Artificial intelligence-based industrial internet data intelligent analysis and early warning method

Abstract

The application discloses an industrial internet data intelligent analysis and early warning method based on artificial intelligence. The analysis and early warning method comprises the following steps: collecting equipment time sequence monitoring, process and work order label data to form an original data set, pre-processing and steady state screening to obtain standardized time sequence data; extracting multi-dimensional degradation features and segment fitting, eliminating working condition fluctuation interference, and outputting persistent degradation feature sequence; determining the real degradation starting point of the equipment through double-criterion inflection point detection, calculating the time sequence offset and correcting the lag fault label in combination with the work order label time, building a segmented interval labeling system, and reconstructing the time sequence matching data set to complete model training. The scheme can reduce the time sequence dislocation problem between the operation and maintenance work order fault label and the actual degradation starting time of the equipment, and improve the early hidden fault identification effect of the industrial equipment.

Description

Technical Field

[0001] This invention relates to the field of industrial internet data processing technology, specifically to an artificial intelligence-based intelligent analysis and early warning method for industrial internet data. Background Technology

[0002] Current industrial internet platforms rely on field sensors to collect time-series operational data from equipment. This data, combined with production process records and maintenance work orders, forms the basis for fault prediction models, enabling the identification of equipment degradation and fault prediction. Conventional industry modeling methods directly extract manually recorded downtime and repair reporting times from maintenance work orders as fault sample annotation nodes. By default, the work order annotation time is aligned with the moment equipment performance begins to degrade, and this correspondence is used to build the dataset and iteratively train the model.

[0003] In real-world industrial scenarios, mechanical wear, electrical aging, and structural loosening are all gradual processes, and equipment performance degradation has a long transition phase. Sensors can detect slight degradation signals in advance. However, maintenance personnel only register work orders after obvious problems such as abnormal equipment noise, reduced production capacity, and downtime occur. The manual annotation time is several hours or even tens of hours behind the actual starting point of equipment degradation, resulting in a significant time misalignment.

[0004] The modeling defects caused by timing discrepancies are significant. Directly using misaligned labels for model training leads to the misclassification of many early, subtle degradation features of equipment as normal production data, severing the temporal correspondence between degradation trends and fault labels. This makes it difficult for the model to learn the decay patterns in the early stages of equipment failure; it can only identify obvious anomalies after a complete failure has occurred, increasing the probability of missing early, latent faults. Consequently, the model cannot fundamentally solve the problem of insufficient early warning accuracy caused by label misalignment, and it is even more difficult to adapt to the actual application requirements of early fault prediction throughout the entire lifecycle of industrial equipment. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an AI-based intelligent data analysis and early warning method for the industrial internet, which can reduce the timing misalignment between maintenance work order fault labels and the actual start time of equipment degradation, and improve the identification of early-stage hidden faults in industrial equipment.

[0006] This application provides an intelligent data analysis and early warning method for the industrial internet based on artificial intelligence, the analysis and early warning method comprising: Collect time-series monitoring data, production process data, and maintenance work order fault labeling data from multiple heterogeneous devices to form a raw dataset. Perform noise reduction, missing data completion, and cross-device clock alignment preprocessing on the raw dataset in sequence. Based on the production process parameters, perform steady-state screening on the preprocessed data to remove non-steady-state interference data and output a standardized and effective industrial time-series dataset. The sliding window operation rules are invoked to extract multi-dimensional equipment degradation features from the standardized effective industrial time series dataset. The segmented trend fitting algorithm is called to iteratively fit the equipment degradation features, separate and filter out the interference features corresponding to random process fluctuations, and output the equipment continuous degradation feature sequence. The continuous degradation feature sequence of the device is input into a preset degradation mutation identification mechanism. The inflection point is detected by combining the feature trend slope mutation criterion and the data distribution offset criterion, and the actual degradation start time of the device is located. Read the manual fault labeling time of the maintenance work order, calculate the time series offset by combining it with the actual degradation start time, and perform adaptive forward correction on the original lagging single-point fault label according to the time series offset to obtain the time series aligned fault label. Based on the time-aligned fault labels, a full-deterioration cycle segmented fault interval labeling system is constructed, a time-matching dataset with features and labels aligned is reconstructed, and the time-matching dataset is input into an artificial intelligence early warning model for iterative training.

[0007] In one aspect, the steps of performing steady-state screening on preprocessed data based on production process parameters, removing non-steady-state interference data, and outputting a standardized and effective industrial time-series dataset include: The steady-state threshold range and minimum steady-state duration corresponding to each production process parameter are preset. Time series data in which the production process parameter falls within the corresponding steady-state threshold range and the duration meets the minimum steady-state duration are retained, while non-steady-state interference data such as equipment start-up and shutdown, sudden load changes, process debugging, and material change transitions are removed.

[0008] In one aspect, the step of retrieving the sliding window operation rules to extract multi-dimensional equipment degradation features from the standardized effective industrial time-series dataset includes: An overlapping sliding sampling method with a step size smaller than the window length is used to extract continuous time segments window by window from the standardized effective industrial time series dataset; For each time segment, multidimensional degradation features of the equipment operating status are calculated. These multidimensional degradation features include mean offset features, variance fluctuation features, peak deviation features, and trend slope features. All features extracted from each window are aggregated to construct a multi-dimensional device degradation feature.

[0009] In one aspect, the steps of iteratively fitting the equipment degradation features using a piecewise trend fitting algorithm, separating and filtering out interference features corresponding to random process fluctuations, and outputting a sequence of continuous equipment degradation features include: The device degradation characteristics are divided into several continuous data segments in chronological order; The trend parameters are repeatedly solved by least squares operation for the feature quantities within a single segment, and the fitting residuals are calculated after each round of solution. Compare the residual values ​​with a preset threshold. Data segments whose residual values ​​exceed the preset threshold are classified as interference features caused by random fluctuations in operating conditions and discarded. Data segments with residual values ​​lower than the preset threshold are retained, and the retained data segments are integrated to form a sequence of continuous degradation characteristics of the equipment.

