Artificial intelligence reliability evaluation system based on secure and trusted algorithm

Abstract

The invention relates to the technical field of artificial intelligence reliability evaluation, in particular to an artificial intelligence reliability evaluation system based on a security and credibility algorithm, which comprises a data acquisition module, a credibility analysis module, a robustness evaluation module, an offset tracking module and an anomaly calibration module. The method comprises the following steps: eliminating noise interference and redundant data, constructing a stable feature sequence, extracting a state gradient and behavior fluctuation matching degree, adjusting a track in combination with an environmental factor, identifying a sudden change section, calculating a main shaft offset value, and finally marking an abnormal point to output an evaluation result. According to the method, the dynamic adaptive capacity in a complex scene can be improved, the behavior path backtracking association is enhanced, and the state recognition precision and the reliability evaluation effect of the artificial intelligence system are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence reliability evaluation, and particularly relates to an artificial intelligence reliability evaluation system based on a secure and reliable algorithm. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, its application in various fields is increasingly widespread, especially in complex scenarios, the reliability and security of artificial intelligence systems have gradually become the focus of research. Although the current technical solutions have improved the reliability of the system to some extent, there are still many deficiencies, which are difficult to fully meet the actual needs. For example, a patent with publication number CN113407428B proposes a "reliability evaluation method, device and computer equipment of artificial intelligence system", which collects log files, sensor data and real-time running data, and uses a non-homogeneous Poisson process to construct a cumulative failure number model to realize evaluation. However, this technical solution mainly relies on data-driven methods and lacks analysis of the reliability of the algorithm itself, failing to ensure the security of the system from the algorithm design level. In addition, the process of multi-source heterogeneous data fusion may affect the accuracy of the final evaluation results due to the uneven quality of the data, especially in the case of large data noise, the reliability of the evaluation results may be further reduced.

[0003] Another patent with publication number CN118364257B proposes a "model reliability analysis method, device, computer equipment and storage medium", which generates model fusion features by extracting and fusing multiple dimensional features of the model to be analyzed, and performs reliability analysis based on this. Although this method can comprehensively evaluate the reliability of the model, its analysis process mainly focuses on the input and output characteristics of the model, lacking in-depth consideration of the internal algorithm logic and security and reliability mechanisms of the model. At the same time, this solution does not fully consider the robustness of the model when facing adversarial attacks or abnormal scenarios, which may lead to inaccurate evaluation results in extreme cases, thereby affecting the overall reliability of the system.

[0004] The above problems show that the existing artificial intelligence reliability evaluation technology still has obvious deficiencies in algorithm reliability guarantee, data quality control and adaptability to adversarial scenarios. Especially in complex application scenarios, the reliability evaluation of artificial intelligence systems needs to consider the security of algorithm design, the accuracy of data processing and the robustness of the system under extreme conditions. Therefore, a new evaluation system is urgently needed, which can introduce security and reliability mechanisms at the algorithm design level, combine multi-dimensional data analysis and adversarial testing methods, and comprehensively improve the reliability evaluation capability of artificial intelligence systems to meet the high security requirements in complex application scenarios. The "artificial intelligence reliability evaluation system based on a secure and reliable algorithm" proposed by the present application aims to solve the above problems through innovative technical means and fill the gap in existing technology. SUMMARY

[0005] The present application aims to solve the problems existing in the prior art, and proposes an artificial intelligence reliability evaluation system based on a secure and reliable algorithm.

[0006] In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme: an artificial intelligence reliability evaluation system based on a secure and reliable algorithm, the system comprising:

[0007] The data acquisition module acquires log files, sensor data and real-time running data of the artificial intelligence system, eliminates noise interference and redundant segments, forms a stable data sequence, inputs a secure and reliable algorithm framework, constructs three groups of feature planes, and obtains an associated feature combination;

[0008] The reliability analysis module extracts and cross-sections the behavior changes based on the running state trajectory in the associated feature combination, calculates the state gradient and behavior fluctuation matching degree, adjusts the trajectory combined with the environmental changes, and forms an overlapping evolution structure;

[0009] The robustness evaluation module extracts the non-overlapping period of the running state and behavior difference in the overlapping evolution structure, analyzes the difference change and amplitude according to time, identifies the mutation segment, and obtains the boundary mutation segment;

[0010] The offset tracking module calls the running cycle segment data in the boundary mutation segment, analyzes the running state, pressure change and wear trajectory, performs offset point tracking analysis, calculates the main shaft offset, maps to the original running segment, and obtains the main shaft track offset result;

[0011] The abnormality calibration module analyzes the running time and behavior fluctuation based on the running segment where the main shaft track offset result is located, identifies the intersection point density and amplitude change, marks the abnormal point, and outputs the reliability evaluation result.

[0012] As a further scheme of the present application, the associated feature combination includes synchronous feature segments, continuous feature segments, and three groups of mapping planes, the overlapping evolution structure includes state change gradient values, behavior fluctuation matching degrees, and environmental adjustment factors, the boundary mutation segment includes running state difference sequences, behavior difference sequences, and mutation change amplitudes, the main shaft track offset result includes main shaft offset values, offset point features, and running state pressure changes, and the reliability evaluation result includes running time features, behavior fluctuation indicators, intersection density and amplitude abnormal points.

[0013] As a further scheme of the present application, the data acquisition module includes a multi-source acquisition sub-module, a noise filtering sub-module, and a feature plane sub-module.

[0014] The multi-source acquisition submodule acquires the timestamp and operation type field of the artificial intelligence system log file, acquires the temperature value and vibration intensity parameter of the sensor data, monitors the CPU occupancy rate and memory consumption of the real-time running data, aligns the three types of data according to the time axis, establishes a cross-device data mapping relationship, and generates an integrated data set;

[0015] The noise filtering submodule calculates the standard deviation of the temperature value and the peak-valley difference of the vibration intensity in the sensor data based on the integrated data set, sets a dynamic threshold to filter abnormal fluctuation segments, eliminates invalid data segments in the low-amplitude vibration interval, retains operation records in the log file that match the running data timestamp, and generates a purified data sequence.

[0016] The feature construction submodule calls the purified data sequence, extracts the temperature change rate and vibration frequency spectrum features as time domain features, calculates the CPU occupancy rate peak interval and memory leakage rate as running features, and calculates the log operation type frequency and response delay as behavior features, normalizes the three types of feature data, establishes a Pearson correlation coefficient matrix between features, and generates a correlation feature combination.

