Server failure prediction sensor optimization deployment method based on edge computing

Abstract

The application discloses a server fault prediction sensor optimization deployment method based on edge calculation and relates to the technical field of server fault monitoring.The application comprises the following steps: an edge calculation node acquires a server structure, operation parameters and a type of fault to be monitored, establishes an airflow noise model and a fault signal propagation topology, forms a position pair on both sides of a structure propagation interface and selects a candidate position pair; an excitation is loaded for each fault in stages, a counteracting compensation amount is generated according to a position pair transmission difference and a path transmission parameter, an effective position pair is determined through virtual cancellation, a fault propagation front, a relay chain and a front fingerprint are formed; the front fingerprints of different faults are compared, an anchor position pair and a downstream position pair combination are determined, a relay confirmation group is formed; common position pairs are combined and redundant positions are virtually removed, and an initial scheme is obtained; and the application can improve the accuracy of fault propagation path judgment and fault differentiation and reduce sensor redundant deployment by using a measured coding correction model.

Description

Technical Field

[0001] This invention relates to the field of server fault monitoring technology, and in particular to an optimized deployment method for server fault prediction sensors based on edge computing. Background Technology

[0002] As data centers continue to expand, server workloads and equipment integration density continue to increase. Fans, hard drives, power supplies, and processor cooling components operate continuously for extended periods. To detect component wear, loosening, and abnormal vibrations in a timely manner, it is necessary to deploy sensors appropriately inside the server to collect and predict early fault signals, while ensuring the accuracy of fault identification while reducing the number of sensors.

[0003] Currently, server fault prediction sensor deployment technology typically selects sensor installation locations based on structural location, historical fault data, signal amplitude, signal-to-noise ratio, or monitoring coverage. It also analyzes the propagation of fault signals within the server structure using simulation models. However, servers contain issues such as airflow noise from cooling fans, local structural resonance, and multipath propagation. The same fault signal can experience amplitude attenuation and phase changes after passing through different structural interfaces, and some interfaces unrelated to fault propagation may also exhibit abnormal responses. Judging the propagation path based on the intensity or correlation of anomalies makes it difficult to verify whether the abnormal signal actually propagates through the candidate structural interface. This can easily lead to misinterpreting responses caused by airflow noise or local resonance as fault propagation signals, resulting in sensor deployment locations deviating from the actual fault propagation path.

[0004] Existing server fault prediction sensor deployment technologies cannot accurately verify the true propagation and reception relationship of early fault signals by the structural interface corresponding to the candidate location, resulting in insufficient fault differentiation capability and reliability of the sensor deployment scheme. To address this, this invention proposes an optimized deployment method for server fault prediction sensors based on edge computing. Summary of the Invention

[0005] The purpose of this invention is to provide an optimized deployment method for server fault prediction sensors based on edge computing, which solves the problem that existing server fault prediction sensor deployment technologies cannot accurately verify the true propagation and reception relationship of candidate location corresponding structural interfaces for early fault signals, resulting in insufficient fault differentiation capability and reliability of sensor deployment schemes.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0007] This invention provides an optimized deployment method for server fault prediction sensors based on edge computing, comprising the following steps:

[0008] S1. Obtain server structure, operating parameters and multiple fault types to be monitored from edge computing nodes, establish airflow noise model and fault signal propagation topology, form pre-selected position pairs on both sides of the structure propagation interface, and determine the position pairs with stable healthy transmission and abnormal early fault transmission as candidate position pairs.

[0009] S2. For each fault, apply excitation step by step and determine the new abnormal interface. Based on the transmission difference on both sides of the candidate position pair and the path transmission parameters from the downstream side of the candidate position pair to the new abnormal interface, generate the inverse compensation amount. Virtually apply it on the downstream side. Mark the position pair that makes the response attenuation of the new abnormal interface meet the cancellation condition as the effective position pair. Use the set of new abnormal interfaces under the same fault excitation level as the fault propagation front. Form a succession chain from the topologically continuous effective position pairs and form the front fingerprint according to the excitation level and topological order.

[0010] Furthermore, when determining the fault propagation front, the edge computing nodes check the effective location pair connections between each newly added abnormal interface and the corresponding faulty component. For newly added abnormal interfaces that can be connected to the corresponding faulty component through topologically continuous effective location pairs, they are identified as newly added abnormal interfaces that have been verified through propagation, and the fault propagation front is formed preferentially by newly added abnormal interfaces that have been verified through propagation. For newly added abnormal interfaces that have not formed continuous effective location pair connections, they are reserved as abnormal interfaces to be reviewed and are not used to form a successor chain.

[0011] S3. Compare the leading fingerprints to determine the common propagation segment and the first differentiation interface set. The candidate position pairs with the smallest topological distance from the first differentiation interface set within the common propagation segment are determined as anchor position pairs. The confirmation codes are formed by determining whether the downstream position pairs belong to the receiving chain. The downstream position pair combination that makes each fault code distinct and has the fewest number is selected and formed with the anchor position pairs to form a relay confirmation group.

[0012] S4. Merge the shared position pairs of each relay confirmation group, retain the anchor position pairs, and virtually remove the remaining position pairs in order of decreasing background noise based on the airflow noise model. If the codes are still different after removal and the relay chain is complete, delete them; otherwise, restore them to form the initial scheme.

[0013] S5. Install the sensor according to the initial plan and apply non-destructive verification excitation to form a measured connection chain and measured code. Correct the model and update the plan by the deviation between the measured and simulated codes. Verify until the code meets the consistency condition and the connection chain is complete, and output the target plan.

[0014] Furthermore, the edge computing nodes are set up in the local network where the target server is located, communicate with the server's data acquisition interface, and communicate with the sensors after the sensors are installed. They are used to perform airflow noise modeling, fault propagation simulation, generation of inverse compensation quantities, generation of confirmation codes, virtual removal of position pairs, and correction of measured models.

[0015] The fault signal propagation topology uses faulty components and structural propagation interfaces as nodes, and the structural connection relationships between adjacent nodes as associated edges. Each associated edge is configured with corresponding path transmission parameters, including amplitude attenuation parameters and phase delay parameters in the fault-sensitive frequency band. The airflow noise model is used to determine the background noise spectrum of each pre-selected location pair under different operating conditions.

[0016] Each preselected position pair includes a first preselected position located upstream of the corresponding structure propagation interface and a second preselected position located downstream. The amplitude ratio, phase difference and coherence coefficient between the first preselected position and the second preselected position are obtained to form the position pair transmission parameters. When the fluctuation of the transmission parameters in the healthy state does not exceed the stability threshold, and the deviation from the healthy transmission baseline in the early fault state exceeds the abnormal threshold, the corresponding preselected position pair is determined as a candidate position pair.

[0017] Furthermore, the fault excitation level is set incrementally based on the amplitude or energy of the excitation component within the corresponding fault characteristic frequency band; the structural propagation interface that did not meet the abnormal conditions under the previous fault excitation level but meets the abnormal conditions under the current fault excitation level is identified as a new abnormal interface.

[0018] The path propagation parameters from the downstream side of the candidate position pair to the newly added anomaly interface are obtained by concatenating the amplitude attenuation parameters and phase delay parameters of each associated edge between the two in topological order. When there is only one propagation path, the path propagation parameters of that propagation path are determined as the equivalent path propagation parameters. When there are multiple parallel propagation paths, the path propagation parameters of each propagation path are determined separately, and the equivalent path propagation parameters are synthesized according to the phase relationship when the response of each path arrives at the newly added anomaly interface.

