Root cause alarm analysis method and system based on time sequence correlation and multi-modal fusion

Abstract

The invention discloses a root cause alarm analysis method and system based on time sequence association and multi-modal fusion, and the method comprises the steps: taking digital twin dynamic threshold cleaning as a core, fusing multi-source data time sequence alignment verification, and forming a multi-source data set of time sequence alignment and preliminary restoration through exception handling and cross-system time consistency verification; a cleaning threshold is optimized in combination with a working condition to filter noise, a label confidence weighting method is fused, confidence is given to a label through twin simulation, and a multi-modal data set with a multi-dimensional label is constructed. A learning model is constructed and trained, exclusive feature extraction branches are designed for multi-source data, a twin working condition adaptation enhancement module is introduced to strengthen core modal extraction, and the proportion of abnormal modals in fusion is reduced by means of a dynamic attention filtering mechanism. Based on the model and a twinning system, root cause penetration type positioning is completed through dynamic sequencing of root cause confidence in combination with a twinning reduction time sequence. And the root cause analysis efficiency and accuracy are improved.

Description

Technical Field

[0001] This invention belongs to the field of production line alarm analysis technology, and more specifically, relates to a root cause alarm analysis method and system based on time-series correlation and multimodal fusion. Background Technology

[0002] In modern industrial production and equipment operation and maintenance systems, root cause alarm analysis is a core component for ensuring continuous and stable system operation and reducing economic losses and safety risks caused by failures. With the advancement of industrial digital transformation, various production equipment and control systems are becoming increasingly complex. A single production line or maintenance unit often generates multi-source heterogeneous data simultaneously. This includes physical quantity data such as temperature, pressure, and current collected in real-time by sensors, as well as operational data such as control system output commands, equipment operation logs, and fault warning records. This vast amount of data, with its diverse dimensions, theoretically provides a rich basis for accurately locating the root cause of failures.

[0003] However, root cause alarm analysis in the industrial sector has long been hampered by technological bottlenecks, making it difficult to fully leverage the value of data and resulting in frequent problems such as low fault diagnosis efficiency and inaccurate root cause identification. On the one hand, traditional analysis methods rely heavily on human experience. Maintenance personnel need to rely on their familiarity with the equipment and their experience in handling past faults to determine the cause of alarms. When faced with cascading faults across modules and systems, it is difficult for humans to quickly sort out the correlation between data from different stages. This not only makes troubleshooting time-consuming but also easily leads to the omission of key factors due to limitations in experience, resulting in misjudgment of the root cause, which in turn prolongs equipment downtime and increases losses caused by production interruptions.

[0004] On the other hand, existing data processing and analysis methods are ill-suited to the demands of complex industrial scenarios. Multi-source data originates from different devices and systems, generally exhibiting inconsistent timestamps and significant format differences. The lack of effective collaborative processing mechanisms makes it difficult for data to form effective linkages over time, hindering the complete reconstruction of equipment operating trajectories before and after a failure. Simultaneously, transient sensor interference noise and invalid data generated by abnormal data acquisition equipment are difficult to remove completely using traditional fixed-threshold cleaning methods, further impacting the reliability of analysis results. Furthermore, traditional methods have weak capabilities in fusing multimodal data, often only utilizing data from a single dimension and failing to comprehensively mine the potential correlations between different data types. This results in a one-sided portrayal of equipment operating status, making it difficult to establish a deep connection between alarm events and the essence of the fault. Ultimately, it fails to achieve precise penetration from alarm symptoms to core faulty components, hindering the transformation of industrial operations and maintenance from "passive response" to "proactive prevention," and failing to meet the urgent needs of modern industrial systems for efficient and precise operations and maintenance. Summary of the Invention

[0005] This invention aims to solve the problems of high reliance on manual intervention, difficulty in coordinating multi-source data, and weak multimodal fusion in industrial root cause alarm analysis. By relying on digital twin and multimodal fusion technologies, it achieves efficient processing of multi-source data, accurate model adaptation to working conditions, and constructs a penetrating positioning link from alarm to faulty component, thereby improving the efficiency and accuracy of root cause analysis and meeting the needs of efficient industrial operation and maintenance.

[0006] To address the aforementioned deficiencies or improvement needs of existing technologies, as a first aspect of this invention, the present invention provides a root cause alarm analysis method based on temporal correlation and multimodal fusion, comprising:

[0007] S1. Based on the dynamic threshold self-optimization cleaning mechanism driven by digital twins, a multi-source data time-series alignment intelligent verification method is integrated; multi-source raw data is mapped in real time through a digital twin mapping model to identify and process collection anomalies. Simultaneously, relying on the intelligent verification method, verification rules are constructed based on the unique device code and timestamp to complete the consistency verification of cross-system data in the time dimension, forming a time-series aligned and initially repaired multi-source data set.

[0008] S2. Continuing the core mechanism driven by digital twins, the cleaning threshold is dynamically and iteratively optimized in combination with the real-time operating conditions of the equipment to filter instantaneous interference noise from the sensors; at the same time, the label confidence weighted optimization method is incorporated to assign confidence weights to manually labeled labels in combination with the simulation results of digital twins, thereby strengthening the role of highly reliable labeled data and finally constructing a multimodal dataset with multi-dimensional labels.

[0009] S3. Construct and complete the training of the learning model; among which, feature extraction branches are designed for each of the multi-source data; then, a digital twin working condition adaptation enhancement module is introduced to automatically enhance the feature extraction capability of the core modal data in the corresponding scenario based on the real-time industrial scenario output by the twin model; at the same time, based on the dynamic attention filtering mechanism, when the twin model detects that there is a transient anomaly in a certain type of modal data, it automatically reduces the influence ratio of that modality in feature fusion;

[0010] S4. Based on the trained model and digital twin system, the root cause penetration localization is completed by combining the dynamic ranking method of root cause confidence with the runtime sequence of twin reconstruction.

[0011] Furthermore, the multi-source data in S1 includes: alarm data, equipment status data, and process parameter data;

[0012] The alarm data includes information directly related to the device alarm event, including alarm ID, alarm type, trigger timestamp, alarm level, and alarm clearing time.

[0013] The equipment status data includes real-time operating parameters during equipment operation, namely, data reflecting the physical operating status of the equipment, including current, voltage, temperature, rotational speed, vibration frequency, component trajectory coordinates, and sensor detection values.

