An artificial intelligence-based device anomaly automatic monitoring method

Abstract

The application discloses an equipment abnormality automatic monitoring method based on artificial intelligence and relates to the technical field of equipment prediction operation and maintenance, and comprises the following steps: reading a multi-source measurement data frame, establishing a measurement channel registration item, determining a resolution quantity, obtaining an original value sequence, constructing a continuous window sequence and performing robust normalization, and writing into a history window library; determining a normalized resolution quantity, generating a parent set and a causal chain, solidifying a time lag table, establishing a near neighbor library, generating an anchor quantity set and an anchor vector library, calibrating an abnormal trigger threshold, generating a benchmark window set and a recursive sequence, and calibrating a maintenance threshold; extracting an anchor vector of an online time window, assembling an observed parent quantity vector, generating a fact prediction, calculating an abnormal score, generating a counter-fact prediction, and generating a degradation label; performing degradation label gating, calculating a health index, constructing a recent window set, generating maintenance information, realizing clear explanation and stable discrimination of abnormal causes, and improving the continuity of health index evolution and the reliability of maintenance information.

Description

Technical Field

[0001] This invention relates to the field of equipment predictive maintenance technology, and in particular to an automated monitoring method for equipment anomalies based on artificial intelligence. Background Technology

[0002] In the operation and maintenance of industrial equipment, multi-source measurements such as vibration, temperature, electrical parameters, pressure, and flow are often used as the basis. These measurements are collected through the Internet of Things to form a continuous time series. Typically, the data is first cleaned with timestamps and uniformly sampled. Within a sliding time window, the data is scaled and features are constructed. Statistical process control or time series prediction models are used to calculate anomaly scores and generate health indicators and maintenance suggestions to support online alarms and preventive maintenance.

[0003] However, in actual deployment, there are still two types of constraints. On the one hand, the difference in sampling rate of multi-source channels, insufficient quantization resolution, and missing data will amplify the fluctuation of normalized residuals and affect the stability of scoring thresholds. On the other hand, conventional alarms lack a discrimination mechanism based on causal chains and counterfactual inference, making it difficult to effectively distinguish between operating condition disturbances and actual degradation, resulting in insufficient consistency in health indicator updates and maintenance information generation. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an automated monitoring method for equipment anomalies based on artificial intelligence to solve the problems of large fluctuations in the alignment and normalization of multi-source data and unstable health assessments caused by operating condition disturbances in existing technologies.

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

[0007] This invention provides an automated monitoring method for equipment anomalies based on artificial intelligence, comprising: reading multi-source measurement data frames, establishing measurement channel registration items, determining resolution quantities, performing timestamp cleaning, generating a unified time scale sequence, obtaining the original value sequence, constructing a continuous window sequence and performing robust normalization, constructing historical window library entries, and writing them into the historical window library; determining the normalized resolution quantity based on the historical window library, solidifying conditional dependency estimation rules, generating parent sets and causal chains, solidifying time delay tables, establishing a nearest neighbor library, generating an anchor quantity set and an anchor vector library, calibrating anomaly trigger thresholds, generating a benchmark window set and a recursive sequence, and calibrating maintenance thresholds; extracting anchor vectors from online time windows, assembling observed parent quantity vectors, generating fact predictions, calculating anomaly scores, retrieving anchor benchmark windows and performing anchor quantity replacement, generating counterfactual predictions, calculating disturbance sensitivity, degradation persistence, and discrimination ratio, and generating degradation markers; performing degradation marker gating based on degradation markers, calculating health indicators through degradation persistence, constructing a recent window set, and generating maintenance information in conjunction with maintenance thresholds.

[0008] As a preferred embodiment of the AI-based automated monitoring method for equipment anomalies described in this invention, the acquisition of the original value sequence includes: reading multi-source measurement data frames, establishing measurement channel registration items, writing them into a channel table, determining the resolution quantity, performing timestamp cleaning, and generating a cleaned timestamp sequence and a cleaned original value sequence; calculating the common coverage starting point based on the cleaned timestamp sequence, fixing the unified sampling interval, and generating a unified time scale sequence; and performing unified sampling of previous values ​​based on the cleaned timestamp sequence and the cleaned original value sequence, combined with the unified time scale sequence, to generate the original value sequence.

[0009] As a preferred embodiment of the AI-based automated monitoring method for device anomalies described in this invention, the step of writing to the historical window library includes: constructing a continuous window sequence based on a unified time scale sequence, extracting original time window segments from the original value sequence, combining the resolution quantity, performing robust normalization, constructing historical window library entries, and writing them to the historical window library.

[0010] As a preferred embodiment of the AI-based automated monitoring method for equipment anomalies described in this invention, the solidified time delay table includes: generating a channel window sequence and a set of positive differences of window sequences based on a historical window library, determining the normalized resolution of each channel; calculating the number of dependent nearest neighbors, solidifying the conditional dependency estimation rule, evaluating candidate parent channels, generating a parent set and a causal chain, and solidifying the time delay table.

[0011] As a preferred embodiment of the AI-based automated monitoring method for equipment anomalies described in this invention, the generation of the anchor set and anchor vector library includes: assembling the parent vector according to the parent set and the time delay table, writing it into the nearest neighbor library and fixing the nearest neighbor distance calculation method, combining the causal chain, calculating the anchor score, and generating the anchor set and anchor vector library.

[0012] As a preferred embodiment of the AI-based automated monitoring method for equipment anomalies described in this invention, the calibration of maintenance thresholds includes: calculating anomaly scores based on a nearest neighbor database and normalized resolution, combined with a parent set and a time delay table; calibrating anomaly trigger thresholds and generating a baseline window set; performing acyclic processing and topological sorting of causal chains to generate a recursive sequence; combining the baseline window set; performing anchor baseline window retrieval and counterfactual recursion to generate counterfactual predictions; calculating degradation persistence and converting it into health indicators; and calibrating maintenance thresholds.

[0013] As a preferred embodiment of the AI-based automated monitoring method for equipment anomalies described in this invention, the calculation of the anomaly score includes: receiving an online time window, extracting normalized sequence fragments, generating a channel window representative value, assembling the anchor vector of the online time window according to the channel order of the anchor set; assembling the observed parent vector based on the parent set and the time delay table, retrieving neighbor entries from the neighbor database, generating a fact prediction, and calculating the anomaly score.