[0019] When generating the inverse compensation quantity, the propagation difference between the two sides of the candidate position pair is projected onto the newly added anomaly interface through the equivalent path propagation parameter to obtain the predicted contribution response of the candidate position pair to the newly added anomaly interface. Based on the equivalent path propagation parameter, the amplitude and initial phase of the virtual loading signal are determined in reverse, so that the compensation response after the virtual loading signal propagates to the newly added anomaly interface is out of phase with the predicted contribution response. ;

[0020] The cancellation condition is determined based on the response energy attenuation rate within the fault-sensitive frequency band. satisfy:

[0021]

[0022] in, To compensate for the response energy of the newly added abnormal interface before virtual loading, To compensate for the virtual loading of the inverse compensation amount, the response energy of the newly added abnormal interface, when If the value is not less than the preset cancellation threshold, the corresponding candidate position pair will be marked as a valid position pair.

[0023] Furthermore, the succession chain consists of adjacent positions from the faulty component to the newly added abnormal interface, all of which are marked as valid; the front fingerprint includes the propagation front interface identifiers arranged according to the fault excitation level, the valid position pair identifiers in the succession chain and their topological arrangement order, and when there are multiple propagation branches at the same fault excitation level, the front sub-fingerprints corresponding to each propagation branch are formed respectively.

[0024] Furthermore, the common propagation segment is a continuous segment in which the successor chains of different faults have the same topological arrangement relationship from the direction of fault propagation. The initial differentiation interface set consists of the structural propagation interfaces corresponding to the first entry of the successor chains of different faults into different topological branches after the common propagation segment. The topological distance between the candidate position pair and the initial differentiation interface set is the minimum value among the topological distances from the candidate position pair to each structural propagation interface in the initial differentiation interface set. The anchoring position pair is the candidate position pair with the smallest topological distance to the initial differentiation interface set in the common propagation segment.

[0025] For a set of faults sharing the same common propagation segment, downstream candidate position pairs are arranged in a preset order. The state of the downstream candidate position pair belonging to the corresponding fault receiving chain is recorded as the first code value, and the state that does not belong to the corresponding fault receiving chain is recorded as the second code value. The confirmation codes of each fault are formed by splicing them according to the fault excitation level. The combination that makes the confirmation codes of the fault set different from each other and has the fewest number of position pairs is selected from the downstream candidate position pairs.

[0026] Furthermore, in step S4, candidate position pairs corresponding to the same location on the server structure are first merged into a shared position pair. Then, the background noise energy in the fault-sensitive frequency band is used as the sorting criterion to perform virtual removal of non-anchored position pairs from high to low. After each virtual removal, the confirmation code is regenerated and the coverage status of the takeover chain is checked. Only when each fault confirmation code is still different from each other, and at least one takeover path consisting of consecutive valid position pairs is retained between each anchored position pair and the corresponding propagation front, the virtually removed position pair is deleted.

[0027] Furthermore, the non-destructive verification excitation has an excitation frequency that matches the corresponding fault-sensitive frequency band, and the excitation amplitude is lower than the damage excitation threshold of the faulty component. Based on the measured responses of paired sensors located on both sides of the same structure propagation interface under the verification excitation, the measured position pairs are determined to have a transmission parameter and response attenuation relationship, and a measured transmission chain and measured code are formed. When there is a deviation between the measured code and the simulation code, the amplitude attenuation parameter and phase delay parameter of the corresponding associated edge are corrected according to the position of the position pair that caused the deviation in the fault signal propagation topology. The airflow noise model is also corrected according to the measured background response of each sensor.

[0028] Furthermore, in step S2, the airflow noise response of the upstream and downstream sides of the candidate position pair and the newly added abnormal interface under the current operating parameters is determined based on the airflow noise model. The response component whose frequency matches the order frequency of the cooling fan speed and whose coherence with the corresponding cooling fan operating parameters meets the preset coherence condition is determined as the same source airflow noise transmission component. The same source airflow noise transmission component is separated from the transmission difference on both sides of the candidate position pair to obtain the fault residual transmission difference. The corresponding airflow noise response is separated from the abnormal response of the newly added abnormal interface to obtain the fault residual response. An inverse compensation amount is generated based on the fault residual transmission difference and path transmission parameters. After virtual loading, when the attenuation of the fault residual response of the newly added abnormal interface meets the cancellation condition and the change in airflow noise response does not exceed the preset noise retention threshold, the corresponding candidate position pair is marked as a valid position pair.

[0029] The present invention has the following beneficial effects:

[0030] 1. This invention forms position pairs on both sides of the server structure propagation interface and filters candidate positions by utilizing changes in transmission parameters under healthy and early fault states. This allows sensor deployment to no longer rely solely on single-point response amplitude or empirical positions, but can instead identify positions that are more sensitive to early fault propagation and have stable transmission under healthy states by combining structural connection relationships, thereby improving the reliability of candidate deployment positions.

[0031] 2. This invention loads fault excitation step by step and generates an inverse compensation amount based on the transmission difference between the two sides of the candidate position pair and the path transmission parameters to the newly added abnormal interface. It uses the response attenuation after virtual loading to verify whether the candidate position pair actually participates in the propagation of the fault signal, thereby reducing the interference caused by local resonance, bypass propagation and background noise on the propagation path judgment, and improving the identification accuracy of effective position pairs and fault propagation front.

[0032] 3. This invention determines the common propagation segment, the first differentiation interface set, and the anchoring position pair by comparing the leading-edge fingerprints of different faults, and uses the downstream position pair to form an acknowledgment code, so that the anchoring position can confirm that the fault signal has reached the common propagation segment, and the downstream position can distinguish the specific propagation branch into which the fault signal enters, thereby reducing the number of redundant sensors while maintaining the distinguishability of fault types.

[0033] 4. This invention merges shared location pairs and virtually removes non-anchored locations according to background noise levels, deleting redundant locations while ensuring that the confirmation codes are distinct and the connection paths are continuous. After installation, non-destructive verification excitation is used to form a measured connection chain and measured codes, and the propagation model and airflow noise model are corrected accordingly, so that the final deployment scheme can adapt to the actual assembly differences and operating conditions of the server, improving the stability of fault prediction and the practicality of the deployment scheme. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is an overall flowchart of the edge computing-based server fault prediction sensor optimization deployment method of the present invention.

[0036] Figure 2 This is a flowchart illustrating the determination of effective position pairs and fault propagation fronts in this invention.

[0037] Figure 3 This is a flowchart of the relay confirmation, redundancy removal, and actual measurement correction process in this invention. Detailed Implementation

[0038] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0039] Example 1

[0040] This invention provides an optimized deployment method for server fault prediction sensors based on edge computing, comprising the following steps:

[0041] See Figure 1This embodiment uses a rack-mounted server as an example. The server includes a server chassis, cooling fan assembly, fan mounting frame, motherboard bracket, hard drive mounting cage, power module bracket, chassis base plate, and chassis side plate. The components are connected by screws, clips, guide rails, or support structures. The types of faults to be monitored include cooling fan bearing wear, cooling fan blade imbalance, cooling fan rotor eccentricity, loose fan mounting frame, hard drive spindle imbalance, and power module fan looseness. Cooling fan bearing wear, blade imbalance, and rotor eccentricity all originate from the cooling fan, but the fault-sensitive frequency bands and main excitation directions of the three are different. Therefore, they may enter different downstream propagation branches after passing through a common structural propagation section. In the early stage, the above faults can generate acoustic and vibration signals with specific frequency bands, amplitudes, and phase characteristics, and propagate outward through the structural connection positions inside the server.

[0042] In this embodiment, the sensor is an acoustic vibration sensor capable of collecting vibration amplitude and phase information within the fault-sensitive frequency band. Specifically, a triaxial accelerometer can be used. Before the sensor is officially installed, the pre-selected and candidate locations are virtual measurement points in the server structure model. Its health transmission data can be obtained through design verification data of the same model server, data collected by temporary test sensors, and structural simulation results.