[0017] As a further scheme of the application, the credibility analysis module includes a trajectory intersection submodule, a gradient coupling submodule, and an instruction generation submodule.

[0018] The trajectory intersection submodule extracts the displacement change rate of the continuous time window in the running state trajectory based on the correlation feature combination, detects the start time point and end time point of the behavior change intersection segment, calculates the synchronization rate of the trajectory curvature and behavior fluctuation amplitude in the intersection segment, and generates an intersection segment feature set.

[0019] The gradient coupling submodule calls the intersection segment feature set and uses the formula:

[0020]

[0021] The operation obtains the coupling strength M of the state gradient and the behavior fluctuation. ij The coupling strength matrix is input into a PID controller to generate a rudder deflection instruction and generate a dynamic coupling instruction set.

[0022] Wherein, represents the gradient change amount of the i-th type of state trajectory at the k-th time point, ΔB jk represents the amplitude difference of the j-th type of behavior fluctuation at the k-th time point, E c represents the current ambient temperature reference value, E tk represents the k-th time point value of the 30-minute ambient temperature change sequence, σ E is the temperature standard deviation threshold;

[0023] The instruction generation submodule parses the rudder deflection angle and the response time parameter according to the dynamic coupling instruction set, reconstructs the mechanical transmission ratio of the track curvature radius, and generates an overlapping evolution structure.

[0024] As a further scheme of the present application, the robustness evaluation module comprises a time period extraction submodule, a difference analysis submodule, and a mutation identification submodule.

[0025] The time period extraction submodule acquires time alignment marks of the running state trajectory and the behavior difference sequence, detects the start time stamp and the end time stamp of the non-overlapping time period, calculates the standard deviation of the state trajectory displacement and the behavior difference amplitude in the non-overlapping time period, and generates non-overlapping time period features based on the overlapping evolution structure.

[0026] The difference analysis submodule extracts the correlation coefficient of the state displacement change rate and the behavior difference change rate, calculates the difference amplitude growth rate of adjacent time windows, sets a dynamic threshold to divide high growth periods and low growth periods, and generates a difference dynamic partition by calling the non-overlapping time period features.

[0027] The mutation identification submodule counts the number of state displacement mutations and the cumulative value of behavior difference mutation amplitudes in the high growth period, compares the cumulative value with a preset mutation reference value, selects continuous time intervals exceeding the reference value, and generates a boundary mutation section.

[0028] As a further scheme of the present application, the offset tracking module comprises a cycle period analysis submodule, an offset coupling submodule, and an instruction generation submodule.

[0029] The cycle period analysis submodule extracts the peak-to-valley difference of the pressure change rate in the running cycle period, detects the spatial coordinates of the wear track curvature mutation point, calculates the Pearson correlation coefficient of the state displacement and the pressure peak, and generates a cycle period feature set by calling the boundary mutation section.

[0030] The offset coupling submodule uses the formula:

[0031]

[0032] The operation obtains the main shaft offset coupling strength ΔS, inputs the coupling strength into a PID controller to generate a servo motor compensation signal, and generates a dynamic offset instruction set.

[0033] Wherein, P t represents the pressure change rate at the t time point, is the average pressure change rate of the last 100 cycle periods, W t represents the wear track curvature, F t is the material fatigue coefficient, D t represents the state displacement, Ct is a Pearson correlation coefficient, mu C is a correlation coefficient mean, sigma C is a correlation coefficient standard deviation, T is the total number of single period segment sampling points;

[0034] The instruction generation submodule parses the amplitude and phase parameters of the servo motor compensation signal according to the dynamic offset instruction set, matches the time and space coordinates of the original running segment, and generates the main shaft track offset result.

[0035] As a further scheme of the application, the anomaly calibration module comprises a time period analysis submodule, a density calculation submodule, and an anomaly calibration submodule.

[0036] The time period analysis submodule calls the main shaft track offset result, extracts the running time length fluctuation of a continuous time window in a running segment, detects the peak-valley difference of a behavior fluctuation sequence, calculates the covariance coefficient of the running time length and the behavior fluctuation, and generates a time period feature set.

[0037] The density calculation submodule, based on the time period feature set, counts the number and position distribution of intersection points within a unit time, calculates the density growth rate of adjacent time windows, sets a dynamic growth threshold to divide a high growth section, and generates a dynamic density.

[0038] The anomaly calibration submodule, according to the dynamic density, matches the spatial coordinates and time stamps of the intersection points in the high growth section, marks the nodes with a coordinate offset greater than a dynamic reference value as abnormal points, and generates a reliability evaluation result.

[0039] Compared with the prior art, the application has the following advantages and positive effects:

[0040] In the application, by removing data segments that do not have synchronicity and continuity, a stable feature sequence is constructed, the data basis of virtual-real mapping is enhanced, a multi-parameter structure is introduced into the feature construction plane, the collaborative expression of state changes is realized, the state gradient is extracted in the running state and behavior intersection section and the trajectory is corrected in combination with environmental factors, the dynamic adaptability of the trajectory to complex scenes is improved, the differential analysis accurately depicts the sudden behavior, the main shaft offset mapping strengthens the behavior path backtracking association, the abnormal points are calibrated in combination with the running time length and behavior fluctuation features, a reliability evaluation system with multi-layer recognition capability is constructed, and the accuracy of state recognition of the artificial intelligence system is improved. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a system flowchart of the application;

[0042] Figure 2 is an acquisition flowchart of the data acquisition module of the application;

[0043] Figure 3Acquisition flowchart of the credibility analysis module of the present application;

[0044] Figure 4 Acquisition flowchart of the robustness evaluation module of the present application;

[0045] Figure 5 Acquisition flowchart of the offset tracking module of the present application;

[0046] Figure 6 Acquisition flowchart of the anomaly labeling module of the present application. DETAILED DESCRIPTION

[0047] The technical solutions in the present application will be described below with reference to the drawings.

[0048] In the embodiments of the present application, the words such as "example", "for example" and the like are used to represent as an example, illustration or description. Any embodiment or design scheme described as "example" in the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the word "example" is intended to present the concept in a specific manner. In addition, in the embodiments of the present application, the meaning expressed by "and / or" can be both, or can be one of the two.

[0049] In the embodiments of the present application, "image" and "picture" can be used interchangeably at times, and it should be pointed out that the meanings expressed are consistent when the distinction is not emphasized. "Of", "corresponding" and "corresponding" can be used interchangeably at times, and it should be pointed out that the meanings expressed are consistent when the distinction is not emphasized.

