An intelligent hotel equipment intelligent maintenance method and system based on AI big data predictive maintenance

Abstract

This invention discloses a smart maintenance method and system for smart hotel equipment based on AI big data predictive maintenance, relating to the fields of fault diagnosis and predictive maintenance. The method includes: collecting control loop setpoints, feedback values, control outputs, execution end observations, and load proxies, and aligning them to generate a deviation time series; segmenting operating segments according to setpoints, outputs, and load changes; filtering healthy segments to form a benchmark set based on deviation stability and output activity; establishing a load-deviation interpretable interval based on the benchmark set to generate a disturbance co-image coefficient, and establishing a response time window to generate time consistency; generating an attribution, confidence, location, and handling list by combining deviation over-limit markers with the disturbance co-image coefficient and time consistency, forming work order elements. Through deviation over-limit triggering and segmented benchmark modeling, the load interpretable interval and response time window are integrated to generate attribution results and form maintenance work order elements.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis and predictive maintenance, specifically to a smart hotel equipment intelligent maintenance method and system based on AI big data predictive maintenance. Background Technology

[0002] Hotels have a wide variety of equipment types, a large number of locations, and frequent changes in operating conditions. Control loop data typically includes setpoints, feedback values, control outputs, actuator observations, and proxy quantities reflecting load changes such as occupancy rates, outdoor weather conditions, and fresh air ratios. In current engineering practices, data collection and alarm linkage are mostly accomplished through building automation systems and operation and maintenance platforms. Threshold alarms, trend statistics, anomaly detection, or predictive models are used to conduct maintenance reminders and work order flow.

[0003] Existing technologies remain susceptible to multi-source disturbances and control semantics during practical application. Deviation changes often occur simultaneously with setpoint adjustments, control output actions, and load fluctuations. Alarm mechanisms based on deviation amplitude or single statistical characteristics struggle to distinguish between normal transients caused by changes in operating conditions and abnormal drifts caused by faults, leading to false alarms and alarm generalization. Fixed-window or simple sliding-window data organization methods fail to establish consistent segment boundaries around setpoints, control outputs, and load changes, resulting in multiple operating conditions mixed within the window. This makes it difficult to establish a stable health baseline, and thresholds rely on manual adjustment. Furthermore, there is a lack of auditable evidence chains between attribution outputs and response actions, often only providing anomaly scores or alarm types, making it difficult to form clear attribution categories, confidence levels, and location points. Insufficient characterization of response time after changes in control outputs can easily lead to issues such as actuator hysteresis, jamming, or measurement link drift being confused with general deviation exceeding limits, reducing maintenance and troubleshooting efficiency. Maintenance records and work order elements are mostly filled in with templated fields, lacking structured connections with segment data and attribution evidence, resulting in limited traceability and reuse capabilities.

[0004] Therefore, there is an urgent need for an intelligent maintenance solution for smart hotel equipment that establishes a clear data transmission relationship between control loop deviation exceeding the limit trigger, segment boundary formation, health benchmark establishment, load-related interpretable interval construction, response time window characterization after control action, and timeliness consistency verification. Without relying on complex model stacking, it can generate a verifiable list of attribution categories, confidence levels, location points, and disposal items, and further structure and generate maintenance records and work order elements to improve the reliability of anomaly identification and the consistency of maintenance decisions. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a smart maintenance method and system for smart hotel equipment based on AI big data predictive maintenance, so as to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a smart hotel equipment intelligent maintenance method based on AI big data predictive maintenance, comprising: Based on the setpoint, feedback value, control output, execution end observation, and load proxy quantity collected by the control loop, a deviation timing sequence is generated by aligning it with the time base. The segment boundaries are determined based on changes in set values, control outputs, and load agents, and the deviation time sequence is segmented to generate operating segments. A baseline set is generated by selecting healthy segments based on the deviation stability and output activity of the running segments; Based on the benchmark set, an interpretable interval for the load surcharge and deviation time series is established, and disturbance co-image coefficients are generated. Establish a response time window after the control output changes based on the benchmark set, and generate timeliness consistency. Based on the deviation time sequence, a deviation exceeding the limit marker is generated. Combined with the disturbance co-image coefficient and timeliness consistency, an attribution category, confidence level, location point and disposal item list are generated, and maintenance records and work order elements are generated.

[0007] The present invention is further configured such that the step of generating the deviation timing sequence by aligning according to the time base includes: The original timing sequence is generated by associating the sampling time identifier with the setpoint, feedback value, control output, execution end observation, and load proxy quantity collected by the control loop. A unified time base is established based on the original timing sequence. The setpoint, feedback value, control output, execution end observation measurement, and load proxy quantity are time aligned to obtain the aligned timing sequence. The deviation timing is formed based on the set value and feedback value in the alignment timing.

[0008] The present invention is further configured such that the step of generating runtime segments by segmenting the deviation time sequence includes: Based on the set values, control outputs, and load proxy quantities in the alignment timing, change markers are determined, and boundary candidate markers are generated. Based on the candidate boundary markers, persistent constraints are applied to generate a set of candidate boundary points; Based on the candidate boundary point set, apply minimum interval constraints and minimum segment length constraints to generate a segment boundary set; Based on the segment boundary set, the deviation time sequence is segmented to generate the running segment corresponding to the segment boundary set.

[0009] The present invention is further configured such that the step of generating a benchmark set by filtering healthy segments based on the deviation stability and output activity of the running segments includes: Statistically analyze the segment length and missing test ratio of the running segments, generate valid segment markers, and retain valid segments based on the valid segment markers; Within the effective segment, the upper quantile deviation value and the lower quantile deviation value are extracted based on the deviation time sequence to generate the segment deviation fluctuation range. The segment deviation fluctuation range is then scaled to generate the deviation stability. Within the effective segment, the cumulative change is calculated based on the control output in the alignment sequence, and the cumulative change is scaled to generate the output activity. The deviation stability is compared with a preset stability threshold, and the output activity is compared with a preset activity threshold to generate a health label. Based on the health label, healthy segments are selected from the valid segments to generate a benchmark set.

[0010] The present invention is further configured such that establishing the interpretable interval of the load proxy quantity and deviation time series based on the reference set includes: Extract corresponding samples of the benchmark centralized load proxy quantity and deviation to generate a benchmark sample pool; Calculate the load coverage range within the baseline sample pool, and generate a segmented support point set according to the load coverage range; The upper and lower boundaries of the deviation are generated based on the segmented support point set. The upper and lower boundaries of the deviation are then processed to achieve continuity, resulting in the upper and lower intervals of the deviation that vary with the load. Perform a consistency check on the upper and lower intervals of the deviation to generate an interval margin. Correct the upper and lower intervals of the deviation based on the interval margin to generate an interpretable interval.

[0011] The present invention is further configured such that the generation of perturbation co-image coefficients includes: Extract the correspondence between load surcharge changes and deviation changes from the benchmark set, calculate the load impact time delay, and generate the time delay corrected load surcharge. Within the running segment, query the interpretable interval based on the time-delay correction load agent quantity, and generate a hit flag based on whether the deviation value is contained within the interpretable interval. Within the running segment, direction-consistent markers are generated based on time-delay correction of load agent quantity changes and deviation changes, and direction-consistent markers are retained at sampling locations where load agent quantity changes meet the significance condition; The perturbation co-image coefficient of the running segment is generated by summarizing the hit marker and the direction-consistent marker.

[0012] The present invention is further configured such that establishing the response time window after the change in control output based on the reference set includes: The system identifies control action moments when the control output reaches a preset change threshold in the benchmark set and generates a benchmark action moment set. Within the time range corresponding to the set of baseline action times, the effective response arrival time is calculated based on the threshold of the change in the execution end observation and feedback value, and execution end delay samples and feedback end delay samples are generated. An execution-end response time window is established based on the execution-end latency sample, and a feedback-end response time window is established based on the feedback-end latency sample.

[0013] The present invention further specifies that the generation timeliness consistency includes: Within a running segment, identify the control action moments when the control output reaches a preset change threshold, and generate a set of segment action moments; Within the time range corresponding to the set of fragment action moments, calculate the arrival time of the effective response at the execution end and the arrival time of the effective response at the feedback end, and generate the fragment execution end delay and the fragment feedback end delay. The execution end delay of the fragment is compared with the execution end response time window to form the execution end time mark, and the feedback end delay of the fragment is compared with the feedback end response time window to form the feedback end time mark; Event-level timeliness markers are formed based on the timeliness markers at the execution end and the timeliness markers at the feedback end. Event-level timeliness markers are then aggregated into timeliness consistency by segment.

