Method for dynamic generation of data center inspection strategy and continuous optimization of inspection path

Abstract

The present application relates to a data center inspection strategy dynamic generation and inspection path continuous optimization method, comprising the following steps: dynamic inspection information graph construction: obtaining the running state data of each device in the data center, financial business key performance indicator data, security event log data and inspection robot state data, and constructing a dynamic inspection information graph; wherein the dynamic inspection information graph takes the device as the node, the feasible channel between the devices as the edge, the node attribute includes the device health risk index and the financial business influence level factor, and the edge attribute includes the edge passing cost and the security area passing permission weight; the feasible channel refers to the channel with a width greater than the minimum width required for robot passing and not being temporarily blocked; the present application generates the inspection strategy dynamically and optimizes the path adaptively, thereby significantly improving the inspection efficiency and resource precise allocation ability.

Description

Technical Field

[0001] This invention relates to the field of inspection optimization, specifically to a method for dynamically generating inspection strategies and continuously optimizing inspection paths for data centers. Background Technology

[0002] As the core infrastructure of information systems, data centers centrally house computing, storage, and network equipment for various critical business operations. Their operational reliability directly impacts the continuity of the supported services and data security. With the continuous expansion of data center scale, the number of internal devices has surged, and the types of equipment have become increasingly diverse, with significant differences in the criticality of the services supported by different devices. Simultaneously, the complex internal network layout of data centers and the dynamic changes in equipment operating status due to multiple factors such as load fluctuations, environmental temperature and humidity, and hardware aging place higher demands on the intelligence, accuracy, and real-time performance of inspection systems.

[0003] Currently, data center inspection methods mainly include two types: manual scheduled inspections and automatic inspections based on fixed rules. Manual inspections rely on the experience and judgment of maintenance personnel, resulting in limited inspection frequency and coverage. They are also difficult to respond to sudden anomalies in real time, and the quality of inspections is greatly affected by the skill level and subjective state of the personnel. Automatic inspections based on fixed rules typically execute according to preset time periods or a fixed list of devices. They cannot adaptively adjust inspection priorities and paths based on dynamic changes in device operating status and the degree of business impact. This can easily lead to delays in the inspection of high-value, high-risk devices, while low-risk devices consume excessive inspection resources, resulting in low inspection efficiency and low resource utilization.

[0004] In terms of path planning, existing methods mostly use the shortest physical path or minimum travel time as optimization objectives, failing to incorporate the business importance of equipment, dynamic changes in health risks, and the energy constraints of the inspection robot itself into the path decision model. This makes it difficult to achieve optimal allocation of inspection resources in complex and ever-changing data center environments. Furthermore, when sudden security incidents or changes in channel status occur during inspection, existing systems typically need to re-plan the global path from the current location. This process is computationally time-consuming, has a slow response time, and lacks a smooth transition mechanism during path switching, affecting the continuity of inspection tasks and the quality of data acquisition.

[0005] To address the problems existing in current technologies, a method for dynamically generating data center inspection strategies and continuously optimizing inspection paths is proposed. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for dynamically generating data center inspection strategies and continuously optimizing inspection paths, wherein the control unit of the inspection robot executes the following steps:

[0007] Step S1: Construction of Dynamic Inspection Information Map: Obtain operational status data of each device in the data center, key performance indicator data of financial business, security event log data, and inspection robot status data to construct a dynamic inspection information map. The dynamic inspection information map uses devices as nodes and feasible channels between devices as edges. Node attributes include device health risk index and financial business impact level factor, while edge attributes include edge passage cost and security area access permission weight. A feasible channel refers to a channel with a width greater than the minimum width required for robot passage that is not temporarily blocked.

[0008] Step S2, Dynamic Inspection Demand Calculation: Based on the equipment health risk index, historical inspection time, and financial business impact level factor, calculate the dynamic inspection demand of each equipment node. The dynamic inspection demand integrates the health risk index time decay term weighted by the financial business impact level factor and the information entropy weight based on the uncertainty of the operating data.

[0009] Step S3: Initial Inspection Path Generation: Based on the dynamic inspection demand, combined with the current position of the inspection robot and the remaining energy constraints, a subset of equipment to be inspected is determined, and an initial inspection path is generated on the dynamic inspection information map using a heuristic search algorithm that integrates the demand potential field and the financial business impact level factor. Determining the subset of equipment to be inspected includes: including equipment with a dynamic inspection demand higher than a preset demand threshold in the subset, and estimating the total energy required for the subset. When the total energy required exceeds the robot's remaining energy, equipment is removed sequentially from low to high dynamic inspection demand until the energy constraint is met.

[0010] Step S4, Online Anomaly Response and Path Replanning: Control the inspection robot to move along the initial inspection path and detect abnormal events related to financial business security in real time; when an abnormal event is detected, update the health risk index and financial business impact level factor of the corresponding equipment according to the type and severity of the abnormal event, and then update the node and edge attributes of the dynamic inspection information graph. Starting from the current node of the robot, use an incremental search-based path optimization algorithm to replan the remaining inspection path online to obtain the optimized inspection path.

[0011] Furthermore, the dynamic inspection demand in step S2 is calculated through the following process:

[0012] Multiply the current health risk index of device i by the time decay factor expressed by the natural exponential function, and then multiply by the financial business impact level factor of the device to obtain the risk decay term. The exponential part of the time decay factor is composed of the negative value of the product of the time decay coefficient and the time difference calculated from the time of the last inspection completion. The financial business impact level factor is a positive number pre-calibrated based on the criticality of the financial business carried by the device.

[0013] Simultaneously, obtain the information entropy weighting factor characterizing the uncertainty of the operating parameters of device i;

[0014] Using the balance coefficient as the weight of the risk attenuation term, and subtracting the balance coefficient from 1 as the weight of the information entropy weight factor, the two are weighted and summed to obtain the dynamic inspection demand of device i at time t:

[0015] ;

[0016] in, The dynamic inspection requirement of device i at time t. The health risk index for device i. For the financial business impact level factor of device i, This refers to the time when device i last completed its inspection. It is a time decay coefficient and is a positive number. α is the information entropy weighting factor, and α is the balance coefficient with a value ranging from 0 to 1.

[0017] Furthermore, the information entropy weighting factor is obtained through the following process:

[0018] The data acquisition device i collects a sequence of multidimensional operating parameters within a sliding time window. The sliding time window has a preset fixed duration and sliding step size. The multidimensional operating parameter sequence includes financial transaction processing latency, transaction success rate, security event alarm frequency, and hardware temperature. The information entropy of each operating parameter in the multidimensional operating parameter sequence is calculated.

[0019] Based on preset weighting coefficients corresponding to the degree of impact of each operating parameter on the continuity of financial business, the information entropy of each operating parameter is weighted and summed to obtain the comprehensive information entropy. The calculation of the weighted sum is expressed as follows:

[0020] ;

[0021] Where M is the total dimension of the running parameters, The weight coefficients corresponding to the m-th dimension running parameters are... Let m be the information entropy of the m-th dimension operating parameter of device i within the sliding time window;

[0022] Obtain the lower and upper limits of the comprehensive information entropy statistically determined under the historical normal operating conditions of the device. Subtract the lower limit from the current comprehensive information entropy, and divide the difference by the difference between the upper and lower limits to obtain a ratio. This ratio is then processed by taking values ​​of 0 for values ​​less than 0 and 1 for values ​​greater than 1. The resulting value is the information entropy weighting factor.

[0023] ;

[0024] In the formula, Let i be the comprehensive information entropy of device i at time t. and These are the lower and upper limits of the device's overall information entropy, respectively.

[0025] Furthermore, the heuristic search algorithm that integrates the demand potential field and the financial business impact level factor in step S3 is an improved A* algorithm. Its evaluation function for node n consists of three parts: the actual path cost from the starting node to node n, the heuristic path cost from node n to the target area, and the gravitational potential field term jointly constructed by the dynamic inspection demand degree of the uninspected equipment node and the financial business impact level factor.

