Machine room equipment monitoring system

Abstract

The invention relates to the technical field of intelligent machine room equipment monitoring, in particular to a machine room equipment monitoring system, which comprises a distributed data acquisition unit, a multi-source data fusion module, an equipment state evaluation module and a dynamic regulation and control execution unit which are sequentially connected through an industrial Ethernet, wherein the distributed data acquisition unit is configured to acquire equipment vibration data by adopting a vibration sensor array deployed in a spatial orthogonality manner; temperature distribution of the equipment surface, the air inlet and the air outlet is obtained through the temperature sensor set deployed in a gradient mode; a non-contact current transformer is used to detect the current waveform of the device. The structured and intelligent monitoring system is beneficial to remote monitoring and management of large-scale machine room equipment, energy consumption is optimized, the service life of the equipment is prolonged, early warning and processing can be carried out before a fault occurs, the maintenance cost is reduced, and the operation and maintenance efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent computer room equipment monitoring, and in particular to a computer room equipment monitoring system. Background Art

[0002] With the rapid development of technologies such as cloud computing and artificial intelligence, the scale of equipment in modern data centers is growing exponentially. Intelligent monitoring of equipment in computer rooms has become a core technical challenge in ensuring high reliability and low energy consumption. Existing computer room monitoring technologies are primarily based on single parameter threshold alarm mechanisms. As equipment complexity increases and multi-physics coupling intensifies, the following technical flaws are gradually being exposed:

[0003] Vibration monitoring uses a single axial accelerometer bolted to the device housing. Temperature detection uses discretely arranged NTC thermistors with a fixed sampling frequency of 1 Hz. Current measurement uses a clamp-on transformer, which only records the effective value parameters. This solution results in a lack of temporal and spatial correlation in the data, making it unable to accurately reflect the device's operating status. The server hard drive vibration signal and the rack fan airflow disturbance produce coupling interference, resulting in a high error rate in fault feature extraction.

[0004] The traditional method uses a threshold-based piecewise linear model, which has the following defects: the threshold setting relies on manual experience and does not consider parameter drift caused by equipment aging. The single parameter alarm mechanism cannot identify complex failure modes, such as the simultaneous occurrence of associated failures caused by bearing wear and coil short circuit. Summary of the Invention

[0005] Based on the above objectives, the present invention provides a computer room equipment monitoring system, comprising:

[0006] Distributed data acquisition unit, multi-source data fusion module, equipment status assessment module, dynamic control execution unit, including:

[0007] The distributed data acquisition unit is configured to: collect equipment vibration data using a spatially orthogonally deployed vibration sensor array; obtain the temperature distribution of the equipment surface, air inlet, and air outlet using a gradient-deployed temperature sensor group; and detect the equipment current waveform using a contactless current transformer.

[0008] The multi-source data fusion module is configured to: establish a spatiotemporal coordinate system to normalize and encode the multi-source data; fuse the sensor data using an evidence theory method with dynamic weight adjustment; and output a standardized vector containing device state characteristics;

[0009] The equipment status assessment module is configured to: generate a health index by processing the standardized vector through a transfer learning model; calculate the remaining life of the equipment based on a multi-stage prediction model;

[0010] The dynamic control execution module is configured to: construct a multi-objective optimization function to generate a control strategy; and distribute control instructions to air conditioners, UPSs, and network equipment actuators.

[0011] Preferably, the deployment of the vibration sensor array includes:

[0012] Three-axis acceleration sensors are arranged at six orthogonal locations in the equipment cabinet to form a detection network surrounding the equipment;

[0013] Methods for determining the distance between adjacent sensors include:

[0014] a. Obtain the natural frequency of the device through tapping test;

[0015] b. Calculate the wavelength of the sound wave corresponding to the resonant frequency;

[0016] c. Set the sensor spacing to 1 / 4 of the wavelength to avoid standing wave interference;

[0017] The signal processing flow includes:

[0018] a. Charge amplification and impedance matching of the original sensor signal;

[0019] b. Adopt adaptive bandpass filter to eliminate environmental noise, and its passband range is dynamically adjusted according to the speed of the equipment;

[0020] c. Separate the equipment body vibration and conducted vibration through cross-correlation analysis.

[0021] Preferably, the gradient deployment of the temperature sensor group includes:

[0022] The surface of the device uses a flexible thin film thermocouple array, and its wiring path is dynamically planned according to the distribution of heating elements;

[0023] The spatial resolution of the distributed fiber optic temperature measurement unit deployed at the air inlet is calibrated using optical time domain reflectometry;

[0024] The calibration method for synchronizing the sampling frequency of the infrared thermal imager array installed at the air outlet with the fan speed includes:

[0025] a. Extract the passing frequency of the fan blades;

[0026] b. Set the sampling frequency to an integer multiple of the passing frequency;

[0027] c. Maintain sampling synchronization through a phase-locked loop.

[0028] Preferably, the evidence theory method includes:

[0029] Define the basic probability distribution function for each sensor data:

[0030] The temperature data allocation function is constructed based on the historical fault pattern library; the vibration data allocation function is determined by the wavelet packet energy entropy distribution; the current data allocation function associates the harmonic distortion rate with the load factor;

[0031] The dynamic weight adjustment process includes:

[0032] a. Calculate the KL divergence between each piece of evidence and the baseline model;

[0033] b. Adjust the weight of evidence in inverse proportion to the scatter value;

[0034] c. When the conflict factor exceeds the threshold, the weight redistribution mechanism is activated.

[0035] Preferably, the construction of the transfer learning model includes:

[0036] The neural network architecture consists of:

[0037] The LSTM layer is used to extract time series features, the attention mechanism layer calculates feature weight distribution, and the fully connected layer outputs the health index;

[0038] Online update methods include:

[0039] a. Freeze the underlying parameters of the pre-trained network;

[0040] b. Dynamically update the attention layer weights using sliding time window data;

[0041] c. Maintain the continuity of model output through knowledge distillation.

