Industrial dataset quality detection method and apparatus, electronic device, and storage medium

Abstract

The present disclosure provides an industrial dataset quality detection method and device, electronic equipment and storage medium, which generates fusion data by fusing and preprocessing the to-be-tested industrial dataset and the associated description information; identifies the modal type, industrial scene type and task target of the to-be-tested industrial dataset based on the fusion data; extracts an index set matched with the modal type and the industrial scene type from a preset index library; dynamically determines the weight of each index in the index set according to the task target; and performs joint detection of the basic quality detection and the industrial characteristic matching detection on the to-be-tested industrial dataset based on the index set and the weight, and generates a quality detection result, thereby improving the quality detection accuracy of the industrial dataset.

Description

Technical Field

[0001] This disclosure relates to data processing technology, and in particular to a method, apparatus, electronic device, and storage medium for quality inspection of industrial datasets. Background Technology

[0002] Industrial datasets are collections of data from industrial production, manufacturing, operation, and maintenance environments, typically used for the development and training of industrial models. An industrial model is an artificial intelligence model developed and trained based on an industrial dataset to perform specific industrial tasks. The quality of the industrial dataset directly affects the accuracy, stability, security, and reliability of the industrial model.

[0003] In related technologies, a unified quality inspection method is used to inspect the quality of different datasets. For example, common quality indicators (such as completeness and consistency) are calculated for datasets from different fields, and the quality inspection results of the datasets are obtained accordingly.

[0004] However, the unified quality inspection methods in related technologies cannot be adapted to the characteristics of industrial datasets themselves, resulting in the quality inspection results of industrial datasets failing to reflect the true quality of industrial datasets and insufficient accuracy in quality inspection. Summary of the Invention

[0005] This disclosure provides an industrial dataset quality inspection method, apparatus, electronic device, and storage medium.

[0006] According to a first aspect of the present disclosure, an industrial dataset quality detection method is provided, comprising: Obtain the industrial dataset to be tested and its associated descriptive information; The industrial dataset to be tested and the associated descriptive information are fused and preprocessed to generate fused data; Based on the fused data, the modal type, industrial scenario type, and task objective of the industrial dataset to be tested are identified; Extract a set of indicators that match the modality type and the industrial scenario type from a preset indicator library; Based on the stated task objectives, the weights of each indicator in the indicator set are dynamically determined; Based on the indicator set and the weights, the industrial dataset to be tested is subjected to joint detection of basic quality detection and industrial characteristic matching detection, and quality detection results are generated.

[0007] In one optional implementation, dynamically determining the weights of each indicator in the indicator set based on the task objective includes: Pre-generate a knowledge graph of task metrics; For any indicator in the indicator set, calculate the correlation coefficient between that indicator and the task objective in the task indicator knowledge graph; The weight of each indicator is determined based on the correlation coefficient between each indicator in the indicator set and the task objective.

[0008] In one optional implementation, dynamically determining the weights of each indicator in the indicator set based on the task objective includes: Based on the stated task objectives, the basic weights of each indicator in the indicator set are dynamically determined; Based on at least one weighting factor, determine the correction coefficient for at least one indicator in the indicator set; Based on the basic weights and the correction coefficients, the weights of each indicator in the indicator set are generated; The weighting influence factors include at least one of the following: The risk level corresponding to the industrial scenario type, the sample distribution characteristics of the industrial dataset to be tested, the dataset size of the industrial dataset to be tested, the historical feedback of the downstream industrial model, and the priority configuration information.

[0009] In one optional implementation, the step of jointly detecting basic quality inspection and industrial characteristic matching inspection on the industrial dataset to be tested based on the indicator set and the weights, and generating quality inspection results, includes: For any indicator in the set of indicators, determine the basic quality testing function and / or industrial characteristic matching function corresponding to the indicator, and calculate the indicator score using the corresponding basic quality testing function and / or industrial characteristic matching function; The quality inspection results are generated based on the index scores and weights of each index in the index set.

[0010] In one optional implementation, generating the quality inspection result based on the index scores and weights of each index in the index set includes: A basic score is generated based on the score and weight of each indicator in the indicator set. Obtain the high-risk defect score of the industrial dataset to be tested, and generate a risk penalty score based on the high-risk defect score and the penalty coefficient. A quality score for the industrial dataset to be tested is generated based on the base score and the risk penalty score.

[0011] In one optional implementation, generating a basic score based on the score and weight of each indicator in the indicator set includes: A basic score is generated based on the score, weight, and credibility of each indicator in the indicator set.

[0012] In one optional implementation, after generating the quality inspection results, the method further includes: Based on the quality inspection results, locate the data defects existing in the industrial dataset to be tested; Root cause analysis is performed on the data defects, and remediation suggestions are generated based on the results of the root cause analysis.

[0013] According to a second aspect of the present disclosure, an industrial dataset quality inspection device is provided, comprising: The acquisition module is used to acquire the industrial dataset to be tested and the associated descriptive information of the industrial dataset to be tested; The preprocessing module is used to perform fusion preprocessing on the industrial dataset to be tested and the associated descriptive information to generate fused data; The identification module is used to identify the modal type, industrial scenario type, and task objective of the industrial dataset to be tested based on the fused data. The indicator set construction module is used to extract an indicator set that matches the modality type and the industrial scenario type from a preset indicator library; The weight determination module is used to dynamically determine the weight of each indicator in the indicator set based on the task objective. The joint detection module is used to perform joint detection of basic quality detection and industrial characteristic matching detection on the industrial dataset to be tested based on the indicator set and the weights, and generate quality detection results.

[0014] According to a third aspect of the present disclosure, an electronic device is provided, comprising: Memory, used to store computer programs; A processor is configured to execute a computer program stored in the memory, and when the computer program is executed, to implement the above-described industrial dataset quality inspection method.

[0015] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the above-described industrial dataset quality detection method.

[0016] According to a fifth aspect of the present disclosure, a computer program product is provided, including computer program instructions that, when executed by a processor, implement the above-described industrial dataset quality detection method.

[0017] Based on the embodiments of this disclosure, by using the industrial dataset to be tested and its associated descriptive information as a basis, the modal type, industrial scenario type, and task objective of the industrial dataset to be tested are identified, and an indicator set matching the modal type and industrial scenario type is dynamically constructed, so that the indicator set is adapted to the industrial dataset to be tested, thereby improving the accuracy of quality detection. Moreover, the weight of each indicator in the indicator set is dynamically determined according to the task objective, further aligning the focus of quality detection with the actual needs of downstream tasks, avoiding the deviation of quality detection results caused by static weights. Furthermore, the joint detection adopts a combination of basic quality detection and industrial characteristic matching detection, which not only evaluates the general quality of the dataset, but also assesses whether the dataset conforms to industrial characteristics, thereby significantly improving the accuracy and industrial applicability of the industrial dataset quality detection results.

[0018] The technical solutions of this disclosure will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings, which form part of this specification, illustrate embodiments of this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0020] This disclosure will become clearer with reference to the accompanying drawings and the following detailed description, wherein: Figure 1 A flowchart illustrating an industrial dataset quality inspection method provided in an embodiment of this disclosure is shown. Figure 2 A flowchart illustrating a dynamic weight determination method provided in an embodiment of this disclosure is shown. Figure 3 A schematic diagram of a task indicator knowledge graph provided in an embodiment of this disclosure is shown; Figure 4 A schematic flowchart of a joint detection method provided in an embodiment of this disclosure is shown; Figure 5 A flowchart illustrating another industrial dataset quality inspection method provided in this disclosure is shown. Figure 6 A flowchart illustrating another industrial dataset quality inspection method provided in an embodiment of this disclosure is shown; Figure 7 A flowchart illustrating another industrial dataset quality inspection method provided in this disclosure is shown. Figure 8 A schematic diagram of the structure of an industrial dataset quality inspection device provided in an embodiment of this disclosure is shown; Figure 9 A schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure is shown. Detailed Implementation

[0021] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present disclosure.

[0022] Those skilled in the art will understand that the terms "first," "second," etc., in the embodiments of this disclosure are only used to distinguish different steps, devices, or modules, and do not represent any specific technical meaning, nor do they indicate a necessary logical order between them.

[0023] It should also be understood that in the embodiments disclosed herein, "a plurality of" may refer to two or more, and "at least one" may refer to one, two or more.

[0024] It should also be understood that any component, data or structure mentioned in the embodiments of this disclosure can generally be understood as one or more unless expressly defined or given to the contrary in the context.

[0025] Furthermore, the term "and / or" in this disclosure is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this disclosure generally indicates that the preceding and following related objects have an "or" relationship.

[0026] It should also be understood that the description of the various embodiments in this disclosure emphasizes the differences between the various embodiments, and the similarities or similarities can be referred to each other. For the sake of brevity, they will not be described in detail.

[0027] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0028] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit this disclosure or its application or use.