[0014] As a preferred embodiment of the AI-based automated monitoring method for device anomalies described in this invention, the generation of degradation markers includes: performing a judgment based on an anomaly score and an anomaly trigger threshold; when the anomaly score is less than the anomaly trigger threshold, writing a counterfactual prediction as a factual prediction, and writing perturbation sensitivity, degradation persistence, discrimination ratio, and degradation marker; when the anomaly score is greater than or equal to the anomaly trigger threshold, reading the baseline window set and anchor vector library, combining the anchor vectors of the online time window, calculating the nearest neighbor distance of the anchor vectors, and determining the anchor baseline window index; performing anchor quantity replacement based on the anchor baseline window index, anchor quantity set, and recursive sequence, assembling the counterfactual parent vector according to the recursive sequence, retrieving the nearest neighbor library to generate a counterfactual prediction, calculating perturbation sensitivity, degradation persistence, and discrimination ratio, and generating a degradation marker.

[0015] As a preferred embodiment of the AI-based automated monitoring method for equipment anomalies described in this invention, the calculation of health indicators includes: reading degradation markers and performing degradation marker gating; marking time windows where degradation is marked as operating condition disturbances as not updating health indicators; calculating health indicators based on degradation persistence for time windows where degradation is marked as true degradation; writing health indicators into a historical window database; and generating online time windows with health indicators.

[0016] As a preferred embodiment of the AI-based automated monitoring method for equipment anomalies described in this invention, the generation of maintenance information includes: based on the window index of online time windows with health indicators, tracing back historical window database entries, constructing a recent window set, and calculating the robust change rate of health indicators; based on the robust change rate of health indicators and combined with maintenance thresholds, calculating the remaining time prediction, and generating maintenance information.

[0017] The beneficial effects of this invention are as follows: by constructing a parent set and causal chain, solidifying the time delay table, and combining the generation of anchor sets and anchor vector libraries with the calibration of anomaly trigger thresholds, the dependencies between multi-source channels can be structurally accumulated, providing a unified reference for online inference, and making anomaly trigger judgment more consistent and easier to trace; by generating counterfactual predictions through anchor replacement, and integrating perturbation sensitivity, degradation persistence, and discrimination ratio to generate degradation markers, clear explanations and stable discrimination of anomaly causes can be achieved, improving the continuity of health indicator evolution and the reliability of maintenance information. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1This is a flowchart of an AI-based automated monitoring method for device anomalies.

[0020] Figure 2 A flowchart for building a historical window library.

[0021] Figure 3 A flowchart for calibrating maintenance thresholds.

[0022] Figure 4 A flowchart for online time window processing.

[0023] Figure 5 This is a sequence diagram for triggering abnormal scoring.

[0024] Figure 6 This is a distribution map of scattered points in the genetic region. Detailed Implementation

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0026] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0027] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0028] Reference Figures 1-6 This is one embodiment of the present invention, which provides an automated monitoring method for device anomalies based on artificial intelligence, including the following steps:

[0029] S1. Read multi-source measurement data frames, establish measurement channel registration items, determine resolution quantities, perform timestamp cleaning, generate a unified time scale sequence, obtain the original value sequence, construct a continuous window sequence and perform robust normalization, construct historical window library entries, and write them into the historical window library.

[0030] Read multi-source measurement data frames, establish measurement channel registration items, write them into the channel table, determine the resolution quantity, perform timestamp cleaning, and generate a cleaned timestamp sequence and a cleaned original value sequence.

[0031] Furthermore, data frames from multiple sources, such as vibration, temperature, electrical parameters, pressure, and flow, are read. For each multi-source measurement, the field name and field type are extracted, and the existence of timestamp and numerical fields is confirmed.

[0032] For each multi-source measurement, create a measurement channel registration entry. The fields written in the measurement channel registration entry include channel identifier, physical unit, timestamp field name, numerical field name, sampling cycle description, communication protocol type, and range description. Write all measurement channel registration entries into the channel table, and record the channel table generation time and channel table version number in the channel table.

[0033] Furthermore, the raw data streams of multi-source measurements are read one by one according to the channel table, and the timestamp field and numerical field are extracted to form the raw sample sequence. The raw sample sequence is sorted in ascending order by timestamp, and samples with missing timestamps and missing values ​​are deleted.

[0034] For the sorted original sample sequence, calculate the difference between two adjacent original values ​​to form an original difference sequence. Select positive differences from the original difference sequence to form a set of positive differences. Select the smallest positive difference from the set of positive differences to determine the resolution value and write the resolution value into the resolution value field of the channel table.

[0035] It should be noted that when the set of positive difference values ​​is empty, the minimum distinguishable increment is read from the sampling beat description of the channel table or the communication protocol parameters to determine the resolution.

[0036] Furthermore, the original sample sequence is read one by one according to the channel table, and then sorted again in ascending order of timestamp. For duplicate samples with the same timestamp, the last arriving sample is retained and the rest of the duplicate samples are discarded. The timestamp backtracking samples are marked as invalid samples and removed from the alignment calculation. Invalid samples are retained in the original log for traceability. The cleaned timestamp sequence and the cleaned original value sequence are generated. The cleaned original value sequence and the cleaned timestamp sequence correspond one-to-one according to the sample order.

[0037] Based on the cleaned timestamp sequence, the common coverage starting point is calculated, a unified sampling interval is fixed, and a unified time scale sequence is generated. Based on the cleaned timestamp sequence and the cleaned original value sequence, combined with the unified time scale sequence, the previous value is sampled in a unified manner to generate the original value sequence.

[0038] Furthermore, read the cleaned timestamp sequence of all channels, and count the earliest and latest timestamps of each cleaned timestamp sequence; calculate the common coverage start point, which is the maximum value among all the earliest timestamps; the common coverage start point ensures that there are valid samples after the common coverage start point for each measurement.

[0039] Read the sampling cycle description from the channel table and select a uniform sampling interval. Specifically, the uniform sampling interval is an integer multiple of the minimum sampling interval of the multi-source measurements to ensure that the uniform sampling interval can cover all channels.

[0040] Furthermore, starting from the common coverage point and using the unified sampling interval as the step size, a unified time scale sequence is generated and written into the unified time reference record table. The unified time reference record table is used to store the unified time scale sequence and the unified sampling interval. The unified time scale sequence and the unified sampling interval constitute the unified time reference. The unified time reference record table is written into the historical window library metadata area.

[0041] Read the cleaned timestamp sequence and the cleaned original value sequence one by one according to the channel table. For each time scale point of the unified time scale sequence, perform uniform sampling of the previous value. Specifically, find the nearest timestamp not greater than the time scale point in the cleaned timestamp sequence; take the original value corresponding to the nearest timestamp and fill it into the original value sequence position of the time scale point; for time scale points that still lack valid samples after the common coverage starting point, mark the missing measurement status and keep the missing measurement status; the missing measurement status participates in the recording as a missing measurement flag when constructing the subsequent sliding time window, generating the original value sequence corresponding to each measurement. The original value sequence is indexed by the unified time scale sequence and retains the missing measurement flag.

