Multi-source heterogeneous data-oriented self-adaptive analytical model construction method and system

Abstract

The invention relates to the technical field of multi-source heterogeneous data analysis, in particular to a multi-source heterogeneous data-oriented adaptive analysis model construction method and system, and the method comprises the steps: collecting a multi-source heterogeneous data family of a plurality of bolts, and screening alternative diagnosis data to generate a multi-modal key feature vector; judging whether the multi-modal key feature vector is abnormal or not based on the multi-modal key feature vector and the clustering feature vector; obtaining an abnormal multi-modal key feature vector, and generating a fusion feature vector based on a deep learning model; generating an initial diagnosis result based on the fused feature vector; comparing the diagnosis accuracy with a preset threshold value, and judging whether to trigger deep learning model retraining or not; and correcting the attention center of gravity based on the error feature vector and the clustering feature vector, and generating a confidence diagnosis result to position abnormal dimensions and abnormal data. According to the method, the multi-source heterogeneous data analysis efficiency and the anomaly detection accuracy in the bolt tightening process of the battery pack production line are improved.

Description

Technical Field

[0001] This invention relates to the field of multi-source heterogeneous data parsing technology, and in particular to an adaptive parsing model construction method and system for multi-source heterogeneous data. Background Technology

[0002] With the rapid expansion of the new energy vehicle industry, the quality of bolt tightening in battery packs, as core energy storage components, directly determines their sealing performance, structural strength, and safety. The industry's need for real-time monitoring and anomaly diagnosis of the bolt tightening process is increasingly urgent, and the collection of multi-source heterogeneous data from various sensors is gradually becoming a core basis for quality diagnosis. However, existing technologies lack a systematic solution for processing multi-source heterogeneous data. Either they rely solely on single-dimensional data, failing to comprehensively reflect the spatial assembly accuracy and hidden faults of the bolts; or they lack a data quality assessment system, allowing missing, invalid, or out-of-time data to directly enter the diagnostic process, resulting in large errors in diagnostic results and failing to meet the stringent quality requirements of battery packs.

[0003] Chinese Patent Publication No. CN120339899A discloses a multi-source heterogeneous data acquisition system and method for production lines. This invention discloses a multi-source heterogeneous data acquisition system for production lines, comprising: a video data preprocessing module, which extracts continuous frames from video and extracts effective features from the video frames; a time synchronization module, which manages the unified timestamps of multi-source heterogeneous data and identifies the temporal relationships between different data; a protocol parsing module, which parses data frames of communication protocols and enables effective data interaction with different devices or systems; a data fusion module, which integrates data into a unified data system model; a data storage module, which performs real-time data acquisition, processing, and distribution; and a data display module, which provides data display, dynamic filtering, and alarm functions. This invention effectively solves the problem of inconsistent data timestamps, enabling all acquired data to be compared and analyzed under a unified time benchmark, efficiently processing diverse data formats from different devices and systems, and achieving the integration and unified management of multi-source heterogeneous data.

[0004] Therefore, it is evident that the existing technology has the following problems: The lack of effective multi-source heterogeneous data parsing, fusion, and adaptive learning mechanisms during the bolt tightening process in the battery pack production line leads to inaccurate anomaly detection and an inability to accurately pinpoint the cause of anomalies. Summary of the Invention

[0005] To address this, the present invention provides an adaptive analytical model construction method and system for multi-source heterogeneous data, which overcomes the problem in the prior art that the lack of an effective multi-source heterogeneous data analysis, fusion and adaptive learning mechanism in the process of tightening bolts in battery pack production lines leads to inaccurate anomaly detection and inability to accurately locate the cause of anomalies.

[0006] To achieve the above objectives, this invention provides a method for constructing an adaptive analytical model for multi-source heterogeneous data, comprising: Step S1: Real-time collection of multi-source heterogeneous data families of several bolts based on multiple types of sensors deployed on the target production line, wherein the multi-source heterogeneous data includes mechanical dimension data, visual dimension data and acoustic dimension data. Step S2: Determine whether the data quality meets the standard based on the data quality characteristic factors of the multi-source heterogeneous data family, so as to screen candidate diagnostic data; wherein, the data quality characteristic factors include data integrity, data validity and data consistency; Step S3: Obtain the candidate diagnostic data and extract the corresponding key feature parameters to generate multimodal key feature vectors for each bolt; Step S4: Based on the comparison between the multimodal key feature vectors and clustering feature vectors of each bolt, determine the feature dissimilarity index of the multimodal key feature vectors of each bolt, so as to determine whether the multimodal key feature vectors of each bolt are abnormal. Step S5: Obtain the multimodal key feature vector of the anomaly, determine the attention center based on the clustering results of the single-dimensional data, and perform weighted fusion of the multimodal features based on the deep learning model of the multi-head attention mechanism to generate the fused feature vector of each bolt. Step S6: Determine a standard fusion feature vector based on the clustering feature vector and the number of attention heads in each dimension of the data, and generate an initial diagnostic result by comparing the fusion feature vector with the standard fusion feature vector. Step S7: Obtain the diagnostic accuracy rate within the historical period and compare it with a preset threshold to determine whether to trigger the retraining of the deep learning model. Step S8: During the retraining process, extract the multi-source heterogeneous data corresponding to the diagnostic errors, cluster them to determine the error feature vector, and correct the attention center based on the degree of deviation between the error feature vector and the clustered feature vector to regenerate the fused feature vector and generate a confidence diagnosis result. Step S9: Based on the confidence diagnosis results, locate the corresponding abnormal dimension and abnormal data in the multi-source heterogeneous data family of a single bolt.

