Edge computing based environmental context processing and model optimization method and system

Abstract

The present application relates to the technical field of data processing, and more particularly to an environment context processing and model optimization method and system based on edge computing, which stores initial data in each time period together with historical data in a target storage, guarantees the continuity and integrity of data in the time dimension, and fuses data of different sources and times to mine the association between current data and past environment states, more comprehensively grasps the state of the target physical place and each target object at the end of each time period, so that the target storage stores continuous historical data over time, facilitates understanding the dynamic changes of the target physical place and the target object in combination with context data, enables the edge world model to better adapt to the characteristics and changes of the target physical place, improves the adaptability between the edge world model and the target physical place, and improves the completion efficiency and completion effect of the target application task.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an environment context processing and model optimization method and system based on edge computing. BACKGROUND

[0002] With the rapid development of Internet of Things and artificial intelligence technology, environment context processing and model application are widely used in intelligent buildings, industrial manufacturing, intelligent transportation and many other fields. For AI application and multi-modal model, context environment information is the key to improve model effect. In the prior art, environment context data extraction and model optimization usually rely on cloud processing and common sense modeling, which is difficult to efficiently and timely perceive the local environment context and understand the unique rules of the local physical place, and difficult to efficiently integrate the independent data of each target object and the overall context data of the target physical place, resulting in poor quality of the obtained context data, making the trained model insufficiently adaptable to the dynamic changes of the target physical place, and unable to accurately reflect the real state and change trend of the physical place, which makes it difficult to provide reliable decision support in industrial production environment monitoring, intelligent traffic flow prediction and many other practical application scenarios.

[0003] Therefore, how to improve the adaptability between the edge world model and the target physical place, so as to improve the completion efficiency and completion effect of the target application task has become a problem to be solved. SUMMARY

[0004] In view of the above technical problems, the technical solution adopted by the present application is an environment context processing and model optimization method based on edge computing, comprising the following steps:

[0005] S1, generating the ID corresponding to each target object in the target physical place and the target physical place, and allocating the corresponding target storage location for each ID in the target storage.

[0006] S2, storing the initial data corresponding to each target object in the target physical place in the current time period into the corresponding target storage location, wherein the target storage also stores the first historical context data corresponding to the target physical place and the second historical context data corresponding to each target object before the start of the current time period, and the current time period refers to a continuous time period with the current time as the end point and the time length T.

[0007] S3, according to the initial data corresponding to each target object in the current time period, the first historical context data corresponding to the target physical place and the second historical context data corresponding to each target object before the start of the current time period, the first reference context data corresponding to the target physical place and the second reference context data corresponding to each target object at the end of the current time period are obtained.

[0008] S4, the first reference context data corresponding to the target physical site and the second reference context data corresponding to each target object at the end of the current time period are stored as the first historical context data corresponding to the target physical site and the second historical context data corresponding to each target object before the start of the next time period, and are stored in the corresponding target storage location.

[0009] S5, according to the first historical context data and the second historical context data stored in the target storage before the preset model training time, the preset world model of the cloud is model distilled at the preset model training time, and an edge world model corresponding to the target physical site is obtained.

[0010] S6, according to the first historical context data and the second historical context data stored in the target storage and the edge world model, a target application interface is called to execute a target application task corresponding to the target application interface.

[0011] The application also provides an environment context processing and model optimization system based on edge computing, comprising:

[0012] An ID generation module is configured to generate IDs corresponding to the target physical site and each target object in the target physical site, and allocate a corresponding target storage location for each ID in the target storage.

[0013] A first data storage module is configured to store initial data corresponding to each target object in the target physical site in the current time period in the corresponding target storage location, wherein the target storage also stores first historical context data corresponding to the target physical site and second historical context data corresponding to each target object before the start of the current time period, and the current time period refers to a continuous time period with the current time as the end point and a time length of T.

[0014] A data processing module is configured to obtain first reference context data corresponding to the target physical site and second reference context data corresponding to each target object at the end of the current time period according to the initial data corresponding to each target object in the current time period, the first historical context data corresponding to the target physical site and the second historical context data corresponding to each target object before the start of the current time period.

[0015] A second data storage module is configured to store the first reference context data corresponding to the target physical site and the second reference context data corresponding to each target object at the end of the current time period as the first historical context data corresponding to the target physical site and the second historical context data corresponding to each target object before the start of the next time period, and store them in the corresponding target storage location.

[0016] The model distillation module is configured to perform model distillation on the preset world model of the cloud at a preset model training time according to the first historical context data and the second historical context data stored in the target memory before the preset model training time, and obtain the edge world model corresponding to the target physical site.

[0017] The task execution module is configured to invoke the target application interface to execute a target application task corresponding to the target application interface according to the first historical context data and the second historical context data stored in the target memory and the edge world model.

[0018] The present application has at least the following advantages: by assigning a unique ID to each target physical site and target object and assigning a corresponding storage location, the data can be organized in order in the target memory, improving the efficiency and accuracy of subsequent data storage, query and update, by storing the initial data in the current time period together with the second historical context data in the target memory, the continuity and integrity of the data in the time dimension are guaranteed, which helps to establish the association between the current data and the past environment state, and to fuse data from different sources and different times, which can fully exploit the advantages of various types of data, and help to mine potential rules and trends in the data, so as to more comprehensively and accurately grasp the state of the target physical site and each target object at the end of the current time period, further enabling the target memory to store continuous historical data over time, facilitating understanding of the dynamic changes of the target physical site and the target object in combination with the context data, improving the characterization degree of the target physical site, enabling subsequent model training to be based on more comprehensive and longer-term data samples, enabling the edge world model trained based on the historical data of the target physical site to better adapt to the characteristics and changes of the target physical site, thereby improving the adaptability between the edge world model and the target physical site, and further improving the completion efficiency and completion effect of the target application task. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0020] Figure 1 A flowchart of an environment context processing and model optimization method based on edge computing is provided for the first embodiment of the present application.