[0042] Preferably, the multi-stage prediction model includes:

[0043] Cointegration analysis is used in the early stages to detect performance degradation trends;

[0044] In the mid-term, an ARIMA model is constructed to predict changes in health indices;

[0045] In the late stage, physical failure models are applied to calculate the remaining life;

[0046] Multi-model fusion strategies include:

[0047] a. Calculate the confidence index of each model's prediction results;

[0048] b. Triggering the manual review process when the prediction difference exceeds the threshold;

[0049] c. Output weighted average lifespan value, where weight is positively correlated with confidence.

[0050] Preferably, the construction of the multi-objective optimization function includes:

[0051] Define the target set: minimize the health index decline rate, minimize the inverse of air conditioning system energy consumption, and maximize the UPS power supply time;

[0052] Constraint settings:

[0053] a. Temperature constraints are set according to the maximum operating temperature allowed by the equipment;

[0054] b. Vibration constraints adopt ISO10816 standard limits;

[0055] c. Network delay constraints are graded according to the urgency of control instructions.

[0056] Preferably, the air conditioning control strategy generation includes:

[0057] Generate candidate air supply solutions based on computational fluid dynamics simulation. Specific steps:

[0058] a. Discretize the 3D model of the computer room into a finite volume grid;

[0059] b. Input real-time temperature distribution and equipment heat data;

[0060] c. Solve the Navier-Stokes equations to obtain the thermal field distribution;

[0061] d. Evaluate the thermal uniformity index of each air supply scheme;

[0062] The optimal solution selection criteria include:

[0063] Pareto front solutions that satisfy temperature constraints;

[0064] The balance coefficient between fan energy consumption and cooling efficiency.

[0065] Preferably, the network QoS control includes:

[0066] Traffic classification method:

[0067] a. Deep packet inspection to identify protocol types;

[0068] b. Statistical traffic burst and periodic characteristics;

[0069] c. Prioritize services based on device health index;

[0070] Dynamic bandwidth allocation process:

[0071] a. Monitor network load rate and end-to-end delay;

[0072] b. When the load rate exceeds 50%, elastic allocation is initiated: i. Allocate real-time control flow bandwidth according to the urgency of the health index; ii. Implement weighted random early detection on the monitoring data flow; iii. Limit the management flow bandwidth to no more than 10% of the total bandwidth. Preferably, the exception handling mechanism includes:

[0073] Multi-level alarm trigger condition settings:

[0074] a. Level 1 alarm: A single indicator exceeds the threshold for 10 seconds;

[0075] b. Level 2 alarm: related indicators are abnormal at the same time;

[0076] c. Level 3 warning: The remaining life is lower than the safety margin;

[0077] Emergency response process:

[0078] a. Start the backup sensor to verify the data;

[0079] b. Switch to the lightweight diagnostic model to quickly locate the fault source; c. Execute the pre-set emergency plan based on the fault type:

[0080] Perform graded load shedding in case of electrical faults;

[0081] Active vibration suppression triggered by mechanical faults;

[0082] Thermal failure initiates emergency cooling protocol.

[0083] Beneficial effects of the present invention:

[0084] Through the coordinated efforts of these connections and steps, the system can provide early warnings before equipment failures occur and provide real-time health assessments and predictions. Multi-source data fusion and dynamic control strategies effectively optimize equipment operating efficiency, reduce the frequency of failures, and extend equipment lifespan. In particular, transfer learning and multi-stage predictive models accurately assess equipment health and remaining lifespan, ensuring that equipment in the computer room is always in optimal working condition. Furthermore, through dynamic control and optimization strategies, the system not only improves equipment performance but also effectively conserves energy and optimizes resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0086] Figure 1 It is a structural block diagram of the system of the present invention;

[0087] Figure 2 A flow chart of the steps of a method for determining the distance between adjacent sensors in the system of the present invention;

[0088] Figure 3 A flowchart of the steps for generating candidate air supply schemes based on computational fluid dynamics simulation in the system of the present invention. DETAILED DESCRIPTION

[0089] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0090] See Figure 1-Figure 3 The embodiment of the present invention provides a computer room equipment monitoring system, wherein a distributed data acquisition unit is responsible for collecting key operating data of the equipment in real time. The system includes:

[0091] Vibration sensor array: Using triaxial accelerometers arranged in an orthogonal pattern, these sensors accurately monitor equipment vibration data. Sensors are positioned at six locations within the cabinet to ensure full coverage of the equipment's vibration profile, and strategic spacing is employed to minimize standing wave interference. This data is crucial for monitoring the mechanical health of the equipment.

[0092] Temperature sensor array: Equipment surface temperature is monitored by a flexible thin-film thermocouple array, air inlet temperature is acquired by a distributed fiber-optic temperature measurement unit, and air outlet temperature is monitored by an infrared thermal imager array. The layout and measurement method of each sensor are carefully designed to ensure a comprehensive and accurate reflection of the equipment's thermal conditions.

[0093] Non-contact current transformer: used to detect the current waveform of the equipment, capable of capturing the dynamic changes of the current and assisting in analyzing the electrical status of the equipment.

[0094] The multi-source data fusion module is used to uniformly process and fuse sensor data from different sources. First, the collected multi-source data is normalized and encoded by establishing a spatiotemporal coordinate system to ensure comparability across sensor types. Next, the evidence theory method employs dynamic weight adjustment to fuse the data from different sensors, adjusting the weights of each sensor to ensure a balanced sensor data set. Finally, the module outputs a standardized vector containing device status characteristics, providing reliable data support for further evaluation of the device status.