[0007] Furthermore, in step S2, a weighted sum of the data integrity, data validity, and data consistency of the multi-source heterogeneous data family is used to determine the comprehensive quality score, and the data quality is determined to be non-compliant with the standard if the comprehensive quality score is less than a preset threshold.

[0008] Furthermore, in step S3, the key feature parameters include key feature parameters of the mechanical dimension, key feature parameters of the visual dimension, and key feature parameters of the acoustic dimension. Among them, the key mechanical dimension features include the torque when the bolt is tightened, the torque rise slope, the torque overshoot rate, and the stability index. The key visual dimension features include bolt position deviation, bolt washer compression uniformity, and bolt-to-shell clearance consistency. The key acoustic dimension features include the dominant frequency energy ratio and harmonic distortion.

[0009] Further, in step S4, the process of determining the clustering feature vector includes, The K-means clustering algorithm is used to cluster all multimodal key feature vectors collected under historical normal operating conditions, and the resulting cluster centers are determined as cluster feature vectors.

[0010] Further, in step S4, the process of comparing the multimodal key feature vectors of each bolt with the clustering feature vectors to determine the feature dissimilarity index of the multimodal key feature vectors of each bolt, in order to determine whether the multimodal key feature vectors of each bolt are abnormal, includes the following steps: The cosine similarity of the feature vectors is calculated and clustered based on the multimodal key feature vectors of each bolt, and the feature dissimilarity index is determined based on the preset interval in which the cosine similarity is located. If the feature dissimilarity index is greater than a preset threshold, the corresponding bolt multimodal key feature vector is determined to be abnormal.

[0011] Furthermore, in step S5, the multimodal key feature vectors of the anomalies are obtained, and the maximum value of the average distance between each dimension of data and the cluster center is determined as the attention centroid.

[0012] Furthermore, in step S6, the process of determining the number of attention heads for each dimension of data based on the clustering results includes: Calculate the standard deviation of each dimension based on the clustering feature vector; The number of attention heads for each dimension of data is determined based on the ratio of the average distance between each dimension of data and the cluster center to the standard deviation.

[0013] Furthermore, in step S7, if the diagnostic accuracy is lower than a preset accuracy threshold and the number of diagnostic errors exceeds a specified number, it is determined that the deep learning model needs to be retrained.

[0014] Furthermore, in step S8, the process of correcting the attention center based on the degree of deviation between the erroneous feature vector and the clustering feature vector includes: Calculate the absolute deviation between the error vector and the cluster center for each feature dimension, and normalize it using the cluster center spacing for the corresponding dimension to obtain the normalized deviation significance vector; Set a significance threshold and mark dimensions with normalization bias higher than the threshold as high-bias dimensions; The dimension with the highest deviation between the high-bias dimension and the original attention center dimension is determined as the new attention center.

[0015] On the other hand, the present invention also provides a system for constructing an adaptive analytical model for multi-source heterogeneous data, comprising: The multi-source data acquisition module is based on the real-time acquisition of multi-source heterogeneous data families of several bolts by multiple types of sensors deployed on the target production line; The quality data screening module determines whether the data quality meets the standard based on the data quality characteristic factors of the multi-source heterogeneous data family, so as to screen candidate diagnostic data. The key feature extraction module is used to obtain the candidate diagnostic data, extract the corresponding key feature parameters, and generate multimodal key feature vectors for each bolt. The anomaly vector determination module compares the multimodal key feature vectors of each bolt with the clustering feature vectors to determine the feature dissimilarity index of the multimodal key feature vectors of each bolt, so as to determine whether the multimodal key feature vectors of each bolt are abnormal. The attention fusion module is used to obtain the multimodal key feature vectors of anomalies, determine the attention center based on the clustering results of single-dimensional data, and perform weighted fusion of multimodal features based on a deep learning model with a multi-head attention mechanism to generate the fused feature vector of each bolt. The initial diagnosis module determines a standard fusion feature vector based on the clustering feature vector and the number of attention heads in each dimension of the data, and generates an initial diagnosis result by comparing the fusion feature vector with the standard fusion feature vector. The retraining trigger module is used to compare the diagnostic accuracy within the historical period with a preset threshold to determine whether to trigger the retraining of the deep learning model. The confidence diagnosis module extracts multi-source heterogeneous data corresponding to the diagnostic errors during the retraining process, performs clustering to determine the error feature vector, corrects the attention center based on the degree of deviation between the error feature vector and the clustered feature vector, regenerates the fused feature vector, and generates a confidence diagnosis result. The anomaly localization module locates the corresponding abnormal dimension and abnormal data in the multi-source heterogeneous data family of a single bolt based on the confidence diagnosis results.