[0014] The process parameter data includes the set and actual process indicators in the production process, namely, data related to the production process, such as production cycle time, material ratio, equipment operating pressure / temperature threshold, product processing accuracy parameters, and production batch number.

[0015] Furthermore, the intelligent verification method for multi-source data temporal alignment in S1 is specifically as follows:

[0016] The device's PLC clock signal serves as the reference time axis. Original timestamps of each system Calibration via time topology mapping:

[0017]

[0018] Where T0 is the original timestamp set; For the target timestamp; To iterate through the j-th timestamp in T0; This indicates the calculation of each timestamp in T0. With target time Find the smallest of the "absolute values ​​of time differences";

[0019] For the calibrated data sequence The value of the k-th type of data in the s-th system at time t is combined with the standard time series curve output by the digital twin. Defining trend alignment:

[0020]

[0021] Where sign(·) is the sign function, which takes the values ​​1 / 0 / -1; Let be the value of the k-th type of data in the s-th system at time t-1; The source data is the k-th type of data collected by the actual device's sensors, and its value at time t. To verify the data, i.e., the k-th type of data output by the digital twin simulation, the value at time t; |T0| is the number of elements in the timestamp set T0;

[0022] It is a quantitative indicator of the consistency of trends between source data and verification data; based on this, abnormal trend data is repaired through time series manifold alignment.

[0023] Furthermore, the temporal manifold alignment repair in S1 specifically includes:

[0024]

[0025] in, The corrected target data; The original target data; For reference only; Let proj(·) be the temporal manifold formed by twin data, and proj(·) be the projection operator. Temporal alignment is achieved by integrating to compensate for the deviation between the actual data and the twin manifold. Δt is the radius of the time window.

[0026] Furthermore, the confidence-weighted optimization method in S2 is specifically as follows:

[0027] Based on typical working conditions reproduced by digital twin models, a standard label benchmark library is constructed; for any labeled category L, multi-dimensional features in its twin simulation process are extracted to form a standard feature vector:

[0028]

[0029] in, This represents the standard feature vector of the labeled category L in the twin simulation process; v Li Let be the standard value of the i-th feature in the twin simulation, k be the feature dimension, and satisfy .

[0030] For a manually labeled data tag Extract the feature vector of its actual collected data:

[0031]

[0032] in, This represents a manually labeled data tag. The feature vector corresponding to the actual collected data; A data label representing a manually labeled entry. The value of the i-th feature in the actual collected data is the actual numerical representation of the i-th dimension feature of the labeled data in the real scene, where i takes the value of 1, 2, ..., k.

[0033] The confidence level of the label is directly quantified by using the cosine similarity between real and virtual features.

[0034]

[0035] because Simplified to The value range is [0,1].

[0036] Furthermore, the digital twin operating condition adaptation enhancement module in S3 specifically comprises:

[0037] Let the key feature set of the real-time industrial scene output by the digital twin be K = {k1, k2, ..., k n}, the i-th type of modal data M i The association link matrix with scene features is L;

[0038] Defining the scene coreness of a modality:

[0039]

[0040] Among them, C(M) i ) represents the i-th type of modal data M i Scene coreness; d(M i ,k j ) is M i The link length with k is used to directly quantify the core value of the mode in the current scenario by taking the reciprocal of the shortest link length;

[0041] Let the basic feature extraction window be W0 = [t0 - Δt, t0 + Δt]:

[0042] If C(M) i If )≥1, then W i =W0∪[t0-d(M i ,k j )·Δt,t0+d(M i ,k j )·Δt];

[0043] If C(M) i If W < 1, then W i =W0;

[0044] Among them, W i M represents the i-th type of modal data i The final feature extraction window; W0 is the basic feature extraction window; t0 is the baseline time point; Δt is the basic time span;

[0045] By leveraging the inherent link between scenes and modalities, the scope of core modality extraction is adaptively expanded, enhancing feature capture capabilities.

[0046] Furthermore, the dynamic attention filtering mechanism in S3 specifically includes:

[0047] Let the digital twin model be M in each mode. i The constructed standard fluctuation baseline is The modal data acquired in real time is x i (t);

[0048] Define modal anomaly deviation:

[0049]

[0050] Among them, D i (t) represents the modal anomaly deviation of the i-th type of modal data at time t, x i (t) represents the value of the i-th modal data acquired in real time at time t; The minimum value of the standard fluctuation baseline constructed for the i-th mode M in the digital twin model. Then it is the maximum value;

[0051] The numerator is the distance between the data and the midpoint of the baseline, and the denominator is the baseline half-width. When the data is within the baseline, D... i (t) = 0, exceeding the time limit D i (t)≥1, quantifying the degree of anomaly based on the intrinsic relationship between baseline and real-time data;

[0052] Suppose the system has k modes. Under normal conditions, that is, all D... i (t) = 0, and the attention percentage for each modality is 0.

[0053] When a certain mode M i When an anomaly occurs, the dynamic attention percentage is:

[0054]

[0055] Abnormal deviation D i The larger (t) is, the more significant the compression of attention share; the released attention share is evenly distributed among other normal modalities, i.e., the normal modality share.

[0056] When digital twin detects M i T consecutive timestamps satisfy D i (t) = 0, and the data matches the twin simulation curve with a degree of Γ. i When (t)≥0.9, a reset is triggered:

[0057]

[0058] Among them, A i (t+τ) represents the proportion of the i-th mode at the τ-th time after the current time t; This represents the default percentage of the i-th modality; A represents the original proportion of the i-th mode; i (t) represents the proportion of the i-th mode at the current time t; τ is the step number, τ = 1, 2, ..., T, indicating that the default proportion is linearly restored through T steps to avoid sudden shocks.

[0059] Furthermore, in step S4, the root cause penetration localization is completed by combining the dynamic ranking method of root cause confidence with the runtime sequence of twin reduction. The specific process is as follows:

[0060] Assume the actual runtime sequence before and after the alarm in the digital twin reconstruction is as follows:

[0061]

[0062] in, Represents the actual runtime sequence dataset; t i x is the i-th timestamp; i Indicates the t-th i The fused value of multi-source data collected in real time;

[0063] The k-th potential root cause R k The corresponding twin simulation standard timing is:

[0064]

[0065] in, t represents the standard time series dataset of the twin simulation corresponding to the k-th potential root cause; i x is the i-th timestamp; ki Indicates the t-th i At time k, the standard data value of the twin simulation related to the k-th potential root cause; n is the total number of time-data pairs;

[0066] Define timing matching degree:

[0067]

[0068] Where card(·) represents the number of data points that meet the deviation condition, and the matching degree is directly quantified by the proportion of matching data points, taking values ​​[0,1]; δ(x ki ) is the adaptive deviation threshold;

[0069] Let the set of device states monitored in real time by the digital twin be S = {s1, s2, ..., s}. m}, the kth root cause R k The corresponding set of key influencing states is

[0070] Define state correlation:

[0071]

[0072] Where, β k Let k be the state correlation degree corresponding to the k-th root cause. For state s j The fault warning interval is quantified by the proportion of warning status points, with a value of [0,1].