[0026] The gravitational potential field term is constructed as follows: For each device j in the subset of devices to be inspected, the dynamic inspection demand of device j is multiplied by its financial business impact level factor to obtain the weighted demand. Then, the ratio of this weighted demand to the shortest path distance from node n to device j plus a small positive number is calculated. The ratio of all devices is summed and the negative value is taken. The gravitational potential field term makes the inspection path more attractive to devices with high financial importance and high demand during the search process.

[0027] The specific forms of the evaluation function and the gravitational potential field term are as follows:

[0028] ;

[0029] ;

[0030] Where g(n) is the actual path cost from the starting node to node n, h(n) is the heuristic cost from node n to the target region, U(n) is the gravitational potential field term, and η is the potential field weight, which is positive. For the subset of equipment to be inspected, D j (t) represents the dynamic inspection demand of equipment j, L j Let d(n,j) be the financial business impact level factor of device j, and let d(n,j) be the shortest path distance from node n to device node j. It is a preset small positive number.

[0031] Furthermore, when calculating the gravitational potential field term U(n), for the subset of equipment to be inspected... For each device j in the process, perform the following processing:

[0032] The dynamic inspection requirement D of equipment j j (t) and the impact factor L of financial business j The product of the two values ​​is compared with the preset threshold θ.

[0033] If the product is greater than the key threshold θ, then proceed according to... Calculate the enhancement factor, where The preset positive enhancement adjustment coefficient;

[0034] If the product is not greater than θ, then take the enhancement coefficient. = 1;

[0035] Based on the enhancement coefficients corresponding to each device, the gravitational potential field term U(n) is calculated as a weighted sum of the enhancements from each device term:

[0036] .

[0037] Furthermore, in step S4, when an abnormal event is detected, the health risk index and financial business impact level factor of the corresponding device are updated, specifically including:

[0038] After detecting an abnormal event, the affected device i is identified, and the corresponding risk increment is determined according to the type of abnormal event. The types of abnormal events include abnormal fluctuations in financial transaction volume, security intrusion alarms, and compliance indicator exceedances. Each type of abnormal event corresponds to a preset risk increment value.

[0039] Using the time of occurrence of the abnormal event as a benchmark, calculate the time difference between the current time and the time of occurrence of the event. Multiply this time difference by the event impact attenuation coefficient and take the negative value as the exponent of the natural exponential function. Multiply the risk increment by the value of this exponential function to obtain the attenuated risk increment.

[0040] The updated health risk index is obtained by adding the health risk index of device i before the event occurs to the risk increment after the decay.

[0041] Based on the type of abnormal event, update the financial business impact factor of device i according to the preset financial business impact factor adjustment rules; the adjustment rules include: security intrusion alarm events will raise the financial business impact factor by two levels, compliance indicator exceedance events will raise it by one level, and abnormal transaction volume fluctuation events will adjust the factor according to the fluctuation amplitude in a linear relationship.

[0042] The updated health risk index is expressed as follows:

[0043] ;

[0044] in, The health risk index of device i before the event occurred. For the risk increment determined based on the type of abnormal event, The time when the event occurred. This is the event impact attenuation coefficient and is a positive number.

[0045] Furthermore, in step S4, the path optimization algorithm based on incremental search adopts the D* Lite algorithm; when the change in the node or edge attributes of the dynamic inspection information graph exceeds the preset trigger threshold, the current node of the robot is taken as the new starting point, and the remaining subset of the devices to be inspected is taken as the target set. The D* Lite algorithm is called to perform incremental replanning on the current path to generate an optimized inspection path.

[0046] The smooth transition during path switching is achieved through the following process: obtaining the pose of the robot's current path point and the pose of the access point on the optimized inspection path; within a preset transition time window, performing linear interpolation according to the time ratio to generate a reference pose sequence for a smooth transition from the current path to the optimized inspection path:

[0047] ;

[0048] in, The robot's current pose. To optimize the pose of access points along the inspection path, T is the length of the transition time window preset based on the robot's movement speed. This is the transition time parameter.

[0049] Furthermore, the edge travel cost in the dynamic inspection information graph is dynamically adjusted in the following ways:

[0050] Using the physical distance of the passage as the baseline, the real-time congestion coefficient of the passage is multiplied by the first adjustment coefficient and then 1 is added to obtain the congestion penalty factor; the real-time congestion coefficient of the passage is obtained from the real-time analysis of the number of people and obstacles in the passage by the data center video surveillance system.

[0051] Calculate the quotient of the robot's nominal energy divided by the current remaining energy, and then take the second adjustment coefficient raised to the power of the quotient to obtain the energy penalty factor; the nominal energy is the total energy of the robot when its battery is fully charged;

[0052] Obtain the access permission weight value of the safe area involved in the robot's entry from node i to node j. This weight value is determined according to the matching degree between the robot's security permission level and the target safe area. It is 1 when matching and a penalty value greater than 1 when not matching.

[0053] Multiplying the baseline value, congestion penalty factor, energy penalty factor, and safe zone access permission weight value, we obtain the edge travel cost from node i to node j at time t:

[0054] ;

[0055] in, Let be the travel cost of the edge between node i and node j at time t. The physical distance of the channel. This represents the real-time congestion coefficient of the passage. This represents the robot's current remaining energy. The robot's nominal energy. and The preset positive adjustment coefficient, Weights for access permissions in secure areas.

[0056] Furthermore, the method also includes:

[0057] After completing an inspection cycle, a loss function is constructed based on the actual equipment status anomalies, financial business interruption records, and missed inspections found during the inspection process. The inspection strategy parameters are then updated with the goal of minimizing this loss function.

[0058] The inspection cycle ends when all equipment in the subset of equipment to be inspected is inspected or when the robot’s remaining energy is lower than the minimum energy required to return to the charging station.

[0059] The process of constructing the loss function is as follows:

[0060] All equipment that was missed during this inspection and resulted in abnormal consequences were counted. The health risk index and the financial business impact level factor were calculated for each of these equipment. The products were then summed to obtain the penalty item for missed inspection risk.

[0061] The sum of the edge travel costs of all edges along the path traversed during this inspection is used as the inspection energy consumption cost item.

[0062] For each abnormal event that causes a financial business interruption, the product of the transaction loss per unit time caused by the event and the duration of the abnormality is used as the financial loss estimate for the event. The financial loss estimates of all related events are summed to obtain the potential loss item for the financial business interruption.

[0063] The total loss function is obtained by weighting and summing the penalty item for missed detection, the energy consumption cost item for inspection, and the potential loss item for financial business interruption using preset first, second, and third weights, respectively; the first, second, and third weights are preset positive numbers and their sum is 1.

[0064] ;

[0065] in, This is a collection of devices that fail to detect defects. Let this be the set of edges traversed by the inspection path. A set of abnormal events that could trigger an interruption of financial services. k Let t be the transaction loss per unit time corresponding to the k-th abnormal event. k The duration of the anomaly is defined by w1, w2, and w3, which are preset weights.

[0066] With the goal of minimizing the total loss, the balance coefficient α and the time decay coefficient are updated using gradient descent. And the potential field weight η, and the updated balance coefficient α and time decay coefficient The potential field weight η is used in the next inspection cycle to achieve adaptive optimization of the inspection strategy parameters;

[0067] Gradient descent uses a preset learning rate and maximum number of iterations until the loss function converges.