[0095] The device health assessment module processes standardized vectors to generate a device health index and continuously updates the device health assessment using a transfer learning model. Transfer learning leverages existing historical data to quickly adapt to the health assessment needs of different devices, eliminating the tedious task of retraining models. This module also uses a multi-stage prediction model to predict the remaining life of the device. From early performance degradation trends to late-stage physical failure models, this module provides accurate lifecycle estimates for the long-term operation of the device.

[0096] The dynamic control execution module uses multi-objective optimization functions to generate effective control strategies and issue control instructions to actuators such as air conditioners, UPS (uninterruptible power supplies), and network devices. These strategies include controlling device temperature, vibration, and power supply to ensure optimal operation. For example, the air conditioning system's control strategy can dynamically adjust the cooling mode based on temperature data to maintain the normal temperature of the equipment; the UPS system can provide backup power when equipment anomalies occur, ensuring stable operation; and network equipment can adjust bandwidth and traffic priority based on real-time status.

[0097] Through the coordinated efforts of these connections and steps, the system can provide early warnings before equipment failures occur and provide real-time health assessments and predictions. Multi-source data fusion and dynamic control strategies effectively optimize equipment operating efficiency, reduce the frequency of failures, and extend equipment lifespan. In particular, transfer learning and multi-stage predictive models accurately assess equipment health and remaining lifespan, ensuring that equipment in the computer room is always in optimal working condition. Furthermore, through dynamic control and optimization strategies, the system not only improves equipment performance but also effectively conserves energy and optimizes resource utilization.

[0098] In one possible implementation, the vibration sensor array uses a spatially orthogonal deployment, with triaxial accelerometers placed in six locations around the equipment cabinet, forming a vibration detection network that fully encompasses the equipment. Each accelerometer can detect the equipment's vibration in three orthogonal directions (X, Y, and Z), ensuring that vibration in every direction and angle of the equipment is accurately captured. This arrangement offers the advantage of more comprehensive visibility into the equipment's vibration status, effectively avoiding vibration blind spots and improving monitoring accuracy and reliability.

[0099] To avoid standing wave interference and ensure the accuracy of vibration data, the spacing between adjacent sensors needs to be properly set. The specific method is as follows:

[0100] By tapping the device and collecting the vibration response signal, the device's natural frequency can be obtained. This is to identify the device's resonant frequency range during operation and prevent resonance within the frequency range that could lead to device failure.

[0101] Based on the natural frequency, the corresponding wavelength of the sound wave is calculated using the basic formula for sound wave propagation (wave speed = frequency × wavelength). This allows you to determine the wavelength of the vibration transmitted by the device and use this information to set the sensor spacing.

[0102] Setting the spacing between adjacent sensors to one-quarter of the calculated wavelength avoids interference from standing waves, which can cause signal distortion and errors, affecting the accuracy of equipment vibration data. Therefore, this measure can effectively improve system monitoring accuracy.

[0103] Signal processing flow:

[0104] The raw vibration signals collected by the sensor are often very weak, so they require charge amplification. Furthermore, considering the impedance matching between the sensor and the signal processing circuit, the circuit design is adjusted to ensure maximum efficiency during signal transmission and avoid signal attenuation.

[0105] Because computer room environments often generate noise from various electrical devices and external factors, adaptive bandpass filters are required for signal processing to eliminate this interference. The filter's passband range dynamically adjusts based on the device's rotational speed, ensuring that even with varying speeds, it can still accurately filter out external noise unrelated to device vibration.

[0106] When equipment is operating, it may generate conducted vibrations, which can mix with the vibration signals from the equipment itself. To accurately analyze the equipment's status, cross-correlation analysis is used to separate the signals. This method distinguishes vibrations originating from the equipment itself from those transmitted to other locations based on the signal's time domain characteristics, extracting accurate information about the equipment's vibrations.

[0107] In one possible implementation, the deployment of the temperature sensor group enables accurate monitoring of the device surface and air flow through different types of sensors. Specifically:

[0108] An array of flexible thin-film thermocouples is used on the device surface. The routing of this array is dynamically planned based on the distribution of heat-generating components within the device. The flexible thin-film thermocouples adhere closely to the device surface, enabling real-time temperature monitoring of all device components. This dynamic routing ensures that temperature monitoring can adapt to varying device layouts and the location of heat-generating components, ensuring high-precision coverage of heat sources. This technology captures temperature changes across all areas of the device surface in real time, providing detailed temperature data, particularly in complex devices or with unevenly distributed heat sources.

[0109] Distributed fiber-optic temperature measurement units are deployed at the air inlets. Fiber-optic temperature sensors offer high accuracy and long-distance transmission capabilities, making them particularly suitable for temperature monitoring over large areas. Using optical time-domain reflectometry (OTDR) to calibrate the spatial resolution of the optical fiber, high-precision temperature monitoring is possible at the air inlet. Fiber-optic temperature sensors are highly durable and reliable, adapting to varying temperature fluctuations within the equipment room and providing stable monitoring data.

[0110] An array of infrared thermal imagers is installed at the air outlet to capture the temperature distribution in the air flow area. This provides a comprehensive temperature map, making it particularly useful for monitoring temperature fluctuations around the air outlet to assess cooling effectiveness. Infrared thermal imaging technology can accurately detect areas of abnormal temperature during equipment operation, helping to identify potential cooling deficiencies or overheating issues.

[0111] The sampling frequency of the infrared thermal imager needs to be synchronized with the fan speed to ensure that the temperature changes in the corresponding area can be stably captured during the fan operation. The calibration method includes the following steps:

[0112] First, a sensor or other measurement method is used to extract the fan blade's passage frequency—the frequency at which the fan blades pass through the fan during each rotation. This frequency is a key parameter in determining the fan's speed, ensuring that synchronous sampling can accurately track temperature changes during fan rotation.