[0016] Compared with existing technologies, the beneficial effects of this invention are that by constructing a full-process method of "multi-source data acquisition - quality screening - feature extraction - anomaly judgment - attention fusion - initial diagnosis - adaptive retraining - confidence diagnosis - anomaly localization", it integrates multi-dimensional data from mechanics, vision, and acoustics, solving the problem of traditional bolt tightening diagnosis relying only on single data and having a fragmented process. At the same time, by introducing an attention mechanism and adaptive retraining, the model can focus on core anomaly dimensions and dynamically adapt to changes in working conditions, ultimately achieving precision and automation from data analysis to anomaly localization. This significantly improves the diagnostic accuracy and efficiency of bolt tightening quality in battery pack production lines, reduces problems such as battery pack sealing failure and safety hazards caused by bolt anomalies, and ensures stable operation of the production line.

[0017] Furthermore, by clearly defining the key feature parameters of mechanical, visual, and acoustic dimensions, core indicators strongly correlated with bolt tightening quality are extracted, such as torque overshoot rate in the mechanical dimension, gasket compression uniformity in the visual dimension, and harmonic distortion in the acoustic dimension. This avoids the problems of parameter redundancy or missing key information in traditional feature extraction. The parameters of each dimension correspond to the three core quality aspects of "connection reliability, spatial accuracy, and latent faults". They can independently characterize anomalies in a certain dimension and collaboratively verify the overall quality status, providing a precise and comprehensive feature foundation for subsequent multimodal feature vector generation, and significantly improving the pertinence and accuracy of anomaly detection.

[0018] Furthermore, the K-means clustering algorithm is used to cluster the multimodal key feature vectors under historical normal working conditions. The cluster centers are used as cluster feature vectors. These cluster centers can objectively condense the core distribution features of historical normal data, automatically adapt to the normal fluctuations of different batches of bolts, provide a unified and reliable normal benchmark for subsequent feature dissimilarity index calculation, reduce abnormal misjudgments caused by benchmark deviation, and improve the universality and stability of the diagnostic benchmark.

[0019] Furthermore, by calculating the cosine similarity between the multimodal key feature vector and the clustering feature vector, and based on the preset interval mapping feature dissimilarity index, the difference in vector direction, i.e. the degree of deviation of feature distribution, can be accurately quantified. Combined with the dissimilarity index of interval division, the anomaly judgment is transformed from "fuzzy threshold comparison" to "quantifiable and standardized" judgment. It can effectively distinguish between normal fluctuations, i.e. low dissimilarity index, and true anomalies, i.e. high dissimilarity index, and also reduce the missed judgment and false judgment rate, thereby improving the reliability of anomaly identification.

[0020] Furthermore, by determining the maximum average distance between each dimension of data and the cluster center as the attention centroid, the problem of random attention allocation and inability to focus on core anomalies in traditional attention mechanisms is solved. The dimension with the largest average distance represents the dimension that deviates most significantly from the normal pattern in the abnormal samples. Using it as the attention centroid can guide the model to concentrate computational resources on the dimension most sensitive to anomalies, such as torque anomalies or positional deviations, avoiding the neglect of key features due to attention dispersion. This provides a clear focus direction for subsequent multimodal feature weighted fusion and improves the anomaly representation capability of the fused features.

[0021] Furthermore, the number of attention heads is determined by the ratio of the average distance between each dimension and the cluster center to the standard deviation of that dimension. This solves the problem that the traditional fixed allocation of attention heads cannot adapt to the differences in the importance of dimensions. The standard deviation reflects the normal fluctuation range of the dimension, and the ratio can comprehensively measure the degree of abnormal deviation and the tolerance of normal fluctuation, making the allocation of attention heads more targeted. For example, more attention heads are allocated to dimensions with significant deviation and small normal fluctuation, avoiding resource waste or insufficiency, strengthening the feature extraction of core abnormal dimensions, and improving the accuracy and differentiated expression ability of fused features.

[0022] Furthermore, by triggering model retraining under the dual conditions of diagnostic accuracy falling below a preset threshold and the number of errors exceeding a specified number, it is possible to accurately identify whether the model has systematic diagnostic biases, such as a continuous decline in accuracy due to tool wear or environmental changes. This ensures that the model is retrained in a timely manner when updates are needed, avoiding resource waste and maintaining the model's long-term diagnostic accuracy, thereby improving the model's adaptability and robustness to changes in production line conditions.

[0023] Furthermore, by calculating the degree of deviation between the erroneous feature vector and the clustering feature vector, high-bias dimensions are screened and the attention centroid is corrected in conjunction with the original centroid, thus solving the diagnostic error problem caused by the omission of key abnormal dimensions in the original model due to the attention centroid. This process can reverse the identification of the root cause of the missed judgment from the erroneous cases, such as the deviation anomaly in the position that was not originally focused on. The corrected attention centroid can specifically cover the erroneous dimension, enabling the subsequent retrained model to more accurately capture the core anomaly, avoid the recurrence of the same type of diagnostic error, and significantly improve the reliability and accuracy of the confidence diagnosis results.

