Artificial intelligence system failure characterization analysis method, device and computer equipment

Abstract

The present application relates to a method, device, computer equipment, storage medium and computer program product for failure characterization analysis of an artificial intelligence system. The method obtains a domain data set and various types of failure characterization elements of the artificial intelligence system to be analyzed, and then determines the value of the failure characterization element of the artificial intelligence system based on the environmental characteristic type characterization elements and data characteristic type characterization elements of the data samples in the domain data set; constructs a sample prone to failure based on the data samples, failure mode type characterization elements and accident case type characterization elements in the domain data set, and obtains the sample label of the sample prone to failure; determines the failure characterization law of the artificial intelligence system based on the failure characterization element value and the sample label. The present application constructs failure characterization around the domain data set, environment, failure mode and accident case, and uses data mining technology to obtain the failure characterization law based on the failure characterization value in the domain data set, which can effectively improve the accuracy of the failure law analysis of the artificial intelligence system.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method, apparatus, computer equipment, storage medium and computer program product for failure characterization and analysis of an artificial intelligence system. Background Art

[0002] With the development of computer technology, artificial intelligence (AI) technology has emerged. Artificial intelligence is an important part of the intelligent discipline. It attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Research in this field includes robots, language recognition, image recognition, natural language processing and expert systems. Since the birth of artificial intelligence, the theory and technology of artificial intelligence have become increasingly mature, and the application field has continued to expand. It can be imagined that the technological products brought by artificial intelligence in the future will be the "container" of human wisdom. Artificial intelligence can simulate the information process of human consciousness and thinking. Artificial intelligence is not human intelligence, but it can think like a human and may also exceed human intelligence. However, the core technology of artificial intelligence software system - deep learning technology has the characteristics of probabilistic system and technical black box, and the technical principle is very different from traditional software. Artificial intelligence software system has a series of new characteristics such as strong adaptability and evolution, weak generalization, weak interpretability, high uncertainty, unclear failure mechanism, etc., which makes it easier to make unpredictable wrong decisions. Traditional software reliability analysis, testing, evaluation and other technologies are difficult to adapt to these new characteristics. If the reliability of artificial intelligence software systems is not implemented properly, it may cause frequent software product failures, a significant increase in total costs, and damage to brand image. It may also cause casualties, serious property losses, and environmental damage. Therefore, it is necessary and urgent to conduct research on the reliability of artificial intelligence software systems.

[0003] The core AI components in AI software systems are essentially probabilistic systems with high inherent uncertainty, and are sensitive to changes in the external environment. They are prone to significant declines in predictive capabilities in new scenarios not included in the training set. AI components do not have typical control structures, and the failure mechanism is unclear, making it difficult to conduct comprehensive coverage testing of AI software systems. The test inputs for the core components of AI software systems are mostly unstructured data, such as images, text, and voice. The reliability test data generation process is not intuitive, and it is difficult to determine whether it reflects the actual usage scenario. The above problems make AI software systems more likely to make unpredictable erroneous decisions in rare scenarios than traditional software, and it is more difficult to eliminate software failures through prior testing. Summary of the invention

[0004] Based on this, it is necessary to provide an artificial intelligence system failure characterization analysis method, device, computer equipment, computer-readable storage medium and computer program product that can effectively and accurately analyze the failure laws of artificial intelligence systems in response to the above-mentioned technical problems.

[0005] In a first aspect, the present application provides a method for analyzing failure characterization of an artificial intelligence system, comprising:

[0006] Obtaining a domain data set and various types of failure characterization elements of the artificial intelligence system to be analyzed, wherein the failure characterization elements include environmental feature characterization elements, data feature characterization elements, failure mode characterization elements, and accident case characterization elements;

[0007] Determine the value of the failure characterization factor of the artificial intelligence system based on the environmental characteristic characterization factors and data characteristic characterization factors of the data samples in the domain data set;

[0008] Constructing a failure-prone sample based on the data sample in the domain data set, the failure mode class characterization element, and the accident case class characterization element, and obtaining a sample label of the failure-prone sample;

[0009] The failure characterization law of the artificial intelligence system is determined according to the failure characterization factor value and the sample label.

[0010] In a second aspect, the present application also provides an artificial intelligence system failure characterization analysis device, comprising:

[0011] A data acquisition module is used to acquire a domain data set of the artificial intelligence system to be analyzed and various types of failure characterization elements, wherein the failure characterization elements include environmental feature characterization elements, data feature characterization elements, failure mode characterization elements, and accident case characterization elements;

[0012] A first analysis module is used to determine the value of the failure characterization element of the artificial intelligence system based on the environmental characteristic characterization elements and data characteristic characterization elements of the data samples in the domain data set;

[0013] The second analysis module is used to construct a sample prone to failure based on the data sample in the domain data set, the failure mode characterization element and the accident case characterization element, and obtain a sample label of the sample prone to failure;

[0014] A rule analysis module is used to determine the failure characterization rule of the artificial intelligence system according to the failure characterization factor value and the sample label.

[0015] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0016] Obtaining a domain data set and various types of failure characterization elements of the artificial intelligence system to be analyzed, wherein the failure characterization elements include environmental feature characterization elements, data feature characterization elements, failure mode characterization elements, and accident case characterization elements;

[0017] Determine the value of the failure characterization factor of the artificial intelligence system based on the environmental characteristic characterization factors and data characteristic characterization factors of the data samples in the domain data set;

[0018] Constructing a failure-prone sample based on the data sample in the domain data set, the failure mode class characterization element, and the accident case class characterization element, and obtaining a sample label of the failure-prone sample;

[0019] The failure characterization law of the artificial intelligence system is determined according to the failure characterization factor value and the sample label.

[0020] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the following steps are implemented:

[0021] Obtaining a domain data set and various types of failure characterization elements of the artificial intelligence system to be analyzed, wherein the failure characterization elements include environmental feature characterization elements, data feature characterization elements, failure mode characterization elements, and accident case characterization elements;

[0022] Determine the value of the failure characterization factor of the artificial intelligence system based on the environmental characteristic characterization factors and data characteristic characterization factors of the data samples in the domain data set;

[0023] Constructing a failure-prone sample based on the data sample in the domain data set, the failure mode class characterization element, and the accident case class characterization element, and obtaining a sample label of the failure-prone sample;

[0024] The failure characterization law of the artificial intelligence system is determined according to the failure characterization factor value and the sample label.

[0025] In a fifth aspect, the present application further provides a computer program product, including a computer program, which implements the following steps when executed by a processor:

[0026] Obtaining a domain data set and various types of failure characterization elements of the artificial intelligence system to be analyzed, wherein the failure characterization elements include environmental feature characterization elements, data feature characterization elements, failure mode characterization elements, and accident case characterization elements;

[0027] Determine the value of the failure characterization factor of the artificial intelligence system based on the environmental characteristic characterization factors and data characteristic characterization factors of the data samples in the domain data set;

[0028] Constructing a failure-prone sample based on the data sample in the domain data set, the failure mode class characterization element, and the accident case class characterization element, and obtaining a sample label of the failure-prone sample;

[0029] The failure characterization law of the artificial intelligence system is determined according to the failure characterization factor value and the sample label.