[0042] Based on a unified time scale sequence, a continuous window sequence is constructed. The original time window segments are extracted from the original value sequence, and robust normalization is performed in combination with the resolution value to construct historical window library entries and write them into the historical window library.

[0043] Furthermore, a time window length is set on the unified time scale sequence, the time window length being the time span corresponding to at least 30 unified sampling points, and a sliding step size is set on the unified time scale sequence, the sliding step size being the step size corresponding to a unified sampling interval.

[0044] Starting from the beginning of a unified time scale sequence, continuous time scale points are extracted according to the length of the time window to form the first time window; the time window records the start and end times of the time window and the set of indexes of the time scale points contained in the time window.

[0045] The time window is advanced along a uniform time scale sequence with a sliding step size, and the time window is repeatedly constructed to form a continuous window sequence. For each time window, a segment of the original value within the coverage area of ​​the time window is extracted from the original value sequence to form the original segment of the time window. The original segment of the time window contains the missing value flag and retains the missing value position index.

[0046] For each time window in the continuous window sequence, read the original segments of the time window one by one, and perform robust scaling estimation on the original segments of the time window. Specifically, calculate the median of the original segments of the time window, calculate the absolute value sequence of the difference between the original segments of the time window and the median, and use the median of the absolute value sequence as the median absolute deviation.

[0047] The resolution values ​​are read from the channel table, and normalization calculations are performed on each original value of the original segment of the time window to generate a normalized sequence of values.

[0048] Specifically, the normalized sequence values ​​are represented as follows:

[0049] ;

[0050] in, Indicates the values ​​of the normalized sequence. This represents the original value of the time window. This indicates the resolution.

[0051] It should be noted that 1.4826 is the consistency conversion factor from the median absolute deviation to the standard deviation scale. 1.4826 is taken as the reciprocal of the 0.67449 quantile corresponding to the 0.75 percentile of the standard normal distribution, so that the median absolute deviation conversion result is consistent with the standard deviation estimation result under the condition that the error is approximately normally distributed.

[0052] For each time window in the continuous window sequence, construct a historical window database entry. The fields written to the historical window database entry include the window index, the start and end time of the time window, the normalized sequence segment, the missing measurement location index set, the channel table version number, and the resolution reference identifier.

[0053] Historical window entries are written to the historical window database in ascending order of the window index. The historical window database is written using an append-only method to ensure that the historical window database is irreversible and easy to audit and review.

[0054] S2. Based on the historical window library, determine the normalized resolution quantity, solidify the conditional dependency estimation rules, generate the parent set and causal chain, solidify the time delay table, establish the nearest neighbor library, generate the anchor set and anchor vector library, calibrate the anomaly trigger threshold, generate the benchmark window set and recursive sequence, and calibrate the maintenance threshold.

[0055] Based on the historical window library, generate channel window sequences and a set of positive differences in window sequences, determine the normalized resolution of each channel, calculate the dependent nearest neighbor number, solidify the conditional dependency estimation rule, evaluate candidate parent channels, generate parent sets and causal chains, and solidify the time delay table.

[0056] Furthermore, the historical window database metadata area is read to obtain the channel set, window index order, time window start and end time, unified sampling interval, and number of time window samples. The historical window database entry area is read, and the normalized sequence fragments corresponding to each channel are extracted window by window in ascending order of window index. A channel window sequence is generated for each channel. Specifically, the median of the normalized sequence fragments within each time window is taken, and the median is used as the representative value of the channel window. The representative values ​​of the channel windows corresponding to each time window are concatenated in the order of window index to form the channel window sequence.

[0057] For each channel, a set of positive difference values ​​for the window sequence is generated. Specifically, the absolute value of the difference between two adjacent windows in the channel window sequence is taken, and the differences that are greater than zero are filtered out to form a set of positive difference values ​​for the window sequence.

[0058] Determine the normalized resolution for each channel. Specifically, when the set of positive differences in the window sequence contains elements, take the minimum positive difference as the normalized resolution; when the set of positive differences in the window sequence is empty, take the reciprocal of the number of time window samples as the normalized resolution. Write the normalized resolution for each channel into the historical window database metadata area.

[0059] Furthermore, read the normalized resolution values ​​of all channels in the historical window database metadata area, take the median of the normalized resolution values ​​of all channels, and determine the ratio denominator protection value; write the ratio denominator protection value into the historical window database metadata area.

[0060] Furthermore, the number of time windows in the historical window database is read, and the number of dependent neighbors is calculated. Specifically, the square root of the number of time windows is taken and rounded down. When the number of dependent neighbors is less than 10, it is taken as 10, and when the number of dependent neighbors is greater than 200, it is taken as 200.

[0061] Specifically, the number of dependent nearest neighbors is represented as:

[0062] ;

[0063] in, Indicates dependency on nearest neighbors. This indicates the number of time windows in the historical window database.

[0064] It should be noted that the lower bound for the number of dependent nearest neighbors is set to 10 to ensure statistical stability of the nearest neighbor count estimate at the level of the number of neighbors; the upper bound for the number of dependent nearest neighbors is set to 200 to limit the computational cost of nearest neighbor retrieval and to limit the nonlocality of the neighborhood caused by radius expansion; these are the constraints. Ensure that the number of dependent neighbors does not exceed one-fifth of the number of time windows in the historical window database, and ensure that the neighborhood size used for neighbor count estimation is proportional to the size of the historical window database.

[0065] The fixed conditional dependency estimation rule is the nearest neighbor count estimation. The input for the nearest neighbor count estimation is fixed as the parent channel window sequence, the child channel window sequence, the conditional channel window sequence, and the number of dependent nearest neighbors. The output of the nearest neighbor count estimation is fixed as the conditional dependency value. The number of dependent nearest neighbors and the conditional dependency estimation rule are written into the metadata area of ​​the historical window library.

[0066] Specifically, for each time window index in the historical window database, the parent channel window vector, child channel window vector, and conditional channel window vector are extracted; the joint window vector is obtained by concatenating the parent channel window vector, child channel window vector, and conditional channel window vector; the parent conditional window vector is obtained by concatenating the parent channel window vector and conditional channel window vector; the child conditional window vector is obtained by concatenating the child channel window vector and conditional channel window vector; and the conditional window vector is taken from the conditional channel window vector.

[0067] In the joint window vector space, a distance metric is used that averages the absolute differences of each component. The nth nearest neighbor is retrieved from the joint window vector space, and the distance between the nth nearest neighbor is taken as the radius. The radius is used to count the number of neighbors in the parent conditional window vector space to obtain the parent conditional neighbor count. The radius is used to count the number of neighbors in the child conditional window vector space to obtain the child conditional neighbor count. The radius is used to count the number of neighbors in the conditional window vector space to obtain the conditional neighbor count. The parent conditional neighbor count, child conditional neighbor count, and conditional neighbor count all count the number of window indices whose distance is not greater than the radius. Items whose window index is equal to the current time window index are not included in the neighbor count.