[0068] The present invention has the following advantages over the prior art:

[0069] This case addresses the extremely high requirements for business continuity and security in financial and banking data centers. By constructing a dynamic inspection information graph that integrates equipment health risk indices and financial business impact level factors, inspection strategies are deeply linked to the criticality of financial business, enabling automatic allocation of inspection resources towards high-value, high-risk equipment. The dynamic inspection demand model weights and fuses the time decay characteristics of the health risk index with the uncertainty quantification based on multi-dimensional operating parameter information entropy. This allows inspection triggering to simultaneously respond to the gradual risks of equipment aging and the sudden anomalies of business fluctuations, significantly improving the early detection capability of potential faults such as abnormal financial transaction processing latency and frequent security authentication failures. The improved A* algorithm introduces a gravitational potential field term jointly constructed by the dynamic inspection demand degree of equipment and the financial business impact level factor in path search. Through a segmented processing mechanism using key thresholds and enhancement coefficients, it generates a stronger path attraction bias for high-demand, high-business-impact equipment, enabling inspection paths to proactively approach core financial equipment while satisfying energy constraints, effectively shortening the inspection response latency of critical equipment under abnormal conditions. When abnormal events such as unusual fluctuations in financial transaction volume, security intrusion alarms, or compliance indicator exceedances are detected, the device health risk index is updated using a time-decaying risk increment model. Based on preset financial business impact level factor adjustment rules, the inspection priority of affected devices is automatically upgraded. Simultaneously, incremental path replanning and smooth transition processing based on the D* Lite algorithm are triggered, allowing the robot to dynamically adjust the inspection sequence and path without interrupting the inspection task. This ensures that inspection resources can cover affected devices as quickly as possible after a financial security incident. The edge access cost model comprehensively considers the physical distance of the channel, real-time congestion level, remaining robot energy, and access permission weights for secure areas. This enables path planning to avoid congested areas and prevent the robot from interrupting its task due to energy depletion. It also adheres to access control policies for different security levels within the data center, ensuring the physical security and compliance of the inspection process itself. After the inspection cycle ends, a loss function is constructed based on the penalty for missed inspection, the energy consumption cost of inspection, and the potential loss of financial business interruption. The balance coefficient, time decay coefficient, and potential field weight are iteratively updated through gradient descent, so that the inspection strategy parameters can adaptively converge to the optimal configuration that minimizes business loss and inspection cost, thereby continuously optimizing the inspection effect under constantly changing business load and equipment status. Attached Figure Description

[0070] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0071] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0072] like Figure 1 As shown, the method for dynamically generating data center inspection strategies and continuously optimizing inspection paths involves the following steps executed by the control unit of the inspection robot:

[0073] Step S1: Construction of Dynamic Inspection Information Map: Obtain operational status data of each device in the data center, key performance indicator data of financial business, security event log data, and inspection robot status data to construct a dynamic inspection information map. The dynamic inspection information map uses devices as nodes and feasible channels between devices as edges. Node attributes include device health risk index and financial business impact level factor, while edge attributes include edge passage cost and security area access permission weight. A feasible channel refers to a channel with a width greater than the minimum width required for robot passage that is not temporarily blocked.

[0074] Step S2, Dynamic Inspection Demand Calculation: Based on the equipment health risk index, historical inspection time, and financial business impact level factor, calculate the dynamic inspection demand of each equipment node. The dynamic inspection demand integrates the health risk index time decay term weighted by the financial business impact level factor and the information entropy weight based on the uncertainty of the operating data.

[0075] Step S3: Initial Inspection Path Generation: Based on the dynamic inspection demand, combined with the current position of the inspection robot and the remaining energy constraints, a subset of equipment to be inspected is determined, and an initial inspection path is generated on the dynamic inspection information map using a heuristic search algorithm that integrates the demand potential field and the financial business impact level factor. Determining the subset of equipment to be inspected includes: including equipment with a dynamic inspection demand higher than a preset demand threshold in the subset, and estimating the total energy required for the subset. When the total energy required exceeds the robot's remaining energy, equipment is removed sequentially from low to high dynamic inspection demand until the energy constraint is met.

[0076] Step S4, Online Anomaly Response and Path Replanning: Control the inspection robot to move along the initial inspection path and detect abnormal events related to financial business security in real time; when an abnormal event is detected, update the health risk index and financial business impact level factor of the corresponding equipment according to the type and severity of the abnormal event, and then update the node and edge attributes of the dynamic inspection information graph. Starting from the current node of the robot, use an incremental search-based path optimization algorithm to replan the remaining inspection path online to obtain the optimized inspection path.

[0077] The dynamic inspection demand in step S2 is calculated through the following process:

[0078] Multiply the current health risk index of device i by the time decay factor expressed by the natural exponential function, and then multiply by the financial business impact level factor of the device to obtain the risk decay term. The exponential part of the time decay factor is composed of the negative value of the product of the time decay coefficient and the time difference calculated from the time of the last inspection completion. The financial business impact level factor is a positive number pre-calibrated based on the criticality of the financial business carried by the device.

[0079] Simultaneously, obtain the information entropy weighting factor characterizing the uncertainty of the operating parameters of device i;

[0080] Using the balance coefficient as the weight of the risk attenuation term, and subtracting the balance coefficient from 1 as the weight of the information entropy weight factor, the two are weighted and summed to obtain the dynamic inspection demand of device i at time t:

[0081] ;

[0082] in, The dynamic inspection requirement of device i at time t. The health risk index for device i. For the financial business impact level factor of device i, This refers to the time when device i last completed its inspection. It is a time decay coefficient and is a positive number. α is the information entropy weighting factor, and α is the balance coefficient with a value ranging from 0 to 1;

[0083] By weighting and integrating the time-cumulative growth characteristics of the health risk index with the information entropy weight that reflects operational uncertainty, the inspection demand reflects both the gradual accumulation of risks since the last inspection and the fluctuations and anomalies in the current operating status. This effectively avoids inspection delays or misjudgments caused by relying on a single indicator, making it particularly suitable for scenarios in financial and banking data centers where business continuity requirements are extremely high.

[0084] Suppose a bank's data center has two servers, A and B, that handle core payment transactions. The current time is 14:00, and the balance coefficient α is set to 0.6. Server A handles real-time clearing transactions, and the financial business impact factor L... A =3;

[0085] Server B handles the historical report query service, L B =1. The financial business impact level factor is a pre-defined positive integer based on the criticality of the financial business carried by the equipment, used to give higher-level equipment greater weight in demand calculations.

[0086] The current health risk index H(t) of both servers is 0.4, and the last inspection was completed at 10:00 AM for both servers. The time decay coefficient is... Set to 0.5 (unit: per hour).

[0087] After a 4-hour interval, the time risk accumulation factor is: .

[0088] The formula for calculating the risk attenuation term is as follows: The specific values ​​are as follows:

[0089] Risk attenuation factors for server A:

[0090] ;

[0091] Risk attenuation factor for server B:

[0092] ;

[0093] At 14:00, the multi-dimensional operating parameter sequence of servers A and B within the sliding time window is collected, including transaction processing latency, transaction success rate, security event alarm frequency and hardware temperature, totaling 4 dimensions, i.e., M=4.

[0094] For each dimension of the operating parameter, the parameter value range is first divided into several discrete intervals, and the frequency p of occurrence in each interval is counted. k Then, according to the discrete information entropy formula Calculate the real-time information entropy of this dimension parameter.

[0095] After obtaining the information entropy of each dimension, it is then weighted according to preset coefficients. The weighted summation yields the comprehensive information entropy. ,in satisfy Take equal weights, that is =0.25.

[0096] The calculation process is as follows: ;

[0097] Assuming server A experiences a surge in real-time clearing transaction volume, resulting in significant fluctuations in transaction latency, its four-dimensional information entropy, after weighting, yields the comprehensive information entropy S. A (t) = 0.85;

[0098] Server B is running stably, and the overall information entropy S is... B (t)=0.2.

[0099] The lower limit of the comprehensive information entropy, S, was statistically determined based on historical normal operating conditions for this server model. low =0.1, normal upper limit value S up =0.9.

[0100] Calculate the information entropy weighting factor using the linear mapping formula:

[0101] ;

[0102] Substituting, we get:

[0103] Server A:

[0104] ;

[0105] Server B:

[0106] ;

[0107] Finally, the calculation is performed using the weighted fusion formula based on the dynamic inspection demand level:

[0108] The dynamic inspection requirement of server A:

[0109]

[0110] =0.6226 + 0.375 = 0.9976;

[0111] The dynamic inspection requirement of server B:

[0112] ;

[0113] It is evident that although the static health risk indices and the time since the last inspection are exactly the same for both devices, server A, which handles higher-level financial transactions, has greater uncertainty in its current operational status, and exhibits a more significant cumulative risk effect over time, has an inspection requirement approximately 3.9 times that of server B. Based on this, the inspection robot prioritizes inspection tasks for server A, rather than distributing limited inspection resources equally or relying solely on fixed-cycle inspections, significantly improving the timeliness of fault detection for core business equipment.