[0029] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0030] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0031] The embodiments disclosed herein can be applied to electronic devices such as terminal devices, computer systems, and servers, and can operate together with a wide range of other general-purpose or special-purpose computing system environments or configurations. Examples of well-known terminal devices, computing systems, environments, and / or configurations suitable for use with electronic devices such as terminal devices, computer systems, and servers include, but are not limited to: personal computer systems, server computer systems, thin clients, thick clients, handheld or laptop devices, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputer systems, mainframe computer systems, and distributed cloud computing environments including any of the above systems, etc.

[0032] Electronic devices such as terminal devices, computer systems, and servers can be described in the general context of computer system executable instructions (such as program modules) executed by a computer system. Typically, program modules can include routines, programs, object programs, components, logic, data structures, etc., which perform specific tasks or implement specific abstract data types. Computer systems / servers can be implemented in distributed cloud computing environments, where tasks are executed by remote processing devices linked through communication networks. In distributed cloud computing environments, program modules can reside on local or remote computing system storage media, including storage devices.

[0033] In related technologies, general data quality standards (such as GB / T 36344) are typically used to inspect the quality of different types of datasets. However, general data quality standards often only include common indicators such as completeness, consistency, and accuracy. Industrial datasets, on the other hand, have unique characteristics such as strong temporal correlation (e.g., industrial sensor data has strict temporal dependencies, and issues such as data breakpoints and timestamp drift can directly lead to the failure of downstream industrial models), multi-source heterogeneity (e.g., the same production process often involves data from multiple sources such as PLCs, monitoring and data acquisition systems, vision cameras, and quality inspection instruments, and the data from each source has differences in acquisition frequency, protocol differences, and alignment difficulties), scarcity of operating scenarios (e.g., industrial fault data and extreme operating condition data are scarce), and safety and compliance (e.g., industrial data often involves industrial proprietary information such as production processes and equipment operating parameters, and compliance requirements are high). Therefore, relying on general data quality standards cannot effectively assess the unique quality attributes of industrial datasets, such as temporal continuity, multi-source data alignment, process consistency, operating condition coverage, and mechanistic rationality, and it is also difficult to discover hidden defects in data that appear normal but violate process mechanisms. Therefore, the general data quality standards used in related technologies cannot accurately detect the quality of industrial datasets.

[0034] Figure 1The diagram illustrates a flowchart of an industrial dataset quality inspection method provided in this embodiment. This method can be executed in a local server, edge computing node, industrial internet platform, enterprise private cloud, or hybrid cloud environment (a combination of public and private clouds).

[0035] like Figure 1 As shown, the method specifically includes the following steps: Step S101: Obtain the associated description information of the industrial dataset to be tested and the industrial dataset to be tested.

[0036] The industrial dataset to be tested is the industrial dataset to be subjected to quality inspection. For example, the industrial dataset to be tested may include, but is not limited to, at least one of the following: industrial visual image data (such as product surface defect images, assembly inspection images, or character recognition images collected using industrial cameras or quality inspection equipment), equipment sensing time-series data (such as sequences of temperature, pressure, vibration, current, voltage, etc., collected by sensors in chronological order over time), quality inspection structured tabular datasets (such as tabular data such as product size measurements and pass / fail labels), process logs and operation text data (such as production batch records, operator logs, alarm information, maintenance work orders, etc.), voice inspection data (inspection data presented in voice format), multimodal fusion data (such as a combination of equipment status images and corresponding sensor data that simultaneously contain multiple modalities such as images, time series, and text), and / or datasets used for industrial model pre-training, fine-tuning, instruction alignment, and knowledge enhancement (such as classification labels, detection boxes, segmentation masks, time series labels, etc.).

[0037] Obtain the associated descriptive information of the industrial dataset under test. This associated descriptive information is auxiliary data that is outside the dataset itself but related to it. For example, the associated descriptive information may include, but is not limited to, at least one of the following: metadata (data describing the attributes of the industrial dataset under test, such as dataset structure information, storage format, etc.), annotation specifications (data describing the annotation rules, label system, annotation format, and annotation process of the industrial dataset under test), task description (describing the industrial tasks that the industrial dataset under test is used to perform, such as defect detection, life prediction, etc.), business scenario description (describing the industrial environment information from which the industrial dataset under test originates or is applied, such as industry type, production process, equipment type, operating conditions, risk level, etc.), equipment and process ledger (data recording the basic parameters, operating characteristics, and process flow information of industrial equipment, such as equipment model, rated power, and process parameter ranges for each process, etc.), and data acquisition log (log information recording the dataset acquisition process, such as acquisition time, acquisition frequency, and data source equipment). The data processing logs include: data collection environment parameters, data processing logs (log information recording the preprocessing, cleaning, transformation, labeling, and other processing operations performed on the dataset after collection), historical version information (information recording the differences and evolution process between different versions of the dataset, such as version number, version release time, change description, etc.), access control information (description of the dataset's access permissions, usage restrictions, authorized users and roles, such as authorized user list, operation permissions, etc.), compliance constraint information (information describing the laws, regulations, industry standards, and internal rules that the dataset must comply with during use, storage, and transmission, such as de-identification requirements, data retention period, etc.), and / or historical quality inspection feedback information (recording the quality inspection results generated during the historical quality inspection process of this dataset or similar datasets, such as historical quality inspection results, user feedback on historical quality inspection results, etc.).

[0038] In one optional implementation, this step can obtain the industrial dataset to be tested and its associated descriptive information through methods such as local file reading, database querying, data stream access, or API interface calls. The obtained data can be part or all of the industrial dataset to be tested; for example, the entire industrial dataset can be obtained, or a portion of the industrial dataset can be obtained through random sampling or stratified sampling to improve the efficiency of quality inspection of the industrial dataset.

[0039] Step S102: The industrial dataset to be tested and the associated descriptive information are fused to generate fused data.

[0040] The industrial dataset to be tested and its associated descriptive information are fused to obtain fused data. This fused data contains both the industrial dataset to be tested and the associated descriptive information, thereby improving the processing efficiency of subsequent steps.

[0041] In one optional implementation, a structured mapping approach is used to generate the fused data. Specifically, the industrial dataset to be tested and its associated descriptive information are standardized and integrated. The originally scattered, diverse, and ambiguous industrial dataset and its associated descriptive information are parsed and processed according to predefined rules, ultimately forming a structured data object with unified fields, unified semantics, and a unified format. Thus, the fused data retains the complete content of the original data while achieving multi-dimensional and structured representation, which is beneficial to improving the processing efficiency of subsequent steps.

[0042] For example, a structured data object is constructed, which contains multiple fields. These fields may include, but are not limited to, at least one of the following: data entity identifier, modality type, sampling frequency, timestamp, tag category, operating condition code, source equipment, process segment identifier, data version number, annotation personnel identifier, processing link, compliance level, risk level, and mechanism constraint association items. This maps the industrial dataset to be tested and its associated descriptive information to the various fields of the structured data object, specifically using the corresponding content as the field content of the corresponding field. Furthermore, the structured data object also contains multiple data containers, each used to store the field content of the corresponding field. For example, structured data objects also include structured data containers with data layer, metadata layer, annotation layer, document layer, process layer, knowledge layer, and feedback layer. The data layer is used to store fields related to the dataset ontology (such as data entity identifiers, data content, etc.), the metadata layer is used to store metadata field information (such as data size, etc.), the annotation layer is used to store annotation-related field information (such as label categories, annotation user identifiers, etc.), the document layer is used to store key field information extracted from unstructured or semi-structured documents (such as task descriptions, compliance document summaries, etc.), the process layer is used to store fields related to the processing history of the dataset lifecycle (such as dataset collection process, cleaning process, etc.), the knowledge layer is used to store knowledge-related fields (such as industrial mechanism knowledge, process rules, physical constraints, etc.), and the feedback layer is used to store feedback-related fields (such as historical quality inspection results, user feedback, etc.).

[0043] Step S103: Based on the fused data, identify the modal type, industrial scenario type, and task objective of the industrial dataset to be tested.

[0044] Based on the fused data, the modal type, industrial scenario type, and task objective of the industrial dataset to be tested are identified.

[0045] Modality type refers to the data format category contained in the industrial dataset under test. Modality type can include, but is not limited to, at least one of the following: image modality, video modality, time-series modality, tabular modality, text modality, audio modality, and multimodality. Specifically, image modality presents the dataset as static images, such as surface defect images; video modality presents the dataset as video, such as equipment operation status videos; time-series modality refers to the dataset being numerical sequence data sampled in chronological order, such as sequences of temperature, pressure, vibration, and current changes over time; tabular modality presents the dataset as tables, such as process parameter tables; text modality presents the dataset as text, such as operation logs and alarm messages; audio modality presents the dataset as audio, such as equipment operation audio and voice inspection records; multimodality is a dataset containing multiple single modalities (image modality, video modality, time-series modality, tabular modality, text modality, and audio modality), such as a joint set of images and corresponding sensor time-series data. In the embodiments of this disclosure, the industrial dataset under test typically corresponds to one modality type.