[0010] In one aspect, the steps for detecting inflection points and locating the true starting moment of device degradation by combining the characteristic trend slope abrupt change criterion and the data distribution offset criterion include: Based on the fitting results of each time segment in the continuous degradation feature sequence of the equipment, the trend slope change status of the corresponding time period is statistically analyzed. When the slope fluctuation amplitude of multiple consecutive time segments exceeds the fluctuation range corresponding to the steady-state operation of the equipment, it is determined that the current time period meets the slope change condition. The system retrieves the baseline of historical normal data from the equipment and compares the distribution differences of the characteristic data for the current period. When the overall data offset exceeds the preset critical threshold, it is determined that the current period meets the distribution offset condition. The earliest time-series node that simultaneously satisfies the slope abruptness condition and the distribution offset condition is initially selected as the quasi-degenerate inflection point. A second review is conducted on the short-term window data adjacent to the quasi-degradation inflection point to determine the true starting time of substantial equipment degradation.

[0011] In one aspect, the step of reading the manually marked fault time of the maintenance work order and calculating the time series offset in combination with the actual degradation start time includes: Extract the fault reporting time recorded in the maintenance work order as the manual fault labeling time, and retrieve the actual degradation start time of the equipment; Calculate the time difference between the artificial fault labeling time and the actual degradation start time, and use the time difference as the initial time series offset; The offset statistics of similar deterioration cases in the equipment's historical fault file are retrieved, and the initial time offset is corrected by mean to obtain the time offset used for tag calibration.

[0012] In one aspect, the step of adaptively shifting and correcting the original lagging single-point fault labels based on the time offset to obtain time-aligned fault labels includes: The system classifies and matches the corresponding type of the current device fault, retrieves the preset fault lag correction coefficient, and adapts and corrects the timing offset to obtain the optimized effective offset duration. Using the time of the manual fault labeling as the reference time point, the marking time of the original single-point fault label is shifted forward and corrected in combination with the effective offset duration; The corrected tag time node is matched and verified with the device's continuous degradation feature sequence. Once the verification is successful, the time-aligned fault tag is generated.

[0013] In one aspect, the steps of constructing a full-deterioration cycle segmented fault interval labeling system based on the time-aligned fault labels include: Using the fault outbreak time corresponding to the time-aligned fault label as the endpoint, and combining the historical deterioration and evolution patterns of the equipment, the continuous equipment operation segments are divided backward. Based on the characteristic fluctuation amplitude and the degree of change in degradation trend, the back-traced equipment operation period is divided into a slight deterioration stage, a fault development stage, and a fault imminent outbreak stage. Independent category identifiers are configured for the minor degradation stage, the fault development stage, and the fault imminent outbreak stage, replacing the original single time point labeling method, forming a segmented fault interval labeling system that covers the equipment degradation process.

[0014] In one aspect, the steps for reconstructing a temporal matching dataset with features aligned to labels include: Using the full degradation cycle segmented fault interval labeling system as the labeling benchmark, match the time series feature segments corresponding to the standardized effective industrial time series dataset; Filter out effective feature data that are associated with the equipment degradation trend within the time-series feature segment interval, and remove noisy samples with disordered and mismatched feature changes; The effective time-series features that are filtered and retained are bound one-to-one with the corresponding segmented fault labels, and the data dimension and time-series dimension are aligned to obtain the time-series matching dataset.

[0015] In one aspect, the step of inputting the time-series matching dataset into an artificial intelligence early warning model for iterative training includes: The time-series matching dataset is divided into training samples and validation samples, and the divided sample data is fed into the artificial intelligence early warning model. The artificial intelligence early warning model updates its internal weight parameters layer by layer based on the segmented interval labeling information within the sample. During the training process, a pre-set constraint loss function is used to correct the fitting bias, weaken the interference weight of random process fluctuation characteristics, and enhance the learning ability of the equipment's early weak degradation characteristics. The parameters are iteratively updated until the model error converges, thus obtaining an early warning model for early hidden faults.

[0016] The beneficial effects of this invention are as follows: This application collects multi-source data and preprocesses it through steady-state screening, sliding window feature extraction, and segmented trend fitting to eliminate interference from operating condition fluctuations, purifying the continuous degradation feature sequence and providing clean data; then, it uses slope mutation and distribution offset as dual criteria to lock the true degradation start time of the equipment, calculates the time series offset between work order labeling and degradation start point, and solves the problem of insufficient quantification of time series deviation; it uses the offset to correct lagging fault labels forward, builds a segmented labeling system covering the degradation cycle, and repairs the defect of feature and label time series separation; finally, it reconstructs the time series matching dataset for model training, completely retains the early degradation corresponding samples, broadens the effective fault feature range, alleviates the small sample overfitting phenomenon, and enables the model to fully learn the degradation law in the bud stage, reducing the probability of early fault missed reporting. This solution forms an integrated closed loop from data regularization, degradation source tracing, label correction to dataset reconstruction and model training, eradicating various modeling defects caused by the time series lag of work order labels from the data source, and improving the reliability of early fault warning for industrial equipment. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0018] Figure 1 This is a schematic diagram illustrating the process steps of the AI-based industrial internet data intelligent analysis and early warning method proposed in this application. Detailed Implementation