[0113] To ensure the synchronization and accuracy of the thermal imager's sampling data, the sampling frequency was set to an integer multiple of the fan blade's frequency. This setting allows the thermal imager to stably sample within each fan rotation cycle, capturing temperature distribution data synchronized with the fan's rotation.

[0114] Phase-locked loop (PLL) technology is used to synchronize the sampling frequency with the fan speed. PLL can automatically adjust the sampling timing to ensure that the infrared thermal imager can still accurately perform synchronous sampling when the fan speed changes.

[0115] In one possible implementation, the application of the evidence theory method in this system is mainly to conduct a comprehensive analysis through the probability distribution function of sensor data, and then enhance the accuracy of the system's assessment of the equipment health status through a dynamic weight adjustment process.

[0116] Specifically, the temperature data distribution function is constructed based on a historical failure pattern library. This means the system analyzes the device's historical failure records and combines the relationships between different temperature data and failure patterns to construct a probability distribution function for the temperature data. This historical failure pattern library helps the system determine during real-time monitoring whether the current temperature state is within the normal range or is approaching certain known failure modes.

[0117] The distribution function for vibration data is determined by the energy entropy distribution of wavelet packets. Wavelet packet transforms effectively extract information about different frequency components from vibration signals, allowing for precise analysis of the energy distribution of equipment vibration. This helps identify possible abnormal vibration patterns during equipment operation and, in turn, infer the risk of mechanical failure.

[0118] The allocation function for current data is determined by correlating harmonic distortion and load factor. Harmonic distortion is a common sign of electrical faults in power systems, while load factor reflects the workload of the equipment. By combining these two parameters, the system can assess whether the current data is abnormal and thus determine whether the electrical system is facing overload or other electrical problems.

[0119] In the evidence theory approach, after the data provided by each sensor is initially processed, the weight of each piece of evidence needs to be dynamically adjusted to more accurately reflect the credibility and importance of each piece of evidence in the decision-making process. The specific steps are as follows:

[0120] a. Calculate the KL divergence between each piece of evidence and the baseline model: KL divergence (Kullback-Leibler divergence) is a common method for measuring the difference between two probability distributions. Here, KL divergence is used to measure the difference between sensor data (such as temperature, vibration, and current data) and the baseline model (typically an ideal model of the normal operating state of the device). By calculating KL divergence, the system can identify which pieces of evidence deviate from the expected normal pattern.

[0121] b. Adjust the weight of evidence inversely proportional to the divergence value: A larger KL divergence value indicates that the evidence differs more from the baseline model, reflecting lower credibility. Therefore, the weight should be adjusted inversely proportionally. That is, the larger the divergence value, the lower the weight of the evidence. Conversely, evidence with a smaller divergence value indicates that it is closer to the normal model and has higher credibility, so it is given a larger weight.

[0122] c. Activate a weight redistribution mechanism when the conflict factor exceeds a threshold: In actual operation, if conflicting evidence arises between multiple sensors (for example, some sensors indicate normal temperature, while others indicate abnormal temperature), a conflict factor is calculated. When the conflict factor exceeds a preset threshold, the system activates a weight redistribution mechanism, readjusting the weights of the evidence to resolve the conflict and ensure a more reasonable final decision. This mechanism allows the system to find the optimal balance between multiple pieces of evidence, ensuring more robust decision-making.

[0123] By introducing evidence theory methods and combining the probability distribution and dynamic weight adjustment mechanism of different sensor data, the system can effectively improve the intelligence level and fault prediction ability of computer room equipment monitoring, reduce the risk of equipment failure, and ensure the stability and safety of the computer room environment.

[0124] In one possible implementation, a computer room equipment monitoring system uses a transfer learning model for fault detection and health assessment. By combining LSTM layers, attention mechanisms, and fully connected layers in a neural network architecture, supplemented by online update methods, the system can monitor equipment health in real time and gradually improve prediction accuracy.

[0125] Specifically, LSTM (Long Short-Term Memory) networks excel at processing time series data. The operating status of equipment in a computer room often exhibits temporal correlations, with changes in temperature, vibration, current, and other data exhibiting certain regularities. The LSTM layer effectively captures these temporal characteristics and learns the evolution of equipment status over time. In this way, the LSTM layer can identify dynamic changes in equipment health and extract key time series features, providing important evidence for subsequent health assessments.

[0126] The attention mechanism assigns different weights to different features, enabling the model to focus on important features and ignore irrelevant ones. In this system, the attention mechanism layer dynamically calculates the weight distribution of features based on the features extracted by the LSTM layer, weighting them according to their importance at different time points. This enables the model to focus on time points or data relevant to the health status of the equipment, thereby improving the accuracy and reliability of fault prediction.

[0127] After extracting and weighting key features, the fully connected layer transforms these features into a final health index output. The health index is a comprehensive indicator of the device's operating status that intuitively reflects its health. The fully connected layer learns the nonlinear relationships between different features to produce a final health assessment result, helping the monitoring system determine whether the device has potential faults or requires maintenance.

[0128] The online update method enables the model to be continuously optimized and adjusted according to new data in actual applications, improving the real-time and adaptability of the model. The specific steps are as follows:

[0129] In transfer learning, a network pre-trained on a large dataset is typically used, its underlying parameters frozen, and only the upper layer of the network structure is adjusted. This approach avoids training the entire network from scratch, conserving computing resources and accelerating model convergence. Freezing the underlying parameters leverages the general features learned by the previous model while focusing on adjusting the higher-level features relevant to the current computer room equipment monitoring, improving the system's performance for specific tasks.