[0046] Industrial scenario types refer to the industrial environment information from which the industrial dataset to be tested originates or is applied. Optionally, industrial scenario types can be categorized according to at least one dimension, which may include, but is not limited to, at least one of the following: industry dimension, production process dimension, equipment type dimension, operating condition dimension, and risk level dimension. Specifically, if it is an industry dimension, industrial scenario types may include, but are not limited to, at least one of the following: automobile manufacturing, chemical industry, steel industry, power industry, mining industry, metallurgy industry, equipment manufacturing industry, energy and power industry, etc. If it is a production process dimension, industrial scenario types may include, but are not limited to, at least one of the following: stamping, welding, painting, assembly, testing, packaging, refining, reaction, etc. If it is an equipment type dimension, industrial scenario types may include, but are not limited to, at least one of the following: stamping press, CNC machine tool, fan, compressor, reactor, industrial robot, conveyor belt, etc. If it is an operating condition dimension, industrial scenario types may include, but are not limited to, at least one of the following: normal operating condition, start-up and shutdown operating condition, fault operating condition, extreme load operating condition, transitional operating condition, etc. If it is a risk level dimension, industrial scenario types may include, but are not limited to, at least one of the following: safety critical, highly compliance sensitive, general industrial application, etc. In this embodiment of the disclosure, the industrial scenario type corresponding to the industrial dataset to be tested can be one or more.

[0047] The task objective is a specific type of downstream task to be completed using the dataset. For example, the task objective may include, but is not limited to, at least one of the following: image classification, object detection, temporal classification, fault prediction, remaining life prediction, text classification, visual question answering, image generation, instruction alignment, and knowledge enhancement. In this embodiment of the disclosure, the task objective corresponding to the industrial dataset to be tested may be one or more.

[0048] This embodiment does not limit the specific identification methods for modality types, industrial scenario types, and task objectives. For example, identification can be performed through field reading, feature matching, large language models, or classifiers.

[0049] For example, modality types can be identified based on information in the data layer and metadata layer. Specifically, modality types can be identified by reading fields, such as directly reading the modality type field content to obtain the modality type of the industrial dataset to be tested; or, modality types can be identified based on modality type features (which may include, but are not limited to, at least one of the following: file extension, encoding format, number of fields, field type, sample size, number of channels, sequence length, sampling interval, text word frequency, syntactic structure, term density, image annotation type, time series label form, etc.), such as the modality type corresponding to the file extension ".jpg" being image, or, for example, if there is a feature that corresponds one-to-one between images and text, then its corresponding modality type is multimodal; or, modality types can be extracted first by reading fields or modality type features, and if extraction fails, machine learning models such as large language models can be used to determine the modality type.

[0050] For example, industrial scenario types are identified based on information from the document layer, knowledge layer, and metadata layer. Specifically, industrial scenario types can be identified by reading fields, such as directly reading the content of the industrial scenario type field to obtain the industrial scenario type of the industrial dataset under test; or, industrial scenario types can be identified through scenario knowledge graph reasoning, such as extracting scenario keywords (e.g., scenario description, equipment code, etc.) from the document layer, knowledge layer, and metadata layer. The scenario knowledge graph records the association between different keywords and different industrial scenario types. By searching the scenario knowledge graph for industrial scenario types associated with the extracted scenario keywords, the industrial scenario type of the industrial dataset under test can be obtained.

[0051] For example, task objectives can be identified based on information in the document layer and / or annotation layer. Specifically, task objectives can be identified by reading fields, such as directly reading the task description field to obtain the task objectives of the industrial dataset under test; or, task objectives can be identified through task knowledge graph reasoning, such as extracting task keywords (e.g., task configuration) from the document layer and annotation layer. The task knowledge graph can also record the association between different keywords and different task objectives. By searching for task nodes associated with the extracted task keywords in the task knowledge graph, the task objectives of the industrial dataset under test can be obtained; or, corresponding prompt words can be generated, and the information in the document layer and / or annotation layer and the prompt words can be input into a large language model to obtain the task objectives output by the large language model.

[0052] Step S104: Extract the set of indicators that match the modal type and industrial scenario type from the preset indicator library.

[0053] At least one preset indicator library is generated in advance, which contains multiple indicators for dataset quality detection.

[0054] In one optional implementation, the preset indicator library, in addition to containing various indicators, further includes indicator description information, which may include, but is not limited to, at least one of the following: indicator identifier (a unique code or name of the indicator), indicator definition (a textual description of the meaning of the indicator, such as "the proportion of non-null samples in the dataset to the total number of samples"), indicator calculation method (the quantitative calculation method of the indicator, such as missing rate = number of null values ​​ / total number of records), data unit (the unit used in the indicator result, such as percentage, second, dimensionless, etc.), applicable modality label (a list of data modality types to which the indicator is applicable), applicable scenario label (a list of industrial scenarios to which the indicator is applicable), value range (the minimum and maximum values ​​of the indicator calculation result), evaluation direction (the relationship between the indicator value and the quality of the dataset, such as the smaller the indicator value, the better the quality of the dataset), and data source (the data required for indicator calculation).

[0055] In one optional implementation, the preset indicator library is specifically a hierarchical indicator library, with each level containing different subsets of indicators or different sub-levels within that level. For example, different sub-libraries can be constructed according to the purpose of the indicators, such as a basic quality indicator library (including general quality indicators such as standardization, completeness, and accuracy), an industrial attribute indicator library (including indicators specific to industrial data such as timeliness, diversity, balance, interpretability, and traceability), a security and compliance indicator library (including indicators used to evaluate the security, legality, and compliance of datasets such as content security, privacy protection, prevention of leakage of sensitive industrial information, access control, legal and regulatory compliance, trade secret compliance, and ethical bias control), a downstream application indicator library (including indicators used to evaluate the suitability of datasets for industrial models such as model adaptability, cross-condition generalization, scalability, and deployment portability), and an industrial mechanism constraint indicator library (including indicators to verify whether datasets conform to known industrial mechanism knowledge such as consistency of physical laws, consistency of process logic, rationality of sensor association, consistency of control timing, rationality of energy balance, rationality of material balance, and legality of equipment state transitions). Alternatively, corresponding sub-libraries can be generated based on modality type or industrial scenario type. This hierarchical setting facilitates the management and maintenance of the indicator library and improves the efficiency of subsequent indicator set acquisition. Furthermore, this preset indicator library can be updated in response to indicator library update operations.

[0056] Indicators that match the modality type and industrial scenario type of the current industrial dataset to be tested are extracted from the preset indicator library. The extracted indicators constitute the indicator set of the current industrial dataset to be tested.

[0057] For example, for a visual defect detection dataset (modal type: image, industrial scenario type: quality inspection), the corresponding indicator set includes: annotation accuracy, class balance, small sample defect coverage, bounding box consistency, image sharpness, scene illumination diversity, etc.; for a predictive maintenance time series dataset (modal type: time series, industrial scenario type: equipment operation and maintenance), the corresponding indicator set includes: sampling continuity, time synchronization, fault evolution integrity, anomaly segment reliability, mechanism coupling consistency, etc.

[0058] Compared to assigning the same set of indicators to different datasets, the embodiments of this disclosure dynamically select an indicator set that matches the modality type and industrial scenario type of the industrial dataset to be tested. The extracted indicator set is highly adapted to the industrial dataset to be tested, thereby achieving differentiated and automated configuration of the indicator set.

[0059] In one alternative implementation, during the extraction of the indicator set matching the modality type and the industrial scenario type, one or more combinations of the following methods may be used, but are not limited to: Keyword matching method: The description information of indicators in the preset indicator library is matched with the modality type and / or industrial scenario type. If the description information of an indicator contains keywords related to the current modality type or industrial scenario type, the indicator is added to the indicator set.

[0060] Rule mapping method: Predefined mapping rules between modal types, industrial scene types, and indicators, such as "IF modality = image AND scene = stamping THEN, add annotation accuracy and image clarity". Then, the indicator set of the industrial dataset to be tested is determined according to at least one pre-generated mapping rule.

[0061] Similarity matching method: Generate corresponding feature vectors for the modality and scenario types of the industrial dataset to be tested, and calculate the similarity between these vectors and the corresponding indicator vectors in the indicator set. Determine the indicator set of the industrial dataset to be tested based on the similarity calculation results. For example, sort the indicators from high to low similarity and select the top N (N is an integer greater than 1) indicators to add to the indicator set of the industrial dataset to be tested.

[0062] Large Language Model (LLM) recommendation method: Generate prompt words based on the indicator library, modality type and industrial scenario type, input the prompt words into the large language model, and obtain the indicator set recommended by the large language model as the indicator set of the industrial dataset to be tested.

