IDC machine room fault point prediction method and system based on artificial intelligence and internet of things

Abstract

The application discloses an IDC machine room fault point prediction method and system based on artificial intelligence and the Internet of Things, relates to the technical field of computers, and aims at the problems of the existing IDC machine room in the aspects of heat dissipation and air flow organization management, such as the inability to identify implicit risks before temperature overrun, passive fault response, and the lack of fault cause explanation capability, etc.The application analyzes and standardizes PDU, UPS, air conditioners, temperature and humidity probes, BMC and access control data through Internet of Things data access, and builds a unified time axis to carry out quality labeling and credible normal segment screening;based on the cabinet topology and air supply partition atlas, the effective air volume and theoretical temperature rise are estimated, and the cabinet-level air supply temperature field is solved under the constraints of anchoring and spatial smoothing, and the short-circuit return air ratio and vertical hot spot height section are further inverted;the early prediction, accurate positioning and explainable output of the fault point are realized, and the stability and operation and maintenance efficiency of the machine room are improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically, to a method and system for predicting fault points in IDC (Internet Data Center) data centers based on artificial intelligence and the Internet of Things. Background Technology

[0002] With the rapid development of technologies such as cloud computing and big data, the scale of large Internet Data Centers (IDCs) is growing larger and larger, with server racks being densely deployed. In this high-density environment, the internal equipment operates under continuous high load. Any abnormal heat dissipation or local hotspots may cause the equipment to overheat, leading to downtime, which seriously affects business continuity and causes huge economic losses.

[0003] Existing monitoring systems in IDC (Internet Data Center) data centers typically employ alarm mechanisms based on a single temperature threshold. For example, when a temperature probe inside a server rack detects a temperature exceeding a preset upper limit, the system triggers an alarm. However, this passive monitoring approach has significant limitations. First, a single temperature threshold cannot effectively identify hidden risks such as airflow short circuits or the accumulation of localized hotspots. An airflow short circuit refers to hot air exhausted from the server rack not being fully recovered by the air conditioner but instead being drawn back in at the front of the rack, leading to increased intake air temperature and decreased heat dissipation efficiency. This phenomenon may occur before the temperature reaches the alarm threshold and continue to worsen. Second, localized hotspots may only affect a few devices or a specific height section within the rack, and the overall temperature probe may not be able to detect such localized anomalies in time, resulting in a passive fault response and an inability to intervene in the early stages of the problem. Third, existing systems often lack the ability to deeply analyze and explain the causes of faults. When an alarm occurs, maintenance personnel find it difficult to quickly locate the fault point and take effective measures, increasing the time and difficulty of troubleshooting.

[0004] Therefore, existing technologies for heat dissipation and airflow organization management in IDC data centers have technical shortcomings such as the inability to identify hidden risks before temperatures exceed limits, passive fault response, and lack of ability to explain fault causes. There is an urgent need for an intelligent monitoring method and system that can predict fault points in advance and provide explanations and suggestions. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for predicting IDC data center fault points based on artificial intelligence and the Internet of Things, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for predicting fault points in IDC data centers based on artificial intelligence and the Internet of Things includes the following steps;

[0008] Collect multi-source operational data from the IDC data center and generate structured data records; perform time alignment on the structured data records and generate data quality labels and alignment reliability markers; obtain the rack adjacency relationship and air supply zoning relationship to construct a rack topology diagram;

[0009] Estimate the rack-level supply air temperature field on the rack topology diagram; the estimation simultaneously satisfies: selecting anchor racks with normal heat dissipation and reliable data so that their inlet air temperature is close to their supply air temperature, and applying smoothing constraints to the supply air temperature of adjacent racks or racks in the same supply air zone; calculate the theoretical temperature rise based on rack power and effective air volume, and combine the rack inlet air temperature and the supply air temperature field to invert the short-circuit return air ratio according to the energy balance relationship;

[0010] For each of the preset risk physical assumptions, self-consistent residuals are calculated. The weights of the residual terms of the self-consistent residuals are gated by the data quality labeling and alignment confidence mark. The assumption with the smallest self-consistent residual is selected to determine the risk type and failure point of the target cabinet. The risk indicators are predicted for future time windows and the failure point prediction results containing the target cabinet, risk type and prediction lead time are output.

[0011] In a preferred embodiment, the structured data record includes at least: asset identifier, location identifier, reporting time, collection time, record type, numerical value or event body, source protocol, and quality mark.

[0012] In a preferred embodiment, the time alignment includes: obtaining a typical delay range by statistically analyzing the arrival intervals for each location; estimating the time by combining the acquisition time and the typical delay when the reported time is missing or abnormal, and mapping it to a unified time axis; the data quality labeling includes at least complete missing data, intermittent missing data, long-term constant values, and out-of-physical range, and the alignment confidence labeling includes at least high confidence and low confidence.

[0013] In a preferred embodiment, the method further includes screening for reliable normal segments, wherein the reliable normal segments simultaneously satisfy at least two of the following conditions: the access control status is stable, the air intake temperature of most cabinets in the air supply area shows a consistent trend and the magnitude is less than a threshold, and the cabinet power change is within a stable range.

[0014] In a preferred embodiment, the effective air volume is obtained through a monotonic mapping relationship between server fan speed and effective heat dissipation air volume, and the monotonic mapping relationship is calibrated based on power change excitation within the trusted normal segment; when calibration fails or calibration conditions are not met, the effective air volume is obtained by backtracking using a default curve or a general curve.

[0015] In a preferred embodiment, the supply air temperature field estimation is achieved by constructing and solving an optimization problem, the objective function of which includes: a weighted residual term between the inlet air temperature and the supply air temperature and a smoothing regularization term defined by the edge set of the cabinet topology graph; wherein the edge set consists of at least one of the adjacency relationship and the same supply air zone relationship or their union, and when the solution fails, it reverts to the average supply air temperature of the supply air zone.

[0016] In a preferred embodiment, an identifiability determination is performed before inverting the short-circuit return air ratio. The identifiability determination includes at least the following: the theoretical temperature rise is not less than the minimum temperature rise threshold, the difference between the inlet air temperature and the supply air temperature is not less than the minimum temperature difference threshold, and the data quality and alignment reliability of the relevant points are not low. Physical boundary constraints are applied to the inversion results, negative values ​​are truncated to 0, and results exceeding 1 are marked as inconsistent and trigger data verification.

