operating status, comprising: a running data preprocessing module for collecting chip operating status data and generating an operating status sequence; a process sequence construction module for constructing a state evolution process sequence based on the operating status sequence; a reference plane offset construction module for constructing a steady-state reference plane and generating a state offset sequence based on the operating status sequence; a model construction module for constructing a process offset CfC model; an offset propagation calculation module for generating process driving variables, correcting the state offset sequence, and offset propagation representation; a closed-loop propagation calculation module for constructing a functionally coupled closed-loop propagation topology and generating state propagation results; and a state monitoring output module for generating chip operating status monitoring results based on the state propagation results. This invention improves the accuracy and propagation positioning capability of chip operating status monitoring.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of integrated circuit operation monitoring, and in particular to an intelligent monitoring system for chip operation status based on deep learning. Background Technology

[0002] With the widespread application of high-performance chips, systems-on-a-chip (SoCs), and heterogeneous computing platforms, the operational stability of chips under complex loads, frequency switching, power consumption fluctuations, and thermal coupling conditions is becoming increasingly prominent. Existing chip operational status monitoring technologies mainly focus on threshold alarms, trend analysis, and anomaly detection for individual operational parameters such as temperature, voltage, current, and power consumption. While these technologies can detect local anomalies to some extent, they lack a unified capability to characterize the interconnected changes between multiple operational states within the chip.

[0003] In existing technologies, some solutions can collect chip operating parameters and identify anomalies, while others can analyze and model fault propagation paths, hotspot distributions, and power consumption distributions. Some continuous-time neural network models have also been applied to time-series state modeling scenarios. However, when these technologies are directly applied to chip operating status monitoring, there are still significant shortcomings: First, multi-source operating status data differ in sampling granularity, time scale, and field organization, making it difficult to form a unified operating status representation that can be directly used for subsequent analysis. Second, existing solutions typically make anomaly judgments based on raw monitoring data, which is difficult to accurately reflect the degree of deviation of the chip's state under different operating modes, resulting in insufficient stability of monitoring results in complex load scenarios. Third, there is a lack of integrated analysis methods for the coupling and transmission relationships between computation execution, memory access, on-chip interconnects, power supply, and thermal response, making it difficult to accurately identify the anomaly propagation process and the anomaly initiation location.

[0004] Therefore, how to provide a chip operation status intelligent monitoring system based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent chip operating status monitoring system based on deep learning. This invention collects chip operating status data, generates operating status sequences, state evolution process sequences, steady-state reference surfaces, and state offset sequences, constructs a process offset CfC model, and combines this with a functionally coupled closed-loop transmission topology to generate chip operating status monitoring results. Compared to methods that rely solely on a single monitoring quantity for judgment, this invention can simultaneously output the current operating status category, abnormal offset intensity, risk propagation path, and abnormal starting location, offering advantages such as high monitoring accuracy, strong propagation analysis capabilities, and clear location results.

[0006] According to an embodiment of the present invention, a deep learning-based intelligent monitoring system for chip operating status includes:

[0007] The runtime data preprocessing module is used to collect chip runtime status data, perform alignment, cleaning, and standardization processing, and generate runtime status sequences.

[0008] The process sequence construction module is used to construct a state evolution process sequence based on the running state sequence;

[0009] The reference surface offset construction module is used to construct the steady-state reference surface corresponding to each operating mode based on the operating state sequence, and generate the state offset sequence based on the difference relationship between the steady-state reference surface and the operating state sequence.

[0010] The model building module is used to build the process offset CfC model, including the process encoding unit, reference plane alignment unit, offset propagation unit, closed propagation unit and state determination unit.

[0011] The offset propagation calculation module is used to input the state evolution process sequence into the process encoding unit to generate process driving variables, input the running state sequence and steady-state reference surface into the reference surface alignment unit, perform offset correction processing to generate a corrected state offset sequence, input the corrected state offset sequence into the offset propagation unit, perform continuous propagation calculation of state offset, and generate an offset propagation representation.

[0012] The closed-loop conduction calculation module is used to construct a functionally coupled closed-loop conduction topology based on the internal operational coupling relationship of the chip, and combine process driving variables and offset propagation representation to perform closed-loop conduction calculation of state offset in the closed-loop conduction unit to generate state conduction results.

[0013] The status monitoring output module is used to generate chip operating status monitoring results by inputting the status transmission results into the status determination unit.

[0014] Optionally, the runtime data preprocessing module includes:

[0015] Collect chip operating status data, including temperature data, voltage and current data, frequency data, power consumption data, memory access data, on-chip interconnect data, task scheduling data, and operation log data, and record the sampling timestamp;

[0016] The sampling timestamp is converted into a unified time index, the chip operating status data is mapped to the time step corresponding to the unified time index, and the continuous time steps are organized into windows according to a fixed time window length.

[0017] Based on a unified sampling alignment rule, time step alignment processing is performed on the chip operating status data mapped to a unified time index;

[0018] Perform cleaning processing on the data that has completed time step alignment;

[0019] The cleaned data is standardized, and the chip operating status data is uniformly quantified, numerically mapped and rearranged according to the unified monitoring field rules. The chip operating status data corresponding to each time step is combined into a multi-dimensional vector according to a fixed field order.

[0020] The multidimensional vectors are continuously arranged according to the time step order under a unified time index to generate a running state sequence.

[0021] Optionally, the process sequence construction module includes:

[0022] Based on the unified time index corresponding to the running state sequence, the running state sequence is divided into multiple time periods according to continuous time steps;

[0023] The thermal change is determined based on the cumulative trend and rise / fall conversion of temperature data within each time period; the power supply change is determined based on the combined fluctuation of voltage, current and power consumption data within each time period; and the frequency change is determined based on the frequency switching of frequency data within each time period.

[0024] The amount of interconnection change is determined based on the changes in link occupancy, transmission load migration, and congestion expansion of on-chip interconnection data within each time period. The amount of execution load change is determined based on the stage switching and resource occupancy migration of task scheduling data, storage access data, and power consumption data within each time period.

[0025] The task phase switching order and resource usage change order between different time periods are determined based on the running status sequence;

[0026] According to the task phase switching order and resource usage change order, the thermal change, power supply change, frequency change, interconnection change and execution load change are sequentially arranged and combined into a structured vector.

[0027] Following the time step order under a unified time index, the structured vectors corresponding to each time period are arranged continuously, and the advancement relationship between adjacent time periods is established according to the task stage switching order and the resource consumption change order, thus generating a state evolution process sequence.

[0028] Optionally, the reference plane offset construction module includes:

[0029] Based on the running state sequence, load configuration data, resource usage data, frequency configuration data and task stage data are extracted to form a combined state sequence;

[0030] Based on the combined state sequence, continuous time periods with consistent combined states are divided into the same operating mode, and an operating mode identifier is established for each operating mode.

[0031] For each operating mode, extract the temperature data, voltage data, current data, frequency data, power consumption data, memory access data, and on-chip interconnect data within the time period corresponding to the operating mode identifier, and form temperature distribution, voltage distribution, current distribution, frequency distribution, power consumption distribution, memory access distribution, and on-chip interconnect distribution respectively. Perform steady-state aggregation and arrange the steady-state aggregation results into a matrix form of steady-state reference surface.