[0043] S1. Obtain server structure, operating parameters and multiple fault types to be monitored from edge computing nodes, establish airflow noise model and fault signal propagation topology, form pre-selected position pairs on both sides of the structure propagation interface, and determine the position pairs with stable healthy transmission and abnormal early fault transmission as candidate position pairs.

[0044] The server structure includes the installation location, dimensions, material properties, connection methods, contact range, and connection rigidity of each component. The operating parameters include cooling fan speed, fan drive duty cycle, processor load, hard disk operating status, power supply output load, server intake air temperature, and chassis intake and exhaust air volume. The edge computing node retrieves the corresponding basic structure data according to the server model and uses the actual assembly information of the target server to correct the basic structure data.

[0045] Edge computing nodes are set up in the local network where the target server is located, communicate with the server's data acquisition interface, and communicate with the sensor after the sensor is installed. They are used to perform airflow noise modeling, fault propagation simulation, generation of inverse compensation quantity, generation of confirmation code, virtual removal of position pairs, and correction of measured models.

[0046] Before the sensors are installed, the edge computing nodes acquire fan speed, processor load, hard drive operating status, and temperature data through the server management controller and related data acquisition interfaces. After the sensors are installed, the edge computing nodes receive synchronous sampling data from each sensor. Two sensors at the same location use a unified sampling clock to reduce the impact of sampling time difference on the phase difference calculation results.

[0047] The fault signal propagation topology uses faulty components and structural propagation interfaces as nodes, and the structural connection relationships between adjacent nodes as associated edges. Each associated edge is configured with corresponding path transmission parameters, including amplitude attenuation parameters and phase delay parameters in the fault-sensitive frequency band. The airflow noise model is used to determine the background noise spectrum of each pre-selected location pair under different operating conditions.

[0048] When establishing the fault signal propagation topology, the cooling fans, hard drives, power module fans and their mounting components that may generate the fault to be monitored are set as fault component nodes; the connection positions between the fan body and the fan mounting frame, the fan mounting frame and the chassis base plate, the chassis base plate and the motherboard bracket, the chassis base plate and the chassis side plate, and the chassis base plate and the hard drive mounting cage are set as structural propagation interface nodes. When there is a connection relationship between adjacent nodes that can transmit structural acoustic and vibration signals, an associated edge is established between them.

[0049] The amplitude attenuation parameters and phase delay parameters of each associated edge are determined based on the component material, connection stiffness, connection length, and acoustic and vibration transmission data under healthy conditions. For multiple associated edges connected in sequence, the path transmission parameters of each associated edge are combined according to the topological order in which the fault signal actually passes through, so as to represent the total amplitude attenuation and total phase change generated when the fault signal propagates from one structural position to another structural position.

[0050] In this embodiment, the amplitude attenuation parameter is used to represent the degree of amplitude reduction of the acoustic signal after passing through the corresponding associated edge. Increasing the amplitude attenuation parameter indicates that the amplitude reduction of the acoustic signal after passing through the corresponding associated edge is greater, and decreasing the amplitude attenuation parameter indicates that the amplitude reduction of the acoustic signal after passing through the corresponding associated edge is less. The phase delay parameter is used to represent the degree of phase lag generated by the acoustic signal after passing through the corresponding associated edge.

[0051] The airflow noise model is established based on the calibration data of the same model of healthy server. During the calibration process, the server is placed under low load, medium load and high load states respectively, and the background noise spectrum of each pre-selected location is collected at multiple cooling fan speeds to form a correspondence between operating parameters and background noise spectrum. After the edge computing node obtains the current operating parameters of the target server, it selects data that matches the current operating state from the calibration data to obtain the background noise spectrum of each pre-selected location under the current operating conditions.

[0052] The noise represented by the airflow noise model includes the rotational frequency noise generated by the cooling fan, the blade passing frequency noise, and the broadband disturbance generated when the airflow passes through the motherboard bracket, hard drive mounting cage, and cable channel. This model is used to determine the background noise level of different locations in the corresponding fault-sensitive frequency band, providing a basis for subsequent candidate location screening and virtual removal.

[0053] Each preselected position pair includes a first preselected position located upstream of the corresponding structure propagation interface and a second preselected position located downstream. The amplitude ratio, phase difference and coherence coefficient between the first preselected position and the second preselected position are obtained to form the position pair transmission parameters. When the fluctuation of the transmission parameters in the healthy state does not exceed the stability threshold, and the deviation from the healthy transmission baseline in the early fault state exceeds the abnormal threshold, the corresponding preselected position pair is determined as a candidate position pair.

[0054] The upstream side of the structural propagation interface refers to the side closer to the faulty component along the current fault signal propagation direction, while the downstream side refers to the side farther away from the faulty component. When the same structural propagation interface has different propagation directions for different faults, the upstream and downstream sides are determined according to the location of the corresponding faulty component.

[0055] The first and second pre-selected positions are respectively set in regions with stable stiffness on both sides of the structural propagation interface, avoiding the edges of fastening holes, areas of suspended thin walls, and locations where it is inconvenient to install sensors. The edge computing node obtains the amplitude ratio, phase difference, and coherence coefficient in the fault-sensitive frequency band based on the response data of the first and second pre-selected positions, and uses all three as the position pair transmission parameters for the pre-selected position pair.

[0056] Healthy transmission stability refers to the fact that when the target server is under different normal operating loads and different normal fan speeds, the amplitude ratio and phase difference of the pre-selected position pair remain within a preset fluctuation range, and the responses on both sides have high coherence. In this embodiment, the fluctuation of the amplitude ratio not exceeding 10% of the healthy average, the change in phase difference not exceeding 15 degrees, and the coherence coefficient not less than 0.8 can be used as a way to judge healthy transmission stability. The specific threshold is determined according to the server structure, sensor accuracy, and fault-sensitive frequency band.

[0057] In the fault signal propagation model, an early fault excitation below the fault alarm level is loaded, and the position pair transmission parameters of each pre-selected position pair are obtained again. When the change of amplitude ratio, phase difference or coherence coefficient relative to the healthy transmission baseline reaches the preset abnormal condition, it is considered that the pre-selected position pair has a transmission abnormality in the early fault state. Only position pairs that simultaneously meet the requirements of healthy transmission stability and early fault transmission abnormality are determined as candidate position pairs, thereby excluding positions where the transmission relationship itself fluctuates greatly in the healthy state.

[0058] In this embodiment, the preset abnormal conditions can be set as follows: at least one of the following conditions must be met: the rate of change of amplitude ratio relative to the healthy transmission baseline is not less than 15%, the change of phase difference relative to the healthy transmission baseline is not less than 20 degrees, and the decrease of coherence coefficient relative to the healthy transmission baseline is not less than 0.15, and the corresponding change is maintained within two consecutive sampling windows; the above thresholds can be adjusted according to the server structure, fault-sensitive frequency band, sampling period and data acquisition accuracy.

[0059] In this embodiment, candidate position pairs P1 are formed on both sides of the structural propagation interface I1 between the cooling fan body and the fan mounting frame; candidate position pairs P2 are formed on both sides of the structural propagation interface I2 between the fan mounting frame and the chassis bottom plate; candidate position pairs P3 are formed on both sides of the structural propagation interface I3 between the chassis bottom plate and the motherboard bracket; candidate position pairs P4 are formed on both sides of the structural propagation interface I4 between the chassis bottom plate and the chassis side plate; candidate position pairs P5 are formed on both sides of the structural propagation interface I5 between the chassis side plate and the power module bracket; candidate position pairs P6 are formed on both sides of the structural propagation interface I6 between the chassis bottom plate and the hard drive mounting cage; and candidate position pairs P7 are formed on both sides of the structural propagation interface I7 between the hard drive mounting cage and the chassis front panel.