[0017] In a preferred embodiment, the method further includes inversion of the vertical hot spot height segment of the cabinet: the cabinet is divided into multiple height segments, a temperature baseline is constructed based on the air supply temperature field and theoretical temperature rise, and a non-negative constraint least squares problem is solved based on the observation deviation of multiple vertical temperature measurement points to obtain the hot spot intensity of each height segment, thereby refining the fault point from the cabinet level to the height segment within the cabinet.

[0018] In a preferred embodiment, the future time window prediction is performed on at least two of the following: short-circuit return air ratio, hot spot intensity, or theoretical temperature rise. A safety threshold is set based on the data center level or equipment temperature resistance level. When the prediction exceeds the safety threshold within a preset lead time, a high-risk prediction result is output.

[0019] An IDC (Internet Data Center) fault location prediction system based on artificial intelligence and the Internet of Things includes:

[0020] The IoT data access module is used to collect multi-source operational data and generate structured data records;

[0021] The time alignment and quality labeling module is used for time alignment and to generate data quality labels and alignment confidence markers.

[0022] The rack topology and spatial map module is used to construct rack topology maps;

[0023] The thermal risk state estimation module is used to estimate the cabinet-level supply air temperature field that satisfies anchoring and smoothing constraints on the topology map, and obtain the theoretical temperature rise and the short-circuit return air ratio based on power and effective air volume.

[0024] The fault point prediction and interpretation module is used to calculate self-consistent residuals for preset risk physical assumptions, and to perform weight gating on residual terms based on data quality labeling and alignment confidence marking. The assumption with the smallest self-consistent residual is selected to determine the risk type and fault point, and the prediction lead is output for future time window prediction.

[0025] The technical effects and advantages of the IDC (Internet Data Center) data center fault point prediction method and system based on artificial intelligence and the Internet of Things are as follows:

[0026] This invention performs in-depth analysis of multi-source heterogeneous data to invert hidden risk indicators such as short-circuit return air ratio and hot spot height segment. Combined with time series prediction models, it can predict potential risks hours or even days before a failure occurs, giving maintenance personnel ample time to intervene. This avoids the drawback of traditional monitoring systems that can only passively alarm after the temperature exceeds the limit, and significantly improves the operational stability of IDC data centers.

[0027] Meanwhile, the present invention adopts a risk discrimination mechanism based on verifiable hypotheses, which can clearly identify the dominant type of current risk, and through the inversion of the vertical hot spot height segment of the cabinet, it refines the fault location from the cabinet to the degree of comprehensive judgment of the cabinet and height, which greatly improves the accuracy of fault location and operation and maintenance efficiency.

[0028] Furthermore, this invention introduces a strict time alignment and quality labeling mechanism, which can filter out reliable and normal segments for model training and calibration, ensuring the reliability of the input data.

[0029] Finally, this invention can integrate data from various devices such as power distribution units, UPS, environmental probes, server BMC, switches, and access control systems. Through multi-dimensional data fusion analysis, it can construct a comprehensive profile of the rack's thermal risk, overcoming the limitations of insufficient information from a single data source. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the IDC data center fault point prediction system based on artificial intelligence and the Internet of Things according to the present invention;

[0031] Figure 2 This is a flowchart of the IDC data center fault point prediction method based on artificial intelligence and the Internet of Things according to the present invention;

[0032] Figure 3 This is a timing diagram of the IDC (Internet Data Center) fault point prediction method based on artificial intelligence and the Internet of Things, as presented in this invention. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] This invention provides a method and system for predicting fault points in IDC (Internet Data Center) data centers. By integrating IoT data collection, artificial intelligence analysis, and closed-loop calibration mechanisms, it aims to predict, accurately locate, and explain the causes of potential fault points in IDC data centers, thereby improving the stability and operational efficiency of the data center.

[0035] Example 1, see Figure 1 One embodiment of the present invention provides an IDC (Internet Data Center) data center fault prediction system, which includes an Internet of Things (IoT) data access unit 100, a time alignment and quality labeling unit 200, a cabinet topology and spatial mapping unit 300, a cabinet thermal risk status estimation unit 400, a fault prediction and interpretation unit 500, and a closed-loop calibration unit 600. These units work together to form a complete prediction process from data acquisition, status estimation, risk prediction to feedback calibration.

[0036] Specifically, the present invention further describes in detail each unit included in the system.

[0037] The IoT data access unit 100 is responsible for establishing communication connections with various physical devices in the data center and collecting their operational data. This unit collects data from multiple devices through standard or proprietary protocols, including but not limited to:

[0038] Rack power distribution unit (PDU): Collects real-time current, voltage and power data;

[0039] Uninterruptible power supply (UPS) equipment: collects input and output voltage, frequency, and load rate data;

[0040] Environmental probes: Deployed in cold aisles, hot aisles, and front and rear doors of server racks to collect temperature and humidity data;

[0041] Computer room air conditioning equipment: collects data on supply air temperature, return air temperature, fan speed, and operating mode;

[0042] Server Baseboard Management Controller (BMC): Collects CPU temperature, fan speed, power consumption, and alarm events;

[0043] Core and access switch equipment: collects port traffic and packet loss rate data;

[0044] Access control system: Collects cabinet door opening and closing events.

[0045] All collected raw data is typically asynchronous and heterogeneous in format; the network data access unit 100 has a built-in multi-protocol parsing engine 110 that decodes the received data packets and extracts valid information fields; subsequently, the standardization processing module 120 is executed to uniformly convert data from different sources and in different formats into structured data records containing the following core fields:

[0046] Asset identifier: Used to uniquely identify a physical device, such as a server or air conditioner;

[0047] Point identifier: Used to uniquely identify a specific data measurement point, such as the temperature of the first CPU of a server, or the temperature in the middle of the front door of a server rack;

[0048] Reporting time: refers to the timestamp of the event recorded by the data source device itself;

[0049] Collection time: refers to the timestamp when this system receives this data;

[0050] Record type: Distinguish between periodically collected numerical data and event-triggered alarm data;

[0051] Numerical or event-based records: Numerical records store the measured value and unit; event-based records store the event code and descriptive text.