[0032] Based on the operating mode identifier corresponding to each time step in the operating state sequence, a steady-state reference surface corresponding to the operating mode is matched for each time step;

[0033] According to the fixed field order and distribution statistical position, for the temperature data, voltage data, current data, frequency data, power consumption data, memory access data and on-chip interconnect data of each time step, a dimension-wise difference relationship is established with the matched steady-state reference surface to generate the state offset vector corresponding to that time step.

[0034] The state offset vectors corresponding to each time step are arranged continuously according to the time step order under a unified time index to generate a state offset sequence.

[0035] Optionally, the model building module includes:

[0036] The basic structure of the original CfC model is invoked, including the input mapping structure, the continuous-time hidden state update structure, the closed-form state evolution structure, and the output mapping structure.

[0037] The process coding unit is constructed by replacing the mapping layer in the input mapping structure that is oriented towards general temporal inputs with a process mapping layer that is oriented towards state evolution process sequences.

[0038] Based on the input interface between the input mapping structure of the original CfC model and the continuous-time hidden state update structure, a reference face alignment unit is constructed.

[0039] Based on the continuous-time hidden state update structure, an offset propagation unit is constructed.

[0040] Based on the closed state evolution structure, the closed state evolution path in the original CfC model that is oriented towards the general hidden state update is transformed into a closed propagation path that is oriented towards the closed propagation of state offset, and a closed propagation unit is constructed.

[0041] The readout layer in the output mapping structure that is oriented towards general prediction results is replaced with a decision mapping layer that is oriented towards chip operating status monitoring results, and a status decision unit is constructed.

[0042] The process encoding unit, reference plane alignment unit, offset propagation unit, closed-loop transmission unit, and state determination unit are networked together to generate a process offset CfC model.

[0043] Optionally, the offset propagation calculation module includes:

[0044] Based on the state evolution process sequence, the thermal change, power change, frequency change, interconnection change and execution load change corresponding to each time period are written into the process coding unit at fixed positions, and continuous time recursive coding is performed to generate process driving quantities.

[0045] Match the running mode identifier corresponding to each time step in the running state sequence with the corresponding running mode identifier in the steady-state reference surface to determine the task stage boundary position and resource occupation boundary position, and generate the reference surface alignment processing result.

[0046] Based on the reference plane alignment processing results, offset correction processing is performed on the state offset sequence. The offset direction and offset magnitude corresponding to each time step are corrected by amplitude correction. At the task stage boundary position and resource occupation boundary position, the offset data of cross-boundary connection is segmented and corrected to generate the corrected state offset sequence.

[0047] Based on the corrected state offset sequence, the offset propagation unit is input, and continuous propagation calculation is performed to generate the propagation results corresponding to each time step.

[0048] Based on the propagation results corresponding to each time step, the time steps are arranged continuously according to the time step order corresponding to the unified time index, and combined into a time step vector sequence according to the fixed field positions to generate the offset propagation representation.

[0049] Optionally, the closed-loop conduction calculation module includes:

[0050] Based on the sequence of operating states, the computation execution relationships, memory access relationships, on-chip interconnection relationships, power supply relationships, and thermal response relationships within the chip are determined.

[0051] A unified location mapping is performed on computation execution relationships, memory access relationships, on-chip interconnect relationships, power supply relationships, and thermal response relationships. The computation activity location, memory access location, link location, power supply location, and temperature location are mapped to functional location nodes, and the start location and destination location in each relationship are mapped to node connection relationships.

[0052] Based on the frequency of occurrence, duration, associated power consumption intensity, and temperature response intensity of each node connection relationship, the coupling weight corresponding to each node connection relationship is determined.

[0053] Construct a functionally coupled closed-loop transmission topology based on functional location nodes, node connection relationships, and coupling weights;

[0054] Write the process-driven variable into the node input position corresponding to the functional location node, write the offset propagation representation into the offset input position corresponding to the functional location node, and connect the functionally coupled closed-loop transmission topology to the topology input position of the closed-loop transmission unit to establish a joint input correspondence.

[0055] Based on the closed-loop transmission unit, the closed-loop transmission calculation of state offset is performed on each functional position node in the joint input correspondence to generate the node transmission result corresponding to each time step;

[0056] The node transmission results corresponding to each time step are arranged in the order of time steps corresponding to the unified time index, and combined into state transmission results according to the order of functional position nodes.

[0057] Optionally, the status monitoring output module includes:

[0058] The state transmission results are input into the state determination unit and organized into state determination input data.

[0059] Based on the state determination unit, a determination mapping process is performed on the state determination input data to determine the current running state category;

[0060] Based on the continuous change of the transmission value at each time step in the state determination input data under a unified time index, the cumulative expansion of the transmission value at each functional position under the node connection relationship, and the transmission concentration of the transmission value corresponding to each node connection relationship in the functionally coupled closed transmission topology, the intensity of abnormal offset is determined.

[0061] Based on the transmission record of state transmission results in the functionally coupled closed transmission topology, the state transmission results are backtracked to determine the risk propagation path.

[0062] Based on the connection relationship of the starting node, the starting time step position and the starting function position corresponding to the risk propagation path, and combined with the first concentrated position of the abnormal offset intensity in the risk propagation path, the abnormal starting position is determined.

[0063] The current operating status category, abnormal offset intensity, risk propagation path, and abnormal starting position are combined in a fixed field order to generate chip operating status monitoring results.

[0064] The beneficial effects of this invention are:

[0065] First, this invention constructs an operating state sequence, a state evolution process sequence, a steady-state reference surface, and a state offset sequence, and on this basis, constructs a process offset CfC model. This transforms the chip operating state monitoring process from simple parameter detection into a continuous analysis process centered on the relationship between the progress of the operating process and the relationship between state offsets. This can more accurately distinguish between normal fluctuations and abnormal offsets, thereby improving the accuracy of the current operating state category determination and reducing the misjudgment caused by relying solely on single monitoring data.

[0066] Secondly, this invention performs corresponding analysis between the operating state sequence and the steady-state reference plane, and combines the state offset sequence, offset propagation representation and functionally coupled closed-loop conduction topology to carry out state propagation processing. This enables the unified inclusion of computation execution relationships, memory access relationships, on-chip interconnection relationships, power supply relationships and thermal response relationships between different functional locations inside the chip into the monitoring link, thereby improving the continuity of abnormal offset intensity analysis and the completeness of propagation relationship identification, and enhancing the ability to characterize the abnormal propagation process under complex operating conditions.

[0067] Furthermore, this invention generates risk propagation paths and anomaly initiation locations based on state transmission results. It can not only provide monitoring conclusions on whether the chip is in an abnormal state, but also provide the propagation direction and initiation location of the anomaly within the chip. This improves the interpretability and localization capability of chip operation status monitoring results, facilitating subsequent targeted operation analysis, risk investigation, and status handling around the anomaly location. Attached Figure Description

[0068] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0069] Fig. 1 This is an overall flowchart of a deep learning-based intelligent monitoring method for chip operating status proposed in this invention.

[0070] Fig. 2 The flowchart for constructing the state evolution process sequence, steady-state reference surface, and state offset sequence proposed in this invention is shown below.