[0030] The above-mentioned artificial intelligence system failure characterization analysis method, device, computer equipment, storage medium and computer program product obtain the domain data set and various types of failure characterization elements of the artificial intelligence system to be analyzed, and then determine the failure characterization element values ​​of the artificial intelligence system based on the environmental characteristic type characterization elements and data characteristic type characterization elements of the data samples in the domain data set; construct easy-to-fail samples based on the data samples, failure mode type characterization elements and accident case type characterization elements in the domain data set to obtain sample labels of easy-to-fail samples; determine the failure characterization law of the artificial intelligence system based on the failure characterization element values ​​and sample labels. This application constructs various types of failure characterization elements of the artificial intelligence system to be analyzed, and then analyzes and determines the corresponding failure characterization element values ​​and sample labels of easy-to-fail samples in combination with the domain data set of the artificial intelligence system, constructs failure characterizations around the domain data set, environment, failure mode and accident case, and uses data mining technology to obtain failure characterization laws based on the failure characterization values ​​in the domain data set, which can effectively improve the accuracy of the analysis of the failure laws of artificial intelligence systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0032] Figure 1 A diagram of an application environment of a failure characterization and analysis method for an artificial intelligence system in one embodiment;

[0033] Figure 2 A schematic diagram of the difference between a failure mechanism study and a failure characterization law study in one embodiment;

[0034] Figure 3 A schematic diagram of a process of a failure characterization analysis method for an artificial intelligence system in one embodiment;

[0035] Figure 4 In one embodiment Figure 3 Schematic diagram of a sub-process of step 302;

[0036] Figure 5 A schematic diagram of the application flow of a model development process in one embodiment;

[0037] Figure 6 It is a structural block diagram of a failure characterization and analysis device for an artificial intelligence system in one embodiment;

[0038] Figure 7 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0040] The artificial intelligence system failure characterization analysis method provided in the embodiment of the present application can be applied to Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. When the user of the terminal 102 wants to analyze the artificial intelligence system applied to the specified field, such as the machine learning model, and determine its failure law, the failure analysis can be performed through the artificial intelligence system failure characterization analysis method of the present application. First, the user can submit the domain data set and various types of failure characterization elements of the artificial intelligence system to be analyzed to the server 104 through the terminal 102. The failure characterization elements include environmental feature characterization elements, data feature characterization elements, failure mode characterization elements, and accident case characterization elements. Then the server 104 obtains the domain data set and various types of failure characterization elements of the artificial intelligence system to be analyzed, and determines the value of the failure characterization element of the artificial intelligence system based on the environmental feature characterization elements and data feature characterization elements of the data samples in the domain data set; constructs a sample that is prone to failure based on the data samples, failure mode characterization elements, and accident case characterization elements in the domain data set, and obtains the sample label of the sample that is prone to failure; determines the failure characterization law of the artificial intelligence system according to the failure characterization element value and the sample label. The terminal 102 may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. The portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers.

[0041] The applicant found through research that since the test inputs of the core components of the artificial intelligence software system are mostly unstructured data, such as images, text, and voice, the reliability test data generation process is not intuitive, and it is difficult to determine whether it reflects the actual usage scenario. In addition, the machine learning model, the foundation of the artificial intelligence software system, has a technical black box nature and weak interpretability, making it difficult to conduct failure mechanism research from the perspective of failure physics, similar to traditional reliability. Figure 2The difference between the study of failure mechanism and the study of failure characterization law is shown in the figure. The study of failure mechanism mainly finds out the failure factors based on failure physics, clarifies the relationship between failure factors (linear, exponential, etc.), establishes a formulaic description, obtains experimental data through reliability tests, and estimates the parameters of the fitting failure formula. The study of failure characterization law is based on the dimension of analysis, that is, measurement, to find out all the characterization factors that may affect the prediction results of the model, and obtains the values ​​and prediction results of the failure characterization in the training set during the model training process, so as to identify the failure characterization law. The difference between failure mechanism and failure characterization is that the failure mechanism is based on failure physics and can be directly applied to hardware and components with the same physical properties; while the failure characterization law is based on probability statistics and is often limited to specific functions or tasks (recognizing specific types of images, such as facial images). Traditional software has a low degree of standardization, and it is almost impossible to find software with the same functions and structures, so the failure characterization law is often only applicable to specific objects; while the neural network model has certain standardized characteristics, and the prediction results of well-trained models on most data samples are very similar, so the failure characterization law can be applied to a class of neural network models with similar functions. Therefore, the artificial intelligence system failure characterization analysis method of the present application is proposed to effectively analyze the failure characterization laws of the artificial intelligence system, which can be applied to the fields of training, testing and optimization of the artificial intelligence system.

[0042] In an exemplary embodiment, Figure 3 As shown in the figure, a failure characterization analysis method for an artificial intelligence system is provided. Figure 1 The server 104 in the example is used as an example to illustrate, including the following steps 302 to 308. Among them:

[0043] Step 302, obtain the domain data set of the artificial intelligence system to be analyzed and various types of failure characterization elements, where the failure characterization elements include environmental feature characterization elements, data feature characterization elements, failure mode characterization elements, and accident case characterization elements.

[0044] Among them, the artificial intelligence system to be analyzed is the analysis target of this application. The artificial intelligence system can be artificial intelligence software created based on artificial intelligence, or a system such as a machine learning model. The domain data set refers to the training data set constructed by the application field of the artificial intelligence system to be analyzed. For example, for an artificial intelligence system of the image processing type, its domain data set can be an image sample set, and for an artificial intelligence system of the text processing type, its domain data set can be a text sample set. As for the failure characterization elements, it specifically includes environmental feature characterization elements, data feature characterization elements, failure mode characterization elements, and accident case characterization elements, which respectively describe environmental features (such as geographical or weather conditions), data features (typical entities in the data set and relationships between entities), failure modes (such as data pipelines, human-computer interaction, natural disturbances, external attacks, etc.) and accident cases (specific combinations of environmental features, data features, and failure modes).

[0045] For example, when a user wishes to analyze a specified artificial intelligence system, the terminal 102 can provide the server 104 with information related to the artificial intelligence system to be analyzed, such as the system application field, the system initial model, and the field data set, and the server 104 performs basic analysis based on the information provided by the terminal 102, and can first obtain the field data set of the artificial intelligence system to be analyzed, and determine various types of failure characterization elements based on the provided information. Of course, these failure characterization elements can also be collected by the user and provided to the server 104. In one embodiment, for typical artificial intelligence fields such as images, texts, and audios, the artificial intelligence failure characterization elements under various types are shown in Table 1.

[0046] Table 1 Table of artificial intelligence failure characteristics

[0047]

[0048] Step 304, based on the environmental characteristic type representation elements and data characteristic type representation elements of the data samples in the domain data set, determine the value of the failure representation element of the artificial intelligence system.

[0049] Among them, the value of the failure characterization factor refers to the range of environmental characteristic characterization factors or data characteristic characterization factors that will cause the artificial intelligence system to fail. For example, for environmental characteristic characterization factors, when a certain environmental characteristic characterization factor reaches the corresponding extreme value, the processing accuracy of the artificial intelligence system will drop significantly. At this time, it can be determined that the artificial intelligence system has failed, and this value is the value of the failure characterization factor.