[0063] The fusion and determination method employs a combination of rules, knowledge graphs, and large language models. Specifically, firstly, preliminary screening is performed through rule mapping, adding the initially selected indicators to the indicator set. A knowledge graph is pre-constructed connecting modality types, industrial scenario types, and indicators, and secondary screening is conducted based on this knowledge graph. For example, if an indicator has a node edge with a specific modality type or industrial scenario type, that indicator is added to the indicator set. Further screening is performed using prompt words and the large language model, adding the selected indicators to the indicator set. After deduplicating the indicators in the indicator set, the indicator set for the industrial dataset to be tested is obtained.

[0064] In one optional implementation, multiple indicators are initially selected based on modal type and industrial scenario type. These indicators are then validated (e.g., deduplication, merging, and / or generation of derived indicators), and an indicator set is generated based on the validated indicators. Specifically, deduplication removes duplicate indicators; merging combines related indicators, such as merging temperature range and temperature gradient indicators into a temperature compliance indicator; and derived indicators are calculated and generated based on existing indicators, such as generating a sensor fusion consistency indicator based on multiple sensor-related indicators. The specific validation methods can be customized as needed.

[0065] Step S105: Dynamically determine the weight of each indicator in the indicator set according to the task objective.

[0066] Weights are used to characterize the degree of influence of an indicator on the quality inspection results during the quality inspection process. The higher the weight value, the greater the influence of the indicator on the quality inspection results; the lower the weight value, the smaller the influence of the indicator on the quality inspection results. Optionally, the weight values ​​are usually normalized to between 0 and 1, and the sum of the weights of all indicators is 1.

[0067] The weights of each indicator in the indicator set are dynamically determined based on the task objectives of the industrial dataset to be tested. Even for the same indicator set, the weights of each indicator in the indicator set may differ under different task objectives.

[0068] In one alternative implementation, during the process of dynamically determining the weights of each indicator in the indicator set, one or more of the following methods may be used, but not limited to: Weight configuration data determination method: Weight configuration data for at least one task objective is pre-generated. For example, if the weight configuration data records the specified weights of at least one indicator under each task objective, then the weights of each indicator are generated based on the weight configuration data. For instance, if the weight configuration data records that the labeling accuracy indicator has a weight of 0.45 and the class balance indicator has a weight of 0.25 under the defect detection task objective, then the weight of the labeling accuracy indicator in the indicator set is determined to be 0.45, the weight of the class balance indicator is determined to be 0.25, and the remaining indicators in the indicator set are evenly distributed with the remaining weight of 0.3. This weight configuration data can be determined based on expert experience, the analytic hierarchy process (AHP), inferring indicator weights based on historical model training results, determining weights based on Bayesian updates, or determining weights based on reinforcement learning, among other methods. For example, in the process of back-deriving index weights based on historical model training results, multiple historical datasets can be obtained, recording the scores of each index in the historical datasets and the model performance (such as F1 score) of downstream industrial models. Using the index scores as independent variables and model performance as the dependent variable, a regression model (such as linear regression) can be constructed to obtain the regression coefficients of each index. These regression coefficients are then normalized to obtain the corresponding index weights. Another example is in the Bayesian update process, where an initial set of weights can be pre-defined as a prior distribution. When a new dataset is obtained, the index scores of each index on that dataset and the performance feedback from downstream industrial models are used to update the posterior distribution of the weights using the Bayesian formula, thereby continuously updating the weights. In the reinforcement learning-based weight determination process, the current weight is the state, fine-tuning the weights is the action, and changes in the performance of downstream industrial models are the reward. The weight model is trained using a policy gradient algorithm (e.g., an Actor-Critic model architecture can be used, including an input layer, an output layer, and a fully connected layer; the input layer receives the state feature vector of the dataset, and the output layer outputs the value estimate of the current state), thereby continuously optimizing the weights. This method of determining weight configuration data can record the weights of important indicators under the task objectives in the weight configuration data, ensuring that the weights of important indicators meet expectations.

[0069] Knowledge graph determination method: Reference Figure 2 The weights can be dynamically determined using the following steps S1051-S1053: S1051, pre-generate task indicator knowledge graph.

[0070] The task indicator knowledge graph contains two types of graph nodes: task nodes and indicator nodes. Task nodes correspond one-to-one with task objectives, and indicator nodes correspond one-to-one with indicators. An edge is established between any two related graph nodes. This edge can be an edge without a weight or an edge with a weight. The weight describes the degree of association between the two graph nodes; the higher the degree of association between the two graph nodes, the higher the weight of the edge between them.

[0071] For example, such as Figure 3 As shown, the task indicator knowledge graph contains the following nodes: Task 1, Task 2, Indicator A, Indicator B, Indicator C, and Indicator D. Task 1 has nodes connected to Indicator A and Indicator B, Task 2 has nodes connected to Indicator C and Indicator D, and Indicator B has a node connected to Indicator C.

[0072] S1052, for any indicator in the indicator set, calculate the correlation coefficient between that indicator and the task objective in the task indicator knowledge graph.

[0073] The correlation coefficient is used to quantify the degree of association between the indicator node corresponding to the metric and the task node corresponding to the task objective in the task indicator knowledge graph. The correlation coefficient can be, but is not limited to, a combination of one or more of the following methods: Direct association method: If there is a direct node edge between the task node and the indicator node, the important weight carried by that node edge is used as the association coefficient; if the node edge does not carry an important weight, the association coefficient takes the default value (for example, the default value can be 1). If there is no direct node edge between the task node and the indicator node, the association coefficient takes the set minimum association coefficient (for example, the minimum association coefficient can be 0.05).

[0074] Path accumulation method: If the node edges have significant weights, the correlation coefficient is obtained by multiplying the significant weights of all node edges between the task node and the indicator node; if there are no significant weights, the correlation coefficient is obtained by the reciprocal of the number of node edges between the task node and the indicator node. For example... Figure 3 In the graph, if the importance weight of the node edge between Task 1 and Indicator B is 0.7, and the importance weight of the node edge between Indicator B and Indicator C is 0.5, then the correlation coefficient between Task 1 and Indicator C is 0.7 × 0.5. If the node edges in the graph have no importance weight, and the path between Task 1 and Indicator C contains two node edges, then the correlation coefficient between Task 1 and Indicator C is 1 / 2; if the path between Task 1 and Indicator B contains one node edge, then the correlation coefficient between Task 1 and Indicator B is 1 / 1. If there is no reachable path between a task node and an indicator node in the graph, then the correlation coefficient is the minimum set correlation coefficient.

[0075] S1053, determine the weight of each indicator based on the correlation coefficient between each indicator in the indicator set and the task objective.

[0076] The weights of each indicator are positively correlated with their corresponding correlation coefficients. For example, the correlation coefficients of each indicator can be normalized, and the normalized values ​​can be used as the weights of the indicators.

[0077] By using a task indicator knowledge graph, we can accurately determine the degree of correlation between each indicator in the indicator set and the current task objective, so that the weight of the indicator is positively correlated with the degree of correlation, which is beneficial to improving the quality detection accuracy of industrial datasets.

[0078] Step S106: Based on the indicator set and the weight of each indicator, perform joint detection of basic quality detection and industrial characteristic matching detection on the industrial dataset to be tested, and generate quality detection results.

[0079] Basic quality inspection is the quantitative analysis of the basic quality attributes of the industrial dataset to be tested. This basic quality inspection may include, but is not limited to, at least one of the following: basic rule inspection (such as detecting missing values, outliers, field formats, unit specifications, label ranges, timestamp order, file naming rules, etc. through preset rules) and statistical analysis inspection (such as detecting completeness, duplication rate, consistency, etc.).

[0080] Industrial characteristic matching detection involves matching the dataset to industrial characteristics to determine the degree of match between the dataset and the industrial characteristics. These industrial characteristics may include, but are not limited to, at least one of the following: industrial mechanism characteristics (such as physical laws, process logic, energy balance, etc.), industrial time-series characteristics (such as equipment state transition time series), correlation and consistency characteristics (such as consistency between multi-source equipment data, label and event data, and process steps and parameter changes), safety and compliance characteristics (such as compliance constraints on sensitive fields, de-identification strategies, and permission identifiers), and downstream industrial model application characteristics (such as the degree to which the dataset supports the performance of downstream industrial models).

[0081] Joint testing involves performing basic quality testing and industrial characteristic matching testing on the same set of indicators, and then combining the results of the two types of testing to obtain the final quality test result.

[0082] In one alternative implementation, specifically, it can be adopted Figure 4 The steps S1061-S1063 shown are for joint detection: S1061, for any indicator in the indicator set, determine the basic quality testing function and / or industrial characteristic matching function corresponding to that indicator.