[0052] Source protocol: indicates the communication protocol from which the data originates;

[0053] Quality marking: Initial marking to check whether the data is complete and whether the format is correct.

[0054] To support the calculation of the observed temperature rise of the server rack, the temperature points in this embodiment include at least the temperature points on the rack's air intake side and the temperature points on the rack's air exhaust side. The temperature points on the rack's air intake side can be temperature probes or cold aisle probes at the top, middle, and bottom heights of the rack's front door. The temperature points on the rack's air exhaust side can be temperature probes, hot aisle probes, or server exhaust side temperature points (if provided by the equipment). The system distinguishes between the air intake and exhaust side measurement points in the structured data record by using point identifiers. When the exhaust side temperature point is missing, has a long-term constant value, or is determined to be outside the physical range, the system marks the observed temperature rise at that moment as unusable and does not use the data at that moment in the calculation of the supply air temperature anchoring rack selection and mapping calibration residual.

[0055] After being processed by the IoT data access unit 100, the massive and messy raw IoT data is transformed into a structured data stream with a unified format and clear semantics, laying the foundation for subsequent in-depth analysis.

[0056] The time alignment and quality labeling unit 200 is used to receive a structured data record stream from the Internet of Things data access unit 100.

[0057] Its primary task is to construct a unified timeline. Due to differences in acquisition cycles, network latency, and clock synchronization errors among different devices, data generated at the same physical moment may arrive at the system at different times. The timeline construction module 210 of this unit maintains a time series buffer for each monitoring point. For periodically acquired points, such as temperature probes, the system counts the arrival time intervals of data packets in the most recent period and calculates the typical delay range. When performing time alignment, the reporting time in the data record is used as the benchmark. If the reporting time is missing or obviously abnormal, the median of the acquisition time minus the typical delay is used for estimation, and the data points are aligned to the nearest standard time grid. For event-type points, such as access control switch events, the precise timestamp inherent in the event body is used for alignment first. Through the above methods, the observation data of all points are mapped onto a unified timeline with equal or unequal intervals, forming an aligned unified timeline sequence.

[0058] After time alignment is completed, the data quality assessment module 220 performs rigorous data quality assessment and annotation. This unit determines the status of each point point by point:

[0059] Completely missing data: No data was received within the expected time window;

[0060] Intermittent missing data: Data is sometimes missing and sometimes not, with gaps in the data.

[0061] Long-term constant value: Data remains unchanged for an extended period, exceeding the normal fluctuation range of the equipment;

[0062] Beyond the physical range: Data values ​​that clearly do not conform to common sense in physics, such as negative temperatures or temperatures exceeding 200 degrees Celsius.

[0063] For the identified missing or defective pixels, the system further analyzes and labels the reasons for the missing test, mainly classifying them into three categories:

[0064] Category 1: Device or gateway is offline, usually accompanied by loss of network heartbeat packets;

[0065] The second category is: data collection timeout or network packet loss, which manifests as a single or a few failed data collection attempts, but the device status is normal.

[0066] The third category: location faults or abnormal sensor values, which manifest as continuous abnormal data but normal communication links.

[0067] These quality labeling details will serve as an important basis for subsequent calculations of credibility weights.

[0068] In order to obtain high-quality training and calibration data, the trusted normal segment screening module 230 performs trusted normal segment screening; trusted normal segments refer to the time period when the computer room is relatively stable and there are few interference factors; the screening is based on a comprehensive judgment of multiple dimensions, including but not limited to at least two of the following categories.

[0069] In one optional implementation, four categories are listed in this embodiment:

[0070] Condition 1: The access control system records show that the cabinet door is closed, and there are no frequent door opening events in the area, excluding instantaneous airflow disturbances introduced by human operation;

[0071] Condition 2: Within the same air supply zone, the data from the air intake temperature probes of most cabinets show a consistent trend and a small amplitude, indicating that the air supply environment of the zone is stable.

[0072] Condition 3: The real-time power data of the server rack changes stably, with no instances of servers being powered on or off in batches or drastic load fluctuations.

[0073] Condition 4: The readings of the three height temperature probes at the top, middle, and bottom of the front door of the cabinet are within the common range of historical statistics, indicating that the airflow distribution inside the cabinet is relatively uniform.

[0074] The system scans a unified timeline sequence, marks continuous time periods that simultaneously meet multiple screening criteria, and outputs them as a set of reliable normal segments. The data in these segments are of high quality and have clear physical relationships, making them the preferred data for subsequent parameter estimation and model calibration.

[0075] After completing time alignment, the system generates an alignment confidence label for each point, which is used for subsequent gating and weight settings. In an optional implementation, the system establishes a typical delay distribution based on the arrival interval statistics of the point over a recent period and generates an alignment error index to measure the deviation of the difference between the acquisition time and the reporting time from the median of the typical delay. When there is a missing reporting time, a jump in reporting time, an alignment error index that continuously exceeds a preset threshold, or a significant drift in the point's delay distribution, the alignment confidence of that point within the corresponding time window is marked as low confidence. When the alignment error index is within an acceptable range but has a slight drift, it is marked as medium confidence. When the alignment error is stable and the alignment error index is less than a preset threshold, it is marked as high confidence. This alignment confidence label is used for the identifiability determination of the short-circuit return air ratio inversion and for weight gating of each residual term.

[0076] Rack topology and spatial mapping unit 300 is a static or semi-static knowledge base that stores and maintains a digital model of the data center's physical layout; this model contains at least three layers of relationships:

[0077] Rack adjacency: Record which racks are physically adjacent to each other (left and right) or front and back;

[0078] Cold aisle relationships: Identify which closed or open cold aisle each rack belongs to;

[0079] Air supply zoning relationship: Define which computer room air conditioning unit or group is responsible for supplying air to each cabinet, and the corresponding relationship of the air conditioning outlets.

[0080] It should be noted that, in this embodiment, the air supply area and the air supply sub-area refer to the same air supply zone or its manageable subset under the air supply zone relationship. Unless otherwise stated, they are collectively referred to as air supply zones.