[0019] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0020] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0021] like Figure 1 As shown, this application provides an intelligent data analysis and early warning method for the industrial internet based on artificial intelligence. This embodiment selects the processing process in a parts processing workshop, using a three-phase asynchronous spindle motor and an automated guided vehicle as monitoring objects. On-site data is collected using an edge acquisition gateway paired with an integrated temperature and vibration sensor. The analysis and early warning method includes: Step S10 involves collecting time-series monitoring data from multi-source heterogeneous equipment, production process data, and maintenance work order fault labeling data to form a raw dataset. This raw dataset undergoes preprocessing including noise removal, missing data completion, and cross-device clock alignment. Based on production process parameters, the preprocessed data is subjected to steady-state screening to remove non-steady-state interference data, resulting in a standardized and valid industrial time-series dataset. This step requires the simultaneous collection of three types of industrial field data: time-series monitoring data is acquired by piezoelectric vibration sensors, platinum resistance temperature acquisition elements, and current transformers. Vibration sensors continuously output acceleration waveforms at a default sampling frequency of 5120Hz, while temperature and current data are uploaded at a 1Hz frequency. Production process data is taken from the real-time records of spindle speed, feed rate, and cutting load range stored by the workshop's programmable logic controller. Maintenance work order data is stored in the manufacturing execution system, including manually entered fault reports, downtime maintenance, and component replacement time text records. All three types of data are uniformly collected and merged in the gateway's local cache to generate the raw dataset. The preprocessing operations described in this step include noise removal, missing data filling, and cross-device clock alignment. Clock calibration employs three lower-level execution methods. Noise removal utilizes two algorithms: wavelet threshold denoising and three-fold standard difference constant removal. For missing data shorter than 30 seconds, linear interpolation is used for completion; for data longer than 30 seconds, historical averages under the same operating conditions are retrieved for filling. Clock calibration relies on the gateway's built-in network time synchronization service to uniformly correct the timestamps of each sensor, eliminating time offsets from multiple devices. In the steady-state screening stage, standard process ranges for each process are pre-entered. For example, in the roughing process, the standard range for spindle speed is 8000–12000 rpm, and the load threshold is 20–45%. Time series segments where the speed and load fall entirely within the standard range and are maintained continuously for at least 120 seconds are retained. Non-steady-state operating data such as equipment start-up and shutdown, tool change debugging, and idle standby are automatically filtered out. After screening, a standardized and effective industrial time series dataset with a unified sampling interval and timestamp alignment is output. The purpose of this preprocessing and screening step is to remove interference data with no analytical value in advance, avoid introducing invalid samples in subsequent feature calculation and trend fitting stages, reduce the computational load of subsequent algorithms, and ensure that all data entering the feature extraction stage corresponds to stable equipment operating conditions.

[0022] Step S20: The sliding window operation rules are invoked to extract multi-dimensional equipment degradation features from the standardized and effective industrial time-series dataset. A segmented trend fitting algorithm is called to iteratively fit the equipment degradation features, separating and filtering out interference features corresponding to random process fluctuations, and outputting a sequence of continuous equipment degradation features. An overlapping sliding sampling mechanism is used to process the standardized time-series dataset. The window length is fixed at 10 minutes, the single sliding step size is configured at 5 minutes, and the overlap ratio of adjacent windows is 50%. This parameter combination can fully capture the continuous change pattern of slow equipment degradation and expand the number of feature samples, avoiding the loss of early degradation features due to non-overlapping windows. For each segment of the extracted time-series dataset, four types of degradation feature values ​​are calculated: mean offset feature to characterize the degree of drift of the long-term operating baseline value of the equipment; variance fluctuation feature to reflect the short-term operating oscillation amplitude; peak deviation feature to capture instantaneous impact anomalies; and trend slope feature to quantify the rate of index decay per unit time. The features calculated from all windows are summarized to form a complete multi-dimensional equipment degradation feature set. Then, the segmented... The trend fitting algorithm iteratively solves all degradation features, dividing the feature sequence into continuous sub-data segments of 3 minutes each according to time sequence. For each feature segment, the least squares algorithm is used to repeatedly solve the linear trend parameters. After each round of parameter update, the fitting residual of this segment is calculated. The residual critical threshold is pre-calibrated according to the equipment model. For conventional power equipment, this threshold is set to 0.08. Segments with residual values ​​exceeding the critical threshold are directly judged as random process interference data caused by cutting amount fluctuations and raw material material differences and are discarded. Segments with residual values ​​consistently below the threshold are all retained. All retained segments are integrated to obtain a continuous degradation feature sequence that only reflects the long-term aging and decay of the equipment. The reason for setting up the segmented fitting and residual screening links is to distinguish between two types of time series fluctuations: one is short-term oscillations caused by random changes in the production process, and the other is a long-term unidirectional deterioration trend formed by mechanical wear and electrical aging. After removing process interference features by residual comparison, irrelevant fluctuations can be removed, and the change patterns related to the aging of the equipment itself can be retained, eliminating the interference of operating condition disturbances on the subsequent determination of the deterioration starting point.

[0023] Step S30: Input the continuous degradation feature sequence of the equipment into the preset degradation mutation identification mechanism, and combine the feature trend slope mutation criterion and data distribution offset criterion to detect inflection points and locate the actual degradation start time of the equipment; simultaneously load two sets of judgment logic to complete the inflection point retrieval. The first set is the trend slope mutation judgment condition, which iterates through the slope values ​​obtained by fitting each time series segment. If the slope fluctuation amplitude of five consecutive adjacent segments exceeds the fluctuation range corresponding to the steady-state operation state of the equipment (the upper limit of slope fluctuation for general equipment steady-state operation is 0.003), it is determined that the current time period meets the slope mutation condition; the second set is to retrieve the historical normal data baseline of the equipment and compare the distribution differences of the feature data of the current time period. The Kolmogorov-Smirnov distribution test is used to determine the distribution difference between two sets of data. When the overall deviation exceeds the preset critical threshold of 0.12, the current time period is determined to meet the distribution deviation condition. The earliest time node that simultaneously meets both judgment conditions within the same time interval is retrieved as the initial quasi-degradation inflection point. To avoid single misjudgments caused by instantaneous impacts and electromagnetic interference, the data in the adjacent 2-minute short window before the inflection point is automatically extracted and the dual judgment criteria are repeatedly checked. Only after both judgment results meet the criteria can the node be determined as the true starting time of substantial equipment degradation. This step relies on the degradation characteristics of the equipment itself to complete the inflection point location and can objectively quantify the time difference between the time marked on the work order and the actual degradation starting point.