[0114] Information entropy weighting factors are obtained through the following process:

[0115] The data acquisition device i collects a sequence of multidimensional operating parameters within a sliding time window. The sliding time window has a preset fixed duration and sliding step size. The multidimensional operating parameter sequence includes financial transaction processing latency, transaction success rate, security event alarm frequency, and hardware temperature. The information entropy of each operating parameter in the multidimensional operating parameter sequence is calculated.

[0116] Based on preset weighting coefficients corresponding to the degree of impact of each operating parameter on the continuity of financial business, the information entropy of each operating parameter is weighted and summed to obtain the comprehensive information entropy. The calculation of the weighted sum is expressed as follows:

[0117] ;

[0118] Where M is the total dimension of the running parameters, The weight coefficients corresponding to the m-th dimension running parameters are... Let m be the information entropy of the m-th dimension operating parameter of device i within the sliding time window;

[0119] Obtain the lower and upper limits of the comprehensive information entropy statistically determined under the historical normal operating conditions of the device. Subtract the lower limit from the current comprehensive information entropy, and divide the difference by the difference between the upper and lower limits to obtain a ratio. This ratio is then processed by taking values ​​of 0 for values ​​less than 0 and 1 for values ​​greater than 1. The resulting value is the information entropy weighting factor.

[0120] ;

[0121] In the formula, Let i be the comprehensive information entropy of device i at time t. and These are the lower and upper limits of the normal comprehensive information entropy of the device, respectively.

[0122] By collecting multi-dimensional operational parameters such as financial transaction processing latency, transaction success rate, security event alarm frequency, and hardware temperature, each parameter is divided into discrete intervals, and information entropy is calculated based on frequency distribution. The entire information entropy calculation uniformly adopts a base-2 logarithm, and historical upper and lower limit quantiles are statistically based on the same algorithm, ensuring system consistency. Entropy values ​​uniformly measure the fluctuation degree of parameters with different dimensions, avoiding distortion caused by directly weighting the original values ​​of different dimensions. Pre-set business influence weights are used for weighted fusion and linear mapping to generate information entropy weight factors, enabling the inspection demand to adaptively reflect the current operational disorder of the equipment. Simultaneously, by using upper and lower limits and amplitude limiting processing determined based on historical normal data quantile statistics, the factors are ensured to always remain between 0 and 1. This prevents the impact of abnormal extreme values ​​on demand calculations and clearly distinguishes between three states: highly stable, normal fluctuation, and significant disorder. This significantly improves the sensitivity of financial and banking data centers to early abnormal signs such as performance jitter and accumulated security risks.

[0123] The information entropy weight factor calculation for two core servers, A (real-time clearing) and B (report query), in a bank's data center at 14:00.

[0124] A multidimensional sequence of operating parameters is collected within a sliding time window. The sliding window has a fixed duration of 10 minutes and a sliding step size of 1 minute. Operating parameters are sampled every 5 seconds within the window, and a total of 4-dimensional parameters are collected (M=4).

[0125] Financial transaction processing latency (unit: milliseconds), transaction success rate (percentage), security event alarm frequency (times / minute), hardware temperature (degrees Celsius).

[0126] Weighting coefficients of each dimension parameter The preset values ​​based on the degree of business impact are: β1=0.3 (transaction processing latency), β2=0.3 (transaction success rate), β3=0.25 (security event alarm frequency), and β4=0.15 (hardware temperature), which meet the requirements. .

[0127] For each dimension parameter, a discrete interval is pre-divided according to its physical characteristics, and the frequency p of the sampled values ​​falling into each interval within the statistical window is used as the basis for the calculation. k According to the discrete information entropy formula Calculate the information entropy of this dimension (base 2, unit: bits).

[0128] The discrete intervals for each parameter are divided as follows: transaction processing latency is divided into three intervals: low latency (0-50ms), medium latency (50-200ms), and high latency (above 200ms); transaction success rate is divided into three intervals: low success rate (below 95%), normal success rate (95%-99%), and high success rate (above 99%); security event alarm frequency is divided into three intervals: zero alarms (0 times), low alarms (1-2 times), and high alarms (3 times and above); hardware temperature is divided into three intervals: normal (30-50℃), slightly high (50-70℃), and overheated (above 70℃).

[0129] The calculation process of server A (real-time settlement, malfunctioning):

[0130] Server A experienced a sudden surge in traffic, causing drastic fluctuations in all its parameters. Of the 100 sampling points for transaction processing latency, p1=0.1 occurred 10 times in the low-latency range, p2=0.2 occurred 20 times in the medium-latency range, and p3=0.7 occurred 70 times in the high-latency range. Calculate the latency entropy:

[0131] .

[0132] Transaction success rate sampling statistics: Low success rate occurred 15 times (p1=0.15), normal success rate occurred 60 times (p2=0.6), and high success rate occurred 25 times (p3=0.25). Calculate the success rate information entropy:

[0133] .

[0134] Security incident alarm frequency sampling statistics: 20 occurrences in the zero alarm interval, p1=0.2; 20 occurrences in the low alarm interval, p2=0.2; 60 occurrences in the high alarm interval, p3=0.6. Calculate the alarm information entropy:

[0135] .

[0136] Hardware temperature sampling statistics: 40 occurrences in the normal range (p1=0.4), 35 occurrences in the slightly high range (p2=0.35), and 25 occurrences in the overheated range (p3=0.25). Temperature entropy was calculated.

[0137] .

[0138] The overall information entropy of server A is obtained by weighted summation:

[0139] ;

[0140] The calculation process of server B (report query, stable operation):

[0141] Server B is running smoothly, with all parameters highly concentrated. Transaction processing latency: low latency interval 95 transactions p1=0.95, medium latency interval 5 transactions p2=0.05, high latency interval 0 transactions p3=0, according to the definition of information entropy. ,calculate:

[0142] .

[0143] Transaction success rate: High success rate range (98 trades, p1=0.98), Normal success rate range (2 trades, p2=0.02), Low success rate range (0 trades). Calculation:

[0144] .

[0145] Security incident alarm frequency: 100 alarms in the zero alarm interval, p1=1, information entropy .

[0146] Hardware temperature: within normal range, p1=1 100 times, information entropy .

[0147] The overall information entropy of server B is obtained by weighted summation:

[0148] ;

[0149] Statistical determination of the upper and lower limits of normal:

[0150] For this server model, the comprehensive information entropy was calculated window-by-window using the same method during a continuous 30-day period of normal operation (excluding data from known failure periods). This yielded a historical sequence of comprehensive information entropy. The full-text information entropy calculation consistently used a base-2 logarithm, and the historical lower and lower quantiles were statistically analyzed using the same algorithm, ensuring consistency. The 5th percentile of this sequence was taken as the normal lower limit value S. low =0.1, take the 95th percentile as the normal upper limit value S. up =0.9.

[0151] Linear mapping and amplitude limiting:

[0152] Substitute the combined information entropy of servers A and B into the formula respectively:

[0153] ;

[0154] Server A:S A (t) = 1.3295, mapping ratio ,through Limiting .

[0155] Server B: Mapping ratio ,through Limiting .

[0156] When server A experiences sudden traffic surges causing drastic fluctuations in various parameters, its overall information entropy... SA( t) = 1.3295 significantly exceeds the normal upper limit, and the information entropy weighting factor is clamped at 1, indicating that the operational uncertainty has reached the warning extreme value and immediate inspection is required; Server B is operating highly stably, and the comprehensive information entropy S B (t)=0.1283 is close to the lower limit of normal, and the factor is only 0.0354, which hardly generates any additional demand.

[0157] The significant difference between the two allows the inspection demand model to adaptively distinguish between equipment that is "highly disordered and requires priority attention" and equipment that is "highly stable and can be delayed in inspection," guiding inspection resources to be accurately allocated to the links with the highest continuity risks in financial business.

[0158] In step S3, the heuristic search algorithm that integrates the demand potential field and the financial business impact level factor is an improved A* algorithm. Its evaluation function for node n consists of three parts: the actual path cost from the starting node to node n, the heuristic path cost from node n to the target area, and the gravitational potential field term jointly constructed by the dynamic inspection demand degree of the uninspected equipment node and the financial business impact level factor.