[0021] Figure 2A flowchart for acquiring initial data in an environment context processing and model optimization method based on edge computing provided for embodiment one of the present application;

[0022] Figure 3 A flowchart for acquiring initial collected data in an environment context processing and model optimization method based on edge computing provided for embodiment one of the present application;

[0023] Figure 4 A flowchart for acquiring second reference context data in an environment context processing and model optimization method based on edge computing provided for embodiment one of the present application;

[0024] Figure 5 A flowchart for acquiring first reference context data in an environment context processing and model optimization method based on edge computing provided for embodiment one of the present application;

[0025] Figure 6 A flowchart for acquiring edge world model in an environment context processing and model optimization method based on edge computing provided for embodiment one of the present application;

[0026] Figure 7 A flowchart for acquiring first target context data, second target context data and target scene feature in an environment context processing and model optimization method based on edge computing provided for embodiment one of the present application;

[0027] Figure 8 A flowchart for executing prediction control task in an environment context processing and model optimization method based on edge computing provided for embodiment one of the present application;

[0028] Figure 9 A flowchart for executing target query task in an environment context processing and model optimization method based on edge computing provided for embodiment one of the present application;

[0029] Figure 10 A flowchart for executing natural language interaction task in an environment context processing and model optimization method based on edge computing provided for embodiment one of the present application;

[0030] Figure 11 A module schematic diagram of an environment context processing and model optimization system based on edge computing provided for embodiment two of the present application. DETAILED DESCRIPTION

[0031] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0032] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It can be understood that the above-described terms for distinguishing similar objects can be interchanged under appropriate circumstances, so that the present application can also be implemented in other embodiments in addition to the above-described illustrated embodiments or described embodiments. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to the process, method, product, or device.

[0033] Embodiment one

[0034] The embodiment one provides an edge computing-based environment context processing and model optimization method, which includes the following steps, as shown in Figure 1

[0035] S1, generating the ID corresponding to the target physical site and each target object in the target physical site, and allocating the corresponding target storage location for each ID in the target storage.

[0036] The target physical site refers to a specific physical space or system that needs to be processed and analyzed, such as a smart factory, a smart building, or a specific natural ecological area, which is the object range of data research and processing.

[0037] The target object refers to an individual or component with independent characteristics and behaviors in the target physical site, which can be a sensor, a device, a person, an object, a robot, etc. For example, in a smart factory, the target object can be a robot on the production line, various sensors, and transport vehicles, etc.

[0038] The target storage refers to a storage device or system for storing the data corresponding to the target physical site and each target object, which can be a database, a hard disk array, etc., providing a centralized management and storage space for data, facilitating subsequent data query, reading, and processing.

[0039] ​In order to ensure independent data storage and tracking of the target physical site and each physical object, a unique corresponding ID is generated for the target physical site and each target object, and a corresponding target storage location is allocated for each ID in the target storage, that is, a special area is demarcated for each ID in the target storage to store the relevant data of the target physical site or target object corresponding to each ID. In addition, the target storage usually adopts a specific data structure to manage the storage location, such as a hash table, an index tree, etc. Through the specific data structure, the corresponding storage location can be quickly found according to the ID, and the data storage, query and update can be more efficient and accurate.

[0040] In the embodiment, the ID can be a string of numbers, letters or a combination of numbers and letters, and can be generated by a Universally Unique Identifier (UUID) or Re-Identification (ReID) method.

[0041] By allocating a unique ID to each target physical site and target object and allocating a corresponding storage location, the data can be organized in order in the target storage, and the efficiency and accuracy of data storage, query and update can be improved.

[0042] S2, store the initial data of each target object in the target physical site in the corresponding target storage location within the current time period, wherein the target storage also stores the first historical context data corresponding to the target physical site before the start of the current time period and the second historical context data corresponding to each target object, and the current time period refers to a continuous time period with the current time as the end point and the time length T.

[0043] The current time period is a continuous time period with the current time as the end point and the time length T, and the specific value of T is determined according to the actual application scenario and demand, representing a specific time window for data collection and processing, for analyzing the changes and status of the physical site within a period of time. Through the change of the current time period, the target physical site and the target object can be regularly collected, processed, stored, updated, summarized and applied.

[0044] The initial data refers to the data collected or generated by each target object within the current time period, which is the basis for subsequent processing and analysis, and contains the original information of the target object within the time period, such as temperature, humidity, pressure, image data collected by the camera, etc.

[0045] The first historical context data refers to a series of background data and environmental information related to the target physical site before the start of the current time period, which describes the overall state and related conditions of the physical site in the past, such as the trend of environmental temperature change in the past period, the historical record of equipment operation state, etc., which helps to understand the evolution process of the current physical site.

[0046] The second historical context data corresponding to each target object refers to the data record generated by the target object in the past period before the start of the current time period, which reflects the historical behavior and characteristic change of the target object, such as the numerical sequence obtained by the sensor through multiple measurements in the past. It can be used to analyze the performance change and trend of the target object. The target memory has saved the first historical context data corresponding to the target physical site and the second historical context data corresponding to each target object before the start of the current time period.

[0047] As mentioned above, the target object data in the physical site changes over time, and by storing the initial data in the current time period together with the historical data in the target memory, the continuity and integrity of the data in the time dimension are ensured, and it is helpful to establish the association between the current data and the past environmental state, facilitate the mining of the association and change trend between the target objects and the physical site, at the same time, the continuous time series data can more accurately reflect the dynamic change process of the target object and the physical site, providing comprehensive data support for subsequent analysis and processing.

[0048] In a specific embodiment, a plurality of data acquisition devices are arranged in the target physical site, each data acquisition device corresponding to a plurality of target objects in the target physical site, such as Figure 2 As shown in FIG. 2, S2 includes the following steps:

[0049] S21, obtaining the initial acquisition data of each target object corresponding to each data acquisition device in the current time period.

[0050] S22, performing vectorization processing on the initial acquisition data of each target object in the target physical site corresponding to the current time period, to obtain the initial vector feature corresponding to each target object in the current time period.

[0051] S23, performing semantic summarization on the initial acquisition data of each target object in the target physical site corresponding to the current time period, to obtain the initial semantic description corresponding to each target object in the current time period.

[0052] S24, performing rule summarization on the initial acquisition data, initial vector feature and initial semantic description corresponding to each target object in the current time period, to obtain the local rule data corresponding to each target object in the current time period.

[0053] S25, taking the initial acquisition data, initial vector features, initial semantic description, local regularity data and the current time period of each target object as the initial data of each target object in the current time period.

[0054] Among them, the data acquisition device can be various sensors such as temperature sensors, humidity sensors, light sensors, pressure sensors, video acquisition devices such as cameras, audio acquisition devices such as microphones, and detection devices such as radars. The data acquisition device is deployed in the target physical site for real-time monitoring and acquisition of relevant data of the target object corresponding thereto. In the current time period, the data acquisition device continuously collects data at a certain sampling frequency to reflect the state and change of the target object in the time period.

[0055] In machine learning and data analysis, vectors are a data representation method that is convenient for computer processing and analysis. For the initial acquisition data of each target object in the current time period, the initial acquisition data is converted into initial vector features through a specific vector conversion method, which facilitates the computer to process the data in a unified and standardized form, thereby improving the efficiency and accuracy of data processing.