[0014] The present invention is further configured such that, in generating deviation exceedance markers based on deviation time sequence, and combining disturbance co-location coefficients and timeliness consistency to generate attribution categories, confidence levels, location points, and a list of disposal items, the maintenance record and work order elements include: Within the running segment, a deviation exceeding the limit marker is generated based on the deviation timing and a preset deviation threshold. Based on the deviation exceeding the limit marker, segment exceeding the limit coverage information and segment exceeding the limit duration information are generated, and an attribution effective marker is generated. When the attribution validity marker meets the conditions, a mutually exclusive attribution evidence set is generated based on the perturbation co-image coefficient and the timeliness consistency. Attribution categories are generated based on the mutually exclusive attribution evidence set, forming the attribution evidence difference quantity. The confidence level is determined by combining the difference in attribution evidence with information on fragment over-coverage. Based on the attribution category and confidence level, location points are generated in the closed-loop location correlation relationship; based on the attribution category and confidence level, a list of disposal items is generated in the disposal rule set. The maintenance record is generated by associating the operation segment identifier, deviation exceeding the limit mark, disturbance co-image coefficient, timeliness consistency, attribution category, confidence level, location point, and disposal item list, and the work order elements are extracted from the maintenance record.

[0015] This invention also provides a smart hotel equipment intelligent maintenance system based on AI big data predictive maintenance, the system comprising: Loop alignment module: Based on the control loop's collected setpoints, feedback values, control outputs, execution end measurements, and load surcharges, it aligns the time data to generate a deviation timing sequence. Boundary segmentation module: Determines segment boundaries based on changes in setpoints, control outputs, and load agents, and generates running segments by segmenting the deviation time sequence; Health screening module: Generates a baseline set by screening healthy segments based on the deviation stability and output activity of the running segments; Interval modeling module: Establishes interpretable intervals for load surcharge and deviation time series based on the baseline set, and generates disturbance co-image coefficients; Time window modeling module: Establishes response time windows after changes in control output based on the baseline set, and generates timeliness consistency. Attribution Work Order Module: Generates deviation over-limit markers based on deviation time sequence, and generates attribution categories, confidence levels, location points, and a list of disposal items by combining disturbance co-image coefficient and timeliness consistency, and generates maintenance records and work order elements.

[0016] This invention provides a smart hotel equipment intelligent maintenance method and system based on AI big data predictive maintenance. The method generates a deviation time series by collecting setpoints, feedback values, control outputs, execution end observations, and load surcharges from the control loop and aligning them with a time reference; it determines segment boundaries based on changes in setpoints, control outputs, and load surcharges, and segments the deviation time series to generate operational segments; it filters healthy segments based on the deviation stability and output activity of the operational segments to generate a benchmark set; it establishes an interpretable interval between the load surcharge and the deviation time series based on the benchmark set, generating a disturbance co-occurrence coefficient; it establishes a response time window after changes in control output based on the benchmark set, generating a timeliness consistency; it generates a deviation over-limit marker based on the deviation time series, and combines the disturbance co-occurrence coefficient and timeliness consistency to generate an attribution category, confidence level, location point, and disposal item list, generating maintenance records and work order elements. The beneficial effects include: 1. Evidence Chain Attribution Output: Taking the deviation exceeding the limit mark as the trigger entry, the disturbance co-image coefficient and the timeliness consistency are linked to form mutually exclusive attribution evidence, output the attribution category and confidence level, and link with the location point and the list of disposal items to generate maintenance records and work order elements, so as to realize the continuous data transmission from abnormal facts to disposal elements. 2. Fragmented modeling driven by operating conditions: Fragment boundaries are formed based on changes in setpoints, control outputs, and load agents. Deviation time series are consistently segmented, and healthy fragments are selected within the fragments to form a benchmark set. Benchmark learning and online discrimination are supported to be performed under the same fragment caliber, ensuring consistency between the analysis object and the data source. 3. Interpretable Interval and Timeliness Consistency: Based on the benchmark set, an interpretable interval for load proxy quantity and deviation time series is established and a disturbance co-image coefficient is generated. At the same time, a response time window after the control output changes is established and a timeliness consistency degree is generated. The load interpretability and response timeliness characteristics are incorporated into the same discrimination framework, providing a structured basis for locating points and handling item lists.

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart illustrating an exemplary embodiment of the present invention is provided for a smart hotel equipment intelligent maintenance method based on AI big data predictive maintenance. Figure 2 This is a schematic diagram illustrating the structure of an intelligent maintenance system for smart hotel equipment based on AI big data predictive maintenance, as an exemplary embodiment of the present invention. Detailed Implementation

[0019] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0020] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0021] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0022] Example 1: A smart hotel equipment intelligent maintenance method based on AI big data predictive maintenance, such as... Figure 1 As shown, it includes: Based on the setpoint, feedback value, control output, execution end observation, and load proxy quantity collected by the control loop, a deviation timing sequence is generated by aligning it with the time base. The segment boundaries are determined based on changes in set values, control outputs, and load agents, and the deviation time sequence is segmented to generate operating segments. A baseline set is generated by selecting healthy segments based on the deviation stability and output activity of the running segments; Based on the benchmark set, an interpretable interval for the load surcharge and deviation time series is established, and disturbance co-image coefficients are generated. Establish a response time window after the control output changes based on the benchmark set, and generate timeliness consistency. Based on the deviation time sequence, a deviation exceeding the limit marker is generated. Combined with the disturbance co-image coefficient and timeliness consistency, an attribution category, confidence level, location point and disposal item list are generated, and maintenance records and work order elements are generated.

[0023] The present invention is further configured such that the step of generating the deviation timing sequence by aligning according to the time base includes: Based on the control loop's collected setpoints, feedback values, control outputs, execution end measurements, and load surcharges, a raw time sequence is generated by associating sampling time identifiers. Specifically, within the same control loop, five types of point data are collected: setpoints, feedback values, control outputs, execution end measurements, and load surcharges. Each sampling record is written with a sampling time identifier during acquisition. The sampling time identifier is generated using the same acquisition-side clock, with an accuracy of no less than the second level, and retains both the point identifier and the loop identifier to prevent cross-loop mixing. A validity screening is performed synchronously on each type of point data: range verification is performed based on the engineering range configuration, and records exceeding the range are marked as invalid; interval verification is performed based on the point refresh cycle configuration, and the time interval between two adjacent records exceeding the refresh cycle limit is marked as a sampling interruption. After screening, raw time sequences with time identifiers are generated according to point type. The raw time sequence includes the sampled value, sampling time identifier, validity marker, and interruption marker. A unified time base is established based on the original timing sequence. The setpoint, feedback value, control output, execution end observation, and load proxy quantity are time-aligned to obtain the aligned timing sequence. Specifically, the control output is selected as the reference timing for the unified time base in the original timing sequence because the control output is most closely related to the controller's action time and is suitable as the time anchor point for closed-loop data alignment. First, the typical values ​​of adjacent sampling time intervals of the control output are statistically analyzed, and the median is used as the unified sampling period. Then, the unified time range is determined by the start and end times of the effective data coverage. Within this range, a continuous unified time scale sequence is generated according to the unified sampling period. Subsequently, five types of points are aligned on a unified time scale: For setpoints, control outputs, and load proxy quantities, a preservative alignment method is used, where the most recent valid sampled value before each unified time scale is taken as the alignment value for that scale; for feedback values ​​and execution-end observations, a nearest-neighbor interpolation alignment method is used, where the two most recent valid sampled records before and after each unified time scale are found, and the alignment value for that scale is estimated according to the time distance ratio to reduce step errors caused by refresh frequency differences; when the interval between valid samples before and after a certain unified time scale exceeds the preset interpolation upper limit, that point is determined to be uninterpolable at that scale, a missing marker is generated, and no interpolation result is output; for other points, the alignment value is still output according to the above rules. Through the above process, five types of aligned time sequences are formed, all indexed by the unified time scale, and the missing marker for each scale is retained. The deviation timing sequence is formed based on the setpoint and feedback values ​​in the alignment timing sequence. Specifically, at each unified time scale, the setpoint alignment value and feedback value alignment value are read, and the setpoint alignment value minus the feedback value alignment value is taken as the deviation value for that scale. If either the setpoint or feedback value for that scale has a missing flag, the deviation value for that scale is simultaneously marked as missing and is not included in the statistical caliber of subsequent segment segmentation and benchmark learning. The deviation timing sequence records are mapped one-to-one with the unified time scale, and each record simultaneously carries the control output alignment value, execution end observation alignment value, load proxy quantity alignment value, and the missing flag and sampling interruption flag for that scale. After completion, the deviation timing sequence and alignment timing sequence are indexed by loop identifier and time range and written into the timing series dataset.