[0159] The gravitational potential field term is constructed as follows: For each device j in the subset of devices to be inspected, the dynamic inspection demand of device j is multiplied by its financial business impact level factor to obtain the weighted demand. Then, the ratio of this weighted demand to the shortest path distance from node n to device j plus a small positive number is calculated. The ratio of all devices is summed and the negative value is taken. The gravitational potential field term makes the inspection path more attractive to devices with high financial importance and high demand during the search process.

[0160] The specific forms of the evaluation function and the gravitational potential field term are as follows:

[0161] ;

[0162] ;

[0163] Where g(n) is the actual path cost from the starting node to node n, h(n) is the heuristic cost from node n to the target region, U(n) is the gravitational potential field term, and η is the potential field weight, which is positive. For the subset of equipment to be inspected, D j (t) represents the dynamic inspection demand of equipment j, L j Let d(n,j) be the financial business impact level factor of device j, and let d(n,j) be the shortest path distance from node n to device node j. It is a preset, tiny positive number;

[0164] By introducing a gravitational potential field term, jointly constructed by the dynamic inspection demand of equipment and the impact level of financial business, into the evaluation function of the traditional A* algorithm, the evaluation of nodes during the path search process depends not only on the distance cost, but also on the potential field attraction of the business importance and urgency of the equipment to be inspected. This automatically tilts towards high-value, high-risk equipment in the inspection sequence and travel route, achieving dynamic matching between inspection resources and the needs of financial business protection.

[0165] Taking a local area of ​​a bank's data center as an example, the inspection robot is currently located at the starting point S, and the subset of equipment to be inspected... .

[0166] Device A is the core payment server, and its dynamic inspection requirement is D. A (t)=0.9, Financial Business Impact Level Factor L A =3, weighted demand ;

[0167] Device B is a log backup server, D B (t)=0.5, L B =1, weighted demand level .

[0168] The evaluation function of the improved A* algorithm is as follows: The potential field weight η = 1, a small positive number. =0.01 is used to prevent the denominator from being zero.

[0169] g(n) is the actual path cost from the starting point S to node n, and h(n) is the heuristic path cost from node n to the target area. In this example, h(n) is the straight-line distance from node n to the nearest device to be inspected.

[0170] To verify the guiding role of the gravitational potential field in path selection, two symmetrical candidate extension nodes X and Y are set near the starting point S. The planar coordinates of each node are set as follows: S(0,0), A(4,0), B(0,4), X(2,0), Y(0,2).

[0171] The distance from S to X is 2, and the distance from S to Y is also 2; the distance from X to device A is 2, the distance from Y to device B is 2, and the distance from X to B is... The distance from Y to A is also 4.4721.

[0172] As can be seen, under the traditional A* algorithm, the g values ​​of nodes X and Y are both 2, the h values ​​are both 2 (distance to the nearest device), and the total estimated value f is equal, so the algorithm needs to select them randomly.

[0173] After adding the gravitational potential term, calculate the potential fields of X and Y. .

[0174] For node X:

[0175] = -1.4549;

[0176] Then the valuation function value for node X is:

[0177] ;

[0178] For node Y:

[0179] ;

[0180] Then the evaluation function value for node Y is:

[0181] ;

[0182] The comparison shows that f(X) = 2.5451 < f(Y) = 3.1488. Therefore, the improved A* algorithm prioritizes expanding node X, driving the robot to move along the X direction, thus prioritizing its access to the core equipment A with higher weighted demand. This process demonstrates that even with a symmetrical physical distance layout, the gravitational potential field can be negatively enhanced, allowing nodes closer to high-value equipment to obtain smaller valuation functions. The inspection path proactively moves towards the equipment with the greatest risk to financial business continuity, significantly shortening the response latency of critical equipment under abnormal conditions and ensuring the stable operation of financial transactions.

[0183] When calculating the gravitational potential field term U(n), for the subset of equipment to be inspected... For each device j in the process, perform the following processing:

[0184] The dynamic inspection requirement D of equipment j j (t) and the impact factor L of financial businessj The product of the two values ​​is compared with the preset threshold θ.

[0185] If the product is greater than the key threshold θ, then proceed according to... Calculate the enhancement factor, where The preset positive enhancement adjustment coefficient;

[0186] If the product is not greater than θ, then take the enhancement coefficient. = 1;

[0187] Based on the enhancement coefficients corresponding to each device, the gravitational potential field term U(n) is calculated as a weighted sum of the enhancements from each device term:

[0188] };

[0189] By introducing a key threshold and an enhancement adjustment coefficient into the gravitational potential field, equipment whose product of dynamic inspection demand and financial business impact level exceeds the threshold is given an enhancement coefficient greater than 1. This allows the potential field contribution of key equipment to be amplified in a differentiated manner when multiple high-value equipment exist at the same time and have similar demand. This not only widens the gap in the valuation function, but also fundamentally reverses the path priority direction when the original potential field direction is ambiguous. This accurately directs inspection resources to the equipment that poses the greatest threat to business continuity, avoiding the path selection ambiguity caused by the superposition of potential fields of multiple equipment.

[0190] Taking a local area of ​​a bank's data center as an example, the inspection robot is currently located at the starting point S, and the subset of equipment to be inspected... .

[0191] Device A is the core transaction server, with a dynamic inspection requirement of D. A (t)=0.9, Financial Business Impact Level Factor L A =3, product D A L A =2.7;

[0192] Device B is a risk control decision server, D B (t)=0.8, L B =3, product D B L B =2.4;

[0193] Device C is a log backup server, D C (t)=0.5, L C =1, product D C L C =0.5.

[0194] Preset key threshold θ=2, enhance adjustment coefficient =0.5, a tiny positive number =0.01.

[0195] Enhancement coefficient The segmented calculation rule is: when hour, ;

[0196] when hour, .

[0197] Therefore .

[0198] Two symmetrical candidate nodes X and Y are set near the starting point S, with planar coordinates S(0,0), A(3,0), B(0,4), C(5,5), X(1,0), Y(0,2).

[0199] The distance from node X to A is d(X,A)=2, and the distance to B is... Distance to C ;

[0200] Distance from node Y to A The distance to B is d(Y,B)=2, and the distance to C is d(Y,C)= .

[0201] First, calculate the unenhanced standard gravitational potential field. .

[0202] Standard potential field of node X:

[0203] ;

[0204] Standard potential field of node Y:

[0205] ;

[0206] If we assume that the sum of the actual path cost from the starting point S to X and Y is equal to the sum of the heuristic cost, i.e. g(n) + h(n) = 5, and the potential field weight η = 1, then the evaluation function f(n) = 5 + U(n).

[0207] Standard potential field , .

[0208] at this time The algorithm prioritizes expanding node Y, with the inspection path biased towards the direction of device B. However, since device A is more important than device B, the standard potential field fails to reflect this difference, and the direction selection does not perfectly match the business priority.

[0209] After introducing the enhancement mechanism, the enhanced potential field is calculated. .

[0210] Enhanced potential field of node X:

[0211] ;

[0212] .

[0213] Enhanced potential field of node Y:

[0214] ;

[0215] .

[0216] Enhanced The algorithm prioritizes expanding node X, and the path direction is reversed to the direction of device A.

[0217] This reversal proof enhancement mechanism utilizes the higher enhancement coefficient of key equipment A to apply a stronger gravitational correction to its nearby nodes, causing the valuation function value of X to change from a relative disadvantage to an advantage under the standard potential field. This completely solves the directional ambiguity problem caused by the superposition of potential fields of multiple key equipment under the standard potential field, ensuring that inspection resources are accurately concentrated on the equipment with the highest business value and the most urgent risks, and maximizing the effect of financial business continuity protection.

[0218] In step S4, when an abnormal event is detected, the health risk index and financial business impact level factor of the corresponding device are updated, specifically including:

[0219] After detecting an abnormal event, the affected device i is identified, and the corresponding risk increment is determined according to the type of abnormal event. The types of abnormal events include abnormal fluctuations in financial transaction volume, security intrusion alarms, and compliance indicator exceedances. Each type of abnormal event corresponds to a preset risk increment value.

[0220] Using the time of occurrence of the abnormal event as a benchmark, calculate the time difference between the current time and the time of occurrence of the event. Multiply this time difference by the event impact attenuation coefficient and take the negative value as the exponent of the natural exponential function. Multiply the risk increment by the value of this exponential function to obtain the attenuated risk increment.