[0056] The embodiment converts the initial acquisition data of each modality through a text embedding model, an image embedding model, and a multi-modal embedding model. The text embedding model is used to convert the initial acquisition data of the text modality into a low-dimensional vector space representation, extract the semantic and syntactic features of the initial acquisition data, and facilitate the computer to process and analyze the text. The image embedding model is used to convert the initial acquisition data of the image modality into a low-dimensional vector representation, extract the image features, and facilitate the computer to process and analyze the image. The multi-modal embedding model is used to fuse the embedding models of text, image, audio and other multi-modal data, capture the correlation and complementary information between different modal data, and facilitate the computer to better understand and process multi-modal data.

[0057] Semantic summarization involves natural language processing and semantic understanding. Specifically, the initial acquisition data is converted into natural language text description through semantic analysis and conversion of the acquisition data, which facilitates the combination of data and domain knowledge and provides more rich environmental data. Semantic summarization can be realized by using a large language model to generate appropriate text description according to the characteristics of the data. For example, in the intelligent factory scenario, the temperature sensor collects temperature data corresponding to the warehouse shelf as 22℃, and the large language model can generate the corresponding initial semantic description as "The temperature of the shelf is maintained at 22℃, which is conducive to the storage of goods and avoids damage to goods due to excessive or low temperature."

[0058] The local rule refers to rule information with specificity and locality mined by analyzing and processing the initial collection data, the initial vector feature and the initial semantic description of each target object in the current time period, and used to reflect the characteristics and behavior model of the corresponding target object. The local rule is comprehensively analyzed and summarized from multiple dimensions to fully depict the local rule of the target object.

[0059] The above, the initial data is obtained by comprehensively analyzing the initial collection data, the initial vector feature, the initial semantic description, the local rule data and the current time period, so that the initial data has a more comprehensive and multi-dimensional representation, and the state of the physical place and the target object can be more comprehensively understood, thereby providing a rich data basis for deep mining and comprehensive utilization of data.

[0060] In a specific embodiment, as shown in Figure 3 S21 includes the following steps:

[0061] S211, according to the business scene type corresponding to the target physical place, obtaining the preset data extraction rule corresponding to the target physical place.

[0062] S212, according to the preset data extraction rule, screening the original data collected by each data collection device in the current time period, and obtaining the initial collection data of each target object corresponding to each data collection device in the current time period.

[0063] Among them, people, things and robots under different business scene types have different business characteristics and needs, pay attention to different dimensions of business data, and the corresponding data extraction rules are different.

[0064] Therefore, the implementer can preform the preset data extraction rule matched with each business scene type according to the monitoring target, data characteristics and analysis requirements of each business scene type, and store it in the rule library or configuration file of the target storage, so as to obtain the preset data extraction rule corresponding to the target physical place from the target storage according to the business scene type corresponding to the target physical place, and use it to screen the original data collected by each data collection device in the current time period, so as to ensure that the extracted data matches the characteristics and needs of the target physical place, thereby improving the pertinence and effectiveness of the data.

[0065] For example, in a shopping mall retail scenario, consumers are more concerned about product prices, promotions, store location conditions, and efficient selection of desired products. Merchants are more concerned about product sales, inventory, and customer purchase preferences to optimize product display and marketing strategies. Mall managers are more concerned about customer flow, consumer stay time in each area, and store operation conditions to evaluate mall operation efficiency and plan commercial layout. Correspondingly, the data extraction rules are: for consumers, extract their action trajectory, stay in stores, browse products, and other data in the mall. For merchants, extract product sales records, inventory changes, customer purchase time and frequency, and other data. For mall managers, extract data such as customer flow at each entrance, real-time customer flow on each floor and in each store, and average customer stay time.

[0066] In a hospital inpatient department scenario, patients are more concerned about their own symptom changes, treatment plans, and rehabilitation progress. Medical staff are more concerned about patient vital signs, test results, and treatment reactions to adjust treatment measures. Hospital management is more concerned about bed usage, medical resource allocation, and department operation efficiency to optimize management decisions. Correspondingly, the data extraction rules are: for patients, extract their symptom description, rehabilitation self-evaluation data, and feedback on treatment. For medical staff, extract patient vital sign data such as heart rate, blood pressure, and body temperature, as well as examination reports and medication records. For hospital management, extract data such as bed occupancy rate in each department, medical equipment usage time, and drug consumption.

[0067] According to the above, the preset data extraction rules are obtained according to the type of the business scenario, and the initial collection data is filtered from the original collection data according to the preset data extraction rules, ensuring that the initial collection data extracted is more suitable for the type and needs of the target physical place, avoiding the extraction of irrelevant data, and when the type of the business scenario changes or the needs are adjusted, the initial collection data extracted can be optimized by modifying and updating the corresponding data extraction rules, without the need to modify each data collection device separately. The collection and processing, thereby improving the efficiency and accuracy of subsequent data analysis and processing.

[0068] S3, according to the initial data of each target object in the current time period, the first historical context data corresponding to the target physical place before the start of the current time period, and the second historical context data corresponding to each target object, the first reference context data corresponding to the target physical place at the end of the current time period and the second reference context data corresponding to each target object are obtained.

[0069] Among them, the initial data of each target object in the current time period reflects the real-time state of the target object in that period, the first historical context data reflects the overall condition of the target physical place in the past, and the second historical context data of each target object summarizes and records the past state of the target object.

[0070] The first reference context data of the target physical site at the end of the current time period and the second reference context data of each target object are obtained by integrating the initial data of each target object in the current time period, the first historical context data, and the second historical context data of each target object. Specifically, the first reference context data is a description of the overall state of the target physical site at the end of the current time period after comprehensively considering the current data and the second historical context data, and the second reference context data is the state data of each target object at the end of the current time period, which comprehensively considers the current performance and historical development of the target object.

[0071] By fusing data from different sources and at different times, the advantages of various types of data can be fully utilized, which helps to mine potential rules and trends in the data, so that the state of the target physical site and each target object at the end of the current time period can be more comprehensively and accurately grasped.

[0072] In a specific embodiment, as shown in Figure 4 S3 includes the following steps:

[0073] S31, for any target object in the target physical site, feature extraction is performed on the initial data corresponding to the current target object in the current time period to obtain the first initial feature corresponding to the current target object in the current time period.

[0074] S32, feature extraction is performed on the second historical context data corresponding to the current target object before the start of the current time period to obtain the first historical feature corresponding to the current target object before the start of the current time period.

[0075] S33, feature fusion is performed on the first initial feature corresponding to the current target object in the current time period and the first historical feature corresponding to the current target object before the start of the current time period to obtain the second reference context data corresponding to the current target object at the end of the current time period.

