An intelligent analysis method and system based on multi-modal data fusion
By employing an intelligent analysis method that integrates multimodal data fusion with temperature and wind speed data, and dynamically adjusting the observation noise covariance matrix of the Kalman filter, the problem of false overheating alarms in high-density data centers is solved, achieving highly reliable and accurate thermal management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JIAJIE TECH CO LTD
- Filing Date
- 2026-03-17
- Publication Date
- 2026-06-09
Smart Images
Figure CN121858396B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing, and more particularly to an intelligent analysis method and system based on multimodal data fusion. Background Technology
[0002] High-density data centers, as the core carriers of modern information technology infrastructure, are experiencing exponential growth in scale and computing density driven by the high computing power demands of artificial intelligence, big data, and other fields. With the continuous increase in power density per rack, the heat flux per unit area of server chips is rising sharply, leading to increasingly complex and localized heat load distribution within the data center. This makes it highly susceptible to forming temperature hotspots in areas with poor airflow organization. Failure in thermal management not only triggers server frequency reduction protection mechanisms and reduces computing efficiency, but also accelerates the aging of electronic components due to prolonged high-temperature operation, significantly shortening equipment lifespan and increasing the risk of failure.
[0003] In high-density data centers, when the server load suddenly increases, causing a surge in local heat load, or when precision air conditioning fans start and stop or adjust their speed rapidly in response to heat demand, the airflow organization in the cabinet inlet / outlet area will be strongly disturbed. According to the principles of fluid mechanics, high-velocity airflow is prone to boundary layer separation when bypassing obstacles such as cabinet structure and cable trays, which in turn induces local turbulence and vortex phenomena.
[0004] Existing Kalman filtering algorithms employ a fixed measurement noise covariance matrix, which directly determines the system's level of trust in sensor observations during the filtering process. However, in the actual operating environment of high-density data centers, the reliability of observations is significantly reduced when substantial non-thermal interference is superimposed. Because fixed-parameter filters lack the ability to sense environmental disturbances, they cannot dynamically increase the measurement noise covariance matrix value to reduce the weight of noise observations, thus triggering false overheating alarms and failing to accurately reflect the true steady-state temperature trend. Summary of the Invention
[0005] To address the technical problem that fixed-parameter filters lack the ability to sense environmental disturbances and cannot dynamically adjust the observed noise covariance matrix value, thereby triggering false overheating alarms, this invention provides solutions in the following aspects.
[0006] In a first aspect, the present invention provides an intelligent analysis method based on multimodal data fusion, comprising: collecting temperature and wind speed measurements at a target location in a data center at a target time, wherein the target location is any sampling location in the data center and the target time is any sampling time; constructing a target window for the target time; calculating the instability of the target time based on the temperature and wind speed measurements; calculating the confidence level of the target time based on the instability of the target time; obtaining a preset number of comparison locations, taking the previous sampling time adjacent to the target time as the control time, and taking the difference between the temperature measurement at the target time and the temperature measurement at the control time as the temperature change gradient at the target time; calculating the spatial consistency of the temperature sensor at the target location at the target time based on the temperature change gradient at the target location, the temperature change gradient at the comparison location, and the instability; calculating the noise variance correction value of the temperature sensor at the target location at the target time based on the confidence level, spatial consistency, a preset adjustment threshold, and the reference noise variance of the temperature sensor at the target location; and completing the multimodal data analysis based on the noise variance correction value.
[0007] Preferably, the target window for constructing the target time includes: using the target time as the last sampling time in the window and constructing the target time with a preset length as the window scale.
