A robot active perception immersed cabinet adaptive diagnosis method

Abstract

This application discloses an adaptive diagnostic method for immersion cabinets based on robot active perception. The method includes acquiring real-time perception data collected by the robot moving along a preset inspection path during inspection, and forming a spatial real-time signal sequence according to the spatial location of the acquisition points; calculating a difference sequence based on the spatial real-time signal sequence and a preset spatial basic signal sequence; performing spatial difference calculation on the difference sequence to obtain a spatial gradient change rate sequence; acquiring the current real-time total power consumption of the cabinet, and generating a spatial adaptive threshold based on the real-time total power consumption and the difference sequence; comparing the gradient change rate of each spatial location with the spatial adaptive threshold based on the spatial gradient change rate sequence to identify abnormal spatial location intervals, and further determining the abnormal diagnostic result based on the abnormal spatial location intervals. This application enables comprehensive and early diagnosis of the health status of the cabinet through robot active perception.

Description

Technical Field

[0001] This application relates to the field of data center operation and maintenance technology, and in particular to an adaptive diagnostic method for immersion cabinets based on robot active perception. Background Technology

[0002] Immersion liquid-cooled server racks are high-efficiency heat dissipation devices that completely immerse servers in insulating coolant, and are widely used in high heat flux density scenarios such as data centers and intelligent computing centers. To ensure the safe and stable operation of the rack, real-time monitoring and fault diagnosis of the internal status of the rack are necessary.

[0003] Existing diagnostic methods for immersion cabinets mainly rely on fixedly deployed sensors. When the sensor reading exceeds a preset threshold, an alarm is triggered. However, since fixed sensors can only reflect the physical state of their single-point area, and the physical field distribution inside the immersion cabinet is uneven, local anomalies are difficult to be captured by remote sensors in the early stages, which can easily lead to the failure of early warnings. At the same time, changes in cabinet load can cause significant changes in the normal fluctuation range. Fixed thresholds cannot take into account both high and low load conditions, which can easily lead to missed or false alarms. Summary of the Invention

[0004] The purpose of this application is to provide an adaptive diagnostic method for immersion cabinets based on robot active perception, in order to solve the problems of blind spots, insufficient early warning capabilities, and high false alarm rates in existing fixed sensor diagnostic methods.

[0005] Firstly, this application provides an adaptive diagnostic method for immersion cabinets based on robot active perception, employing the following technical solution: The robot acquires real-time perception data as it moves along a preset inspection path during the inspection process, and forms a real-time spatial signal sequence according to the spatial location of the acquisition points. Based on the real-time spatial signal sequence, the difference sequence is calculated by using a preset spatial basic signal sequence. The preset spatial basic signal sequence is reference data that the robot collects and stores in advance when the cabinet is in a healthy state. Spatial difference calculation is performed on the difference sequence to obtain the spatial gradient rate of change sequence; Obtain the real-time total power consumption of the current rack, and generate a space adaptive threshold based on the real-time total power consumption and the difference sequence; Based on the spatial gradient change rate sequence, the gradient change rate of each spatial location is compared with the spatial adaptive threshold to identify abnormal spatial location intervals, and the abnormal diagnosis results are further determined according to the abnormal spatial location intervals.

[0006] The above technical solution utilizes the robot's mobile perception capability to construct a spatial signal sequence characterizing the physical field distribution along the space of the outer surface of the cabinet. By calculating the difference sequence and the rate of change of the spatial gradient, it captures the local physical field distortion caused by internal faults. At the same time, it introduces a diagnostic threshold that is adaptively adjusted according to the operating conditions to generate the real-time total power consumption of the cabinet, effectively distinguishing between load fluctuations and real anomalies, and realizing the proactive and accurate detection of early faults in submerged cabinets.

[0007] Optionally, before calculating the difference sequence based on the real-time spatial signal sequence and a preset spatial base signal sequence, the following steps are included: Acquire the current environmental parameters of the immersion cabinet, as well as the baseline environmental parameters when the preset space basic signal sequence is acquired; Environmental compensation correction is performed on the real-time spatial signal sequence based on the difference between the current environmental parameters and the reference environmental parameters.

[0008] Optionally, generating a spatially adaptive threshold based on the real-time total power consumption and the difference sequence includes: Divide the difference sequence into sliding windows, calculate the local standard deviation of the difference data within each sliding window, and form a local standard deviation sequence; The local standard deviation sequence is sorted by value, and a preset quantile is extracted as the spatial benchmark standard deviation; Based on the real-time total power consumption, the sensitivity coefficient corresponding to the current load is determined by a preset load-sensitivity mapping function; A spatially adaptive threshold is generated based on the baseline standard deviation and sensitivity coefficient.

[0009] Optionally, the step of further determining the abnormal diagnosis result based on the abnormal spatial location interval includes: Based on the abnormal spatial location range, a fixed-point exploration command is generated. The fixed-point exploration command is used to control the robot to move back and forth between the start and end points of the exploration range at a preset speed, and to perform a static pause for a preset duration each time it reaches the start and end points. Acquire the sequence of sensory data generated by the robot during the execution of fixed-point exploration instructions, at a preset sampling rate; Based on the perceived data sequence, dynamic features in the time domain are extracted, and based on these dynamic features, anomaly diagnosis results for the abnormal spatial location intervals are generated.