[0073] The dynamic confidence of the k-th root cause is the time-series matching degree α. k correlation with state β k Natural coupling:

[0074] γ k =α k ·β k

[0075] Sort all potential root causes in descending order of confidence: γ(1)≥γ(2)≥...≥γ(p); take the root cause R(1) corresponding to the highest confidence γ(1), and combine it with twin time series. Construct a location link; starting from the alarm event, trace back along the actual time sequence to the time point with the smallest deviation from the optimal root cause standard time sequence, lock the fault node corresponding to the root cause, and achieve penetrating location.

[0076] As a second aspect of the present invention, the present invention provides a root cause alarm analysis system based on temporal correlation and multimodal fusion, comprising:

[0077] The twin-driven data cleaning and verification unit is designed with a dynamic threshold self-optimization cleaning mechanism driven by digital twins as its core, and incorporates a multi-source data time-series alignment intelligent verification method. It maps multi-source raw data in real time through a digital twin mapping model, identifies and processes collection anomalies, and simultaneously relies on the intelligent verification method to build verification rules based on the device's unique code and timestamp. This completes the consistency verification of cross-system data in the time dimension, forming a time-series aligned and initially repaired multi-source data set.

[0078] The multimodal dataset construction optimization unit is used to continue the core mechanism driven by digital twins. It combines the real-time operating conditions of the equipment to dynamically iterate and optimize the cleaning threshold to filter the instantaneous interference noise of the sensor. At the same time, it incorporates the label confidence weighted optimization method, which assigns confidence weights to manually labeled labels in combination with the digital twin simulation results, strengthens the role of high-confidence labeled data, and finally constructs a multimodal dataset with multi-dimensional labels.

[0079] The twin enhancement model training unit is used to build and complete the training of the learning model. It designs corresponding feature extraction branches for multi-source data. Then, a digital twin working condition adaptation enhancement module is introduced to automatically enhance the feature extraction capability of core modal data in the corresponding scenario based on the real-time industrial scenario output by the twin model. At the same time, based on the dynamic attention filtering mechanism, when the twin model detects that there is a transient anomaly in a certain type of modal data, it automatically reduces the influence ratio of that modality in feature fusion.

[0080] The root cause localization unit is used to complete root cause penetration localization by relying on the trained model and digital twin system, through the dynamic ranking method of root cause confidence combined with the runtime sequence of twin reconstruction.

[0081] As a third aspect of the invention, the invention provides a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor of any step of the root cause alarm analysis method based on temporal correlation and multimodal fusion.

[0082] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0083] 1. The root cause alarm analysis method based on temporal correlation and multimodal fusion of this invention uses a dynamic threshold self-optimization cleaning mechanism driven by digital twins as its core, and incorporates a multi-source data temporal alignment intelligent verification method to complete the preprocessing and optimization of multi-source data. Specifically, it uses a digital twin mapping model to map multi-source raw data in real time, identify and process collection anomalies, and builds verification rules based on the unique device code and timestamp to complete the consistency verification of cross-system data time dimension, forming a temporally aligned and initially repaired multi-source data set. Subsequently, it dynamically iteratively optimizes the cleaning threshold based on the real-time operating conditions of the equipment, filters instantaneous interference noise from sensors, incorporates a label confidence weighted optimization method, and assigns confidence weights to manually labeled labels based on twin simulation results, strengthening the role of high-confidence labeled data and constructing a multimodal dataset with multi-dimensional labels. This process effectively solves problems such as temporal misalignment, noise interference, and uneven labeling quality of multi-source data, providing high-quality and standardized data support for subsequent steps and ensuring the reliability of data input.

[0084] 2. The root cause alarm analysis method based on temporal correlation and multimodal fusion of this invention improves the model's ability to process industrial scenario data by constructing and training a learning model adapted to multimodal data. Dedicated feature extraction branches are designed for each source of data to fully explore the unique information of various modalities. A digital twin working condition adaptation enhancement module is introduced to automatically strengthen the feature extraction capability of core modal data in the corresponding scenario based on the real-time industrial scenario output by the twin model. Simultaneously, a dynamic attention filtering mechanism is incorporated, which automatically reduces the influence ratio of a certain modality in feature fusion when the twin model detects a transient anomaly in a certain type of data. These designs enable the model to comprehensively capture multimodal data features and dynamically adjust processing strategies according to the actual scenario, enhancing the model's adaptability and anti-interference ability to different working conditions, and providing a reliable model foundation for accurate root cause localization.

[0085] 3. The root cause alarm analysis method based on temporal correlation and multimodal fusion of the present invention achieves accurate root cause location by relying on a trained model and a digital twin system, and using dynamic ranking of root cause confidence combined with twin reconstruction of the runtime sequence. The digital twin system reconstructs the full temporal sequence of equipment operation before and after the alarm, and uses timestamps to connect multi-source data to form a time-series graph linking "time-data-equipment status". Based on historical fault data, a "preceding anomaly-alarm event-root cause" association rule base is constructed. After screening potential root causes, the root cause confidence is calculated and ranked from two dimensions: temporal matching degree and state correlation degree, constructing a location link of "alarm event → preceding anomaly → potential root cause → faulty component". This process breaks through the limitations of traditional root cause location, which relies on experience and is difficult to trace the complete link, achieving accurate penetration from alarm symptoms to the core faulty component, and improving the efficiency and accuracy of root cause location. Attached Figure Description

[0086] Figure 1 This is a flowchart of the root cause alarm analysis method based on temporal correlation and multimodal fusion according to an embodiment of the present invention;

[0087] Figure 2 This is a schematic diagram showing the main interface of the digital twin according to an embodiment of the present invention;

[0088] Figure 3 This is a schematic diagram of fault location in an embodiment of the present invention;

[0089] Figure 4 This is a system unit diagram of an embodiment of the present invention. Detailed Implementation

[0090] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0091] Example 1

[0092] Please refer to Figure 1 This embodiment 1 provides a root cause alarm analysis method based on temporal correlation and multimodal fusion, including:

[0093] S1. Based on the dynamic threshold self-optimization cleaning mechanism driven by digital twins, a multi-source data time-series alignment intelligent verification method is integrated; multi-source raw data is mapped in real time through a digital twin mapping model to identify and process collection anomalies. Simultaneously, relying on the intelligent verification method, verification rules are constructed based on the unique device code and timestamp to complete the consistency verification of cross-system data in the time dimension, forming a time-series aligned and initially repaired multi-source data set.