[0008] Preferably, the calculation of the instability at the target time includes: taking the difference between the wind speed measurement value at any sampling time in the target window and the wind speed measurement value at the previous sampling time adjacent to any sampling time as the wind speed difference value at any sampling time, and calculating the absolute value of the wind speed difference value at any sampling time as the first absolute value; constructing a temperature sequence from the temperature measurement values at each sampling time in the target window, smoothing the temperature sequence of the target window using the moving average method to obtain the smoothed temperature measurement value at any sampling time in the target window, and taking the difference between the temperature measurement value at any sampling time and the smoothed temperature measurement value as the temperature residual at any sampling time; calculating the mean of the first absolute values of all sampling times in the target window, and normalizing the mean of the first absolute values to obtain the first normalized value; calculating the mean of the temperature measurement values of all sampling times in the target window, and calculating the standard deviation of the temperature residuals of all sampling times in the target window, calculating the first ratio of the standard deviation of the temperature residuals to the mean of the temperature measurement values, and normalizing the first ratio to obtain the second normalized value; and taking the product of the first normalized value and the second normalized value as the instability at the target time. For example, normalization is performed using the norm function.
[0009] Preferably, the step of calculating the confidence level of the target time based on the instability of the target time includes: calculating a first difference between the instability of the target time and a preset instability threshold; constructing a selection function, wherein the value of the selection function is the maximum value between 0 and the first difference; calculating a first sum of the value of the selection function and 1, and using the reciprocal of the first sum as the confidence level of the target time.
[0010] Preferably, obtaining a preset number of comparison positions includes: calculating the Euclidean distance between any sampling position other than the target position and the target position, sorting all Euclidean distances, and taking the preset number of comparison positions closest to the target position as sampling positions.
[0011] Preferably, the calculation of the spatial consistency of the temperature sensor at the target location at the target time includes: constructing a set of locations from a preset number of comparison locations; calculating the mean of the temperature change gradient of each location in the set at the target time; calculating the absolute difference between the temperature change gradient of the target location at the target time and the mean of the temperature change gradient; using 1 and the sum of the instabilities of the target location at the target time as the denominator; calculating a second ratio of the absolute difference to the denominator; calculating a first product of the second ratio and a preset first adjustment coefficient; and using the negative exponent of the first product as the spatial consistency of the temperature sensor at the target location at the target time.
[0012] Preferably, the calculation of the noise variance correction value of the temperature sensor at the target location at the target time includes: calculating the second product of the confidence level and spatial consistency of the temperature sensor at the target location at the target time; calculating the second difference between the second product and a preset second adjustment coefficient; mapping the inverse of the second difference using a hyperbolic tangent function to obtain a mapped value; and calculating the second sum of the mapped value and 1; and using the product of the second sum and the reference noise variance of the temperature sensor at the target location as the noise variance correction value of the temperature sensor at the target location.
[0013] Preferably, the step of performing multimodal data analysis based on the noise variance correction value includes: using the noise variance correction value as the measurement noise covariance matrix, obtaining the temperature measurement value of the temperature sensor at the target location at the target time, and calculating the temperature prediction value at the target time based on the temperature estimate value at the reference time; inputting the measurement noise covariance matrix, the temperature measurement value at the target time, and the temperature prediction value at the target time into a Kalman filter, and outputting the temperature estimate value at the target time; calculating the temperature change rate at the target time, using the temperature estimate value at the target time and the temperature change rate at the target time as the state vector at the target time, and using the LOF detection algorithm to obtain the anomaly score of the state vector at the target time, thereby completing the multimodal data analysis.
[0014] Secondly, the present invention also provides an intelligent analysis system based on multimodal data fusion, comprising: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned intelligent analysis method based on multimodal data fusion is implemented.
[0015] The present invention has the following effects:
[0016] This invention decouples real heat load changes from airflow disturbance artifacts at the fundamental level of fluid mechanics by coupled sensing of temperature and wind speed: when high-speed airflow causes boundary layer separation due to air conditioning step response or load change to bypass cabinet obstacles, wind speed mode can a priori identify the source of turbulence disturbance, while temperature residual analysis accurately removes high-frequency jitter components, thereby avoiding misjudging instantaneous temperature fluctuations caused by airflow scouring as equipment overheating; at the same time, a spatial consistency verification mechanism is introduced, which effectively distinguishes real thermal events with spatial diffusion characteristics from airflow dead zones or sensor drift interference limited to a single point by comparing the temperature change gradient of neighboring sensors.