[0068] For all time window indexes, a local logarithmic ratio is generated based on the parent conditional neighbor count, child conditional neighbor count, and conditional neighbor count. The local logarithmic ratio is then used to obtain the conditional dependency value by taking the median of the time window index.

[0069] Specifically, the conditional dependency is represented numerically as follows:

[0070] ;

[0071] in, This indicates that the condition depends on the value. Indicates conditional neighbor counting. Indicates the parent condition neighbor count. This indicates the subconditional neighbor count.

[0072] Furthermore, the channel set, channel window sequence, conditional dependency estimation rule, uniform sampling interval, and time window length are read; the maximum time delay step is calculated. Specifically, the ratio of the time window length to the uniform sampling interval is calculated and rounded to obtain the maximum time delay step, which is at least 1.

[0073] Select one channel from the channel set as a child channel to construct a candidate parent channel set. The candidate parent channel set is obtained by removing the child channels from the channel set.

[0074] Perform candidate evaluation for each parent channel in the candidate parent channel set. Specifically, calculate the conditional dependency values ​​one by one within the time delay step range from 1 to the maximum time delay step, and record the time delay step corresponding to the maximum conditional dependency value and the maximum value of the conditional dependency value.

[0075] The candidate parent channel set is sorted from largest to smallest according to the maximum value of the conditional dependency, forming a candidate sorting table. The parent set increment is generated. Specifically, starting from the first parent channel in the candidate sorting table, the parent set is tried to be added in turn. After the parent set is added, the channel window sequence corresponding to the parent set is used as part of the conditional channel window sequence. The maximum value of the conditional dependency of the parent channel that has not been added to the parent set is recalculated. The addition of the parent set is stopped when all the maximum values ​​of the conditional dependency of the parent channel that has not been added to the parent set are less than or equal to 0.

[0076] Write the parent set corresponding to the child channel into the metadata area of ​​the history window library, write the directed relationship between the parent channel and the child channel into the causal chain, and write the parent channel identifier, child channel identifier, maximum value of the condition dependency, and the time delay steps corresponding to the maximum value of the condition dependency into the causal chain entries.

[0077] Read the causal chain and conditional dependency estimation rules. For each directed relationship from a parent channel to a child channel in the causal chain, calculate the conditional dependency values ​​one by one within the maximum time delay step range. Select the time delay step corresponding to the conditional dependency value reaching the maximum as the solidified time delay step. Write the solidified time delay step into the time delay table. Write the time delay table into the historical window library metadata area.

[0078] Based on the parent set and the time delay table, the parent vector is assembled, written into the nearest neighbor library, and the nearest neighbor distance calculation method is fixed. Combined with the causal chain, the anchor score is calculated, and the anchor set and anchor vector library are generated.

[0079] Furthermore, the parent set and time delay table are read, and the nearest neighbor entries are written for each time window and each sub-channel in the historical window library.

[0080] Furthermore, the parent vector is assembled in the order of the parent set. Specifically, for each parent channel in the parent set, the value of the parent channel window sequence at the position of window index minus the solidification delay step is read and written into the parent vector components in the order of the parent set.

[0081] Read the value of the sub-channel window sequence at the window index position as the sub-quantity value, and write the parent vector, sub-quantity value, and window index into the nearest neighbor database.

[0082] The nearest neighbor distance calculation method is solidified. Specifically, for any two parent vectors, the absolute difference is calculated component by component, all absolute differences are accumulated, and the ratio of the accumulated absolute difference to the dimension of the parent vector is calculated to obtain the nearest neighbor distance.

[0083] Furthermore, the causal chain entries are read, the number of child channels pointed to by each channel as a parent channel is counted, and the anchor score of each channel is calculated. Specifically, the maximum value of the conditional dependency of all causal chain entries corresponding to the channel as a parent channel is read, and the median of the set of maximum value of conditional dependency is taken to obtain the anchor score.

[0084] Take the median of all anchor scores in the channel set to obtain the anchor score boundary; select the channel set whose anchor scores are greater than or equal to the anchor score boundary to generate the anchor value set; write the anchor value set into the metadata area of ​​the historical window database.

[0085] Generate an anchor vector library. Specifically, for each window index in the historical window library, read the values ​​of the anchor vector set channel window sequence at the window index position according to the anchor vector set channel order, and assemble the anchor vectors in order; write the anchor vectors and window indexes into the anchor vector library.

[0086] Based on the nearest neighbor database and normalized resolution, combined with the parent set and time delay table, anomaly scores are calculated, anomaly trigger thresholds are calibrated, and a baseline window set is generated. The causal chain is decycliced ​​and topologically sorted to generate a recursive sequence. Combined with the baseline window set, anchor baseline window retrieval and counterfactual recursion are performed to generate counterfactual predictions. Degradation persistence is calculated and health indicators are converted, and maintenance thresholds are calibrated.

[0087] Furthermore, the nearest neighbor database, nearest neighbor distance, normalized resolution, channel window sequence, parent set, time delay table, and number of dependent nearest neighbors are read, and anomaly score calculation is performed on each window index of the historical window database.

[0088] For each sub-channel in the channel set, fact prediction calculation is performed. Specifically, the parent vector is assembled according to the parent set and the time delay table; the nearest distance between the parent vector and the parent vector in the nearest neighbor library is calculated; the number of dependent nearest neighbor entries is selected in ascending order of nearest distance; the median of the sub-value set of the nearest neighbor entries is taken to obtain the fact prediction; and the anomaly score is calculated by using the fact prediction of all channels and the representative value of the channel window.

[0089] Specifically, the anomaly score and the anomaly trigger threshold are expressed as follows:

[0090] ;

[0091] ;

[0092] in, Indicates abnormal rating. Indicates window index, Indicates the channel index. This indicates the value represented by the channel window. Indicates factual predictions, Indicates the first Path normalized resolution quantity, Indicates the threshold for triggering an anomaly. This represents the higher quantile operator. Indicates the higher quantile parameter, subscript Indicates the abnormal trigger threshold marker, superscript Indicates factual prediction markers.

[0093] Furthermore, the abnormal scores of all window indexes are combined into an abnormal score sample sequence, and the high percentile of the abnormal score sample sequence is used as the abnormal trigger threshold; the set of window indexes with abnormal scores less than the abnormal trigger threshold is selected to generate a benchmark window set, which is written into the historical window database metadata area.