[0094] S2. Continuing the core mechanism driven by digital twins, the cleaning threshold is dynamically and iteratively optimized in combination with the real-time operating conditions of the equipment to filter instantaneous interference noise from the sensors; at the same time, the label confidence weighted optimization method is incorporated to assign confidence weights to manually labeled labels in combination with the simulation results of digital twins, thereby strengthening the role of highly reliable labeled data and finally constructing a multimodal dataset with multi-dimensional labels.

[0095] S3. Construct and complete the training of the learning model; among which, feature extraction branches are designed for each of the multi-source data; then, a digital twin working condition adaptation enhancement module is introduced to automatically enhance the feature extraction capability of the core modal data in the corresponding scenario based on the real-time industrial scenario output by the twin model; at the same time, based on the dynamic attention filtering mechanism, when the twin model detects that there is a transient anomaly in a certain type of modal data, it automatically reduces the influence ratio of that modality in feature fusion;

[0096] S4. Based on the trained model and digital twin system, the root cause penetration localization is completed by combining the dynamic ranking method of root cause confidence with the runtime sequence of twin reconstruction.

[0097] This embodiment 1 further elaborates on the above steps.

[0098] (1) Twin-driven data cleaning and verification

[0099] Please refer to Figure 2 In root cause alarm analysis of industrial equipment operation and maintenance, multi-source data is the core analytical basis. However, this data often comes from different systems and suffers from problems such as inconsistent timestamps, significant format differences, and the inclusion of invalid information, directly affecting the accuracy of subsequent analysis. Therefore, the preprocessing of multi-source data must first address the challenges of data collaboration and quality optimization to lay a solid foundation for accurate subsequent analysis.

[0100] The multi-source data involved in preprocessing encompasses three key types of information: alarm data, equipment status data, and process parameter data. Alarm data includes information directly related to equipment alarm events, such as alarm ID, alarm type, trigger timestamp, alarm level, and alarm clearing time. Equipment status data includes real-time operating parameters during equipment operation, reflecting the physical operating status of the equipment, including current, voltage, temperature, rotational speed, vibration frequency, component trajectory coordinates, and sensor readings. Process parameter data includes set and actual process indicators in the production process, including production cycle time, material ratio, equipment operating pressure / temperature thresholds, product processing accuracy parameters, and production batch numbers—data related to the production process. These three types of data collectively constitute a complete data profile of equipment operation and alarm events, but they require system processing to realize their value.

[0101] During processing, a dynamic threshold self-optimization cleaning mechanism driven by digital twins is used as the core, while incorporating a multi-source data time-series alignment intelligent verification method. The time-series alignment verification uses the device PLC clock signal as a unified reference time axis to perform topological mapping calibration on the original timestamps of each system, matching each original timestamp with the closest time point on the reference time axis, achieving consistency of cross-system data in the time dimension, and avoiding data association failure due to time misalignment.

[0102] After calibration, the standard time-series curve output by the digital twin is combined with the actual collected data and the simulation standard data. The trend of change (rising, falling or unchanged) at adjacent time points is compared, and the trend is statistically analyzed and normalized to obtain the trend consistency, thereby determining whether there are any anomalies in the data.

[0103] In a preferred embodiment, the multi-source data time-series alignment intelligent verification method is specifically as follows:

[0104] The device's PLC clock signal serves as the reference time axis. Original timestamps of each system Calibration via time topology mapping:

[0105]

[0106] Where T0 is the original timestamp set; For the target timestamp; To iterate through the j-th timestamp in T0; This indicates the calculation of each timestamp in T0. With target time Find the smallest of the "absolute values ​​of time differences";

[0107] For the calibrated data sequence The value of the k-th type of data in the s-th system at time t is combined with the standard time series curve output by the digital twin. Defining trend alignment:

[0108]

[0109] Where sign(·) is the sign function, which takes the values ​​1 / 0 / -1; Let be the value of the k-th type of data in the s-th system at time t-1; The source data is the k-th type of data collected by the actual device's sensors, and its value at time t. To verify the data, i.e., the k-th type of data output by the digital twin simulation, the value at time t; |T0| is the number of elements in the timestamp set T0;

[0110] It is a quantitative indicator of the consistency of trends between source data and verification data; based on this, abnormal trend data is repaired through time series manifold alignment.

[0111] For data with abnormal trends, the original data is projected onto the time-series manifold formed by digital twin data as a reference. The difference between the standard data and the projected value is calculated and integrated within a set time window. The integrated result is used to compensate for the original data, thus completing the anomaly repair.

[0112] In a preferred embodiment, the temporal manifold alignment repair specifically includes:

[0113]

[0114] in, The corrected target data; The original target data; For reference only; Let proj(·) be the temporal manifold formed by twin data, and proj(·) be the projection operator. Temporal alignment is achieved by integrating to compensate for the deviation between the actual data and the twin manifold. Δt is the radius of the time window.

[0115] This series of processes not only filtered out noise and invalid information in the data and unified the time dimension of multi-source data, but also repaired data with abnormal trends, ultimately forming a time-aligned and reliable multi-source data set. This lays a solid data foundation for accurate feature extraction and establishing the correlation between alarm events and equipment status during subsequent model training.

[0116] (2) Optimization of multimodal dataset construction

[0117] After time-series alignment and outlier data repair, the multi-source data has initially achieved temporal consistency and basic reliability. However, transient interference noise from sensors still affects the data's portrayal of the device's true state, and the reliability of manually labeled data varies due to differences in personnel experience. These issues directly constrain the accuracy of subsequent model training. Therefore, it is necessary to continue the core mechanism driven by digital twins and further optimize data quality and label reliability to lay a solid foundation for accurate feature extraction and the establishment of effective associations in subsequent models.