[0024] Step S40: Read the manual fault labeling time of the maintenance work order, calculate the time series offset by combining it with the actual degradation start time, and perform adaptive forward correction on the original lagging single-point fault label based on the time series offset to obtain the time series aligned fault label; read the fault reporting time manually filled in in the maintenance work order as the manual fault labeling time, retrieve the actual degradation start time of the equipment obtained by the aforementioned inflection point detection, and calculate the time difference between the two time nodes as the initial time series offset. The initial offset corresponding to the fault of conventional industrial equipment is generally distributed in the range of 4.5–18 hours; retrieve the historical fault files of the same model of equipment stored locally, extract the offset statistical mean of more than 300 sets of similar degradation cases, and use the mean correction method to numerically calibrate the initial time series offset to obtain the final standard time series offset used for label correction. The historical mean correction step is set to eliminate the offset distortion caused by the deviation of a single fault work order record and improve the time series matching accuracy of subsequent label correction.

[0025] Step S50: Construct a segmented fault interval labeling system for the entire degradation cycle based on time-aligned fault labels, reconstruct a time-series matching dataset with aligned features and labels, and input the time-series matching dataset into the artificial intelligence early warning model for iterative training. Using the fault outbreak time corresponding to the corrected time-aligned fault labels as the time-series endpoint, retrieve the evolution patterns formed by years of equipment degradation history data from the workshop, and trace back the complete degradation interval along the time axis. Based on the feature fluctuation amplitude and the gradient of the trend slope, the complete interval is divided into three continuous cycles: a weak degradation stage, a fault development stage, and a fault imminent outbreak stage. The weak degradation stage corresponds to the initial wear stage, where the feature fluctuation amplitude is less than 1.2 times the steady-state threshold. In the fault development stage, the indicators continuously decay in one direction. In the fault imminent outbreak stage, the feature fluctuation exceeds twice the steady-state threshold. Each of the three stages is assigned an independent digital category identifier, replacing the traditional single-time-point labeling method, forming a comprehensive labeling system covering the entire equipment. A segmented fault interval labeling system for the entire degradation process is proposed. Based on this segmented labeling system, time-series feature segments corresponding to standardized and effective industrial time-series datasets are matched. Effective feature data positively correlated with the equipment degradation trend within the interval are automatically selected, and noisy samples with disordered feature changes or those that do not match the segmented interval change patterns are removed. The remaining effective time-series features are bound to the corresponding segmented fault identifiers segment by segment, achieving dual alignment of the time-series and feature dimensions, and constructing a time-series matching dataset. The segmented labeling and sample selection steps can completely retain fault samples corresponding to the early weak degradation of the equipment, expand the effective fault feature interval, and alleviate the small sample modeling defects caused by the natural scarcity of industrial fault samples.

[0026] In this embodiment, multi-source data is collected and preprocessed, including steady-state screening, sliding window feature extraction, and segmented trend fitting to eliminate interference from operating condition fluctuations, thus refining the continuous degradation feature sequence and providing clean data. Then, the true degradation start time of the equipment is determined using slope abrupt change and distribution offset as dual criteria. The temporal offset between work order labeling and the degradation start point is calculated to address the inability to quantify temporal deviations. The offset is used to correct lagging fault labels, building a segmented labeling system covering the degradation cycle and repairing the defect of feature and label temporal separation. Finally, the temporal matching dataset is reconstructed for model training, fully preserving early degradation corresponding samples, broadening the effective fault feature range, alleviating small-sample overfitting, and enabling the model to fully learn the degradation patterns in the nascent stage, reducing the probability of early fault underreporting. This solution forms an integrated closed loop from data regularization, degradation tracing, label correction to dataset reconstruction and model training, eradicating various modeling defects caused by lagging work order label temporalities from the data source and improving the reliability of early fault warnings for industrial equipment.

[0027] In one embodiment of this application, the step of performing steady-state screening on preprocessed data based on production process parameters, removing non-steady-state interference data, and outputting a standardized and effective industrial time-series dataset includes: Step S110: Preset the steady-state threshold range and minimum steady-state duration for each production process parameter. Retain time-series data where the production process parameter falls within the corresponding steady-state threshold range and the duration meets the minimum steady-state duration, while removing non-steady-state interference data such as equipment start-up and shutdown, sudden load changes, process debugging, and material changeovers. Before equipment operation, maintenance personnel calibrate the steady-state threshold range and minimum steady-state duration for each production process parameter based on long-term stable production records of the production line. Under general processing conditions, the load steady-state range is set to 20% to 45%, the speed steady-state range is set to 8000 to 12000 revolutions per minute, and the minimum steady-state duration is uniformly set to 120 seconds. Continuously monitor each process indicator by traversing each segment of time-series data. Retain time-series segments where all indicators fall within the corresponding calibration range and the continuous stable running time meets the preset standard. For equipment start-up and shutdown moments, and sudden increases in processing load, the data is excluded. All time-series data generated during transitional periods such as downsizing, process parameter adjustments, and material changes are identified as non-steady-state interference data and directly discarded. The core reason for setting up this filtering operation is that only monitoring data generated during stable operation can objectively reflect the continuous aging and degradation trend of the equipment. Fluctuations in indicators during transitional conditions such as start-up, shutdown, and debugging are short-term disturbances caused by human operation. If mixed into the dataset, they will interfere with the accuracy of subsequent degradation feature extraction and trend fitting. The standardized and effective industrial time-series dataset output after filtering can remove worthless interference samples from the source, reducing the computational burden on subsequent algorithms.

[0028] In one embodiment of this application, the step of retrieving sliding window operation rules to extract multi-dimensional equipment degradation features from a standardized and valid industrial time-series dataset includes: Step S210: Using an overlapping sliding sampling method with a step size smaller than the window length, continuous time segments are extracted window by window from the standardized effective industrial time series dataset. Step S211: For each time segment, calculate the multidimensional degradation features of the equipment operating status. The multidimensional degradation features include mean offset features, variance fluctuation features, peak deviation features, and trend slope features. Step S212: Aggregate all features extracted from each window to construct multi-dimensional device degradation features.