[0050] In the embodiments of the present application, sometimes the subscript such as W1 can be written in the form of non-subscript such as W1, and the meanings expressed are consistent when the distinction is not emphasized.

[0051] In order to make the technical problems, technical solutions and advantages to be solved by the present application more clear, the following will be described in detail with reference to the drawings and specific embodiments.

[0052] Please refer to Figure 1 The present application provides a technical solution: an artificial intelligence reliability evaluation system based on a secure and credible algorithm, which comprises:

[0053] The data acquisition module acquires log files, sensor data and real-time running data of the artificial intelligence system, eliminates noise interference and redundant segments, forms a stable data sequence, inputs a secure and credible algorithm framework, constructs three characteristic dimensions, and obtains an associated characteristic combination;

[0054] The credibility analysis module extracts the cross-section of behavior change based on the running state trajectory in the associated feature combination, calculates the state gradient and behavior fluctuation matching degree, adjusts the trajectory combined with the environmental change, and forms an overlapping evolution structure;

[0055] The robustness evaluation module extracts the non-coincidence period of running state and behavior difference in the overlapping evolution structure, analyzes the difference change and amplitude according to time, identifies the mutation section, and obtains the boundary mutation section;

[0056] The offset tracking module calls the running cycle segment data in the boundary mutation section, analyzes the running state, pressure change and wear trajectory, performs offset point tracking analysis, calculates the main shaft offset, maps it to the original running section, and obtains the main shaft track offset result;

[0057] The abnormality calibration module analyzes the running time and behavior fluctuation based on the running section where the main shaft track offset result is located, identifies the intersection point density and amplitude change, marks the abnormal point, and outputs the reliability evaluation result.

[0058] The associated feature combination includes synchronicity feature segment, continuity feature segment, three groups of mapping facets, the overlapping evolution structure includes state change gradient value, behavior fluctuation matching degree, environmental adjustment factor, the boundary mutation section includes running state difference sequence, behavior difference sequence, mutation change amplitude, the main shaft track offset result includes main shaft offset value, offset point feature, running state pressure change, and the reliability evaluation result includes running time feature, behavior fluctuation index, intersection density and amplitude abnormal point.

[0059] Please refer to Figure 2 , the data acquisition module includes a multi-source acquisition submodule, a noise filtering submodule, and a feature facet submodule;

[0060] The multi-source acquisition submodule collects the timestamp and operation type field of the artificial intelligence system log file, obtains the temperature value and vibration intensity parameter of the sensor data, monitors the CPU occupancy rate and memory consumption of real-time running data, aligns the three types of data according to the time axis, establishes a cross-device data mapping relationship, and generates an integrated data set;

[0061] The NTP time synchronization protocol is established to control the timestamp error of the log server, sensor array and running monitoring terminal within ±5ms, and the temperature value (unit: ℃) and vibration intensity (unit: m / s 2), real-time running data to collect CPU occupancy rate (percentage) and memory consumption (unit: GB) at 1 second intervals, when the sensor records the temperature value 82.3℃ at 2023-05-27T10:00:00.000Z, synchronously acquire the CPU usage rate 78% and the memory consumption 4.2GB within the time window of 50ms before and after this time, retrieve the "model inference" operation type within the corresponding timestamp ±100ms range in the log, compensate the time axis deviation of the three types of data to within ±10ms by timestamp alignment algorithm, for example, when the sensor clock is offset by +8ms, add a time compensation mark [10:00:00.008] to the temperature value 82.3℃, in the data mapping stage, create an associated key-value pair {sensorID: "TS-01", cpuID: "CORE-3", logID: "OP-28741"} for the temperature value 82.3℃, when it is detected that the sensor node S-05 loses data packets at 10:05:23.455Z due to network delay, trigger the data interpolation program, generate a compensation value 79.95℃ according to the previous data point (10:05:23.355Z, 79.8℃) and the subsequent data point (10:05:23.555Z, 80.1℃) using linear interpolation method, complete the cross-device data mapping to generate a structured data table containing timestamp, temperature value, vibration intensity, CPU%, memory GB, operation type, and store it as an integrated dataset in Parquet format.

[0062] Noise filtering submodule, based on the integrated dataset, calculate the standard deviation of the temperature value and the peak-to-valley difference of the vibration intensity in the sensor data, set a dynamic threshold to filter abnormal fluctuation segments, eliminate invalid data segments in the low-amplitude vibration interval, retain operation records in the log file that match the running data timestamp, generate a purified data sequence;

[0063] Calculate the sliding window standard deviation of the temperature value sequence, set the window size to 300 sampling points (corresponding to 30 seconds), mark as an abnormal fluctuation segment when the window temperature standard deviation exceeds 2.5℃, for example, measure the temperature standard deviation 3.1℃ from 10:15:00 to 10:15:30, trigger threshold filtering, calculate the peak-to-valley difference (maximum value - minimum value) of the vibration intensity data within each 60-second time window, set a dynamic threshold of window mean ±1.5 times standard deviation, when the vibration intensity peak-to-valley difference reaches 215m / s 2 (mean 150m / s 2 , standard deviation 43m / s 2) is determined as an abnormal fluctuation, timestamp matching verification is performed on the log operation record, when the time of a certain "parameter update" operation record is 10:25:33.120Z, it is checked whether there is a sudden increase record of CPU usage rate in the corresponding time window [10:25:33.100Z, 10:25:33.140Z], if there is no matching data, the log entry is removed, in the data purification stage, the temperature value sequence is filtered by the 3σ principle, when the temperature value of a certain sampling point deviates from the sliding mean value by more than 3 times the standard deviation (for example, when the mean value is 80°C and the standard deviation is 2°C, the data point is above 86°C or below 74°C), the removal is performed, and finally the purified data sequence containing the effective temperature value, vibration intensity, CPU%, memory GB and matching operation record is generated.

[0064] The feature plane submodule calls the purified data sequence, extracts the temperature change rate and vibration frequency spectrum feature as the time domain plane, calculates the CPU occupation rate peak interval and memory leakage rate as the running plane, and counts the log operation type frequency and response delay as the behavior plane. The three types of plane data are normalized, a Pearson correlation coefficient matrix between features is established, and an associated feature combination is generated.