[0118] To address transient interference noise from sensors, this method leverages digital twins to capture real-time dynamic changes in equipment operating conditions and iteratively optimizes the cleaning threshold: when the equipment is under high load or experiencing significant fluctuations in operating conditions, the threshold is broadened to avoid mistakenly deleting valid fluctuating data; when the equipment is under stable, low load conditions, the threshold is narrowed to precisely filter out minute transient interference. This approach preserves the equipment's true operating status data while efficiently removing invalid noise, ensuring the data accurately reflects the equipment's condition.

[0119] For manually labeled data, a weighted optimization method based on label confidence is incorporated: First, various typical working conditions are reproduced using digital twins, and multi-dimensional standard features of each label category in the simulation are extracted to construct a standard label benchmark library. The standard feature vectors are then normalized to ensure balanced feature weights. Next, for each manually labeled data, the feature vector corresponding to the actual data is extracted. The label confidence is obtained by calculating the cosine similarity between this vector and the corresponding category's standard feature vector in the benchmark library—the higher the similarity, the stronger the label's credibility. Finally, weights are assigned to labels based on their confidence, strengthening the role of high-confidence labeled data in subsequent training and weakening the interference of low-confidence labels. This process achieves quality screening and weight calibration of labeled data, making the dataset's labeling information more reliable.

[0120] In a preferred embodiment, the confidence-weighted optimization method is specifically as follows:

[0121] Based on typical working conditions reproduced by digital twin models, a standard label benchmark library is constructed; for any labeled category L, multi-dimensional features in its twin simulation process are extracted to form a standard feature vector:

[0122]

[0123] in, This represents the standard feature vector of the labeled category L in the twin simulation process; v Li Let be the standard value of the i-th feature in the twin simulation, k be the feature dimension, and satisfy .

[0124] For a manually labeled data tag Extract the feature vector of its actual collected data:

[0125]

[0126] in, This represents a manually labeled data tag. The feature vector corresponding to the actual collected data; A data label representing a manually labeled entry. The value of the i-th feature in the actual collected data is the actual numerical representation of the i-th dimension feature of the labeled data in the real scene, where i takes the value of 1, 2, ..., k.

[0127] The confidence level of the label is directly quantified by using the cosine similarity between real and virtual features.

[0128]

[0129] because Simplified to The value range is [0,1].

[0130] Through the above optimizations, not only was the data quality further improved, but the credibility of the labeled information was also accurately controlled. Ultimately, a multi-dimensional labeled multimodal dataset with both real validity and label reliability was constructed, providing high-quality data support for subsequent models to accurately mine data features and establish a deep correlation between alarm events and the root causes of failures.

[0131] (3) Training of twin augmentation model

[0132] After data cleaning, optimization, and the construction of a multimodal dataset, high-quality data has been generated, providing a foundation for analysis. At this point, it is necessary to build and train a learning model to transform the data into effective features that can be used for root cause localization. This step is the core link connecting data with precise analysis and directly determines the efficiency and accuracy of subsequent root cause identification.

[0133] The model construction and training are centered on "precise extraction and dynamic adaptation," and achieve efficient processing of multimodal data through three key designs: First, dedicated feature extraction branches are designed for data from different sources such as alarms, equipment status, and process parameters to ensure that the unique information of each type of data is not obscured by uniform processing, thus preserving a complete feature foundation for subsequent fusion analysis; Second, a digital twin working condition adaptation enhancement module is introduced to allow the model to optimize the extraction strategy to fit the real-time scenario; Third, a dynamic attention filtering mechanism is incorporated to ensure that feature fusion is not interfered with by abnormal data.

[0134] The role of the digital twin operational condition adaptation enhancement module is to enable the model to focus on core data. The digital twin outputs a set of key features for real-time industrial scenarios. By calculating the shortest link length between each type of modality data and these key features, the reciprocal of this length is used to determine the scenario coreness of the modality, quantifying its importance in the current scenario. The basic feature extraction window has a fixed time range. If the modality coreness meets the standard, the extraction window will be expanded accordingly to enhance the feature capture of core data; otherwise, the basic window will be maintained to avoid wasting resources and achieve targeted feature extraction.

[0135] In a preferred embodiment, the digital twin operating condition adaptation enhancement module specifically comprises:

[0136] Let the key feature set of the real-time industrial scene output by the digital twin be K = {k1, k2, ..., k n}, the i-th type of modal data M i The association link matrix with scene features is L;

[0137] Defining the scene coreness of a modality:

[0138]

[0139] Among them, C(M) i ) represents the i-th type of modal data M i Scene coreness; d(M i ,k j ) is M i The link length with k is used to directly quantify the core value of the mode in the current scenario by taking the reciprocal of the shortest link length;

[0140] Let the basic feature extraction window be W0 = [t0 - Δt, t0 + Δt]:

[0141] If C(M) i If )≥1, then W i =W0∪[t0-d(M i ,k j )·Δt,t0+d(M i ,k j )·Δt];

[0142] If C(M) i If W < 1, then W i =W0;

[0143] Among them, W i M represents the i-th type of modal data i The final feature extraction window; W0 is the basic feature extraction window; t0 is the baseline time point; Δt is the basic time span;

[0144] By leveraging the inherent link between scenes and modalities, the scope of core modality extraction is adaptively expanded, enhancing feature capture capabilities.

[0145] A dynamic attention filtering mechanism safeguards feature fusion. The digital twin sets a standard fluctuation baseline for each modality, calculating the degree of deviation between real-time data and the baseline to determine if the data is abnormal. Under normal conditions, the attention allocation for each modality during fusion is balanced; when a modality is abnormal, its allocation decreases as the deviation increases, with the released portion being shared by normal modalities, reducing abnormal interference. When the abnormal modality returns to normal and closely matches the simulation curve, its allocation is linearly restored through multiple steps, avoiding sudden changes that could impact the model.

[0146] In a preferred embodiment, the dynamic attention filtering mechanism is specifically as follows:

[0147] Let the digital twin model be M in each mode. i The constructed standard fluctuation baseline is The modal data acquired in real time is x i (t);

[0148] Define modal anomaly deviation:

[0149]

[0150] Among them, D i (t) represents the modal anomaly deviation of the i-th type of modal data at time t, x i (t) represents the value of the i-th modal data acquired in real time at time t; The minimum value of the standard fluctuation baseline constructed for the i-th mode M in the digital twin model. Then it is the maximum value;

[0151] The numerator is the distance between the data and the midpoint of the baseline, and the denominator is the baseline half-width. When the data is within the baseline, D... i (t) = 0, exceeding the time limit D i (t)≥1, quantifying the degree of anomaly based on the intrinsic relationship between baseline and real-time data;

[0152] Suppose the system has k modes. Under normal conditions, that is, all D... i (t) = 0, and the attention percentage for each modality is 0.