[0024] The present invention is further configured such that the step of generating runtime segments by segmenting the deviation time sequence includes: Based on the setpoints, control outputs, and load surcharges in the alignment time series, change markers are determined, and boundary candidate markers are generated. Specifically, under a unified time base, setpoint alignment time series, control output alignment time series, load surcharge alignment time series, and deviation time series have been obtained. First, at each time scale, it is determined whether the setpoints, control outputs, and load surcharges have undergone "valid changes." The determination of valid changes is achieved by "comparing the difference between adjacent scales with a fluctuation scale threshold": for each type of alignment time series, the typical fluctuation level of the difference between adjacent scales is first statistically analyzed within a continuous time window. The typical fluctuation level uses the median absolute deviation as a scale, which can suppress the influence of a small number of spikes and outliers; then, the absolute value of the difference between the current scale and the previous scale is taken and compared with this scale multiplied by a multiplier factor. When the difference exceeds the comparison result, the scale is marked as having undergone a valid change in the variable. The multiplier factor is an engineering configurable parameter, with a value between three and ten. A larger multiplier indicates a more conservative approach. When any one of the setpoint, control output, or load agent quantity undergoes a valid change at a certain scale, a boundary candidate mark is generated for that scale; when none of the three undergoes a valid change, no boundary candidate mark is generated for that scale. To avoid boundary candidate marks being frequently triggered in a short period of time, a "minimum change duration threshold" can be superimposed when generating candidate marks. That is, a candidate mark is only retained when the direction of the difference remains consistent within a few adjacent scales or the difference continuously exceeds the threshold, in order to reduce false triggering caused by single-point jitter. Based on the boundary candidate markers, persistent constraints are applied to generate a candidate boundary point set. Specifically, the boundary candidate markers may include isolated single-point triggers or instantaneous triggers caused by noise; therefore, persistent constraints are applied to the boundary candidate markers. Persistent constraints are implemented using a consecutive hit threshold: boundary candidate markers are sequentially scanned on the time axis, and the length of consecutive occurrences of candidate markers is counted. When the consecutive length reaches a preset threshold, the starting scale of that consecutive segment is registered as a candidate boundary point. The consecutive hit threshold is a configurable parameter, typically two to five scales. A larger threshold makes it less sensitive to noise but may delay the boundary position. A fixed rule is used to select candidate boundary points: the first scale of the consecutive segment is taken as the candidate boundary point so that subsequent segmentation times are as close as possible to the starting point of the change. After scanning is completed, a candidate boundary point set is formed in chronological order, and the trigger source type corresponding to each candidate boundary point is retained for tracing the boundary originating from setpoint changes, control output actions, or load mutations when necessary. A fragment boundary set is generated by applying minimum interval constraints and minimum fragment length constraints to the candidate boundary point set. Specifically, multiple similar boundaries may still appear in the candidate boundary point set within a short period of time. To make the fragments analyzable, minimum interval constraints and minimum fragment length constraints need to be applied. The minimum interval constraint is implemented using time scale difference: boundary points are selected sequentially from the candidate boundary point set in chronological order. If the time interval between the current candidate boundary point and the most recently retained boundary point is less than the preset minimum interval, the current candidate boundary point is discarded. The minimum interval is a configurable parameter used to remove repeated triggers in the same change process, and its value typically covers one to several control refresh cycles. The minimum fragment length constraint is implemented using the number of scales within the fragment: on the boundary sequence obtained after applying the minimum interval, the number of scales between two adjacent boundaries is checked. If it is less than the preset minimum fragment length, a merging process is performed. The merging process uses fixed rules: the next boundary point that makes the fragment length insufficient is deleted first, and the short fragment is merged into the previous fragment; when the short fragment is at the beginning or end of the sequence, the short fragment is merged into the adjacent available fragment. The minimum fragment length is a configurable parameter for engineering purposes, typically chosen as the lowest sample size sufficient to support subsequent stability and activity assessments. After completing both types of constraints, the final fragment boundary set is obtained, which is a set of boundary scales arranged in ascending order of time. Based on the segment boundary set, the deviation time series is segmented, generating running segments corresponding to the segment boundary set. Specifically, the deviation time series is segmented based on the segment boundary set. During segmentation, the segment boundary set is first expanded into a boundary sequence containing a start and end scale: the start scale of a unified time range is used as the first boundary, and the next scale after the end of the unified time range is used as the last boundary. Then, the scale interval between adjacent boundaries is used as a time index range for a running segment. For each running segment, the corresponding deviation sequence is extracted from the deviation time series according to the segment's time index range to form segment deviation data; at the same time, the setpoint alignment sequence, control output alignment sequence, and load proxy quantity alignment sequence within the same time index range are retained as segment metadata, so that subsequent benchmark set selection, interpretable interval establishment, and response time window construction can directly reference a consistent data source within the segment. If a segment contains a missing marker, the missing ratio is recorded at the segment level for use in subsequent effective segment selection to remove missing abnormal segments. Thus, a set of running segments corresponding one-to-one with the segment boundary set is obtained, and the running segment set contains the segment start and end scales, segment deviation sequence, and segment metadata index.