[0094] Furthermore, the number of entries in the baseline window set in the historical window database metadata area is read, and the number of backtracking windows is calculated. Specifically, the square root of the number of entries in the baseline window set is taken and rounded down. When the number of backtracking windows is less than 10, it is taken as 10, and when the number of backtracking windows is greater than 200, it is taken as 200. The number of backtracking windows is then written into the historical window database metadata area.

[0095] It should be noted that, for example, in this embodiment, the high quantile parameter is set to 0.995. A high quantile parameter of 0.995 can reduce the false alarm rate in the context of industrial noise while retaining the ability to trigger significant anomalies, thus meeting the efficiency requirements of online monitoring. Specifically, the high quantile parameter is determined by the target false alarm rate. The value of the high quantile parameter is equal to 1 minus the target false alarm rate value. The target false alarm rate is determined by the operation and maintenance strategy. The operation and maintenance strategy sets an upper limit on the number of false alarms allowed per unit time. The target false alarm rate is the ratio of the upper limit on the number of false alarms allowed to the number of time windows participating in anomaly judgment per unit time.

[0096] Read the causal chain, perform directed acyclic processing, detect directed loops in the causal chain, locate the directed relation with the smallest conditional dependency value within the directed loop, and delete the directed relation with the smallest conditional dependency value; repeat detection and deletion until there are no directed loops in the causal chain; perform topological sorting on the acyclic causal chain to obtain the channel recursive order; write the channel recursive order into the recursive sequence and store it in the historical window database metadata area.

[0097] Read the baseline window set, anchor vector library, anchor quantity set, recursive sequence, nearest neighbor library, and normalized resolution quantity. Perform anchor baseline window retrieval for each time window in the baseline window set. Specifically, calculate the anchor vector nearest neighbor distance in the baseline window set, and select the time window with the smallest anchor vector nearest neighbor distance and smaller window index as the anchor baseline window.

[0098] Furthermore, anchor replacement is performed. Specifically, the channel window representative value of the anchor set channel within the time window is replaced with the corresponding value of the anchor reference window. The parent vector is assembled channel by channel according to the recursive sequence, and the median is taken from the nearest neighbor library to generate counterfactual predictions.

[0099] Furthermore, the degradation persistence is calculated, converted into a health index, and the low quantile of the health index sample sequence of the baseline window set is taken to obtain the maintenance threshold; the maintenance threshold is then written into the metadata area of ​​the historical window library.

[0100] Specifically, the persistence of degradation and health indicators are expressed as follows:

[0101] ;

[0102] ;

[0103] in, Indicates degradation durability. Indicates the first Road contrary to fact prediction, superscript Indicates a counterfactual prediction marker. Indicates health indicators.

[0104] It should be noted that, for example, in this embodiment, the value of the low quantile parameter is 0.05, which can cover about 95% of normal fluctuations, suppress occasional noise false triggers, and maintain trigger sensitivity in the early stage of actual degradation. Specifically, the low quantile parameter is selected by calculating the maintenance threshold for each candidate low quantile parameter set based on the health index sample sequence of the benchmark window set, and statistically analyzing the frequency of maintenance information generation in the benchmark window set replay evaluation. The candidate low quantile parameter that satisfies the condition that the frequency of maintenance information generation does not exceed the upper limit of the allowable false triggering limit of maintenance and has the fewest number of false triggering of maintenance information is selected.

[0105] S3. Extract the anchor vector of the online time window, assemble the observation parent vector, generate fact prediction, calculate the anomaly score, retrieve the anchor reference window and perform anchor replacement, generate counterfactual prediction, calculate the perturbation sensitivity, degradation persistence and discrimination ratio, and generate degradation marker.

[0106] Receive the online time window, extract the normalized sequence fragment, generate the channel window representative value, and assemble the anchor vector of the online time window according to the channel order of the anchor set; assemble the observation parent vector according to the parent set and the time delay table, retrieve the neighbor entries in the neighbor database, generate fact prediction, and calculate the anomaly score.

[0107] Furthermore, an online time window is received, which corresponds to a window index on a unified time base. The unified time base is determined by the unified time base record table in the metadata area of ​​the historical window library. The normalized sequence segment corresponding to the online time window is read from the historical window library. For each channel, the median is calculated within the normalized sequence segment of the online time window to generate a channel window representative value. The window index and the channel window representative value are written into the online time window entry in the historical window library.

[0108] The system reads the channel window representative value from the online time window entries in the historical window database, reads the anchor set from the metadata area of ​​the historical window database, and extracts the corresponding channel window representative value for each online time window in the channel order of the anchor set to assemble the anchor vector of the online time window. The window index and anchor vector are written to the anchor vector database query cache area. The anchor vector database query cache area is only used for anchor base window retrieval and does not change the historical entries of the anchor vector database.

[0109] Read the parent set and time delay table from the historical window database metadata area. Perform observation parent vector assembly for each sub-channel. Specifically, process the parent channels one by one in the order of the parent set channels. For each parent channel, read the number of solidified time delay steps from the parent channel to the sub-channel from the time delay table. Calculate the parent channel value window index. The parent channel value window index is the online time window index minus the number of solidified time delay steps. Read the channel window representative value of the time window entry corresponding to the parent channel value window index from the historical window database and write it to the corresponding component of the observation parent vector.

[0110] The nearest neighbor database, the number of dependent nearest neighbors, and the calculation method of the nearest neighbor distance are read from the metadata area of ​​the historical window database. For each sub-channel, fact prediction generation is performed. Specifically, the observed parent vector is used as the search key to calculate the nearest neighbor distance between the observed parent vector and all parent vectors in the nearest neighbor database. The nearest neighbor distances are sorted in ascending order. The number of dependent nearest neighbor entries is selected, and the median value of the sub-quantity corresponding to the number of dependent nearest neighbor entries is calculated to obtain the fact prediction of the sub-channel. The fact prediction of the entire channel is written into the online time window entries of the historical window database.

[0111] Read the normalized resolution and anomaly trigger threshold from the metadata area of ​​the historical window library. Read the channel window representative value and fact prediction from the online time window entries in the historical window library. Calculate the anomaly score for each channel according to the anomaly scoring formula and write the anomaly score into the online time window entries in the historical window library.

[0112] Based on the anomaly score and the anomaly trigger threshold, a judgment is made. If the anomaly score is less than the anomaly trigger threshold, the counterfactual prediction is written as the factual prediction, and the perturbation sensitivity, degradation persistence, discrimination ratio, and degradation flag are written. If the anomaly score is greater than or equal to the anomaly trigger threshold, the baseline window set and anchor vector library are read, and the anchor vectors of the online time window are combined to calculate the nearest neighbor distance of the anchor vectors and determine the anchor baseline window index.