[0050] For example, after obtaining various types of failure characterization elements of the artificial intelligence system, the reasons for the failure of the artificial intelligence system due to the failure characterization elements should be analyzed. For environmental characteristic characterization elements and data characteristic characterization elements, they contain normal value ranges and failure value ranges. Within the normal value range, environmental characteristic characterization elements and data characteristic characterization elements will not cause failure problems in the artificial intelligence system, but when they are in the failure value range, environmental characteristic characterization elements and data characteristic characterization elements may cause failure problems in the artificial intelligence system. Therefore, the environmental characteristic characterization elements and data characteristic characterization elements can be analyzed through data samples in the domain data set to determine the range within which the values ​​of these two elements will affect the effectiveness of the artificial intelligence system, and obtain the failure characterization element values ​​corresponding to each of the two characteristic elements.

[0051] Step 306 , constructing a failure-prone sample based on the data samples in the domain data set, the failure mode characterization elements, and the accident case characterization elements, and obtaining a sample label of the failure-prone sample.

[0052] Among them, samples prone to failure refer to sample data that will cause the failure of the artificial intelligence system. Because both the failure mode characterization factors and the accident case characterization factors need to be analyzed in combination with specific sample scenarios, it is necessary to construct samples prone to failure to analyze and process the failure rules under the failure mode characterization factors and the accident case characterization factors.

[0053] Exemplarily, for failure mode characterization elements and accident case characterization elements, failure pattern analysis can be performed by constructing samples that are prone to failure. Specifically, the failure mode characterization elements and accident case characterization elements can be used to process data samples in the domain data set, enhance the samples, construct corresponding enhanced samples, and then determine the samples that are prone to failure through analysis of the enhanced samples. The failure mode characterization elements and accident case characterization elements corresponding to the sample labels of the failed samples are the failure elements of the machine learning model determined after the analysis.

[0054] Step 308, determining the failure characterization law of the artificial intelligence system according to the failure characterization factor values ​​and sample labels.

[0055] Exemplarily, after obtaining the failure characterization factor values ​​and sample labels of samples prone to failure, the failure characterization factor values ​​or failure sample labels corresponding to each type of factor in the four types of factors, namely, environmental feature characterization factors, data feature characterization factors, failure mode characterization factors, and accident case characterization factors, are determined. These data can be used as the final failure characterization laws. Through these failure characterization laws, the samples used in the training process of the artificial intelligence system can be analyzed, such as screening out unqualified samples, etc., to improve the training efficiency of the artificial intelligence system and ensure the use effect of the artificial intelligence system.

[0056] The above-mentioned artificial intelligence system failure characterization analysis method obtains the domain data set and various types of failure characterization elements of the artificial intelligence system to be analyzed, and then determines the failure characterization element values ​​of the artificial intelligence system based on the environmental characteristic type characterization elements and data characteristic type characterization elements of the data samples in the domain data set; constructs failure-prone samples based on the data samples, failure mode type characterization elements and accident case type characterization elements in the domain data set to obtain sample labels of failure-prone samples; and determines the failure characterization law of the artificial intelligence system based on the failure characterization element values ​​and sample labels. This application constructs various types of failure characterization elements of the artificial intelligence system to be analyzed, and then analyzes and determines the corresponding failure characterization element values ​​and sample labels of failure-prone samples in combination with the domain data set of the artificial intelligence system, constructs failure characterizations around the domain data set, environment, failure mode and accident case, and uses data mining technology to obtain failure characterization laws based on the failure characterization values ​​in the domain data set, which can effectively improve the accuracy of the analysis of the failure laws of artificial intelligence systems.

[0057] In an exemplary embodiment, Figure 4 As shown, step 302 includes:

[0058] Step 401, determine the technical field type of the artificial intelligence system to be analyzed.

[0059] Step 403, determining the failure characterization element extraction method and basic system information of the artificial intelligence system based on the technical field type.

[0060] Step 405 , performing element extraction processing on the basic system information according to the failure characterization element extraction method to obtain various types of failure characterization elements of the artificial intelligence system.

[0061] The technical field type refers to the application field of the artificial intelligence system to be analyzed, such as specific fields such as images, text or audio. Basic system information refers to the basic information of the artificial intelligence system, such as the application field of the artificial intelligence system, the model category used, etc. Since different application fields or model categories focus on different fields and have different failure characterization elements, it is necessary to determine the corresponding failure characterization elements based on the basic system information.

[0062] Exemplarily, the key point of this application is the analysis and extraction operation of various types of failure characterization elements of the artificial intelligence system to be analyzed. When extracting failure characterization elements, a set of corresponding failure characterization element extraction methods can be summarized for artificial intelligence systems of different technical field types to extract failure characterization elements, and the extraction process of failure characterization elements should be analyzed in combination with the actual basic system information. Therefore, after obtaining the technical field type of the artificial intelligence system, the failure characterization element extraction method and basic system information of the artificial intelligence system can be determined based on the technical field type; then, the basic system information is subjected to element extraction processing according to the failure characterization element extraction method to obtain various types of failure characterization elements of the artificial intelligence system. After that, the specific artificial intelligence system is combined for analysis to determine the failure characterization law of the artificial intelligence system from various types of failure characterization elements. In this embodiment, the failure characterization element extraction method and basic system information are determined by the technical field type of the artificial intelligence system to be analyzed, and then the element extraction processing is performed to obtain various types of failure characterization elements of the artificial intelligence system, which can effectively improve the extraction accuracy of various types of failure characterization elements.

[0063] In an exemplary embodiment, the failure characterization element extraction method includes an environmental element extraction method. Performing element extraction processing on basic system information according to the failure characterization element extraction method to obtain environmental feature characterization elements of the artificial intelligence system includes: determining environmental impact factors of the artificial intelligence system based on the type of technical field; converting the environmental impact factors into quantitative data of environmental factors; performing feature extraction processing on the quantitative data of environmental factors to obtain environmental feature characterization elements of the artificial intelligence system.

[0064] Among them, environmental factors refer to environmental factors that may have a potential impact on the prediction results of the artificial intelligence system, such as weather, lighting, shadows, textures, geographical features, race, age and other high-level information in the picture, text source, language style, application field and other information of the text; such as the overall recording environment assessment of the audio, the typical environmental conditions when the user uses the voice system, etc. The quantitative data of environmental factors are quantifiable feature data converted from environmental factors, such as quantitative features such as light intensity, color contrast, and skin color value in the picture. For example, the text source in the text can be divided into "social media", "news", "academic", etc., and the audio scene can evaluate the overall recording environment through the quietness of the recording scene, the size of the space, and whether there is an echo. For the representation elements of environmental characteristics, they are the data obtained by feature extraction processing of environmental factor quantitative data. For example, for picture data, image processing techniques such as edge detection and color space conversion can be used to identify and extract environmental features. For text data, it is possible to apply text analysis techniques such as keyword extraction and syntactic analysis to convert environmental factors into quantifiable features. For audio data, frequency analysis and time series analysis of sound signals can be used to extract key environmental features.

[0065] Exemplarily, for environmental feature characterization elements, the environmental impact factors of the artificial intelligence system can be determined based on the type of technical field; then the environmental impact factors are converted into environmental factor quantitative data; finally, the environmental factor quantitative data is subjected to feature extraction processing to obtain environmental feature characterization elements of the artificial intelligence system, thereby ensuring the effect of element extraction. In this embodiment, the environmental feature characterization elements of the artificial intelligence system are determined by quantitative analysis of environmental impact factors, which can effectively ensure the analysis effect of environmental feature characterization elements.