[0081] This spatial topology information is stored in the form of a graph structure or a relational table, providing basic data support for spatial analysis and coupled calculation in the cabinet thermal risk status estimation unit 400; for example, when performing spatial estimation of supply air temperature, it is necessary to utilize the relationship between supply air zones; when analyzing risk transmission, it is necessary to utilize the adjacency relationship.

[0082] The rack thermal risk state estimation unit 400 takes the aligned unified time axis sequence and rack topology and spatial map as input, performs a series of physical principle-based state inversion and estimation calculations, and outputs a sequence of key state variables characterizing the thermal risk state of each rack. Specifically, this includes:

[0083] The effective airflow estimation module 410 is used to perform effective airflow estimation of the server rack. Since directly measuring the intake airflow of each server is costly and difficult to implement, this invention indirectly estimates the effective airflow by establishing a monotonic mapping relationship between server fan speed and effective heat dissipation airflow. The parameter calibration of this monotonic mapping is only performed in a set of reliable normal segments and when the following conditions are met: the access control system records and shows that the rack door is stably closed; the real-time power change of the rack meets the preset excitation conditions, such as the power change amplitude exceeding a threshold to ensure that the fan speed has a significant response; the reading difference of the three height temperature probes at the top, middle and bottom of the front door of the rack is within the common range of historical statistics, indicating that the airflow distribution inside the rack is relatively uniform; and the temperature points on the intake side and exhaust side are determined to be usable during this period to support residual evaluation.

[0084] Among the reliable normal segments that meet the above conditions, the system filters out the periods in which both fan speed and overall power consumption change significantly. Based on the principle of heat conservation, under good heat dissipation conditions, an increase in power consumption requires more airflow to remove heat, so an increase in fan speed should correspond to an increase in airflow. The system constructs a mapping function from fan speed percentage to estimated airflow according to the server model. This function satisfies strict monotonicity constraints and allows the use of piecewise linear form. The construction process can be regarded as a constrained optimization: at the data points of the reliable normal segments, the theoretical consistency residual formed by the estimated airflow, power consumption, and intake and exhaust temperature difference is minimized, while the mapping function is forced to remain monotonically constant.

[0085] When parameter calibration fails or the current data does not meet the calibration conditions, such as insufficient excitation leading to non-convergence, the mapping output is marked as unavailable and an alternative process is triggered, such as degenerating to the server model's default airflow-speed curve or a common mapping curve for the same type of server.

[0086] For a server rack, the estimated airflow of all online servers inside is summed to obtain the effective airflow estimate of the rack at the current moment. When the server model is unknown, the fan speed is missing, or the server is in a standby low-power mode, resulting in greater uncertainty, the system generates confidence labels such as high confidence, medium confidence, and low confidence for subsequent weight gating.

[0087] The theoretical temperature rise estimation module 420 is used to perform theoretical temperature rise estimation. According to the principle of heat conservation, the temperature of air rises after absorbing heat as it flows through the cabinet. The theoretical temperature rise is determined by the relationship between the real-time power of the cabinet, the air mass flow rate, and the specific heat of air at constant pressure. The air mass flow rate is obtained from the effective air volume and air density, and the air density can be corrected by combining the temperature and humidity of the computer room. This theoretical temperature rise reflects the expected temperature rise of the exhaust air relative to the intake air under ideal and uniform heat dissipation conditions. It is used to compare with the observed temperature rise, that is, the exhaust side temperature minus the intake side temperature, to identify abnormal heat dissipation efficiency. If the exhaust side temperature point is unavailable at that moment, the observed temperature rise is marked as unavailable and is not included in the relevant criterion calculation.

[0088] The air supply temperature spatial estimation module 430 is used to perform rack-level spatial estimation of the air supply temperature. The air supply temperature of the computer room air conditioner is usually measured at the air conditioner outlet, but the airflow may mix and be conducted when it is delivered to the rack through the floor plenum and perforated floor, resulting in spatial distribution differences. Obtaining the actual air supply temperature in front of the rack is crucial for accurate analysis. This invention uses a physical constraint-based optimization method for estimation. The known observation data include: air conditioner air supply temperature, air inlet temperature of each rack, and rack spatial topology.

[0089] The estimation problem is modeled as solving for a set of rack-level supply air temperature estimates that simultaneously satisfy two physical constraints:

[0090] Constraint 1: Normal Anchoring; For cabinets that are judged to have normal heat dissipation in the trusted normal segment, and whose inlet air temperature data quality is marked as high or medium confidence, their inlet air temperature should be very close to their supply air temperature, i.e., less than the corresponding preset threshold; the criteria for normal heat dissipation may include: the difference between the theoretical temperature rise and the observed temperature rise is within the preset threshold, the access control is stably closed, and the observed temperature rise is available at that time; cabinets that meet the conditions are selected as anchor cabinets.

[0091] It should be noted that, in this specific embodiment, the inlet air temperature refers to the real-time temperature of the plane where the air inlet port of the anchor cabinet is located, and the supply air temperature refers to the real-time detected temperature at the center of the outlet section of the main air supply duct of the air supply zone to which the anchor cabinet belongs. The difference between the two is the temperature deviation after taking the absolute value, which is used to eliminate the detection direction error of the inlet air temperature and the supply air temperature. The corresponding preset threshold is the maximum deviation range set for screening anchor cabinets with normal heat dissipation, and the value range is 1℃~2℃ (preferably 2℃). This value is determined with reference to the cold aisle temperature control requirements in GB50174-2017 Data Center Design Specification and combined with the airflow stability test of the IDC computer room air supply system. When this absolute value difference is not greater than the threshold, the inlet air temperature of the anchor cabinet can be approximately equivalent to the actual supply air temperature, and can be used as a reliable reference anchor point for the cabinet-level supply air temperature field.

[0092] Constraint 2: Spatial smoothness; cabinets that are physically adjacent or belong to the same air supply zone should have their air supply temperature change smoothly and should not have drastic jumps.