[0024] Furthermore, by calculating the comprehensive quality score through a weighted summation of data integrity, validity, and consistency, the multi-dimensional weighted evaluation can ensure that the candidate diagnostic data is not only continuous and without missing data, but also physically reasonable and time-synchronized with multi-source data. This provides a high-quality data foundation for subsequent feature extraction and anomaly detection, reducing diagnostic errors caused by poor data quality from the source and ensuring the reliability of the overall diagnostic process. Attached Figure Description

[0025] Figure 1This is a flowchart of an adaptive analytical model construction method for multi-source heterogeneous data according to an embodiment of the present invention; Figure 2 This is a logic diagram for determining anomalies in the multimodal key feature vectors of bolts according to an embodiment of the present invention; Figure 3 A logic decision diagram for determining whether to retrain a deep learning model in an embodiment of the present invention; Figure 4 This is a structural block diagram of the adaptive analytical model construction system for multi-source heterogeneous data according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0027] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0028] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0029] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0030] Please see Figure 1 The diagram shows a flowchart of an adaptive analytical model construction method for multi-source heterogeneous data according to an embodiment of the present invention. The present invention provides an adaptive analytical model construction method for multi-source heterogeneous data, comprising: Step S1: Real-time collection of multi-source heterogeneous data families of several bolts based on multiple types of sensors deployed on the target production line, wherein the multi-source heterogeneous data includes mechanical dimension data, visual dimension data and acoustic dimension data. Step S2: Determine whether the data quality meets the standard based on the data quality characteristic factors of the multi-source heterogeneous data family, so as to screen candidate diagnostic data; wherein, the data quality characteristic factors include data integrity, data validity and data consistency; Step S3: Obtain the candidate diagnostic data and extract the corresponding key feature parameters to generate multimodal key feature vectors for each bolt; Step S4: Based on the comparison between the multimodal key feature vectors and clustering feature vectors of each bolt, determine the feature dissimilarity index of the multimodal key feature vectors of each bolt, so as to determine whether the multimodal key feature vectors of each bolt are abnormal. Step S5: Obtain the multimodal key feature vector of the anomaly, determine the attention center based on the clustering results of the single-dimensional data, and perform weighted fusion of the multimodal features based on the deep learning model of the multi-head attention mechanism to generate the fused feature vector of each bolt. Step S6: Determine a standard fusion feature vector based on the clustering feature vector and the number of attention heads in each dimension of the data, and generate an initial diagnostic result by comparing the fusion feature vector with the standard fusion feature vector. Step S7: Obtain the diagnostic accuracy rate within the historical period and compare it with a preset threshold to determine whether to trigger the retraining of the deep learning model. Step S8: During the retraining process, extract the multi-source heterogeneous data corresponding to the diagnostic errors, cluster them to determine the error feature vector, and correct the attention center based on the degree of deviation between the error feature vector and the clustered feature vector to regenerate the fused feature vector and generate a confidence diagnosis result. Step S9: Based on the confidence diagnosis results, locate the corresponding abnormal dimension and abnormal data in the multi-source heterogeneous data family of a single bolt.

[0031] In this embodiment, a sensor array is deployed at the bolt tightening station of the battery pack production line. This array includes several high-precision torque sensors for collecting mechanical data, several industrial cameras for collecting visual data, and several microphones for collecting acoustic data. All sensors are synchronized via a unified NTP time server to ensure a time error of no more than 10 milliseconds. The collected mechanical data includes real-time torque values ​​and rotation angles throughout the bolt tightening process. The collected visual data includes the center coordinates of the bolt head, the image grayscale of the bolt sealing gasket, and the gap between the lower surface of the bolt head and the upper surface of the battery pack casing. The collected acoustic data includes audio signals acquired during the tightening process.

[0032] Specifically, in step S2, a weighted sum of the data integrity, data validity, and data consistency of the multi-source heterogeneous data family is used to determine the comprehensive quality score, and the data quality is determined to be non-compliant with the standard if the comprehensive quality score is less than a preset threshold.

[0033] In this embodiment, data quality is assessed through data integrity, data validity, and data consistency.

[0034] The data integrity calculation process includes: In the mechanical dimension, the missing rate is the proportion of data points with missing values ​​and over-range values ​​in the torque and angle time series data of the entire tightening process to the total data points. Data integrity is equal to the difference between 1 and the missing rate. If the missing rate is greater than 10%, it is assigned a value of 0. In the visual dimension, the missing rate is the proportion of consecutive frames with more than 2 lost frames and blurred frames in the total number of frames in a statistical image sequence. Data integrity is equal to the difference between 1 and the missing rate. If the missing rate is greater than 8%, it is assigned the value of 0. In the acoustic dimension, the proportion of the duration of silent segments and strong interference segments in the total duration of the audio waveform is the missing rate. Data integrity is equal to the difference between 1 and the missing rate. If the missing rate is greater than 5%, it is assigned a value of 0. The average data integrity values ​​of the mechanical, visual, and acoustic dimensions are calculated to determine the data integrity of the multi-source heterogeneous data families of each bolt.

[0035] The calculation process for data validity includes: In terms of mechanics, the verification torque is between 10% and 90% of the equipment's range and the angle increases monotonically. The percentage of valid data points determines the data validity. If the percentage is less than 70%, the value is assigned to 0. In the visual dimension, the verification image grayscale is within the normal range of 80 to 200 and the bolt and gasket features are clear. The percentage of valid frames indicates data validity. If the percentage is less than 60%, the value is assigned to 0. In the acoustic dimension, the sound pressure level is verified to be 30-80dB and the frequency domain capability is concentrated in 200 to 5000Hz. The effective duration percentage is the data validity. If the percentage is less than 65%, it is assigned a value of 0. The average value of data validity in the mechanical, visual, and acoustic dimensions is calculated to determine the data validity of the multi-source heterogeneous data family for each bolt.