[0066] In one embodiment, the present application is applied to the extraction process of failure characterization elements of environmental feature types in images. At this time, it is necessary to first identify and define environmental factors that may affect the prediction performance of artificial intelligence models. Specifically, the impact of environmental factors can be analyzed from multiple dimensions, including research on specific environmental variables that may exist in different application fields. For example, in a vehicle recognition system, environmental factors may include different weather conditions, road types, traffic density, vehicle types and colors, etc. Then the environmental factors defined in the previous step are converted into numbers or category labels that can be understood by the computer. For example, light intensity can be quantified by a light meter, color contrast can be calculated by image analysis software, and skin color values ​​can be determined by skin detection algorithms. These quantification methods need to consider accuracy, repeatability, and whether they can be automated. Then use advanced image processing technology to extract features from the quantified environmental factors. This may include applying edge detection algorithms to identify object contours, using color space conversion technology to analyze color distribution in images, using texture analysis to evaluate detailed features of the environment, and extracting spatial layout and visual depth features in images. This step requires quality assessment of the extracted features to ensure their effectiveness and accuracy. The algorithm and parameters for feature extraction need to be adjusted and optimized according to the specific application scenario. Finally, the extracted environmental features can be integrated into the failure characterization library. When updating the library, the newly added features need to be classified, labeled, and stored to ensure that they can be correctly identified and called. At the same time, it is also necessary to ensure that all newly added features have appropriate metadata descriptions, such as source, extraction date, extraction method, etc.

[0067] In one embodiment, the present application is applied to the extraction process of failure characterization elements of environmental feature types in texts. At this time, it is possible to first determine which external environmental factors may directly or indirectly affect the performance of the text processing model. This includes an investigation of the source of the text (such as social media, news reports, academic articles), as well as an analysis of the language style of the text (such as formal, informal, technical language) and application fields (such as finance, medical, and law). This step requires the classification of texts of different sources and styles and the evaluation of their potential impact on model performance. For example, social media texts may contain a large amount of non-standard language and slang, which may pose a challenge to the parsing ability of natural language processing models. The qualitative environmental description is then converted into numerical or categorical data that can be used for model analysis. For example, the source of the text can be classified as "social media", "news", "academic", etc., and the corresponding labels can be assigned; the language style can be quantified by indicators such as vocabulary diversity and sentence complexity; and the application field can be classified according to keywords and topics. The quantification process may require the use of a text classification algorithm to automatically identify and tag text data. Use text analysis techniques such as keyword extraction and syntactic analysis to extract the quantified environmental factor features. Keyword extraction can reveal the main content and themes of the text, while syntactic analysis helps understand the structure and language complexity of the text. These features help to gain a deeper understanding of the nature and potential challenges of text data, which is crucial for predicting the performance of models in different environments. Finally, the extracted environmental features can also be integrated into the failure characterization library. When updating the library, the newly added features need to be classified, labeled, and stored to ensure that they can be correctly identified and called. At the same time, it is also necessary to ensure that all new features have appropriate metadata descriptions, such as source, extraction date, extraction method, etc.

[0068] In one embodiment, the present application is applied to the extraction process of failure characterization elements of environmental features in audio. At this time, it is necessary to first determine the key environmental factors that affect the performance of the speech recognition system. This includes the evaluation of the overall recording environment (for example, recording studio, open space, vehicle interior, etc.) and the typical environmental conditions when users usually use the voice system (such as indoor quiet environment, outdoor noisy environment, etc.). This step needs to consider the propagation characteristics of sound in different environments and possible interference factors, such as indoor echo, outdoor wind noise, etc., and how these factors affect the quality and recognition accuracy of the voice signal. Then the environmental factors are converted into specific and quantifiable features. For example, the quietness of the recording scene can be quantified by the decibel level, the size of the space can be evaluated by the duration and intensity of the echo, and the presence or absence of the echo can be judged by the effect of the acoustic echo cancellation technology. The purpose of quantification is to more objectively evaluate the impact of environmental factors on speech recognition and provide practical metrics for subsequent data analysis. Then apply acoustic analysis technology to extract key environmental features. Through a detailed analysis of the frequency, the characteristics of the environmental noise and its impact on the voice signal can be revealed; and the time series analysis of the sound signal helps to understand the change pattern of the voice under different environmental conditions. The focus of this step is to extract key acoustic features that can reflect the impact of the environment on speech recognition. Finally, the extracted environmental features are integrated into the failure characterization library. When updating the library, the newly added features need to be classified, labeled, and stored to ensure that these features can be correctly identified and called. At the same time, it is also necessary to ensure that all newly added features have appropriate metadata descriptions, such as source, extraction date, extraction method, etc.

[0069] In an exemplary embodiment, the failure characterization element extraction method includes a data type element extraction method; step 302 includes: performing semantic analysis processing on the artificial intelligence system based on the technical field type to obtain a semantic analysis result; performing data feature extraction processing on the semantic analysis result to obtain a data feature type characterization element.

[0070] Exemplarily, semantic analysis processing refers to the process of performing corresponding data semantic analysis on the data involved in the artificial intelligence system. For example, for the semantic analysis of image data, deep learning models can be applied to perform entity recognition, image resolution and color space analysis and other data semantic analysis. For the semantic analysis of text data, natural language processing models, such as the bidirectional encoder representation model (BERT) based on transformers, the generative pre-trained transformer model (GPT), etc., are applied to perform semantic analysis in text, including entity recognition, sentiment analysis, topic detection, etc. For the semantic analysis of audio, speech recognition technology can be applied, such as the semantic information in speech based on models such as convolutional neural networks (CNN) and recurrent neural networks (RNN), including speaker recognition, keyword detection, sentiment analysis, etc. For the analysis and processing process of data feature representation elements, it can be quantified by semantic analysis. First, the artificial intelligence system is semantically analyzed based on the type of technical field to obtain the semantic analysis result; then the semantic analysis result is subjected to data feature extraction processing to obtain the data feature representation element. In this embodiment, the data feature representation element is constructed by semantic analysis and feature extraction processing of the data, which can effectively improve the accuracy of the recognition of the representation element.

[0071] In one of the embodiments, for the construction process of the feature-class representation elements of the image data, pre-trained or self-trained deep learning models, such as convolutional neural networks (CNNs), Transformer models, etc., can be used to perform data semantic analysis content such as entity recognition in the image and detailed analysis of image quality. For example, various entities can be identified on different data sets based on pre-trained deep learning models. The analysis of image resolution needs to determine the clarity and detail level of the image, while the analysis of color space should evaluate whether the color distribution of the image conforms to the color range of the real world. Focus should be placed on subtle differences in the image, such as color changes and detail capture under different lighting conditions. Based on the image semantic analysis, quantitative feature information about the image content needs to be further extracted. This includes but is not limited to the labels of entities (e.g., vehicles, pedestrians, traffic signs, etc.), the size and shape of entities, and the spatial relationship between entities (e.g., a pedestrian is located in front of a vehicle). In addition, the contextual information and quality information of the image should also be considered, such as the general category of the scene (e.g., city streets, suburban roads), the resolution and quality of the image (for low-resolution or high-noise images, the difficulty of recognition may increase), etc. The extraction of all these features should be automated and ensure that they reflect the key semantic information in the image sample. Finally, the extracted data features are integrated into the failure characterization library. When updating the library, the newly added features need to be classified, annotated, and stored to ensure that they can be correctly identified and called. At the same time, it is also necessary to ensure that all newly added features have appropriate metadata descriptions, such as source, extraction date, extraction method, etc.