[0093] The system's optimization objective is to minimize the weighted sum of squared residuals between the observed and estimated air supply temperatures of all racks, and to incorporate a regularization term to penalize excessive differences between the estimated air supply temperatures of adjacent racks. The edge set of the graph structure is derived from the union of the rack topology and the adjacency relationships and air supply zone relationships provided by the spatial graph unit 300. The weights are jointly gated by data quality labeling, effective airflow estimation confidence, and alignment confidence labeling: for example, if the air supply temperature is labeled as completely missing or outside the physical range, the weight is zero; if the effective airflow estimation has low confidence, the weight is reduced; if the alignment confidence is low, the weight of that point is reduced or removed from the corresponding window.

[0094] This optimization problem is typically a convex quadratic optimization or a sparse linear least squares problem, which can be solved using the conjugate gradient method or a sparse linear equation solver, with an upper limit for iteration and a convergence stopping criterion set. If the solution fails, i.e. it does not converge or the result does not conform to physical common sense, a fallback strategy is triggered, such as degenerating to use the average air supply temperature of the air supply zone to which the cabinet belongs as the estimate. By solving, the estimated air supply temperature sequence in front of each cabinet can be obtained, thereby constructing a refined air supply temperature field.

[0095] The short-circuit return air ratio inversion module 440 is used to perform short-circuit return air ratio inversion; airflow short-circuit refers to the hot air exhausted from the cabinet not being completely recovered by the air conditioner, but instead detouring and entering the front of the cabinet again to be sucked in, resulting in increased intake air temperature and decreased heat dissipation efficiency; the short-circuit return air ratio is used to quantify the degree of this defect.

[0096] This embodiment uses a testable physical relationship of hybrid conservation for inversion, and its derivation process is as follows:

[0097] First, the intake airflow of the server rack is considered to be a mixture of two parts: one part is cold air from the air supply zone, the temperature of which is the estimated value of the supply air temperature in front of the rack; the other part is hot air from the short-circuit return, the temperature of which can be approximated as the rack exhaust air temperature. The short-circuit return air ratio can be understood as the proportion of hot return air in the mixture. Second, according to the law of conservation of heat, the rack exhaust air temperature can be considered as the result of the intake air temperature plus the theoretical temperature rise. By combining the relationship between the intake air temperature as a proportional mixture of cold air and hot return air and the relationship between the exhaust air temperature as the intake air temperature plus the theoretical temperature rise, the correspondence between the short-circuit return air ratio and the following two types of quantities can be obtained:

[0098] One type of quantity is the deviation of the intake air temperature from the supply air temperature, which is used to describe the degree to which the intake air of the cabinet is raised.

[0099] Another type of quantity is the sum of the deviation and the theoretical temperature rise, which is used to describe the interpretable space of the hot return air on the mixing rise.

[0100] By rearranging and eliminating terms in the simultaneous equations, the short-circuit return air ratio can be expressed as the ratio of the deviation to the sum of the deviation and the theoretical temperature rise, thus allowing the numerical value of the short-circuit return air ratio to be obtained at each time point.

[0101] Before performing the short-circuit return air ratio inversion, the system first checks the quality and identifiability of the relevant input data. If any of the following conditions are met, the short-circuit return air ratio at that moment is marked as unidentifiable, the inversion is not performed, and the data quality verification process is triggered:

[0102] If the theoretical temperature rise is less than the preset minimum temperature rise threshold, such as 0.5 degrees Celsius, it indicates that the rack load is too low or the heat dissipation is too good. The short circuit ratio is highly sensitive to noise.

[0103] The estimated values ​​of the cabinet's inlet or supply air temperature have quality issues such as missing measurements, long-term constant values, or values ​​exceeding the physical range.

[0104] The alignment confidence level of the cabinet inlet air temperature or supply air temperature estimate is marked as low confidence.

[0105] If the difference between the inlet air temperature and the supply air temperature is less than the preset minimum temperature difference threshold, it indicates that the deviation is too small and the inversion results are sensitive to measurement noise.

[0106] For identifiable data, the system applies physical boundary constraints to the inversion results: the short-circuit return air ratio should be non-negative; if the inversion yields a negative value, it is truncated to zero and recorded as no short circuit; the short-circuit return air ratio should not be greater than one; if the inversion result exceeds one, it is marked as inconsistent, indicating possible data quality issues, distorted theoretical temperature rise estimation, or inapplicability of the local physical model, and triggering the data verification and rollback process; finally, the system outputs the time series of the short-circuit return air ratio for each cabinet and its identifiable / inconsistent markings.

[0107] The vertical hotspot height segment inversion module 450 is used to perform vertical hotspot height segment inversion in the cabinet. Since the equipment inside the cabinet does not heat up uniformly, there may be local high-power devices that form vertical hotspots. It is difficult to accurately locate them with only a few temperature probes. This invention divides the vertical of the cabinet into several height segments and treats the contribution intensity of the hotspot in each height segment to the temperature probe as an unknown quantity.

[0108] First, based on the estimated cabinet air supply temperature and theoretical temperature rise, the baseline temperature distribution of each probe is constructed according to the probe installation location. Then, the deviation between the probe observation temperature and the baseline temperature is calculated, and this deviation is interpreted as the superposition of hot spot contributions at different heights. The influence of each hot spot at each height on each probe can be described by a heat transfer model or an empirical influence matrix. The system finds a set of hot spot intensities by solving the least squares fitting under non-negative constraints, such that the overall error between the prediction deviation and the observation deviation obtained by superimposing these intensities is minimized. The height segment with the largest value in the solved hot spot intensity vector is located as the most likely hot spot area, i.e., the candidate location of the fault point, thus refining the location from the cabinet level to the cabinet and height segment joint decision-making level.

[0109] The fault point prediction and interpretation unit 500 receives key state quantity sequences from the cabinet thermal risk state estimation unit 400, including the short-circuit return air ratio, hot spot height segment inversion results, theoretical temperature rise, and estimated supply air temperature, and combines them with contextual information such as access control events. This unit is used to comprehensively determine the risk type and perform short-term predictions; specifically, it includes:

[0110] The risk type identification module 510 employs a mechanism based on testable hypotheses; the system pre-sets at least four testable physical hypotheses to explain the abnormal state, namely:

[0111] Assume H1 (short-circuit return air is the main type) 511: The proportion of short-circuit return air is consistently high, and the inlet air temperature is higher than that of normal cabinets in the same air supply zone.