[0025] The present invention is further configured such that the step of generating a benchmark set by filtering healthy segments based on the deviation stability and output activity of the running segments includes: The process involves statistically analyzing the segment length and missing rate of running segments, generating valid segment markers, and retaining valid segments based on these markers. Specifically, for each running segment in the acquired set of segments, the segment length is first calculated, represented by the number of uniform time scales contained within that segment. Then, the missing rate is calculated, with the missing rate criteria aligned with the preceding steps: if a missing marker exists for a deviation value at a given time scale, or if a missing marker exists for the control output at that scale, that scale is included in the missing rate count. The missing rate count is divided by the segment length to obtain the missing rate. Valid segment markers are generated using a "double threshold" method: a segment is marked as valid if its length is not less than a preset minimum segment length threshold and its missing rate does not exceed a preset upper limit threshold; otherwise, it is marked as invalid. The minimum segment length threshold is set to ensure stable subsequent quantile extraction, typically ensuring at least thirty usable sampling points within a segment; the upper limit threshold for the missing rate is set to guarantee the integrity of the segment information, typically not exceeding 20%. After marking, only valid segments are retained for subsequent steps, and the segment length, missing rate, and valid markers are recorded in the segment metadata. Within the valid segments, the upper and lower quantile deviation values ​​are extracted based on the deviation time series to generate the segment deviation fluctuation range. Scale normalization is then performed on the segment deviation fluctuation range to generate deviation stability. Specifically, for each valid segment, all valid deviation values ​​within the segment are first extracted and sorted in ascending order. Then, the upper and lower quantile deviation values ​​are extracted. The upper quantile is the deviation value located at the 90th percentile after sorting, and the lower quantile is the deviation value located at the 10th percentile after sorting. When the number of valid samples is insufficient to precisely fall at this position, the nearest position rule is adopted to ensure reproducibility. The difference between the upper and lower quantile deviation values ​​is used as the deviation fluctuation range of the segment, characterizing the main fluctuation bandwidth of the deviation within the segment and avoiding a single peak value dominating the fluctuation evaluation. Scale normalization is performed on the deviation fluctuation range. The reference scale for normalization is the median of the deviation fluctuation ranges of all valid segments. Using the median can suppress the scale pull of a small number of abnormal segments. If the median is close to zero, the minimum positive value corresponding to the deviation acquisition resolution is used as the lower limit compensation to avoid zero scale during the normalization process. Deviation stability is generated based on the principle that "the smaller the fluctuation range is relative to the reference scale, the higher the stability": when the fluctuation range of the segment deviation is significantly smaller than the reference scale, the stability is close to one; when the fluctuation range of the segment deviation is close to or greater than the reference scale, the stability decreases accordingly. Within a valid segment, the cumulative change is calculated based on the control output in the aligned time sequence. The cumulative change is then scale-normalized to generate output activity. Specifically, for each valid segment, the aligned value sequence of the control output within that segment is extracted, and after removing missing scales, it is arranged in chronological order. Then, the absolute amount of control output change is calculated between two adjacent time scales, and these absolute changes are accumulated within the segment to obtain the cumulative change for that segment. The cumulative change is used to characterize the intensity of the control output action within the segment, reflecting states such as frequent adjustments, oscillations, or continuous actions, without relying on a single amplitude or mean. Scale normalization is also used to form the output activity. The normalized reference scale is the median of the cumulative changes across all valid segments, with the minimum positive value corresponding to the control output acquisition resolution used as a lower limit compensation to ensure the reference scale is always positive. The output activity is generated based on the principle that "the larger the cumulative change relative to the reference scale, the higher the activity": when the cumulative change of a segment is much smaller than the reference scale, the activity approaches zero; when the cumulative change of a segment approaches or exceeds the reference scale, the activity increases. The deviation stability is compared with a preset stability threshold, and the output activity is compared with a preset activity threshold to generate a health label. Based on the health label, healthy segments are selected from the valid segments to generate a benchmark set. Specifically, for each valid segment, the deviation stability is compared with a preset stability threshold, and the output activity is compared with a preset activity threshold. A health label is generated when both conditions are met. The stability and activity thresholds are selected using a verifiable method: the distribution of deviation stability in the effective fragment set is statistically analyzed, and the higher quantile is used as the stability threshold to ensure that fragment deviation fluctuations are controlled; the distribution of output activity is statistically analyzed, and the lower quantile is used as the activity threshold to ensure that fragment control actions are relatively smooth; the stability threshold usually falls within the 70th to 90th percentile range, and the activity threshold usually falls within the 10th to 30th percentile range; alternatively, it can be fixed by the equipment type parameter table to ensure that different equipment maintains a consistent screening caliber; healthy fragments are selected from the effective fragments based on health markers, and the set of healthy fragments constitutes the benchmark set; in addition to storing the fragment start and end time scales and fragment indexes, the benchmark set also retains the alignment sequence indexes of intra-fragment deviation, control output, execution end observations, and load proxy quantities, and archives fragment length, missing measurement ratio, deviation stability, and output activity.

[0026] The present invention is further configured such that establishing the interpretable interval of the load proxy quantity and deviation time series based on the reference set includes: A baseline sample pool is generated by extracting corresponding samples of load surcharge and deviation from the baseline set. Specifically, in the baseline set, each healthy segment has its load surcharge alignment value and deviation value stored under a unified time base. All healthy segments in the baseline set are traversed one by one, and the load surcharge alignment value and the deviation value at each unified time scale are read simultaneously. These are written into the sample pool as a pair of samples corresponding to the same time scale. To avoid contamination of the interval by missing measurements and outliers, a consistent validity screening is performed before writing: if the deviation value or load surcharge has a missing marker, it is not written; if the load surcharge exceeds its engineering range or the deviation exceeds the preset reasonable limit, the pair of samples is marked as abnormal and removed. The reasonable limit is obtained by quantile statistics of historical healthy data, using higher and lower quantiles as boundaries and reserving a safety margin to ensure that obvious anomalies are excluded without compressing normal fluctuations. After traversal, the baseline sample pool is obtained, containing several corresponding samples of "load value - deviation value". Calculate the load coverage area within the baseline sample pool, and generate segmented support point sets based on the load coverage area. Specifically, extract all load proxy quantity samples from the baseline sample pool, and first perform coverage area statistics: take the minimum and maximum values ​​of the load samples as the two ends of the load coverage area. To avoid the influence of extremely sparse endpoints, the coverage area can be determined using a quantile pruning method, i.e., take the lower quantile as the lower end and the higher quantile as the upper end, and configure the pruning ratio to be between 1% and 5%. Generate segmented support point sets within the load coverage area; the segmentation method adopts a "quantile equalization" approach. "Frequency segmentation" involves sorting the baseline sample pool by load value from smallest to largest, then dividing it into several segments with an equal number of samples in each segment. The number of segments is driven by the sample size; when the sample size is small, fewer segments are used to ensure that the number of samples in each segment is not less than the preset minimum number of samples, while when the sample size is sufficient, the number of segments can be increased to improve the interval resolution. The support point of each segment is taken as the representative value of the load value of that segment, and the representative value is the median of the load value of that segment to suppress endpoint anomalies. This forms a set of segmented support points, and each support point corresponds to a load interval and its interval sample set. The upper and lower boundaries of the deviation are generated based on the segmented support point set. The upper and lower boundaries of the deviation are then processed to obtain the upper and lower intervals of the deviation that vary with the load. Specifically, for each load segment interval, all deviation values ​​are extracted from the sample set corresponding to the interval and sorted by size. The deviation value with the lower quantile is taken as the lower boundary of the load interval, and the deviation value with the higher quantile is taken as the upper boundary of the load interval. The values ​​of lower and higher quantiles are used to characterize the main fluctuation bands of the deviation under healthy conditions. The quantile ratio is configured from 5% to 15%, with a larger ratio indicating a more convergent interval and a smaller ratio indicating a more relaxed interval. After obtaining the upper and lower boundaries of each segment interval, continuous processing is performed on the upper and lower boundaries to make the intervals exhibit a smooth and interpretable trend with load changes. The continuous processing adopts the "segmented monotonic constraint smoothing" method: first, the changes in the upper and lower boundaries of adjacent segments are checked for jumps. If the difference between the boundaries of adjacent segments exceeds the preset jump threshold, the jump amplitude is reduced by the weighted average of adjacent segments. Then, moving median smoothing is performed on the boundary sequence, with the smoothing window covering three to five adjacent segments to suppress the jagged boundaries caused by insufficient local segment samples. If the business rules require the deviation to change monotonically with the load, a monotonic consistency constraint is further applied to the boundary sequence. Segments that violate monotonicity are corrected by backtracking to the nearest feasible value to ensure that the boundary shape is interpretable. After completion, the lower and upper deviation intervals that change with the load are obtained, both of which are composed of segment support points and corresponding boundaries. Consistency checks are performed on the upper and lower intervals of the deviation to generate interval margins. These margins are then used to correct the deviation intervals, generating interpretable intervals. Specifically, consistency checks are performed on the continuous deviation intervals. The goal of the checks is to ensure that the intervals reach a preset coverage level within the baseline sample pool while avoiding local over-narrowing. During the checks, the baseline sample pool is replayed line by line: for each sample, it is determined whether its deviation value falls within the upper and lower intervals corresponding to its load value; the overall hit rate is calculated as the coverage rate, and segmented coverage rates are calculated by load segment to identify local weak areas; the preset coverage level is set based on the principle of tolerating healthy fluctuations, typically between 90% and 98%; when the overall coverage rate or any segmented coverage rate is lower than the preset coverage level, interval margins are generated to correct the intervals; the interval margins are generated using "quantile margins of out-of-bounds quantities": for all missed samples, the out-of-bounds quantities where the deviation exceeds the upper boundary and the out-of-bounds quantities where the deviation is below the lower boundary are calculated separately. Boundary quantity; take the higher quantile of the two types of out-of-boundary quantities as the unified margin, and set a minimum margin lower limit to absorb the acquisition resolution and quantization error; then shift the upper boundary upward according to the margin and the lower boundary downward according to the margin to obtain the corrected deviation upper and lower intervals; after correction, perform consistency check again until the coverage level is met or the preset maximum number of iterations is reached; the maximum number of iterations is used to limit the infinite expansion of the interval, usually two to three times; finally, the interpretable interval is obtained and stored in the structure of "load segment support point - deviation lower boundary - deviation upper boundary", while recording the number of segments, quantile ratio, smoothing window, coverage level and margin value rules.