[0010] Optionally, the sensing data sequence includes sensing signal values ​​collected at fixed sampling time intervals, the robot's real-time spatial position, and motion direction identifiers; the temporal dynamic features include a reciprocating consistency coefficient and a response time constant; and the extraction of temporal dynamic features based on the sensing data sequence includes: Based on the direction of motion, the sensing data sequence is divided into a forward passage sequence and a reverse passage sequence; Based on the forward and reverse passage sequences, calculate the reciprocating consistency coefficient between the forward and reverse sensing signal data at the same spatial location; Obtain the data segment in the static pause period of the sensing data sequence, and extract the response time constant of the sensing signal data in the data segment as time changes.

[0011] Optionally, generating anomaly diagnosis results for anomaly spatial location intervals based on temporal dynamic characteristics includes: Based on the current total power consumption of the rack, the health characteristic distribution corresponding to the current load is determined through a preset constraint relationship model. The preset constraint relationship model characterizes the joint distribution law of the time-domain dynamic characteristics under the health state as the real-time total power consumption changes. Calculate the eigenvectors formed by the time-domain dynamic features, and their Mahalanobis distance and deviation direction angle under the health feature distribution; If the Mahalanobis distance exceeds the preset distance threshold, the abnormal dominant type is determined based on the deviation direction angle, and the abnormal diagnosis result of the abnormal spatial location interval is marked as an abnormal state, while the abnormal dominant type is output. If the Mahalanobis distance does not exceed the preset distance threshold, the abnormal diagnosis result of the abnormal spatial location interval will be marked as a healthy state.

[0012] Optionally, after labeling the abnormal diagnostic results of the abnormal spatial location interval as a healthy state, the process includes: Using the abnormal spatial location interval as an index, basic signal data and real-time signal data are extracted from the preset spatial basic signal sequence and spatial real-time signal sequence, respectively; Based on real-time signal data, the basic signal data is updated by pre-setting update weight coefficients.

[0013] Optionally, before updating the basic signal data based on real-time signal data using preset update weight coefficients, the method further includes: Calculate the absolute value of the deviation between real-time signal data and basic signal data; If the absolute value of the deviation exceeds the preset maximum permissible deviation threshold, the basic signal data will be updated using the preset maximum permissible deviation threshold as the upper limit for the update.

[0014] Secondly, this application provides a robot-driven adaptive diagnostic system for immersion cabinets, comprising: The data acquisition module 101 is used to acquire real-time perception data collected by the robot as it moves along the preset inspection path during the inspection process, and to form a real-time spatial signal sequence according to the spatial location of the acquisition point. The difference sequence calculation module 102 is used to calculate the difference sequence based on the real-time spatial signal sequence and through a preset spatial basic signal sequence. The preset spatial basic signal sequence is reference data that the robot collects and stores in advance when the cabinet is in a healthy state. The spatial difference calculation module 103 is used to perform spatial difference calculation on the difference sequence to obtain the spatial gradient change rate sequence; The adaptive threshold generation module 104 is used to obtain the real-time total power consumption of the current cabinet and generate a spatial adaptive threshold based on the real-time total power consumption and the difference sequence. The abnormal spatial diagnosis module 105 is used to compare the gradient change rate of each spatial location with the spatial adaptive threshold based on the spatial gradient change rate sequence, identify abnormal spatial location intervals, and further determine the abnormal diagnosis results based on the abnormal spatial location intervals.

[0015] Thirdly, this application provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above for a robot-actively-perceived adaptive diagnostic method for immersion cabinets.

[0016] In summary, this application first constructs a spatial signal sequence through robot mobile inspection and uses the spatial gradient change rate to achieve preliminary screening of abnormal spatial intervals. In addition, it obtains temporal dynamic features through fixed-point reciprocating exploration and combines them with Mahalanobis distance determination based on power consumption constraints to achieve two-level accurate diagnosis from suspected abnormality to confirmed abnormality. Furthermore, through a baseline sequence adaptive update mechanism driven by diagnostic results, the diagnostic system can continuously evolve with the normal aging of the equipment, ensuring diagnostic accuracy throughout its entire life cycle. Attached Figure Description

[0017] Figure 1 This is a flowchart of an adaptive diagnostic method for an immersion cabinet based on robot active perception, provided in an embodiment of this application. Figure 2 This is a flowchart provided in an embodiment of the present application for generating a spatially adaptive threshold based on real-time total power consumption and a difference sequence; Figure 3 This is a flowchart provided in an embodiment of the present application for further determining the abnormal diagnosis result based on the abnormal spatial location interval; Figure 4 This is a flowchart of extracting temporal dynamic features from a sensing data sequence, provided in an embodiment of this application. Figure 5 This is a schematic diagram of an adaptive diagnostic system for an immersion cabinet with robot active perception, provided in an embodiment of this application. Detailed Implementation

[0018] The following is in conjunction with the appendix Figure 1 -Appendix Figure 5 This application will be described in further detail below.