[0221] The updated health risk index is obtained by adding the health risk index of device i before the event occurs to the risk increment after the decay.

[0222] Based on the type of abnormal event, update the financial business impact factor of device i according to the preset financial business impact factor adjustment rules; the adjustment rules include: security intrusion alarm events will raise the financial business impact factor by two levels, compliance indicator exceedance events will raise it by one level, and abnormal transaction volume fluctuation events will adjust the factor according to the fluctuation amplitude in a linear relationship.

[0223] The updated health risk index is expressed as follows:

[0224] ;

[0225] in, The health risk index of device i before the event occurred. For the risk increment determined based on the type of abnormal event, The time when the event occurred. This is the event impact attenuation coefficient, and it is a positive number.

[0226] When abnormal events such as unusual fluctuations in financial transaction volume, security intrusion alarms, or compliance indicator exceedances are detected, the device health risk index is dynamically updated using a risk increment model with time decay. This ensures that the contribution of the event's impact on health risk gradually weakens over time, preventing a single event from causing a permanent increase in risk value or misjudgment. At the same time, based on the event type, preset financial business impact level factor adjustment rules are executed, automatically upgrading the inspection priority of affected devices. This ensures that inspection resources can be concentrated on devices with the highest security threats and business interruption risks as soon as possible, improving the response speed and accuracy of the financial data center to emergencies.

[0227] A security intrusion alert was detected on one of the online banking front-end servers in a bank's data center at 14:00.

[0228] Prior to the incident, the device's health risk index was... The impact factor for financial business is L=2.

[0229] Based on the pre-defined correspondence between abnormal event types and risk increments, the risk increments for various abnormal events are... Pre-defined criteria based on statistical analysis of the impact of historical events on business interruptions, including security intrusion alarms. =0.6, corresponding to abnormal fluctuations in financial trading volume =0.4, compliance indicator exceeds the standard. =0.3.

[0230] Event impact attenuation coefficient Its value is determined based on the timeliness requirements for data center security incident response, ensuring that the impact of the incident decays to below 5% of the initial value within approximately 15 minutes. The general formula for updating the health risk index over time is: , where t event This represents the time the event occurred. Current time t = 14:05, time interval tt event =5 minutes, then the risk increment after decay is calculated as follows:

[0231] ;

[0232] The updated health risk index is:

[0233] H(t) = 0.25 + 0.2207 = 0.4707

[0234] Meanwhile, the initial level range for the financial business impact factor is preset as follows: L=1 for general business equipment, L=2 for important business equipment, L=3 for core business equipment, and L=4 for extremely core business equipment. According to the preset adjustment rules, a security intrusion alarm event will raise the financial business impact factor by two levels, adjusting the equipment from L=2 to L=4; abnormal fluctuations in financial transaction volume will raise it by one level; and exceeding compliance indicators will raise it by one level.

[0235] If no new anomalies are detected after 14:05, the increment continues to decrease over time, until 14:10:

[0236] ;

[0237] The health risk index dropped to H=0.25+0.0812=0.3312, maintaining a record of recent security incidents while preventing the permanent accumulation of risk values. The financial business impact factor remained at its increased value of L=4 until manually reset after the inspection, ensuring that the equipment would have a higher weight in the next round of inspection demand calculations. This guides the inspection robot to prioritize inspecting equipment under security threats, effectively shortening security incident response time and preventing the risk of financial transaction interruption.

[0238] In step S4, the incremental search-based path optimization algorithm uses the D* Lite algorithm. When the change in the node or edge attributes of the dynamic inspection information graph exceeds the preset trigger threshold, the robot's current node is taken as the new starting point, and the remaining subset of devices to be inspected is taken as the target set. The D* Lite algorithm is then called to perform incremental replanning on the current path to generate an optimized inspection path.

[0239] The smooth transition during path switching is achieved through the following process: obtaining the pose of the robot's current path point and the pose of the access point on the optimized inspection path; within a preset transition time window, performing linear interpolation according to the time ratio to generate a reference pose sequence for a smooth transition from the current path to the optimized inspection path:

[0240] ;

[0241] in, The robot's current pose. To optimize the pose of access points along the inspection path, T is the length of the transition time window preset based on the robot's movement speed. For transition time parameters;

[0242] Employing the D* Lite incremental search algorithm, when abnormal events cause changes in the attributes of nodes or edges in the inspection graph, only local path updates are performed on the affected areas, eliminating the need to replan the global path from the starting point. This significantly reduces the computation time for online replanning, enabling the inspection robot to respond in milliseconds to sudden channel congestion or changes in security area permissions in the financial data center. Simultaneously, by generating smooth transition pose sequences through linear interpolation, the robot maintains continuous speed and stable movement when switching paths, avoiding blurred inspection images and equipment vibration caused by sudden turns or stops. This ensures that the financial data center meets the high requirements for inspection data acquisition quality and equipment operational stability.

[0243] Inside a bank's data center, an inspection robot is traveling along its initial path and is currently located at node N. cur The planar pose is P cur =(10, 20) (unit: meters), facing angle .

[0244] The remaining subset of devices to be inspected includes server D1 and storage array D2. In the original planned path, from node N... cur To the next node N next The cost of passage C ij =5.

[0245] Suddenly, the security system detected a change in the access permissions of the security zone on this side of the passage; the access permission weight W of the security zone on this side was changed. ij The change from 1 to 10 causes the edge travel cost to be instantly updated to C. ij =50.

[0246] The preset threshold for triggering attribute changes is 10, and the cost change amount is... = 45 > 10, immediately trigger path replanning.

[0247] The D* Lite algorithm uses the robot's current node N as an example. cur Starting from a new point of view, with the remaining subsets of devices to be inspected, D1 and D2, as the target set, an incremental search is performed based on the updated edge passage cost, and only the optimal path value of the affected nodes is recalculated to generate an optimized path.

[0248] Optimize the access point P of the path new =(18, 28), facing angle .

[0249] The preset transition time window T=3 seconds, the pose sequence of the robot smoothly switching from the current pose to the new path is generated by linear interpolation:

[0250] for ,

[0251] ;

[0252] Substitute the values, when =1.5 seconds, interpolation ratio ,but

[0253] P(1.5) = (10, 20) + 0.5 ×(8, 8) = (14, 24);

[0254] The orientation angle is also linearly interpolated to ensure a smooth transition in the movement direction. This smooth transition allows the robot to move smoothly from (10,20) to (18,28) along a straight line within 3 seconds, avoiding the impact of sudden turns. The entire process completes replanning and trajectory generation within tens of milliseconds, ensuring that the inspection task is not interrupted and that high-priority equipment is not missed due to temporary road closures, effectively improving the adaptability of the financial data center inspection system to dynamic environments.

[0255] The edge travel cost in the dynamic inspection information graph is dynamically adjusted in the following ways:

[0256] Using the physical distance of the passage as the baseline, the real-time congestion coefficient of the passage is multiplied by the first adjustment coefficient and then 1 is added to obtain the congestion penalty factor; the real-time congestion coefficient of the passage is obtained from the real-time analysis of the number of people and obstacles in the passage by the data center video surveillance system.

[0257] Calculate the quotient of the robot's nominal energy divided by the current remaining energy, and then take the second adjustment coefficient raised to the power of the quotient to obtain the energy penalty factor; the nominal energy is the total energy of the robot when its battery is fully charged;

[0258] Obtain the access permission weight value of the safe area involved in the robot's entry from node i to node j. This weight value is determined according to the matching degree between the robot's security permission level and the target safe area. It is 1 when matching and a penalty value greater than 1 when not matching.