[0072] In one embodiment, for the construction process of the feature-class representation elements of text-class data, it is necessary to first use a natural language processing (NLP) model, such as BERT or GPT, to conduct an in-depth semantic analysis of the text. This includes but is not limited to entity recognition (for example, identifying names of people, places, and organization names), sentiment analysis (determining the sentiment tendency of the text, such as positive, negative, and neutral), and topic detection (identifying the main topic or content area of ​​the text). In text semantic analysis, it should be ensured that the language characteristics in a specific data set can be accurately processed, such as correctly understanding the terminology and expressions unique to texts in specific fields such as medical or legal documents. Quantitative feature information about the content of the text is further extracted from the results of the text semantic analysis. This includes analysis of sentence structure (such as sentence length and complexity), frequency of vocabulary use (identifying high-frequency words and professional terms), etc. The purpose of this step is to understand the composition of the text from a micro perspective and how these constituent elements may affect the performance of the model. For example, an overly complex sentence structure may cause difficulties in parsing for some NLP models, and the frequency of use of specific vocabulary may reveal deviations in specific types of text or specific fields. The extraction of all these features should be automated and ensure that they can reflect the key semantic information in the text sample. Finally, the extracted data features should be integrated into the failure representation library. When updating the library, the newly added features need to be classified, labeled, and stored to ensure that they can be correctly identified and called. At the same time, it is also necessary to ensure that all new features have appropriate metadata descriptions, such as source, extraction date, extraction method, etc.

[0073] In one embodiment, for the construction process of the feature representation elements of audio data, it is necessary to first use advanced speech recognition technology, such as models based on convolutional neural networks (CNN), recurrent neural networks (RNN), and transformers, to analyze the semantic information contained in the speech. This includes identification of the speaker, detection of keywords, and analysis of the speaker's emotional state. Speaker identification not only involves confirming the speaker's identity, but may also include identification of his gender, age, and even emotional state. Keyword detection aims to identify important words and phrases in speech, which is crucial to understanding the main content of the speech. Emotional analysis attempts to interpret the speaker's emotional state from aspects such as intonation, speaking speed, and emphasis, which is very important in certain application scenarios, such as customer service or mental health analysis. Then, key data features are extracted from the results of speech semantic analysis. These features include speech rhythm (i.e., speaking speed and rhythm changes), pitch (changes in high and low tones), and volume (loudness of speech). The purpose of extracting these features is to understand the characteristics of speech in more detail and how they affect the effect of speech recognition and processing. For example, changes in volume may affect the clarity of the speech signal, while changes in rhythm and pitch may reflect the speaker's emotional state or emphasized content. Finally, the extracted environmental features are integrated into the failure characterization library. When updating the library, the newly added features need to be classified, labeled, and stored to ensure that they can be correctly identified and called. At the same time, it is also necessary to ensure that all newly added features have appropriate metadata descriptions, such as source, extraction date, extraction method, etc.

[0074] In an exemplary embodiment, the failure characterization element extraction method includes a failure mode element extraction method. Step 302 includes: based on the type of technical field, performing system analysis processing on the artificial intelligence system through the failure mode and effect analysis method to determine the key failure mode of the artificial intelligence system; performing feature mapping processing on the key failure mode to obtain the failure mode characterization element of the artificial intelligence system.

[0075] Exemplarily, the failure mode and effects analysis (FMEA) refers to a systematic activity in which the subsystems and parts that constitute the product and each process of the constituent process are analyzed one by one during the product design stage and the process design stage to find all potential failure modes and analyze their possible consequences, so as to take necessary measures in advance to improve the quality and reliability of the product. For image-type data, the failure mode and effects analysis method can identify key failure modes, such as defocus, motion blur, compression distortion, etc. For text-type data, the failure mode and effects analysis method can identify possible failure modes in text processing, such as information loss, misinterpretation, contextual errors, etc. For audio-type data, the failure mode and effects analysis method can identify possible failure modes in speech processing, such as echo and feedback problems, intermittent speech signals, etc. Feature mapping processing refers to the process of mapping the identified failure mode to specific failure characterization elements. For failure mode characterization elements, the failure mode and effects analysis method can be used for analysis. Specifically, based on the type of technical field, the artificial intelligence system can be systematically analyzed and processed by the failure mode and effects analysis method to determine the key failure modes of the artificial intelligence system. Then, feature mapping is performed based on the determined key failure modes to obtain the failure mode characterization elements of the artificial intelligence system. In this embodiment, the failure mode and effect analysis method is used to implement system analysis and processing of the artificial intelligence system, which can effectively improve the accuracy of failure mode analysis.

[0076] In one embodiment, for the construction process of image failure mode characterization elements, failure mode and effect analysis can be used to identify key failure modes that may affect the accuracy of artificial intelligence models. First, an interdisciplinary team of experts, including software engineers, data scientists, domain experts, etc., is organized to discuss and determine a list of potential failure modes. Then, a qualitative and quantitative risk assessment of these modes is performed, considering their likelihood, severity, and probability of detection. This requires a meticulous analysis process that may involve past accident cases, experimental data, and expert experience. The ultimate goal is to establish a priority list to clarify which failure modes have the most significant impact on model performance. In the process of failure mode identification, failure modes introduced by image capture devices and processing algorithms, such as defocus, motion blur, compression distortion, etc., should be considered. Then, the identified failure modes are converted into specific failure characterization elements that can be quantified. For example, a failure mode may be "the sensor causes overexposure of the image due to excessive light", which will be mapped to a corresponding failure characterization element, such as a numerical range of light intensity. When performing feature mapping, it is necessary to ensure the accuracy and interpretability of the mapping logic so that these failure modes can be accurately identified and addressed in subsequent analysis. Finally, the failure characterization elements of the failure mode mapping are integrated into the failure characterization library. When updating the library, the newly added features need to be classified, labeled, and stored to ensure that they can be correctly identified and called. At the same time, it is also necessary to ensure that all newly added features have appropriate metadata descriptions, such as source, extraction date, extraction method, etc.

[0077] In one embodiment, for the construction process of the text-based failure mode characterization element, the failure mode and effect analysis method can be used to identify the failure modes that may occur during the text processing process. Typical failure modes include information loss (key information is not correctly identified or extracted), misinterpretation (text content is misunderstood or classified), context error (ignoring or misunderstanding the context information of the text), etc. In order to identify these modes, it is necessary to organize an interdisciplinary team, including natural language processing experts, data scientists, and application domain experts, to jointly analyze and determine potential risks and their impacts. In this process, different types of text data and multiple possible processing scenarios should be considered to ensure comprehensive coverage of potential failure modes. Then the identified failure mode is converted into a specific failure characterization element that can be quantified. For example, if the identified failure mode is "misinterpretation", it may be mapped to a specific language feature, such as the frequency of use of ambiguous words, complex sentence structure, etc. When performing feature mapping, it is necessary to ensure the accuracy and interpretability of the mapping logic so that these failure modes can be accurately identified and dealt with in subsequent analysis. Finally, the failure characterization elements of the failure mode mapping are integrated into the failure characterization library. When updating the library, the newly added features need to be classified, labeled and stored to ensure that they can be correctly identified and called. At the same time, it is also necessary to ensure that all new features have appropriate metadata descriptions, such as source, extraction date, extraction method, etc.