[0153] When a certain mode M i When an anomaly occurs, the dynamic attention percentage is:

[0154]

[0155] Abnormal deviation D iThe larger (t) is, the more significant the compression of attention share; the released attention share is evenly distributed among other normal modalities, i.e., the normal modality share.

[0156] When digital twin detects M i T consecutive timestamps satisfy D i (t) = 0, and the data matches the twin simulation curve with a degree of Γ. i When (t)≥0.9, a reset is triggered:

[0157]

[0158] Among them, A i (t+τ) represents the proportion of the i-th mode at the τ-th time after the current time t; This represents the default percentage of the i-th modality; A represents the original proportion of the i-th mode; i (t) represents the proportion of the i-th mode at the current time t; τ is the step number, τ = 1, 2, ..., T, indicating that the default proportion is linearly restored through T steps to avoid sudden shocks.

[0159] This model building and training scheme enables the model to fully mine multimodal data features and dynamically adjust strategies according to the scenario to resist abnormal interference, significantly improving its adaptability to complex industrial scenarios and providing reliable model support for accurately locating the root cause of failure from features.

[0160] (4) Root cause localization

[0161] Please refer to Figure 3 After model building and training, it has the ability to extract effective features from multimodal data. The next step is to rely on the trained model and digital twin system to achieve penetrating localization from alarm symptoms to fault core by dynamically ranking the root cause confidence and combining the runtime sequence of twin reconstruction. This step is the key to transforming data features into actual operation and maintenance decision-making basis, directly solving the problems of ambiguous root cause localization and unclear links in traditional analysis.

[0162] The entire localization process is based on the core logic of "time sequence matching + state association" and achieves accurate localization in three steps: First, the digital twin system restores the actual operating sequence of the device before and after the alarm, forming a complete data chain containing each timestamp and the corresponding multi-source data fusion value. At the same time, it generates a corresponding simulation standard time sequence for each potential root cause as a reference benchmark for root cause matching. Then, the degree of fit between the potential root cause and the actual situation is quantified from two dimensions: time sequence matching degree and state association degree. Finally, the core root cause is locked by confidence ranking, and a complete localization link is constructed.

[0163] The digital twin's reconstruction of the actual runtime sequence fully records the trajectory of the equipment's state changes from normal operation to alarm triggering, providing a temporal basis for tracing the source of the fault. Meanwhile, the simulation standard timing sequences for each potential root cause clearly define the equipment data performance corresponding to different root causes, providing a clear reference for root cause matching. When calculating the timing matching degree, the proportion of deviations between actual data and root cause standard data within an adaptive threshold is statistically analyzed to accurately measure the temporal consistency between the two. When calculating the state correlation degree, the focus is on the key equipment state parameters corresponding to the root cause, statistically analyzing the proportion of these parameters within the fault warning range, reflecting the correlation between the root cause and the current equipment state.

[0164] The root cause dynamic confidence score is obtained by multiplying the temporal matching degree and the state correlation degree. After sorting the confidence scores in descending order, the root cause corresponding to the highest value is selected as the core root cause. Combined with the actual runtime sequence, the system traces back from the alarm event to the time point with the smallest deviation from the standard temporal sequence of the core root cause to locate the fault node and construct a penetrating positioning link of "alarm event → abnormal time point → core root cause → fault node".

[0165] In a preferred embodiment, root cause penetration localization is achieved by combining a dynamic root cause confidence ranking method with the runtime sequence of twin reconstruction. The specific process is as follows:

[0166] Assume the actual runtime sequence before and after the alarm in the digital twin reconstruction is as follows:

[0167]

[0168] in, Represents the actual runtime sequence dataset; t i x is the i-th timestamp; i Indicates the t-th i The fused value of multi-source data collected in real time;

[0169] The k-th potential root cause R k The corresponding twin simulation standard timing is:

[0170]

[0171] in, t represents the standard time series dataset of the twin simulation corresponding to the k-th potential root cause; i x is the i-th timestamp; ki Indicates the t-th i At time k, the standard data value of the twin simulation related to the k-th potential root cause; n is the total number of time-data pairs;

[0172] Define timing matching degree:

[0173]

[0174] Where card(·) represents the number of data points that meet the deviation condition, and the matching degree is directly quantified by the proportion of matching data points, taking values ​​[0,1]; δ(x ki ) is the adaptive deviation threshold;

[0175] Let the set of device states monitored in real time by the digital twin be S = {s1, s2, ..., s}. m}, the kth root cause R k The corresponding set of key influencing states is

[0176] Define state correlation:

[0177]

[0178] Where, β k Let k be the state correlation degree corresponding to the k-th root cause. For state s j The fault warning interval is quantified by the proportion of warning status points, with a value of [0,1].

[0179] The dynamic confidence of the k-th root cause is the time-series matching degree α. k correlation with state β k Natural coupling:

[0180] γ k =α k ·β k

[0181] Sort all potential root causes in descending order of confidence: γ(1)≥γ(2)≥...≥γ(p); take the root cause R(1) corresponding to the highest confidence γ(1), and combine it with twin time series. Construct a location link; starting from the alarm event, trace back along the actual time sequence to the time point with the smallest deviation from the optimal root cause standard time sequence, lock the fault node corresponding to the root cause, and achieve penetrating location.

[0182] This positioning method not only relies on digital twins to achieve a complete backtracking of the operation process, but also ensures the objectivity of root cause judgment through dual-dimensional quantification. It avoids the limitations of traditional analysis that relies on experience, greatly improves the accuracy and efficiency of root cause positioning, and provides reliable support for quickly troubleshooting and reducing equipment downtime.

[0183] Example 2

[0184] Please refer to Figure 4 This embodiment 2 also provides a root cause alarm analysis system based on temporal correlation and multimodal fusion, including:

[0185] The twin-driven data cleaning and verification unit is designed with a dynamic threshold self-optimization cleaning mechanism driven by digital twins as its core, and incorporates a multi-source data time-series alignment intelligent verification method. It maps multi-source raw data in real time through a digital twin mapping model, identifies and processes collection anomalies, and simultaneously relies on the intelligent verification method to build verification rules based on the device's unique code and timestamp. This completes the consistency verification of cross-system data in the time dimension, forming a time-series aligned and initially repaired multi-source data set.