[0113] Furthermore, if the anomaly score is less than the anomaly trigger threshold, the counterfactual prediction field is written as the fact prediction field, the perturbation sensitivity, degradation persistence, and discrimination ratio are all written as zero, and the degradation flag is written as the operating condition perturbation; the online time window processing flow ends, and the process waits for the next online time window to arrive.

[0114] If the abnormal score is greater than or equal to the abnormal trigger threshold, the baseline window set and anchor vector library are read from the metadata area of ​​the historical window library; the anchor vector of the online time window is read from the query cache area of ​​the anchor vector library; the candidate window index is read one by one from the baseline window set; for each candidate window index, the candidate anchor vector is read from the anchor vector library; and the anchor vector nearest neighbor distance between the anchor vector of the online time window and the candidate anchor vector is calculated.

[0115] The candidate window index with the smallest nearest neighbor distance of the anchor vector is selected as the anchor reference window index. If the nearest neighbor distance of the anchor vector has the same minimum value, the candidate window index with the smaller window index is selected as the anchor reference window index, and the anchor reference window index is written into the online time window entry of the historical window database.

[0116] Based on the anchor reference window index, anchor set, and recursive sequence, anchor replacement is performed, counterfactual parent vectors are assembled according to the recursive sequence, counterfactual predictions are generated by retrieving the nearest neighbor library, perturbation sensitivity, degradation persistence, and discrimination ratio are calculated, and degradation markers are generated.

[0117] Furthermore, the channel window representative value corresponding to the time window entry of the anchor reference window index is read from the historical window library, the anchor quantity set is read, and anchor quantity replacement is performed on each channel in the anchor quantity set. Specifically, the channel window representative value corresponding to the online time window is replaced with the channel window representative value corresponding to the anchor reference window, and the channel window representative value of the anchor quantity set after anchor quantity replacement is written into the counterfactual input buffer.

[0118] Read the recursive sequence, parent set, time delay table, nearest neighbor database, number of dependent nearest neighbors, and nearest neighbor distance calculation method from the historical window database metadata area, and initialize the counterfactual prediction set.

[0119] Furthermore, channels are processed one by one in the order of the recursive sequence. Specifically, the counterfactual input buffer is read to determine whether the current channel of the recursive sequence belongs to the anchor set.

[0120] When the current channel of the recursive sequence belongs to the anchor set, the channel window representative value after anchor replacement is read from the counterfactual input buffer and written to the corresponding channel position in the counterfactual prediction set. When the current channel of the recursive sequence does not belong to the anchor set, the counterfactual parent vector assembly is performed. Specifically, the parent channels are processed one by one according to the order of the parent set channels. When the parent channel belongs to the anchor set, the parent channel window representative value after anchor replacement is read from the counterfactual input buffer and written to the counterfactual parent vector component. When the parent channel does not belong to the anchor set, the counterfactual prediction of the parent channel is read from the counterfactual prediction set and written to the counterfactual parent vector component to obtain the counterfactual parent vector.

[0121] Using the counterfactual parent vector as the search key, calculate the nearest neighbor distance between the counterfactual parent vector and all parent vectors in the nearest neighbor database. Sort the nearest neighbor distances from smallest to largest and select several nearest neighbor entries that depend on the nearest neighbor. Calculate the median of the values ​​of the sub-quantities corresponding to the several nearest neighbor entries that depend on the nearest neighbor to obtain the counterfactual prediction for the current channel of the recursive sequence. Write the counterfactual prediction for the current channel of the recursive sequence into the corresponding channel position of the counterfactual prediction set, and write the counterfactual prediction set into the online time window entries of the historical window database.

[0122] Read the normalized resolution and ratio denominator protection from the historical window database metadata area, and read the channel window representative value, fact prediction, and counterfactual prediction from the online time window entries in the historical window database.

[0123] For each channel, the difference between factual prediction and counterfactual prediction, and the difference between observed value and counterfactual prediction are calculated separately. The median value is then taken across all channels to form the perturbation sensitivity and degradation persistence. The discrimination ratio is then further calculated.

[0124] Specifically, the perturbation sensitivity and the discrimination ratio are expressed as follows:

[0125] ;

[0126] ;

[0127] in, Indicates sensitivity to disturbances. Indicates the discriminant ratio. This represents the protection quantity in the denominator of the ratio.

[0128] A degradation marker is generated. Specifically, when the discrimination ratio is greater than or equal to 1, the degradation marker is written to the actual degradation; when the discrimination ratio is less than 1, the degradation marker is written to the operating condition disturbance.

[0129] Perturbation sensitivity, degradation persistence, discrimination ratio, and degradation flag are written into the online time window entries of the history window library.

[0130] S4. Based on the degradation marker, perform degradation marker gating, calculate health indicators through degradation persistence, construct a recent window set, and generate maintenance information by combining the maintenance threshold.

[0131] Read the degradation flag and perform degradation flag gating. Mark the time window that is marked as a condition disturbance as not updating the health index. For the time window that is marked as true degradation, calculate the health index based on the degradation persistence and write the health index into the historical window library to generate an online time window with health index.

[0132] Furthermore, the corresponding entries for online time windows in the historical window database are read, including the window index, start and end times of the time window, degradation flag, and degradation durability field. The degradation flag and degradation durability fields are checked for valid values. If the degradation flag field is missing, the degradation flag value is written to the historical window database as "operating condition disturbance". If the degradation durability field is missing, the degradation durability value is written to the historical window database as "0".

[0133] The system reads the maintenance threshold and backtracking window number from the historical window database metadata area, compares the degradation flag values ​​of online time window entries in the historical window database. Specifically, if the degradation flag value is "operating condition disturbance," it reads the health indicator field and remaining time prediction field of the previous time window entry in the historical window database, reads the time window end time field of the previous time window entry and online time window entry in the historical window database, writes the health indicator field of the previous time window entry into the health indicator field of the online time window entry, reads the remaining time prediction field of the previous time window entry as the time advancement update input, and writes it into the remaining time prediction field of the online time window entry according to the remaining time prediction time advancement update rules. The maintenance information generation flag for writing to online time window entries in the historical window database is set to "not generate."

[0134] Specifically, the remaining time prediction time advancement update rules are as follows: when the remaining time prediction field of the previous time window is positive infinity, the remaining time prediction field of the online time window is set to positive infinity; when the remaining time prediction field of the previous time window is 0, the remaining time prediction field of the online time window is set to 0; when the remaining time prediction field of the previous time window is a finite positive number, the remaining time prediction field of the online time window is set to the difference between the remaining time prediction field of the previous time window and the end time field of the time window. If the result is less than zero, the remaining time prediction field of the online time window is set to 0. The difference between the end time field of the time window and the end time field of the previous time window is taken as the end time field of the online time window minus the end time field of the previous time window; when the difference between the end time field of the time window and the end time field of the previous time window is less than 0, the difference between the end time field of the time window is set to 0.