[0071] Fig. 3 This is a flowchart of the process offset CfC model, offset propagation calculation, and closed-loop propagation calculation proposed in this invention. Detailed Implementation

[0072] 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.

[0073] refer to Figs. 1-3A deep learning-based intelligent monitoring system for chip operating status includes:

[0074] The runtime data preprocessing module is used to collect chip runtime status data, which includes temperature data, voltage data, current data, frequency data, power consumption data, memory access data, on-chip interconnect data, task scheduling data, and runtime log data. Based on unified time indexing rules, unified monitoring field rules, and unified sampling alignment rules, the module performs alignment, cleaning, and standardization processing on the chip runtime status data to generate a runtime status sequence.

[0075] The process sequence construction module is used to extract the thermal changes, power changes, frequency changes, interconnection changes and execution load changes corresponding to each time period based on the running state sequence, construct the state evolution process sequence based on the combination results of each change, and use the state evolution process sequence as the basis for continuously updating the chip's running state.

[0076] The reference surface offset construction module is used to construct a steady-state reference surface corresponding to each operating mode based on load configuration data, resource usage data, frequency configuration data and task stage data in the operating state sequence, and generate a state offset sequence based on the difference relationship between the steady-state reference surface and the operating state sequence.

[0077] The model building module is used to build the process offset CfC model. The process offset CfC model includes a process encoding unit, a reference surface alignment unit, an offset propagation unit, a closed propagation unit, and a state determination unit. The reference surface alignment unit is connected to the offset propagation unit. The output of the process encoding unit and the output of the offset propagation unit are connected to the closed propagation unit. The closed propagation unit is connected to the state determination unit.

[0078] The offset propagation calculation module is used to input the state evolution process sequence into the process encoding unit based on the process offset CfC model, generate process driving variables, input the running state sequence and steady-state reference surface into the reference surface alignment unit, perform reference surface alignment processing, perform offset correction processing on the state offset sequence based on the reference surface alignment processing result, generate a corrected state offset sequence, input the corrected state offset sequence into the offset propagation unit, perform continuous propagation calculation of state offset, and generate an offset propagation representation.

[0079] The closed-loop conduction calculation module is used to construct a functionally coupled closed-loop conduction topology based on the operational coupling relationship within the chip. The process driving quantity, offset propagation representation and functionally coupled closed-loop conduction topology are input into the closed-loop conduction unit to perform closed-loop conduction calculation of state offset and generate state conduction results. The operational coupling relationship includes calculation execution relationship, memory access relationship, on-chip interconnect relationship, power supply relationship and thermal response relationship.

[0080] The status monitoring output module is used to input the status transmission results into the status determination unit based on the status transmission results to generate chip operation status monitoring results. The chip operation status monitoring results include the current operation status category, abnormal offset intensity, risk propagation path and abnormal starting position.

[0081] In this embodiment, the runtime data preprocessing module includes:

[0082] Collect chip operating status data, including collecting temperature data from the chip temperature monitoring interface, voltage and current data from the power monitoring interface, frequency data from the clock monitoring interface, power consumption data from the power consumption monitoring interface, memory access data from the memory control interface, on-chip interconnect data from the on-chip interconnect monitoring interface, task scheduling data from the task scheduling interface, and operation log data from the operation log output interface. Record sampling timestamps for temperature data, voltage data, current data, frequency data, power consumption data, memory access data, on-chip interconnect data, task scheduling data, and operation log data.

[0083] Based on the unified time indexing rule, the sampling timestamp is converted into a unified time index, the chip operating status data is mapped to the time step corresponding to the unified time index, and the continuous time steps are organized into windows according to a fixed time window length.

[0084] Based on unified monitoring field rules, a fixed field position correspondence is established for chip operating status data. Event markers, module markers, status markers and time markers in the operating log data are parsed and written into the corresponding field positions of the chip operating status data according to the fixed field position correspondence.

[0085] Based on the unified sampling alignment rules, time step alignment processing is performed on the chip operating status data mapped to the unified time index, time step merging is performed on data with a sampling frequency higher than the unified time index, time step padding is performed on data with a sampling frequency lower than the unified time index, and multiple records within the same time step are put back in place according to the sampling timestamp order.

[0086] The data that has completed time step alignment is cleaned by deleting data records corresponding to null fields, deleting duplicate data records corresponding to duplicate time indices, filling in data items corresponding to missing fields, removing data items that exceed the corresponding monitoring range, and splitting fields in the unstructured text content of the operation log data.

[0087] The cleaned data is standardized, and the chip operating status data is uniformly quantified, numerically mapped and rearranged according to the unified monitoring field rules. The chip operating status data corresponding to each time step is combined into a multi-dimensional vector according to a fixed field order.

[0088] The multidimensional vectors are continuously arranged according to the time step order under a unified time index to generate a running state sequence.

[0089] In this embodiment, the process sequence construction module includes:

[0090] Based on the unified time index corresponding to the running state sequence, the running state sequence is divided into multiple time periods according to continuous time steps, and the start time step and end time step of each time period are determined.

[0091] The thermal change is determined based on the cumulative trend and rise / fall conversion of temperature data within each time period; the power supply change is determined based on the combined fluctuation of voltage, current and power consumption data within each time period; and the frequency change is determined based on the frequency switching of frequency data within each time period.

[0092] The amount of interconnection change is determined based on the changes in link occupancy, transmission load migration, and congestion expansion of on-chip interconnection data within each time period. The amount of execution load change is determined based on the stage switching and resource occupancy migration of task scheduling data, storage access data, and power consumption data within each time period.

[0093] The task phase switching order between time periods is determined based on the task scheduling data in the running state sequence, and the resource occupancy change order between time periods is determined based on the storage access data, on-chip interconnect data and power consumption data in the running state sequence.

[0094] According to the task phase switching order and resource usage change order, the thermal change, power change, frequency change, interconnection change and execution load change are sequentially arranged, and the thermal change, power change, frequency change, interconnection change and execution load change corresponding to each time period are combined into a structured vector at fixed positions.

[0095] According to the time step order under the unified time index, the structured vectors corresponding to each time period are arranged continuously, and the advancement relationship between adjacent time periods is established according to the task stage switching order and the resource consumption change order to generate a state evolution process sequence.

[0096] The progression relationship in the state evolution process sequence is used as the basis for continuously updating the chip's operating state.

[0097] In this embodiment, the reference plane offset construction module includes:

[0098] Based on the running state sequence, load configuration data, resource usage data, frequency configuration data, and task stage data are extracted. A combined state sequence is formed according to the value changes of load configuration data, resource usage data, frequency configuration data, and task stage data in continuous time steps. Among them, load configuration data is obtained by organizing task scheduling data and power consumption data, resource usage data is obtained by organizing storage access data, on-chip interconnect data, and power consumption data, frequency configuration data is obtained by organizing frequency data, and task stage data is obtained by organizing task scheduling data.

[0099] Based on the combined state sequence, a continuous time period in which the combined state of load configuration data, resource usage data, frequency configuration data and task phase data is consistent is divided into the same operating mode. An operating mode identifier is established for each operating mode and written into the time step position of the corresponding time period.

[0100] For each operating mode, extract the temperature data, voltage data, current data, frequency data, power consumption data, memory access data, and on-chip interconnect data within the time period corresponding to the operating mode identifier, and form temperature distribution, voltage distribution, current distribution, frequency distribution, power consumption distribution, memory access distribution, and on-chip interconnect distribution respectively.