[0029] In this embodiment, an overlapping sliding sampling mechanism is used to process standardized and effective industrial time-series datasets. The actual configured window length is 10 minutes, and the single sliding step size is set to 5 minutes. The step size is smaller than the overall window length, resulting in a 50% data overlap between adjacent sampling windows. The complete time-series data stream is sequentially segmented according to this truncation rule, resulting in continuous, uninterrupted time-series segments. For each individually extracted time-series segment, numerical calculations are performed sequentially to generate four types of multi-dimensional degradation features that characterize the long-term operational degradation state of the equipment. Using overlapping sliding windows for sampling avoids the loss of early, weak degradation signals due to uninterrupted window segmentation. Simultaneous multi-layer feature calculation captures different dimensions of equipment aging performance, and the feature set formed after aggregating all window features can completely carry information on the equipment's degradation changes throughout the entire process.

[0030] In one embodiment of this application, the steps of iteratively fitting equipment degradation features using a piecewise trend fitting algorithm, separating and filtering out interference features corresponding to random process fluctuations, and outputting a sequence of continuous equipment degradation features include: Step S220: Divide the device degradation characteristics into several continuous data segments according to the time sequence; Step S221: Use least squares to repeatedly solve for the trend parameters of the feature quantities within a single segment, and calculate the fitting residuals after each round of solving. Step S222: Compare the residual values ​​with the preset threshold. Data segments with residual values ​​exceeding the preset threshold are classified as interference features caused by random fluctuations in operating conditions and discarded. Step S223: Data segments with residual values ​​lower than a preset threshold are retained, and the retained data segments are integrated to form a sequence of continuous degradation characteristics of the equipment.

[0031] This embodiment divides the data into multiple segments of continuous time series, each three minutes long, according to chronological order. This allows for segmented and refined analysis of the degradation trend. For any segment containing all feature values, the least squares method is used to iteratively solve for the linear trend parameters corresponding to that segment. After each parameter update, the fitting residuals for all features in that segment are calculated immediately. Before the equipment is officially put into analysis and early warning, technicians calibrate the residual judgment threshold based on the equipment model and normal operating conditions. For general power equipment, this threshold is uniformly set to 0.08. The residual values ​​calculated in real time for each segment are compared with the preset threshold. If the residual value of a certain segment is higher than the calibrated threshold, it indicates that the fluctuation of the index in that segment is caused by random production conditions such as raw material differences and changes in cutting amount. This type of data segment is directly judged as an interference feature and discarded. For each segment where the residual value is consistently lower than the preset threshold, all data are retained. Finally, all the retained valid data segments are integrated and concatenated to obtain a sequence of continuous degradation features of the equipment that reflects only the long-term aging and decay law of the equipment itself.

[0032] In one embodiment of this application, the step of detecting inflection points and locating the true starting moment of device degradation by combining abrupt change in characteristic trend slope criteria and data distribution offset criteria includes: Step S310: Based on the fitting results of each time segment in the equipment continuous degradation characteristic sequence, the trend slope change status of the corresponding time period is statistically analyzed. When the slope fluctuation amplitude of multiple consecutive time segments exceeds the fluctuation range corresponding to the steady-state operation of the equipment, it is determined that the current time period meets the slope change condition. Step S320: Retrieve the historical normal data baseline of the equipment, compare the distribution differences of the characteristic data of the current period, and determine that the current period meets the distribution offset condition when the overall data offset exceeds the preset critical threshold. Step S330: The earliest time-series node that simultaneously satisfies the slope abrupt change condition and the distribution shift condition is initially selected as the quasi-degenerate inflection point. Step S340 involves a secondary review of the short-term window data adjacent to the degradation inflection point to determine the true starting time of substantial equipment degradation. In this embodiment, the slope abrupt change judgment logic corresponding to S310 is executed first. It iterates through the trend parameters obtained from fitting calculations of each time series data segment within the sequence, continuously calculating the fluctuation amplitude of the trend slope for each time period. If the slope fluctuation amplitude of five consecutive adjacent time series segments exceeds the pre-defined upper limit of the equipment's steady-state operation fluctuation range (this upper limit is generally 0.003 for general equipment), then the current time period is marked as meeting the slope abrupt change condition. Subsequently, the distribution offset judgment logic corresponding to S320 is executed, retrieving the normal data baseline formed by 720 hours of continuous fault-free operation of the equipment stored locally. The Kolmogorov-Smirnov distribution test is used to determine the characteristics of the current time period to be detected. The data is compared with the baseline to determine the distribution difference. Once the overall data offset exceeds the preset critical threshold of 0.12, the current time period is deemed to meet the distribution offset condition. All time-series nodes are traversed to select the earliest time node that simultaneously meets both judgment conditions within the same time period. This node is used as the preliminary quasi-degradation inflection point. To avoid single misjudgments caused by instantaneous electromagnetic interference and short-term impacts, operation S340 is executed to extract a short window of data two minutes before the quasi-degradation inflection point. The above two sets of judgment logic are repeated to complete two verifications. After both verification results meet the standard simultaneously, the node is confirmed as the true starting point of substantial degradation in equipment performance. This step objectively locates the degradation starting point based on the equipment's own long-term degradation data, accurately quantifying the time difference between the manually labeled time on the work order and the actual degradation starting point of the equipment, thus identifying label timing misalignment deviations that traditional solutions cannot detect at their root.