[0076] The initial data may contain a large amount of redundant information, and feature extraction can convert high-dimensional original data into low-dimensional feature vectors, reduce the amount of data while retaining key information, and improve the efficiency of subsequent processing.

[0077] The first initial feature reflects the current state of the target object, and the first historical feature reflects the long-term trend of the target object. By fusing the two, the complementarity of the data can be fully utilized, and the current state and historical development of the target object can be considered comprehensively to obtain the second reference context data corresponding to each target object, which is used to represent more comprehensive and accurate context information of each target object in the before and after time periods.

[0078] In a specific embodiment, as shown inFigure 5 As shown, S3 includes the following steps:

[0079] S34, feature extraction is performed on the initial data of all target objects in the target physical site corresponding to the current time period, to obtain the second initial feature corresponding to the target physical site in the current time period.

[0080] S35, feature fusion is performed on the second initial feature corresponding to the target physical site in the current time period and the first historical context data corresponding to the beginning of the current time period, to obtain the first reference context data corresponding to the target physical site at the end of the current time period.

[0081] Among them, the initial data of all target objects in the environment in the current time period is taken as a whole for feature extraction, so as to extract the key features that can represent the state of the entire target physical site in the current time period, i.e. the second initial feature.

[0082] The first historical context data contains the past state, trend and background information of the target physical site, and through the operation of feature fusion, the state features of the environment in the current time period are combined with the past historical information, so as to generate a more comprehensive and more capable of reflecting the overall state of the target physical site at the end of the current time period. The first reference context data.

[0083] S4, the first reference context data corresponding to the target physical site at the end of the current time period and the second reference context data corresponding to each target object are stored as the first historical context data corresponding to the target physical site and the second historical context data corresponding to each target object before the start of the next time period, and are stored in the corresponding target storage location.

[0084] Among them, the state of the target physical site and the target objects therein changes continuously over time, by converting the first reference context data and the second reference context data obtained at the end of the current time period into the second historical context data before the start of the next time period, and storing them in the corresponding target storage location of the target storage, to provide historical basis for data processing in the subsequent time period, to form a process of cyclic update and continuous accumulation of data, so as to better track and analyze the dynamic changes of the target physical site and each target object over time, to discover the trend, periodicity and abnormal situation of the target physical site based on the context data changing over time. Dynamic information such as.

[0085] S5, according to the first historical context data and the second historical context data stored by the target storage before the preset model training time, model distillation is performed on the preset world model of the cloud at the preset model training time, to obtain the edge world model corresponding to the target physical site.

[0086] The preset model training time is a time node preset by an implementer and used to determine the data range for model training. Correspondingly, the first historical context data and the second historical context data stored in the target storage before this time will be used to train the model.

[0087] The cloud is a remote server cluster or computing platform based on cloud computing technology, which has strong computing and storage capabilities and can process large-scale data and complex computing tasks. In the model training process, the cloud is used to provide a preset world model and related computing resources.

[0088] The preset world model is a pre-constructed general model framework that has been pre-trained on a large amount of general data and has learned some general features and patterns. The preset world model can be used as a base model, and based on the first historical context data and the second historical context data stored in the target storage before the preset model training time, the preset world model of the cloud can be post-trained. By combining the general features and patterns learned by the preset world model with the characteristic historical data of the target physical site, the preset world model learns the characteristics, trends and laws of the target physical site, and teaches the simpler and lighter edge world model in a transferable manner, and finally obtains an edge world model that fits the target physical site.

[0089] Correspondingly, the number of parameters of the preset world model is greater than the number of parameters of the edge world model corresponding to the edge world model, and the edge world model can more accurately reflect the characteristics and change rules of the target physical site, and can provide strong support for subsequent environmental monitoring, prediction and decision-making tasks. The edge world model is more suitable for edge devices such as smartphones, Internet of Things sensors and intelligent cameras, and has more advantages in hardware purchase cost and energy consumption. When deploying artificial intelligence applications on a large scale, it can greatly reduce the server hardware cost and operating cost.

[0090] The present embodiment migrates the knowledge of the complex and large preset world model to the simpler and lighter edge world model through model distillation. Those skilled in the art know that the specific model distillation process in the prior art falls within the scope of the present application, and will not be described here.

[0091] The target memory stores continuous historical data over time, which facilitates understanding the dynamic changes of the target physical site and the target object in combination with the context data, improves the characterization degree of the target physical site, enables subsequent model distillation to be based on more comprehensive and longer-term data samples, and enables the edge world model trained based on the historical data of the target physical site to better adapt to the characteristics and changes of the target physical site compared to the general preset world model, thereby improving the adaptation degree between the edge world model and the target physical site, making the prediction and analysis results in the target physical site more accurate, and being able to provide more reliable basis for decision-making.

[0092] In a specific embodiment, as shown in Figure 6 S5 includes the following steps:

[0093] S51, according to the business scene type corresponding to the target physical site, obtaining the basic scene feature corresponding to the target physical site.

[0094] S52, according to the basic scene feature, reducing the dimension of the first historical context data and the second historical context data stored by the target memory before the preset model training time, obtaining the first target context data, the second target context data and the target scene feature corresponding to the target physical site.

[0095] S53, taking the first target context data, the second target context data and the target scene feature corresponding to the target physical site as training data, performing model distillation on the preset world model of the cloud at the preset model training time, obtaining the edge world model corresponding to the target physical site, wherein the parameter quantity of the preset world model is greater than the parameter quantity corresponding to the edge world model.

[0096] The basic scene feature is a key feature extracted for a specific business scene type, which can reflect the basic scene and main attributes of the environment. For example, in an intelligent factory, the basic scene feature may include the device running parameter range, the key node of the production process, etc. In an intelligent home, it may be the normal range of indoor temperature and humidity, the common working mode of household appliances, etc.

[0097] Different business scene types have different characteristics and operation rules. By classifying and identifying the business scene types, the corresponding basic scene features are extracted, which can provide targeted guidance for subsequent data processing and model training, make the processing process more consistent with the actual situation of the target physical site, and improve the data utilization efficiency and the accuracy of the model.

[0098] The first historical context data and the second historical context data can contain a large amount of redundant information and noise, and high-dimensional data not only increases the calculation cost, but also can affect the training effect of the model. By dimension reduction, the key information in the data can be highlighted, and the data is more concise and easy to process. At the same time, combined with the basic scene features for dimension reduction, important information related to the target physical site can be ensured to be retained, and the processed data is more in line with the needs of actual application.

[0099] The first target context data, the second target context data and the target scene feature obtained after dimension reduction are integrated as training data and input into the preset world model in the cloud, and by model distillation on the preset world model, an edge world model that can accurately reflect the characteristics of the target physical site is finally obtained.