[0060] S2. For each fault, apply excitation step by step and determine the new abnormal interface. Based on the transmission difference on both sides of the candidate position pair and the path transmission parameters from the downstream side of the candidate position pair to the new abnormal interface, generate the inverse compensation amount. Virtually apply it on the downstream side. Mark the position pair that makes the response attenuation of the new abnormal interface meet the cancellation condition as the effective position pair. Use the set of new abnormal interfaces under the same fault excitation level as the fault propagation front. Form a succession chain from the topologically continuous effective position pairs and form the front fingerprint according to the excitation level and topological order.

[0061] See Figure 2 In step S2, the fault excitation and anti-phase compensation are loaded in the simulation model corresponding to the fault signal propagation topology. No destructive fault excitation is applied to the actual components of the target server. The edge computing node performs fault propagation simulation for each type of fault to be monitored and establishes the leading edge fingerprint corresponding to each fault.

[0062] The fault excitation level is set incrementally based on the amplitude or energy of the excitation component within the corresponding fault characteristic frequency band; the structural propagation interface that did not meet the abnormal conditions under the previous fault excitation level but meets the abnormal conditions under the current fault excitation level is identified as a new abnormal interface;

[0063] Taking the wear failure of the cooling fan bearing as an example, a virtual excitation matching the bearing failure characteristic frequency band is applied at the corresponding fan bearing position, and multiple fault excitation levels are set in order from weak to strong. The excitation intensity of each level increases progressively, with the lowest level corresponding to the early wear state and the highest level still below the complete failure state of the component.

[0064] Each time a fault stimulus level is loaded, the edge computing node recalculates the position pairs of the propagation interfaces of each structure and the transmission parameters. If a certain structure propagation interface did not meet the abnormal conditions in the previous fault stimulus level, but meets the abnormal conditions for the first time in the current fault stimulus level, the structure propagation interface is identified as a new abnormal interface for the current level. If multiple propagation branches at the same level simultaneously have new abnormal interfaces, they are collectively formed into a set of new abnormal interfaces for that level.

[0065] For a set of newly identified anomalous interfaces determined under the same fault excitation level, after completing the inverse compensation verification of each candidate location pair, the edge computing node further checks whether there is a connection relationship consisting of topologically continuous valid location pairs between the newly identified anomalous interface and the corresponding faulty component. When a connection relationship exists, it indicates that the anomalous response of the newly identified anomalous interface can be transmitted by the faulty component along the verified structural propagation path, and it is used as the fault propagation front under this fault excitation level; when no connection relationship exists, it indicates that the anomalous response of the newly identified anomalous interface may be caused by local resonance, bypass propagation, or model error, and it is marked as an anomalous interface to be reviewed, and this interface is not used to form the current fault's succession chain and front fingerprint.

[0066] In this invention, the transmission difference between the two sides of the candidate position pair is not obtained by directly subtracting the amplitude ratio, phase difference, and coherence coefficient. Instead, the transmission relationship of the position pair under healthy conditions is used as a baseline. Based on the upstream response under the current fault excitation, the predicted response that should be generated on the downstream side under healthy transmission relationship is determined. The difference between the actual response and the predicted response on the downstream side is determined as the transmission difference response. The transmission difference response retains the amplitude and phase information within the fault-sensitive frequency band and is used to generate the anti-phase compensation amount. The coherence coefficient is only used to determine whether the transmission relationship between the upstream and downstream responses is reliable and does not directly participate in the amplitude and phase calculation of the anti-phase compensation amount.

[0067] The path propagation parameters from the downstream side of the candidate position pair to the newly added anomaly interface are obtained by concatenating the amplitude attenuation parameters and phase delay parameters of each associated edge between the two in topological order. When there is only one propagation path, the path propagation parameters of that propagation path are determined as the equivalent path propagation parameters. When there are multiple parallel propagation paths, the path propagation parameters of each propagation path are determined separately, and the equivalent path propagation parameters are synthesized according to the phase relationship when the response of each path arrives at the newly added anomaly interface.

[0068] When there is only one propagation path between the downstream side of the candidate location pair and the newly added abnormal interface, the amplitude attenuation parameter and the phase delay parameter are combined according to the order of the associated edges on the path to obtain the equivalent path transmission parameter of the propagation path.

[0069] When there are multiple parallel propagation paths between the downstream side of the candidate location pair and the newly added abnormal interface, the edge computing nodes adjust the response amplitude according to the amplitude attenuation parameter of each associated edge in the topological arrangement order of the associated edges in each propagation path, and adjust the response phase according to the phase delay parameter to obtain the predicted response when each propagation path reaches the newly added abnormal interface. Then, each predicted response is converted to the same sampling direction and the same fault-sensitive frequency band, and synthesized according to the arrival phase. Predictive responses with similar arrival phases are enhanced and synthesized, and predictive responses with opposite arrival phases are canceled and synthesized. For propagation paths whose response amplitude after transmission is lower than the preset path threshold, they are marked as weak propagation paths and are not used to generate anti-phase compensation quantities separately. The synthesis results of the remaining propagation paths are used as the equivalent path transmission results from the downstream side of the candidate location pair to the newly added abnormal interface.

[0070] When generating the inverse compensation quantity, the edge computing node first determines the predicted contribution response formed after the transmission difference response propagates to the new abnormal interface through the equivalent path based on the transmission difference response of the candidate location pair and the transmission parameters of the equivalent path from its downstream side to the new abnormal interface. Then, it determines the initial amplitude of the virtual loading signal in reverse according to the amplitude attenuation degree of the equivalent path, and adjusts the initial phase of the virtual loading signal according to the phase delay of the equivalent path so that the frequency of the virtual loading signal is consistent with the frequency of the predicted contribution response and the phase is opposite when the virtual loading signal reaches the new abnormal interface. The virtual loading signal increases step by step from the preset minimum amplitude until the response of the new abnormal interface reaches the minimum value or the virtual loading signal reaches the preset upper limit of amplitude.

[0071] To avoid generating excessively large virtual loading signals when the equivalent path transmission capability is weak, edge computing nodes set a virtual loading amplitude upper limit. When the virtual loading signal determined in the reverse direction exceeds the amplitude upper limit, it is limited to the amplitude upper limit and cancellation verification is performed. If the cancellation condition cannot be met after the limitation, the corresponding candidate position pair is marked as a position pair to be verified.

[0072] It should be noted that the inverse compensation amount is only used as a verification signal for the fault propagation path and is virtually loaded in the simulation model. Its purpose is not to reduce the actual vibration or noise of the target server. The edge computing nodes determine whether the candidate position pairs actually participate in the propagation of the fault signal by observing the response changes of the newly added abnormal interfaces before and after the virtual loading.

[0073] The cancellation condition is determined based on the response energy attenuation rate within the fault-sensitive frequency band. satisfy:

[0074]

[0075] in, To compensate for the response energy of the newly added abnormal interface before virtual loading, For the response energy of the newly added abnormal interface after virtual loading of the inverse compensation amount, when the response energy attenuation rate is not less than the preset offset threshold, the corresponding candidate position pair is marked as a valid position pair according to the basic implementation method.

[0076] The response energy of the newly added abnormal interface is obtained by statistically analyzing the response spectrum within the fault-sensitive frequency band. The preset cancellation threshold is determined based on the fault propagation model error, background noise level, and sensor sampling accuracy. For example, the preset cancellation threshold can be set to 60%. When the response energy attenuation rate of the newly added abnormal interface within the fault-sensitive frequency band reaches or exceeds 60% after the inverse compensation is virtually applied, the corresponding candidate position pair is marked as a valid position pair.