[0017] This invention significantly reduces the false alarm rate caused by airflow dynamics, preventing maintenance personnel from falling into alarm fatigue and ignoring real hotspots. It also prevents slowly changing thermal anomalies from being missed due to noise masking, thus improving the reliability and response accuracy of thermal management in high-density data centers and providing a reliable foundation for data analysis. Attached Figure Description
[0018] Figure 1 This is a flowchart of an intelligent analysis method based on multimodal data fusion according to an embodiment of the present invention. Detailed Implementation
[0019] 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 some embodiments of the present invention, but not all embodiments.
[0020] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0021] Specifically, this invention focuses on the refined thermal environment monitoring and management of high-density data center server rooms. In a typical deployment, multimodal sensor groups are installed at preset intervals on the rack pillars or inlet / outlet areas. Each group integrates a high-precision temperature sensor and a wind speed sensor to capture real-time temperature-wind speed coupled data of the local microenvironment at a synchronous sampling frequency. This scenario is suitable for areas with high heat flux density. In such areas, sudden load fluctuations of servers or step speed adjustments of precision air conditioning fans can easily lead to local turbulence and mixing of hot and cold airflows in the gaps between racks and behind cable trays, causing high-frequency fluctuations in the readings of a single temperature sensor. This invention identifies airflow disturbance characteristics by fusing wind speed data at the same location and combines spatial temperature gradient consistency analysis of adjacent locations to dynamically distinguish between "real heat load increase" and "airflow interference artifacts." It then adaptively adjusts the observation noise covariance of the Kalman filter, ultimately providing a high-confidence steady-state temperature estimate for the data center infrastructure management system.
[0022] Reference Figure 1 A smart analysis method based on multimodal data fusion includes steps S1-S4, as detailed below:
[0023] S1: Collect temperature and wind speed measurements at the target location in the data center at the target time. The target location is any sampling location in the data center, and the target time is any sampling time. Construct a target window for the target time and calculate the instability at the target time based on the temperature and wind speed measurements.
[0024] It should be noted that in high-density data centers, server racks have high power density and large heat loads, easily forming local hotspots within the server room. Furthermore, the failure of hot and cold aisle isolation can lead to airflow mixing. When equipment load changes abruptly or air conditioning fans start and stop abruptly, high-speed airflow bypasses obstacles such as rack structures and cable trays, generating local turbulence and vortices based on the boundary layer separation principle of fluid mechanics. This causes irregular scouring and instantaneous mixing of the air around the sensor probe. At this time, the data collected by a single temperature sensor not only includes inherent measurement errors but also incorporates high-frequency nonlinear jitter caused by airflow disturbances. Relying solely on observation data from a single mode or isolated location can easily lead to misjudging airflow artifacts as overheating events and triggering false alarms, or causing noise masking and missing real hotspots. Therefore, this invention deploys a temperature-wind speed coupled multimodal sensor. It uses wind speed prior identification to identify disturbance sources and combines spatial consistency to verify the authenticity of temperature trends, achieving multi-dimensional data cross-validation.
[0025] In one embodiment, temperature and wind speed sensors are deployed at predetermined locations on the uprights of server racks or inlet / outlet areas in a high-density data center. The temperature acquisition unit acquires time-series temperature measurements at a 5Hz frequency to estimate the steady-state temperature. The wind speed acquisition unit simultaneously acquires wind speed measurements at the same location. Although not directly involved in the temperature estimation, these measurements serve as crucial prior environmental information input to the feature construction module to quantify airflow disturbance intensity. When wind speed undergoes a step change or high-frequency fluctuations, this information can be used to identify unsteady airflow events such as turbulence or vortices.