[0065] The temperature change rate is calculated as the temperature difference between adjacent sampling points divided by the sampling interval (0.1 seconds), when the temperature is 80.2°C at 10:30:00.000Z and the temperature is 80.5°C at 10:30:00.100Z, the change rate is (80.5-80.2) / 0.1=3°C / s, when extracting the vibration frequency spectrum feature, FFT transformation is performed on each 1-second data segment (10 sampling points), and the first 5 main frequency component amplitudes are taken, for example, the main frequency 25Hz component amplitude measured at 10:35:00 is 58m / s 2, record the duration of the interval when the CPU occupancy rate exceeds the threshold of 85%, mark the interval as 15 seconds when there are 15 consecutive seconds of peak values from 10:40:00 to 10:40:15, calculate the memory leakage rate as the increase in memory consumption per minute, when the memory is 4.1 GB at 10:45:00 and 4.3 GB at 10:46:00, the leakage rate is 0.2 GB / min, when counting the frequency of the "model reload" operation in a 5-minute window, when there are 3 such operations from 10:50:00 to 10:55:00, record the frequency as 3 times / 5min, when performing normalization, linearly map the temperature change rate from the original range [0, 5℃ / s] to [0, 1], when the original value is 3℃ / s, the normalized value is (3-0) / (5-0) = 0.6, when constructing the Pearson correlation coefficient matrix, calculate the covariance of the temperature change rate and the CPU usage rate divided by the product of the standard deviations of the two, when the covariance is 2.15, the temperature standard deviation is 1.2, and the CPU standard deviation is 1.8, the correlation coefficient is 2.15 / (1.2×1.8) = 0.996, generate a correlation feature combination containing the time domain aspect, the running aspect, and the behavior aspect correlation coefficients.

[0066] Please refer to Figure 3 , the trustworthiness analysis module includes a trajectory intersection submodule, a gradient coupling submodule, and an instruction generation submodule;

[0067] The trajectory intersection submodule extracts the displacement change rate of the running state trajectory in the continuous time window based on the correlation feature combination, detects the start time point and the end time point of the behavior change intersection segment, calculates the synchronization rate of the trajectory curvature and the behavior fluctuation amplitude in the intersection segment, and generates an intersection segment feature set;

[0068] From the correlation feature combination, extract the running state trajectory in three dimensions of temperature change rate (0-5℃ / s), CPU usage rate (0-100%), and vibration frequency amplitude (0-200m / s 2 ) three dimensions of running state trajectory, set the time window length to 30 seconds, calculate the displacement change rate in each window, take the maximum and minimum difference of the temperature change rate sequence in the window (for example, when the temperature change rate in the window is [2.1, 3.5, 4.0, 3.8], the displacement change rate is 4.0-2.1 = 1.9℃ / s), when the displacement change rate of the continuous three windows exceeds the threshold of 1.5℃ / s, mark the starting point of the behavior change intersection segment, for example, when the change rate sequence [1.6, 1.7, 1.8] is detected in the period from 10:00:00 to 10:01:30, determine the starting time as 10:00:00, calculate the trajectory curvature using the three-point formula where Δx is the temperature rate difference, Δy is the CPU usage rate difference, when the temperature rate increases from 3.2℃ / s to 4.1℃ / s, the CPU usage rate increases from 78% to 85%, the curvature value is calculated as 0.15m -1 , the synchronization rate calculation uses the Pearson correlation coefficient, when the curvature sequence and the vibration amplitude sequence correlation coefficient exceeds 0.8, it is determined as high synchronization, for example, the curvature sequence [0.12, 0.15, 0.18] and the vibration amplitude [58, 62, 65] correlation coefficient 0.87, generate a cross-section feature set containing start time, end time, average curvature, synchronization rate.

[0069] Table 1: Trajectory crossing feature parameter table

[0070]

[0071] As shown in Table 1, the parameter changes of three consecutive time windows constitute the cross-section feature set, in which the temperature rate and the vibration amplitude show a positive correlation trend.

[0072] Gradient coupling sub-module, call cross-section feature set, use formula:

[0073]

[0074] Operation obtains the coupling strength M of state gradient and behavior fluctuation ij , input the coupling strength matrix into the PID controller to generate rudder deflection instructions, generate dynamic coupling instruction set;

[0075] where, represents the gradient change amount of the i-th state trajectory at the k-th time point (unit: m / s 2 ), ΔB jk represents the amplitude difference of the j-th behavior fluctuation at the k-th time point (unit: dB), E c represents the current environmental temperature reference value (unit: ℃), E tk represents the k-th time point value of the recent 30-minute environmental temperature change sequence, σ E is the temperature standard deviation threshold;

[0076] Collect the recent 30-minute environmental temperature sequence to build a sliding window, the window length is set to 5-minute intervals, when the time stamp 2023-05-27T11:00:00 to 11:30:00 obtains the temperature sequence [24.8, 25.1, 24.9, 25.3, 25.0, 24.7]℃, calculate the mean E c =24.97℃, standard deviation σ E= 0.21℃, the standard deviation threshold is determined by experiment: collect 100 groups of temperature data continuously in a constant temperature environment of 25℃, calculate the standard deviation distribution interval as 0.18-0.25℃, take the median 0.21℃ as the threshold, state gradient The calculation uses the central difference method, when the temperature change rate at the k = 15 time point is 3.8℃ / s and the k = 16 time point is 4.2℃ / s, Behavior fluctuation amplitude difference ΔB jk Take the vibration intensity difference of adjacent time points, when k = 15, the vibration intensity is 58m / s 2 , k = 16 is 62m / s 2 , then ΔB j,16 = 62-58 = 4m / s 2 , when calculating the molecular term, iterate k = 15 to k = 17 three time points:

[0077] k = 15: (0.4 x 4) 2 = 2.56;

[0078] k = 16: (0.3 x 5) 2 = 2.25 (when ΔB j,17 = 63-62 = 1m / s 2 );

[0079] k = 17: (0.2 x 1) 2 = 0.04;

[0080] The sum of the molecules ∑ = 2.56 + 2.25 + 0.04 = 4.85, and the denominator is calculated as When calculating the environmental temperature compensation term, take the environmental temperature at the k = 16 time point as 24.9℃, then the exponential term is e - | 24.97-24.9| / 0.21 = e -0.07 / 0.21 = e -0.333 = 0.716, the final coupling strength M ij = (4.85 / 29.920) x 0.716 = 0.116, which indicates that the coupling strength between the state gradient and the behavior fluctuation exceeds the set threshold 0.1, meaning that the system detects abnormal coupling phenomenon, and needs to generate compensation instructions through the PID controller, when the error threshold of the set PID controller is 0.1, 0.116 > 0.1 triggers the control amount calculation, the proportional term K p×e(t) = 0.8 × (0.116 - 0.1) = 0.0128, the integral term accumulates the past 5 seconds error integral value 0.0083, the derivative term calculates the error rate of change (0.116 - 0.105) / 0.1 = 0.11, the final output rudder deflection angle θ = 0.0128 + 0.0083 + 0.8 × 0.11 = 0.109 rad ≈ 6.25°, generate instruction set record timestamp 11:00:16, angle 6.25°, response time 0.5 seconds control parameters.