[0077] To further eliminate the response reduction caused by simulation fluctuations, path superposition, and non-phase cancellation factors, in a preferred embodiment, after the edge computing node completes the marking of valid location pairs, it continues to perform comparison loading verification on the valid location pairs. During the comparison loading verification, the valid marking of the valid location pairs is temporarily retained and set to the verification state.

[0078] For valid position pairs in the verification state, the edge computing node generates a reference loading signal with the same initial amplitude and frequency as the inverse compensation amount, and adjusts the initial phase of the reference loading signal so that it remains in phase with the predicted contribution response after propagating to the newly added abnormal interface. The edge computing node loads the inverse compensation amount and the reference loading signal from the same simulation initial state, and compares the response of the newly added abnormal interface under the two loading states.

[0079] When the response energy attenuation rate corresponding to the inverse compensation amount reaches the preset cancellation threshold, while the response energy attenuation rate corresponding to the control loading signal does not reach the cancellation threshold, the valid mark of the corresponding candidate position pair is retained; when the response energy attenuation rates corresponding to both the inverse compensation amount and the control loading signal reach the preset cancellation threshold, it indicates that the response reduction may not be caused by the inverse cancellation effect, the valid mark of the corresponding candidate position pair is removed, and it is marked as a position pair to be reviewed and not used to form a succession chain; when the response energy attenuation rate corresponding to the inverse compensation amount does not reach the preset cancellation threshold, the valid mark of the corresponding candidate position pair is removed, and it is determined as an invalid position pair.

[0080] The edge computing node performs initial cancellation determination and comparison loading verification on each candidate location pair located upstream of the newly added abnormal interface. When multiple candidate location pairs pass the comparison loading verification, they are determined as valid location pairs respectively. When the valid location pairs are distributed in different propagation branches, the valid location pairs and their topological relationships in each propagation branch are retained according to the fault signal propagation topology.

[0081] The fault chain starts from the faulty component and connects adjacent valid position pairs in the direction of fault propagation until the corresponding newly added abnormal interface. Position pairs that are on the same geometric path but have not passed the offset verification do not participate in the formation of the fault chain. Thus, the fault chain represents the fault propagation relationship verified by the reverse compensation, rather than the propagation path inferred only from the server structure connection relationship.

[0082] In this embodiment, for the wear failure of the cooling fan bearing, under the first-level fault excitation, the structure propagation interface I1 first becomes abnormal, forming a connection chain from the faulty fan to the candidate position pair P1; under the second-level fault excitation, the structure propagation interface I2 first becomes abnormal, forming a connection chain that passes through the candidate position pairs P1 and P2 in sequence; under the third-level fault excitation, the structure propagation interface I3 first becomes abnormal, forming a connection chain that passes through the candidate position pairs P1, P2 and P3 in sequence. The edge computing node records the leading edge fingerprint of the fault according to the above excitation level and position pair topology order.

[0083] Record the newly added abnormal interface identifier, valid location pair identifier, and topological arrangement order of valid location pairs corresponding to each fault excitation level in sequence to form the leading edge fingerprint of the fault. When there are multiple propagation branches under the same level, record the leading edge sub-fingerprint of each branch respectively.

[0084] See Figure 3 S3. Compare the leading edge fingerprints to determine the common propagation segment and the first differentiation interface set. The candidate position pair with the smallest topological distance from the first differentiation interface set within the common propagation segment is determined as the anchor position pair. The confirmation code is formed by determining whether the downstream position pair belongs to the receiving chain. The downstream position pair combination that makes each fault code distinct and has the fewest number is selected and formed with the anchor position pair to form a relay confirmation group.

[0085] The succession chain consists of adjacent positions from the faulty component to the newly added abnormal interface, all of which are marked as valid. The front fingerprint includes the propagation front interface identifiers arranged according to the fault excitation level, the valid position pair identifiers in the succession chain and their topological arrangement order. When there are multiple propagation branches at the same fault excitation level, the front sub-fingerprints corresponding to each propagation branch are formed respectively.

[0086] Edge computing nodes align and compare the leading edge fingerprints of different faults from low to high according to the fault excitation level. During the comparison process, it is first determined whether each fault passes through the same effective position pair under the same excitation level, and then it is determined whether the arrangement order of each effective position pair in the topology is the same. Only segments with the same position pairs and continuous and consistent topological arrangement are identified as common propagation segments.

[0087] The common propagation segment is a continuous segment in which the successor chains of different faults have the same topological arrangement relationship from the direction of fault propagation; the first differentiation interface set consists of the structural propagation interfaces corresponding to the first entry of the successor chains of different faults into different topological branches after the common propagation segment.

[0088] The topological distance between a candidate location pair and the first differentiation interface set is the minimum of the topological distances from the candidate location pair to each structure propagation interface in the first differentiation interface set. In this embodiment, the topological distance is represented by the number of associated edges traversed between the candidate location pair and the corresponding structure propagation interface. The anchored location pair is the candidate location pair with the smallest topological distance to the first differentiation interface set in the common propagation segment. When multiple candidate location pairs have the same topological distance, the location pair with higher health transmission stability, lower background noise energy, and available sensor installation space is given priority.

[0089] In this embodiment, the cooling fan bearing wear fault is denoted as F1, the cooling fan blade imbalance fault as F2, and the cooling fan rotor eccentricity fault as F3. Based on the leading-edge fingerprints of the three faults under the same excitation level, all three faults sequentially pass through candidate position pairs P1 and P2. After passing through P2, due to the different fault sensitive frequency bands and main excitation directions, the receiving chain corresponding to F1 continues through P3 located on the motherboard bracket branch, the receiving chain corresponding to F2 continues through P4 located on the chassis side panel branch, and the receiving chain corresponding to F3 continues through P6 located on the hard drive mounting cage branch. Therefore, the continuous propagation segment between P1 and P2 is determined as the common propagation segment, and the structural propagation interfaces downstream of P2 corresponding to P3, P4, and P6 are jointly determined as the first differentiation interface set. Each structural propagation interface from P2 to the first differentiation interface set passes through only one associated edge, and P2 is the candidate position pair with the smallest topological distance to the first differentiation interface set in the common propagation segment. Therefore, P2 is determined as the anchoring position pair.

[0090] The anchor position is used to confirm that the fault propagation signal has passed through the common propagation section and is approaching the differentiation position, while the downstream position is used to confirm the specific propagation branch into which the fault signal has entered. This forms a relay relationship in which the anchor position first confirms the arrival of the fault front and then the downstream position confirms the fault type.

[0091] For a set of faults sharing the same common propagation segment, downstream candidate position pairs are arranged in a preset order. The state of the downstream candidate position pair belonging to the corresponding fault receiving chain is recorded as the first code value, and the state that does not belong to the corresponding fault receiving chain is recorded as the second code value. The confirmation codes of each fault are formed by splicing them according to the fault excitation level. The combination that makes the confirmation codes of the fault set different from each other and has the fewest number of position pairs is selected from the downstream candidate position pairs.

[0092] In this embodiment, the first encoding value is recorded as 1 and the second encoding value is recorded as 0. The downstream candidate position pairs are arranged according to their node numbers in the fault signal propagation topology and maintain the same order in different fault encoding processes. When there are multiple fault excitation levels, each level's encoding segment is formed and then spliced ​​together in order from low to high excitation level.