[0026] It's important to note that in data center environment monitoring, when servers are running smoothly, temperature changes are gradual and airflow is stable. The data collected by sensors exhibits low-frequency smoothness, reflecting the true steady-state thermal field distribution. However, when a sudden increase in server load causes a rise in localized heat load, the precision air conditioning system responds quickly and increases fan speed to enhance cooling, resulting in a sharp increase in airflow velocity at the rack inlet / outlet areas. Based on the principle of boundary layer separation in fluid mechanics, high-speed airflow bypassing rack structures, cables, and other obstacles easily induces localized turbulence and vortices, causing irregular instantaneous mixing and scouring of the air around the sensor probe, leading to high-frequency fluctuations in temperature readings. These fluctuations do not originate from actual changes in server heat generation but are "artifact noise" superimposed on the observed signal by airflow dynamic disturbances. If such distorted data is directly input into a traditional Kalman filter, because it uses a fixed noise covariance, it cannot distinguish between actual temperature rise and airflow interference, misjudging it as a drastic temperature change, leading to oscillations in state estimation or false alarms. To this end, the present invention introduces instability at any sampling time to characterize the correlation between airflow disturbance intensity and observation noise, providing a key basis for subsequent dynamic adjustment of the filter's trust weight for the observation value, thereby effectively removing non-thermal interference and restoring the true temperature trend.
[0027] Any sampling time is used as the target time, and the target time is used as the last sampling time in the window. A target window for the target time is constructed with a preset length as the window size. For example, the preset length is 20, that is, 20 consecutive sampling times, including the target time, are selected as the target window, and the target time is the last sampling time at the end of the target window.
[0028] It should be explained that environmental instability is essentially a high-frequency dynamic property in the time dimension. Scalar data at a single sampling moment loses frequency domain information and cannot characterize the perturbation state. Therefore, this invention constructs a target window for the target time.
[0029] The difference between the wind speed measurement value at any sampling time in the target window and the wind speed measurement value at the previous sampling time adjacent to any sampling time is taken as the wind speed difference value at any sampling time, and the absolute value of the wind speed difference value at any sampling time is calculated as the first absolute value.
[0030] The temperature measurement values at each sampling time in the target window are constructed into a temperature sequence. The temperature sequence of the target window is smoothed using the moving average method to obtain the smoothed temperature measurement value at any sampling time in the target window. The difference between the temperature measurement value at any sampling time and the smoothed temperature measurement value is taken as the temperature residual at any sampling time.
[0031] It's important to explain that the raw signal collected by the temperature sensor is the result of a superposition of multiple physical processes: it includes both slow, trend-based temperature drift caused by changes in server load and instantaneous high-frequency fluctuations caused by air conditioning start-up / shutdown or airflow disturbances. If the standard deviation of the temperature sequence within a window is directly used as the stability criterion, it will be difficult to distinguish between these two modes of change: while a genuine increase in heat load manifests as a continuous temperature rise, its change is relatively smooth; whereas airflow turbulence can cause irregular and violent fluctuations in a short period. To accurately isolate the interfering components, this invention calculates the temperature residual, which reflects the energy of high-frequency disturbances and is insensitive to slow, trend-based changes. Therefore, even if the system is in a load ramp-up phase, as long as the residual amplitude remains at a low level, it can be determined that the current temperature change originates from a genuine thermal event rather than airflow disturbance, thus avoiding misjudging a reasonable temperature rise as environmental instability.
[0032] Calculate the mean of the first absolute values of all sampling times in the target window, and normalize the mean of the first absolute values to obtain the first normalized value; calculate the mean of the temperature measurements of all sampling times in the target window, and calculate the standard deviation of the temperature residuals of all sampling times in the target window, calculate the first ratio of the standard deviation of the temperature residuals to the mean of the temperature measurements, and normalize the first ratio to obtain the second normalized value; use the product of the first normalized value and the second normalized value as the instability at the target time.
[0033] The first absolute mean reflects the wind speed variation within the target window. A larger first absolute mean indicates drastic wind speed changes, i.e., turbulence, leading to inaccurate observations. The first ratio represents the degree of fluctuation in the temperature residual. A larger first ratio indicates greater fluctuation in the temperature residual, meaning high-frequency noise appears in the temperature series. Dividing by the mean of temperature measurements helps eliminate the influence of high or low temperature values, focusing on the relative fluctuations of temperature measurements. Multiplying the first normalized value by the second normalized value reveals that the instability at the target time only increases significantly when both drastic wind speed changes and high-frequency noise in the temperature measurements occur simultaneously.
[0034] S2: Calculate the confidence level of the target time based on the instability of the target time.