[0019] This application provides an adaptive diagnostic method for immersion cabinets based on robot active perception. See [link to relevant documentation] Figure 1 This includes the following steps: S100: Acquire real-time perception data collected by the robot as it moves along the preset inspection path during the inspection process, and form a spatial real-time signal sequence according to the spatial location of the collection point.

[0020] S200: Based on the real-time spatial signal sequence, calculate the difference sequence using a preset spatial basic signal sequence.

[0021] S300. Perform spatial difference calculation on the difference sequence to obtain the spatial gradient rate of change sequence.

[0022] S400: Obtain the real-time total power consumption of the current rack, and generate a space adaptive threshold based on the real-time total power consumption and the difference sequence.

[0023] S500: Based on the spatial gradient change rate sequence, the gradient change rate of each spatial location is compared with the spatial adaptive threshold to identify abnormal spatial location intervals, and the abnormal diagnosis results are further determined based on the abnormal spatial location intervals.

[0024] In this embodiment, the robot moves at a constant speed along a preset path (such as an inspection guide rail) on the outer surface of the immersion cabinet. The onboard sensing module collects sensing signals (such as temperature and sound pressure levels) in a spatially triggered manner (e.g., sampling once every fixed distance moved), forming a real-time spatial signal sequence corresponding to each spatial position, denoted as... ,in For spatial indexing place, or the first The real-time sensing signal value at each spatial sampling point, where n is the total number of spatial sampling points divided along the preset inspection path.

[0025] After acquiring the real-time spatial signal sequence, a difference sequence is calculated based on it and a preset spatial base signal sequence. The preset spatial base signal sequence is reference data pre-collected and stored by the robot while the rack is in a healthy state, denoted as... , For the robot in the Sensing signal values ​​collected at each spatial sampling point under the health status of the server rack.

[0026] Since the preset spatial basic signal sequence is collected under specific environmental conditions, if the environmental conditions during the current inspection differ from the baseline environmental conditions, there will be a systematic deviation between the real-time signal sequence and the basic signal sequence, even if the internal state of the cabinet is exactly the same.

[0027] Therefore, in order to eliminate the impact of environmental condition differences on the difference calculation, environmental compensation correction will be performed on the real-time spatial signal sequence.

[0028] Therefore, in this embodiment of the application, before calculating the difference sequence based on the real-time spatial signal sequence and a preset spatial basic signal sequence, the following steps are also included: S210. Obtain the current environmental parameters of the immersion cabinet, as well as the baseline environmental parameters when the preset space basic signal sequence is collected.

[0029] S220. Based on the difference between the current environmental parameters and the reference environmental parameters, perform environmental compensation correction on the real-time spatial signal sequence.

[0030] First, obtain the current environmental parameters of the immersion cabinet, as well as the baseline environmental parameters when the preset space basic signal sequence is collected. These environmental parameters include at least one of the ambient temperature and ambient humidity.

[0031] Then, based on the difference between the current environmental parameters and the reference environmental parameters, environmental compensation correction is performed on the real-time spatial signal sequence. Let the current environmental parameters be denoted as... The baseline environmental parameters are Therefore, environmental compensation correction for the real-time spatial signal sequence can be expressed as:

[0032] in, For the first Real-time sensed signal values ​​at each spatial sampling point after environmental compensation. The environmental coupling coefficient characterizes the change in the sensed signal caused by a unit change in environmental parameters. Determined through laboratory calibration, the compensated real-time spatial signal sequence and the preset basic spatial signal sequence are within the same environmental reference system, eliminating the interference of environmental parameter changes on the rack health status diagnosis. For ease of explanation, all subsequent real-time spatial signal sequences can be considered as those after environmental compensation correction.

[0033] By subtracting the signal values ​​at the same spatial location from the real-time spatial signal sequence, a difference sequence can be obtained, denoted as . , , indicating spatial location The offset of the current state relative to the health baseline state can eliminate spatial differences caused by the rack's own structure. For example, the temperature at the bottom of the rack is lower than the temperature at the top. If the sensing signals at different spatial locations are all set to the same warning threshold, it is easy to have false alarms or missed alarms.

[0034] Spatial difference calculation of the difference sequence yields the spatial gradient rate of change sequence. The second-order spatial difference, defined as the difference, can be expressed as:

[0035] The spatial gradient change rate sequence characterizes the degree of drastic change of the difference along the spatial direction. Under healthy conditions, the difference sequence changes smoothly and gradually, and the amplitude of the spatial gradient change rate is small. When there is a local fault inside, the spatial continuity of the physical field of the outer wall near the fault source is disrupted, and the gradient change rate at the corresponding position shows a significant peak.

[0036] Therefore, in this embodiment of the application, the spatial gradient change rate is used as the preliminary diagnostic basis. By comparing the spatial gradient change rate of each spatial location with the set threshold, it is determined whether there is an anomaly and the spatial location range where the anomaly is located.