[0112] Assume H2 (access control disturbance is the main factor) 512: Access control opening and closing events are highly correlated with the timing of sudden rises and falls in intake air temperature, and may recover after the event ends;

[0113] Assumption H3 (adjacent rack coupling is the main factor) 513: The intake air temperature of the abnormal rack is significantly affected by the exhaust air of the adjacent high-power rack, and spatial topology analysis shows that a thermal coupling path exists;

[0114] Assumption H4 (predominantly hot air supply zone) 514: Multiple cabinets within the same air supply zone simultaneously have excessively high air intake, and the estimated cabinet-level air supply temperature rises overall.

[0115] For each cabinet at each time point, the system interprets the current observed state quantities using the above assumptions: theoretical predictions are generated based on the expected physical relationships, and then compared with actual observations to calculate self-consistent residuals.

[0116] The self-consistent residuals of each hypothesis are composed of residual terms of multiple key state quantities, for example:

[0117] Short-circuit return air ratio residual: The deviation between the inverted short-circuit return air ratio and the assumed expected level;

[0118] Residual term of air intake and supply difference: the deviation between the estimated difference between the air intake temperature and the supply temperature and the assumed expected difference;

[0119] Hotspot height segment intensity residual term: the deviation between the hotspot intensity distribution and the assumed expected distribution;

[0120] Access control event window sudden change residual term: the deviation between the change in air temperature inside the window before and after access control and the expected sudden change (used for H2).

[0121] Supply air temperature field offset residual term: the deviation between the estimated supply air temperature in the same supply air zone and the zone average and the expected increase (used for H4).

[0122] The weights of each residual item are gating by data quality labeling, the confidence level of effective air volume estimation, and the confidence level of alignment: if the inlet air temperature is completely missing or exceeds the physical range, the weight of that item is zero; if the confidence level of effective air volume estimation is low, the weight of short-circuit related items is reduced; if the confidence level of alignment is low, the weight of the residual item in the corresponding window is reduced or removed.

[0123] The smaller the self-consistent residual, the stronger the explanatory power. The system selects the hypothesis with the smallest self-consistent residual and whose derived key parameters satisfy the physical boundary constraints as the current dominant risk type output, while retaining the suboptimal hypothesis and its residual for uncertainty assessment, thereby avoiding unexplainable conclusions that rely solely on multiple index thresholds.

[0124] The Future Lead Prediction Module 520 is used to predict short-term trends of key risk indicators such as short-circuit return air ratio, maximum hot spot intensity, and theoretical temperature rise. The prediction model is built based on historical state quantity sequences and combined with physical evolution constraints: for example, the short-circuit return air ratio is related to structural defects such as cable obstruction, floor vent opening, and missing blind plates, showing a relatively stable trend; hot spot intensity is strongly correlated with the power consumption of specific servers and can be correlated with power consumption prediction; theoretical temperature rise is related to the ratio of power consumption to effective airflow. The system can be implemented using a state-space model or a constrained recurrent neural network, with the goal of determining whether the above indicators will exceed preset safety thresholds within a specified window such as 2 hours, 6 hours, or 24 hours in the future; the safety thresholds can be set comprehensively based on the data center level, equipment temperature resistance level, and historical fault data; if the prediction shows that the short-circuit return air ratio will enter a high-risk range within a certain lead time, or the hot spot intensity exceeds the preset temperature difference and is accompanied by an increase in theoretical temperature rise, a high-risk prediction output is triggered.

[0125] Predictive output module 530, used to output a structured report, includes at least:

[0126] Fault location information: rack number and height range;

[0127] Lead time forecast: the estimated time when risks will manifest;

[0128] Dominant risk type: For example, short-circuit return flow is the main risk.

[0129] Chain of evidence: key data segments and intermediate state quantities that support the conclusion, such as the continuous increase in the proportion of short-circuit return air over a certain period of time while other cabinets in the same area remain stable;

[0130] Recommended inspection items: For example, check whether the air vents on the front floor of the cabinet are blocked by cables, and whether the rear blind flanges are complete.

[0131] A closed-loop calibration unit 600 is used for continuous system self-optimization. This includes:

[0132] The feedback receiving module 610 is used to receive rectification feedback information submitted by operation and maintenance personnel through the system interface, including operation content and timestamp.

[0133] The data comparison and analysis module 620 is used to extract the key status quantity sequences of the cabinet and related areas before and after rectification, and to screen time periods with similar load conditions for comparison to evaluate the rectification effect, such as whether the short-circuit return air ratio has decreased after rectification, whether the hot spot intensity has disappeared, and whether the theoretical temperature rise is closer to the ideal.

[0134] The model parameter calibration module 630 is used to calibrate the internal model parameters of the system, including server fan speed-airflow mapping parameters, smoothing weights in the supply air temperature spatial estimation model, and influence matrix parameters in the hotspot inversion model. The calibration process aims to improve the physical self-consistency of the rectified data and can update parameters using incremental learning or batch retraining.

[0135] The parameter version management module 640 is used to record the reason, effect and scope of each update, so as to adapt to long-term changes brought about by equipment aging and minor adjustments to the computer room layout.

[0136] Example 2, as Figure 2 , Figure 3 As shown, Embodiment 2 of the present invention also provides a method for predicting fault points in an IDC (Internet Data Center) server room, which includes the following steps:

[0137] S1: Collect and standardize IDC data center operation data;

[0138] The IoT data access unit 100 establishes communication connections with various physical devices in the computer room to collect operational data, including PDU current / voltage / power, UPS voltage / frequency / load rate, ambient temperature and humidity, air conditioning supply and return air temperature and fan status, server BMC temperature / fan / power consumption / alarms, switch port traffic and packet loss rate, access control switch events, etc. The multi-protocol parsing engine 110 decodes the protocols, and the standardized processing module 120 outputs structured records containing fields such as asset identifier, location identifier, reporting time, collection time, record type, value or event body, source protocol, and quality mark. Among them, the temperature location includes at least the inlet side and the exhaust side measurement point. When the exhaust side is unavailable, the corresponding observed temperature rise is marked as unavailable.