[0035] It's important to note that in existing Kalman filtering, the observation noise covariance matrix essentially represents the system's prior judgment on the reliability of sensor readings: a smaller observation noise covariance matrix value indicates that the system trusts the observation data more, and the filter assigns higher weights to the observations to quickly track changes; a larger observation noise covariance matrix value indicates that the system relies more on the state prediction model, and the output tends to be smoother. Traditional methods typically set the observation noise covariance matrix to a fixed constant, implicitly assuming that the observation environment is always stable. However, in actual data center operation, when fans start or stop or sudden load changes cause local turbulence, sensor readings will be mixed with significant non-thermal interference. If a small, fixed observation noise covariance matrix value is still used in this case, the filter will incorrectly treat airflow disturbances as real temperature changes and overfit, leading to non-physical oscillations in the estimation results. To address this, the present invention introduces a dynamic confidence mechanism, which adaptively adjusts the observation noise covariance matrix value based on the real-time calculated environmental instability index. When enhanced airflow disturbance is detected, the observation noise covariance matrix value is actively increased to weaken the impact of noise observation, so that the filter can effectively suppress spurious fluctuations caused by environmental disturbances while maintaining the sensitivity of the filter to real thermal events.
[0036] In one embodiment, a first difference between the instability at the target time and a preset instability threshold is calculated, and a selection function is constructed. The value of the selection function is the maximum value between 0 and the first difference. A first sum of the selection function value and 1 is calculated, and the reciprocal of the first sum is used as the confidence level at the target time. For example, the preset instability threshold is 0.2, and the specific value can be set by those skilled in the art based on experience.
[0037] The confidence level calculation logic is as follows: The difference between the instability at the target time and a preset instability threshold is calculated, and this difference is truncated using a selection function. When the instability is within the threshold range, the selection function outputs zero, indicating that the environmental disturbance is at an acceptable level, and the confidence level remains at 1, meaning the sensor observations are fully trusted. Once the instability exceeds the threshold, the selection function starts to output a positive value, which increases linearly with the disturbance intensity. Taking the reciprocal of the first sum makes the confidence level exhibit a monotonically decreasing characteristic. This confidence level calculation process ensures that the system maintains high sensitivity under stable operating conditions to quickly track real temperature changes, while automatically reducing its dependence on observation data during periods of severe airflow disturbance.
[0038] S3: Obtain a preset number of comparison positions, take the previous sampling time adjacent to the target time as the comparison time, and take the difference between the temperature measurement value at the target time and the temperature measurement value at the comparison time as the temperature change gradient at the target time. Calculate the spatial consistency of the temperature sensor at the target position at the target time based on the temperature change gradient at the target position, the temperature change gradient at the comparison position, and the instability.
[0039] It should be noted that while relying solely on time-dimensional instability and confidence analysis can effectively filter out high-frequency fluctuations caused by airflow turbulence, it is insufficient to handle two types of operating conditions that exhibit smooth changes: one is the real heat load increase caused by the synchronous rise of server cluster load, whose temperature changes show a consistent diffusion characteristic in space; the other is isolated temperature rise caused by local airflow dead zones, sensor drift, or heat recirculation, whose changes are limited to a single point and are disconnected from the surrounding area. Since both exhibit low-frequency, slowly varying characteristics in the time series, residual analysis cannot effectively distinguish between them. Therefore, this invention calculates spatial consistency by comparing the temperature change gradient differences between the temperature sensor at the target location and the temperature sensors at neighboring comparison locations, quantifying the degree of agreement between local observations and regional trends. When the temperature rise phenomenon exhibits spatial synergy, it is determined to be a real thermal event and its observation weight is increased; when it appears isolated and out of place, it is considered a local disturbance and its influence is suppressed, thereby maintaining system response sensitivity while avoiding misjudging local artifacts as global heat load changes.
[0040] In one embodiment, the Euclidean distance between any sampling location other than the target location and the target location is calculated, and all Euclidean distances are sorted. A preset number of comparison locations that are closest to the target location are used as sampling locations.