[0100] In a specific embodiment, as shown in Figure 7 S52 includes the following steps:

[0101] S521, for any preset index corresponding feature value in the first historical context data and the second historical context data stored by the target memory before the preset model training time, calculate the feature difference value between the current preset index corresponding feature value and the basic scene feature.

[0102] S522, statistics of the feature difference values corresponding to all preset indexes in the first historical context data and the second historical context data stored by the target memory before the preset model training time are obtained.

[0103] S523, according to the feature difference distribution diagram, all preset indexes are screened to obtain a plurality of target indexes.

[0104] S524, the target memory stores the first historical context data, the second historical context data and the basic scene feature before the preset model training time, and determines the target index corresponding data as the first target context data, the second target context data and the target scene feature corresponding to the target physical site.

[0105] Among them, for each preset index, the difference value between the corresponding feature value and the basic scene feature is calculated. For example, in the intelligent factory scene, the preset indexes can be device temperature, production efficiency, etc., and the normal temperature range of the device in the basic scene feature is an interval. The minimum difference value between the current device temperature feature value and the upper and lower limits of the normal temperature range is the corresponding feature difference value.

[0106] The feature difference value can find the part deviating from the typical feature in the first historical context data and the second historical context data, and the feature difference value often contains special information of the target physical site, such as abnormal situation, change trend, etc., and the feature difference distribution diagram can intuitively show the distribution of the difference, providing a basis for subsequent index screening.

[0107] According to the result of the feature difference distribution diagram, all preset indexes are screened. Specifically, the embodiment selects indexes with larger feature difference values or special distribution rules as target indexes, because the selected target indexes can better reflect the uniqueness and change of the target physical site, and are more valuable for subsequent model training and analysis. For example, in the intelligent agricultural scene, if the feature difference value of the soil humidity index is large, it means that the index changes significantly at different times or in different environments, and is selected as a target index.

[0108] The first historical context data, the second historical context data, and the data corresponding to the target index in the basic scene feature are extracted to determine the first target context data, the second target context data, and the target scene feature, respectively, realizing dimension reduction of the data, removing some unimportant information for describing the characteristics of the target physical site, making the data more refined and targeted, and thus improving the adaptability between the edge world model and the target physical site.

[0109] S6, according to the first historical context data and the second historical context data stored in the target memory and the edge world model, calling the target application interface to execute the target application task corresponding to the target application interface.

[0110] The target application interface is a set of pre-defined functions or methods for interacting with the target application, providing a standardized way for external systems to request the target application to perform specific tasks.

[0111] According to the information provided by the first historical context data, the second historical context data and the edge world model in the target memory, the specific task to be executed is determined, and the corresponding operation is triggered by calling the target application interface to execute the corresponding target application task to achieve a specific business goal.

[0112] In a specific embodiment, the target application task corresponding to the target application interface includes a behavior prediction control task, and the edge world model includes an edge prediction model, as shown in Figure 8 S6 includes the following steps:

[0113] S611, receiving a behavior prediction control request through the target application interface, wherein the behavior prediction control request includes a target control object and a prediction control requirement.

[0114] S612, acquire the first target data corresponding to the behavior prediction and control request from the first historical context data and the second historical context data stored in the target memory according to the behavior prediction and control request.

[0115] S613, input the first target data into the edge prediction model to acquire the behavior prediction data corresponding to the behavior prediction and control request, wherein the behavior prediction data is used to control the behavior of the target control object.

[0116] The behavior prediction and control request is received through the target application interface, so as to explicitly indicate that the operation related to the behavior prediction and control is needed. The target control object is specified in the behavior prediction and control request, i.e. the object that needs to be specifically predicted and controlled, such as a robot, a person or an intelligent household appliance, etc. Meanwhile, the specific requirements of the behavior prediction and control of the target control object are indicated in the behavior prediction and control request, for example, predicting and controlling the position, moving track, running state, etc. of the target control object at a certain time in the future.

[0117] The first target data corresponding to the behavior prediction and control request is filtered from the large amount of first historical context data and second historical context data stored in the target memory. The edge prediction model analyzes the first target data according to the learned rules and patterns, so as to predict the future behavior of the target control object under the condition of meeting the prediction and control requirements, and output the behavior prediction data for controlling the future behavior of the target control object. For example, for a robot, the edge prediction model can predict the moving speed, position change, etc. of the robot in the next period of time according to the past motion data and the environmental information of the target physical space, so as to control the moving speed and track of the robot.

[0118] Based on the rich first historical context data and second historical context data in the target memory, the potential rules and influencing factors behind the behavior of the target control object are mined through the edge prediction model, so that the prediction and control process of the target control object is more scientific and reasonable.

[0119] In a specific embodiment, the target application task corresponding to the target application interface further includes a target query task, such as Figure 9 As shown in FIG. 6, S6 further includes the following steps:

[0120] S621, receive a target query request through the target application interface, wherein the target query request includes a target query object, a target query time and a target query condition.

[0121] S622, index the first historical context data and the second historical context data stored in the target memory.

[0122] S623, obtaining target query data corresponding to the target query request from the first historical context data and the second historical context data stored in the target memory according to the index.

[0123] The target query object indicates a specific object that the user wants to query, for example, in a company, it can be a certain employee or a certain product. The target query time limits the time range of the query data, for example, the user may want to query the work record of a certain employee in the past month, or the sales data of a certain product in a certain quarter, etc. The target query condition further refines the query requirements, for example, when querying the work record of an employee, the condition may be that the working time exceeds a certain standard, the performance score reaches a certain level, etc. When querying product sales data, the condition may be that the sales amount is within a certain interval, the sales area meets a certain range, etc.

[0124] Since the amount of the first historical context data and the second historical context data stored in the target memory can be very large, in order to improve the efficiency of data query, an index needs to be established for these data to record the storage location of the data in the memory and some key feature information of the data. Among them, those skilled in the art know that the index establishment method in the prior art falls within the protection scope of the present application. For example, for time series data, an index can be established according to time order, so as to quickly locate the storage location of the related data when querying the data in a certain time range. For classification data, an index can be established according to categories, so as to query data of a certain category.

[0125] After the index is established, according to the target query object, the target query time and the target query condition in the target query request, the index is used to quickly locate and filter the data meeting the requirements, which can avoid traversing the entire data set, and greatly improves the speed of data retrieval.

[0126] As described above, by establishing an index, traversal of the entire data set can be avoided, efficient and accurate data query function is realized, the query demand of the user can be quickly responded, and accurate query results can be provided.