[0037] Considering that the temperature and flow fields inside the submerged cabinet are naturally non-uniform and dynamically changing, the flow state and thermal field distribution of the internal coolant vary significantly under different load conditions. For example, under low load, the coolant flow rate is low, the temperature field inside the cabinet changes smoothly and gradually along the space, and the spatial fluctuation amplitude of the sensing signal on the outer wall of the cabinet is small; under high load, the flow rate increases, the flow field instability increases, and the spatial fluctuation amplitude of the sensing signal on the outer wall of the cabinet increases accordingly.

[0038] Therefore, if a fixed threshold is used to judge the abnormality of the gradient change rate, false alarms are likely to occur because the spatial gradient change rate is naturally high under the high-debt healthy state; and false alarms are likely to occur because the absolute value of the spatial gradient change rate is small in the early stage of low-load faults.

[0039] Therefore, in this embodiment of the application, the real-time total power consumption of the current cabinet will be obtained, and a spatial adaptive threshold will be generated based on the real-time total power consumption and the difference sequence. The spatial adaptive threshold here is a quantitative calibration of the maximum normal fluctuation range allowed by the spatial gradient change rate of the physical field on the outer surface of the healthy cabinet along the inspection path under the current working conditions.

[0040] Specifically, see Figure 2 Based on the real-time total power consumption and the difference sequence, a spatially adaptive threshold is generated, including the following steps: S410. Divide the difference sequence into sliding windows, calculate the local standard deviation of the difference data within each sliding window, and form a local standard deviation sequence.

[0041] S420. Sort the local standard deviation sequence by value and extract the preset quantile as the spatial benchmark standard deviation.

[0042] S430. Based on the real-time total power consumption, determine the sensitivity coefficient corresponding to the current load through a preset load-sensitivity mapping function.

[0043] S440. Generate a spatially adaptive threshold based on the baseline standard deviation and sensitivity coefficient.

[0044] First, the difference sequence is divided into overlapping segments using a sliding window. For example, the sliding window width is set to... That is, the number of differences included is With a step size of 1, the difference sequence can be divided into: , ,…, .

[0045] Then calculate the local standard deviation of the difference data within each window. This forms a local standard deviation sequence, denoted as m is the total number of windows. .

[0046] Then, the local standard deviation sequence is sorted in ascending order of value, and a preset quantile is extracted as the spatial benchmark standard deviation. The preset quantile here is a statistical parameter used to robustly extract the health fluctuation benchmark from the local standard deviation sequence. It is usually set to 0.5, which is the median. Because the abnormal region is spatially sparse, the median can better reflect the inherent spatial fluctuation level of the healthy region under the current operating conditions.

[0047] Then, based on the real-time total power consumption, the sensitivity coefficient corresponding to the current load is determined through a preset load-sensitivity mapping function. This sensitivity coefficient... , is a dimensionless magnification factor used to proportionally amplify the extracted spatial baseline standard deviation, generating a spatial adaptive threshold for determining whether the spatial rate of change is abnormal, i.e., the spatial adaptive threshold. , can be represented as: .

[0048] The preset load-sensitivity mapping function here is a monotonically decreasing function. The higher the real-time total power consumption, the smaller the corresponding sensitivity coefficient. For example, it can be set to... , here Where P is the low-load limit sensitivity coefficient, and P is the current real-time total power consumption. The rated total power consumption of the rack is given by α, which is the attenuation coefficient. This mapping function effectively mitigates the spatial reference standard deviation under high load conditions. The problem is that raising the threshold leads to a passive increase in the spatial adaptive threshold.

[0049] Once the spatial adaptive threshold is determined, the gradient change rate at each spatial location can be calculated based on the spatial gradient change rate sequence. With spatial adaptive threshold Compare and identify abnormal spatial location intervals. Then the position The locations marked as abnormal, and the consecutive locations marked as abnormal, constitute the spatial location interval of the abnormality.

[0050] An abnormal spatial location interval, as a preliminary screening result, indicates a suspected abnormality. Further investigation is needed to confirm the preliminary screening result, that is, to determine the abnormal diagnosis result based on the abnormal spatial location interval.

[0051] Specifically, see Figure 3 Based on the abnormal spatial location range, the abnormal diagnosis result is further determined, including the following steps: S510. Generate a fixed-point exploration command based on the abnormal spatial location interval.

[0052] S520. Acquire the sensor data sequence generated by the robot during the execution of the fixed-point exploration command, according to the preset sampling rate.

[0053] S530. Based on the sensing data sequence, extract the time-domain dynamic features, and based on the time-domain dynamic features, generate the abnormal diagnosis results of the abnormal spatial location interval.

[0054] First, based on the abnormal spatial location range, a fixed-point exploration command will be generated. This fixed-point exploration command is used to control the robot to move back and forth between the start and end points of the exploration range at a preset speed, and to perform a static pause for a preset duration each time it reaches the start and end points.

[0055] Based on the spatial location index of the endpoints of the abnormal spatial location interval, the exploration start point and exploration end point can be determined. The exploration start point, exploration end point, as well as the preset speed, number of reciprocating cycles, turnaround pause duration, and sampling frequency are encapsulated into a fixed-point exploration command.

[0056] Once the fixed-point exploration command is sent to the robot, and the robot responds and executes it, the sensor data sequence generated by the robot during the execution of the fixed-point exploration command can be obtained. The sensor data sequence includes sensor signal values ​​collected at fixed sampling time intervals, the robot's real-time spatial position, and motion direction indicators.