[0041] The previous sampling time adjacent to the target time is used as the control time, and the difference between the temperature measurement value at the target time and the temperature measurement value at the control time is used as the temperature change gradient at the target time. A preset number of comparison locations are constructed into a location set. The mean value of the temperature change gradient at the target time for each location in the set is calculated. The absolute difference between the temperature change gradient at the target location at the target time and the mean value of the temperature change gradient is calculated. The sum of 1 and the instability of the target location at the target time is used as the denominator. A second ratio of the absolute difference to the denominator is calculated. A first product of the second ratio and a preset first adjustment coefficient is calculated. The negative exponent of the first product is used as the spatial consistency of the temperature sensor at the target location at the target time. For example, the preset first adjustment coefficient is 0.1, which is used to avoid the second ratio being too large, causing the negative exponent value to converge too quickly.
[0042] The calculation logic for spatial consistency is as follows: the absolute deviation between the temperature change gradient at the target location and its mean reflects the degree of coordination with the surrounding environment. A smaller deviation indicates that the heating / cooling phenomenon has spatial diffusion characteristics, and is more likely to originate from real thermal events such as synchronous changes in server cluster load. A larger deviation suggests that the change is limited to a single point, possibly caused by a local airflow dead zone or sensor drift. To avoid misjudgment due to instantaneous spatial differences caused by turbulence during periods of severe airflow disturbance, 1 and instability are used as the denominator to dynamically relax the tolerance threshold for spatial outliers. Finally, the normalized deviation is transformed into a range between 0 and 1 through negative exponential mapping. The closer the spatial consistency is to 1, the higher the observation reliability, and the filter should reduce noise covariance to quickly track real changes. A smaller spatial consistency suggests questionable data, requiring an increase in covariance to suppress local interference.
[0043] S4: Based on confidence level, spatial consistency, preset adjustment threshold, and the baseline noise variance of the temperature sensor at the target location, calculate the noise variance correction value of the temperature sensor at the target time, and complete the multimodal data analysis based on the noise variance correction value.
[0044] It's important to note that the measurement noise covariance matrix, as a core parameter of the Kalman filter, essentially defines the system's trade-off strategy between observed data and the prediction model during state updates. When the measurement noise covariance matrix is large, the filter tends to weaken the influence of the current observation and rely more on the state transition model for prediction, thereby suppressing noise interference and maintaining output smoothness. Conversely, when the measurement noise covariance matrix is small, the filter assigns higher weights to the observed values to improve the sensitivity to tracking changes in the actual state. However, traditional methods set the measurement noise covariance matrix to a fixed constant, assuming a consistently stable observation environment. This is difficult to adapt to dynamic conditions such as airflow disturbances and sudden load changes in data centers. Using a small measurement noise covariance matrix value in turbulent environments can lead to overfitting distorted data and causing oscillations. Conversely, if the measurement noise covariance matrix value is too large when the actual heat load increases, it can delay the response and miss hot spots.
[0045] To this end, this invention innovatively integrates confidence level and spatial consistency to construct an adaptive adjustment mechanism: when both indicate that the data is reliable, the observation noise covariance matrix value is actively compressed to accelerate the tracking of the actual temperature rise; when local outliers or high disturbances are detected, the observation noise covariance matrix value is appropriately amplified to suppress artifact interference, thereby achieving dynamic optimization of filtering performance under different operating conditions, taking into account both response speed and anti-interference capability.
[0046] In one embodiment, a second product of the confidence level of the temperature sensor at the target location at the target time and spatial consistency is calculated; a second difference between the second product and a preset second adjustment coefficient is calculated; the inverse of the second difference is mapped using a hyperbolic tangent function to obtain a mapped value; and a second sum of the mapped value and 1 is calculated; the product of the second sum and the reference noise variance of the temperature sensor at the target location is used as the noise variance correction value of the temperature sensor at the target location. For example, the preset second adjustment coefficient is 0.3. The reference noise variance of the temperature sensor is obtained from the sensor datasheet (typically determined by the sensor's factory precision).