[0027] The present invention is further configured such that the generation of perturbation co-image coefficients includes: The correspondence between load surcharge changes and deviation changes is extracted from the benchmark set, the load impact time delay is calculated, and the time delay correction load surcharge is generated. Specifically, in the benchmark set, all benchmark segments are in a healthy state and the load surcharge and deviation have been saved under a unified time benchmark. First, the load change and deviation change are calculated on a scale-by-scale basis in the benchmark set. The load change is the difference between the load surcharge of the current scale and the load surcharge of the previous scale, and the deviation change is the difference between the deviation of the current scale and the deviation of the previous scale. To avoid missing measurements and quantification noise interference, changes are calculated only on scales where both the load change and deviation are valid values, and outliers that significantly exceed the upper limit of the engineering range are removed. A time-delay search is then performed to characterize the propagation delay of load changes on deviation changes. The time-delay search uses a discrete enumeration method: a maximum time-delay range is pre-defined, calculated using the sampling period and the upper limit of the system response time, typically covering several to dozens of sampling scales. Within this range, different lag steps are tried one by one, pairing the delayed load change with the current deviation change on the same time scale. For each lag step, the proportion of paired samples with consistent direction and the coverage of significant change amplitude are statistically analyzed: consistent direction means the load change and deviation change have the same sign, and significant change amplitude means the absolute value of the load change exceeds a significance threshold. The significance threshold is obtained from the high quantile of the absolute value of the baseline load change, typically between the 70th and 90th percentiles, to ensure that gating only applies to identifiable load changes. The directional consistency ratio and the degree of significant coverage are weighted according to preset criteria to form a score, with the highest-scoring lag step used as the load impact lag. After obtaining the load impact lag, time-lag correction is performed on the subsequent load proxy quantities, i.e., the load proxy quantities are shifted forward on the time axis by this lag to align with the deviation response; missing markers are written for any initial gaps in the scale resulting from the shift. Within each runtime segment, the interpretable range is queried based on the time-delay corrected load surcharge. A hit flag is generated based on whether the deviation value is contained within the interpretable range. Specifically, within each runtime segment, the time-delay corrected load surcharge value is read scale by scale, and the corresponding load segment is located within the interpretable range based on this load value. The load segment location adopts the same rules as the range construction: if the load value falls into a certain segment support point range, the lower and upper boundaries of the deviation of that segment are selected as the interpretable range of that scale; if the load value is between segment boundaries, a linear transition method between adjacent segment boundaries is used to obtain the interpretable range of that scale to avoid jumps caused by segmentation; after completing the interpretable range location, the deviation value of that scale is subjected to inclusion relationship judgment: a hit flag is generated when the deviation value falls into the interpretable range, and a hit flag is not generated when the deviation value exceeds the interpretable range; if there are missing flags for the load value or deviation value of that scale, no hit judgment is generated for that scale and an invalid flag is written; Within the runtime segment, direction consistency markers are generated based on changes in load surcharge and deviation due to time delay. These markers are retained at sampling locations where the load surcharge changes meet the significance criteria. Specifically, the time delay-corrected load change and deviation change are calculated scale-by-scale within the runtime segment, using the same calculation method as the baseline time delay search, employing the difference between adjacent scales. To ensure that direction consistency judgment is not misled by minor fluctuations, a significance gating is applied to each scale: if the absolute value of the time delay-corrected load change is less than a significance threshold, the scale is marked as not participating in direction consistency judgment; when its absolute value reaches or exceeds the significance threshold, it enters the direction consistency judgment process. Direction consistency judgment uses a sign consistency rule: if the load change is positive and the deviation change is positive, or the load change is negative and the deviation change is negative, a direction consistency marker is generated; otherwise, they are inconsistent. Scales with missing deviation or load change values ​​are also marked as not participating. Finally, a sequence of direction consistency markers valid only for scales with significant load changes within the segment is obtained and bound to the segment index for storage. The running segments are summarized based on the hit mark and the direction consistency mark to form the perturbation co-image coefficient of the running segments. Specifically, when summarizing each running segment, only the set of scales that simultaneously have valid hit discrimination and direction consistency discrimination are counted to ensure that the statistical caliber of the numerator and denominator is consistent. For each scale in the scale set, if the hit mark is "hit" and the direction is consistent, it is counted as a valid scale with the same shadow; otherwise, it is counted as an invalid scale with the same shadow. To make the same shadow coefficient more sensitive to scales with strong disturbances, a weighted summation rule based on the significance of load changes is introduced: when the load change just reaches the significance threshold, the weight is taken as a lower value, and the larger the load change, the higher the weight. The upper limit of the weight is taken as one to avoid extreme values ​​dominating. The weight is calculated by proportional truncation based on the significance threshold, that is, when the ratio of the absolute value of the load change to the significance threshold exceeds one, it is counted as one. Then, the same shadow validity of each scale in the segment is summed according to the weight, and the sum of the weights in the segment is normalized to obtain the disturbance same shadow coefficient of the segment. When there is no scale in the segment that meets the significance threshold, it means that the segment lacks load change events that can be used to determine the same shadow relationship. At this time, the disturbance same shadow coefficient is written into the preset default value and an insufficient evidence mark is added.

[0028] The present invention is further configured such that establishing the response time window after the change in control output based on the reference set includes: The control action moment when the control output reaches the preset change threshold is identified in the benchmark set, and a benchmark action moment set is generated. Specifically, the control output alignment sequence is traversed in time order in each healthy segment of the benchmark set, and the control output difference between the current scale and the previous scale is compared on a scale-by-scale. The absolute value of the control output difference is then compared with the preset change threshold. The preset change threshold is not a fixed constant, but is jointly given by the quantization resolution and noise level of the control output within the reference set: the minimum resolution of the control output is used as the lower limit, and the higher quantile of adjacent differences within the reference set is superimposed as the upper limit, so that the threshold can filter out quantization jitter and only respond to the real adjustment action; when the control output difference reaches or exceeds the preset change threshold, the scale is registered as a candidate moment of control action; considering that the control output may change in multiple consecutive scales, in order to avoid the same action process being recorded repeatedly, a minimum interval constraint is applied to the candidate moments of control action: the first candidate moment is retained in chronological order, and no repeated registration is made in the subsequent scale ranges. The minimum interval length is calculated by converting the smaller of the control refresh cycle and the typical response time; after traversal, a set of reference action moments is formed, and each action moment is associated with its action amplitude level. The action amplitude level is divided into several levels according to the ratio of the action difference to the threshold. Within the time range corresponding to the set of baseline action times, the effective response arrival time is calculated based on the threshold values ​​for the change amplitude of the execution-side observed values ​​and feedback values, generating execution-side delay samples and feedback-side delay samples. Specifically, for each action time in the set of baseline action times, a fixed observation time range is established, with the starting point being the action time itself and the ending point being the maximum observation length after the action time. The maximum observation length is determined by the upper limit of the response of the device control loop, and its value is converted into several scales based on the sampling period, typically covering several times the typical response time to ensure that hysteresis or slow response can be captured. Within this observation range, the effective response arrival determination is performed on the execution-side observed values ​​and feedback values ​​respectively. The effective response arrival determination uses a change amplitude threshold: first, the execution-side observed value and the feedback value benchmark value at the action time are calculated, and then the search proceeds backward step by step from the action time. When the deviation of the execution-side observed value relative to its benchmark value first reaches or exceeds the execution-side change amplitude threshold, this scale is recorded as the effective response arrival time of the execution end; when the deviation of the feedback value relative to its benchmark value first reaches or exceeds the feedback change amplitude threshold, this scale is recorded as the effective response arrival time of the feedback end. The threshold for the magnitude of change is not a fixed constant, but is adaptively generated by the natural fluctuation scale within the benchmark set: the typical fluctuation level of the adjacent difference between the execution end observation and the adjacent difference between the feedback value is statistically analyzed within the benchmark set. The typical fluctuation level uses the median absolute deviation to suppress a small number of outliers. Based on this, a multiplier factor is multiplied to form the threshold for the magnitude of change. The multiplier factor is configured to be three to eight to ensure that the threshold is higher than the noise jitter but not so high that it misses the true response. When no scale reaching the threshold is found within the maximum observation length, the corresponding delay sample of the action moment is marked as missing, and the reason for the missing is recorded as no valid arrival within the observation window. When the arrival moment is found, the scale difference between the arrival moment and the action moment is converted into a delay, forming the execution end delay sample and the feedback end delay sample respectively. An execution-end response time window is established based on the execution-end delay sample, and a feedback-end response time window is established based on the feedback-end delay sample. Specifically, statistical windows are built for all execution-end delay samples and feedback-end delay samples. Before window building, abnormal samples are removed: when the delay is zero and the action amplitude level is higher than the preset level, it is determined that there may be a synchronization jump or sampling alignment error, and the sample is marked as abnormal and removed. When the delay exceeds the fixed upper limit of the maximum observation length, it is determined as an extreme hysteresis sample, and is also marked and removed. To avoid insufficient samples due to overly strict removal rules, abnormal removal only applies to a small number of tail samples, and removal records are retained to support traceability. After removing abnormal samples, a quantile method is used to build response time windows for delay samples: the lower quantile delay value is taken as the lower bound of the time window, and the higher quantile delay value is taken as the upper bound of the time window. The lower and higher quantiles are configured in the range of 5% to 20% to reduce the widening effect of extreme samples on the window while ensuring coverage of mainstream response behaviors. If the sample size is small and the quantiles are unstable, a fixed rule of taking the nearest neighbor value according to the sort position is adopted to obtain the boundary to ensure reproducibility. Finally, the execution end response time window and the feedback end response time window are obtained and archived together with the baseline set version identifier, action threshold generation method, change amplitude threshold generation method, maximum observation length, and anomaly removal rules.