[0057] Then, based on the sensing data sequence, temporal dynamic features can be extracted. These temporal dynamic features include the reciprocating consistency coefficient and the response time constant. The reciprocating consistency coefficient reflects the degree of difference in the sensing signal value when passing through the same spatial position in the forward and reverse directions. The response time constant reflects the decay rate of the sensing signal value over time during the robot's motion reversal pause.

[0058] Specifically, see Figure 4 Based on the sensing data sequence, extract the time-domain dynamic features, including the following steps: S531. Based on the direction of motion, the sensing data sequence is divided into a forward passage sequence and a reverse passage sequence.

[0059] S532. Based on the forward pass sequence and the reverse pass sequence, calculate the reciprocating consistency coefficient between the forward sensing signal data and the reverse sensing signal data at the same spatial location.

[0060] S533. Obtain the data segment in the sensing data sequence that is in a static pause period, and extract the response time constant of the sensing signal data in the data segment as time changes.

[0061] First, based on the direction of motion, the sensing data sequence can be divided into a forward passage sequence and a reverse passage sequence.

[0062] It is worth noting that the reciprocating consistency coefficient must compare forward and reverse signals at the same spatial location. However, since the original acquisition is time-based, the spatial locations of the sampling points during forward and reverse passage are neither uniform nor aligned. Therefore, spatial interpolation is required to map them onto a unified spatial grid before point-by-point comparison can be performed. This involves dividing the physical space range corresponding to the exploration interval into a preset number of spatial grid points. Through spatial interpolation, the sensing signal data in the forward and reverse passage sequences are mapped onto the spatial grid points respectively.

[0063] In this way, based on the forward and reverse passage sequences, the reciprocating consistency coefficient between the forward and reverse sensing signal data at the same spatial location can be calculated. Let the reciprocating consistency coefficient be C, and it can be expressed as:

[0064] in, It is a forward sequence. This is a reverse sequence, where M is the number of discrete space grid points. For the first The sensing signal of each spatial grid point. The value of C ranges from [0,1]. The closer C is to 1, the higher the consistency between the forward and reverse signals, and the more likely the anomaly source is to have static attributes, such as fixed anomalies such as surface stains and structural defects. The lower C is, the more directional or time-varying the anomaly source is, such as flow change anomalies such as flow channel blockage, gas accumulation, and boiling anomalies.

[0065] Then, the data segment in the sensing data sequence that is in a static pause period is obtained, and the response time constant of the sensing signal data in the data segment as a function of time is extracted. This response time constant is denoted as... It can be obtained by calculating the data segment during the static pause using an exponential decay fitting model, which can be expressed as:

[0066] in, This represents the signal value acquired at time t after the pause. The signal value that eventually stabilizes. The initial signal value at the moment of stopping corresponds to the sampling time as follows. The response time constant can be calculated by solving the problem. , The smaller the value, the faster the signal response and the smaller the corresponding thermal inertia, which usually indicates a surface or shallow anomaly. The larger the value, the slower the signal response and the greater the corresponding thermal inertia, which usually indicates an internal heat source or deep structural anomaly.

[0067] After determining the time-domain dynamic characteristics, the anomaly diagnosis results for the abnormal spatial location interval can be generated based on these characteristics.

[0068] Specifically, based on the dynamic characteristics in the time domain, anomaly diagnosis results are generated for the spatial location intervals of anomalies, including the following steps: S534. Based on the current total power consumption of the rack, determine the health characteristic distribution corresponding to the current load through a preset constraint relationship model.

[0069] S535. Calculate the eigenvectors formed by the time-domain dynamic features, and their Mahalanobis distance and deviation direction angle under the health feature distribution.

[0070] S536. If the Mahalanobis distance exceeds the preset distance threshold, the abnormal dominant type is determined according to the deviation direction angle, and the abnormal diagnosis result of the abnormal spatial location interval is marked as an abnormal state, while the abnormal dominant type is output.

[0071] S537. If the Mahalanobis distance does not exceed the preset distance threshold, the abnormal diagnosis result of the abnormal spatial location interval will be marked as healthy.

[0072] The preset constraint relationship model characterizes the joint distribution law of time-domain dynamic characteristics under healthy conditions as a function of real-time total power consumption. This can be achieved by controlling the rack load within a preset range during the confirmed healthy operation of the immersion rack, acquiring samples of the reciprocating consistency coefficient and response time constant collected by the robot at each load level, and then establishing the load P in relation to the reciprocating consistency coefficient C and response time constant based on the statistical analysis of the samples at each load level. The joint probability distribution, with mean vector Covariance Matrix express, , Let be the mean of the cyclic consistency coefficient under load P. Let be the mean of the response time constant under load P.

[0073] Based on the current total power consumption of the rack, the health characteristic distribution corresponding to the current load, i.e., the distribution mean vector, can be determined through a preset constraint relationship model. Covariance Matrix .