[0130] The sliding time window approach regularly updates the model's weights by continuously acquiring the latest time series data. For equipment monitoring, the sliding time window ensures that the model constantly adapts to changes in equipment status and captures new failure modes or operating patterns. Dynamically updating the attention layer's weights means the system adjusts the feature importance scores based on new time series data, enabling the model to more accurately identify the impact of the current time point on the equipment's health. This approach allows the system to continuously optimize performance over time and adapt to long-term changes in equipment.

[0131] Knowledge distillation is a technique that transfers knowledge from a large model (the teacher model) to a smaller model (the student model). In this system, knowledge distillation is used to ensure output continuity during transfer learning. When updating a model, the student model learns from the output behavior of the teacher model to ensure that the new model's output does not differ significantly from the old model's, thus avoiding performance fluctuations caused by network updates. This has the beneficial effect of making the model's output more stable, enabling a smooth transition to new learning states and avoiding unnecessary disruption to system operations.

[0132] In one possible implementation, in the early stages of equipment development, cointegration analysis is used to identify long-term trends and relationships between multiple variables. Cointegration analysis can detect common trends between multiple time series variables, even if their respective time series may fluctuate to a certain extent. Applied to equipment monitoring in computer rooms, cointegration analysis can reveal the potential relationships between various equipment indicators (such as temperature, pressure, current, etc.). When deviations from the long-term common trends between certain indicators are detected, the system can issue an early warning and determine that the equipment may be experiencing a performance degradation trend. This stage is mainly used to promptly detect and take measures when minor degradation of the equipment occurs, thereby preventing early failures.

[0133] Specifically, for each time series x t and y t , first perform a unit root test (such as ADF test) to confirm that they are non-stationary time series.

[0134] Δx t =α+βx t-1 +∈ t ;

[0135] where Δx t is the difference of the time series, α is a constant term, β is the parameter of the unit root test, ∈ t is the error term.

[0136] Cointegration regression: If the test passes, it means x t and y tThere may be a cointegration relationship between them. Then, a cointegration regression model is constructed to test the long-term stable relationship between the two time series.

[0137] y t =θ0+θ1x t +∈ t ;

[0138] Among them, θ0 and θ1 are regression coefficients, ∈ t is the error term of the regression model. The cointegration regression model can reveal x t and g t The long-term equilibrium relationship between them.

[0139] At an early stage, when cointegration analysis reveals significant changes in the relationships between performance indicators, it can be inferred that equipment is showing signs of degradation. For example, if the relationship between temperature and current deviates from its long-term stable state, this could be a precursor to equipment failure. In this case, the system will issue an early warning, indicating a potential performance degradation issue.

[0140] During the equipment's mid-term operation phase, the ARIMA (Autoregressive Integrated Moving Average) model is used to predict changes in the health index. The ARIMA model is suitable for processing time series data and can predict future health status by analyzing trends, seasonality, and randomness in historical data. Based on the equipment's operating data and current status, the model can predict future changes in the health index, helping the monitoring system assess future health risks to the equipment. The advantage of the ARIMA model lies in its ability to fit the changing trends of the equipment health index based on historical data and predict possible future anomalies or failures.

[0141] Specifically, the basic form of the ARIMA model is:

[0142] X t =φ1X t-1 +φ2X t-2 +…+φ p X t-p +θ1∈ t-1 +θ2∈ t-2 +…+θ q ∈ t-q +∈ t ;

[0143] X t It is a time series value (such as a device health index).

[0144] φ1,φ2,...,φ p are the parameters of the autoregressive (AR) part.

[0145] θ1, θ2, ..., θ qIt is the parameter of the moving average (MA) part.

[0146] ∈ t is the white noise error term.

[0147] In the medium term, the ARIMA model is used to predict the health index of equipment. For example, if the equipment's health index exhibits a trend or cyclical fluctuations over time, the ARIMA model can predict the health index value for the next few months. This is important for mid-term maintenance and predicting potential equipment failures.

[0148] When equipment reaches its advanced stages, physical failure models are used to calculate the remaining lifespan. This model assesses the equipment's remaining useful life in its current state through an in-depth analysis of factors such as its operating principle, material properties, and operating environment. Physical failure models are typically based on equipment degradation patterns and failure modes, combined with historical operating data, to accurately estimate the equipment's likely failure time within a certain period of time. This provides a scientific basis for equipment maintenance and replacement, avoiding unnecessary premature or delayed replacement.

[0149] The fusion strategy for multiple prediction models first relies on calculating the confidence level of each model's predictions. Different models may produce different predictions about device health, so evaluating the confidence level of each model's predictions is crucial. By analyzing the model's prediction error, historical performance, and training quality, a confidence level can be assigned to each model. This confidence level reflects the reliability of the model's predictions and provides a foundation for subsequent fusion.

[0150] When significant discrepancies between the predictions of multiple models exceed a set threshold, the system automatically triggers a manual review process. This process is typically performed by experienced engineers or technicians, who conduct a detailed inspection of the equipment's status, verify the predictions of each model, and adjust the prediction plan based on actual conditions. This process effectively avoids errors or uncertainties in model predictions, ensuring high accuracy and safety of equipment monitoring.

[0151] Finally, the system outputs a weighted average lifespan value. The results of each prediction model are weighted according to their confidence level, with models with higher confidence levels having a greater impact on the final lifespan value. This weighted fusion approach allows the system to combine the strengths of each model to produce a more accurate and reliable prediction of the remaining lifespan of the equipment. This weighted average lifespan value provides maintenance personnel with a reasonable window for equipment replacement or repair, helping to reduce the risk of equipment failure and optimize maintenance strategies for equipment in the computer room.