[0029] The present invention further specifies that the generation timeliness consistency includes: Within each running segment, the control action moments when the control output reaches a preset change threshold are identified, and a set of segment action moments is generated. Specifically, within each running segment, there are already control output alignment sequences, execution end observation alignment sequences, and feedback value alignment sequences under a unified time scale, and the execution end response time window and feedback end response time window have been obtained from the reference set. First, the control output alignment sequences are traversed in chronological order within the running segment, and the control output difference between the current scale and the previous scale is compared scale by scale. When the control output difference reaches or exceeds the preset change threshold, the scale is recorded as a candidate control action moment. The preset change threshold is consistent with the baseline windowing to avoid incomparability of action events due to different thresholds. This threshold is jointly given by the minimum resolution of the control output and the high quantile of adjacent differences in the control output within the baseline set, enabling the threshold to filter out quantization jitter and retain only the true action. To prevent the same action process from being repeatedly registered due to continuous refresh, a minimum interval constraint is applied to the candidate moments of the control action: the first candidate moment is retained in chronological order, and no new action moments are registered within a certain range after that candidate moment. The minimum interval length is consistent with the baseline windowing and usually covers one or more control refresh cycles. After traversal, a set of fragment action moments is obtained, and each action moment is associated with an action amplitude level, which is divided according to the ratio of the difference in that scale to the threshold. Within the time range corresponding to the set of fragment action moments, calculate the arrival time of the effective response from the execution end and the arrival time of the effective response from the feedback end, generating the fragment execution end delay and fragment feedback end delay. Specifically, for each action moment in the set of fragment action moments, establish a fixed observation range, with the starting point of the observation range being the action moment itself and the ending point being the maximum observation length after the action moment. The maximum observation length is taken from the upper limit configuration used in the baseline windowing to ensure consistent delay caliber. Within this observation range, search for the effective response arrival time of the execution end observations and feedback values ​​respectively. The search rule is consistent with the baseline windowing: based on the execution end observations at the action moment. The measured value serves as a reference point. Starting from the action moment, the deviation of the observed value at the execution end relative to the reference point is checked progressively backward on each scale. When the deviation first reaches or exceeds the execution end's change amplitude threshold, that scale is recorded as the effective arrival time of the execution end's response. Similarly, the feedback value at the action moment is used as the reference point. When the deviation first reaches or exceeds the feedback change amplitude threshold, the effective arrival time of the feedback end's response is recorded. Both types of change amplitude thresholds follow the configuration of adaptive noise scale generation from the reference set, i.e., multiplied by a multiplier factor from the robust scale of adjacent differences within the reference set. The multiplier factor is a preset parameter that ensures the threshold is higher than the noise jitter. If no arrival time is found within the maximum observation length, the corresponding end's delay is recorded as missing and no effective arrival is recorded within the observation window. When an arrival time exists, the scale difference between the arrival time and the action moment is converted into the segment execution end delay and the segment feedback end delay, and archived along with the action moment, action amplitude level, and segment identifier to form the input for subsequent timeliness marking. The execution end delay is compared with the execution end response time window to form an execution end timeliness mark, and the feedback end delay is compared with the feedback end response time window to form a feedback end timeliness mark. Specifically, for each action moment, the execution end delay and the feedback end delay corresponding to that action are read respectively, and the execution end response time window and the feedback end response time window output by the reference set are read. The generation of the execution end timeliness mark follows the following judgment: when the execution end delay of the fragment exists and falls within the execution end response time window range, the generated execution end timeliness mark is valid; when the execution end delay of the fragment exists but exceeds the time window range, the generated execution end timeliness mark is invalid; when the execution end delay of the fragment is missing, the generated execution end timeliness mark is invalid and an "delay missing" reason mark is attached. Similarly, the feedback end timeliness mark is generated based on the inclusion relationship between the segment feedback end delay and the feedback end response time window, and a reason mark is written for missing cases; to avoid frequent flipping caused by boundary jitter, a fixed tolerance band can be added to the time window boundary: when the difference between the delay and the time window boundary does not exceed the tolerance band, it is treated as falling into the time window; the tolerance band is taken as a scale corresponding to the control output sampling period or converted from the upper limit of the alignment error to ensure that the judgment is robust to the sampling quantization error; Event-level timeliness markers are formed based on the execution-end and feedback-end timeliness markers. These event-level timeliness markers are then aggregated by segment to form a timeliness consistency score. Specifically, at each action moment, the execution-end and feedback-end timeliness markers are combined to form an event-level timeliness marker. The event-level timeliness markers employ a rule of simultaneous validity at both ends: an action event is recorded as time-consistent only when both the execution-end and feedback-end timeliness markers are valid; otherwise, it is recorded as time-inconsistent, and the reason for the inconsistency is retained as originating from the execution end, the feedback end, or a missing marker at either end. Subsequently, the event-level timeliness markers of all action events within the same runtime segment are aggregated to form the timeliness consistency score of that segment. Aggregation uses a weighted ratio method: each action event is assigned a weight, which increases with the action amplitude level, and a weight cap is set to avoid extreme actions dominating. The rules for dividing positions are consistent with those in step one, and the weight configuration is fixed in the equipment type parameter table. The weights of events with consistent timeliness within a segment are summed and then divided by the sum of the weights of all valid events within that segment to obtain the timeliness consistency. When there are no action events or all event delays are missing in a segment, the timeliness consistency is written to a preset default value and marked with "insufficient evidence" so that subsequent attribution steps can identify that the segment lacks timely evidence and adopt a gating strategy. Finally, the timeliness consistency, segment identifier, default mark, and cause statistics are archived together as inputs for subsequent deviation over-limit attribution and work order element generation.