[0036] The calculation process for data consistency includes: Extract the timestamps of three key events: peak mechanical torque, maximum visual gasket compression, and peak acoustic friction sound. Calculate the ratio of the maximum time difference to the total tightening time. The difference between 1 and the ratio is determined as data consistency. If the time difference is greater than 100ms, it is assigned a value of 0.

[0037] The overall quality score is determined by weighting data integrity, data validity, and data consistency with a weight of 4:4:2 and then summing the results. If the overall quality score is less than 0.8, the data is considered to be of non-compliance with the standards.

[0038] After quality assessment of the multi-source heterogeneous data families of several bolts, data that meet the standards are selected as candidate diagnostic data.

[0039] Specifically, in step S3, the key feature parameters include key feature parameters of the mechanical dimension, key feature parameters of the visual dimension, and key feature parameters of the acoustic dimension. Among them, the key mechanical dimension features include the torque when the bolt is tightened, the torque rise slope, the torque overshoot rate, and the stability index. The key visual dimension features include bolt position deviation, bolt washer compression uniformity, and bolt-to-shell clearance consistency. The key acoustic dimension features include the dominant frequency energy ratio and harmonic distortion.

[0040] Understandably, the extracted key mechanical dimension features are used to characterize the reliability of bolt connections. Among them, the torque at the completion of bolt tightening reflects the basic fastening strength of the bolt; the torque rise slope reflects the smoothness during the tightening process; the torque overshoot rate reflects the risk of over-tightening; and the stability index is used to assess the ability to retain torque after final tightening.

[0041] Understandably, the extracted key visual dimension features are used to reflect the accuracy of spatial assembly. Among them, bolt position deviation reflects the bolt assembly condition; bolt gasket compression uniformity reflects the sealing and buffering effectiveness; and bolt-to-shell gap consistency reflects the tightness of the bolt-to-shell fit.

[0042] Understandably, the extracted key acoustic dimension features are used to capture latent state signals, where the dominant frequency energy ratio corresponds to the normal vibration frequency characteristics; harmonic distortion reflects the vibration regularity and can identify potential faults such as tool wear and thread impurities.

[0043] The extracted multi-dimensional key feature parameters are used to form the multi-modal key feature vectors for each bolt.

[0044] Specifically, in step S4, the process of determining the clustering feature vector includes, The K-means clustering algorithm is used to cluster all multimodal key feature vectors collected under historical normal operating conditions, and the resulting cluster centers are determined as cluster feature vectors.

[0045] In this embodiment, historical data from the battery pack bolt tightening station over the past three months are selected. Normal samples without quality feedback are identified; these are multi-source heterogeneous data sets of bolts that have passed manual re-inspection, subsequent assembly, and factory testing without issues such as loose bolts or poor sealing. Key feature parameters are extracted from each normal sample to form a key feature vector. All key feature vectors from normal samples constitute a normal feature vector set. Since K-means clustering is based on Euclidean distance calculation, it is necessary to eliminate the dimensional differences between different features. 0-1 standardization is performed on each dimension of the feature vector set. The minimum and maximum values ​​of each dimension in the normal sample set are selected. The standardized feature value of each dimension is equal to the ratio of the difference between the original feature value and the minimum value to the difference between the maximum value and the minimum value. The K-means clustering algorithm is used to cluster all standardized normal feature vector sets, and the resulting cluster centers are determined as the cluster feature vectors.

[0046] Please see Figure 2 As shown, it is a logic determination diagram for determining the anomaly of the multimodal key feature vector of a bolt in an embodiment of the present invention.

[0047] Specifically, in step S4, the process of comparing the multimodal key feature vectors of each bolt with the clustering feature vectors to determine the feature dissimilarity index of the multimodal key feature vectors of each bolt, in order to determine whether the multimodal key feature vectors of each bolt are abnormal, includes... The cosine similarity of the feature vectors is calculated and clustered based on the multimodal key feature vectors of each bolt, and the feature dissimilarity index is determined based on the preset interval in which the cosine similarity is located. If the feature dissimilarity index is greater than a preset threshold, the corresponding bolt multimodal key feature vector is determined to be abnormal.

[0048] In this embodiment, the similarity is mapped to a feature dissimilarity index according to a preset interval: The similarity is greater than or equal to 0.9, and the feature dissimilarity index is determined to be 0.1; The similarity falls within the range of [0.7, 0.9), and the feature dissimilarity index is determined to be 0.3; The similarity falls within the range of [0.5, 0.7), and the feature dissimilarity index is determined to be 0.6; If the similarity is less than 0.5, the feature dissimilarity index is determined to be 0.9.

[0049] When the feature dissimilarity index is greater than 0.5, the corresponding bolt multimodal key feature vector is determined to be abnormal.

[0050] Specifically, in step S5, the multimodal key feature vectors of the anomalies are obtained, and the maximum value of the average distance between each dimension of data and the cluster center is determined as the attention centroid.