[0033] In one embodiment of this application, the step of reading the manually marked fault time of the maintenance work order and calculating the time series offset in combination with the actual degradation start time includes: Step S410: Extract the fault reporting time recorded in the maintenance work order as the manual fault labeling time, and retrieve the actual degradation start time of the equipment. Step S411: Calculate the time difference between the time of manual fault labeling and the actual time of degradation start, and use the time difference as the initial time series offset. Step S412: Retrieve the offset statistics of similar deterioration cases in the equipment's historical fault file, and perform mean correction on the initial time offset to obtain the time offset used for tag calibration. The system retrieves maintenance work order records, extracts the manually registered fault reporting time from the work order as the manual fault labeling time, and simultaneously retrieves the actual equipment degradation start time output from the aforementioned inflection point detection stage. Both sets of time nodes carry uniformly calibrated timestamps, allowing for direct difference calculation. The system calculates the time difference between the manually labeled fault time and the actual equipment degradation start time, using this time difference as the uncorrected initial time series offset. In typical industrial operation scenarios, the initial time series offsets corresponding to similar wear and aging faults are mostly distributed in the range of 4.5 hours to 18 hours. A single fault record is easily affected by the delay in registration by maintenance personnel and the time required for on-site fault handling, resulting in extreme deviations. The system reads historical fault files of the same model of equipment stored in the local database for a long time, summarizes the historical offset statistics of more than 300 sets of similar equipment and the same fault mode, and corrects the initial time series offset by calculating the average value, eliminating numerical distortion caused by abnormal records of individual work orders. Finally, it outputs a time series offset with higher accuracy and adaptability to the common operating conditions of similar equipment. Setting up a historical case mean correction step can eliminate random biases caused by a single fault registration, allowing the calculated time offset to conform to the real lag pattern of long-term deterioration of this type of equipment. This provides an objective and stable time correction benchmark for subsequent fault label forwarding correction operations, quantifies the inherent time misalignment between work order labeling and the starting point of equipment deterioration from a numerical perspective, and solves the shortcomings of existing technologies that cannot quantify time lag and can only directly adopt the original time of the work order.

[0034] In one embodiment of this application, the step of adaptively shifting and correcting the original lagging single-point fault labels based on the time offset to obtain time-aligned fault labels includes: Step S420: Classify and match the corresponding type of the current equipment fault, retrieve the preset fault hysteresis correction coefficient, and adapt and correct the timing offset to obtain the optimized effective offset duration. Step S421: Using the time of manual fault labeling as the reference time point, the labeling time of the original single-point fault label is shifted forward and corrected in combination with the effective offset duration. Step S422: Perform time-series matching verification between the corrected tag time node and the equipment's continuous degradation feature sequence. If the verification passes, generate a time-aligned fault tag. Read the time-series offset obtained after mean correction. First, match the fault code to the fault category corresponding to the current equipment degradation. Pre-calibrate a specific fault hysteresis correction coefficient within each fault type. The correction coefficient for wear-related faults is uniformly set to 0.92, and the correction coefficient for electrical winding aging faults is uniformly set to 0.96. Multiply the corresponding fault hysteresis correction coefficient by the aforementioned time-series offset to obtain the effective offset duration adapted to the equipment's operating conditions. The introduction of a fault classification correction coefficient considers the differences in degradation evolution rates among different fault types. Using a uniform, uncorrected offset would cause the tag's forward shift to be too large or too small. Classification coefficient calibration can reduce the difference between the offset and the corrected time. The error lies between the actual degradation evolution pattern of the equipment; after calculating the effective offset duration, the manually registered fault reporting time in the maintenance work order is used as the baseline endpoint. The original single-point fault label's annotation time is shifted forward along the time axis, and the shift duration is matched with the calculated effective offset duration to eliminate the overall time lag caused by manual annotation. After the label time shift correction is completed, the equipment's continuous degradation feature sequence output from the previous segmented trend fitting is automatically retrieved for time-series matching verification. The verification standard is limited to a continuously increasing degradation trend within a 10-minute interval before and after the corrected label. If the degradation trend within the interval is stable and unidirectionally decaying, the verification is considered passed. After the verification is qualified, the time-series aligned fault label is output. This step relies on the quantified time offset to complete the label shift correction, and then combines it with the time-series matching verification mechanism to ensure that the label corresponds synchronously with the equipment degradation trend, thus correcting the core defect of label time-series misalignment from the data annotation level.

[0035] In one embodiment of this application, the step of constructing a segmented fault interval labeling system based on time-aligned fault labels for the entire degradation cycle includes: Step S510: Using the fault outbreak time corresponding to the time sequence aligned fault label as the endpoint, and combining the historical deterioration and evolution patterns of the equipment, backtrack to divide the continuous equipment operation segments. Step S511: Based on the characteristic fluctuation amplitude and the degree of change in degradation trend, the back-traced equipment operating period is divided into the slight deterioration stage, the fault development stage, and the fault imminent outbreak stage. Step S512 involves configuring independent category identifiers for the minor degradation stage, the fault development stage, and the fault imminent outbreak stage, replacing the original single-time-point labeling method to form a segmented fault interval labeling system covering the equipment degradation process. The fault outbreak time corresponding to the time-aligned fault label after time-series correction is used as the endpoint of the entire degradation time-series chain. The complete performance degradation evolution pattern is summarized from years of accumulated degradation history data from long-term equipment operation. The entire continuous operating period from the initial performance degradation to the complete fault outbreak is traced backward along the time axis, no longer limited to a single time point of the work order to extract sample intervals. This covers the entire evolution cycle of the equipment from minor degradation to severe fault. The characteristic fluctuation amplitude and trend slope change gradient of each time period are read to divide the complete degradation interval, thereby defining three consecutive time-series equipment operating cycles. The system is divided into three stages: slight degradation, fault development, and impending fault. The slight degradation stage corresponds to the early stages of equipment wear and aging, with indicator fluctuations within 1.2 times the steady-state operating threshold. The degradation trend is gradual and difficult to detect manually. In the fault development stage, indicators show a continuous unidirectional decline, and the fluctuation amplitude steadily increases. In the impending fault stage, indicator fluctuations exceed twice the steady-state threshold, and the fault characteristics become clearly identifiable. Each of these three stages corresponds to a different evolution level of equipment degradation. Independent, non-repeating numerical category identifiers are assigned to each of the three stages, forming a segmented fault interval labeling system that covers the entire equipment degradation process. This step splits the label intervals into multiple stages according to the degradation evolution gradient, ensuring that each early slight degradation feature can be matched with a corresponding fault labeling interval. This eliminates sample segmentation errors caused by label temporal misalignment from the data labeling logic level, providing a segmented and accurate labeling benchmark for subsequent screening of effective fault features and reconstructing the temporal matching dataset. This fundamentally expands the effective feature coverage of fault samples and alleviates the model training defects caused by the scarcity of industrial fault samples.