[0030] The present invention is further configured such that, in generating deviation exceedance markers based on deviation time sequence, and combining disturbance co-location coefficients and timeliness consistency to generate attribution categories, confidence levels, location points, and a list of disposal items, the maintenance record and work order elements include: Within each runtime segment, deviation exceedance markers are generated based on the deviation timing and a preset deviation threshold. These markers then generate segment exceedance coverage and duration information, and an attribution activation marker. Specifically, for each runtime segment, valid sampling points of the deviation timing are traversed in a uniform time scale order. The deviation threshold uses a dual-source setting method that is feasible to implement: the permissible deviation standard threshold of the device or loop is used first; when the standard threshold is missing, the high quantile of the deviation distribution within the reference set is used as a substitute threshold, and a safety margin corresponding to the quantization resolution is added to avoid false triggering due to the threshold being close to the noise level. At each valid sampling point, the absolute value of the deviation is compared with the deviation threshold; if it exceeds the threshold, it is marked as exceedance, forming a deviation exceedance marker for that sampling point; otherwise, it is marked as not exceedance. Subsequently, two types of segment overview information are generated within the segment range: coverage information is represented by the ratio of the number of out-of-limit sampling points to the total number of valid sampling points, used to characterize the breadth of out-of-limit occurrences; persistence information is represented by the ratio of the longest consecutive out-of-limit segment length to the effective segment length. Continuous segment statistics adopt a sequential scanning method, ending the current continuous segment and updating the maximum continuous length when encountering a segment that does not exceed the limit. Attribution activation markers are generated based on coverage information and persistence information. Attribution activation adopts a dual-threshold gating system: it takes effect when coverage information reaches a preset coverage threshold and persistence information reaches a preset persistence threshold. Both the coverage threshold and persistence threshold are configurable parameters. The coverage threshold is used to exclude occasional spikes, and the persistence threshold is used to exclude instantaneous jumps. Their values ​​can be adaptively converted into the minimum number of out-of-limit points and the minimum number of consecutive out-of-limit points according to the segment length to ensure consistent gating caliber for segments of different lengths. When the attribution activation flag meets the conditions, a mutually exclusive attribution evidence set is generated based on the disturbance co-occurrence coefficient and the timeliness consistency. An attribution category is then generated based on the mutually exclusive attribution evidence set, forming an attribution evidence difference quantity. Specifically, when the attribution activation flag is active, the disturbance co-occurrence coefficient and timeliness consistency corresponding to the running segment are read. Both are derived from the output of previous steps, with values ​​limited to zero to one and associated with the segment identifier. To avoid stacking multiple common features, the mutually exclusive attribution evidence set constructs only three types of mutually exclusive evidence around these two coefficients, corresponding to three maintainable paths: The first type of evidence targets the load disturbance co-occurrence type, using a multiplicative fusion method. When both the disturbance co-occurrence coefficient and the timeliness consistency are simultaneously high, the evidence is enhanced, used to characterize deviation changes that are in the same direction as load disturbances and where the response time is not significantly biased. The first type of evidence addresses the situation of deviation; the second type targets execution link timeliness anomalies, employing a reverse multiplicative fusion method. This involves inverting the disturbance co-occurrence coefficient and the timeliness consistency, then multiplicatively fusing them. When both are simultaneously low, the evidence is strengthened, used to characterize situations where deviation anomalies are difficult to explain by load and response time is significantly abnormal. The third type targets measurement link bias, employing a combination of low co-occurrence and high timeliness, and introducing continuous information as a constraint factor. This strengthens the evidence when the disturbance co-occurrence coefficient is low, the timeliness consistency is high, and the exceedance is persistent. This is used to characterize situations where load explanation is insufficient but system response time is normal while deviation continues to deviate. After the mutually exclusive attribution evidence set is formed, attribution categories are generated according to the rule of prioritizing the highest-ranking evidence. The category corresponding to the highest-ranking evidence is used as the attribution category. To provide auditable discriminative strength, the attribution evidence difference is calculated. The difference is represented by the difference between the highest and second-highest evidence values; a larger difference indicates stronger evidence discrimination. The confidence level is formed based on the difference in attribution evidence combined with information on excessive coverage of fragments. Specifically, to avoid overlooking the sufficiency of excessive facts based solely on the difference in evidence, the generation of the confidence level is divided into two stages. The first stage forms the confidence score: the difference in attribution evidence is used as the main factor, while coverage information is introduced as a consistency correction factor. When the coverage information is below the coverage threshold, even if the difference in evidence is large, the confidence score is lowered to avoid a small number of excessive points triggering a high-confidence conclusion. The second stage maps the confidence score to discrete confidence levels, which include at least three levels: high, medium, and low. The level threshold is a configurable parameter, and the threshold setting follows the rule that high confidence requires both significant difference in evidence and sufficient excessive coverage. When the attribution effectiveness flag is not effective, the confidence level is fixed as ineffective, and the cause field is recorded as insufficient coverage or persistent insufficiency. Location points are generated based on the attribution category and confidence level in the closed-loop point association relationship, and a list of disposal items is generated in the disposal rule set based on the attribution category and confidence level. Specifically, the generation of location points relies on the closed-loop point association relationship. The closed-loop point association relationship is established in the configuration phase by the point list and loop topology, and includes at least the association edges between setpoints, feedback points, control output points, execution end observation points, and load proxy points, as well as point level attributes. Candidate point sets are selected based on the attribution category: when the attribution category is load disturbance co-occurrence type, candidate points are screened from the load proxy point and its upstream acquisition link points, and points directly related to the generation of load proxy quantity are given priority; when the attribution category is execution link timeliness anomaly type, candidate points are screened from the control output point, execution end observation point, and execution link related points between the two, and sorted from near to far according to link level; when the attribution category is measurement link bias type, candidate points are screened from the feedback point and its acquisition link related points, and points sensitive to calibration and zero drift are given priority. Within the candidate point set, point convergence is performed based on the confidence level: For high confidence, the first point in the ranking is used as the location point; for medium confidence, the first and second points form a location point group; for low confidence, only the minimum point set is output with a "requires verification" label to avoid excessive dispatching. The list of actions is generated in the action rule set. The action rule set is indexed by attribution category and confidence level. Each rule contains a minimum set of necessary actions and the order of action execution. High-confidence rules output a "must-do list," with the number of actions limited by the minimum set principle to avoid overloading; medium-confidence rules output a "must-do plus verification list," adding a link consistency verification; low-confidence rules output a "verification list," prioritizing confirmation of data link and point validity. Maintenance records are generated by associating the operation segment identifier, deviation exceeding limit marker, disturbance co-image coefficient, timeliness consistency, attribution category, confidence level, location point, and disposal item list. Work order elements are extracted from the maintenance records. Specifically, the segment identifier and segment start and end time scale are used as primary keys, and the deviation exceeding limit marker sequence summary information, coverage information, duration information, disturbance co-image coefficient, timeliness consistency, attribution effective marker, attribution category, attribution evidence difference, confidence level, location point, and disposal item list are associated and written to form a structured maintenance record. Deviation exceeding limits is not repeatedly written to the work order with full details. Instead, the maintenance record retains summary fields such as the list of start and end intervals of exceeding limits, the longest continuous interval of exceeding limits, and the index of peak values ​​exceeding limits, making the record traceable and easy to store and retrieve. Based on this, work order elements are extracted: when the attribution effective flag is effective and the confidence level reaches the preset dispatch threshold, the equipment identifier, loop identifier, location point, list of disposal items, segment start and end time, attribution category, and confidence level are extracted from the maintenance record as work order elements; when the confidence level does not reach the dispatch threshold, only the maintenance record is generated without generating work order elements, and the "reason for not dispatching" field is written. The reason field comes from the specific triggers of insufficient coverage, continuous insufficiency, or insufficient confidence, ensuring a closed-loop process and auditability.

[0031] Example 2: Please see Figure 2 This exemplary smart hotel equipment intelligent maintenance system based on AI big data predictive maintenance includes: Loop alignment module: Based on the control loop's collected setpoints, feedback values, control outputs, execution end measurements, and load surcharges, it aligns the time data to generate a deviation timing sequence. Boundary segmentation module: Determines segment boundaries based on changes in setpoints, control outputs, and load agents, and generates running segments by segmenting the deviation time sequence; Health screening module: Generates a baseline set by screening healthy segments based on the deviation stability and output activity of the running segments; Interval modeling module: Establishes interpretable intervals for load surcharge and deviation time series based on the baseline set, and generates disturbance co-image coefficients; Time window modeling module: Establishes response time windows after changes in control output based on the baseline set, and generates timeliness consistency. Attribution Work Order Module: Generates deviation over-limit markers based on deviation time sequence, and generates attribution categories, confidence levels, location points, and a list of disposal items by combining disturbance co-image coefficient and timeliness consistency, and generates maintenance records and work order elements.

[0032] It should be noted that the intelligent maintenance system for smart hotel equipment based on AI big data predictive maintenance provided in the above embodiments and the intelligent maintenance method for smart hotel equipment based on AI big data predictive maintenance provided in the above embodiments belong to the same concept. The specific methods of operation of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the intelligent maintenance system for smart hotel equipment based on AI big data predictive maintenance provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above, and this is not a limitation.