[0083] The calculation function for each indicator in the indicator set is determined. This function can be a basic quality detection function (a predefined program or mathematical expression used to quantify the basic quality attributes of the industrial dataset under test), an industrial characteristic matching function (used to verify whether the industrial dataset under test conforms to predefined constraint rules or calculation logic of preset industrial characteristics), or a combination of the basic quality detection function and the industrial characteristic matching function. The calculation function corresponding to an indicator can be obtained by looking up a preset indicator function mapping table or by obtaining it from the indicator's descriptive information.

[0084] S1062, calculate the index score of the indicator using the corresponding basic quality inspection function and / or industrial characteristic matching function.

[0085] For each indicator, using the industrial dataset to be tested as the data basis, the industrial dataset to be tested is automatically processed through the built-in processing logic of the indicator's calculation function (such as calculation formula, constraint verification rules, graph inference rules, etc.) to output a quantitative indicator score.

[0086] For example, if the indicator is a completeness indicator, its corresponding calculation function is the completeness calculation formula in the basic quality inspection function. The process of obtaining the indicator score is as follows: read the field null value statistics of the industrial dataset to be tested, calculate the proportion of the number of null values ​​to the total number of data units to obtain the missing rate, and then obtain the completeness rate based on the missing rate. The completeness rate is converted into a value between 0 and 100, thereby outputting the indicator score of the completeness indicator.

[0087] If the indicator is an energy balance indicator, its corresponding calculation function is the energy balance verification function in the industrial characteristic matching function. The process of obtaining the indicator score is as follows: extract the input power, output power, and internal energy change within the corresponding time window from the industrial dataset to be tested, calculate the energy deviation rate = |input power - output power - internal energy change| / input energy, and use a preset mapping function to map the energy deviation rate to a value between 0 and 100, thereby outputting the indicator score of the energy balance indicator.

[0088] If the indicator is a temperature rationality indicator, its corresponding calculation function includes both a basic quality detection function and an industrial characteristic matching function. The process for obtaining the indicator score is as follows: The temperature range detection function in the basic quality detection function is called to calculate the proportion of samples in the industrial dataset whose temperature falls within a preset temperature range, thus obtaining the first indicator score; the temperature gradient detection function in the industrial characteristic matching function is called to calculate the proportion of samples in the industrial dataset whose temperature change gradient falls within a preset temperature gradient range, thus obtaining the second indicator score; the first and second indicator scores are then fused according to preset fusion weights to obtain the final indicator score.

[0089] S1063 generates quality inspection results based on the indicator scores and weights of each indicator in the indicator set.

[0090] Specifically, the scores of the indicators in the indicator set are weighted and summed, and the quality inspection result is obtained based on the weighted summation result. The quality inspection result specifically includes the quality score of the industrial dataset to be tested.

[0091] For example, the quality score of the industrial dataset to be tested can be calculated using the following formula (1): Formula (1)

[0092] in, Indicates quality score; Indicates the number of indicators in the indicator set; Indicates the first The score of each indicator; Indicates the first The weight of each indicator.

[0093] Therefore, the industrial dataset quality inspection method provided in this disclosure, based on the industrial dataset to be tested and its associated descriptive information, identifies the modality type, industrial scenario type, and task objective of the industrial dataset to be tested, dynamically constructs an indicator set that matches the modality type and industrial scenario type, and adapts the indicator set to the industrial dataset to be tested, thereby improving the accuracy of quality inspection. Moreover, the weight of each indicator in the indicator set is dynamically determined according to the task objective, further aligning the focus of quality inspection with the actual needs of downstream tasks, avoiding the deviation in quality inspection results caused by static weights. Furthermore, the joint inspection adopts a combination of basic quality inspection and industrial characteristic matching inspection, which not only evaluates the general quality attributes of the dataset, but also verifies whether the dataset conforms to industrial characteristics, significantly improving the accuracy and industrial applicability of the industrial dataset quality inspection results.

[0094] Figure 5 A flowchart illustrating another industrial dataset quality inspection method provided in this disclosure is shown. Specifically, as... Figure 5 As shown, the method includes the following steps: Step S501: Construct an index set that matches the modal type and industrial scenario type of the industrial dataset to be tested.

[0095] The dynamic construction process of the indicator set can be referred to the description in other embodiments, and will not be repeated here.

[0096] Step S502: Based on the task objective, dynamically determine the basic weights of each indicator in the indicator set.

[0097] The indicator weights determined based on the task objectives are called basic weights. The method for determining these basic weights can be referred to in other embodiments (such as in step S105).

[0098] Step S503: Determine the correction coefficient of at least one indicator in the indicator set based on at least one weighting influence factor.

[0099] Based on the generated basic weights, additional weight influence factors are introduced to fine-tune the weights. Specifically, each weight influence factor has a corresponding range of influence, which describes which indicators' basic weights are affected by that weight influence factor. Furthermore, a mapping relationship between different values ​​of the weight influence factors and their influence coefficients (referred to as an influence coefficient table) is generated; that is, the influence coefficient table records the influence coefficients corresponding to different values ​​of the weight influence factors.

[0100] In the specific implementation process, for each weighted influence factor, the current value of the weighted influence factor is determined. By consulting the influence coefficient table and influence range of the weighted influence factor, the influence coefficient of the weighted influence factor on each indicator within that influence range is determined. If an indicator in a certain indicator set is not within the influence range of the weighted influence factor, the influence coefficient of the weighted influence factor on that indicator is set to the default value (e.g., the default value can be 1).

[0101] The weighting factors include at least one of the following: the risk level corresponding to the industrial scenario type (such as high risk, low risk, safe, etc.), the sample distribution characteristics of the industrial dataset to be tested (such as sample category balance, working condition coverage), the dataset size of the industrial dataset to be tested (such as the total number of samples), the historical feedback of the downstream industrial model (such as the correlation coefficient between the performance of the downstream industrial model and the indicators), and priority configuration (such as configuring the priority of different indicators).

[0102] For example, if the risk level corresponding to the industrial scenario type is high risk, then the safety compliance indicator and the labeling accuracy indicator are assigned an influence coefficient greater than 1 (e.g., 1.2~1.5), while the influence coefficient of other indicators not within the influence range of the risk level is the default value of 1; if the coverage of a certain working condition in the industrial dataset to be tested is low, then the indicators associated with that working condition are assigned an influence coefficient greater than 1 (i.e., a gain is applied to long-tail samples with low coverage); if the dataset size of the industrial dataset to be tested is small, then the integrity indicator and the like are assigned a correction coefficient greater than 1; if the historical record shows that the downstream industrial model has a high correlation with the time series continuity indicator, then the time series continuity indicator is assigned a correction coefficient greater than 1 (e.g., 1.3); if the priority of compliance indicators is higher than that of efficiency priority indicators in the priority configuration information, then the compliance indicators are assigned a higher influence coefficient, and the efficiency indicators are assigned a lower influence coefficient.

[0103] After obtaining the influence coefficients of each weighted influencing factor on each indicator in the indicator set, an influence coefficient matrix with different dimensions is constructed. In the influence coefficient matrix, each column corresponds one-to-one with each indicator in the indicator set, and each row corresponds one-to-one with each weighted influencing factor. For any indicator in the indicator set, the correction coefficient of the indicator is obtained by multiplying the influence coefficients of each weighted influencing factor on that indicator.

[0104] Step S504: Generate the weights of each indicator in the indicator set based on the basic weights and correction coefficients.

[0105] For any indicator in the indicator set, calculate the product of its base weight and correction coefficient, and use this product as the candidate weight for that indicator. Normalize the candidate weights of all indicators in the indicator set; the normalized values ​​are the weights of the indicators.

[0106] Step S505: Based on the indicator set and weights, perform joint detection of basic quality detection and industrial characteristic matching detection on the industrial dataset to be tested, and generate quality detection results.

[0107] Therefore, the industrial dataset quality detection method provided in this disclosure, based on the determination of basic weights according to the task objectives, further introduces weight influence factors of different dimensions. By applying the influence factors to the indicators within the corresponding influence range, the correction coefficients of the indicators are obtained. The basic weights are dynamically adjusted using the correction coefficients, so that the final indicator weights are more in line with the characteristics of the actual dataset and the application scenario requirements, significantly improving the refinement of weight allocation and the accuracy of quality detection results.

[0108] Figure 6 A flowchart illustrating another industrial dataset quality inspection method provided in this disclosure is shown. Specifically, as... Figure 6 As shown, the method includes the following steps: Step S601: Construct an index set that matches the modal type and industrial scenario type of the industrial dataset to be tested.

[0109] Step S602: Determine the weight and reliability of each indicator in the indicator set.

[0110] The dynamic construction process of the indicator set and the dynamic determination process of the weights can be referred to the description in other embodiments, and will not be repeated here.

[0111] Indicator credibility is a quantitative value representing the reliability or confidence level of an indicator, typically a value greater than 0 and less than or equal to 1. A higher indicator credibility value indicates a more reliable indicator score.