[0036] In one embodiment of this application, the step of reconstructing a temporal matching dataset with feature and label alignment includes: Step S520: Using the full degradation cycle segmented fault interval labeling system as the labeling benchmark, match the time series feature segments corresponding to the standardized and effective industrial time series dataset; Step S521: Filter the effective feature data that are associated with the equipment degradation trend within the time-series feature segment interval, and remove noisy samples with disordered and mismatched feature changes. Step S522: The selected and retained valid time-series features are bound one-to-one with the corresponding segmented fault labels to perform dual alignment of data dimension and time-series dimension, thus obtaining the time-series matching dataset. Each deterioration time series interval within the labeling system is matched and associated one-to-one with time series feature segments of the same length and timestamp in the standardized effective industrial time series dataset, thus achieving a correspondence between the labeled interval and the time series range of the original monitoring data. Subsequently, the data filtering operation in step S521 is performed to analyze the data change patterns within the matched time series feature segments segment by segment. Valid feature data with a change trend within the interval that is consistent with the direction of continuous aging and decay of the equipment and has stable unidirectional numerical changes are retained. Noisy samples with chaotic and irregular fluctuations and changes that are obviously contrary to the deterioration development logic of that stage are directly removed, thereby filtering out residual irrelevant disturbance data and reducing the interference of invalid samples on model training. After completing the sample filtering, step S522 is performed to bind and store all the remaining valid time series feature segments and the segmented fault independent category identifiers corresponding to their respective time periods in groups. Simultaneously, the dual alignment processing of the data's feature dimension and the preceding and following time series dimension is completed, generating a time series matching dataset that perfectly matches the feature change patterns with the segmented fault labels. This step relies on a complete segmented labeling system to accurately match the characteristics of each degradation stage, filter out irrelevant noise and complete bidirectional alignment, retain effective fault samples corresponding to the slight degradation in the early stage of equipment, broaden the effective coverage range of fault features, and alleviate the model overfitting problem caused by the scarcity of industrial fault samples.

[0037] In one embodiment of this application, the step of inputting a time-series matching dataset into an artificial intelligence early warning model for iterative training includes: Step S530: Divide the time-series matching dataset into training samples and validation samples, and send the divided sample data into the artificial intelligence early warning model. Step S531: The artificial intelligence early warning model updates its internal weight parameters layer by layer based on the segmented interval labeling information within the sample. Step S532: During the training process, the fitting bias is corrected by combining the preset constraint loss function, the interference weight of random process fluctuation characteristics is weakened, and the learning ability of early weak degradation characteristics of equipment is strengthened. Step S533: Iteratively update the parameters until the model error converges to obtain an early warning model for latent faults. For the reconstructed time-series matching dataset, samples are split according to their temporal order, using a 7:2 split ratio to distinguish between training and validation samples. The splitting is done sequentially along the time axis without randomly shuffling the sample order to avoid leaking temporal correlation information. After splitting, both types of samples are uniformly input into the AI ​​early warning model for training. During step S531, the AI ​​early warning model reads the segmented interval category labels carried by the samples, distinguishing the characteristics of different stages—weak degradation, fault development, and impending fault outbreak—based on the labels. The model then backpropagates the changes in the network's internal weight parameters layer by layer, enabling the AI ​​early warning model to gradually learn the indicator change patterns corresponding to different degradation stages. The dedicated constraint loss preset in step S532 is used throughout the training process. The function corrects fitting bias. The loss function incorporates two types of control constraints. The first constraint reduces the weight of features corresponding to random process fluctuations, weakening the impact of operating clutter on the model's judgment. The second constraint increases the loss weight of features corresponding to weak degradation stages, forcing the model to prioritize capturing subtle attenuation signals in the early stages of equipment degradation, thus specifically strengthening the model's ability to identify and learn early weak degradation features. The model weight parameters are continuously updated iteratively. At the end of each iteration, the overall prediction error of training and validation samples is simultaneously calculated. When the error fluctuation amplitude of 12 consecutive iterations is less than 0.005, the model is considered to have reached the convergence criterion, the parameter update process is stopped, and a dedicated early warning model that can accurately identify early hidden degradation faults in equipment is output. This step relies on a segmented, precisely labeled dataset and a loss function with dual constraints for training. This not only suppresses the identification interference caused by irrelevant process disturbances but also forces the model to focus on learning early degradation features, reducing the probability of missed detection of hidden faults in industrial equipment and improving the model's generalization and early warning performance in multi-operating conditions and multi-equipment scenarios.

[0038] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A data intelligent analysis and early warning method for the industrial internet based on artificial intelligence, characterized in that, The analysis and early warning methods include: Collect time-series monitoring data, production process data, and maintenance work order fault labeling data from multiple heterogeneous devices to form a raw dataset. Perform noise reduction, missing data completion, and cross-device clock alignment preprocessing on the raw dataset in sequence. Based on the production process parameters, perform steady-state screening on the preprocessed data to remove non-steady-state interference data and output a standardized and effective industrial time-series dataset. The sliding window operation rules are invoked to extract multi-dimensional equipment degradation features from the standardized effective industrial time series dataset. The segmented trend fitting algorithm is called to iteratively fit the equipment degradation features, separate and filter out the interference features corresponding to random process fluctuations, and output the equipment continuous degradation feature sequence. The continuous degradation feature sequence of the device is input into a preset degradation mutation identification mechanism. The inflection point is detected by combining the feature trend slope mutation criterion and the data distribution offset criterion, and the actual degradation start time of the device is located. Read the manual fault labeling time of the maintenance work order, calculate the time series offset by combining it with the actual degradation start time, and perform adaptive forward correction on the original lagging single-point fault label according to the time series offset to obtain the time series aligned fault label. Based on the time-aligned fault labels, a full-deterioration cycle segmented fault interval labeling system is constructed, a time-matching dataset with features and labels aligned is reconstructed, and the time-matching dataset is input into an artificial intelligence early warning model for iterative training.