[0081] The instruction generation submodule parses the rudder deflection angle and response time parameters according to the dynamic coupling instruction set, reconstructs the mechanical transmission ratio of the track curvature radius, and generates an overlapping evolution structure.

[0082] When the rudder deflection angle is 6.25°, according to the harmonic reducer transmission ratio 120:1, the curvature radius adjustment amount ΔR = θ / (360 × N) = 6.25 / (360 × 120) = 0.0001447 m, when the original track curvature radius R0 = 0.15 m is detected, the new curvature radius R new = 0.15 + 0.0001447 = 0.1501447 m, the response time parameter 0.5 seconds is used to calculate the angular acceleration α = 2θ / t 2 = 2 × 6.25 / (0.5 2 ) = 50° / s 2 , converted to linear acceleration a = α × R0 = 50 × 0.15 × (π / 180) = 0.1309 m / s 2 , the curvature radius adjustment amount sequence 0.0001447 m, 0.0001389 m, 0.0001421 m is obtained in the next three control periods, the average value 0.0001419 m is calculated as the mechanical transmission ratio calibration reference, according to the ISO 1328-1 standard, the gear clearance error 0.00005 m is checked, the adjustment amount is corrected to 0.0001419 - 0.00005 = 0.0000919 m, finally an overlapping evolution structure containing deflection angle 6.25°, response time 0.5 seconds, corrected curvature radius 0.1500919 m is generated, which is stored in the control database with time stamp, curvature radius, adjustment sequence, transmission ratio as the core field.

[0083] Please refer to Figure 4 , the robustness evaluation module includes time period extraction submodule, difference analysis submodule, mutation recognition submodule;

[0084] The time period extraction submodule obtains the time alignment mark of the running state trajectory and the behavior difference sequence based on the overlapping evolution structure, detects the start and end time stamps of the non-overlapping period, calculates the standard deviation of the state trajectory displacement and the behavior difference amplitude in the non-overlapping period, and generates the non-overlapping period feature.

[0085] The tolerance threshold of the time alignment marker is ± 50 ms. When the running state trajectory records a curvature radius of 0.1503 m at 2023-05-28T14:00:00.000Z, the behavior difference sequence database is retrieved to find the nearest record of 5 m / s at 14:00:00.045Z 2 , the time deviation is calculated as 45 ms. When the deviation of the last 5 sampling points exceeds 80% (40 ms) of the tolerance threshold, the non-coincidence period start time is marked as 14:00:00.000Z, and the scanning continues to 14:00:02.150Z when the time deviation drops to 30 ms (60% below the threshold). The period end marker is terminated, and the curvature radius sequence [0.1503, 0.1505, 0.1508, 0.1506, 0.1507] m within the 2.15 seconds is extracted. The displacement amount range is calculated as 0.0005 m (0.1508-0.1503), and the behavior difference sequence is [5, 7, 8, 6, 9] m / s 2 When calculating the standard deviation, the mean is 6.8 m / s 2 , the sum of squared differences (5-6.8) 2 +(7-6.8) 2 +(8-6.8) 2 +(6-6.8) 2 +(9-6.8) 2 = 3.24 + 0.04 + 1.44 + 0.64 + 4.84 = 10.2, and the standard deviation is √(10.2 / 5) = 1.43 m / s 2 When the displacement amount 0.0005 m exceeds the threshold 0.0004 m (this threshold is obtained by analyzing 1000 groups of normal data: 99.7% of the sample displacement amount ≤0.0004 m), the standard deviation 1.43 m / s 2 exceeds the threshold 1.2 m / s 2 (set according to the XYZ-2000 technical manual of the device model), the non-coincidence period feature {start: "14:00:00.000", end: "14:00:02.150", displacement amount: 0.0005, standard deviation: 1.43} is generated.

[0086] The difference analysis submodule calls the non-coincidence period feature, extracts the correlation coefficient of the state displacement change rate and the behavior difference change rate, calculates the difference amplitude growth rate of adjacent time windows, sets a dynamic threshold to divide high growth periods and low growth periods, and generates a dynamic partition of differences;

[0087] The state displacement change rate is calculated every 0.1 seconds within the time period of 14:00:00.000-14:00:02.150, when t1=14:00:01.000, the radius of curvature is 0.1503m, and when t2=14:00:01.100, it is 0.1505m, the change rate is (0.1505-0.1503) / 0.1=0.002m / s, and the behavior difference value within the time window is from 5m / s 2 to 7m / s 2 , the change rate is (7-5) / 0.1=20m / s 3 , the correlation coefficient is calculated using a 20-second sliding window (200 sample points), the average state change rate within the window is 0.0015m / s, and the average behavior change rate is 18m / s 3 , the covariance is calculated as Σ[(s_i-0.0015)(b_i-18)] / 200=2.1, the state standard deviation is 0.0007m / s, and the behavior standard deviation is 4.8m / s 3 , the correlation coefficient r=2.1 / (0.0007×4.8)=625 (after standardization, it is 0.82), when |r|>0.8, it is determined as strong correlation, and the time period of 14:00:01.200-14:00:01.500 is detected, r=-0.83, which is marked as strong negative correlation, when calculating the growth rate of adjacent windows, the average difference value of the front window (14:00:01.000-14:00:01.250) is 6.2m / s 2 , and the average difference value of the rear window (14:00:01.250-14:00:01.500) is 8.4m / s 2 , the growth rate is (8.4-6.2) / 6.2×100%=35.5%, the dynamic threshold is set to 30% (according to historical fault data statistics: 85% of abnormal event growth rates are greater than or equal to 30%), when 35.5%>30%, it is marked as a high growth area, and a difference dynamic partition record is generated {start: "14:00:01.200", end: "14:00:01.500", growth rate: 35.5%, correlation coefficient: -0.83}.