[0186] The multimodal dataset construction optimization unit is used to continue the core mechanism driven by digital twins. It combines the real-time operating conditions of the equipment to dynamically iterate and optimize the cleaning threshold to filter the instantaneous interference noise of the sensor. At the same time, it incorporates the label confidence weighted optimization method, which assigns confidence weights to manually labeled labels in combination with the digital twin simulation results, strengthens the role of high-confidence labeled data, and finally constructs a multimodal dataset with multi-dimensional labels.

[0187] The twin enhancement model training unit is used to build and complete the training of the learning model. It designs corresponding feature extraction branches for multi-source data. Then, a digital twin working condition adaptation enhancement module is introduced to automatically enhance the feature extraction capability of core modal data in the corresponding scenario based on the real-time industrial scenario output by the twin model. At the same time, based on the dynamic attention filtering mechanism, when the twin model detects that there is a transient anomaly in a certain type of modal data, it automatically reduces the influence ratio of that modality in feature fusion.

[0188] The root cause localization unit is used to complete root cause penetration localization by relying on the trained model and digital twin system, through the dynamic ranking method of root cause confidence combined with the runtime sequence of twin reconstruction.

[0189] Example 3

[0190] This embodiment 3 also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it can implement any step of the root cause alarm analysis method based on time-series correlation and multimodal fusion.

[0191] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0192] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.

[0193] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A root cause alarm analysis method based on temporal correlation and multimodal fusion, characterized in that, include: S1. Based on a digital twin-driven dynamic threshold self-optimization cleaning mechanism, it incorporates a multi-source data time-series alignment intelligent verification method; By mapping multi-source raw data in real time through a digital twin mapping model, identifying and processing collection anomalies, and simultaneously relying on intelligent verification methods, verification rules are constructed based on the unique device code and timestamp to complete the consistency verification of cross-system data in the time dimension, forming a multi-source data set that is time-aligned and initially repaired. S2. Continuing the core mechanism driven by digital twins, the cleaning threshold is dynamically and iteratively optimized in combination with the real-time operating conditions of the equipment to filter instantaneous interference noise from the sensors; at the same time, the label confidence weighted optimization method is incorporated to assign confidence weights to manually labeled labels in combination with the simulation results of digital twins, thereby strengthening the role of highly reliable labeled data and finally constructing a multimodal dataset with multi-dimensional labels. S3. Construct and complete the training of the learning model; among which, feature extraction branches are designed for each of the multi-source data; then, a digital twin working condition adaptation enhancement module is introduced to automatically enhance the feature extraction capability of the core modal data in the corresponding scenario based on the real-time industrial scenario output by the twin model; at the same time, based on the dynamic attention filtering mechanism, when the twin model detects that there is a transient anomaly in a certain type of modal data, it automatically reduces the influence ratio of that modality in feature fusion; S4. Based on the trained model and digital twin system, the root cause penetration localization is completed by combining the dynamic ranking method of root cause confidence with the runtime sequence of twin reconstruction.

2. The root cause alarm analysis method based on temporal correlation and multimodal fusion according to claim 1, characterized in that, The multi-source data in S1 includes: alarm data, equipment status data, and process parameter data; The alarm data includes information directly related to the device alarm event, including alarm ID, alarm type, trigger timestamp, alarm level, and alarm clearing time. The equipment status data includes real-time operating parameters during equipment operation, namely, data reflecting the physical operating status of the equipment, including current, voltage, temperature, rotational speed, vibration frequency, component trajectory coordinates, and sensor detection values. The process parameter data includes the set and actual process indicators in the production process, namely, data related to the production process, such as production cycle time, material ratio, equipment operating pressure / temperature threshold, product processing accuracy parameters, and production batch number.

3. The root cause alarm analysis method based on temporal correlation and multimodal fusion according to claim 1, characterized in that, The intelligent verification method for multi-source data temporal alignment in S1 is specifically as follows: The device's PLC clock signal serves as the reference time axis. Original timestamps of each system Calibration via time topology mapping: Where T0 is the original timestamp set; For the target timestamp; To iterate through the j-th timestamp in T0; This indicates the calculation of each timestamp in T0. With target time Find the smallest of the "absolute values ​​of time differences"; For the calibrated data sequence The value of the k-th type of data in the s-th system at time t is combined with the standard time series curve output by the digital twin. Defining trend alignment: Where sign(·) is the sign function, which takes the values ​​1 / 0 / -1; Let be the value of the k-th type of data in the s-th system at time t-1; The source data is the k-th type of data collected by the actual device's sensors, and its value at time t. To verify the data, i.e., the k-th type of data output by the digital twin simulation, the value at time t; |T0| is the number of elements in the timestamp set T0; It is a quantitative indicator of the consistency of trends between source data and verification data; based on this, abnormal trend data is repaired through time series manifold alignment.

4. The root cause alarm analysis method based on temporal correlation and multimodal fusion according to claim 3, characterized in that, The temporal manifold alignment repair in S1 specifically includes: in, The corrected target data; The original target data; For reference only; Let proj(·) be the temporal manifold formed by twin data, and proj(·) be the projection operator. Temporal alignment is achieved by integrating to compensate for the deviation between the actual data and the twin manifold. Δt is the radius of the time window.

5. The root cause alarm analysis method based on temporal correlation and multimodal fusion according to claim 1, characterized in that, The confidence-weighted optimization method in S2 is specifically as follows: Based on typical working conditions reproduced by digital twin models, a standard label benchmark library is constructed; for any labeled category L, multi-dimensional features in its twin simulation process are extracted to form a standard feature vector: in, This represents the standard feature vector of the labeled category L in the twin simulation process; v Li Let be the standard value of the i-th feature in the twin simulation, k be the feature dimension, and satisfy . For a manually labeled data tag Extract the feature vector of its actual collected data: in, This represents a manually labeled data tag. The feature vector corresponding to the actual collected data; A data label representing a manually labeled entry. The value of the i-th feature in the actual collected data is the actual numerical representation of the i-th dimension feature of the labeled data in the real scene, where i takes the value of 1, 2, ..., k. The confidence level of the label is directly quantified by using the cosine similarity between real and virtual features. because Simplified to The value range is [0,1].