2. The analysis and early warning method according to claim 1, characterized in that, The steps involved in performing steady-state screening on preprocessed data based on production process parameters, removing non-steady-state disturbance data, and outputting a standardized and valid industrial time-series dataset include: The steady-state threshold range and minimum steady-state duration corresponding to each production process parameter are preset. Time series data in which the production process parameter falls within the corresponding steady-state threshold range and the duration meets the minimum steady-state duration are retained, while non-steady-state interference data such as equipment start-up and shutdown, sudden load changes, process debugging, and material change transitions are removed.

3. The analysis and early warning method according to claim 1, characterized in that, The steps of retrieving the sliding window operation rules to extract multi-dimensional equipment degradation features from the standardized and effective industrial time-series dataset include: An overlapping sliding sampling method with a step size smaller than the window length is used to extract continuous time segments window by window from the standardized effective industrial time series dataset; For each time segment, multidimensional degradation features of the equipment operating status are calculated. These multidimensional degradation features include mean offset features, variance fluctuation features, peak deviation features, and trend slope features. All features extracted from each window are aggregated to construct a multi-dimensional device degradation feature.

4. The analysis and early warning method according to claim 1, characterized in that, The steps of iteratively fitting the equipment degradation features using a segmented trend fitting algorithm, separating and filtering out interference features corresponding to random process fluctuations, and outputting a sequence of continuous equipment degradation features include: The device degradation characteristics are divided into several continuous data segments in chronological order; The trend parameters are repeatedly solved by least squares operation for the feature quantities within a single segment, and the fitting residuals are calculated after each round of solution. Compare the residual values ​​with a preset threshold. Data segments whose residual values ​​exceed the preset threshold are classified as interference features caused by random fluctuations in operating conditions and discarded. Data segments with residual values ​​lower than the preset threshold are retained, and the retained data segments are integrated to form a sequence of continuous degradation characteristics of the equipment.

5. The analysis and early warning method according to claim 1, characterized in that, The steps for detecting inflection points and locating the true starting moment of device degradation by combining characteristic trend slope abrupt change criteria and data distribution offset criteria include: Based on the fitting results of each time segment in the continuous degradation feature sequence of the equipment, the trend slope change status of the corresponding time period is statistically analyzed. When the slope fluctuation amplitude of multiple consecutive time segments exceeds the fluctuation range corresponding to the steady-state operation of the equipment, it is determined that the current time period meets the slope change condition. The system retrieves the baseline of historical normal data from the equipment and compares the distribution differences of the characteristic data for the current period. When the overall data offset exceeds the preset critical threshold, it is determined that the current period meets the distribution offset condition. The earliest time-series node that simultaneously satisfies the slope abruptness condition and the distribution offset condition is initially selected as the quasi-degenerate inflection point. A second review is conducted on the short-term window data adjacent to the quasi-degradation inflection point to determine the true starting time of substantial equipment degradation.

6. The analysis and early warning method according to claim 1, characterized in that, The steps of reading the manually marked fault time of the maintenance work order and calculating the time series offset in combination with the actual degradation start time include: Extract the fault reporting time recorded in the maintenance work order as the manual fault labeling time, and retrieve the actual degradation start time of the equipment; Calculate the time difference between the artificial fault labeling time and the actual degradation start time, and use the time difference as the initial time series offset; The offset statistics of similar deterioration cases in the equipment's historical fault file are retrieved, and the initial time offset is corrected by mean to obtain the time offset used for tag calibration.

7. The analysis and early warning method according to claim 1, characterized in that, The step of adaptively shifting and correcting the original lagging single-point fault labels based on the time offset to obtain time-aligned fault labels includes: The system classifies and matches the corresponding type of the current device fault, retrieves the preset fault lag correction coefficient, and adapts and corrects the timing offset to obtain the optimized effective offset duration. Using the time of the manual fault labeling as the reference time point, the marking time of the original single-point fault label is shifted forward and corrected in combination with the effective offset duration; The corrected tag time node is matched and verified with the device's continuous degradation feature sequence. Once the verification is successful, the time-aligned fault tag is generated.

8. The analysis and early warning method according to claim 1, characterized in that, The steps for constructing a full-deterioration cycle segmented fault interval labeling system based on the time-aligned fault labels include: Using the fault outbreak time corresponding to the time-aligned fault label as the endpoint, and combining the historical deterioration and evolution patterns of the equipment, the continuous equipment operation segments are divided backward. Based on the characteristic fluctuation amplitude and the degree of change in degradation trend, the back-traced equipment operation period is divided into a slight deterioration stage, a fault development stage, and a fault imminent outbreak stage. Independent category identifiers are configured for the minor degradation stage, the fault development stage, and the fault imminent outbreak stage, replacing the original single time point labeling method, forming a segmented fault interval labeling system that covers the equipment degradation process.

9. The analysis and early warning method according to claim 1, characterized in that, The steps for reconstructing a temporal matching dataset with feature-label alignment include: Using the full degradation cycle segmented fault interval labeling system as the labeling benchmark, match the time series feature segments corresponding to the standardized effective industrial time series dataset; Filter out effective feature data that are associated with the equipment degradation trend within the time-series feature segment interval, and remove noisy samples with disordered and mismatched feature changes; The effective time-series features that are filtered and retained are bound one-to-one with the corresponding segmented fault labels, and the data dimension and time-series dimension are aligned to obtain the time-series matching dataset.

10. The analysis and early warning method according to claim 1, characterized in that, The steps of inputting the time-series matching dataset into the artificial intelligence early warning model for iterative training include: The time-series matching dataset is divided into training samples and validation samples, and the divided sample data is fed into the artificial intelligence early warning model. The artificial intelligence early warning model updates its internal weight parameters layer by layer based on the segmented interval labeling information within the sample. During the training process, a pre-set constraint loss function is used to correct the fitting bias, weaken the interference weight of random process fluctuation characteristics, and enhance the learning ability of the equipment's early weak degradation characteristics. The parameters are iteratively updated until the model error converges, thus obtaining an early warning model for early hidden faults.

Images