[0101] For each operating mode, steady-state aggregation is performed on the temperature distribution, voltage distribution, current distribution, frequency distribution, power consumption distribution, memory access distribution, and on-chip interconnect distribution. The steady-state aggregation results are arranged into a matrix form of a steady-state reference surface according to a fixed field order. The row order of the matrix corresponds to the temperature distribution, voltage distribution, current distribution, frequency distribution, power consumption distribution, memory access distribution, and on-chip interconnect distribution, respectively, while the column order corresponds to the distribution statistics position of the unified time index under the operating mode identifier. Specifically, steady-state aggregation involves, according to the distribution statistics position corresponding to the unified time index, performing steady-state aggregation on each temperature distribution, voltage distribution, current distribution, frequency distribution, power consumption distribution, memory access distribution, and on-chip interconnect distribution under the same operating mode. Data at the same location in the frequency distribution, power consumption distribution, memory access distribution, and on-chip interconnect distribution are grouped. Sudden jump values, sampling jitter values, and instantaneous interface glitches are removed from the grouped data at the same location, and valid data is retained. Based on the valid data, the stable representative value and stable fluctuation range of each distribution at the corresponding distribution statistical position are determined. The stable representative values ​​of each distribution at each distribution statistical position are arranged according to the fixed field order and the distribution statistical position order to form a matrix-form steady-state reference surface. When the same operating mode corresponds to multiple time periods, the stable representative values ​​formed by multiple time periods are merged to obtain the final steady-state reference surface corresponding to the operating mode.

[0102] Among them, the distribution statistical position refers to the pre-defined data statistical landing point for the same type of monitoring data under a unified time index. It is used to identify the specific value position of this type of monitoring data in the steady-state aggregation process. The distribution statistical position is arranged according to the time step corresponding to the unified time index. It is used to ensure that the sampling data of each time under the same operating mode are grouped, compared and aggregated in the same position. The fixed field order refers to the pre-defined field arrangement order of temperature data, voltage data, current data, frequency data, power consumption data, storage access data and on-chip interconnect data. It is used to unify the field organization method of operating state sequence, temperature distribution, voltage distribution, current distribution, frequency distribution, power consumption distribution, storage access distribution and on-chip interconnect distribution. The fixed field order remains unchanged throughout the data processing process. It is used to ensure that the data corresponding to different time steps and different operating modes are aligned, written and calculated in the same field position.

[0103] Based on the operating mode identifier corresponding to each time step in the operating state sequence, a steady-state reference surface corresponding to the operating mode is matched for each time step, and temperature data, voltage data, current data, frequency data, power consumption data, memory access data and on-chip interconnect data corresponding to that time step are extracted from the operating state sequence.

[0104] According to the fixed field order and distribution statistical position, the temperature data, voltage data, current data, frequency data, power consumption data, memory access data and on-chip interconnect data extracted at each time step are respectively established with the matched steady-state reference surface to generate the state offset vector corresponding to that time step;

[0105] The state offset vectors corresponding to each time step are arranged continuously according to the time step order under a unified time index to generate a state offset sequence.

[0106] In this embodiment, the model building module includes:

[0107] The basic structure of the original CfC model is invoked. The original CfC model includes an input mapping structure, a continuous-time hidden state update structure, a closed-form state evolution structure, and an output mapping structure. The input mapping structure is used to receive temporal input and complete dimension mapping. The continuous-time hidden state update structure is used to receive the input mapping result, historical hidden states, and time interval data. The closed-form state evolution structure is used to express the state update process in the continuous-time hidden state update structure in a closed form. The output mapping structure is used to read out the output result of the closed-form state evolution structure.

[0108] Based on the input mapping structure of the original CfC model, the mapping layer oriented to general temporal input is replaced with a process mapping layer oriented to the state evolution process sequence. A process coding unit is constructed, and the state evolution process sequence is input into the process mapping layer in the order of time steps corresponding to a unified time index. The connection relationship between the continuous time hidden state update structure and the closed state evolution structure in the original CfC model is retained to form a continuous time coding structure inside the process coding unit.

[0109] Based on the input interface between the input mapping structure of the original CfC model and the continuous-time hidden state update structure, a reference surface alignment unit is constructed. Specifically, the running state sequence and the steady-state reference surface are jointly connected to the front end of the input interface. A correspondence is established according to a fixed field order, a unified time index position, and a distribution statistical position. The data at the corresponding positions are organized into an aligned input structure that is consistent with the input dimension of the continuous-time hidden state update structure. The output end of the aligned input structure is connected to the corresponding input interface of the continuous-time hidden state update structure to complete the construction of the reference surface alignment unit. The running state sequence and the steady-state reference surface are used as the input data of the reference surface alignment unit. The output end of the reference surface alignment unit is connected to the input end of the offset propagation unit. A one-to-one correspondence connection relationship is established between the running state sequence field position, the steady-state reference surface matrix position, and the unified time index position within the reference surface alignment unit.

[0110] Based on the continuous-time hidden state update structure of the original CfC model, an offset propagation unit is constructed. The output of the reference face alignment unit is used as the input of the offset propagation unit. Inside the offset propagation unit, the connection form of the continuous-time hidden state update structure receiving time interval data and performing hidden state recursion is retained. At the same time, the output of the offset propagation unit is set as the offset propagation representation output.

[0111] Based on the closed state evolution structure of the original CfC model, the closed state evolution path for general hidden state update in the original CfC model is transformed into a closed transmission path for closed transmission of state offset. A closed transmission unit is constructed, and the output of the process encoding unit and the output of the offset propagation unit are connected to the input of the closed transmission unit. An access interface for functionally coupled closed transmission topology is reserved in the closed transmission unit to form a joint input structure of process driving quantity, offset propagation representation and functionally coupled closed transmission topology.

[0112] Based on the output mapping structure of the original CfC model, the readout layer oriented to the general prediction result in the output mapping structure is replaced with the judgment mapping layer oriented to the chip operating status monitoring result. A state judgment unit is constructed, and the output end of the closed conduction unit is connected to the input end of the state judgment unit to form a unidirectional output connection structure from the state conduction result to the chip operating status monitoring result.

[0113] According to the connection method of connecting the reference plane alignment unit to the offset propagation unit, the output of the process coding unit to the closed conduction unit, and the output of the offset propagation unit to the state determination unit, the process coding unit, the reference plane alignment unit, the offset propagation unit, the closed conduction unit and the state determination unit are networked together to generate the process offset CfC model.

[0114] During the training of the process offset CfC model, historical chip operating status data, corresponding operating status category records, abnormal offset intensity records, risk propagation path records, and abnormal start position records are collected. These data are then merged according to the same unified time index, the same time period identifier, and the same abnormal event identifier to form a historical training sample set. The historical training sample set undergoes preprocessing consistent with the current chip operating status data to generate a historical operating status sequence. Based on this historical operating status sequence, a historical state evolution process sequence, a historical steady-state reference surface, a historical state offset sequence, and a historical functional coupling closed-loop transmission topology are constructed.