[0127] In a specific embodiment, the target application task corresponding to the target application interface further includes a natural language interaction task, and the edge world model further includes an edge language interaction model, as shown in Figure 10 S6 further includes the following steps:

[0128] S631, receiving a natural language text input by a target user through a target application interface.

[0129] S632, obtaining a target problem corresponding to the target user according to the natural language text.

[0130] S633, obtaining the second target data corresponding to the target request from the first historical context data and the second historical context data stored in the target memory according to the target question.

[0131] S634, inputting the second target data into the edge language interaction model to obtain a target reply corresponding to the target question, wherein the target reply corresponds to a natural language form.

[0132] The natural language text can be text input by the user through a keyboard, or text information converted after voice recognition.

[0133] The received natural language text is processed and analyzed through technologies such as word segmentation, part-of-speech tagging, syntax analysis, and semantic understanding, and the core question that the target user really wants to ask or express, i.e., the target question, is extracted from the natural language text. The second target data related to the target question is filtered from the large amount of first historical context data and second historical context data stored in the target memory. The second target data can include solutions to similar questions as the target question, user feedback information, product-related instructions, etc. For example, if the target question is about the solution to product failure, the relevant information such as the processing records of other users encountering the same or similar failure, the troubleshooting manual of the product, etc. can be found from the historical data. By obtaining the second target data, the user's question can be answered by using the past experience and knowledge.

[0134] The edge language interaction model combines the language patterns and knowledge learned by itself to understand and analyze the input second target data, and generates a target reply in the natural language form corresponding to the target question. For example, the edge language interaction model may generate a reply similar to "You can first check if the power connection of the product is normal. If the power supply is not a problem, you can try restarting the product. If the problem still exists, please contact our after-sales service personnel, the contact information is XXX" according to the previous failure processing records, and provide it to the target user to complete a natural language interaction task.

[0135] The above, by processing the natural language text input by the user, accurately understanding the user's question, and generating a suitable reply by using historical data and an edge language interaction model, the function of natural language interaction is realized, and the communication efficiency and experience between the user and the system are improved.

[0136] As described above, by assigning a unique ID to each target physical location and target object and allocating a corresponding storage location, data can be organized in an orderly manner in the target memory, improving the efficiency and accuracy of subsequent data storage, querying, and updating. By storing initial data within the current time period together with historical data in the target memory, the continuity and integrity of data in the time dimension are guaranteed, which helps to establish the correlation between current data and past environmental states. Furthermore, by fusing data from different sources and at different times, the advantages of various types of data can be fully utilized, which helps to uncover potential patterns and trends in the data. This allows for a more comprehensive and accurate understanding of the state of the target physical location and each target object at the end of the current time period. In addition, the target memory stores continuous historical data over time, which facilitates understanding the dynamic changes of the target physical location and target objects in conjunction with contextual data, improving the representation of the target physical location. This allows subsequent model distillation to be based on more comprehensive and longer-term data samples, enabling the edge world model trained based on the historical data of the target physical location to better adapt to the characteristics and changes of the target physical location compared to a general preset world model. This improves the adaptability between the edge world model and the target physical location, thereby enhancing the efficiency and effectiveness of the target application task.

[0137] Example 2

[0138] This second embodiment provides an environment context processing and model optimization system based on edge computing, such as... Figure 11 As shown, the edge computing-based environment context processing and model optimization system includes:

[0139] ID generation module 111 is used to generate the target physical location and the ID corresponding to each target object in the target physical location, and to allocate a corresponding target storage location for each ID in the target storage.

[0140] The first data storage module 112 is used to store the initial data corresponding to each target object in the target physical location within the current time period to the corresponding target storage location. The target storage also stores the first historical context data corresponding to the target physical location before the start of the current time period and the second historical context data corresponding to each target object. The current time period refers to a continuous time period with the current time as the end point and a time length of T.

[0141] The data processing module 113 is used to obtain the first reference context data corresponding to the target physical location and the second reference context data corresponding to each target object at the end of the current time period based on the initial data corresponding to each target object in the current time period, the first historical context data corresponding to the target physical location before the start of the current time period, and the second historical context data corresponding to each target object.

[0142] The second data storage module 114 is configured to store the first reference context data corresponding to the target physical site and the second reference context data corresponding to each target object at the end of the current time period as the first historical context data corresponding to the target physical site and the second historical context data corresponding to each target object before the start of the next time period, and store the first historical context data and the second historical context data in the corresponding target storage location.

[0143] The model distillation module 115 is configured to perform model distillation on the preset world model of the cloud at a preset model training time according to the first historical context data and the second historical context data stored in the target storage before the preset model training time, and obtain the edge world model corresponding to the target physical site.

[0144] The task execution module 116 is configured to invoke the target application interface to execute the target application task corresponding to the target application interface according to the first historical context data and the second historical context data stored in the target storage and the edge world model.

[0145] In an embodiment, a plurality of data collection devices are arranged in the target physical site, each data collection device corresponding to a plurality of target objects in the target physical site, and the first data storage module 112 includes:

[0146] The data collection sub-module is configured to obtain initial collection data of each target object corresponding to each data collection device in the target physical site within the current time period.

[0147] The vector conversion sub-module is configured to perform vectorization processing on the initial collection data of each target object in the target physical site within the current time period, and obtain initial vector features of each target object within the current time period.

[0148] The semantic summary sub-module is configured to perform semantic summarization on the initial collection data of each target object in the target physical site within the current time period, and obtain initial semantic descriptions of each target object within the current time period.

[0149] The rule summary sub-module is configured to perform rule summarization on the initial collection data, the initial vector features, and the initial semantic descriptions of each target object within the current time period, and obtain local rule data of each target object within the current time period.

[0150] The initial data acquisition sub-module is configured to store the initial collection data, the initial vector features, the initial semantic descriptions, the local rule data, and the current time period of each target object within the current time period as initial data of each target object within the current time period.

[0151] In an embodiment, the data collection submodule comprises:

[0152] The extraction rule acquisition unit is configured to acquire a preset data extraction rule corresponding to the target physical site according to a business scenario type corresponding to the target physical site.

[0153] The data screening unit is configured to screen the original data collected by each data collection device in the current time period according to the preset data extraction rule, and acquire initial collection data of each target object corresponding to each data collection device in the current time period.

[0154] In an embodiment, the data processing module 113 comprises:

[0155] The first initial feature acquisition submodule is configured to perform feature extraction on the initial data corresponding to the current target object in the current time period, and acquire a first initial feature corresponding to the current target object in the current time period.

[0156] The first historical feature acquisition submodule is configured to perform feature extraction on the second historical context data corresponding to the current target object before the start of the current time period, and acquire a first historical feature corresponding to the current target object before the start of the current time period.