[0135] If the degradation flag is set to true degradation, then the health index is calculated. Specifically, the degradation persistence field of the online time window entry in the historical window library is read, the health index is calculated according to the health index formula, and written into the health index field of the online time window entry in the historical window library. The health index generation flag is then written into the online time window entry in the historical window library as generated.

[0136] Based on the window index of online time windows with health indicators, backtrack the historical window database entries, construct a recent window set, and calculate the robust change rate of health indicators; based on the robust change rate of health indicators and the maintenance threshold, calculate the remaining time prediction and generate maintenance information.

[0137] Furthermore, the backtracking window count value is read from the historical window database metadata area. Starting from the online time window index, the process backtracks backward according to the window index, reading the health indicator field and time window start and end time field of each historical window database entry. Historical window database entries with health indicator generation flags that have been generated are selected, and backtracking stops when the cumulative number of selected entries reaches the backtracking window count. The recent window set is constructed using the selected set of historical window database entries. The time window end time and health indicator of each entry in the recent window set are read to form a set of time pairs and health pairs.

[0138] Furthermore, the set of time pairs and health pairs is read, and for any two historical window entries in the recent window set, the slope is calculated. Specifically, the slope is the ratio of the difference in health indicators between the two historical window entries to the difference in the end time of the time window, where the unit of the difference in the end time of the time window is seconds.

[0139] Furthermore, all slopes are aggregated to form a slope set, the median of the slope set is taken, the robust change rate of health indicators is obtained, and written into the change rate field of online time window entries in the historical window library.

[0140] Specifically, the robust rate of change of health indicators is expressed as:

[0141] ;

[0142] in, This indicates the rate of steady change in health indicators. The window index is Health indicators The window index is Health indicators The window index is End time of the time window The window index is End time of the time window This represents the set of recent windows.

[0143] Read the maintenance threshold value from the metadata area of ​​the historical window database, and read the health indicator field and change rate field of the online time window entry in the historical window database. When the health indicator is less than or equal to the maintenance threshold, write a value of 0 to the remaining time prediction field of the online time window entry in the historical window database. When the health indicator field value is greater than the maintenance threshold and the change rate field value is less than 0, calculate the remaining time prediction by extrapolation and write it to the remaining time prediction field of the online time window entry in the historical window database. When the health indicator field value is greater than the maintenance threshold and the change rate field value is greater than or equal to 0, write a value of positive infinity to the remaining time prediction field.

[0144] Specifically, the remaining time forecast is expressed as:

[0145] ;

[0146] in, Indicates the remaining time forecast. This indicates the maintenance threshold.

[0147] Furthermore, the maintenance threshold and health indicators are read. If the health indicators are less than or equal to the maintenance threshold, a maintenance information entry is generated and written to the maintenance information list. The fields written to the maintenance information entry include the start and end time of the time window, the window index, the degradation flag, the degradation persistence, the health indicators, the robust change rate of the health indicators, and the remaining time prediction. The remaining time prediction field of the maintenance information entry is set to 0.

[0148] If the health indicator is greater than the maintenance threshold, no new entries will be written to the maintenance information list, and the flag for writing maintenance information to online time window entries in the historical window library will be set to not generate.

[0149] In this embodiment, simulation experiments were conducted to verify that the abnormal trigger threshold was calibrated based on the abnormal score distribution of normal stable segments, ensuring consistency in online trigger judgment. By generating counterfactual predictions through anchor replacement and combining disturbance sensitivity and degradation persistence, the ability to interpretably distinguish between operating condition disturbances and actual degradation was achieved. Specifically, the simulation detection environment is geared towards continuously operating rotating machinery, represented by a centrifugal pump-motor-bearing assembly. A multi-source channel monitoring scenario with a unified sampling period is adopted. Multiple operating conditions are set during online operation, including normal stable operation, pure operating condition disturbance (frequent step switching of speed and load), and pure degradation evolution. The simulation covers the entire chain of scenarios, including stable operating conditions but gradual degradation and mixed segments (superimposed disturbances and degradation), encompassing anomaly triggering, cause identification, and health indicator updates. To closely resemble engineering practice, the simulation incorporates inconsistent sampling, random missing measurements, blocky missing measurements, and differences in quantization resolution. Data is uniformly sampled, aligned, and missing measurements are supplemented. The anomaly triggering threshold calibration and online inference processes are performed under more realistic data quality conditions. The multi-source channel data consists of twelve monitoring parameters, including speed, load, motor current, motor voltage, bearing temperature, oil temperature, root mean square vibration, vibration kurtosis, pressure, flow rate, acoustics, and humidity.

[0150] The historical time window library is extracted from normal and stable segments, constructs parent sets and causal chains, solidifies time delay tables, and generates anchor vector libraries corresponding to anchor sets. Anomaly trigger thresholds are calibrated on the abnormal score distribution of normal windows. In the online stage, a time window sequence is formed according to a fixed time window length and sliding step size. The abnormal score, trigger flag, anchor reference window number, disturbance sensitivity, degradation persistence, discrimination ratio, degradation mark, health indicators, and remaining time prediction are recorded window by window.

[0151] Figure 5The diagram illustrates the overall temporal relationship between anomaly scores, anomaly trigger thresholds, and trigger flags calculated by time windows. The anomaly trigger threshold is determined by high-quantile calibration of the anomaly score distribution in normal stable segments, maintaining stability and serving as a unified trigger reference. In normal stable segments, the anomaly score is generally at a low level, with only a few isolated spikes and few trigger flags. In pure operating condition disturbance segments, frequent step switching of speed and load occurs, and the statistical characteristics of multiple source channels change synchronously, resulting in a significant increase in the anomaly score and dense fluctuations, generating more trigger flags. In pure degradation and mixed segments, the anomaly score maintains a high level over a longer period, accompanied by multiple peaks, reflecting the continuous impact of degradation evolution on channels such as vibration, temperature, and electrical parameters. Figure 5 This indicates that the online trigger determination is consistent by calibrating the abnormal trigger threshold based on the normal window distribution; in conjunction with the time window number and trigger flag record, each trigger can be located and traced, providing a trigger entry point for subsequent anchor reference window retrieval and counterfactual prediction.

[0152] Figure 6 Scatter plots of disturbance sensitivity and degradation persistence under only the trigger window condition were used to verify the interpretability and distinguishability of anomaly causes. Specifically, disturbance sensitivity represents the degree of difference between the factual prediction and the counterfactual prediction after anchor replacement. When the anomaly is mainly caused by operating condition disturbances, anchor replacement will significantly change the prediction results, and the disturbance sensitivity is usually high. Degradation persistence represents the mean of the residual between the observed value and the factual prediction. When the anomaly is driven by actual degradation, the residual has persistence and accumulation, and the degradation persistence is generally high.