[0115] The process offset input samples corresponding to each time period are extracted from the historical state evolution process sequence, historical steady-state reference surface, historical state offset sequence, and historical functionally coupled closed-loop transmission topology. The corresponding operation state category record, abnormal offset intensity record, risk propagation path record, and abnormal start position record after actual operation of each time period are used as output labels. After quantization, encoding, and normalization of the process offset input samples, they are input into the process offset CfC model. The process encoding unit, reference surface alignment unit, offset propagation unit, closed-loop transmission unit, and state determination unit complete the training calculation in sequence, and output the current operation state category prediction result, abnormal offset intensity prediction result, risk propagation path prediction result, and abnormal start position prediction result.

[0116] The total loss is constructed using the classification error between the current operating status category prediction result and the operating status category record, the numerical deviation between the anomaly offset intensity prediction result and the anomaly offset intensity record, the path position error between the risk propagation path prediction result and the risk propagation path record, and the node position error between the anomaly start position prediction result and the anomaly start position record.

[0117] The Adam optimizer is used to iteratively update the model parameters. The initial learning rate is set to 0.0003, the batch size is set to 16, and the maximum number of training epochs is set to 140. After each training epoch, the validation loss is calculated on the validation sample set, and the current validation loss is compared with the historical minimum validation loss. When the validation loss decreases by less than 0.00005 for 12 consecutive training epochs, the process offset CfC model is considered to have reached the convergence condition. When the number of training epochs reaches 140 and the validation loss does not continue to decrease for 8 consecutive training epochs, the process offset CfC model is also considered to have reached the convergence condition. After training, the model parameters corresponding to the minimum validation loss are saved to obtain the trained process offset CfC model.

[0118] In this embodiment, the offset propagation calculation module includes:

[0119] Based on the state evolution process sequence, the thermal changes, power changes, frequency changes, interconnection changes and execution load changes corresponding to each time period are written into the input position of the process coding unit according to a fixed position. The input sequence between each time period is organized according to the time step order corresponding to the unified time index, and the time interval data between adjacent time periods are synchronously written into the time input position corresponding to the process coding unit.

[0120] Based on the process coding unit, continuous-time recursive coding is performed on the thermal changes, power changes, frequency changes, interconnection changes, and execution load changes input in time step order. The coding results corresponding to each time period are then continuously arranged according to the time step order corresponding to a unified time index to generate process-driven variables. Specifically, the continuous-time recursive coding is performed as follows: First, the thermal changes, power changes, frequency changes, interconnection changes, and execution load changes corresponding to each time period are combined into the current time period input vector at fixed positions. Then, the current time period input vector, the coding state corresponding to the previous time period, and the time interval data corresponding to the current time period are input into the process coding unit. The coding result corresponding to the current time period is obtained according to the continuous-time state update method. For the first time period, the encoding is completed by using a preset initial coding state and the current time period input vector. Finally, the coding results corresponding to each time period are continuously arranged according to the time step order corresponding to a unified time index to generate process-driven variables.

[0121] Based on the running state sequence and the steady-state reference surface, the running mode identifier corresponding to each time step in the running state sequence is matched with the running mode identifier corresponding to the steady-state reference surface. The matched steady-state reference surface position is written into the reference position corresponding to each time step. The task stage boundary position is determined according to the task stage data, and the resource occupation boundary position is determined according to the resource occupation data. The reference surface alignment processing result is generated. The reference position corresponding to each time step refers to the data position in the steady-state reference surface that is consistent with the running mode identifier of the current time step, the position of the current monitoring field in the fixed field order, and the distribution statistics position of the current time step under the unified time index.

[0122] Based on the alignment results of the reference plane, offset correction processing is performed on the state offset sequence. Specifically, the field positions corresponding to each time step in the state offset sequence are corrected according to the reference position, the offset direction and offset magnitude corresponding to each time step are corrected according to the stable representative value of the corresponding position in the steady-state reference plane, and the offset data of cross-boundary connection are segmented and corrected at the task stage boundary position and resource occupation boundary position to generate the corrected state offset sequence.

[0123] Based on the corrected state offset sequence, the corrected state offset sequence is input into the offset propagation unit in the order of time steps corresponding to the unified time index, and the time interval data between adjacent time steps is written into the continuous time recursion position corresponding to the offset propagation unit.

[0124] Based on the offset propagation unit, continuous propagation calculation is performed on the corrected state offset sequence. Specifically, the task stage switching relationship between adjacent time periods is determined based on task scheduling data, the expansion relationship between resource occupancy locations within the same time period is determined based on storage access data, power consumption data, and resource occupancy data, the transmission relationship between link locations is determined based on on-chip interconnect data, and the thermal diffusion relationship between temperature locations is determined based on temperature data. According to the task stage switching relationship, expansion relationship, transmission relationship, and thermal diffusion relationship, the offset in the corrected state offset sequence is transmitted sequentially between adjacent time periods, between resource occupancy locations, between link locations, and between temperature locations. The offset transmitted to the same location is cumulatively updated to generate the propagation result corresponding to each time step.

[0125] Based on the propagation results corresponding to each time step, the time steps are arranged continuously according to the time step order corresponding to the unified time index, and combined into a time step vector sequence according to the fixed field positions to generate the offset propagation representation.

[0126] In this embodiment, the closed-loop conduction calculation module includes:

[0127] Based on task scheduling data, operation log data, frequency data, and power consumption data in the running state sequence, the sequential triggering relationship and load sharing relationship between various computing activities within the chip are determined, generating computing execution relationships; based on memory access data, task scheduling data, and power consumption data in the running state sequence, the memory access initiation position, memory access arrival position, and memory access occupancy change relationship corresponding to each computing activity are determined, generating memory access relationships; based on on-chip interconnect data in the running state sequence, the transmission start position, transmission arrival position, and link occupancy transmission relationship between each link location are determined, generating on-chip interconnect relationships; based on voltage data, current data, power consumption data, and task scheduling data in the running state sequence, the power supply association relationship between each power supply location and each computing activity is determined, generating power supply relationships; based on temperature data, power consumption data, and on-chip interconnect data in the running state sequence, the heat transfer association relationship between each location is determined, generating thermal response relationships;

[0128] A unified location mapping is performed on computation execution relationships, memory access relationships, on-chip interconnect relationships, power supply relationships, and thermal response relationships. The computation activity location, memory access location, link location, power supply location, and temperature location are mapped to functional location nodes, and the start position and arrival position in each relationship are mapped to node connection relationships. Specifically, the computation execution relationship corresponds to a directed connection relationship from the preceding computation activity location to the following computation activity location; the memory access relationship corresponds to a directed connection relationship from the memory access initiation location to the memory access arrival location; the on-chip interconnect relationship corresponds to a directed connection relationship from the link sending location to the link receiving location; the power supply relationship corresponds to a directed connection relationship from the power supply location to the computation activity location; and the thermal response relationship corresponds to a directed connection relationship from the high heat location to the heat-affected location.

[0129] Based on the frequency of occurrence, duration, associated power consumption intensity and temperature response intensity of each node connection relationship, the coupling weight corresponding to each node connection relationship is determined, and the coupling weight is written into the corresponding node connection relationship according to the relationship source. For the case where multiple types of relationship connections exist between nodes at the same functional location, the node connection relationship and coupling weight are retained according to the relationship source.