[0259] Multiplying the baseline value, congestion penalty factor, energy penalty factor, and safe zone access permission weight value, we obtain the edge travel cost from node i to node j at time t:

[0260] ;

[0261] in, Let be the travel cost of the edge between node i and node j at time t. The physical distance of the channel. This represents the real-time congestion coefficient of the passage. This represents the robot's current remaining energy. The robot's nominal energy. and The preset positive adjustment coefficient, Weighting of access permissions for secure areas;

[0262] The edge access cost model integrates four factors: physical distance of the passage, real-time congestion level, remaining robot energy, and access rights to safe areas. It uses a congestion penalty factor to automatically avoid passages with dense crowds or many obstacles in path planning, an energy penalty factor to reduce the priority of long-distance or high-energy-consuming paths when the robot's energy is insufficient to ensure the sustainability of inspection tasks, and a security permission weight to prevent the robot from accidentally entering unauthorized high-security areas. Thus, it achieves global optimization of inspection paths in the complex and ever-changing physical and security environment of financial data centers, while meeting the requirements of business continuity, robot endurance, and compliance.

[0263] In a bank's data center, an inspection robot needs to travel from node i (currently located in the core transaction server area) to node j (the online banking front-end server area).

[0264] There are two selectable channels: Channel 1 physical distance =15 meters, which is the daily inspection passage, located in the general equipment area; the physical distance of passage 2 =10 meters, but it requires passing through a high-security encrypted server area, which is currently occupied by personnel and obstacles due to temporary maintenance.

[0265] The system obtains the congestion coefficient of Channel 1 through real-time analysis of video surveillance. =0.2 (only 1 employee passed through), congestion coefficient of passage 2 (t) = 0.8 (with 3 employees and construction obstacles). Congestion adjustment coefficient =2, energy regulation coefficient =1, and all the above parameters are fitted and calibrated based on the historical traffic data of the data center channel.

[0266] Robot nominal energy E nom =100 units (100 when fully charged), current remaining energy E remain (t) = 60 units, the energy penalty factor is a globally uniform value, reflecting the amplification effect of the robot's overall endurance pressure on the total path cost. The same value is used in the calculation of both channels, i.e. 1.6667. The access permission weight for safe areas is determined by the matching relationship between the robot's security permission level and the security level of the target area:

[0267] The robot's current security permission level is level 2, while the security level of the ordinary device area where channel 1 is located is level 1 (lower than the robot's permission). Matching is successful. =1; The security level of the encrypted server area where channel 2 is located is level 3 (higher than the robot's permissions), which does not match. =5 (preset mismatch penalty value).

[0268] The complete formula for calculating the cost of passage is:

[0269] ;

[0270] Substitute the parameters for the two channels respectively.

[0271] Channel 1 cost:

[0272] ;

[0273] Channel 2 cost:

[0274] ;

[0275] Although Channel 2 is physically shorter, its overall cost is approximately 6.2 times that of Channel 1 due to severe congestion and the lack of safe passage for robots in the area.

[0276] The inspection robot therefore chose channel 1, which avoided temporary congestion areas and complied with the compliance requirements for cross-security area access in the data center, effectively preventing security risks caused by unauthorized access. At the same time, the energy penalty factor has little impact on path selection when the robot has plenty of power, but its amplification effect gradually increases as the power continues to decrease, prompting the system to reserve the cost space for returning to recharge in advance in the global path planning, ensuring the sustainable execution of the inspection task.

[0277] The method also includes:

[0278] After completing an inspection cycle, a loss function is constructed based on the actual equipment status anomalies, financial business interruption records, and missed inspections found during the inspection process. The inspection strategy parameters are then updated with the goal of minimizing this loss function.

[0279] The inspection cycle ends when all equipment in the subset of equipment to be inspected is inspected or when the robot’s remaining energy is lower than the minimum energy required to return to the charging station.

[0280] The process of constructing the loss function is as follows:

[0281] All equipment that was missed during this inspection and resulted in abnormal consequences were counted. The health risk index and the financial business impact level factor were calculated for each of these equipment. The products were then summed to obtain the penalty item for missed inspection risk.

[0282] The sum of the edge travel costs of all edges along the path traversed during this inspection is used as the inspection energy consumption cost item.

[0283] For each abnormal event that causes a financial business interruption, the product of the transaction loss per unit time caused by the event and the duration of the abnormality is used as the financial loss estimate for the event. The financial loss estimates of all related events are summed to obtain the potential loss item for the financial business interruption.

[0284] The total loss function is obtained by weighting and summing the penalty item for missed detection, the energy consumption cost item for inspection, and the potential loss item for financial business interruption using preset first, second, and third weights, respectively; the first, second, and third weights are preset positive numbers and their sum is 1.

[0285] ;

[0286] in, This is a collection of devices that fail to detect defects. Let this be the set of edges traversed by the inspection path. A set of abnormal events that could trigger an interruption of financial services. k Let t be the transaction loss per unit time corresponding to the k-th abnormal event. k The duration of the anomaly is defined by w1, w2, and w3, which are preset weights.

[0287] With the goal of minimizing the total loss, the balance coefficient α and the time decay coefficient are updated using gradient descent. And the potential field weight η, and the updated balance coefficient α and time decay coefficient The potential field weight η is used in the next inspection cycle to achieve adaptive optimization of the inspection strategy parameters;

[0288] Gradient descent uses a preset learning rate and maximum number of iterations until the loss function converges.

[0289] After completing an inspection cycle, a loss function is constructed that integrates the penalties for missed inspections, the energy costs of inspections, and the potential losses from financial business interruptions. This function uses the feedback from actual business impacts and inspection costs to drive gradient descent optimization of the balance coefficient, time decay coefficient, and potential field weights. This allows the inspection strategy parameters to adaptively converge towards reducing business losses and conserving inspection energy. Consequently, in subsequent inspection cycles, this results in more accurate assessments of inspection demand, more reasonable path biases, and better resource allocation, continuously improving the financial data center inspection system's adaptability to dynamically changing business loads and equipment status.

[0290] After a bank's data center completes an inspection cycle, the system records the following actual data: the set of devices missed during inspection. It includes a report server with a health risk index H=0.4 and a financial business impact level factor L=2, resulting in a financial business interruption that was not detected in time; the set of edges traversed by the inspection path. It contains three edges, with edge travel costs of C1=35, C2=20, and C3=15 respectively. The total inspection energy consumption cost is... =70; A set of abnormal events that could trigger an interruption of financial services. Including a core trading system outage, the unit time trading loss is l = 5000 yuan / minute, the abnormal duration is t = 10 minutes, and the estimated financial loss is... = 50,000 yuan. Preset weights w1=0.5, w2=0.2, w3=0.3, satisfying w1+w2+w3=1.

[0291] The current strategy parameters are the values ​​from the previous iteration: balance coefficient α = 0.6, time decay coefficient... =0.5, potential field weight η=1.

[0292] Construct the loss function:

[0293] ;

[0294] Substitute the values:

[0295] ;

[0296] To minimize this loss function, the parameters are updated using gradient descent.

[0297] Set learning rate =0.01. Due to the loss function With parameter α, The explicit functional relationship between η and η is complex. In practice, the gradients of each parameter can be estimated using numerical methods (such as finite difference).

[0298] Assume that, based on calculations, the gradient at the current parameter point is approximately: .

[0299] Update rule based on gradient descent:

[0300] ;

[0301] ;

[0302] ;

[0303] The updated parameters will be used in the next inspection cycle.

[0304] For example, a slight decrease in the balance coefficient α means that the weight of the risk decay term will be slightly reduced in the next cycle, while the proportion of the information entropy weight will be slightly increased, thus enhancing the sensitivity to the uncertainty of operational fluctuations.

[0305] Time decay coefficient A slight increase makes the risk approach saturation more quickly over time;

[0306] The potential field weight η is slightly reduced to avoid excessive path bias caused by an overly strong potential field, which would increase unnecessary travel costs. If similar inspection tasks are repeatedly executed under the new parameters, the expected loss function value will gradually decrease. The inspection system achieves closed-loop learning of "evaluation-execution-feedback-optimization," continuously approaching the optimal inspection strategy and significantly improving the return on investment of financial data center inspections.