[0093] In this embodiment, the downstream candidate location pairs are P3, P4, and P6. The relay chain for fault F1 passes through P3, the relay chain for fault F2 passes through P4, and the relay chain for fault F3 passes through P6. When P3 is selected alone, the confirmation codes for F1, F2, and F3 are 1, 0, and 0, respectively, resulting in a code collision between F2 and F3. When P4 or P6 is selected alone, at least two fault codes are identical, therefore one downstream location pair cannot distinguish the three faults. After selecting P3 and P4, the confirmation codes for F1, F2, and F3 are 10, 01, and 00, respectively, and the three codes are distinct. Therefore, two downstream location pairs constitute the minimum number of location pairs that can distinguish the three faults. When other combinations with the same number of P3 and P6, P4 and P6, etc., exist, one of these combinations is selected based on background noise, health transmission stability, and installation conditions.

[0094] Edge computing nodes start encoding verification from combinations containing one downstream location pair. If it is not possible to make all fault codes distinct, the number of downstream location pairs is increased sequentially until a combination is obtained that makes each fault code different from the others. When there are multiple combinations with the same number of location pairs, the combination with lower background noise, higher health transmission stability and easier sensor installation is selected first.

[0095] The anchoring position pair P2 is combined with the selected downstream position pairs P3 and P4 to form a relay confirmation group for the corresponding fault set. Different fault sets form corresponding relay confirmation groups.

[0096] S4. Merge the shared position pairs of each relay confirmation group, retain the anchor position pairs, and virtually remove the remaining position pairs in order of decreasing background noise based on the airflow noise model. If the codes are still different after removal and the relay chain is complete, delete them; otherwise, restore them to form the initial scheme.

[0097] Different relay confirmation groups may contain location pairs located at the same physical installation location. If a location pair belongs to both the first and second relay confirmation groups, only one physical location pair will be retained during merging, and its corresponding set of all faults and confirmation codes will be recorded.

[0098] First, candidate position pairs corresponding to the same location on the server structure are merged into a shared position pair. Then, the background noise energy in the fault-sensitive frequency band is used as the sorting criterion, and the non-anchored position pairs are virtually removed from high to low. After each virtual removal, the confirmation code is regenerated and the coverage status of the takeover chain is checked. The virtually removed position pairs are deleted only when each fault confirmation code is still different from each other and at least one takeover path consisting of consecutive valid position pairs is retained between each anchored position pair and the corresponding propagation front.

[0099] Edge computing nodes obtain the background noise energy of each non-anchored position pair in the corresponding fault-sensitive frequency band based on the airflow noise model, and arrange them in order from high to low. They prioritize the virtual removal of position pairs with higher background noise, which helps to reduce the interference of airflow noise on fault signal acquisition after actual deployment.

[0100] Virtual removal refers to temporarily canceling the corresponding location pair in the current deployment combination without deleting the corresponding structural node and associated edge from the fault signal propagation topology. For each location pair removed, the edge computing node uses the remaining location pairs to reconstruct the confirmation code of each fault and checks whether there is still a continuous connection path between the anchored location pair and the corresponding fault propagation front.

[0101] If at least two fault confirmation codes are the same after removal, it means that the removed location pair still has the function of fault differentiation and the location pair should be restored. If there is no continuous relay path between the anchor location pair and the corresponding propagation front after removal, it means that the location pair has the necessary function of maintaining the relay confirmation relationship and the location pair should also be restored. Only when the codes are still different and the relay path remains continuous should the location pair be deleted from the deployment combination.

[0102] In addition to the relay confirmation groups corresponding to faults F1, F2 and F3, the relay confirmation group for the power module fan loosening fault also includes downstream position pair P5, and the relay confirmation group for the hard disk spindle imbalance fault also includes downstream position pair P6; after merging the relay confirmation groups, the anchored position pair P2 and the non-anchored position pairs P3, P4, P5 and P6 are all included in the set of positions to be removed.

[0103] In this embodiment, the merged deployment combination includes anchored position pair P2 and non-anchored position pairs P3, P4, P5 and P6. Among them, P5 has the highest background noise energy. After the virtual removal of P5, the fault confirmation codes are still different, and the anchored position pair P2 still retains a continuous receiving path to the front edge of each fault propagation. Therefore, P5 is deleted. After the virtual removal of P6 continues, if two faults form the same confirmation code, P6 is restored. The virtual removal test of all non-anchored position pairs is completed in the above manner to obtain the initial scheme.

[0104] The initial scheme includes the physical installation location of the reserved location pair, upstream and downstream side identification, sensor installation direction, sampling frequency, fault-sensitive frequency band, and the correspondence with the data channel of the edge computing node. In principle, each reserved location pair corresponds to two sensors installed on the upstream and downstream sides of the structure propagation interface respectively. When different location pairs contain the same physical measurement point, only one shared sensor is installed at the physical measurement point, and the total number of sensors is counted after the physical measurement points are merged.

[0105] S5. Install the sensor according to the initial plan and apply non-destructive verification excitation to form a measured connection chain and measured code. Correct the model and update the plan by the deviation between the measured and simulated codes. Verify until the code meets the consistency condition and the connection chain is complete, and output the target plan.

[0106] According to the initial plan, triaxial accelerometers are installed on the upstream and downstream sides of each reserved position pair. The two sensors in the same position pair are of the same model, the same installation method and the same coordinate direction. The sensor sampling frequency is higher than the highest frequency of the sensitive frequency band of the fault to be monitored, and synchronous sampling is adopted.

[0107] The non-destructive verification excitation has an excitation frequency that matches the corresponding fault-sensitive frequency band, and the excitation amplitude is lower than the damage excitation threshold of the faulty component. Based on the measured response of the paired sensors located on both sides of the same structure propagation interface under the verification excitation, the relationship between the measured position and the transmission parameters and response attenuation is determined, and a measured transmission chain and a measured code are formed. When there is a deviation between the measured code and the simulation code, the amplitude attenuation parameter and phase delay parameter of the corresponding associated edge are corrected according to the position of the position of the deviation in the fault signal propagation topology. The airflow noise model is also corrected according to the measured background response of each sensor.

[0108] Non-destructive verification excitation can be applied to the mounting base of the faulty component or adjacent support structure using a detachable miniature exciter. The frequency of the verification excitation covers the corresponding fault-sensitive frequency band, the excitation amplitude is lower than the component damage threshold, and it does not cause fastener loosening, bearing damage, hard drive read / write abnormalities, or server protection shutdown. For cooling fan-related faults, the fan speed can also be adjusted in small steps within the allowable range, and the controlled acoustic vibration response generated by the fan speed change can be used for verification.

[0109] Edge computing nodes acquire the measured responses of upstream and downstream sensors at each location, determine the measured amplitude ratio, measured phase difference, and measured coherence coefficient according to the method in S1, and then determine whether each location pair truly receives the verification signal based on the response changes of each location pair under the verification excitation. The topologically continuous measured valid location pairs are connected to form a measured receiving chain, and a measured code is formed according to the location pair order and encoding rules specified in S3.

[0110] When there is a deviation between the measured encoding and the simulated encoding, the edge computing node first identifies the location pairs that cause the encoding difference and determines the associated edges between the verification excitation position and the location pairs based on the fault signal propagation topology. It then sequentially checks each associated edge along the fault propagation direction from the verification excitation position, identifying the upstream associated edge adjacent to the first encoding difference location as the priority for correction. When the measured response amplitude is lower than the simulated response amplitude, it indicates that the actual amplitude reduction of the fault signal after passing through the corresponding associated edge is greater than the model prediction, and the amplitude attenuation parameter of that associated edge is increased. When the measured response amplitude is higher than the simulated response amplitude, it indicates that the actual amplitude reduction of the fault signal after passing through the corresponding associated edge is less than the model prediction, and the amplitude attenuation parameter of that associated edge is decreased. The phase delay parameter is adjusted according to the direction of advance or lag of the measured phase relative to the simulated phase. Only one associated edge is adjusted at a time, and the transmission parameters and verification encoding of each location pair are recalculated after adjustment. If the measured deviation of the location pair causing the encoding difference decreases, and no new encoding differences appear in other location pairs that already meet the consistency conditions, the adjustment is retained; otherwise, the original parameters of the associated edge are restored, and the next associated edge is checked sequentially.