[0152] In one possible implementation, the multi-objective optimization function in the computer room equipment monitoring system is primarily used to balance multiple objectives to ensure equipment health, energy efficiency, and power supply reliability. The system incorporates multiple objectives and constraints to ensure optimal equipment operation, avoiding excessive resource consumption or service interruptions caused by equipment failure.

[0153] The health index of a device is an important indicator of its operating status. The rate of decline in the health index reflects the rate of device aging and performance degradation. This goal aims to improve device reliability and service life by optimizing operating conditions, slowing degradation, and slowing down performance.

[0154] The air conditioning system in the computer room consumes a lot of energy, so minimizing its energy consumption is a key goal. "Minimizing the inverse of energy consumption" here can be understood as minimizing air conditioning energy consumption, thereby improving the system's energy efficiency.

[0155] The UPS system is an important power supply for equipment in the computer room. Maximizing its power supply time can ensure that the equipment can continue to operate normally in the event of a power outage, thereby avoiding data loss or service outage caused by power outages.

[0156] These goals can conflict with each other. For example, reducing air conditioning energy consumption may cause temperatures to rise, affecting the health of the equipment. Therefore, the system needs to use optimization algorithms to find a balance point so that multiple goals can be optimized simultaneously within a reasonable range.

[0157] To ensure that the equipment can still operate under safe and reasonable conditions while meeting different goals, the system sets the following constraints:

[0158] The maximum operating temperature of the equipment is a key factor in temperature constraints. Exceeding this temperature can lead to equipment failure or performance degradation due to overheating. Therefore, the system monitors the room temperature in real time and adjusts the air conditioning system to ensure that the temperature remains within the permitted range for the equipment.

[0159] According to the ISO10816 standard, equipment in computer rooms (especially UPS and air conditioning equipment) must be maintained within specific vibration limits to prevent damage from excessive vibration. The system monitors vibration sensor data to ensure that the equipment does not exceed these limits.

[0160] Network latency is crucial to device control responses. The urgency of control commands directly determines the acceptable range of latency. The system prioritizes control commands by urgency, prioritizing high-priority commands to ensure a fast response time, ensuring network latency does not exceed the specified upper limit.

[0161] In the embodiment of the present invention, the weight of each goal is determined according to actual needs. For example, if device health is a priority goal, then "minimizing the rate of decline of health index" may be given a higher weight.

[0162] Using an optimization algorithm (such as a multi-objective genetic algorithm or particle swarm optimization), we can find the optimal solution within the constraints of a set of objectives. For example, we can calculate the optimal combination of objectives while satisfying constraints such as temperature, vibration, and network latency.

[0163] The system dynamically adjusts its optimization strategies based on real-time data inputs (such as sensor data on temperature, vibration, and network latency). For example, if the system detects that a device's temperature is approaching its upper limit, it will automatically reduce the energy consumption of the air conditioning system or increase the UPS power supply time to ensure safe operation of the device.

[0164] In one possible implementation, the process generates multiple candidate air supply solutions through computational fluid dynamics (CFD) simulation to achieve the optimal air conditioning control strategy. The specific steps are as follows:

[0165] a. Discretize the 3D model of the computer room into a finite volume grid

[0166] First, the 3D model of the computer room needs to be converted into a finite volume mesh. The accuracy of the mesh directly affects the accuracy of the simulation results. Through meshing, the computer room space can be divided into multiple small volume units, providing a foundation for subsequent fluid dynamics calculations.

[0167] b. Input real-time temperature distribution and equipment heat data

[0168] The system then inputs the room's temperature distribution and the heat output of each device based on real-time monitoring data. This data helps the system more accurately simulate thermal field changes within the room and predict cooling needs based on the heat output of different devices.

[0169] c. Solve the Navier-Stokes equations to obtain the thermal field distribution

[0170] After inputting temperature and heat generation data, the system simulates the dynamic process of fluid (air conditioning air supply) and heat transfer by solving the Navier-Stokes equations. This equation describes the changes in the velocity and temperature fields of air flow, thereby determining the thermal field distribution within the computer room. The key to this step is to accurately reflect the relationship between air flow and heat distribution.

[0171] d. Evaluate the thermal uniformity index of each air supply scheme

[0172] The system evaluates different air supply solutions based on the calculated thermal field distribution. The primary evaluation metric is thermal uniformity, which determines whether the temperature distribution throughout the equipment room is balanced. Uneven temperature distribution can cause some equipment to overheat or undercool, impacting normal operation. Therefore, the optimization goal is to achieve the most uniform temperature distribution possible within the equipment room.

[0173] After generating and evaluating multiple candidate air supply solutions, the system needs to select the best solution. The selection criteria include:

[0174] The optimal solution must satisfy temperature constraints, ensuring that the equipment operates within its maximum allowable temperature range. The Pareto front solution is a concept in multi-objective optimization, indicating that no single objective can be improved without sacrificing other objectives. In air conditioning control strategies, the Pareto front solution indicates the optimal balance between temperature control and other objectives, such as energy efficiency.

[0175] When selecting an air supply solution, it's also important to consider fan energy consumption and the cooling efficiency of the air conditioning system. The optimal solution not only ensures an appropriate temperature within the computer room, but also minimizes air conditioning system energy consumption and maximizes cooling efficiency. This means finding the optimal balance between energy efficiency and temperature control.

[0176] In one possible implementation, a traffic classification method is used to identify and classify different types of data flows in a network and to assign appropriate priorities and bandwidths to different flows.

[0177] Specifically, deep packet inspection (DPI) analyzes the content of each packet in the network to identify different protocol types. This step helps the system accurately identify application layer protocols within traffic, such as HTTP, FTP, and VoIP. This identification enables the network to classify different services based on traffic type, enabling precise control and prioritization.