[0307] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for dynamically generating data center inspection strategies and continuously optimizing inspection paths, characterized in that, Includes the following steps: Step S1: Construction of Dynamic Inspection Information Map: Obtain operational status data of each device in the data center, key performance indicator data of financial business, security event log data, and inspection robot status data to construct a dynamic inspection information map. The dynamic inspection information map uses devices as nodes and feasible channels between devices as edges. Node attributes include device health risk index and financial business impact level factor, while edge attributes include edge passage cost and security area access permission weight. A feasible channel refers to a channel with a width greater than the minimum width required for robot passage that is not temporarily blocked. Step S2, Dynamic Inspection Demand Calculation: Based on the equipment health risk index, historical inspection time, and financial business impact level factor, calculate the dynamic inspection demand of each equipment node. The dynamic inspection demand integrates the health risk index time decay term weighted by the financial business impact level factor and the information entropy weight based on the uncertainty of the operating data. Step S3: Initial Inspection Path Generation: Based on the dynamic inspection demand, combined with the current position of the inspection robot and the remaining energy constraints, a subset of equipment to be inspected is determined, and an initial inspection path is generated on the dynamic inspection information map using a heuristic search algorithm that integrates the demand potential field and the financial business impact level factor. Determining the subset of equipment to be inspected includes: including equipment with a dynamic inspection demand higher than a preset demand threshold in the subset, and estimating the total energy required for the subset. When the total energy required exceeds the robot's remaining energy, equipment is removed sequentially from low to high dynamic inspection demand until the energy constraint is met. Step S4, Online Anomaly Response and Path Replanning: Control the inspection robot to move along the initial inspection path and detect abnormal events related to financial business security in real time; when an abnormal event is detected, update the health risk index and financial business impact level factor of the corresponding equipment according to the type and severity of the abnormal event, and then update the node and edge attributes of the dynamic inspection information graph. Starting from the current node of the robot, use an incremental search-based path optimization algorithm to replan the remaining inspection path online to obtain the optimized inspection path.

2. The method for dynamically generating data center inspection strategies and continuously optimizing inspection paths according to claim 1, characterized in that: The dynamic inspection demand in step S2 is calculated through the following process: Multiply the current health risk index of device i by the time decay factor expressed by the natural exponential function, and then multiply by the financial business impact level factor of the device to obtain the risk decay term. The exponential part of the time decay factor is composed of the negative value of the product of the time decay coefficient and the time difference calculated from the time of the last inspection completion. The financial business impact level factor is a positive number pre-calibrated based on the criticality of the financial business carried by the device. Simultaneously, obtain the information entropy weighting factor characterizing the uncertainty of the operating parameters of device i; Using the balance coefficient as the weight of the risk attenuation term, and subtracting the balance coefficient from 1 as the weight of the information entropy weight factor, the two are weighted and summed to obtain the dynamic inspection requirement of device i at time t.

3. The method for dynamically generating data center inspection strategies and continuously optimizing inspection paths according to claim 2, characterized in that: Information entropy weighting factors are obtained through the following process: The acquisition device i collects a sequence of multidimensional operating parameters within a sliding time window. The sliding time window has a preset fixed duration and sliding step size. The sequence of multidimensional operating parameters includes financial transaction processing latency, transaction success rate, security event alarm frequency, and hardware temperature. Calculate the information entropy of each dimension of the operating parameter in the multidimensional operating parameter sequence; According to the preset weight coefficients corresponding to the degree of impact of each dimension of operating parameters on the continuity of financial business, the information entropy of each dimension of operating parameters is weighted and summed to obtain the comprehensive information entropy; Obtain the normal lower limit and normal upper limit of the comprehensive information entropy statistically determined under the historical normal operating conditions of the device. Subtract the normal lower limit from the current comprehensive information entropy, divide the difference by the difference between the normal upper limit and the normal lower limit, and obtain a ratio. Apply the restriction that the ratio is 0 if it is less than 0 and 1 if it is greater than 1. The result is the information entropy weighting factor.

4. The method for dynamically generating data center inspection strategies and continuously optimizing inspection paths according to claim 3, characterized in that: In step S3, the heuristic search algorithm that integrates the demand potential field and the financial business impact level factor is an improved A* algorithm. Its evaluation function for node n consists of three parts: the actual path cost from the starting node to node n, the heuristic path cost from node n to the target area, and the gravitational potential field term jointly constructed by the dynamic inspection demand degree of the uninspected equipment node and the financial business impact level factor. The gravitational potential field term is constructed as follows: For each device j in the subset of devices to be inspected, the dynamic inspection demand of device j is multiplied by its financial business impact level factor to obtain the weighted demand. Then, the ratio of this weighted demand to the shortest path distance from node n to device j plus a small positive number is calculated. The sum of this ratio for all devices is then taken as a negative value, which is the gravitational potential field term. The gravitational potential field term makes the inspection path more attractive to devices with high financial importance and high demand during the search process.

5. The method for dynamically generating data center inspection strategies and continuously optimizing inspection paths according to claim 4, characterized in that: When calculating the gravitational potential field term, for each device j in the subset of devices to be inspected, the following processing is performed: Compare the product of the dynamic inspection demand of equipment j and the financial business impact level factor with the preset key threshold. If the product is greater than the key threshold, the enhancement coefficient will be calculated according to the preset calculation process; If the product is not greater than the key threshold, then the enhancement coefficient is set to 1; Based on the enhancement coefficients corresponding to each device, the gravitational potential field term U(n) is calculated as the weighted sum of the enhancements of each device term.

6. The method for dynamically generating data center inspection strategies and continuously optimizing inspection paths according to claim 5, characterized in that: In step S4, when an abnormal event is detected, the health risk index and financial business impact level factor of the corresponding device are updated, specifically including: After detecting an abnormal event, identify the affected device i and determine the corresponding risk increment based on the type of abnormal event; The types of abnormal events include abnormal fluctuations in financial transaction volume, security intrusion alarms, and exceeding compliance indicators. Each type of abnormal event corresponds to a preset risk increment value. Using the time of occurrence of the abnormal event as a benchmark, calculate the time difference between the current time and the time of occurrence of the event. Multiply this time difference by the event impact attenuation coefficient and take the negative value as the exponent of the natural exponential function. Multiply the risk increment by the value of this exponential function to obtain the attenuated risk increment. The updated health risk index is obtained by adding the health risk index of device i before the event occurs to the risk increment after the decay. Based on the type of abnormal event, update the financial business impact level factor of device i according to the preset financial business impact level factor adjustment rules; The adjusted rules include: security intrusion alarm events will raise the financial business impact factor by two levels; compliance indicator exceedance events will be raised by one level; and abnormal trading volume fluctuation events will have their level factor adjusted linearly based on the fluctuation amplitude.

7. The method for dynamically generating data center inspection strategies and continuously optimizing inspection paths according to claim 6, characterized in that: In step S4, the incremental search-based path optimization algorithm uses the D* Lite algorithm. When the change in the node or edge attributes of the dynamic inspection information graph exceeds the preset trigger threshold, the robot's current node is taken as the new starting point, and the remaining subset of devices to be inspected is taken as the target set. The D* Lite algorithm is then called to perform incremental replanning on the current path to generate an optimized inspection path. The smooth transition during path switching is achieved through the following process: obtaining the pose of the robot's current path point and the pose of the access point on the optimized inspection path, and performing linear interpolation according to the time ratio within a preset transition time window to generate a reference pose sequence for a smooth transition from the current path to the optimized inspection path.

8. The method for dynamically generating data center inspection strategies and continuously optimizing inspection paths according to claim 7, characterized in that: The edge travel cost in the dynamic inspection information graph is dynamically adjusted in the following ways: Using the physical distance of the passage as the baseline, the real-time congestion coefficient of the passage is multiplied by the first adjustment coefficient and then 1 is added to obtain the congestion penalty factor; the real-time congestion coefficient of the passage is obtained from the real-time analysis of the number of people and obstacles in the passage by the data center video surveillance system. Calculate the quotient of the robot's nominal energy divided by the current remaining energy, and then take the second power of the adjustment coefficient of the quotient to obtain the energy penalty factor; The nominal energy is the total energy of the robot when its battery is fully charged; Obtain the access permission weight value of the safe area involved in the robot's entry from node i to node j. This weight value is determined according to the matching degree between the robot's security permission level and the target safe area. It is 1 when matching and a penalty value greater than 1 when not matching. Multiply the baseline value, congestion penalty factor, energy penalty factor, and safe zone access permission weight value together to obtain the edge travel cost from node i to node j at time t.

Images