[0047] The logic for constructing the noise variance correction value is as follows: When the second product is lower than the preset second adjustment coefficient, it indicates that the sensor is simultaneously facing airflow disturbances (low confidence) or spatial outliers (low consistency). In this case, the mapping value approaches 1, making the noise variance correction value approach 2, amplifying the noise variance, and the filter automatically suppresses the observation to avoid interference. Conversely, when the second product is higher than the threshold, it indicates that the temperature change is both stable and spatially consistent. The mapping value approaches negative 1, the noise variance correction value approaches 0, the noise variance is compressed, and the filter enhances its dependence on the observation to quickly track real thermal events. The S-shaped nonlinearity of the hyperbolic tangent function ensures a smooth transition in the adjustment process, avoiding estimation jumps caused by abrupt changes in covariance. At the same time, the design of the second sum value constrains the adjustment coefficient to the range of 0 to 2, ensuring that the physical meaning of the baseline variance is not destroyed, and achieving continuous adjustability from complete suppression to complete trust. This allows the filter to have both anti-interference capability and response sensitivity in complex airflow environments.
[0048] The noise variance correction value of the temperature sensor at the target location at the target time is directly constructed as the measurement noise covariance matrix at the target time. It should be noted that the dimension of the measurement noise covariance matrix strictly corresponds to the dimension of the system observation vector: in this invention, the measurement input of the Kalman filter is only the temperature measurement value (the wind speed measurement value is only used as an environmental state assessment parameter in the noise covariance adjustment calculation and is not included in the state estimation object). Therefore, the observation vector is equal to 1, and the measurement noise covariance matrix is a 1-by-1 scalar matrix. This matrix quantifies the system's confidence in the current temperature measurement value. The larger the value, the stronger the observation noise and the less reliable the data. The filter will reduce the Kalman gain to suppress the noise effect. The smaller the value, the more reliable the observation. The filter will increase the Kalman gain to quickly track the real temperature change.
[0049] The system acquires the temperature measurement value of the temperature sensor at the target location at the target time, and calculates the predicted temperature value at the target time based on the temperature estimate at the reference time. The measurement noise covariance matrix, the measured temperature value at the target time, and the predicted temperature value at the target time are input into a Kalman filter to output the estimated temperature value at the target time. The temperature change rate at the target time is calculated, and the temperature estimate and the temperature change rate are used as the state vector at the target time. It should be noted that calculating the predicted temperature at the target time from the temperature estimate at the reference time is a technique well-known to those skilled in the art, and will not be elaborated upon here.
[0050] Based on this, the Local Outlier Factor (LOF) anomaly detection algorithm is used to calculate the anomaly score of the state vector of the temperature sensor at the target location at the target time. When the anomaly score exceeds a preset threshold (for example, the preset threshold is 1.2), it is determined that there is a potential thermal anomaly risk at the target location, and the system automatically issues an early warning to the operation and maintenance personnel, thereby achieving early identification and intervention of real hotspots and avoiding false alarms caused by airflow disturbances from interfering with normal operation and maintenance decisions.
[0051] The system includes a processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement an intelligent analysis method based on multimodal data fusion according to the first aspect of the present invention.
[0052] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.
[0053] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.
Claims
1. An intelligent analysis method based on multimodal data fusion, characterized in that, include: The system collects temperature and wind speed measurements at a target location in the data center at a target time. The target location is any sampling location in the data center, and the target time is any sampling time. A target window is constructed for the target time, and the instability at the target time is calculated based on the temperature and wind speed measurements. The instabilities at the target time of calculation include: The difference between the wind speed measurement value at any sampling time in the target window and the wind speed measurement value at the previous sampling time adjacent to any sampling time is taken as the wind speed difference value at any sampling time, and the absolute value of the wind speed difference value at any sampling time is calculated as the first absolute value. The temperature measurement values at each sampling time in the target window are constructed into a temperature sequence. The temperature sequence of the target window is smoothed using the moving average method to obtain the smoothed temperature measurement value at any sampling time in the target window. The difference between the temperature measurement value at any sampling time and the smoothed temperature measurement value is taken as the temperature residual at any sampling time. Calculate the mean of the first absolute values at all sampling times in the target window, and normalize the mean of the first absolute values to obtain the first normalized value; Calculate the mean of temperature measurements at all sampling times in the target window, and calculate the standard deviation of the temperature residuals at all sampling times in the target window. Calculate the first ratio of the standard deviation of the temperature residuals to the mean of the temperature measurements, and normalize the first ratio to obtain the second normalized value. The product of the first normalized value and the second normalized value is taken as the instability at the target time. Calculate the confidence level at the target time based on the instability at the target time; A preset number of comparison locations are obtained. The previous sampling time adjacent to the target time is used as the reference time. The difference between the temperature measurement value at the target time and the temperature measurement value at the reference time is used as the temperature change gradient at the target time. Based on the temperature change gradient at the target location, the temperature change gradient at the comparison location, and the instability, the spatial consistency of the temperature sensor at the target location at the target time is calculated. Based on confidence level, spatial consistency, preset adjustment threshold, and the baseline noise variance of the temperature sensor at the target location, the noise variance correction value of the temperature sensor at the target time is calculated, and multimodal data analysis is completed based on the noise variance correction value.