[0130] Based on functional location nodes, node connection relationships, and coupling weights, a functionally coupled closed-loop transmission topology is constructed. In this topology, the connection relationships of each node retain the transmission direction, and the coupling weights correspond to the transmission strength of the node connection relationships. The functionally coupled closed-loop transmission topology is organized according to a unified node order to form a data structure that can be positionally correlated with process driving variables and offset propagation representations.

[0131] The process-driven variables are written into the node input positions corresponding to the functional position nodes in the order of time steps corresponding to the unified time index. The offset propagation representation is written into the offset input positions corresponding to the functional position nodes in the order of fixed field positions. The functionally coupled closed-loop transmission topology is connected to the topology input position of the closed-loop transmission unit to establish a joint input correspondence between the process-driven variables, the offset propagation representation and the functionally coupled closed-loop transmission topology.

[0132] Based on the closed-loop transmission unit, the closed-loop transmission calculation of state offset is performed on each functional position node in the joint input correspondence. Specifically, according to the transmission direction in the functionally coupled closed-loop transmission topology, the process driving quantity and offset propagation representation corresponding to each functional position node are transmitted to the adjacent functional position nodes along the node connection relationship. The transmitted quantity is weighted and accumulated according to the coupling weight corresponding to the node connection relationship. For a functional position node that receives the transmitted quantity from multiple upstream nodes, the transmitted quantity from multiple upstream nodes is merged according to the corresponding position of the node connection relationship. The node transmission state of the functional position node is updated in combination with the process driving quantity and offset propagation representation corresponding to the current functional position node. The node transmission state of each time step is continuously calculated according to the time step order corresponding to the unified time index to generate the node transmission result corresponding to each time step.

[0133] The node transmission results corresponding to each time step are arranged in the order of time steps corresponding to the unified time index, and combined into state transmission results according to the order of functional position nodes.

[0134] In this embodiment, the status monitoring output module includes:

[0135] The state transmission results are input into the state determination unit. According to the time step order corresponding to the unified time index, the transmission values ​​of each time step, each functional position, and each node connection relationship in the state transmission results are read sequentially and organized into state determination input data according to fixed field positions.

[0136] Based on the state determination unit, a determination mapping process is performed on the state determination input data to map the state determination input data to the category determination result corresponding to the current running state category, and the current running state category is determined according to the maximum determination value in the category determination result.

[0137] Based on the continuous change of the transmission value at each time step in the state determination input data under a unified time index, the cumulative expansion of the transmission value at each functional position under the node connection relationship, and the transmission concentration of the transmission value corresponding to each node connection relationship in the functionally coupled closed transmission topology, the degree of offset accumulation in the state transmission result is extracted to determine the intensity of abnormal offset.

[0138] Based on the state transmission results in the transmission records of the functionally coupled closed transmission topology, the state transmission results are backtracked in reverse according to the transmission sequence corresponding to the node connection relationship, the order of coupling weight, and the continuous expansion order of transmission value. The continuous transmission chain traced from the high transmission value position to the low time index position is extracted to determine the risk propagation path.

[0139] Based on the connection relationship of the starting node, the starting time step position and the starting function position corresponding to the risk propagation path, and combined with the first concentrated position of the abnormal offset intensity in the risk propagation path, the abnormal starting position is determined.

[0140] The current operating status category, abnormal offset intensity, risk propagation path, and abnormal starting position are combined in a fixed field order to generate chip operating status monitoring results.

[0141] Example 1: To verify the feasibility of this invention in practice, it was applied to a scenario of intelligent monitoring of the operating status of a chip with a multi-core computing unit, on-chip interconnect structure, memory control unit, power management unit, and temperature monitoring interface. In this scenario, when the chip continuously executes multi-task loads, its operating status changes continuously with task switching, resource usage changes, on-chip interconnect load migration, power fluctuations, and heat accumulation. Traditional monitoring methods typically rely on threshold judgments based on upper limits of temperature, current, and power consumption, which can detect some obvious anomalies but struggles to identify the gradually accumulating state shifts during operating mode switching. While some monitoring methods based on ordinary timing models can utilize continuous sampling data to identify trends, they lack constraints on steady-state reference benchmarks. When faced with normal fluctuations of the same chip under different operating modes, they easily misjudge differences in operating modes as anomalies, and when faced with the actual propagation process of anomalies, they struggle to accurately identify the starting location of the anomaly and the risk propagation path.

[0142] In practical applications, chip operating status data is synchronously collected from temperature monitoring, power monitoring, clock monitoring, power consumption monitoring, memory control, on-chip interconnect monitoring, task scheduling, and operation log output interfaces. The collected data includes temperature, voltage, current, frequency, power consumption, memory access, on-chip interconnect, task scheduling, and operation log data. After sampling, all sampling timestamps are converted according to a unified time indexing rule, mapping data from different sources to a unified time step, and then organized into windows with fixed time window lengths. After time alignment, cleaning and standardization processes are performed, combining the data corresponding to each time step into a multi-dimensional vector with a fixed field order, ultimately forming the operating status sequence. This processing method aims to organize discrete monitoring quantities collected from multiple interfaces into continuous inputs that can be directly calculated at the same time scale, avoiding time misalignment problems caused by inconsistent interface sampling frequencies.

[0143] Based on the operational state sequence, further extraction of thermal changes, power supply changes, frequency changes, interconnection changes, and execution load changes is performed. These changes are then jointly orchestrated according to the task phase switching order and resource occupancy change order to form a state evolution process sequence. This does not simply record the instantaneous value of a single field, but rather the process of the chip progressing from one operational state to the next. After this processing, the task progression trajectory, resource migration trajectory, and link load transfer trajectory during chip operation are uniformly compressed into a single temporal expression. The monitoring system can distinguish between "numerical changes" and "process changes," thereby reducing the likelihood of misjudging normal fluctuations as abnormalities.

[0144] During the reference plane offset construction phase, operating modes are divided based on load configuration data, resource usage data, frequency configuration data, and task stage data in the operating state sequence, and a steady-state reference plane is constructed for each operating mode. The steady-state reference plane does not represent a single value, but rather the stable statistical results of temperature distribution, voltage distribution, current distribution, frequency distribution, power consumption distribution, memory access distribution, and on-chip interconnect distribution under that operating mode. After establishing the field position correspondence between the operating state sequence and the steady-state reference plane, offset calculations can be performed on the operating state at each time step, thus forming a state offset sequence. This process transforms chip monitoring from directly judging instantaneous sampled values ​​to judging the degree of offset of the current state relative to the steady state of that operating mode. Therefore, it can still maintain high judgment consistency even under frequent task switching and varying load levels.

[0145] During the model execution phase, the state evolution process sequence is input into the process encoding unit to obtain the process driving quantity; the running state sequence and the steady-state reference surface are input into the reference surface alignment unit, and offset correction processing is performed on the state offset sequence to form a corrected state offset sequence; then the corrected state offset sequence is input into the offset propagation unit, and continuous propagation calculation is performed according to the task stage advancement relationship, resource occupation expansion relationship, on-chip interconnect transmission relationship, and thermal response relationship to obtain the offset propagation representation; then, based on the computation execution relationship, memory access relationship, on-chip interconnect relationship, power supply relationship, and thermal response relationship, a functionally coupled closed conduction topology is constructed, and the process driving quantity, offset propagation representation, and functionally coupled closed conduction topology are input together into the closed conduction unit to generate the state conduction result; after the state conduction result is input into the state determination unit, the chip running state monitoring result is output, which includes the current running state category, abnormal offset intensity, risk propagation path, and abnormal starting position.