6. The root cause alarm analysis method based on temporal correlation and multimodal fusion according to claim 1, characterized in that, The digital twin operating condition adaptation enhancement module in S3 specifically includes: Let the key feature set of the real-time industrial scene output by the digital twin be K = {k1, k2, ..., k n }, the i-th type of modal data M i The association link matrix with scene features is L; Defining the scene coreness of a modality: Among them, C(M) i ) represents the i-th type of modal data M i Scene coreness; d(M i ,k j ) is M i The link length with k is used to directly quantify the core value of the mode in the current scenario by taking the reciprocal of the shortest link length; Let the basic feature extraction window be W0 = [t0 - Δt, t0 + Δt]: If C(M) i If )≥1, then W i =W0∪[t0-d(M i ,k j )·Δt,t0+d(M i ,k j )·Δt]; If C(M) i If W < 1, then W i =W0; Among them, W i M represents the i-th type of modal data i The final feature extraction window; W0 is the basic feature extraction window; t0 is the baseline time point; Δt is the basic time span; By leveraging the inherent link between scenes and modalities, the scope of core modality extraction is adaptively expanded, enhancing feature capture capabilities.

7. The root cause alarm analysis method based on temporal correlation and multimodal fusion according to claim 1, characterized in that, The dynamic attention filtering mechanism in S3 is specifically as follows: Let the digital twin model be M in each mode. i The constructed standard fluctuation baseline is The modal data acquired in real time is x i (t); Define modal anomaly deviation: Among them, D i (t) represents the modal anomaly deviation of the i-th type of modal data at time t, x i (t) represents the value of the i-th modal data acquired in real time at time t; The minimum value of the standard fluctuation baseline constructed for the i-th mode M in the digital twin model. Then it is the maximum value; The numerator is the distance between the data and the midpoint of the baseline, and the denominator is the baseline half-width. When the data is within the baseline, D... i (t) = 0, exceeding the time limit D i (t)≥1, quantifying the degree of anomaly based on the intrinsic relationship between baseline and real-time data; Suppose the system has k modes. Under normal conditions, that is, all D... i (t) = 0, and the attention percentage for each modality is 0. When a certain mode M i When an anomaly occurs, the dynamic attention percentage is: Abnormal deviation D i The larger (t) is, the more significant the compression of attention share; the released attention share is evenly distributed among other normal modalities, i.e., the normal modality share. When digital twin detects M i T consecutive timestamps satisfy D i (t) = 0, and the data matches the twin simulation curve with a degree of Γ. i When (t)≥0.9, a reset is triggered: Among them, A i (t+τ) represents the proportion of the i-th mode at the τ-th time after the current time t; This represents the default percentage of the i-th modality; A represents the original proportion of the i-th mode; i (t) represents the proportion of the i-th mode at the current time t; τ is the step number, τ = 1, 2, ..., T, indicating that the default proportion is linearly restored through T steps to avoid sudden shocks.

8. The root cause alarm analysis method based on temporal correlation and multimodal fusion according to claim 1, characterized in that, In S4, the root cause penetration localization is completed by combining the dynamic ranking method of root cause confidence with the runtime sequence of twin reduction. The specific process is as follows: Assume the actual runtime sequence before and after the alarm in the digital twin reconstruction is as follows: in, Represents the actual runtime sequence dataset; t i x is the i-th timestamp; i Indicates the t-th i The fused value of multi-source data collected in real time; The k-th potential root cause R k The corresponding twin simulation standard timing is: in, t represents the standard time series dataset of the twin simulation corresponding to the k-th potential root cause; i x is the i-th timestamp; ki Indicates the t-th i At time k, the standard data value of the twin simulation related to the k-th potential root cause; n is the total number of time-data pairs; Define timing matching degree: Where card(·) represents the number of data points that meet the deviation condition, and the matching degree is directly quantified by the proportion of matching data points, taking values ​​[0,1]; δ(x ki ) is the adaptive deviation threshold; Let the set of device states monitored in real time by the digital twin be S = {s1, s2, ..., s}. m }, the kth root cause R k The corresponding set of key influencing states is Define state correlation: Where, β k Let k be the state correlation degree corresponding to the k-th root cause. For state s j The fault warning interval is quantified by the proportion of warning status points, with a value of [0,1]. The dynamic confidence of the k-th root cause is the time-series matching degree α. k correlation with state β k Natural coupling: c k =a k ·b k Sort all potential root causes in descending order of confidence: γ(1)≥γ(2)≥...≥γ(p); take the root cause R(1) corresponding to the highest confidence γ(1), and combine it with twin time series. Construct a location link; starting from the alarm event, trace back along the actual time sequence to the time point with the smallest deviation from the optimal root cause standard time sequence, lock the fault node corresponding to the root cause, and achieve penetrating location.

9. A root cause alarm analysis system based on temporal correlation and multimodal fusion, characterized in that, include: The twin-driven data cleaning and verification unit is used to integrate a multi-source data time-series alignment intelligent verification method with a dynamic threshold self-optimization cleaning mechanism driven by digital twins as its core. By mapping multi-source raw data in real time through a digital twin mapping model, identifying and processing collection anomalies, and simultaneously relying on intelligent verification methods, verification rules are constructed based on the unique device code and timestamp to complete the consistency verification of cross-system data in the time dimension, forming a multi-source data set that is time-aligned and initially repaired. The multimodal dataset construction optimization unit is used to continue the core mechanism driven by digital twins. It combines the real-time operating conditions of the equipment to dynamically iterate and optimize the cleaning threshold to filter the instantaneous interference noise of the sensor. At the same time, it incorporates the label confidence weighted optimization method, which assigns confidence weights to manually labeled labels in combination with the digital twin simulation results, strengthens the role of high-confidence labeled data, and finally constructs a multimodal dataset with multi-dimensional labels. The twin enhancement model training unit is used to build and complete the training of the learning model. It designs corresponding feature extraction branches for multi-source data. Then, a digital twin working condition adaptation enhancement module is introduced to automatically enhance the feature extraction capability of core modal data in the corresponding scenario based on the real-time industrial scenario output by the twin model. At the same time, based on the dynamic attention filtering mechanism, when the twin model detects that there is a transient anomaly in a certain type of modal data, it automatically reduces the influence ratio of that modality in feature fusion. The root cause localization unit is used to complete root cause penetration localization by relying on the trained model and digital twin system, through the dynamic ranking method of root cause confidence combined with the runtime sequence of twin reconstruction.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor as described in any one of claims 1-8: the root cause alarm analysis method based on temporal correlation and multimodal fusion.

Images