[0112] In one alternative implementation, the credibility of the indicator can be determined by a combination of one or more of the following methods: credibility configuration data, evaluation sample coverage, consistency of multi-source evidence, stability of quality detection methods, and historical calibration errors. The evaluation sample coverage can be calculated in real time, while the credibility configuration data, consistency of multi-source evidence, stability of quality detection methods, and historical calibration errors can be pre-generated.

[0113] Credibility configuration data refers to the credibility level configured by the user through the corresponding entry point. For example, the user can determine the credibility level of each indicator based on expert experience, thereby generating credibility configuration data.

[0114] The evaluation sample coverage rate is the ratio of the number of samples used in the calculation of this indicator to the total number of samples in the industrial dataset to be tested. The reliability of the indicator is positively correlated with the evaluation sample coverage rate. If the indicator is calculated based on the full industrial dataset to be tested, the evaluation sample coverage rate is 100%, and the reliability of the indicator can be 1. If the indicator is calculated by extracting 50% of the industrial dataset to be tested, the evaluation sample coverage rate is 50%, and the reliability of the indicator can be 0.8.

[0115] Multi-source evidence consistency refers to the quantitative value of consistency in the scores of the same indicator obtained by using multiple quality testing methods (such as basic rule testing, statistical analysis testing, industrial mechanism matching testing, compliance testing, etc.). The reliability of the indicator is positively correlated with this consistency score. That is, the higher the consistency score, the smaller the difference in scores of the same indicator among multiple quality testing methods, the more mutually corroborating the multi-source evidence, and the higher the reliability of the indicator evaluation result.

[0116] The stability of a quality testing method refers to the quantified value of the fluctuation in the score of an indicator when the same indicator is repeatedly tested. The reliability of an indicator is positively correlated with the stability of its quality testing method; that is, the higher the stability of a quality testing method, the less sensitive it is to random factors, and the more reliable the results.

[0117] Historical calibration error is a quantified value of the error in the quality scores of an indicator across multiple historical datasets. For example, the calibration error can be obtained by comparing historically obtained indicator scores with standard scores set by expert experience. The reliability of an indicator is negatively correlated with this historical calibration error; that is, the larger the historical calibration error, the lower the reliability of the indicator.

[0118] Step S603: Generate a basic score based on the score, weight, and credibility of each indicator in the indicator set.

[0119] For each indicator in the indicator set, calculate the product of the indicator score, weight, and indicator credibility. Use this product as the effective indicator score for that indicator. The basic score of the industrial dataset to be tested is obtained by summing the effective indicator scores of all indicators. For example, the basic score can be calculated using the following formula (2): Formula (2)

[0120] in, Indicates the basic score; Indicates the number of indicators in the indicator set; Indicates the first The weight of each indicator; Indicates the first The score of each indicator; Indicates the first The credibility of each indicator.

[0121] In an alternative implementation, a basic score can also be generated directly based on the score and weight of each indicator in the indicator set, thereby simplifying the basic score generation process.

[0122] Step S604: Obtain the high-risk defect score of the industrial dataset to be tested, and generate a risk penalty score based on the high-risk defect score and the penalty coefficient.

[0123] This step further evaluates the high-risk defects present in the industrial dataset under test and generates corresponding high-risk defect scores. Specifically, high-risk defect identification rules and high-risk defect scoring rules are pre-generated. The high-risk defect identification rules are then used to identify whether high-risk defects exist in the industrial dataset under test. If so, the high-risk defect scoring rules are further used to generate the corresponding high-risk defect scores.

[0124] For example, high-risk defect identification rules may include, but are not limited to, at least one of the following: identifying the number of missing fields, identifying the number of illegal device state transfers, the number of incorrect annotations, and the number of files that have not been anonymized. High-risk defect scoring rules can be determined based on the number of defects detected by the high-risk defect identification rules; the more defects detected, the higher the high-risk defect score.

[0125] Furthermore, for each type of high-risk defect, a penalty coefficient matching that high-risk defect is generated. For example, the penalty coefficients for compliance defects (such as data not being anonymized), labeling defects (such as incorrect or unclear labeling), and industrial mechanism defects (such as energy imbalance exceeding a set threshold) are different.

[0126] For any identified high-risk defect, a penalty score is generated based on the high-risk defect score and penalty coefficient. The risk penalty score of the industrial dataset to be tested is generated based on the sum of the penalty scores of all identified high-risk defects.

[0127] Step S605: Generate a quality score for the industrial dataset to be tested based on the base score and the risk penalty score.

[0128] The quality score of the industrial dataset to be tested is obtained based on the difference between the base score and the risk penalty score. For example, the quality score of the industrial dataset to be tested can be generated using the following formula (3).

[0129] Formula (3)

[0130] in, Indicates quality score; Indicates the basic score; This indicates the number of high-risk defect types in the industrial dataset to be tested; Indicates the first High-risk defect scoring for high-risk defects; Indicates the first The penalty coefficient for high-risk defects.

[0131] Therefore, the industrial dataset quality inspection method provided in this disclosure generates a basic score based on indicator scores, weights, and indicator credibility. It can dynamically adjust the contribution of each indicator to the basic score according to the reliability of the evaluation results, thereby improving the credibility of the basic score. Furthermore, it further detects high-risk defects in the industrial dataset to be tested, generates a risk penalty score based on the high-risk defect score and penalty coefficient, and deducts the risk penalty score from the basic score to obtain the final quality score. This makes high-risk defects have a significant penalty effect on the final score, preventing fatal defects from being masked by high scores of other indicators, and enhancing the accuracy of the quality inspection results.

[0132] Figure 7 A flowchart illustrating another industrial dataset quality inspection method provided in this disclosure is shown. Specifically, as... Figure 7 As shown, the method includes the following steps: Step S701: Generate the quality inspection results of the industrial dataset to be tested.

[0133] The specific process for generating the quality inspection results of the industrial dataset to be tested can be described in other embodiments, and will not be repeated here. The generated quality inspection results include at least a quality score for the industrial dataset to be tested.

[0134] In one alternative implementation, the generated quality inspection results may further include the index scores and / or the corresponding index credibility of each index in the index set.

[0135] In one optional implementation, the generated quality inspection results may further include sub-scores for different dimensional attributes under at least one partitioning dimension. Specifically, the indicators in the indicator set are divided into multiple subsets according to at least one partitioning dimension. Each partitioning dimension contains multiple dimensional attributes, and each dimensional attribute corresponds to a subset containing indicators belonging to that dimensional attribute. For example, partitioning according to the risk level dimension yields high-risk, medium-risk, and low-risk subsets; partitioning according to the data type dimension yields image subsets, time-series subsets, text subsets, etc. For each subset, the weighted score or average score of each indicator in that subset is calculated, thereby obtaining the sub-score corresponding to that dimensional attribute. Thus, the generated quality inspection results contain multi-level and multi-dimensional scoring data, facilitating subsequent data defect localization and data analysis.

[0136] In one optional implementation, corresponding risk warning information can also be generated based on the quality score and / or indicator score. For example, when the quality score is lower than the warning threshold, corresponding risk warning information is generated; or, when the indicator score of a certain indicator is lower than the warning threshold, risk warning information corresponding to that indicator is generated.

[0137] In one alternative implementation, the quality inspection results can be visualized using tables, heatmaps, or sorting (such as sorting by indicator scores).

[0138] Step S702: Locate the data defects in the industrial dataset to be tested based on the quality inspection results.

[0139] Based on the quality inspection results, specific data defects are located in the industrial dataset to be tested. Data defects are actual quality problems existing in the dataset, such as missing fields, incorrect labeling, misaligned timestamps, energy balance violations, and un-de-identified sensitive information.

[0140] In one optional implementation, data defects are identified based on indicator scores. Specifically, abnormal indicators with scores below a set lower limit (e.g., 60 points) can be filtered out. The data defect corresponding to the abnormal indicator is searched from a pre-generated indicator defect mapping table, and this data defect is considered a data defect present in the industrial dataset under test. For example, if the integrity indicator score is below 60 points, by searching the indicator defect mapping table, the corresponding data defect is determined to be a missing field, thus classifying the missing field as a data defect in the industrial dataset under test.