[0074] Then, the eigenvector formed by the time-domain dynamic features is calculated, along with the Mahalanobis distance and deviation direction angle under the health feature distribution. The eigenvector formed by the time-domain dynamic features is... It can calculate the Mahalanobis distance between the eigenvector and the mean vector of the distribution. , can be represented as

[0075] Simultaneously, the deviation direction angle of the feature vector relative to the distribution mean vector can be calculated. , can be represented as: ,in and Standardized deviation vector The two components, , This represents the deviation in dimension C after whitening transformation. After whitening transformation Deviation in dimension. Standardized deviation vector. This is obtained through whitening transformation to eliminate the effects of differences in feature dimensions and correlation, and can be expressed as: .

[0076] If the Mahalanobis distance exceeds the preset distance threshold, it is determined that there is a real anomaly. At this time, the dominant type of anomaly can be determined based on the deviation direction angle. The core logic of determining the anomaly type based on the deviation direction angle is that anomalies with different physical mechanisms will push away from the health center along a specific direction in the feature space. That is, in the feature space formed by the cyclic consistency coefficient and the response time constant, the health center corresponding to the mean is the anchor point of the normal state. Different types of anomalies will produce different directional deviations in the feature space.

[0077] Therefore, the anomaly type can be determined by a preset angle-anomaly mapping database. This database stores the correspondence between different angle ranges and anomaly types. For example, the first angle range is [-30°, 30°], which corresponds to the flow anomaly type, characterized by a significant decrease in the reciprocating consistency coefficient while the response time constant is basically normal; the second angle range is [60°, 120°], which corresponds to the heat accumulation anomaly type, characterized by a significant increase in the response time constant while the reciprocating consistency coefficient is basically normal; and the third angle range is [150°, 210°], which corresponds to the gas lock / boiling anomaly type, characterized by a significant deviation of both the reciprocating consistency coefficient and the response time constant in the same direction.

[0078] Based on the deviation direction angle The location of the angle range can determine the corresponding anomaly type, which is the dominant anomaly type. If the deviation angle does not fall within the above angle range, that is, other angle ranges, it is determined to be an uncertain type, and corresponding prompt information can be output so that relevant maintenance personnel can pay attention.

[0079] After determining the dominant type of the anomaly, the diagnostic results of the anomaly in the spatial location range will be marked as an abnormal state, and the dominant type of the anomaly will be output so that a corresponding diagnostic report can be generated in the end.

[0080] If the Mahalanobis distance does not exceed the preset distance threshold, it indicates that the preliminary screening result is a false trigger. Therefore, the abnormal diagnosis result of the abnormal spatial location interval is marked as a healthy state.

[0081] In this embodiment of the application, after the abnormal spatial location interval is determined to be in a healthy state, the preset spatial basic signal sequence will be adaptively updated using the signal data collected this time. This is mainly because the preset spatial basic signal sequence was collected at the initial stage of the rack's operation or at a moment when it was confirmed to be healthy. It records the spatial distribution characteristics of the physical field on the outer surface of the rack at that time. However, as time goes by, even if the rack has no faults, some factors (such as coolant aging, dust accumulation on the heat exchange surface, server hardware replacement, sensor drift, etc.) can cause the physical field on the outer surface to change slowly and irreversibly.

[0082] Therefore, by updating the preset space basic signal sequence, it is equivalent to maintaining the recent health status as a reference, thus ensuring the continuous accuracy of the diagnostic system throughout the entire immersion cabinet lifecycle.

[0083] Specifically, after classifying the abnormal diagnostic results of abnormal spatial location intervals as healthy, the following steps are also included: S610. Using the abnormal spatial location interval as an index, extract basic signal data and real-time signal data from the preset spatial basic signal sequence and spatial real-time signal sequence, respectively.

[0084] S620 updates the basic signal data based on real-time signal data by pre-setting update weight coefficients.

[0085] First, using the abnormal spatial location interval as an index, basic signal data and real-time signal data are extracted from the preset spatial basic signal sequence and spatial real-time signal sequence, respectively.

[0086] Then, based on the real-time signal data, the basic signal data is updated by pre-setting update weight coefficients. The real-time signal data here, as described above, can be represented as follows: It's just the location index here. Located in an anomalous spatial location interval, denoted as For location index The original reference signal value at that location, For location index The updated reference signal value at that location, It can be represented as:

[0087] in, The preset update weight coefficient is used to control the update speed of the preset spatial base signal sequence. The value range is 0.85~0.95. A higher weight setting can make the update slow, that is, a one-time disturbance is not enough to significantly change the benchmark. Only repeated changes that are confirmed to be healthy will be gradually improved and adjusted to the benchmark.

[0088] In this embodiment of the application, in order to ensure the absolute safety of the benchmark update, a maximum allowable deviation threshold is also set to prevent erroneously judged abnormal data from affecting the health benchmark.

[0089] Therefore, before updating the basic signal data based on real-time signal data by pre-setting update weight coefficients, the following steps are also included: S710, Calculate the absolute value of the deviation between the real-time signal data and the basic signal data.

[0090] S720. If the absolute value of the deviation exceeds the preset maximum permissible deviation threshold, the basic signal data is updated using the preset maximum permissible deviation threshold as the update upper limit.