[0088] The mutation recognition submodule, according to the difference dynamic partition, the cumulative value of the number of state displacement mutations and the mutation amplitude of the behavior difference value within the high growth period is calculated, the cumulative value is compared with the preset mutation reference value, the continuous time interval exceeding the reference value is screened, and the boundary mutation section is generated;

[0089] The scanning state bit displacement mutation occurs in the high growth period 14:00:01.200-14:00:01.500, the single mutation threshold is set to 0.0002 m / 0.1 s, three consecutive sampling point changes of 0.00025 m (14:00:01.300), 0.00022 m (14:00:01.400), and 0.00027 m (14:00:01.500) are detected, and the cumulative value is 0.00074 m. In the behavior difference value mutation detection, the single mutation threshold is set to 3 m / s 2 When the difference value increases from 6 m / s 2 (14:00:01.250) to 10 m / s 2 (14:00:01.300), the change is 4 m / s 2 , and then decreases from 10 m / s 2 (14:00:01.300) to 5 m / s 2 (14:00:01.350), the change is 5 m / s 2 , the cumulative mutation amplitude is 4+5=9 m / s 2 , the preset reference value is set according to the ISO 10816-3 vibration standard: displacement cumulative threshold 0.0007 m (corresponding to the safe operation limit of the equipment), difference cumulative threshold 8 m / s 2 (maximum allowed impact), when the displacement cumulative value is detected to be 0.00074 m> 0.0007 m and the difference cumulative value is 9 m / s 2 > 8 m / s 2 , it is determined that the boundary mutation section is 14:00:01.200-14:00:01.500, and a structured output is generated, including four core parameters of start time, end time, displacement cumulative value, and difference cumulative value. The displacement cumulative value exceeds the reference value by 5.7% ((0.00074-0.0007) / 0.0007x100%), and the difference cumulative value exceeds 12.5% ((9-8) / 8x100%). The data in this section will be imported into the offset tracking module for in-depth analysis.

[0090] Please refer to Figure 5 , the offset tracking module includes a periodic section analysis submodule, an offset coupling submodule, and an instruction generation submodule;

[0091] The periodic section analysis submodule calls the boundary mutation section, extracts the peak-to-valley difference of the pressure change rate in the running period, detects the spatial coordinates of the wear track curvature mutation point, calculates the Pearson correlation coefficient of the state displacement and the pressure peak value, and generates a periodic section feature set.

[0092] Read the time interval 2023-05-29T09:00:00.000 to 09:00:05.000 from the boundary mutation segment database, extract the pressure rate of change sequence, set the sliding window length to 0.5 seconds (50 sampling points), detect the peak-valley difference, the maximum value of the pressure rate of change in the window is 1.8 MPa / s (09:00:01.200), the minimum value is-0.9 MPa / s (09:00:01.700), the difference is 2.7 MPa / s, when the difference of three consecutive windows exceeds the threshold value 2.5 MPa / s, mark it as a curvature mutation candidate area, perform three times of spline interpolation on the wear track data, and calculate the curvature formula in the spatial coordinate system When the x-axis displacement changes at 09:00:02.300 Acceleration y-axis Calculate the curvature κ = |0.12 × 0.03-0.08 × 0.05| / (0.12 2 +0.08 2 ) 3 / 2 = |0.0036-0.004| / (0.0208) 1.5 = 0.0004 / 0.00094 = 0.426 m -1 , when the curvature change exceeds 0.4 m -1 , mark it as a mutation point, calculate the Pearson correlation coefficient of the state displacement sequence [0.12, 0.15, 0.18] mm and the pressure peak sequence [8.2, 8.5, 8.7] MPa, the covariance is 0.015, the state standard deviation is 0.03, the pressure standard deviation is 0.25, the correlation coefficient r = 0.015 / (0.03×0.25) = 2.0 (after standardization, 0.67), generate the period segment feature set {peak-valley difference: 2.7 MPa / s, curvature: 0.426 m -1 , correlation coefficient: 0.67}.

[0093] Offset coupling sub-module, based on the period segment feature set, using the formula:

[0094]

[0095] Operation to obtain the main shaft offset coupling strength ΔS, input the coupling strength into the PID controller to generate the servo motor compensation signal, and generate the dynamic offset instruction set;

[0096] Where, P t represents the pressure rate of change at the t time point (unit: MPa / s), is the average pressure rate of change of the last 100 period segments, W t represents the wear track curvature (unit: m-1), F tF is the material fatigue coefficient (obtained by hardness tester), D t C represents the state displacement amount (unit: mm), C t μ is the Pearson correlation coefficient, μ C σ is the correlation coefficient mean, σ C T is the total number of sampling points in a single period (default value 500);

[0097] During execution, set P is the average pressure change rate of the last 100 periods, and the historical data [1.2, 1.5, 1.3, …, 1.4] MPa / s (a total of 100 values) is taken when calculating, and the average is F is the material fatigue coefficient t Convert through Rockwell hardness HRC value, when the measured hardness HRC45, look up table F t = 0.85 (according to ASTM E140 standard, HRC45 corresponds to fatigue coefficient 0.8-0.9), for t = 300 time point, parameter assignment: P t = 1.8 MPa / s, W t = 0.426 m -1 , F t = 0.85, D t = 0.18 mm, C t = 0.67, μ C = 0.6 (the average of 100 period correlation coefficients in history), σ C = 0.1 (standard deviation);

[0098] The formula is calculated in steps:

[0099] Numerator:

[0100] Denominator:

[0101] Single point contribution value: 1.113 x 0.329 = 0.366;

[0102] Sum for T = 500 sampling points, assuming an average contribution of 0.3, then ΔS = 500 x 0.3 = 150, which indicates that the main shaft offset coupling strength exceeds the safety threshold 120 (according to ISO 1940 balance grade G6.3 setting), which means that the cumulative offset of the mechanical system has reached the critical state, and the PID controller needs to generate a compensation signal for active correction, when the set proportional coefficient K p = 0.5, integral time T i = 2 s, differential time T d=0.05s, when the error e(t)=150-120=30, the control quantity u(t)=0.5×30+(0.5 / 2)×∫30dt+0.5×0.05×d(30) / dt=15+7.5t+0.75, generating a dynamic offset instruction set with a compensation signal amplitude of 15.75V and a phase angle of 32°.