[0139] S2: Time alignment and data quality annotation;

[0140] The time alignment and quality labeling unit 200 constructs a unified timeline for structured records and aligns them; the data quality assessment module 220 identifies anomalies such as complete missing data, intermittent missing data, long-term constant values, and out-of-physical range and labels the causes; the reliable normal segment filtering module 230 filters continuous time periods to form a set of reliable normal segments based on conditions such as stable access control, stable air supply in the same air supply zone, stable power, and normal temperature difference distribution inside the cabinet; the system also generates alignment reliability markers, marking points with continuous abnormal alignment errors or significant delay drift as low reliability for subsequent gate control.

[0141] S3: Obtain physical layout information of the computer room;

[0142] The cabinet topology and spatial map unit 300 maintains the adjacent relationships, cold aisle relationships, and air supply zoning relationships of cabinets, and stores them in a graph structure or relationship table for subsequent spatial analysis and coupling calculations.

[0143] S4: Estimate the thermal risk status of the server rack;

[0144] The rack thermal risk status estimation unit 400 performs the following based on aligned data and topology information:

[0145] 1) Effective airflow estimation: Under the conditions of a reliable normal segment, stable access control, power change satisfying excitation, normal temperature difference distribution inside the cabinet, and usable intake and exhaust air temperature, calibrate the monotonic mapping from server fan speed to effective heat dissipation airflow; if calibration fails or conditions are not met, it degenerates to the default or general curve; output the effective airflow of the cabinet and give a confidence mark.

[0146] 2) Theoretical temperature rise estimation: The theoretical temperature rise is estimated based on cabinet power, effective air volume and air parameters; when the exhaust side temperature is unavailable, the observed temperature rise is not included in the relevant criteria.

[0147] 3) Cabinet-level spatial estimation of supply air temperature: Using the relationship between air conditioning supply air temperature, cabinet inlet air temperature and topology, the estimated value of cabinet supply air temperature is optimized by physical constraints with normal anchoring and spatial smoothness. The anchored cabinets are normal heat dissipation cabinets with stable access control and high inlet air data quality, where the difference between theoretical temperature rise and observed temperature rise is within the threshold. The weights are gated by data quality labeling, effective air volume estimation confidence and alignment confidence. If the solution fails, it will fall back to the average supply air temperature of the supply air zone.

[0148] 4) Short-circuit return air ratio inversion: When the identifiable conditions are met, based on the joint relationship between the intake air being a proportional mixture of supply cold air and short-circuit hot return air, and the exhaust air temperature being obtained by superimposing the intake air temperature with the theoretical temperature rise, the short-circuit return air ratio is determined by elimination of terms as the ratio of the deviation of the intake air relative to the supply air to the deviation and the superposition of the theoretical temperature rise, thus inverting to obtain the short-circuit return air ratio; if the theoretical temperature rise is lower than the minimum temperature rise threshold, the intake or supply air temperature has missing measurements / constant values / extra-physical conditions, the alignment reliability is low, or the difference between the intake and supply air is less than the minimum temperature difference threshold, it is marked as unidentifiable and data verification is triggered; physical boundary constraints are applied to the identifiable results, negative values ​​are truncated to zero, and values ​​exceeding one are marked as inconsistent and verification and rollback processes are triggered.

[0149] 5) Inversion of vertical hot spot height segments in the cabinet: Divide the cabinet into several height segments, construct the probe baseline temperature and calculate the deviation, and invert the hot spot intensity of each height segment through non-negative constraint least square fitting to locate the most likely hot spot height segment.

[0150] The above calculation outputs a sequence of key state variables.

[0151] S5: Predicts fault locations and provides explanations;

[0152] The fault point prediction and interpretation unit 500 receives the key state quantity sequence and context information; the risk type discrimination module 510, based on a testable hypothesis mechanism, presets at least four types of physical hypotheses (short-circuit return air, access control disturbance, adjacent coupling, and air supply zone heat imbalance), calculates the self-consistent residuals of each hypothesis, and selects the one with the smallest residual as the dominant risk type; the weight of the residual term is gated by data quality labeling, effective air volume estimation credibility, and alignment credibility; the future lead prediction module 520 predicts whether the key indicators exceed the boundary within a specified window and triggers an alarm; the prediction output module 530 generates a structured report containing cabinet and height segment positioning, lead time, dominant risk type, evidence chain, and suggested inspection items.

[0153] S6: Closed-loop calibration system model;

[0154] The closed-loop calibration unit 600 receives rectification feedback; the data comparison and analysis module 620 compares and evaluates the rectification before and after the rectification during periods of similar load; the model parameter calibration module 630 optimizes and calibrates the mapping function, spatial estimation smoothing weight, hotspot inversion influence matrix, etc.; the parameter version management module 640 records the reasons for updates, effects and applicable scope to ensure long-term traceability and evolution.

[0155] It should be noted that, for the sake of brevity, the aforementioned method embodiments are described as a series of actions, but this does not mean that the application limits the order of the steps; based on the idea of ​​this application, some steps can change the execution order without affecting the function implementation, or be executed simultaneously in parallel; secondly, those skilled in the art should also understand that the specific embodiments described in the specification are preferred descriptions of the technical solutions of this application, rather than limitations on the scope of protection of this application, and all equivalent improvements or substitutions made within the spirit and principle of this application should be covered within the scope of protection of this application.

[0156] Example 3: Example 3 of the present invention also provides an IDC data center fault point prediction system, which is executed by one or more processors using the above method steps; the system can be deployed on a local server cluster in an IDC data center or a cloud server.

[0157] The system may include:

[0158] Data acquisition server: Deploys IoT data access unit 100 for communication acquisition, protocol parsing and standardization processing;

[0159] Data processing and analysis server cluster: Deployment time alignment and quality labeling unit 200, rack topology and spatial map unit 300, rack thermal risk status estimation unit 400, and fault point prediction and interpretation unit 500; among which the topology and spatial map can be stored in the database server;

[0160] Model training and calibration server: Deploy 600 closed-loop calibration units for incremental training or batch retraining and parameter optimization;

[0161] User interface server: Provides a web interface or client for viewing forecast reports, submitting feedback, and configuring the system;

[0162] Database server: Used to store structured data records, unified timeline sequences, reliable normal segments, topology information, key state quantity sequences, prediction reports, and model parameter versions, etc.