[0091] By calculating the absolute value of the deviation between real-time signal data and basic signal data, i.e. If the absolute value of the deviation exceeds the preset maximum allowable deviation threshold This indicates an abnormal jump in the real-time signal value. The base signal data will be updated using a preset maximum allowable deviation threshold as the update upper limit. In other words, a limited adjustment amount will be added to the base value. The update expression can be adjusted as follows:

[0092] in, For conforming functions, used to control the direction of update adjustment, if This indicates that the real-time value is higher than the base value and needs to be adjusted upwards. ;like This indicates that the real-time value is lower than the base value and needs to be adjusted downwards. .

[0093] The preset maximum allowable deviation threshold here can be determined based on the statistical distribution of signal fluctuations at the same location under healthy conditions. For example, during the healthy operation of the rack, the fluctuation range of the statistical difference sequence can be calculated, and the index of each spatial location under healthy conditions can be determined. The historical distribution; take the 99th percentile or 3 times the standard deviation of the distribution as the preset maximum allowable deviation threshold.

[0094] This application also provides an adaptive diagnostic system for immersion cabinets based on robot active perception, see [link to relevant documentation]. Figure 5 The system includes: a data acquisition module 101, a difference sequence calculation module 102, a spatial difference calculation module 103, an adaptive threshold generation module 104, and an anomaly spatial diagnosis module 105.

[0095] The data acquisition module 101 is used to acquire real-time perception data collected by the robot as it moves along the preset inspection path during the inspection process, and to form a real-time spatial signal sequence according to the spatial location of the acquisition point.

[0096] The difference sequence calculation module 102 is used to calculate the difference sequence based on the real-time spatial signal sequence and a preset spatial basic signal sequence.

[0097] The spatial difference calculation module 103 is used to perform spatial difference calculation on the difference sequence to obtain the spatial gradient rate of change sequence.

[0098] The adaptive threshold generation module 104 is used to obtain the real-time total power consumption of the current cabinet and generate a spatial adaptive threshold based on the real-time total power consumption and the difference sequence.

[0099] The abnormal spatial diagnosis module 105 is used to compare the gradient change rate of each spatial location with the spatial adaptive threshold based on the spatial gradient change rate sequence, identify abnormal spatial location intervals, and further determine the abnormal diagnosis results based on the abnormal spatial location intervals.

[0100] In this embodiment of the application, the data acquisition module 101 is specifically used to acquire real-time perception data collected by the robot as it moves along a preset inspection path during the inspection process, and to form a spatial real-time signal sequence according to the spatial location of the acquisition point.

[0101] The difference sequence calculation module 102 is specifically used to calculate the difference sequence based on the real-time spatial signal sequence obtained by the data acquisition module 101 and the preset spatial basic signal sequence. The preset spatial basic signal sequence is reference data that the robot collects and stores in advance when the cabinet is in a healthy state.

[0102] The spatial difference calculation module 103 is specifically used to perform spatial difference calculation on the difference sequence generated by the difference sequence calculation module 102 to obtain the spatial gradient change rate sequence.

[0103] The adaptive threshold generation module 104 is specifically used to obtain the real-time total power consumption of the current cabinet and generate a spatial adaptive threshold based on the real-time total power consumption and the difference sequence.

[0104] The abnormal spatial diagnosis module 105 is specifically used to compare the gradient change rate of each spatial location with the spatial adaptive threshold based on the spatial gradient change rate sequence, identify abnormal spatial location intervals, and further determine the abnormal diagnosis results based on the abnormal spatial location intervals.

[0105] This application also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above for a robot-actively-perceived adaptive diagnostic method for immersion cabinets.

[0106] The embodiments described in this application are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the principles of this application should be included within the scope of protection of this application.

Claims

1. A robot-driven adaptive diagnostic method for immersion cabinets, characterized in that, include: The robot acquires real-time perception data as it moves along a preset inspection path during the inspection process, and forms a real-time spatial signal sequence according to the spatial location of the acquisition points. Based on the real-time spatial signal sequence, the difference sequence is calculated by using a preset spatial basic signal sequence. The preset spatial basic signal sequence is reference data that the robot collects and stores in advance when the cabinet is in a healthy state. Spatial difference calculation is performed on the difference sequence to obtain the spatial gradient rate of change sequence; Obtain the real-time total power consumption of the current rack, and generate a space adaptive threshold based on the real-time total power consumption and the difference sequence; Based on the spatial gradient change rate sequence, the gradient change rate of each spatial location is compared with the spatial adaptive threshold to identify abnormal spatial location intervals, and the abnormal diagnosis results are further determined according to the abnormal spatial location intervals.

2. The adaptive diagnostic method for immersion cabinets based on robot active perception according to claim 1, characterized in that, Before calculating the difference sequence based on the real-time spatial signal sequence and a preset spatial basic signal sequence, the process includes: Acquire the current environmental parameters of the immersion cabinet, as well as the baseline environmental parameters when the preset space basic signal sequence is acquired; Environmental compensation correction is performed on the real-time spatial signal sequence based on the difference between the current environmental parameters and the reference environmental parameters.