[0103] The instruction generation submodule analyzes the amplitude and phase parameters of the servo motor compensation signal according to the dynamic offset instruction set, matches the time and space coordinates of the original operation segment, and generates the spindle track offset result;

[0104] When the amplitude of the analytical compensation signal is 15.75V, the displacement compensation is calculated as 15.75×0.5=7.875mm based on the servo motor sensitivity of 0.5mm / V. The phase angle of 32° is converted into a time delay of Δt=(32 / 360)×(1 / 50Hz)=0.00178s (50Hz control frequency). When matching the time and space coordinates of the original running segment, the spindle position coordinates X=1250.3mm and Y=980.7mm at 09:00:03.000Z are detected. After applying the compensation, the new coordinates X=1250.3+7.875×cos(32°) =1250.3+6.68=1256.98mm, Y=980.7+7.875×sin(32°)=980.7+4.16=984.86mm. When checking the coordinate boundaries, the maximum allowable displacement of the X-axis is 1260mm (±10mm from the original position). The current 1256.98mm is within the safe range. The spindle track offset result is generated {timestamp: "09:00:03.000", X: 1256.98, Y: 984.86, compensation amount: 7.875mm}, and this result is written to the CNC system for real-time trajectory correction.

[0105] See also Figure 6 ,The anomaly calibration module includes a time period analysis submodule, a density calculation submodule, and anomaly calibration submodule;

[0106] The time period analysis submodule calls the spindle track offset results, extracts the running time fluctuations in the continuous time window within the running segment, detects the peak-to-valley difference of the behavior fluctuation sequence, calculates the covariance coefficient of the running time and behavior fluctuation, and generates a time period feature set;

[0107] Read the 2023-05-30T10:00:00.000 to 10:05:00.000 period data from the spindle orbit deviation result database, set the time window length to 30 seconds, calculate the running time fluctuation of each window, when window 1 (10:00:00-10:00:30) actually runs for 30.2 seconds (theoretical value 30 seconds), the fluctuation is +0.2 seconds, window 2 (10:00:30-10:01:00) runs for 29.8 seconds, the fluctuation is -0.2 seconds, extract the peak-to-valley value of the vibration acceleration of the behavior fluctuation sequence, the maximum acceleration in window 1 is 12.5 m / s 2 , the minimum is 8.3 m / s 2 , the difference is 4.2 m / s 2 , the difference in window 2 is 3.8 m / s 2 , when calculating the covariance coefficient, take the variance of the running time of 10 consecutive windows the variance of the behavior fluctuation difference covariance covariance coefficient ρ=

[0108] When the absolute value of the coefficient is greater than 0.4, it is determined to be moderately correlated, and the period feature set {window number: 10, average fluctuation: ±0.15s, average peak-to-valley difference: 4.0m / s 2 , covariance coefficient: 0.47} is generated.

[0109] The density calculation submodule, based on the period feature set, counts the number of intersection points and position distribution in unit time, calculates the density growth rate of adjacent time windows, sets a dynamic amplitude threshold to divide high growth sections, and generates dynamic density;

[0110] Count the number of intersection points in each 30-second window in the period feature set, window 1 detects 15 intersection points (X-axis coordinate range 1250.1-1255.3mm), window 2 detects 22 intersection points (1253.5-1258.9mm), calculate the density per unit time (every minute): window 1 density 15 / 0.5=30 / minute, window 2 density 22 / 0.5=44 / minute, the growth rate is calculated as (44-30) / 30x100%=46.7%, the dynamic amplitude threshold is set to 40% (based on historical data statistics: 90% of abnormal events have a density growth rate ≥40%), when 46.7%>40% is detected, window 2 is marked as a high growth section, and the dynamic density record {start: "10:00:30", end: "10:01:00", density: 44 / minute, growth rate: 46.7%} is generated, and the subsequent window is continuously scanned, and when the window 3 density reaches 50 / minute (growth rate 13.6%<40%), the marking is stopped.

[0111] Anomaly calibration sub-module, according to dynamic density, matching the spatial coordinates and time stamp of intersection points in high growth section, marking the nodes with coordinate offset exceeding dynamic reference value as anomaly points, generating reliability evaluation results;

[0112] In high growth section 10:00:30-10:01:00, 22 intersection points are analyzed, spatial coordinates are extracted to calculate offset, dynamic reference value is set as average offset of previous 10 windows ±1.5mm (average offset is calculated as 2.1mm, standard deviation is 0.3mm, reference value = 2.1+3×0.3 = 3.0mm), intersection point P15 coordinate X = 1258.2mm (original reference position 1255.0mm), offset 3.2mm>3.0mm, intersection point P19 coordinate X = 1259.1mm, offset 4.1mm>3.0mm, marked as anomaly points, generating reliability evaluation results {anomaly point number:2, maximum offset:4.1mm, belonging section:“10:00:30-10:01:00”}, the results are written into quality management system to trigger equipment shutdown inspection program.

[0113] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited to this, any skilled person in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An artificial intelligence reliability evaluation system based on a secure and reliable algorithm, characterized by: The system comprises: The data acquisition module obtains the AI ​​system's log files, sensor data, and real-time operation data, removes noise interference and redundant fragments, forms a stable data sequence, inputs it into a secure and reliable algorithm framework, constructs three sets of feature facets, and obtains associated feature combinations; The credibility analysis module extracts the intersection segments with behavior changes based on the running state trajectory in the associated feature combination, calculates the matching degree between the state gradient and the behavior fluctuation, adjusts the trajectory in combination with the environmental changes, and forms an overlapping evolution structure; The robustness evaluation module extracts the non-overlapping period of the difference between the operating state and the behavior in the overlapping evolution structure, analyzes the difference change and amplitude according to time, identifies the mutation segment, and obtains the boundary mutation segment; The offset tracking module calls the operating cycle data in the boundary mutation section, analyzes the operating status, pressure change and wear trajectory, performs offset point tracking analysis, calculates the spindle offset, maps it to the original operating section, and obtains the spindle track offset result; The anomaly calibration module analyzes the running time and behavior fluctuations based on the running segment where the spindle track deviation result is located, identifies the intersection density and increase change, marks the anomaly point, and outputs the reliability evaluation result.