[0051] In this embodiment, the average absolute deviation of the multimodal key feature vectors of each anomaly from the corresponding dimension of the cluster center in each dimension is calculated, and the dimension with the largest average deviation is selected as the attention centroid. The number of attention heads corresponding to the attention centroid is no less than 4.

[0052] Specifically, in step S6, the process of determining the number of attention heads for each dimension of data based on the clustering results includes: Calculate the standard deviation of each dimension based on the clustering feature vector; The number of attention heads for each dimension of data is determined based on the ratio of the average distance between each dimension of data and the cluster center to the standard deviation.

[0053] In this embodiment, the standard deviation of each dimension is calculated based on the historical normal feature vectors to reflect the natural fluctuation range of the feature dimensions. The mean absolute deviation of each abnormal feature vector and the cluster feature vector in each dimension is calculated, and the ratio of the mean absolute deviation to the standard deviation is calculated. If the ratio is greater than 2, the number of attention heads is determined to be 4; if the ratio is within the interval (1, 2), the number of attention heads is determined to be 3; if the ratio is within the interval (0.5, 1), the number of attention heads is determined to be 2; if the ratio is less than or equal to 0.5, the number of attention heads is determined to be 1.

[0054] A deep learning model based on multi-head attention mechanism performs weighted fusion of key feature vectors from anomalies across multiple modalities. When calculating attention weights, the attention centroid dimension is weighted by 50% to ensure that key features play a greater role in the fusion process.

[0055] Using clustering feature vectors as a benchmark, a standard fusion feature vector is determined based on the number of attention heads in each dimension of the data. The Euclidean distance is calculated between the fusion feature vector of each abnormal bolt and the standard fusion feature vector. This Euclidean distance is compared to a preset threshold; if it exceeds the preset threshold, the bolt is preliminarily determined to be abnormal. The preset threshold is determined by comparing the fusion feature vectors of confirmed normal bolts and confirmed abnormal bolts within the past month. The Euclidean distances between these normal samples, abnormal samples, and the standard fusion feature vector are calculated separately, and the 95th quantile of the distance for normal samples and the 5th quantile of the distance for abnormal samples are found. The midpoint between these two quantiles is taken as the preset threshold.

[0056] Please see Figure 3 As shown, it is a logic decision diagram for determining whether to retrain a deep learning model according to an embodiment of the present invention.

[0057] Specifically, in step S7, if the diagnostic accuracy is lower than a preset accuracy threshold and the number of diagnostic errors exceeds a specified number, it is determined that the deep learning model needs to be retrained.

[0058] In this embodiment, the diagnostic results of the most recent 100 bolts in the historical diagnostic records are monitored and compared with the uploaded manual re-inspection results to calculate the accuracy rate. When the accuracy rate is less than 85% and the number of errors is greater than 10, model retraining is triggered.

[0059] Specifically, in step S8, the process of correcting the attention center based on the degree of deviation between the erroneous feature vector and the clustering feature vector includes: Calculate the absolute deviation between the error vector and the cluster center for each feature dimension, and normalize it using the cluster center spacing for the corresponding dimension to obtain the normalized deviation significance vector; Set a significance threshold and mark dimensions with normalization bias higher than the threshold as high-bias dimensions; The dimension with the highest deviation between the high-bias dimension and the original attention center dimension is determined as the new attention center.

[0060] In this embodiment, the feature vectors of erroneous cases are extracted and denoted as erroneous feature vectors. The absolute deviation of each dimension from the cluster center is calculated, and the normalized deviation significance is obtained by normalization using the standard deviation of normal samples. A preset significance threshold is set, preferably 1.5. Dimensions greater than the preset threshold are identified as high-biased dimensions. The dimension with the highest deviation significance among the high-biased dimensions and the original attention centroid dimensions is determined as the new attention centroid.

[0061] The multi-head attention model is retrained using the corrected attention center, and the output fused feature vector is compared with the standard fused feature vector to generate a confidence diagnosis result. The core anomaly dimension is located, and a structured output is generated, as shown in the example below: Confidence diagnostic results 1. Basic Information: Bolt ID to be diagnosed (ST05-20240521101530), Diagnosis Time (2024-05-21 10:16:02) 2. Anomaly detection: Abnormal bolts (Euclidean distance 1.5 > threshold 1.3) 3. Core anomaly dimension: Visual dimension - bolt position deviation (contribution 60%) 4. Confidence level: 93% 5. Supporting data for abnormal situations: - The original positional deviation was 1.2 mm (standardized value 1.25), which is outside the normal range (0.2~1.0 mm, standardized value 0.1-0.8). -Related acoustic data: The main frequency energy ratio is 82% (normal range 90%~95%), affected by abnormal friction caused by positional deviation.

[0062] 6. Investigation Recommendations: - Check if the vision camera is offset (recalibrate the camera pose and confirm the accuracy of the position deviation measurement). - Check the positioning accuracy of the bolt delivery track (whether there is any track loosening that could cause the bolts to deviate from their initial position). - Test the alignment of the tightening tool (whether the tool tilt causes a positional deviation when the bolt is tightened).