[0078] In one embodiment, for the construction process of audio failure mode characterization elements, it is necessary to first use the failure mode and effect analysis method to identify various failure modes that may occur in the speech processing process. For example, echo and feedback problems may cause the quality of the speech signal to deteriorate, affecting the accuracy of speech recognition; and the discontinuity of the speech signal may be caused by network instability or recording quality problems. This step involves the cooperation of an interdisciplinary team, including acoustic engineers, data scientists, and user experience experts, to jointly evaluate the probability and potential impact of various failure modes. In this process, different types of speech data and multiple possible processing scenarios should be considered to ensure comprehensive coverage of potential failure modes. The identified failure modes are then converted into specific failure characterization elements. For example, if the failure mode is an echo problem, the corresponding failure characterization elements may include the intensity, duration, and frequency range of the echo. The purpose of this step is to concretize abstract problems so that they can be more easily identified and solved in actual systems. When performing feature mapping, it is necessary to ensure the accuracy and interpretability of the mapping logic so that these failure modes can be accurately identified and dealt with in subsequent analysis. Finally, the failure characterization elements of the failure mode mapping are integrated into the failure characterization library. When updating the library, the newly added features need to be classified, labeled and stored to ensure that they can be correctly identified and called. At the same time, it is also necessary to ensure that all new features have appropriate metadata descriptions, such as source, extraction date, extraction method, etc.

[0079] In an exemplary embodiment, the failure characterization element extraction method includes an accident case type element extraction method. Step 302 includes: obtaining accident cases related to the artificial intelligence system based on the technical field type; performing data analysis and processing on the accident cases to determine the key failure elements corresponding to the accident cases; and determining the accident case type characterization elements of the artificial intelligence system based on the key failure elements.

[0080] Exemplarily, for the case analysis process of the artificial intelligence system, a specific analysis can be conducted in combination with the type of technical field to which the artificial intelligence system belongs, such as traffic accidents caused by misjudgment of image recognition in the image field, incorrect text interpretation and improper information filtering in the text field, and voice assistant misunderstanding of instructions and erroneous operation of the automatic speech recognition system in the audio field. Then, the key failure factors are determined by data analysis of these accident cases, and the accident case characterization factors of the artificial intelligence system are determined based on the key failure factors. In this embodiment, the extraction and feature extraction of key failure factors are performed through data analysis of accident cases, so that the accident case characterization factors related to the system can be accurately determined, thereby ensuring the effectiveness of subsequent failure law analysis.

[0081] In one embodiment, for the construction process of the image-based accident case characterization elements, historical cases of image-based artificial intelligence system accidents can be collected from multiple channels. These channels may include industrial reports, academic papers, news reports, safety accident databases, etc. In the process of collecting cases, attention should be paid not only to the diversity of cases to ensure that accidents in different industries and different application scenarios are covered, but also to the level of detail of the cases, giving priority to those cases that provide sufficient technical details and accident analysis, such as traffic accidents caused by image recognition misjudgment. In addition, case collection should not be limited to accidents that have occurred, but can also include near-accident cases, that is, those events that almost led to disasters but were avoided in time. After the cases are collected, they need to be sorted and classified in preparation for subsequent analysis. Deeply understand each accident case and identify the key elements that led to the failure. This usually involves a multi-faceted analysis of the environment in which the accident occurred, the system status, user behavior, etc., so as to more accurately identify which elements may have caused the failure. This process is iterative and multidisciplinary, and may require the joint efforts of experts such as software engineers, data analysts, and system engineers, who will jointly review the evidence and discuss different failure scenarios to reach a consensus. Finally, based on the case analysis, specific failure characterization elements are identified from the accident. It may include properties of software components (such as software version, artificial intelligence model information), operational elements (such as user input, software configuration), and environmental conditions (such as weather conditions, working hours). When extracting these elements, it is important to distinguish between causes and results, clarify which elements are the direct causes of the accident, which are indirect causes, which are the consequences of the accident, and how these elements are related to image processing and analysis. These failure elements will become the basis for updating the failure characterization library. And integrate the failure characterization elements extracted from the accident into the failure characterization library. When updating the library, the newly added features need to be classified, labeled and stored to ensure that these features can be correctly identified and called. At the same time, it is also necessary to ensure that all new features have appropriate metadata descriptions, such as source, extraction date, extraction method, etc.

[0082] In one embodiment, for the construction process of the characterization elements of text-based accident cases, it is also necessary to first collect and organize accident cases related to text processing, which may include incorrect text interpretation, improper information filtering, misclassification of natural language processing systems, etc. The sources of the cases may be public reports, industry analysis reports, academic research, etc. When collecting cases, attention should be paid to the diversity of the cases to ensure that different types of text processing systems and application scenarios can be covered. The collected cases need to include detailed background information, such as text type, processing scenario, user feedback, etc., to facilitate subsequent in-depth analysis. Then, each collected case is analyzed in depth to identify the key factors that lead to text processing failure. This includes analyzing the input data, processing logic, output results, and differences from the expected results of the system. Case analysis needs to comprehensively consider factors at the technical and application levels, such as the limitations of the algorithm, the quality and representativeness of the data, the accuracy of context understanding, etc. In addition, the perspective of user interaction should also be considered, such as whether the user may misunderstand the output or feedback of the system. And identify specific failure characterization elements from the analyzed accident cases. This includes elements identified from a technical perspective, such as incorrect keyword extraction, inaccurate sentiment analysis results, and elements identified from a user interaction perspective, such as user misunderstanding of system output. When extracting these elements, it is necessary to record in detail their specific manifestations, the context in which they appear, and the possible impact. Finally, the failure characterization elements extracted from the accident are integrated into the failure characterization library. When updating the library, the newly added features need to be classified, labeled, and stored to ensure that these features can be correctly identified and called. At the same time, it is also necessary to ensure that all new features have appropriate metadata descriptions, such as source, extraction date, extraction method, etc.

[0083] In one embodiment, for the construction process of the audio accident case class characterization element, it is necessary to first collect accident cases related to the voice processing system, such as voice assistants misunderstanding user instructions or incorrect operations of automatic voice recognition systems. These cases can come from user feedback, technical support records, online forums, product reviews, etc. The collection of cases is intended to cover a wide range of scenarios and conditions to ensure a comprehensive understanding of the problems that voice systems may encounter in various practical applications. In addition, a detailed description of the case is crucial for subsequent analysis, so as much background information and specific details as possible need to be recorded. It is also necessary to conduct in-depth analysis of the collected cases to identify the key factors that cause the failure of the voice system. This includes analyzing the specific environment in which the accident occurred, the user's interaction method, the system's response, and any related technical issues. For example, when analyzing a case in which a voice assistant misunderstands an instruction, factors such as the speaker's accent, background noise, speaking speed and intonation need to be considered. The purpose of case analysis is to reveal the root cause behind the problem in order to identify key areas that can be improved. And identify specific failure characterization elements from the analyzed cases. This includes extracting information from both the technical and user interaction levels. For example, if a case involves a voice assistant misunderstanding a command in a noisy environment, the relevant failure factors may include the type and intensity of background noise, the clarity of the voice signal, and the sensitivity of the voice recognition algorithm to interference. When extracting these factors, it is necessary to record in detail their specific manifestations, the context in which they appear, and the possible impact. Finally, the failure characterization elements extracted from the accident are integrated into the failure characterization library. When updating the library, the newly added features need to be classified, labeled, and stored to ensure that these features can be correctly identified and called. At the same time, it is also necessary to ensure that all new features have appropriate metadata descriptions, such as source, extraction date, extraction method, etc.