[0146] To ensure sufficient representativeness of the verification process, this embodiment selected 420 continuously running test samples, including 260 normal operation samples, 78 slightly abnormal samples, 52 moderately abnormal samples, and 30 severely abnormal samples. Each sample includes a complete operating state sequence, a state evolution process sequence, a steady-state reference surface, a state offset sequence, state propagation results, and a manual verification label. The manual verification label is jointly determined by chip debugging records, operation log verification results, power consumption change records, temperature rise change records, and on-chip interconnect congestion records. The comparison methods include traditional threshold monitoring methods and monitoring methods based on ordinary timing models. Traditional threshold monitoring methods perform alarm judgments according to fixed temperature thresholds, fixed current thresholds, and fixed power consumption thresholds; the monitoring method based on ordinary timing models uses the operating state sequence as input, directly outputs the anomaly category, does not construct a steady-state reference surface, does not generate a state offset sequence, and does not construct a functionally coupled closed propagation topology.

[0147] Table 1 Comparison of Chip Operation Status Monitoring Performance

[0148]

[0149] As shown in Table 1, the method of this invention achieves a monitoring accuracy of 91.4%, which is 12.8 percentage points higher than traditional threshold monitoring methods and 5.3 percentage points higher than monitoring methods based on ordinary time-series models. This improvement does not stem from simply increasing model complexity, but rather from a change in the way the monitored object is represented. Traditional threshold monitoring methods tend to mistake short-term normal fluctuations as abnormal signals when faced with power consumption jitter, temperature drop, and frequency switching phenomena during task switching. While monitoring methods based on ordinary time-series models can learn time-varying relationships, they cannot provide benchmark differences between operating modes, thus failing to clearly distinguish the normal state boundaries between different operating modes. This invention constructs a steady-state reference surface before entering model calculations and forms a state offset sequence based on it. The model focuses on the offset process itself, thereby reducing interference caused by normal mode switching.

[0150] The accuracy of anomaly initiation location positioning improved from 52.3% and 73.8% to 88.2%, and the risk propagation path matching rate improved from 41.7% and 68.5% to 86.7%. These two indicators directly demonstrate that this invention not only completes anomaly classification but also provides a more complete anomaly evolution chain. Traditional threshold monitoring methods can only indicate that a certain monitored quantity has exceeded its limit, but cannot determine from which functional location the anomaly started. Although monitoring methods based on ordinary time-series models can output anomaly probabilities, their output results lack corresponding constraints with the coupling relationship between the chip's internal operation, thus limiting their ability to reconstruct propagation paths. This invention integrates computation execution relationships, memory access relationships, on-chip interconnect relationships, power supply relationships, and thermal response relationships into a closed-loop propagation calculation process through a functionally coupled closed-loop topology. The state propagation results naturally possess path continuity and positional correlation, thus more closely resembling the results of manual review in anomaly initiation location positioning and risk propagation path backtracking.

[0151] The false alarm rate and false negative rate demonstrate the practicality of this invention. The false alarm rate of the method of this invention is 8.1%, and the false negative rate is 7.4%. The reason for the decrease in the false alarm rate is that the reference surface alignment processing and offset correction processing can position the sequence of operating states with the steady-state reference surface, limiting normal fluctuations in the same operating mode to the range of steady-state fluctuations, and reducing the occurrence of fluctuations within the mode being regarded as abnormal offsets. The reason for the decrease in the false negative rate is that the offset propagation unit does not only look at the data of a single time step, but also performs continuous propagation calculations in combination with task phase advancement, resource occupation expansion, on-chip interconnect transmission, and thermal response propagation. Therefore, for abnormal states where the single-point offset is not obvious but will gradually expand along the functional link, effective identification can be achieved earlier.

[0152] The average early warning lead time reaches 1.86 seconds, which is 1.14 seconds better than traditional threshold monitoring methods and 0.65 seconds better than monitoring methods based on ordinary time series models. This result demonstrates that the method of this invention can determine the propagation trend of state deviation before an anomaly develops into a significant threshold breach. For chip operation monitoring, consistently obtaining an early warning window of more than 1 second provides sufficient response time for frequency back-off, power consumption constraints, task migration, and thermal management adjustments. Meanwhile, the average single-processing latency of the method of this invention is 46 milliseconds, which, although higher than the two comparative methods, is still within the acceptable range for online monitoring of chip operation status. The slight increase in processing latency is due to the addition of several steps in this invention, including steady-state reference surface construction, offset correction processing, continuous propagation calculation, and closed-loop conduction calculation. However, this increase in latency results in higher positioning accuracy and path interpretation capability, making the overall benefits more significant.

[0153] The above are merely preferred embodiments 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 deep learning-based intelligent monitoring system for chip operating status, characterized in that, include: The runtime data preprocessing module is used to collect chip runtime status data, perform alignment, cleaning, and standardization processing, and generate runtime status sequences. The process sequence construction module is used to construct a state evolution process sequence based on the running state sequence; The reference surface offset construction module is used to construct the steady-state reference surface corresponding to each operating mode based on the operating state sequence, and generate the state offset sequence based on the difference relationship between the steady-state reference surface and the operating state sequence. The model building module is used to build the process offset CfC model, including the process encoding unit, reference plane alignment unit, offset propagation unit, closed propagation unit and state determination unit. The offset propagation calculation module is used to input the state evolution process sequence into the process encoding unit to generate process driving variables, input the running state sequence and steady-state reference surface into the reference surface alignment unit, perform offset correction processing to generate a corrected state offset sequence, input the corrected state offset sequence into the offset propagation unit, perform continuous propagation calculation of state offset, and generate an offset propagation representation. The closed-loop conduction calculation module is used to construct a functionally coupled closed-loop conduction topology based on the internal operational coupling relationship of the chip, and combine process driving variables and offset propagation representation to perform closed-loop conduction calculation of state offset in the closed-loop conduction unit to generate state conduction results. The status monitoring output module is used to generate chip operating status monitoring results by inputting the status transmission results into the status determination unit.

2. The intelligent chip operation status monitoring system based on deep learning according to claim 1, characterized in that, The runtime data preprocessing module includes: Collect chip operating status data, including temperature data, voltage and current data, frequency data, power consumption data, memory access data, on-chip interconnect data, task scheduling data, and operation log data, and record the sampling timestamp; The sampling timestamp is converted into a unified time index, the chip operating status data is mapped to the time step corresponding to the unified time index, and the continuous time steps are organized into windows according to a fixed time window length. Based on a unified sampling alignment rule, time step alignment processing is performed on the chip operating status data mapped to a unified time index; Perform cleaning processing on the data that has completed time step alignment; The cleaned data is standardized, and the chip operating status data is uniformly quantified, numerically mapped and rearranged according to the unified monitoring field rules. The chip operating status data corresponding to each time step is combined into a multi-dimensional vector according to a fixed field order. The multidimensional vectors are continuously arranged according to the time step order under a unified time index to generate a running state sequence.