[0157] The first feature fusion submodule is configured to perform feature fusion on the first initial feature corresponding to the current target object in the current time period and the first historical feature corresponding to the current target object before the start of the current time period, and acquire second reference context data corresponding to the current target object at the end of the current time period.

[0158] In an embodiment, the data processing module 113 comprises:

[0159] The second initial feature acquisition submodule is configured to perform feature extraction on the initial data corresponding to all target objects in the target physical site in the current time period, and acquire a second initial feature corresponding to the target physical site in the current time period.

[0160] The second feature fusion submodule is configured to perform feature fusion on the second initial feature corresponding to the target physical site in the current time period and the first historical context data corresponding to the target physical site before the start of the current time period, and acquire first reference context data corresponding to the target physical site at the end of the current time period.

[0161] In an embodiment, the model distillation module 115 comprises:

[0162] The basic scene feature acquisition submodule is configured to acquire a basic scene feature corresponding to the target physical site according to a business scenario type corresponding to the target physical site.

[0163] a data dimension reduction submodule configured to reduce dimensions of first historical context data and second historical context data stored by the target memory before the preset model training time according to the basic scene feature, to obtain first target context data, second target context data, and the target scene feature corresponding to the target physical site.

[0164] a model distillation submodule configured to perform model distillation on a preset world model of the cloud at the preset model training time according to the first target context data, the second target context data, and the target scene feature corresponding to the target physical site as training data, to obtain an edge world model corresponding to the target physical site, wherein a number of parameters of the preset world model is greater than a number of parameters corresponding to the edge world model.

[0165] In an embodiment, the data dimension reduction submodule includes:

[0166] a feature difference value obtaining unit configured to calculate a feature difference value between a feature value corresponding to a preset index and the basic scene feature for any preset index corresponding to a feature value in the first historical context data and the second historical context data stored by the target memory before the preset model training time.

[0167] a feature difference distribution diagram obtaining unit configured to statistically obtain feature difference values corresponding to all preset indexes in the first historical context data and the second historical context data stored by the target memory before the preset model training time, to obtain a feature difference distribution diagram.

[0168] a target index screening unit configured to screen all preset indexes according to the feature difference distribution diagram, to obtain a plurality of target indexes.

[0169] a data screening unit configured to determine data corresponding to the target indexes in the first historical context data, the second historical context data, and the basic scene feature stored by the target memory before the preset model training time as the first target context data, the second target context data, and the target scene feature corresponding to the target physical site.

[0170] In an embodiment, the target application interface corresponds to a behavior prediction control task, the edge world model includes an edge prediction model, and the task execution module 116 includes:

[0171] a first request receiving submodule configured to receive a behavior prediction control request through the target application interface, wherein the behavior prediction control request includes a target control object and a prediction control requirement.

[0172] The first target data acquisition submodule is configured to acquire, according to the behavior prediction control request, first target data corresponding to the behavior prediction control request from the first historical context data and the second historical context data stored in the target storage.

[0173] The behavior prediction submodule is configured to input the first target data into the edge prediction model to acquire behavior prediction data corresponding to the behavior prediction control request, wherein the behavior prediction data is used to control the behavior of the target control object.

[0174] In an embodiment, the target application interface corresponds to a target application task, and the task execution module 116 further includes:

[0175] The second request receiving submodule is configured to receive a target query request through the target application interface, wherein the target query request includes a target query object, a target query time and a target query condition.

[0176] The index establishing submodule is configured to establish an index for the first historical context data and the second historical context data stored in the target storage.

[0177] The target query data acquisition submodule is configured to acquire, according to the index, target query data corresponding to the target query request from the first historical context data and the second historical context data stored in the target storage.

[0178] In an embodiment, the target application interface corresponds to a target application task, and the task execution module 116 further includes:

[0179] The third request receiving submodule is configured to receive natural language text input by a target user through the target application interface.

[0180] The target question acquisition submodule is configured to acquire a target question corresponding to the target user according to the natural language text.

[0181] The second target data acquisition submodule is configured to acquire, according to the target question, second target data corresponding to the target request from the first historical context data and the second historical context data stored in the target storage.

[0182] The target reply acquisition submodule is configured to input the second target data into the edge language interaction model to acquire a target reply corresponding to the target question, wherein the target reply corresponds to a natural language form.

[0183] The data screening submodule is configured to determine data corresponding to the target index in the first historical context data, the second historical context data, and the basic scene feature as first target context data, second target context data, and a target scene feature corresponding to the target physical site.

[0184] It should be noted that the information interaction, execution process, and the like between the above modules are based on the same concept as the method embodiments of the present application, and the specific functions and the resulting technical effects can be referred to the method embodiments part. Therefore, no further description is given here.

[0185] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as above with a preferred embodiment, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content without departing from the scope of the technical solution of the present application, and any simple modification, equivalent change, and modification of the above embodiments according to the technical essence of the present application are still within the scope of the technical solution of the present application.

Claims

1. A method for environmental context processing and model optimization based on edge computing, characterized in that, The method includes the following steps: S1, Generate the target physical location and the ID corresponding to each target object in the target physical location, and allocate a corresponding target storage location for each ID in the target memory; S2, store the initial data corresponding to each target object in the target physical location within the current time period into the corresponding target storage location. The target storage also stores the first historical context data corresponding to the target physical location before the start of the current time period and the second historical context data corresponding to each target object. The current time period refers to a continuous time period with the current time as the end point and a time length of T. S3, based on the initial data corresponding to each target object in the current time period, the first historical context data corresponding to the target physical location before the start of the current time period, and the second historical context data corresponding to each target object, obtain the first reference context data corresponding to the target physical location and the second reference context data corresponding to each target object at the end of the current time period; S4, take the first reference context data corresponding to the target physical location at the end of the current time period and the second reference context data corresponding to each target object as the first historical context data corresponding to the target physical location and the second historical context data corresponding to each target object before the start of the next time period, and store them in the corresponding target storage location; S5, based on the first historical context data and the second historical context data stored in the target memory before the preset model training time, perform model distillation on the preset world model in the cloud during the preset model training time to obtain the edge world model corresponding to the target physical location. S6. Based on the first historical context data and the second historical context data stored in the target memory and the edge world model, the target application interface is invoked to execute the target application task corresponding to the target application interface.