[0178] By analyzing traffic burst and periodicity, the system can analyze traffic patterns. Burst traffic typically represents instantaneous peaks, while periodic traffic exhibits regular patterns. This analysis helps predict traffic demand and optimize bandwidth allocation based on traffic characteristics. For example, periodic traffic can be allocated lower bandwidth during high-load periods, while bursty traffic requires more flexible bandwidth support.

[0179] The device health index is a comprehensive indicator based on multiple factors, including the device's operating status, load, and fault warnings. Based on the device health index, the system can prioritize services. Devices with lower health indexes or critical tasks receive priority bandwidth and resource allocation, while traffic with higher health indexes or non-critical tasks receives lower priority.

[0180] The dynamic bandwidth allocation process is used to flexibly adjust the allocation of bandwidth resources based on real-time network load and latency. The steps of this process are as follows:

[0181] First, the system monitors network load rate and end-to-end latency in real time to understand the network's current status. Load rate indicates current network utilization, while end-to-end latency reflects the delay in data transmission from source to destination. By monitoring these metrics, the system can determine whether the network is congested and provide a basis for bandwidth allocation.

[0182] When the network load exceeds 50%, the system activates the elastic bandwidth allocation mechanism. This indicates that network traffic is approaching or exceeding the carrying capacity, and bandwidth needs to be flexibly allocated to avoid network congestion. The elastic allocation mechanism dynamically adjusts bandwidth resources based on real-time demand.

[0183] Specific measures for flexible allocation include:

[0184] For devices with a lower health index or in an emergency state, the system will prioritize allocating more bandwidth resources to them to ensure that critical businesses or devices can receive timely network support.

[0185] Weighted Random Early Detection (WRED) is a congestion control mechanism that can preemptively discard low-priority packets based on traffic importance and priority, preventing further network congestion. This allows the system to effectively manage network bandwidth usage and prioritize the transmission of critical traffic.

[0186] The system also limits the bandwidth usage of management traffic (such as monitoring, configuration, and control traffic) to no more than 10% of the total bandwidth. This ensures that management traffic does not occupy excessive bandwidth resources and affect the transmission of normal business traffic.

[0187] In one possible implementation, the exception handling mechanism can effectively respond to various failures or abnormal situations and ensure the continuous and stable operation of the system by combining multi-level alarm trigger condition settings and emergency response processes.

[0188] The multi-level alarm mechanism is used to trigger different levels of alarms based on abnormal conditions of equipment monitoring indicators, so as to promptly detect and respond to equipment problems.

[0189] A Level 1 alarm is the most basic alarm triggering condition. When a monitored indicator (such as voltage, temperature, or humidity) exceeds a preset safety threshold for more than 10 seconds, the system triggers a Level 1 alarm. This alarm indicates a potential problem, but it is not necessarily an emergency failure and requires further verification.

[0190] Level 2 alerts have stricter triggering conditions. When multiple related indicators exhibit anomalies simultaneously, this indicates a potentially more serious problem. The system assesses the likelihood of a fault based on these anomalies and triggers a level 2 alert to quickly locate the source of the fault and expedite response.

[0191] A Level 3 alert typically indicates a high risk of impending equipment or system failure. This alert is issued when the remaining life of the equipment falls below a predetermined safety margin. This indicates that the equipment may be entering an extremely high-risk state, requiring immediate preventive action to avoid significant failure or damage.

[0192] The design of the emergency response process enables effective measures to be taken quickly after the alarm is triggered, reducing the risk of failure and restoring the normal operation of the system as soon as possible when a failure occurs.

[0193] When the system detects an anomaly and triggers an alarm, it first activates the backup sensor to verify the data is accurate. The backup sensor can provide auxiliary data when the primary sensor fails or the data is abnormal, ensuring that false alarms are not caused by sensor failure or data deviation.

[0194] After verifying the data, the system switches to a lightweight diagnostic model for rapid fault diagnosis. This model can quickly identify the root cause of the fault, locate the device or system issue, and provide diagnostic results. This process significantly shortens fault location time and ensures that further emergency measures can be implemented promptly.

[0195] Depending on the type of fault, the system will execute the corresponding pre-set emergency plan. These plans are designed based on emergency measures for different fault types and can quickly respond to different fault scenarios. Specifically, they include:

[0196] Execute graded load shedding for electrical faults: When the system detects an electrical fault, it will cut off the power supply to some equipment according to the severity of the fault, gradually reducing the load to prevent the fault from spreading to other equipment.

[0197] Active vibration suppression triggered by mechanical faults: When a mechanical fault occurs, the system will initiate vibration suppression measures to reduce the vibration intensity of mechanical components, thereby preventing damage to mechanical components or worsening of the fault.

[0198] Thermal failure triggers emergency cooling protocol: When the equipment temperature rises abnormally and reaches the risk of overheating, the system automatically triggers the emergency cooling protocol to quickly lower the temperature in the equipment room to avoid equipment damage or data loss due to overheating.

[0199] The exception handling mechanism can improve the reliability and response speed of the equipment monitoring system in the computer room through the reasonable design of multi-level alarms and emergency response processes, ensuring that the equipment can resume normal operation in a timely and effective manner when a failure occurs.

[0200] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0201] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A computer room equipment monitoring system, characterized in that: Includes: Distributed data acquisition unit, multi-source data fusion module, equipment status assessment module, dynamic control execution unit, including: The distributed data acquisition unit is configured to: collect equipment vibration data using a spatially orthogonally deployed vibration sensor array; obtain the temperature distribution of the equipment surface, air inlet, and air outlet using a gradient-deployed temperature sensor group; and detect the equipment current waveform using a contactless current transformer. The multi-source data fusion module is configured to: establish a spatiotemporal coordinate system to normalize and encode the multi-source data; fuse the sensor data using an evidence theory method with dynamic weight adjustment; and output a standardized vector containing device state characteristics; The equipment status assessment module is configured to: generate a health index by processing the standardized vector through a transfer learning model; calculate the remaining life of the equipment based on a multi-stage prediction model; The dynamic control execution module is configured to: construct a multi-objective optimization function to generate a control strategy; and distribute control instructions to air conditioners, UPSs, and network equipment actuators.