3. The adaptive diagnostic method for immersion cabinets based on robot active perception according to claim 1, characterized in that, The step of generating a spatially adaptive threshold based on the real-time total power consumption and the difference sequence includes: Divide the difference sequence into sliding windows, calculate the local standard deviation of the difference data within each sliding window, and form a local standard deviation sequence; The local standard deviation sequence is sorted by value, and a preset quantile is extracted as the spatial benchmark standard deviation; Based on the real-time total power consumption, the sensitivity coefficient corresponding to the current load is determined by a preset load-sensitivity mapping function; A spatially adaptive threshold is generated based on the baseline standard deviation and sensitivity coefficient.

4. The adaptive diagnostic method for immersion cabinets based on robot active perception according to claim 1, characterized in that, The step of further determining the abnormal diagnostic result based on the abnormal spatial location interval includes: Based on the abnormal spatial location range, a fixed-point exploration command is generated. The fixed-point exploration command is used to control the robot to move back and forth between the start and end points of the exploration range at a preset speed, and to perform a static pause for a preset duration each time it reaches the start and end points. Acquire the sequence of sensory data generated by the robot during the execution of fixed-point exploration instructions, at a preset sampling rate; Based on the perceived data sequence, dynamic features in the time domain are extracted, and based on these dynamic features, anomaly diagnosis results for the abnormal spatial location intervals are generated.

5. The adaptive diagnostic method for immersion cabinets based on robot active perception according to claim 4, characterized in that, The sensing data sequence includes sensing signal values ​​collected at fixed sampling time intervals, the robot's real-time spatial position, and motion direction identifiers. The temporal dynamic features include a reciprocating consistency coefficient and a response time constant. Extracting the temporal dynamic features from the sensing data sequence includes: Based on the direction of motion, the sensing data sequence is divided into a forward passage sequence and a reverse passage sequence; Based on the forward and reverse passage sequences, calculate the reciprocating consistency coefficient between the forward and reverse sensing signal data at the same spatial location; Obtain the data segment in the static pause period of the sensing data sequence, and extract the response time constant of the sensing signal data in the data segment as time changes.

6. The adaptive diagnostic method for immersion cabinets based on robot active perception according to claim 4, characterized in that, The step of generating anomaly diagnosis results for anomaly spatial location intervals based on temporal dynamic characteristics includes: Based on the current total power consumption of the rack, the health characteristic distribution corresponding to the current load is determined through a preset constraint relationship model. The preset constraint relationship model characterizes the joint distribution law of the time-domain dynamic characteristics under the health state as the real-time total power consumption changes. Calculate the eigenvectors formed by the time-domain dynamic features, and their Mahalanobis distance and deviation direction angle under the health feature distribution; If the Mahalanobis distance exceeds the preset distance threshold, the abnormal dominant type is determined based on the deviation direction angle, and the abnormal diagnosis result of the abnormal spatial location interval is marked as an abnormal state, while the abnormal dominant type is output. If the Mahalanobis distance does not exceed the preset distance threshold, the abnormal diagnosis result of the abnormal spatial location interval will be marked as a healthy state.

7. The adaptive diagnostic method for an immersion cabinet based on robot active perception according to claim 6, characterized in that, After labeling the abnormal diagnostic results of the abnormal spatial location interval as a healthy state, the process includes: Using the abnormal spatial location interval as an index, basic signal data and real-time signal data are extracted from the preset spatial basic signal sequence and spatial real-time signal sequence, respectively; Based on real-time signal data, the basic signal data is updated by pre-setting update weight coefficients.

8. The adaptive diagnostic method for an immersion cabinet based on robot active perception according to claim 7, characterized in that, Before updating the basic signal data based on real-time signal data using preset update weight coefficients, the process also includes: Calculate the absolute value of the deviation between real-time signal data and basic signal data; If the absolute value of the deviation exceeds the preset maximum permissible deviation threshold, the basic signal data will be updated using the preset maximum permissible deviation threshold as the upper limit for the update.

9. A robot-driven adaptive diagnostic system for immersion cabinets, characterized in that, include: The data acquisition module (101) is used to acquire real-time perception data collected by the robot as it moves along the preset inspection path during the inspection process, and to form a spatial real-time signal sequence according to the spatial location of the acquisition point. The difference sequence calculation module (102) is used to calculate the difference sequence based on the real-time spatial signal sequence and the preset spatial basic signal sequence. The preset spatial basic signal sequence is reference data that the robot collects and stores in advance when the cabinet is in a healthy state. The spatial difference calculation module (103) is used to perform spatial difference calculation on the difference sequence to obtain the spatial gradient change rate sequence; An adaptive threshold generation module (104) is used to obtain the real-time total power consumption of the current cabinet and generate a spatial adaptive threshold based on the real-time total power consumption and the difference sequence. An abnormal spatial diagnosis module (105) is used to compare the gradient change rate of each spatial location with the spatial adaptive threshold based on the spatial gradient change rate sequence, identify abnormal spatial location intervals, and further determine the abnormal diagnosis results based on the abnormal spatial location intervals.

10. A computer-readable storage medium storing a computer program capable of being loaded by a processor and executing a robot-actively-perceived adaptive diagnostic method for an immersion cabinet as described in any one of claims 1 to 8.

Images