[0111] The amplitude attenuation parameter and phase delay parameter are adjusted according to the preset step size, and the single adjustment limit, cumulative adjustment limit and maximum number of corrections are set; if the consistency condition is still not met after the maximum number of corrections is reached, the position pair is supplemented at the structure propagation interface where the first coding difference occurs, and the verification process is re-executed.

[0112] During normal operation without applying verification excitation, the edge computing node acquires the measured background response of each sensor and compares the measured background noise spectrum with the prediction results of the corresponding operating conditions of the airflow noise model. If the background noise at a certain location is consistently higher or lower than the prediction results, the measured data at that location is used to correct the background noise data of the corresponding operating conditions and frequency bands in the airflow noise model.

[0113] After completing the path transmission parameters and airflow noise model correction, S2 to S4 are executed again to update the fault propagation front, the relay chain, the confirmation code, and the initial scheme. If the updated scheme adds a location pair, the sensor is installed at the new location. If the updated scheme deletes the original location pair, the original sensor is retained until the next round of verification is completed. After confirming that the deletion can still maintain the distinct fault codes and the continuity of the relay path, the corresponding sensor is then removed.

[0114] During the retest, non-destructive verification excitation is applied again to re-form the measured transmission chain and measured code. When the measured code of each fault is consistent with the corresponding simulation code, and there is a continuous transmission path composed of measured effective position pairs between each anchor position pair and the corresponding fault propagation front, the model correction and scheme update are stopped.

[0115] The final output target scheme includes at least the number of sensors, the installation location of each sensor, the correspondence between location pairs, the sensor installation direction, the sampling frequency, the fault-sensitive frequency band, the anchor location pair identifier, the downstream confirmation location pair identifier, and the data channel configuration of the edge computing node. Through the above steps, while reducing the number of sensors, the retained sensor locations can cover the verified fault propagation and reception path, and maintain the distinguishability of the confirmation codes corresponding to different faults.

[0116] Example 2

[0117] In the above embodiments, the airflow noise generated by the server cooling fan has order frequencies and harmonic components corresponding to the fan speed. This airflow noise can propagate through structural paths such as the chassis side panel, fan bracket, and server base plate. When the airflow noise frequency band partially overlaps with the sensitive frequency band of the fault to be monitored, the transmission difference between the candidate position pairs simultaneously includes both fault propagation components and airflow noise propagation components. If an inverse compensation amount is directly generated based on the transmission difference, it may primarily cancel out the airflow noise and incorrectly identify a decrease in airflow noise response as a decrease in fault propagation response. This could lead to candidate position pairs unrelated to fault propagation being marked as valid position pairs, and consequently... The accuracy of the transfer chain and the leading edge fingerprint is affected. To solve this problem, in this embodiment, in step S2, the edge computing node determines the airflow noise response of the candidate location to the upstream side, downstream side and newly added abnormal interface under the current operating parameters based on the airflow noise model and the current cooling fan speed, fan drive duty cycle, server load and airflow. The corresponding speed order frequency is determined according to the cooling fan speed. The response component whose frequency is within the preset frequency range of the speed order frequency and its harmonics, and whose coherence with the cooling fan speed change sequence is not less than the preset coherence threshold, is determined as the same source airflow noise transmission component.

[0118] Edge computing nodes extract the same source airflow noise transmission components from the total response of the candidate location to both sides, and subtract the corresponding same source airflow noise transmission components from the transmission difference of the candidate location to both sides to obtain the fault residual transmission difference. At the same time, the same source airflow noise response predicted by the airflow noise model is subtracted from the abnormal response of the newly added abnormal interface to obtain the fault residual response of the newly added abnormal interface. The fault residual transmission difference retains the amplitude and phase information in the fault sensitive frequency band, which is used to characterize the actual transmission change of the fault signal when it passes through the candidate location to the corresponding structural propagation interface.

[0119] Based on the differences in fault residual propagation and the path propagation parameters between the candidate positions and the downstream side to the newly added abnormal interface, the predicted contribution of the candidate positions to the fault residual response of the newly added abnormal interface is determined. The amplitude and initial phase of the virtual loading signal are calculated in reverse so that the phase difference between the compensation response and the predicted contribution after the virtual loading signal propagates to the newly added abnormal interface is π, thereby forming an anti-phase compensation amount. The edge computing node virtually loads the anti-phase compensation amount on the downstream side of the candidate position pair, and determines the fault residual response energy and the same source airflow noise response energy before and after loading.

[0120] When the inverse compensation is virtually loaded, if the residual response energy attenuation rate of the newly added abnormal interface is not less than the preset cancellation threshold, and the response energy change rate in the frequency band dominated by the same source airflow noise is not greater than the preset noise retention threshold, the corresponding candidate position pair will be marked as a valid position pair.

[0121] In this embodiment, the preset cancellation threshold can be set to 60%, and the preset noise retention threshold can be set to 10%. When the residual response energy attenuation rate reaches the preset cancellation threshold and the change rate of the same source airflow noise response energy does not exceed the preset noise retention threshold, the corresponding candidate position pair is determined to pass the airflow noise exclusion verification. Conversely, when the total response is significantly reduced, but the reduction mainly comes from the same source airflow noise, the candidate position pair is not marked as a valid position pair.

[0122] Therefore, the inverse compensation verification is performed on the residual fault component after stripping the airflow noise, and the airflow noise response hold condition is used to constrain the cancellation result. This can avoid mistaking the cancellation of cooling fan airflow noise as the cancellation of fault signal propagation, and improve the accuracy of effective position pair marking, front propagation carry-over chain construction and fault propagation front fingerprint determination.

[0123] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing the deployment of server fault prediction sensors based on edge computing, characterized in that, Includes the following steps: S1. Obtain server structure, operating parameters and multiple fault types to be monitored from edge computing nodes, establish airflow noise model and fault signal propagation topology, form pre-selected position pairs on both sides of the structure propagation interface, and determine the position pairs with stable healthy transmission and abnormal early fault transmission as candidate position pairs. S2. For each fault, apply excitation step by step and determine the new abnormal interface. Based on the transmission difference on both sides of the candidate position pair and the path transmission parameters from the downstream side of the candidate position pair to the new abnormal interface, generate the inverse compensation amount. Virtually apply it on the downstream side. Mark the position pair that makes the response attenuation of the new abnormal interface meet the cancellation condition as the effective position pair. Use the set of new abnormal interfaces under the same fault excitation level as the fault propagation front. Form a succession chain from the topologically continuous effective position pairs and form the front fingerprint according to the excitation level and topological order. S3. Compare the leading fingerprints to determine the common propagation segment and the first differentiation interface set. The candidate position pairs with the smallest topological distance from the first differentiation interface set within the common propagation segment are determined as anchor position pairs. The confirmation codes are formed by determining whether the downstream position pairs belong to the receiving chain. The downstream position pair combination that makes each fault code distinct and has the fewest number is selected and formed with the anchor position pairs to form a relay confirmation group. S4. Merge the shared position pairs of each relay confirmation group, retain the anchor position pairs, and virtually remove the remaining position pairs in order of decreasing background noise based on the airflow noise model. If the codes are still different after removal and the relay chain is complete, delete them; otherwise, restore them to form the initial scheme. S5. Install the sensor according to the initial plan and apply non-destructive verification excitation to form the measured connection chain and measured code. Correct the model and update the plan according to the deviation between the measured and simulated codes. Verify until the code meets the consistency condition and the connection chain is complete, and output the target plan.