3. The intelligent chip operating status monitoring system based on deep learning according to claim 1, characterized in that, The process sequence construction module includes: Based on the unified time index corresponding to the running state sequence, the running state sequence is divided into multiple time periods according to continuous time steps; The thermal change is determined based on the cumulative trend and rise / fall conversion of temperature data within each time period; the power supply change is determined based on the combined fluctuation of voltage, current and power consumption data within each time period; and the frequency change is determined based on the frequency switching of frequency data within each time period. The amount of interconnection change is determined based on the changes in link occupancy, transmission load migration, and congestion expansion of on-chip interconnection data within each time period. The amount of execution load change is determined based on the stage switching and resource occupancy migration of task scheduling data, storage access data, and power consumption data within each time period. The task phase switching order and resource usage change order between different time periods are determined based on the running status sequence; According to the task phase switching order and resource usage change order, the thermal change, power supply change, frequency change, interconnection change and execution load change are sequentially arranged and combined into a structured vector. Following the time step order under a unified time index, the structured vectors corresponding to each time period are arranged continuously, and the advancement relationship between adjacent time periods is established according to the task stage switching order and the resource consumption change order, thus generating a state evolution process sequence.

4. The intelligent chip operation status monitoring system based on deep learning according to claim 1, characterized in that, The reference plane offset construction module includes: Based on the running state sequence, load configuration data, resource usage data, frequency configuration data and task stage data are extracted to form a combined state sequence; Based on the combined state sequence, continuous time periods with consistent combined states are divided into the same operating mode, and an operating mode identifier is established for each operating mode. For each operating mode, extract the temperature data, voltage data, current data, frequency data, power consumption data, memory access data, and on-chip interconnect data within the time period corresponding to the operating mode identifier, and form temperature distribution, voltage distribution, current distribution, frequency distribution, power consumption distribution, memory access distribution, and on-chip interconnect distribution respectively. Perform steady-state aggregation and arrange the steady-state aggregation results into a matrix form of steady-state reference surface. Based on the operating mode identifier corresponding to each time step in the operating state sequence, a steady-state reference surface corresponding to the operating mode is matched for each time step; According to the fixed field order and distribution statistical position, for the temperature data, voltage data, current data, frequency data, power consumption data, memory access data and on-chip interconnect data of each time step, a dimension-wise difference relationship is established with the matched steady-state reference surface to generate the state offset vector corresponding to that time step. The state offset vectors corresponding to each time step are arranged continuously according to the time step order under a unified time index to generate a state offset sequence.

5. The intelligent chip operating status monitoring system based on deep learning according to claim 1, characterized in that, The model building module includes: The basic structure of the original CfC model is invoked, including the input mapping structure, the continuous-time hidden state update structure, the closed-form state evolution structure, and the output mapping structure. The process coding unit is constructed by replacing the mapping layer in the input mapping structure that is oriented towards general temporal inputs with a process mapping layer that is oriented towards state evolution process sequences. Based on the input interface between the input mapping structure of the original CfC model and the continuous-time hidden state update structure, a reference face alignment unit is constructed. Based on the continuous-time hidden state update structure, an offset propagation unit is constructed. Based on the closed state evolution structure, the closed state evolution path in the original CfC model that is oriented towards the general hidden state update is transformed into a closed propagation path that is oriented towards the closed propagation of state offset, and a closed propagation unit is constructed. The readout layer in the output mapping structure that is oriented towards general prediction results is replaced with a decision mapping layer that is oriented towards chip operating status monitoring results, and a status decision unit is constructed. The process encoding unit, reference plane alignment unit, offset propagation unit, closed-loop transmission unit, and state determination unit are networked together to generate the process offset CfC model.

6. The intelligent chip operating status monitoring system based on deep learning according to claim 1, characterized in that, The offset propagation calculation module includes: Based on the state evolution process sequence, the thermal change, power change, frequency change, interconnection change and execution load change corresponding to each time period are written into the process coding unit at fixed positions, and continuous time recursive coding is performed to generate process driving quantities. Match the running mode identifier corresponding to each time step in the running state sequence with the corresponding running mode identifier in the steady-state reference surface to determine the task stage boundary position and resource occupation boundary position, and generate the reference surface alignment processing result. Based on the reference plane alignment processing results, offset correction processing is performed on the state offset sequence. The offset direction and offset magnitude corresponding to each time step are corrected by amplitude correction. At the task stage boundary position and resource occupation boundary position, the offset data of cross-boundary connection is segmented and corrected to generate the corrected state offset sequence. Based on the corrected state offset sequence, the offset propagation unit is input, and continuous propagation calculation is performed to generate the propagation results corresponding to each time step. Based on the propagation results corresponding to each time step, the time steps are arranged continuously according to the time step order corresponding to the unified time index, and combined into a time step vector sequence according to the fixed field positions to generate the offset propagation representation.

7. The intelligent chip operating status monitoring system based on deep learning according to claim 1, characterized in that, The closed-loop conduction calculation module includes: Based on the sequence of operating states, the computation execution relationships, memory access relationships, on-chip interconnection relationships, power supply relationships, and thermal response relationships within the chip are determined. A unified location mapping is performed on computation execution relationships, memory access relationships, on-chip interconnect relationships, power supply relationships, and thermal response relationships. The computation activity location, memory access location, link location, power supply location, and temperature location are mapped to functional location nodes, and the start location and destination location in each relationship are mapped to node connection relationships. Based on the frequency of occurrence, duration, associated power consumption intensity, and temperature response intensity of each node connection relationship, the coupling weight corresponding to each node connection relationship is determined. Construct a functionally coupled closed-loop transmission topology based on functional location nodes, node connection relationships, and coupling weights; Write the process-driven variable into the node input position corresponding to the functional location node, write the offset propagation representation into the offset input position corresponding to the functional location node, and connect the functionally coupled closed-loop transmission topology to the topology input position of the closed-loop transmission unit to establish a joint input correspondence. Based on the closed-loop transmission unit, the closed-loop transmission calculation of state offset is performed on each functional position node in the joint input correspondence to generate the node transmission result corresponding to each time step; The node transmission results corresponding to each time step are arranged in the order of time steps corresponding to the unified time index, and combined into state transmission results according to the order of functional position nodes.

8. The intelligent chip operating status monitoring system based on deep learning according to claim 1, characterized in that, The status monitoring output module includes: The state transmission results are input into the state determination unit and organized into state determination input data. Based on the state determination unit, a determination mapping process is performed on the state determination input data to determine the current running state category; Based on the continuous change of the transmission value at each time step in the state determination input data under a unified time index, the cumulative expansion of the transmission value at each functional position under the node connection relationship, and the transmission concentration of the transmission value corresponding to each node connection relationship in the functionally coupled closed transmission topology, the intensity of abnormal offset is determined. Based on the transmission record of state transmission results in the functionally coupled closed transmission topology, the state transmission results are backtracked to determine the risk propagation path. Based on the connection relationship of the starting node, the starting time step position and the starting function position corresponding to the risk propagation path, and combined with the first concentrated position of the abnormal offset intensity in the risk propagation path, the abnormal starting position is determined. The current operating status category, abnormal offset intensity, risk propagation path, and abnormal starting position are combined in a fixed field order to generate chip operating status monitoring results.