[0084] In an exemplary embodiment, step 304 includes: identifying environmental feature type representation elements and data feature type representation elements of data samples in a domain data set; predicting the data samples in the domain data set through multiple pre-training artificial intelligence systems to obtain data sample prediction results; determining inconsistent target data samples in the data sample prediction results, and determining the values ​​of failure representation elements of the artificial intelligence system based on the target data samples.

[0085] Among them, the pre-trained artificial intelligence system refers to the model obtained by training the artificial intelligence system in the initial state with a large amount of model training data. These models can be used as benchmark models to analyze and process the rules of failure characterization factors.

[0086] Exemplarily, for the analysis process of environmental feature type characterization elements and data feature type characterization elements, the environmental feature type characterization elements and data feature type characterization elements of each data sample in the domain data set can be identified first, and then multiple pre-trained artificial intelligence systems are used as benchmark models to predict the data samples in the domain data set to obtain data sample prediction results. If the data sample prediction results are consistent, it means that the sample is not invalid. When there is an inconsistent target data sample in the data sample prediction result, it means that the prediction of the target data sample has failed, and the value of the target data sample can be used as the value of the invalid characterization element of the artificial intelligence system. In this embodiment, the benchmark model is obtained by pre-training to predict the data samples in the domain data set, so as to effectively determine the value of the invalid characterization element of the data sample, which can effectively improve the accuracy of the recognition of the value of the invalid characterization element.

[0087] In an exemplary embodiment, step 306 includes: performing data enhancement processing on data samples based on failure mode characterization elements and accident case characterization elements to obtain an enhanced sample set; predicting the data samples in the enhanced sample set by an artificial intelligence system obtained through multiple pre-trainings to obtain enhanced sample prediction results; determining inconsistent failure-prone samples in the enhanced sample data sample prediction results to obtain sample labels for failure-prone samples.

[0088] Data augmentation is a technique used to increase the amount of data by adding slightly modified copies of existing data or synthetic data newly created from existing data. In this application, it is a processing method that modifies the data accordingly based on the failure mode characterization elements and accident case characterization elements.

[0089] Exemplarily, for the analysis process of the failure mode characterization elements and the accident case characterization elements, for the data enhancement process, a new data sample can be generated based on the failure mode and accident case applicable to the data sample, so as to determine which data samples will fail due to the data enhancement processing of the failure mode and accident case. Similarly, after the enhanced sample set is constructed, multiple pre-trained artificial intelligence systems can be used as benchmark models to predict the data samples in the enhanced sample set to obtain the enhanced sample prediction results; determine the inconsistent easy-to-fail samples in the enhanced sample data sample prediction results, and obtain the sample labels of the easy-to-fail samples. Thus, the failure law analysis results for the failure mode characterization elements and the accident case characterization elements are obtained. In this embodiment, by processing the data samples based on the failure mode characterization elements and the accident case characterization elements, data enhancement is achieved to obtain easy-to-fail samples, and then according to the sample labels of the easy-to-fail samples, the failure mode characterization elements and the accident case characterization elements are implemented, which can effectively ensure the accuracy of the analysis of the failure mode characterization elements and the accident case characterization elements.

[0090] In one embodiment, the application flow chart of the artificial intelligence system failure characterization analysis method in the model development process can be referred to Figure 5 As shown in the figure, the failure characterization analysis process of the artificial intelligence system corresponds to the model development process. For the analysis process, we can first use four methods based on environmental description, data sample semantic information extraction, failure mode library, and accident case analysis to build an artificial intelligence software system failure characterization element library. These failure characterization elements describe environmental characteristics (such as geographical or weather conditions), data characteristics (typical entities and relationships between entities in the data set), failure modes (such as data pipelines, human-computer interaction, natural disturbances, external attacks, and other failure types) and accident cases (specific combinations of environmental characteristics, data characteristics, and failure modes), which are easy for testers to understand. Then, based on the failure characterization element library, the failure characterization law library is obtained by analyzing the model training set data. Then, the test case of the model is constructed in combination with the test set, thereby effectively improving the application effect of the model from the perspective of model testing.

[0091] It should be understood that, although the steps in the flowcharts involved in the above embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0092] Based on the same inventive concept, the embodiment of the present application also provides an artificial intelligence system failure characterization analysis device for implementing the artificial intelligence system failure characterization analysis method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more artificial intelligence system failure characterization analysis device embodiments provided below can refer to the limitations of the artificial intelligence system failure characterization analysis method above, and will not be repeated here.

[0093] In an exemplary embodiment, Figure 6 As shown, a device for characterizing and analyzing failure of an artificial intelligence system is provided, comprising:

[0094] The data acquisition module 601 is used to obtain the domain data set of the artificial intelligence system to be analyzed and various types of failure characterization elements, where the failure characterization elements include environmental feature characterization elements, data feature characterization elements, failure mode characterization elements, and accident case characterization elements.

[0095] The first analysis module 603 is used to determine the value of the failure characterization factor of the artificial intelligence system based on the environmental characteristic characterization factors and data characteristic characterization factors of the data samples in the domain data set.

[0096] The second analysis module 605 is used to construct easy-to-fail samples based on the data samples in the domain data set, the failure mode characterization elements and the accident case characterization elements, and obtain sample labels of the easy-to-fail samples.

[0097] The rule analysis module 607 is used to determine the failure characterization rule of the artificial intelligence system according to the failure characterization factor values ​​and sample labels.

[0098] In one embodiment, the data acquisition module 601 is specifically used to: determine the technical field type of the artificial intelligence system to be analyzed; determine the failure characterization element extraction method and basic system information of the artificial intelligence system based on the technical field type; perform element extraction processing on the basic system information according to the failure characterization element extraction method to obtain various types of failure characterization elements of the artificial intelligence system.

[0099] In one embodiment, the data acquisition module 601 is specifically used to: determine the environmental impact factors of the artificial intelligence system based on the type of technical field; convert the environmental impact factors into quantitative data of environmental factors; perform feature extraction processing on the quantitative data of environmental factors to obtain environmental feature characterization elements of the artificial intelligence system.

[0100] In one of the embodiments, the data acquisition module 601 is specifically used to: perform semantic analysis on the artificial intelligence system based on the technical field type to obtain semantic analysis results; perform data feature extraction on the semantic analysis results to obtain data feature class representation elements.

[0101] In one embodiment, the data acquisition module 601 is specifically used to: based on the type of technical field, perform system analysis and processing on the artificial intelligence system through the failure mode and effect analysis method to determine the key failure modes of the artificial intelligence system; perform feature mapping processing on the key failure modes to obtain the failure mode class representation elements of the artificial intelligence system.