2. A computer room equipment monitoring system according to claim 1, characterized in that: The deployment of the vibration sensor array includes: Three-axis acceleration sensors are arranged at six orthogonal locations in the equipment cabinet to form a detection network surrounding the equipment; Methods for determining the spacing between adjacent sensors include: a. Obtain the natural frequency of the device through tapping test; b. Calculate the wavelength of the sound wave corresponding to the resonant frequency; c. Set the sensor spacing to 1 / 4 of the wavelength to avoid standing wave interference; The signal processing flow includes: a. Charge amplification and impedance matching of the original sensor signal; b. Adopt adaptive bandpass filter to eliminate environmental noise, and its passband range is dynamically adjusted according to the speed of the equipment; c. Separate the equipment body vibration and conducted vibration through cross-correlation analysis.

3. A computer room equipment monitoring system according to claim 1, characterized in that: The gradient deployment of the temperature sensor group includes: The surface of the device uses a flexible thin film thermocouple array, and its wiring path is dynamically planned according to the distribution of heating elements; The spatial resolution of the distributed fiber optic temperature measurement unit deployed at the air inlet is calibrated using optical time domain reflectometry; The calibration method for synchronizing the sampling frequency of the infrared thermal imager array installed at the air outlet with the fan speed includes: a. Extract the passing frequency of the fan blades; b. Set the sampling frequency to an integer multiple of the passing frequency; c. Maintain sampling synchronization through a phase-locked loop.

4. A computer room equipment monitoring system according to claim 1, characterized in that: The evidence-theoretic approach includes: Define the basic probability distribution function for each sensor data: The temperature data allocation function is constructed based on the historical fault pattern library; the vibration data allocation function is determined by the wavelet packet energy entropy distribution; the current data allocation function associates the harmonic distortion rate with the load factor; The dynamic weight adjustment process includes: a. Calculate the KL divergence between each piece of evidence and the baseline model; b. Adjust the weight of evidence in inverse proportion to the scatter value; c. When the conflict factor exceeds the threshold, the weight redistribution mechanism is activated.

5. A computer room equipment monitoring system according to claim 1, characterized in that: The construction of the transfer learning model includes: The neural network architecture consists of: The LSTM layer is used to extract time series features, the attention mechanism layer calculates feature weight distribution, and the fully connected layer outputs the health index; Online update methods include: a. Freeze the underlying parameters of the pre-trained network; b. Dynamically update the attention layer weights using sliding time window data; c. Maintain the continuity of model output through knowledge distillation.

6. A computer room equipment monitoring system according to claim 1, characterized in that: The multi-stage prediction model includes: Cointegration analysis is used in the early stages to detect performance degradation trends; In the mid-term, an ARIMA model is constructed to predict changes in health indices; In the late stage, physical failure models are applied to calculate the remaining life; Multi-model fusion strategies include: a. Calculate the confidence index of each model's prediction results; b. Triggering the manual review process when the prediction difference exceeds the threshold; c. Output weighted average lifespan value, where weight is positively correlated with confidence.

7. A computer room equipment monitoring system according to claim 1, characterized in that: The construction of the multi-objective optimization function includes: Define the target set: minimize the health index decline rate, minimize the inverse of air conditioning system energy consumption, and maximize the UPS power supply time; Constraint settings: a. Temperature constraints are set according to the maximum operating temperature allowed by the equipment; b. Vibration constraints adopt ISO10816 standard limits; c. Network delay constraints are graded according to the urgency of control instructions.

8. A computer room equipment monitoring system according to claim 7, characterized in that: The air conditioning control strategy generation includes: Generate candidate air supply solutions based on computational fluid dynamics simulation. Specific steps: a. Discretize the 3D model of the computer room into a finite volume grid; b. Input real-time temperature distribution and equipment heat data; c. Solve the Navier-Stokes equations to obtain the thermal field distribution; d. Evaluate the thermal uniformity index of each air supply scheme; The optimal solution selection criteria include: Pareto front solutions that satisfy temperature constraints; The balance coefficient between fan energy consumption and cooling efficiency.

9. A computer room equipment monitoring system according to claim 1, characterized in that: The network QoS control includes: Traffic classification method: a. Deep packet inspection to identify protocol types; b. Statistical traffic burst and periodic characteristics; c. Prioritize services based on device health index; Dynamic bandwidth allocation process: a. Monitor network load rate and end-to-end delay; b. Start elastic allocation when the load rate exceeds 50%: i. Allocate real-time control flow bandwidth according to the urgency of the health index; ii. Implement weighted random early detection on monitoring data streams; iii. Limit the management traffic bandwidth to no more than 10% of the total bandwidth.

10. A computer room equipment monitoring system according to claim 1, characterized in that: The exception handling mechanism includes: Multi-level alarm trigger condition settings: a. Level 1 alarm: A single indicator exceeds the threshold for 10 seconds; b. Level 2 alarm: related indicators are abnormal at the same time; c. Level 3 warning: The remaining life is lower than the safety margin; Emergency response process: a. Start the backup sensor to verify the data; b. Switch to a lightweight diagnostic model to quickly locate the source of the fault; c. Execute the preset emergency plan according to the fault type: Perform graded load shedding in case of electrical faults; Active vibration suppression triggered by mechanical faults; Thermal failure initiates emergency cooling protocol.

Images