[0141] In one alternative implementation, data defects are identified based on sub-ratings of dimensional attributes. Specifically, abnormal dimensional attributes with sub-ratings below a set lower limit can be filtered out, and data defects are identified based on these abnormal dimensional attributes. For example, if divided by data source dimension, sub-scores can be obtained for different data sources (such as equipment, sensors, systems, production lines, or workstations). If the sub-score of a data source is lower than the lower limit, it is determined that the data source has data defects. Similarly, if divided by time dimension, sub-scores can be obtained for different time periods. If the sub-score for a certain time period is lower than the lower limit, it is determined that the time period has data defects. If divided by modality type dimension, sub-scores can be obtained for different modality types. If the sub-score for a certain modality type is lower than the lower limit, it is determined that the modality type has data defects. If divided by label dimension, sub-scores can be obtained for different labels. If the sub-score for a certain label is lower than the lower limit, it is determined that the data corresponding to that label has data defects. If divided by process dimension, sub-scores can be obtained for different processes. If the sub-score for a certain process is lower than the lower limit, it is determined that the data corresponding to that process has data defects. If divided by version dimension, sub-scores can be obtained for different versions. If the sub-score for a certain version is lower than the lower limit, it is determined that the data corresponding to that version has data defects.

[0142] In one alternative implementation, the severity of identified data defects can be quantitatively evaluated based on the absolute value of the difference between the indicator score and / or the sub-score of the dimension attribute and a set lower limit for the score. For example, the larger the absolute value of the difference between an indicator score and the lower limit for the score, the lower the indicator score, and the more severe the data defect corresponding to that indicator score.

[0143] Step S703: Perform root cause analysis on the data defects and generate remediation suggestions based on the root cause analysis results.

[0144] After identifying data defects in the industrial dataset to be tested, the causes of these defects are further analyzed to obtain root cause analysis results. These root cause analysis results may include, but are not limited to, at least one of the following: abnormal data acquisition configuration, out-of-sync device clocks, errors in preprocessing scripts, ambiguous labeling standards, insufficient supplementation of long-tail samples, incomplete desensitization of sensitive information, distribution drift caused by operational condition migration, format conflicts caused by the return of old version data, and failure to synchronize mechanism rule updates to the data governance process.

[0145] In one alternative implementation, root cause analysis can be performed using one or more of the following methods in combination: root cause analysis based on a rule base, root cause analysis based on a causal knowledge graph, root cause analysis based on a process log, and root cause analysis based on a historical question database.

[0146] Root cause analysis is performed based on a rule base. Specifically, a mapping relationship between defects and their causes is generated in the rule base. For example, the mapping relationship could be: data defect "timestamp reversed" corresponds to the cause "device clock asynchrony" or "abnormal acquisition configuration", and data defect "energy balance violation" corresponds to the cause "sensor calibration parameters expired". By searching and comparing this mapping relationship, the cause of the data defect can be quickly determined.

[0147] Root cause analysis based on causal knowledge graphs involves pre-constructing a causal knowledge graph, establishing data defect nodes and cause nodes within it, and creating edges between related data defect nodes and cause nodes. By comparing the located data defects against the causal knowledge graph, the causes of the data defects can be quickly determined.

[0148] Root cause analysis is performed based on process logs. Specifically, process logs such as acquisition logs, data processing logs, annotation logs, and version change records are read. Operation records near the time when the data defect occurred are extracted from these process logs, and the events or operations corresponding to these operation records are taken as the causes of the data defect.

[0149] Root cause analysis is performed based on a historical problem database. Specifically, the database stores historical data defects that have undergone root cause analysis and their corresponding root cause results. The current data defect is compared with the historical data defects in the database to identify the historical data defect with the highest similarity to the current data defect. The root cause result of the current data defect is then generated based on the root cause result of that historical data defect.

[0150] Further, remediation recommendations are generated based on the root cause analysis results. These recommendations may include: remediation actions, target objects, priorities, expected benefits, and implementation cost assessments. The generation of remediation recommendations can be based on a predefined root cause-remediation mapping table or through pre-defined remediation rules. These modification recommendations may include, but are not limited to, at least one of the following: recommending re-collecting data for a specific equipment range; recommending supplementary labeling and dual-person review of a certain type of defect sample; recommending resampling and aligning time-series data according to a unified clock; recommending adding a working condition label field and completing long-tail working condition samples; recommending revising a certain type of boundary definition in the labeling guidelines; recommending graded desensitization of sensitive fields; recommending removing contaminated versions and rolling back to a stable version; and recommending establishing mechanism verification rules for preset information.

[0151] In one optional implementation, at least one calibration reference can be obtained to update at least one calibration object in the quality inspection method provided in this disclosure embodiment, thereby improving the adaptability of the quality inspection method. This achieves a closed-loop architecture encompassing quality inspection, locating data defects based on the quality inspection results, repairing the data defects, performing quality inspection again after repair, and updating parameters in the quality inspection method.

[0152] The calibration basis may include, but is not limited to, at least one of the following: the re-evaluation results after data defect repair, the performance changes of downstream industrial models before and after training with the dataset, online false alarms, missed alarms, drift alarm information, expert review opinions, feedback information on new scenarios and new working conditions, and updated standard and specification information. The calibration objects may include, but are not limited to, at least one of the following: set relevant thresholds, indicator weight rules, mechanism rules, classifiers, and credibility determination rules. For example, if in multiple evaluations it is found that the pass threshold for the integrity indicator (e.g., 95%) is too strict, leading to a large number of actually usable datasets being judged as unqualified, while the downstream industrial model performs well on these datasets (i.e., the model performance improves after training with the dataset), then the pass threshold for the integrity indicator will be automatically adjusted from 95% to 92%.

[0153] Therefore, the industrial dataset quality inspection method provided in this disclosure, after obtaining the quality inspection results, further performs root cause analysis based on the quality inspection results and generates corresponding repair suggestions, thereby helping to improve the quality of the dataset.

[0154] Figure 8 A schematic diagram of the structure of an industrial dataset quality inspection device provided in an embodiment of this disclosure is shown. Figure 8 As shown, the device 800 includes: an acquisition module 810, a preprocessing module 820, an identification module 830, an index set construction module 840, a weight determination module 850, and a joint detection module 860.

[0155] The acquisition module 810 is used to acquire the industrial dataset to be tested and the associated description information of the industrial dataset to be tested; The preprocessing module 820 is used to perform fusion preprocessing on the industrial dataset to be tested and the associated descriptive information to generate fused data; The identification module 830 is used to identify the modal type, industrial scenario type, and task objective of the industrial dataset to be tested based on the fused data. The indicator set construction module 840 is used to extract an indicator set that matches the modality type and the industrial scenario type from a preset indicator library; The weight determination module 850 is used to dynamically determine the weight of each indicator in the indicator set according to the task objective. The joint detection module 860 is used to perform joint detection of basic quality detection and industrial characteristic matching detection on the industrial dataset to be tested based on the indicator set and the weight, and generate quality detection results.

[0156] In one alternative implementation, the weight determination module 850 is used to: pre-generate a task indicator knowledge graph; For any indicator in the indicator set, calculate the correlation coefficient between that indicator and the task objective in the task indicator knowledge graph; The weight of each indicator is determined based on the correlation coefficient between each indicator in the indicator set and the task objective.

[0157] In one optional implementation, the weight determination module 850 is used to: dynamically determine the basic weights of each indicator in the indicator set according to the task objective; Based on at least one weighting factor, determine the correction coefficient for at least one indicator in the indicator set; Based on the basic weights and the correction coefficients, the weights of each indicator in the indicator set are generated; The weighting influence factors include at least one of the following: The risk level corresponding to the industrial scenario type, the sample distribution characteristics of the industrial dataset to be tested, the dataset size of the industrial dataset to be tested, the historical feedback of the downstream industrial model, and the priority configuration information.

[0158] In an optional implementation, the joint detection module 860 is used to: for any indicator in the indicator set, determine the basic quality detection function and / or industrial characteristic matching function corresponding to the indicator, and calculate the indicator score of the indicator using the corresponding basic quality detection function and / or industrial characteristic matching function; The quality inspection results are generated based on the index scores and weights of each index in the index set.

[0159] In one optional implementation, the joint detection module 860 is used to: generate a basic score based on the index scores and weights of each index in the index set; Obtain the high-risk defect score of the industrial dataset to be tested, and generate a risk penalty score based on the high-risk defect score and the penalty coefficient. A quality score for the industrial dataset to be tested is generated based on the base score and the risk penalty score.

[0160] In one optional implementation, the joint detection module 860 is used to generate a basic score based on the index scores, weights, and credibility of each index in the index set.

[0161] In an optional embodiment, the device further includes a post-processing module (not shown in the figure) for locating data defects in the industrial dataset to be tested based on the quality inspection results; Root cause analysis is performed on the data defects, and remediation suggestions are generated based on the results of the root cause analysis.

[0162] The industrial dataset quality inspection device of this disclosure corresponds to the embodiments of the industrial dataset quality inspection method described above, and the relevant content can be referred to each other, which will not be repeated here. The beneficial technical effects of the industrial dataset quality inspection device of this disclosure can be referred to the corresponding beneficial technical effects of the industrial dataset quality inspection method described above, which will not be repeated here.

[0163] In addition, embodiments of this disclosure also provide an electronic device, including: Memory, used to store computer programs; A processor is configured to execute a computer program stored in the memory, wherein, when the computer program is executed, it implements the industrial dataset quality detection method described in any of the above embodiments of the present disclosure.