[0102] In one embodiment, the data acquisition module 601 is specifically used to: acquire accident cases related to the artificial intelligence system based on the technical field type; perform data analysis and processing on the accident cases to determine the key failure factors corresponding to the accident cases; and determine the accident case class characterization factors of the artificial intelligence system based on the key failure factors.

[0103] In one embodiment, the first analysis module 603 is specifically used to: identify environmental feature type representation elements and data feature type representation elements of data samples in a domain data set; obtain an artificial intelligence system through multiple pre-trainings to predict data samples in the domain data set to obtain data sample prediction results; determine inconsistent target data samples in the data sample prediction results, and determine the failure representation element values ​​of the artificial intelligence system based on the target data samples.

[0104] In one embodiment, the second analysis module 605 is specifically used to: perform data enhancement processing on data samples based on failure mode characterization elements and accident case characterization elements to obtain an enhanced sample set; obtain an artificial intelligence system through multiple pre-trainings to predict data samples in the enhanced sample set to obtain enhanced sample prediction results; determine inconsistent failure-prone samples in the enhanced sample data sample prediction results to obtain sample labels for failure-prone samples.

[0105] Each module in the above-mentioned artificial intelligence system failure characterization and analysis device can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0106] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Figure 7 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data related to the failure characterization analysis of the artificial intelligence system. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for failure characterization analysis of an artificial intelligence system is implemented.

[0107] Those skilled in the art will understand that Figure 7 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0108] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0109] In one embodiment, a computer-readable storage medium is provided, storing a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0110] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program includes computer instructions, the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the steps in the above-mentioned method embodiments.

[0111] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0112] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0113] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0114] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for characterizing and analyzing failure of an artificial intelligence system, characterized in that: The method comprises: Acquire a domain data set of an artificial intelligence system to be analyzed, and determine the technical field type of the artificial intelligence system to be analyzed; determine a failure characterization element extraction method and basic system information of the artificial intelligence system based on the technical field type; perform element extraction processing on the basic system information according to the failure characterization element extraction method to obtain various types of failure characterization elements of the artificial intelligence system, wherein the failure characterization elements include environmental feature characterization elements, data feature characterization elements, failure mode characterization elements, and accident case characterization elements; Identify environmental feature type representation elements and data feature type representation elements of data samples in the domain data set; obtain the artificial intelligence system through multiple pre-training to predict the data samples in the domain data set to obtain data sample prediction results; determine inconsistent target data samples in the data sample prediction results, and determine the failure representation element value of the artificial intelligence system based on the target data samples; Based on the failure mode characterization elements and the accident case characterization elements, the data samples are subjected to data enhancement processing to obtain an enhanced sample set; the artificial intelligence system obtained through multiple pre-trainings predicts the data samples in the enhanced sample set to obtain enhanced sample prediction results; and inconsistent failure-prone samples in the enhanced sample data sample prediction results are determined to obtain sample labels of the failure-prone samples; The failure characterization law of the artificial intelligence system is determined according to the failure characterization factor value and the sample label.

2. The method according to claim 1, characterized in that The failure characterization factor extraction method includes an environmental factor extraction method; Performing element extraction processing on the basic system information according to the failure characterization element extraction method to obtain the environmental feature characterization elements of the artificial intelligence system includes: Determining environmental impact factors of the artificial intelligence system based on the technical field type; Converting the environmental impact factors into quantitative data of environmental factors; The quantified data of environmental factors are subjected to feature extraction processing to obtain environmental feature characterization elements of the artificial intelligence system.

3. The method according to claim 1, characterized in that The failure characterization element extraction method includes a data type element extraction method; Performing element extraction processing on the basic system information according to the failure characterization element extraction method to obtain the data feature characterization elements of the artificial intelligence system includes: Performing semantic analysis on the artificial intelligence system based on the technical field type to obtain a semantic analysis result; The semantic analysis result is subjected to data feature extraction processing to obtain data feature class representation elements.

4. The method according to claim 1, characterized in that: The failure characterization element extraction method includes a failure mode element extraction method; Performing element extraction processing on the basic system information according to the failure characterization element extraction method to obtain the failure mode characterization elements of the artificial intelligence system includes: Based on the technical field type, the artificial intelligence system is systematically analyzed and processed by using the failure mode and effect analysis method to determine the key failure modes of the artificial intelligence system; Perform feature mapping processing on the key failure mode to obtain failure mode class characterization elements of the artificial intelligence system.

5. The method according to claim 1, characterized in that: The failure characterization element extraction method includes an accident case element extraction method; The basic system information is subjected to element extraction processing according to the failure characterization element extraction method to obtain the accident case characterization elements of the artificial intelligence system, including: Based on the technical field type, obtain accident cases related to the artificial intelligence system; Performing data analysis and processing on the accident case to determine the key failure factors corresponding to the accident case; Based on the key failure factors, the accident case characterization factors of the artificial intelligence system are determined.

6. An artificial intelligence system failure characterization and analysis device, characterized in that: The device comprises: A data acquisition module is used to acquire a domain data set of an artificial intelligence system to be analyzed, and determine the type of technical field of the artificial intelligence system to be analyzed; determine a failure characterization element extraction method and basic system information of the artificial intelligence system based on the type of technical field; perform element extraction processing on the basic system information according to the failure characterization element extraction method to acquire various types of failure characterization elements of the artificial intelligence system, wherein the failure characterization elements include environmental feature characterization elements, data feature characterization elements, failure mode characterization elements, and accident case characterization elements; The first analysis module is used to identify the environmental characteristic type representation elements and data characteristic type representation elements of the data samples in the domain data set; obtain the artificial intelligence system through multiple pre-training to predict the data samples in the domain data set to obtain data sample prediction results; determine the inconsistent target data samples in the data sample prediction results, and determine the failure representation element value of the artificial intelligence system based on the target data samples; The second analysis module is used to perform data enhancement processing on the data samples based on the failure mode characterization elements and the accident case characterization elements to obtain an enhanced sample set; the artificial intelligence system obtained through multiple pre-trainings predicts the data samples in the enhanced sample set to obtain enhanced sample prediction results; determine the inconsistent failure-prone samples in the enhanced sample data sample prediction results to obtain sample labels of the failure-prone samples; A rule analysis module is used to determine the failure characterization rule of the artificial intelligence system according to the failure characterization factor value and the sample label.

7. The device according to claim 6, characterized in that The failure characterization element extraction method includes an environmental element extraction method; the data acquisition module is specifically used to: determine the environmental impact factors of the artificial intelligence system based on the technical field type; convert the environmental impact factors into environmental factor quantitative data; perform feature extraction processing on the environmental factor quantitative data to obtain the environmental feature type characterization elements of the artificial intelligence system.

8. The device according to claim 6, characterized in that The failure characterization element extraction method includes a data type element extraction method; the data acquisition module is specifically used to: perform semantic analysis processing on the artificial intelligence system based on the technical field type to obtain a semantic analysis result; perform data feature extraction processing on the semantic analysis result to obtain a data feature type characterization element.

9. The device according to claim 6, characterized in that The failure characterization element extraction method includes a failure mode element extraction method; the data acquisition module is specifically used to: based on the technical field type, perform system analysis and processing on the artificial intelligence system through the failure mode and effect analysis method to determine the key failure mode of the artificial intelligence system; perform feature mapping processing on the key failure mode to obtain the failure mode characterization element of the artificial intelligence system.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

Images