2. The method for optimizing the deployment of sensors for server fault prediction based on edge computing according to claim 1, characterized in that, The edge computing node is set up in the local network where the target server is located, communicates with the server's data acquisition interface, and communicates with the sensor after the sensor is installed. It is used to perform airflow noise modeling, fault propagation simulation, inverse compensation quantity generation, confirmation code generation, virtual removal of position pairs, and correction of the measured model. The fault signal propagation topology uses faulty components and structural propagation interfaces as nodes, and structural connection relationships between adjacent nodes as associated edges. Each associated edge is configured with corresponding path transmission parameters, which include amplitude attenuation parameters and phase delay parameters in the fault-sensitive frequency band. The airflow noise model is used to determine the background noise spectrum of each pre-selected location pair under different operating conditions. Each preselected position pair includes a first preselected position located upstream of the corresponding structure propagation interface and a second preselected position located downstream. The amplitude ratio, phase difference and coherence coefficient between the first preselected position and the second preselected position are obtained to form the position pair transmission parameters. When the fluctuation of the transmission parameters in the healthy state does not exceed the stability threshold, and the deviation from the healthy transmission baseline in the early fault state exceeds the abnormal threshold, the corresponding preselected position pair is determined as a candidate position pair.

3. The method for optimizing the deployment of sensors for server fault prediction based on edge computing according to claim 2, characterized in that, The fault excitation level is set incrementally based on the amplitude and energy of the excitation component within the corresponding fault characteristic frequency band; the structural propagation interface that did not meet the abnormal conditions under the previous fault excitation level but meets the abnormal conditions under the current fault excitation level is identified as a new abnormal interface; The path propagation parameters from the downstream side of the candidate position pair to the newly added anomaly interface are obtained by concatenating the amplitude attenuation parameters and phase delay parameters of each associated edge between the two in topological order. When there is only one propagation path, the path propagation parameters of that propagation path are determined as the equivalent path propagation parameters. When there are multiple parallel propagation paths, the path propagation parameters of each propagation path are determined separately, and the equivalent path propagation parameters are synthesized according to the phase relationship when the response of each path arrives at the newly added anomaly interface. When generating the inverse compensation amount, the transmission difference between the two sides of the candidate position pair is projected onto the newly added anomaly interface via the equivalent path transmission parameters to obtain the predicted contribution response of the candidate position pair to the newly added anomaly interface. Based on the equivalent path transmission parameters, the amplitude and initial phase of the virtual loading signal are determined in reverse, ensuring that the compensation response of the virtual loading signal after propagation to the newly added anomaly interface is phase-differentiated with the predicted contribution response. ; The cancellation condition is determined according to the response energy attenuation rate within the fault-sensitive frequency band, and the response energy attenuation rate... satisfy: in, To compensate for the response energy of newly added abnormal interfaces before virtual loading, To compensate for the virtual loading of the inverse compensation amount, the response energy of the newly added abnormal interface, when If the value is not less than the preset cancellation threshold, the corresponding candidate position pair will be marked as a valid position pair.

4. The method for optimizing the deployment of server fault prediction sensors based on edge computing according to claim 3, characterized in that, The receiving chain consists of adjacent position pairs from the faulty component to the newly added abnormal interface, all of which are marked as valid. The front fingerprint includes the propagation front interface identifiers arranged according to the fault excitation level, the valid position pair identifiers in the receiving chain and their topological arrangement order. When there are multiple propagation branches at the same fault excitation level, front sub-fingerprints corresponding to each propagation branch are formed respectively.

5. The method for optimizing the deployment of server fault prediction sensors based on edge computing according to claim 4, characterized in that, The common propagation segment is a continuous segment in which the successor chains of different faults have the same topological arrangement relationship from the direction of fault propagation. The first differentiation interface set consists of the structural propagation interfaces corresponding to the first entry of the successor chains of different faults into different topological branches after the common propagation segment. The topological distance between the candidate position pair and the first differentiation interface set is the minimum value among the topological distances from the candidate position pair to each structural propagation interface in the first differentiation interface set. The anchoring position pair is the candidate position pair with the smallest topological distance to the first differentiation interface set in the common propagation segment. For a set of faults sharing the same common propagation segment, downstream candidate position pairs are arranged in a preset order. The state of the downstream candidate position pair belonging to the corresponding fault receiving chain is recorded as the first code value, and the state of not belonging to the corresponding fault receiving chain is recorded as the second code value. The confirmation codes of each fault are formed by splicing them according to the fault excitation level. From the downstream candidate position pairs, the combination that makes the confirmation codes of the fault set different from each other and has the fewest number of position pairs is selected.

6. The method for optimizing the deployment of sensors for server fault prediction based on edge computing according to claim 5, characterized in that, In step S4, candidate position pairs corresponding to the same location on the server structure are first merged into a shared position pair. Then, the background noise energy in the fault-sensitive frequency band is used as the sorting criterion, and the non-anchored position pairs are virtually removed from high to low. After each virtual removal, the confirmation code is regenerated and the coverage status of the takeover chain is checked. The virtually removed position pairs are deleted only when each fault confirmation code is still different from each other and at least one takeover path consisting of consecutive valid position pairs is retained between each anchored position pair and the corresponding propagation front.

7. The method for optimizing the deployment of sensors for server fault prediction based on edge computing according to claim 6, characterized in that, The non-destructive verification excitation has an excitation frequency that matches the corresponding fault-sensitive frequency band, and the excitation amplitude is lower than the damage excitation threshold of the faulty component. Based on the measured responses of paired sensors located on both sides of the same structural propagation interface under the verification excitation, the measured position pairs are determined to transmit parameters and response attenuation relationships, and a measured transmission chain and measured code are formed. When there is a deviation between the measured code and the simulation code, the amplitude attenuation parameters and phase delay parameters of the corresponding associated edge are corrected according to the position of the position pair that caused the deviation in the fault signal propagation topology. The airflow noise model is also corrected based on the measured background response of each sensor.

8. The method for optimizing the deployment of server fault prediction sensors based on edge computing according to claim 1, characterized in that, In step S2, based on the airflow noise model, the airflow noise responses of the upstream and downstream sides of the candidate position pair and the newly added abnormal interface under the current operating parameters are determined. The response components whose frequencies match the order frequency of the cooling fan speed and whose coherence with the corresponding cooling fan operating parameters meets the preset coherence condition are identified as homogeneous airflow noise transmission components. The homogeneous airflow noise transmission components are extracted from the transmission differences on both sides of the candidate position pair to obtain the fault residual transmission difference. The corresponding airflow noise response is extracted from the abnormal response of the newly added abnormal interface to obtain the fault residual response. An inverse compensation amount is generated based on the fault residual transmission difference and path transmission parameters. After virtual loading, when the attenuation of the fault residual response of the newly added abnormal interface meets the cancellation condition and the change in airflow noise response does not exceed the preset noise retention threshold, the corresponding candidate position pair is marked as a valid position pair.

9. The method for optimizing the deployment of server fault prediction sensors based on edge computing according to claim 3, characterized in that, For candidate position pairs whose response energy decay rate is not less than a preset cancellation threshold, a reference loading signal with the same frequency and initial amplitude as the inverse compensation amount is generated, and the initial phase of the reference loading signal is adjusted so that the response after propagation to the newly added abnormal interface is in phase with the predicted contribution response; when the response energy decay rate corresponding to the inverse compensation amount is not less than the preset cancellation threshold, and the response energy decay rate corresponding to the reference loading signal is less than the preset cancellation threshold, the corresponding candidate position pair is confirmed as a valid position pair.

Images