2. The artificial intelligence reliability evaluation system based on a secure and trustworthy algorithm according to claim 1 is characterized by: The associated feature combination includes synchronous feature segments, continuous feature segments, and three groups of mapping facets; the overlapping evolution structure includes state change gradient values, behavior fluctuation matching degrees, and environmental regulation factors; the boundary mutation segment includes operating state difference sequences, behavior difference sequences, and mutation change amplitudes; the spindle orbit offset results include spindle offset values, offset point features, and operating state pressure changes; and the reliability evaluation results include operating duration features, behavior fluctuation indicators, intersection density, and amplification anomaly points.

3. The artificial intelligence reliability evaluation system based on a secure and trustworthy algorithm according to claim 1 is characterized by: The data acquisition module includes a multi-source acquisition submodule, a noise filtering submodule, and a feature faceting submodule; The multi-source acquisition submodule collects the timestamp and operation type fields of the AI ​​system log files, obtains the temperature value and vibration intensity parameters of the sensor data, monitors the CPU usage and memory consumption of the real-time operation data, aligns the three types of data according to the time axis, establishes a cross-device data mapping relationship, and generates an integrated data set; A noise filtering submodule calculates the standard deviation of temperature values ​​and the peak-to-valley difference of vibration intensity in the sensor data based on the integrated data set, sets a dynamic threshold to filter out abnormal fluctuation segments, removes invalid data segments in the low-amplitude vibration range, retains operation records in the log file that match the operation data timestamp, and generates a cleansed data sequence; The feature faceting submodule calls the purified data sequence, extracts the temperature change rate and vibration spectrum features as the time domain facet, calculates the CPU occupancy peak interval and memory leakage rate as the operation facet, and counts the log operation type frequency and response delay as the behavior facet. The three types of facet data are normalized, and a Pearson correlation coefficient matrix between features is established to generate a correlation feature combination.

4. The artificial intelligence reliability evaluation system based on a secure and trustworthy algorithm according to claim 1 is characterized by: The credibility analysis module includes a trajectory intersection submodule, a gradient coupling submodule, and an instruction generation submodule; The trajectory intersection submodule extracts the displacement change rate of the continuous time window in the running state trajectory based on the combination of the associated features, detects the starting and ending time points of the behavior change intersection segment, calculates the synchronization rate between the trajectory curvature and the behavior fluctuation amplitude in the intersection segment, and generates the intersection segment feature set; The gradient coupling submodule calls the cross-segment feature set and uses the formula: The coupling strength M between the state gradient and the behavior fluctuation is obtained by calculation ij ,The coupling strength matrix is ​​input into the PID controller to generate the servo deflection command and generate the dynamic coupling instruction set; in, Represents the gradient change of the i-th state trajectory at the k-th time point, ΔB jk represents the amplitude difference of the j-th type of behavior fluctuation at the k-th time point, E c Represents the current ambient temperature reference value, E tk Represents the k-th time point value of the ambient temperature change sequence in the last 30 minutes, σ E is the temperature standard deviation threshold; The instruction generation submodule analyzes the steering gear deflection angle and response time parameters according to the dynamic coupling instruction set, reconstructs the mechanical transmission ratio of the trajectory curvature radius, and generates an overlapping evolution structure.

5. The artificial intelligence reliability evaluation system based on a secure and trustworthy algorithm according to claim 1 is characterized by: The robustness evaluation module includes a time period extraction submodule, a difference analysis submodule, and a mutation identification submodule; The time period extraction submodule obtains the time alignment mark of the running state trajectory and the behavior difference sequence based on the overlapping evolution structure, detects the start and end timestamps of the non-overlapping time period, calculates the standard deviation of the state trajectory displacement and the behavior difference amplitude within the non-overlapping time period, and generates the non-overlapping time period feature; The difference analysis submodule calls the non-overlapping period features, extracts the correlation coefficient between the state displacement change rate and the behavior difference change rate, calculates the difference amplitude growth rate of adjacent time windows, sets a dynamic threshold to segment the high-growth period and the low-growth period, and generates a difference dynamic partition; The mutation identification submodule, based on the dynamic partitioning of the difference, counts the cumulative value of the number of state displacement mutations and the behavioral difference mutation amplitude during the high-growth period, compares the cumulative value with the preset mutation benchmark value, screens the continuous time intervals exceeding the benchmark value, and generates the boundary mutation segment.

6. The artificial intelligence reliability evaluation system based on a secure and trustworthy algorithm according to claim 1 is characterized by: The offset tracking module includes a periodic segment analysis submodule, an offset coupling submodule, and an instruction generation submodule; The periodic segment analysis submodule calls the boundary mutation segment, extracts the peak-to-valley difference of the pressure change rate within the operating periodic segment, detects the spatial coordinates of the wear track curvature mutation point, calculates the Pearson correlation coefficient between the state displacement and the pressure peak, and generates a periodic segment feature set; The offset coupling submodule, based on the periodic segment feature set, adopts the formula: The spindle offset coupling strength ΔS is obtained by calculation, and the coupling strength is input into the PID controller to generate a servo motor compensation signal and a dynamic offset instruction set; Among them, P t represents the pressure change rate at time t, is the average pressure change rate of the last 100 cycles, W t represents the wear track curvature, F t is the material fatigue coefficient, D t Represents the state displacement, C t is the Pearson correlation coefficient, μ C is the mean of the correlation coefficient, σ C is the standard deviation of the correlation coefficient, T is the total number of sampling points in a single cycle; The instruction generation submodule analyzes the amplitude and phase parameters of the servo motor compensation signal according to the dynamic offset instruction set, matches the time and space coordinates of the original operation segment, and generates the spindle track offset result.

7. The artificial intelligence reliability evaluation system based on a secure and trustworthy algorithm according to claim 1 is characterized by: The anomaly calibration module includes a time period analysis submodule, a density calculation submodule, and an anomaly calibration submodule; The time period analysis submodule calls the spindle track offset result, extracts the running time fluctuation of the continuous time window within the running segment, detects the peak-to-valley difference of the behavior fluctuation sequence, calculates the covariance coefficient of the running time and the behavior fluctuation, and generates a time period feature set; The density calculation submodule, based on the time period feature set, counts the number and location distribution of intersection points in unit time, calculates the density growth rate of adjacent time windows, sets a dynamic growth threshold to segment high-growth segments, and generates dynamic density; The anomaly calibration submodule matches the spatial coordinates and timestamps of the intersection points in the high-growth section according to the dynamic density, marks the nodes whose coordinate offset exceeds the dynamic benchmark value as anomalies, and generates reliability evaluation results.

Images