[0063] Please see Figure 4 As shown, it is a structural block diagram of the adaptive analytical model construction system for multi-source heterogeneous data according to an embodiment of the present invention.

[0064] Specifically, embodiments of the present invention also provide a system for constructing an adaptive analytical model for multi-source heterogeneous data, comprising: The multi-source data acquisition module is based on the real-time acquisition of multi-source heterogeneous data families of several bolts by multiple types of sensors deployed on the target production line; A quality data screening module, which is connected to the multi-source data acquisition module, determines whether the data quality meets the standard based on the data quality characteristic factors of the multi-source heterogeneous data family, so as to screen candidate diagnostic data. A key feature extraction module, which is connected to the quality data filtering module, is used to obtain the candidate diagnostic data and extract the corresponding key feature parameters to generate a multimodal key feature vector for each bolt. An anomaly vector determination module, which is connected to the key feature extraction module, compares the multimodal key feature vectors of each bolt with the clustering feature vectors to determine the feature dissimilarity index of the multimodal key feature vectors of each bolt, so as to determine whether the multimodal key feature vectors of each bolt are abnormal. The attention fusion module, which is connected to the anomaly vector determination module, is used to obtain the multimodal key feature vectors of anomalies, determine the attention center based on the clustering results of single-dimensional data, and perform weighted fusion of multimodal features based on a deep learning model with a multi-head attention mechanism to generate the fused feature vector of each bolt. An initial diagnosis module, which is connected to the anomaly vector determination module and the attention fusion module, determines a standard fusion feature vector based on the clustering feature vector and the number of attention heads in each dimension of the data, and generates an initial diagnosis result by comparing the fusion feature vector with the standard fusion feature vector. The retraining trigger module is connected to the initial diagnosis module and is used to compare the diagnosis accuracy in the historical period with a preset threshold to determine whether to trigger the retraining of the deep learning model. The confidence diagnosis module, which is connected to the retraining trigger module and the attention fusion module, extracts multi-source heterogeneous data corresponding to the diagnostic errors during the retraining process, performs clustering to determine the error feature vector, corrects the attention center based on the degree of deviation between the error feature vector and the clustered feature vector, regenerates the fused feature vector, and generates a confidence diagnosis result. An anomaly localization module, which is connected to the confidence diagnosis module, locates the corresponding anomaly dimension and anomaly data in the multi-source heterogeneous data family of a single bolt based on the confidence diagnosis results.

[0065] It is understood that the aforementioned system for constructing an adaptive analytical model for multi-source heterogeneous data is used to execute the aforementioned method for constructing an adaptive analytical model for multi-source heterogeneous data. The apparatus and method described in the embodiments can also be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division; in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the connection may be through some interfaces, indirect coupling of devices or modules, or communication connections, and may be electrical, mechanical, or other forms.

[0066] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs. The functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The units composed of the above modules can be implemented in hardware or in a combination of hardware and software functional units.

[0067] The integrated modules implemented as software functional modules described above can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of this application.

[0068] It is understood that the module implementing the above functions can be a processor, which can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as being executed by a hardware processor, or being executed by a combination of hardware and software modules within the processor.

[0069] The storage function can be a memory, including high-speed random access memory (RAM), and also non-volatile memory (NVM), such as at least one disk storage device, or a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0070] The storage medium described above, which provides storage functionality, can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium accessible to general-purpose or special-purpose computers.

[0071] In implementation, the storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. Both the processor and storage medium can reside within an application-specific integrated circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components within an electronic device or host device.

[0072] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for constructing an adaptive analytical model for multi-source heterogeneous data, characterized in that, include: Based on the real-time acquisition of multi-source heterogeneous data of several bolts by multiple types of sensors deployed on the target production line, wherein the multi-source heterogeneous data includes mechanical dimension data, visual dimension data and acoustic dimension data. The data quality characteristics of the multi-source heterogeneous data family are used to determine whether the data quality meets the standards in order to screen candidate diagnostic data; wherein, the data quality characteristics include data integrity, data validity and data consistency. The candidate diagnostic data is obtained, and the corresponding key feature parameters are extracted to generate a multimodal key feature vector for each bolt. By comparing the multimodal key feature vectors of each bolt with the clustering feature vectors, the feature dissimilarity index of the multimodal key feature vectors of each bolt is determined, so as to determine whether the multimodal key feature vectors of each bolt are abnormal. Obtain the multimodal key feature vectors of the anomalies, determine the attention center based on the clustering results of single-dimensional data, and perform weighted fusion of multimodal features based on a deep learning model with multi-head attention mechanism to generate fused feature vectors for each bolt. A standard fusion feature vector is determined based on the clustering feature vector and the number of attention heads in each dimension of the data. An initial diagnostic result is generated by comparing the fusion feature vector with the standard fusion feature vector. The diagnostic accuracy rate within the historical period is compared with a preset threshold to determine whether to trigger retraining of the deep learning model. During the retraining process, multi-source heterogeneous data corresponding to diagnostic errors are extracted and clustered to determine error feature vectors. The attention center is corrected based on the degree of deviation between the error feature vectors and the clustered feature vectors to regenerate the fused feature vectors and generate confidence diagnostic results. Based on the confidence diagnosis results, locate the corresponding abnormal dimensions and abnormal data in the multi-source heterogeneous data family of a single bolt.