2. The intelligent analysis method based on multimodal data fusion according to claim 1, characterized in that, The target window at the construction target time includes: The target time is used as the last sampling time in the window, and a target window of the target time is constructed with a preset length as the window size.
3. The intelligent analysis method based on multimodal data fusion according to claim 1, characterized in that, The calculation of the confidence level at the target time based on the instability at the target time includes: Calculate the first difference between the instability at the target time and the preset instability threshold, construct a selection function, and take the maximum value between 0 and the first difference; Calculate the first sum of the value of the selection function and 1, and use the reciprocal of the first sum as the confidence level at the target time.
4. The intelligent analysis method based on multimodal data fusion according to claim 1, characterized in that, The comparison positions for obtaining the preset number include: Calculate the Euclidean distance between any sampling location other than the target location and the target location, sort all Euclidean distances, and select the preset number of comparison locations closest to the target location as sampling locations.
5. The intelligent analysis method based on multimodal data fusion according to claim 1, characterized in that, The spatial consistency of the temperature sensor at the target location at the target time includes: A set of locations is constructed by constructing a preset number of comparison locations. The mean value of the temperature change gradient at each location in the set at the target time is calculated. The absolute difference between the temperature change gradient at the target location at the target time and the mean value of the temperature change gradient is calculated. Use 1 and the sum of the instabilities of the target position at the target time as the denominator; Calculate the second ratio of the absolute difference to the denominator, calculate the first product of the second ratio and the preset first adjustment coefficient, and use the negative exponent of the first product as the spatial consistency of the temperature sensor at the target location at the target time.
6. The intelligent analysis method based on multimodal data fusion according to claim 1, characterized in that, The noise variance correction value of the temperature sensor at the target location at the target time includes: Calculate the second product of the confidence level of the temperature sensor at the target location at the target time and the spatial consistency, and calculate the second difference between the second product and the preset second adjustment coefficient; The opposite of the second difference is mapped using the hyperbolic tangent function to obtain the mapped value, and the second sum of the mapped value and 1 is calculated. The product of the second sum and the reference noise variance of the temperature sensor at the target location is used as the noise variance correction value of the temperature sensor at the target location.
7. The intelligent analysis method based on multimodal data fusion according to claim 1, characterized in that, The multimodal data analysis based on the noise variance correction value includes: The noise variance correction value is used as the measurement noise covariance matrix. The temperature measurement value of the temperature sensor at the target location at the target time is obtained. The temperature prediction value at the target time is calculated based on the temperature estimate value at the reference time. The measurement noise covariance matrix, the measured temperature value at the target time, and the predicted temperature value at the target time are input into the Kalman filter, and the output is the temperature estimate at the target time. The temperature change rate at the target time is calculated, and the estimated temperature at the target time and the temperature change rate at the target time are used as the state vector at the target time. The LOF detection algorithm is used to obtain the anomaly score of the state vector at the target time, thus completing the multimodal data analysis.
8. An intelligent analysis system based on multimodal data fusion, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement an intelligent analysis method based on multimodal data fusion according to any one of claims 1-7.