[0153] Figure 6 The data is differentiated into pure operating condition disturbances and degradation cases using ground truth labels. The two types of point clouds show a clear separation trend on the plane. Specifically, operating condition disturbance samples are more distributed in areas with relatively low degradation persistence, while degradation samples are more distributed in areas with high degradation persistence, and their distribution range is more concentrated in the intervals of continuous deviation. Figure 6 This indicates that by generating counterfactual predictions through anchor replacement and integrating perturbation sensitivity and degradation persistence to form a discrimination criterion, it can stably identify the causes of anomalies and provide explanatory paths, laying a reliable foundation for degradation marking, health indicator gating updates, and maintenance information generation.

[0154] In summary, this invention constructs a parent set and causal chain, solidifies the time delay table, and combines the generation of anchor sets and anchor vector libraries with the calibration of anomaly trigger thresholds to structurally solidify the dependencies between multi-source channels, providing a unified reference for online inference, making anomaly trigger judgment more consistent and easier to trace. By generating counterfactual predictions through anchor replacement and integrating perturbation sensitivity, degradation persistence, and discrimination ratio to generate degradation markers, it achieves clear explanation and stable discrimination of anomaly causes, improving the continuity of health indicator evolution and the reliability of maintenance information.

[0155] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An automated monitoring method for equipment anomalies based on artificial intelligence, characterized in that, include: Read multi-source measurement data frames, establish measurement channel registration items, determine resolution quantities, perform timestamp cleaning, generate a unified time scale sequence, obtain the original value sequence, construct a continuous window sequence and perform robust normalization, construct historical window library entries, and write them into the historical window library. Based on the historical window library, determine the normalized resolution quantity, solidify the conditional dependency estimation rules, generate the parent set and causal chain, solidify the time delay table, establish the nearest neighbor library, generate the anchor set and anchor vector library, calibrate the anomaly trigger threshold, generate the benchmark window set and recursive sequence, and calibrate the maintenance threshold. Extract the anchor vector of the online time window, assemble the observation parent vector, generate fact prediction, calculate the anomaly score, retrieve the anchor reference window to perform anchor replacement, generate counterfactual prediction, calculate the perturbation sensitivity, degradation persistence and discrimination ratio, and generate degradation label; Based on the degradation marker, degradation marker gating is performed, health indicators are calculated through degradation persistence, a recent window set is constructed, and maintenance information is generated by combining the maintenance threshold. The process of generating the anchor set and anchor vector library includes: assembling the parent vector based on the parent set and the time delay table, writing it into the nearest neighbor library and fixing the nearest neighbor distance calculation method, combining the causal chain to calculate the anchor score, and generating the anchor set and anchor vector library. The calibration and maintenance threshold includes: calculating anomaly scores based on the nearest neighbor database and normalized resolution, combined with the parent set and time delay table, calibrating anomaly trigger thresholds, and generating a baseline window set; The causal chain is decyclicized and topologically sorted to generate a recursive sequence. Combined with the benchmark window set, anchor benchmark window retrieval and counterfactual recursion are performed to generate counterfactual predictions. Degradation persistence is calculated and health indicators are converted to calibrate maintenance thresholds. The calculation of anomaly score includes: receiving an online time window, extracting normalized sequence fragments, generating representative values ​​for channel windows, and assembling anchor vectors for the online time window according to the channel order of the anchor set; Based on the parent set and the time delay table, the observed parent vector is assembled, and the nearest neighbor entries are retrieved from the nearest neighbor database to generate fact predictions and calculate anomaly scores. The generation of degradation markers includes: performing a judgment based on the anomaly score and the anomaly trigger threshold; when the anomaly score is less than the anomaly trigger threshold, writing the counterfactual prediction as the factual prediction; and writing the perturbation sensitivity, degradation persistence, discrimination ratio, and degradation marker. When the abnormal score is greater than or equal to the abnormal trigger threshold, the baseline window set and anchor vector library are read, and the anchor vector nearest neighbor distance is calculated by combining the anchor vector of the online time window to determine the anchor baseline window index. Based on the anchor reference window index, anchor set, and recursive sequence, anchor replacement is performed, counterfactual parent vectors are assembled according to the recursive sequence, counterfactual predictions are generated by retrieving the nearest neighbor library, perturbation sensitivity, degradation persistence, and discrimination ratio are calculated, and degradation markers are generated.

2. The automated equipment anomaly monitoring method based on artificial intelligence as described in claim 1, characterized in that, The process of obtaining the original value sequence includes: reading multi-source measurement data frames, establishing measurement channel registration items, writing them into the channel table, determining the resolution quantity, performing timestamp cleaning, and generating a cleaned timestamp sequence and a cleaned original value sequence. Based on the cleaned timestamp sequence, calculate the common coverage starting point, solidify the unified sampling interval, and generate a unified time scale sequence; Based on the cleaned timestamp sequence and the cleaned original value sequence, combined with a unified time scale sequence, the previous values ​​are sampled in a uniform manner to generate the original value sequence.

3. The automated equipment anomaly monitoring method based on artificial intelligence as described in claim 2, characterized in that, The process of writing to the history window library includes: constructing a continuous window sequence based on a unified time scale sequence, extracting original time window segments from the original value sequence, combining the resolution quantity, performing robust normalization, constructing history window library entries, and writing them into the history window library.

4. The automated equipment anomaly monitoring method based on artificial intelligence as described in claim 1, characterized in that, The solidification time delay table includes: generating a channel window sequence and a set of positive differences in the window sequence based on the historical window library, and determining the normalized resolution of each channel; The calculation of the number of dependent nearest neighbors is performed, the conditional dependency estimation rules are solidified, candidate parent channels are evaluated, parent sets and causal chains are generated, and the time delay table is solidified.

5. The automated equipment anomaly monitoring method based on artificial intelligence as described in claim 1, characterized in that, The calculation of health indicators includes: reading degradation markers and performing degradation marker gating; marking time windows where degradation is marked as operating condition disturbances as not updating health indicators; calculating health indicators based on degradation persistence for time windows where degradation is marked as true degradation; writing health indicators into the historical window library; and generating online time windows with health indicators.

6. The automated equipment anomaly monitoring method based on artificial intelligence as described in claim 5, characterized in that, The generation and maintenance information includes: based on the window index of the online time window with health indicators, tracing back the historical window library entries, constructing a recent window set, and calculating the robust change rate of health indicators; Based on the robust change rate of health indicators and combined with maintenance thresholds, the remaining time prediction is calculated, and maintenance information is generated.

Images