[0033] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A smart hotel equipment intelligent maintenance method based on AI big data predictive maintenance, characterized in that, include: Based on the setpoint, feedback value, control output, execution end observation, and load proxy quantity collected by the control loop, a deviation timing sequence is generated by aligning it with the time base. The segment boundaries are determined based on changes in set values, control outputs, and load agents, and the deviation time sequence is segmented to generate operating segments. A baseline set is generated by selecting healthy segments based on the deviation stability and output activity of the running segments; Based on the benchmark set, an interpretable interval for the load surcharge and deviation time series is established, and disturbance co-image coefficients are generated. Establish a response time window after the control output changes based on the benchmark set, and generate timeliness consistency. Based on the deviation time sequence, a deviation exceeding the limit marker is generated. Combined with the disturbance co-image coefficient and timeliness consistency, an attribution category, confidence level, location point and disposal item list are generated, and maintenance records and work order elements are generated.

2. The intelligent maintenance method for smart hotel equipment based on AI big data predictive maintenance according to claim 1, characterized in that, The generation of offset timing sequences based on a time base includes: The original timing sequence is generated by associating the sampling time identifier with the setpoint, feedback value, control output, execution end observation, and load proxy quantity collected by the control loop. A unified time base is established based on the original timing sequence. The setpoint, feedback value, control output, execution end observation measurement, and load proxy quantity are time aligned to obtain the aligned timing sequence. The deviation timing is formed based on the set value and feedback value in the alignment timing.

3. The intelligent maintenance method for smart hotel equipment based on AI big data predictive maintenance according to claim 2, characterized in that, The process of splitting the time sequence of deviations to generate runtime segments includes: Based on the set values, control outputs, and load proxy quantities in the alignment timing, change markers are determined, and boundary candidate markers are generated. Based on the candidate boundary markers, persistent constraints are applied to generate a set of candidate boundary points; Based on the candidate boundary point set, apply minimum interval constraints and minimum segment length constraints to generate a segment boundary set; Based on the segment boundary set, the deviation time sequence is segmented to generate the running segment corresponding to the segment boundary set.

4. The intelligent maintenance method for smart hotel equipment based on AI big data predictive maintenance according to claim 3, characterized in that, The benchmark set is generated by filtering healthy segments based on their deviation stability and output activity, including: Statistically analyze the segment length and missing test ratio of the running segments, generate valid segment markers, and retain valid segments based on the valid segment markers; Within the effective segment, the upper quantile deviation value and the lower quantile deviation value are extracted based on the deviation time sequence to generate the segment deviation fluctuation range. The segment deviation fluctuation range is then scaled to generate the deviation stability. Within the effective segment, the cumulative change is calculated based on the control output in the alignment sequence, and the cumulative change is scaled to generate the output activity. The deviation stability is compared with a preset stability threshold, and the output activity is compared with a preset activity threshold to generate a health label. Based on the health label, healthy segments are selected from the valid segments to generate a benchmark set.

5. The intelligent maintenance method for smart hotel equipment based on AI big data predictive maintenance according to claim 1, characterized in that, The interpretable intervals for load surcharge and deviation time series established based on the baseline set include: Extract corresponding samples of the benchmark centralized load proxy quantity and deviation to generate a benchmark sample pool; Calculate the load coverage range within the baseline sample pool, and generate a segmented support point set according to the load coverage range; The upper and lower boundaries of the deviation are generated based on the segmented support point set. The upper and lower boundaries of the deviation are then processed to achieve continuity, resulting in the upper and lower intervals of the deviation that vary with the load. Perform a consistency check on the upper and lower intervals of the deviation to generate an interval margin. Correct the upper and lower intervals of the deviation based on the interval margin to generate an interpretable interval.

6. The intelligent maintenance method for smart hotel equipment based on AI big data predictive maintenance according to claim 5, characterized in that, The generated perturbation co-image coefficients include: Extract the correspondence between load surcharge changes and deviation changes from the benchmark set, calculate the load impact time delay, and generate the time delay corrected load surcharge. Within the running segment, query the interpretable interval based on the time-delay correction load agent quantity, and generate a hit flag based on whether the deviation value is contained within the interpretable interval. Within the running segment, direction-consistent markers are generated based on time-delay correction of load agent quantity changes and deviation changes, and direction-consistent markers are retained at sampling locations where load agent quantity changes meet the significance condition; The perturbation co-image coefficient of the running segment is generated by summarizing the hit marker and the direction-consistent marker.

7. The intelligent maintenance method for smart hotel equipment based on AI big data predictive maintenance according to claim 1, characterized in that, Establishing a response time window based on a reference set after a change in control output includes: The system identifies control action moments when the control output reaches a preset change threshold in the benchmark set and generates a benchmark action moment set. Within the time range corresponding to the set of baseline action times, the effective response arrival time is calculated based on the threshold of the change in the execution end observation and feedback value, and execution end delay samples and feedback end delay samples are generated. An execution-end response time window is established based on the execution-end latency sample, and a feedback-end response time window is established based on the feedback-end latency sample.

8. The intelligent maintenance method for smart hotel equipment based on AI big data predictive maintenance according to claim 7, characterized in that, Consistency in generation timeliness includes: Within a running segment, identify the control action moments when the control output reaches a preset change threshold, and generate a set of segment action moments; Within the time range corresponding to the set of fragment action moments, calculate the arrival time of the effective response at the execution end and the arrival time of the effective response at the feedback end, and generate the fragment execution end delay and the fragment feedback end delay. The execution end delay of the fragment is compared with the execution end response time window to form the execution end time mark, and the feedback end delay of the fragment is compared with the feedback end response time window to form the feedback end time mark; Event-level timeliness markers are formed based on the timeliness markers at the execution end and the timeliness markers at the feedback end. Event-level timeliness markers are then aggregated into timeliness consistency by segment.

9. The intelligent maintenance method for smart hotel equipment based on AI big data predictive maintenance according to claim 1, characterized in that, Based on the deviation time sequence, a deviation exceeding limit marker is generated. Combined with the disturbance co-image coefficient and timeliness consistency, an attribution category, confidence level, location point, and list of handling items are generated. Maintenance records and work order elements are then generated, including: Within the running segment, a deviation exceeding the limit marker is generated based on the deviation timing and a preset deviation threshold. Based on the deviation exceeding the limit marker, segment exceeding the limit coverage information and segment exceeding the limit duration information are generated, and an attribution effective marker is generated. When the attribution validity marker meets the conditions, a mutually exclusive attribution evidence set is generated based on the perturbation co-image coefficient and the timeliness consistency. Attribution categories are generated based on the mutually exclusive attribution evidence set, forming the attribution evidence difference quantity. The confidence level is determined by combining the difference in attribution evidence with information on fragment over-coverage. Based on the attribution category and confidence level, location points are generated in the closed-loop location correlation relationship; based on the attribution category and confidence level, a list of disposal items is generated in the disposal rule set. The maintenance record is generated by associating the operation segment identifier, deviation exceeding the limit mark, disturbance co-image coefficient, timeliness consistency, attribution category, confidence level, location point, and disposal item list, and the work order elements are extracted from the maintenance record.

10. A smart hotel equipment intelligent maintenance system based on AI big data predictive maintenance, used to implement the smart hotel equipment intelligent maintenance method based on AI big data predictive maintenance as described in any one of claims 1-9, characterized in that, include: Loop alignment module: Based on the control loop's collected setpoints, feedback values, control outputs, execution end measurements, and load surcharges, it aligns the time data to generate a deviation timing sequence. Boundary segmentation module: Determines segment boundaries based on changes in setpoints, control outputs, and load agents, and generates running segments by segmenting the deviation time sequence; Health screening module: Generates a baseline set by screening healthy segments based on the deviation stability and output activity of the running segments; Interval modeling module: Establishes interpretable intervals for load surcharge and deviation time series based on the baseline set, and generates disturbance co-image coefficients; Time window modeling module: Establishes response time windows after changes in control output based on the baseline set, and generates timeliness consistency. Attribution Work Order Module: Generates deviation over-limit markers based on deviation time sequence, and generates attribution categories, confidence levels, location points, and a list of disposal items by combining disturbance co-image coefficient and timeliness consistency, and generates maintenance records and work order elements.

Images