[0164] Figure 9 A schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure is shown. Below, reference is made to… Figure 9 This describes an electronic device according to embodiments of the present disclosure. The electronic device may be either or both of a first device and a second device, or a standalone device independent of them, which may communicate with the first device and the second device to receive acquired input signals from them.

[0165] like Figure 9 As shown, the electronic device includes one or more processors and memory.

[0166] A processor can be a central processing unit (CPU) or other form of processing unit with data processing and / or instruction execution capabilities, and can control other components in an electronic device to perform desired functions.

[0167] The memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and a processor may execute the program instructions to implement the industrial dataset quality inspection methods of the various embodiments of this disclosure described above, and / or other desired functions.

[0168] In one example, the electronic device may also include input devices and output devices, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0169] In addition, the input device may include, for example, a keyboard, a mouse, etc.

[0170] This output device can output various information to the outside, including determined distance information, direction information, etc. The output device may include, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.

[0171] Of course, for the sake of simplicity, Figure 9 Only some of the components of the electronic device relevant to this disclosure are shown, omitting components such as buses, input / output interfaces, etc. In addition, the electronic device may include any other suitable components depending on the specific application.

[0172] In addition to the methods and apparatus described above, embodiments of this disclosure may also be computer program products comprising computer program instructions that, when executed by a processor, cause the processor to perform the steps in the industrial dataset quality inspection methods according to various embodiments of this disclosure as described in the foregoing portions of this specification.

[0173] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this disclosure. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on a user's computing device, partially on a user's computing device, as a standalone software package, partially on a user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0174] Furthermore, embodiments of this disclosure may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps in the industrial dataset quality inspection method according to various embodiments of this disclosure as described in the foregoing portion of this specification.

[0175] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0176] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as ROM, RAM, magnetic disk, or optical disk.

[0177] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0178] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments, since they largely correspond to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0179] The block diagrams of devices, apparatuses, devices, and systems disclosed herein are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0180] The methods and apparatus of this disclosure may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the methods is for illustrative purposes only, and the steps of the methods of this disclosure are not limited to the order specifically described above unless otherwise specifically stated. Furthermore, in some embodiments, this disclosure may also be implemented as a program recorded on a recording medium, the program including machine-readable instructions for implementing the methods according to this disclosure. Thus, this disclosure also covers recording media storing programs for performing the methods according to this disclosure.

[0181] It should also be noted that in the apparatus, devices, and methods of this disclosure, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions to this disclosure.

[0182] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0183] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A method for quality inspection of industrial datasets, characterized in that, include: Obtain the industrial dataset to be tested and its associated descriptive information; The industrial dataset to be tested and the associated descriptive information are fused and preprocessed to generate fused data; Based on the fused data, the modal type, industrial scenario type, and task objective of the industrial dataset to be tested are identified; Extract a set of indicators that match the modality type and the industrial scenario type from a preset indicator library; Based on the stated task objectives, the weights of each indicator in the indicator set are dynamically determined; Based on the indicator set and the weights, the industrial dataset to be tested is subjected to joint detection of basic quality detection and industrial characteristic matching detection, and quality detection results are generated. The step of dynamically determining the weights of each indicator in the indicator set based on the task objective includes: A task indicator knowledge graph is pre-generated. The task indicator knowledge graph contains two types of graph nodes: task nodes and indicator nodes. Each task node corresponds to a task objective, and each indicator node corresponds to an indicator. A node edge is established between two related graph nodes. The node edge can be a node edge without a significant weight or a node edge with a significant weight. The significant weight is used to describe the degree of association between the two graph nodes. For any indicator in the indicator set, calculate the correlation coefficient between the indicator and the task target in the task indicator knowledge graph. The correlation coefficient is used to quantify the degree of correlation between the indicator node corresponding to the indicator and the task node corresponding to the task target in the task indicator knowledge graph. The correlation coefficient is calculated using a path accumulation method, which includes: if a node edge has significant weights, multiplying the significant weights of all node edges between the task node and the indicator node to obtain the correlation coefficient; if a node edge does not have significant weights, obtaining the correlation coefficient based on the reciprocal of the number of node edges between the task node and the indicator node; if there is no reachable path between the task node and the indicator node in the task indicator knowledge graph, the correlation coefficient is taken as the minimum set correlation coefficient. The weight of each indicator is determined based on the correlation coefficient between each indicator in the indicator set and the task objective, and the weight of each indicator is positively correlated with the corresponding correlation coefficient.

2. The method according to claim 1, characterized in that, The step of dynamically determining the weights of each indicator in the indicator set based on the task objective includes: Based on the stated task objectives, the basic weights of each indicator in the indicator set are dynamically determined; Based on at least one weighting factor, determine the correction coefficient for at least one indicator in the indicator set; Based on the basic weights and the correction coefficients, the weights of each indicator in the indicator set are generated; The weighting influence factors include at least one of the following: The risk level corresponding to the industrial scenario type, the sample distribution characteristics of the industrial dataset to be tested, the dataset size of the industrial dataset to be tested, the historical feedback of the downstream industrial model, and the priority configuration information.

3. The method according to any one of claims 1-2, characterized in that, The process of jointly detecting basic quality and industrial characteristic matching based on the indicator set and the weights, and generating quality detection results, includes: For any indicator in the set of indicators, determine the basic quality testing function and / or industrial characteristic matching function corresponding to the indicator, and calculate the indicator score using the corresponding basic quality testing function and / or industrial characteristic matching function; The quality inspection results are generated based on the index scores and weights of each index in the index set.

4. The method according to claim 3, characterized in that, The process of generating quality inspection results based on the index scores and weights of each index in the index set includes: A basic score is generated based on the score and weight of each indicator in the indicator set. Obtain the high-risk defect score of the industrial dataset to be tested, and generate a risk penalty score based on the high-risk defect score and the penalty coefficient. A quality score for the industrial dataset to be tested is generated based on the base score and the risk penalty score.

5. The method according to claim 4, characterized in that, The process of generating a basic score based on the score and weight of each indicator in the indicator set includes: A basic score is generated based on the score, weight, and credibility of each indicator in the indicator set.

6. The method according to any one of claims 1-2, characterized in that, After generating the quality inspection results, the method further includes: Based on the quality inspection results, locate the data defects existing in the industrial dataset to be tested; Root cause analysis is performed on the data defects, and remediation suggestions are generated based on the results of the root cause analysis.

7. An industrial dataset quality inspection device, characterized in that, include: The acquisition module is used to acquire the industrial dataset to be tested and the associated descriptive information of the industrial dataset to be tested; The preprocessing module is used to perform fusion preprocessing on the industrial dataset to be tested and the associated descriptive information to generate fused data; The identification module is used to identify the modal type, industrial scenario type, and task objective of the industrial dataset to be tested based on the fused data. The indicator set construction module is used to extract an indicator set that matches the modality type and the industrial scenario type from a preset indicator library; The weight determination module is used to dynamically determine the weight of each indicator in the indicator set based on the task objective. The joint detection module is used to perform joint detection of basic quality detection and industrial characteristic matching detection on the industrial dataset to be tested based on the indicator set and the weights, and generate quality detection results. Specifically, the weight determination module is used for: A task indicator knowledge graph is pre-generated. The task indicator knowledge graph contains two types of graph nodes: task nodes and indicator nodes. Each task node corresponds to a task objective, and each indicator node corresponds to an indicator. A node edge is established between two related graph nodes. The node edge can be a node edge without a significant weight or a node edge with a significant weight. The significant weight is used to describe the degree of association between the two graph nodes. For any indicator in the indicator set, calculate the correlation coefficient between the indicator and the task target in the task indicator knowledge graph. The correlation coefficient is used to quantify the degree of correlation between the indicator node corresponding to the indicator and the task node corresponding to the task target in the task indicator knowledge graph. The correlation coefficient is calculated using a path accumulation method, which includes: if a node edge has significant weights, multiplying the significant weights of all node edges between the task node and the indicator node to obtain the correlation coefficient; if a node edge does not have significant weights, obtaining the correlation coefficient based on the reciprocal of the number of node edges between the task node and the indicator node; if there is no reachable path between the task node and the indicator node in the task indicator knowledge graph, the correlation coefficient is taken as the minimum set correlation coefficient. The weight of each indicator is determined based on the correlation coefficient between each indicator in the indicator set and the task objective, and the weight of each indicator is positively correlated with the corresponding correlation coefficient.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute a computer program stored in the memory, wherein when the computer program is executed, it implements the industrial dataset quality inspection method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the industrial dataset quality inspection method according to any one of claims 1-6.

10. A computer program product comprising computer program instructions, characterized in that, When the computer program instructions are executed by the processor, they implement the industrial dataset quality inspection method according to any one of claims 1-6.

Images