2. The adaptive analytical model construction method for multi-source heterogeneous data according to claim 1, characterized in that, The key feature parameters include key feature parameters in the mechanical dimension, key feature parameters in the visual dimension, and key feature parameters in the acoustic dimension. Among them, the key mechanical dimension features include the torque when the bolt is tightened, the torque rise slope, the torque overshoot rate, and the stability index. The key visual dimension features include bolt position deviation, bolt washer compression uniformity, and bolt-to-shell clearance consistency. The key acoustic dimension features include the dominant frequency energy ratio and harmonic distortion.

3. The adaptive analytical model construction method for multi-source heterogeneous data according to claim 2, characterized in that, The process of determining the clustering feature vectors includes, The K-means clustering algorithm is used to cluster all multimodal key feature vectors collected under historical normal operating conditions, and the resulting cluster centers are determined as cluster feature vectors.

4. The adaptive analytical model construction method for multi-source heterogeneous data according to claim 3, characterized in that, The process of comparing the multimodal key feature vectors of each bolt with the clustering feature vectors to determine the feature dissimilarity index of the multimodal key feature vectors of each bolt, and then determining whether the multimodal key feature vectors of each bolt are abnormal, includes... The cosine similarity of the feature vectors is calculated and clustered based on the multimodal key feature vectors of each bolt, and the feature dissimilarity index is determined based on the preset interval in which the cosine similarity is located. If the feature dissimilarity index is greater than a preset threshold, the corresponding bolt multimodal key feature vector is determined to be abnormal.

5. The adaptive analytical model construction method for multi-source heterogeneous data according to claim 4, characterized in that, Obtain the multimodal key feature vectors of the anomalies, and determine the attention centroid based on the maximum average distance between each dimension of data and the cluster center.

6. The adaptive analytical model construction method for multi-source heterogeneous data according to claim 5, characterized in that, The process of determining the number of attention heads for each dimension of data based on clustering results includes: Calculate the standard deviation of each dimension based on the clustering feature vector; The number of attention heads for each dimension of data is determined based on the ratio of the average distance between each dimension of data and the cluster center to the standard deviation.

7. The adaptive analytical model construction method for multi-source heterogeneous data according to claim 1, characterized in that, If the diagnostic accuracy is lower than a preset accuracy threshold and the number of diagnostic errors exceeds a specified number, the deep learning model will be retrained.

8. The adaptive analytical model construction method for multi-source heterogeneous data according to claim 7, characterized in that, The process of correcting the attention center based on the degree of deviation between the erroneous feature vector and the clustering feature vector includes, Calculate the absolute deviation between the error vector and the cluster center for each feature dimension, and normalize it using the cluster center spacing for the corresponding dimension to obtain the normalized deviation significance vector; Set a significance threshold and mark dimensions with normalization bias higher than the threshold as high-bias dimensions; The dimension with the highest deviation between the high-bias dimension and the original attention center dimension is determined as the new attention center.

9. The method for constructing an adaptive analytical model for multi-source heterogeneous data according to claim 1, characterized in that, The comprehensive quality score is determined by weighted summation of data integrity, data validity, and data consistency based on the multi-source heterogeneous data family, and the data quality is determined to be non-compliant with the standard if the comprehensive quality score is less than a preset threshold.

10. An adaptive analytical model construction system for multi-source heterogeneous data, applied to the adaptive analytical model construction method for multi-source heterogeneous data as described in any one of claims 1-9, characterized in that, The multi-source data acquisition module is based on the real-time acquisition of multi-source heterogeneous data families of several bolts by multiple types of sensors deployed on the target production line; The quality data screening module determines whether the data quality meets the standard based on the data quality characteristic factors of the multi-source heterogeneous data family, so as to screen candidate diagnostic data. The key feature extraction module is used to obtain the candidate diagnostic data, extract the corresponding key feature parameters, and generate multimodal key feature vectors for each bolt. The anomaly vector determination module compares the multimodal key feature vectors of each bolt with the clustering feature vectors to determine the feature dissimilarity index of the multimodal key feature vectors of each bolt, so as to determine whether the multimodal key feature vectors of each bolt are abnormal. The attention fusion module is used to obtain the multimodal key feature vectors of anomalies, determine the attention center based on the clustering results of single-dimensional data, and perform weighted fusion of multimodal features based on a deep learning model with a multi-head attention mechanism to generate the fused feature vector of each bolt. The initial diagnosis module determines a standard fusion feature vector based on the clustering feature vector and the number of attention heads in each dimension of the data, and generates an initial diagnosis result by comparing the fusion feature vector with the standard fusion feature vector. The retraining trigger module is used to compare the diagnostic accuracy within the historical period with a preset threshold to determine whether to trigger the retraining of the deep learning model. The confidence diagnosis module extracts multi-source heterogeneous data corresponding to the diagnostic errors during the retraining process, performs clustering to determine the error feature vector, corrects the attention center based on the degree of deviation between the error feature vector and the clustered feature vector, regenerates the fused feature vector, and generates a confidence diagnosis result. The anomaly localization module locates the corresponding abnormal dimension and abnormal data in the multi-source heterogeneous data family of a single bolt based on the confidence diagnosis results.

Images