[0163] The functions of each of the above units can be implemented by software modules and run on the corresponding hardware platform, such as multi-protocol parsing engine 110, standardization processing module 120, time axis construction module 210, data quality assessment module 220, reliable normal segment screening module 230, effective air volume estimation module 410, theoretical temperature rise estimation module 420, supply air temperature spatial estimation module 430, short-circuit return air ratio inversion module 440, vertical hot spot height segment inversion module 450, risk type discrimination module 510, future lead prediction module 520, prediction output module 530, feedback receiving module 610, data comparison and analysis module 620, model parameter calibration module 630, and parameter version management module 640, etc.

[0164] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention; any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solutions of the present invention.

[0165] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting fault points in IDC (Internet Data Center) data centers based on artificial intelligence and the Internet of Things, characterized in that, Includes the following steps: Collect multi-source operational data from the IDC data center and generate structured data records; perform time alignment on the structured data records and generate data quality labels and alignment reliability markers; obtain the rack adjacency relationship and air supply zoning relationship to construct a rack topology diagram; Estimate the rack-level supply air temperature field on the rack topology diagram; the estimation simultaneously satisfies: selecting anchor racks with normal heat dissipation and reliable data, ensuring that the difference between their inlet air temperature and supply air temperature is no greater than the corresponding preset threshold, and applying smoothing constraints to the supply air temperature of adjacent racks or racks within the same supply air zone; calculate the theoretical temperature rise based on rack power and effective air volume, and invert the short-circuit return air ratio according to the energy balance relationship based on the rack inlet air temperature and the supply air temperature field; For each of the preset risk physical assumptions, self-consistent residuals are calculated. The weights of the residual terms of the self-consistent residuals are gated by the data quality labeling and alignment confidence mark. The assumption with the smallest self-consistent residual is selected to determine the risk type and failure point of the target cabinet. The risk indicators are predicted for future time windows and the failure point prediction results containing the target cabinet, risk type and prediction lead time are output. It also includes screening trusted normal segments, wherein the trusted normal segments must simultaneously satisfy at least two conditions: the access control status is stable, the air intake temperature of most cabinets in the air supply area has a consistent trend and the amplitude is less than the corresponding threshold, and the cabinet power change is in a stable range; the effective air volume is obtained through the monotonic mapping relationship from server fan speed to effective heat dissipation air volume, and the monotonic mapping relationship is calibrated based on power change excitation within the trusted normal segments. When calibration fails or the calibration conditions are not met, the effective air volume is obtained by using the default curve or the general curve to roll back.

2. The method for predicting IDC data center fault points based on artificial intelligence and the Internet of Things according to claim 1, characterized in that: The structured data record shall include at least: asset identifier, location identifier, reporting time, collection time, record type, numerical value or event body, source protocol, and quality mark.

3. The method for predicting IDC data center fault points based on artificial intelligence and the Internet of Things according to claim 2, characterized in that: The time alignment includes: statistically analyzing the arrival intervals of each location to obtain a typical delay range; when the reported time is missing or abnormal, using a combination of the acquisition time and the typical delay to estimate the time and map it to a unified time axis; the data quality labeling includes at least complete missing data, intermittent missing data, long-term constant values ​​and out-of-physical range, and the alignment confidence labeling includes at least high confidence and low confidence.

4. The method for predicting IDC data center fault points based on artificial intelligence and the Internet of Things as described in claim 3, characterized in that: The supply air temperature field estimation is achieved by constructing and solving an optimization problem. The objective function of the optimization problem includes: a weighted residual term between the inlet air temperature and the supply air temperature, and a smoothing regularization term defined by the edge set of the cabinet topology graph. The edge set consists of at least one of the adjacency relationship and the same supply air zone relationship, or their union, and if the solution fails, it reverts to the average supply air temperature of the supply air zone.

5. The method for predicting IDC data center fault points based on artificial intelligence and the Internet of Things according to claim 4, characterized in that: Before inverting the short-circuit return air ratio, an identifiability determination is performed. The identifiability determination includes at least the following: the theoretical temperature rise is not less than the minimum temperature rise threshold, the difference between the inlet air temperature and the supply air temperature is not less than the minimum temperature difference threshold, and the data quality and alignment reliability of the relevant points are not low. Physical boundary constraints are applied to the inversion results, negative values ​​are truncated to 0, and results exceeding 1 are marked as inconsistent and trigger data verification.

6. The method for predicting IDC data center fault points based on artificial intelligence and the Internet of Things according to claim 1, characterized in that: It also includes the inversion of vertical hot spot height segments in the cabinet: the cabinet is divided into multiple height segments, a temperature baseline is constructed based on the air supply temperature field and theoretical temperature rise, and a non-negative constraint least squares problem is solved based on the observation deviation of multiple vertical temperature measurement points to obtain the hot spot intensity of each height segment, which is used to refine the fault point from the cabinet level to the height segment inside the cabinet.

7. The method for predicting IDC data center fault points based on artificial intelligence and the Internet of Things according to claim 1, characterized in that: The future time window prediction is to predict at least two of the following: short-circuit return air ratio, hot spot intensity, or theoretical temperature rise. A safety threshold is set according to the data center level or equipment temperature resistance level. When the prediction exceeds the safety threshold within the preset lead time, a high-risk prediction result is output.

8. An IDC (Internet Data Center) fault prediction system based on artificial intelligence and the Internet of Things, used to implement the IDC fault prediction method based on artificial intelligence and the Internet of Things as described in any one of claims 1-7, characterized in that, include: The IoT data access module is used to collect multi-source operational data and generate structured data records; The time alignment and quality labeling module is used for time alignment and to generate data quality labels and alignment confidence markers. The rack topology and spatial map module is used to construct rack topology maps; The thermal risk state estimation module is used to estimate the cabinet-level supply air temperature field that satisfies anchoring and smoothing constraints on the topology map, and obtain the theoretical temperature rise and the short-circuit return air ratio based on power and effective air volume. The fault point prediction and interpretation module is used to calculate self-consistent residuals for preset risk physical assumptions, and to perform weight gating on residual terms based on data quality labeling and alignment confidence marking. The assumption with the smallest self-consistent residual is selected to determine the risk type and fault point, and the prediction lead is output for future time window prediction.

Images