2. The edge computing-based environment context processing and model optimization method according to claim 1, characterized in that, The target physical location is equipped with several data acquisition devices, each data acquisition device corresponding to several target objects in the target physical location. S2 includes the following steps: S21, Obtain the initial acquisition data of each target object corresponding to each data acquisition device in the current time period; S22, the initial collected data of each target object in the target physical location within the current time period is vectorized to obtain the initial vector features of each target object within the current time period; S23, Summarize the initial collected data of each target object in the target physical location within the current time period to obtain the initial semantic description of each target object within the current time period; S24, summarize the patterns of the initial collected data, initial vector features and initial semantic description of each target object in the current time period, and obtain the local pattern data of each target object in the current time period; S25, take the initial collected data, initial vector features, initial semantic description, local pattern data and current time period corresponding to each target object in the current time period as the initial data corresponding to each target object in the current time period.

3. The edge computing-based environment context processing and model optimization method according to claim 2, characterized in that, S21 includes the following steps: S211, Based on the business scenario type corresponding to the target physical location, obtain the preset data extraction rules corresponding to the target physical location; S212, according to the preset data extraction rules, the raw data collected by each data acquisition device in the current time period is filtered to obtain the initial acquisition data of each target object corresponding to each data acquisition device in the current time period.

4. The edge computing-based environment context processing and model optimization method according to claim 1, characterized in that, S3 includes the following steps: S31, for any target object in the target physical location, perform feature extraction on the initial data corresponding to the current target object in the current time period to obtain the first initial feature corresponding to the current target object in the current time period; S32, extract features from the second historical context data of the current target object before the start of the current time period to obtain the first historical features of the current target object before the start of the current time period; S33, perform feature fusion on the first initial feature corresponding to the current target object within the current time period and the first historical feature corresponding to the current time period before the start of the current time period to obtain the second reference context data corresponding to the current target object at the end of the current time period.

5. The edge computing-based environment context processing and model optimization method according to claim 1, characterized in that, S3 includes the following steps: S34, extract features from the initial data of all target objects in the target physical location within the current time period to obtain the second initial features of the target physical location within the current time period; S35, perform feature fusion on the second initial feature corresponding to the target physical location within the current time period and the first historical context data corresponding to the target physical location before the start of the current time period to obtain the first reference context data corresponding to the target physical location at the end of the current time period.

6. The edge computing-based environment context processing and model optimization method according to claim 1, characterized in that, S5 includes the following steps: S51, Based on the business scenario type corresponding to the target physical location, obtain the basic scenario features corresponding to the target physical location; S52, based on the basic scene features, the first historical context data and the second historical context data stored in the target memory before the preset model training time are reduced in dimensionality to obtain the first target context data, the second target context data and the target scene features corresponding to the target physical location; S53, using the first target context data, the second target context data, and the target scene features corresponding to the target physical location as training data, perform model distillation on the preset world model in the cloud during a preset model training time to obtain the edge world model corresponding to the target physical location, wherein the number of parameters in the preset world model is greater than the number of parameters in the edge world model.

7. The edge computing-based environment context processing and model optimization method according to claim 6, characterized in that, S52 includes the following steps: S521, for the feature value corresponding to any preset indicator in the first historical context data and the second historical context data stored in the target memory before the preset model training time, calculate the feature difference value between the feature value corresponding to the current preset indicator and the basic scene feature. S522, Statistical analysis is performed on the feature difference values ​​corresponding to all preset indicators in the first historical context data and the second historical context data stored in the target memory before the preset model training time to obtain a feature difference distribution map. S523, Based on the feature difference distribution map, all preset indicators are filtered to obtain several target indicators; S524, the first historical context data, the second historical context data, and the data corresponding to the target indicators in the basic scene features stored in the target memory before the preset model training time are determined as the first target context data, the second target context data, and the target scene features corresponding to the target physical location.

8. The edge computing-based environment context processing and model optimization method according to claim 1, characterized in that, The target application task corresponding to the target application interface includes a behavior prediction and control task, and the edge world model includes an edge prediction model. S6 includes the following steps: S611, Receive a behavior prediction control request through the target application interface, wherein the behavior prediction control request includes a target control object and prediction control requirements; S612, according to the behavior prediction control request, obtain the first target data corresponding to the behavior prediction control request from the first historical context data and the second historical context data stored in the target memory; S613, the first target data is input into the edge prediction model to obtain the behavior prediction data corresponding to the behavior prediction control request, wherein the behavior prediction data is used to control the behavior of the target control object.

9. The edge computing-based environment context processing and model optimization method according to claim 8, characterized in that, The target application task corresponding to the target application interface also includes a target query task, and S6 further includes the following steps: S621, Receive a target query request through the target application interface, wherein the target query request includes a target query object, a target query time, and target query conditions; S622, establish an index for the first historical context data and the second historical context data stored in the target memory; S623, according to the index, obtain the target query data corresponding to the target query request from the first historical context data and the second historical context data stored in the target memory.

10. The edge computing-based environment context processing and model optimization method according to claim 8, characterized in that, The target application task corresponding to the target application interface also includes a natural language interaction task, and the edge world model also includes an edge language interaction model. S6 also includes the following steps: S631, Receive natural language text input by the target user through the target application interface; S632, obtain the target question corresponding to the target user based on the natural language text; S633, based on the target problem, obtain the second target data corresponding to the target request from the first historical context data and the second historical context data stored in the target memory; S634, the second target data is input into the edge language interaction model to obtain the target response corresponding to the target question, wherein the target response corresponds to natural language form.

11. An environment context processing and model optimization system based on edge computing, characterized in that, The edge computing-based environment context processing and model optimization system includes: The ID generation module is used to generate the target physical location and the ID corresponding to each target object in the target physical location, and to allocate a corresponding target storage location for each ID in the target memory; The first data storage module is used to store the initial data corresponding to each target object in the target physical location within the current time period to the corresponding target storage location. The target storage also stores the first historical context data corresponding to the target physical location before the start of the current time period and the second historical context data corresponding to each target object. The current time period refers to a continuous time period with the current time as the end point and a time length of T. The data processing module is used to obtain the first reference context data corresponding to the target physical location and the second reference context data corresponding to each target object at the end of the current time period based on the initial data corresponding to each target object in the current time period, the first historical context data corresponding to the target physical location before the start of the current time period, and the second historical context data corresponding to each target object. The second data storage module is used to store the first reference context data corresponding to the target physical location at the end of the current time period and the second reference context data corresponding to each target object as the first historical context data corresponding to the target physical location and the second historical context data corresponding to each target object before the start of the next time period, and store them in the corresponding target storage location. The model distillation module is used to perform model distillation on the preset world model in the cloud during the preset model training time, based on the first historical context data and the second historical context data stored in the target memory before the preset model training time, to obtain the edge world model corresponding to the target physical location. The task execution module is used to call the target application interface to execute the target application task corresponding to the target application interface based on the first historical context data and the second historical context data stored in the target memory and the edge world model.

Images