Fire safety agent construction method based on large model

Abstract

According to the fire safety agent construction method based on the large model, sensor data, video stream data and historical data are subjected to fusion modeling by adopting a dynamic weighted tensor fusion analysis method to serve as a reasoning data source of the large model; combining the fused data with a fire-fighting knowledge base to form an input prompt of a large model through a template mode; and finally, analyzing and verifying the reasoning result of the large model, and adopting a function call mode to correspond to the platform capability, and automatically executing a disposal means. According to the fire safety agent construction method based on the large model, through deep combination of multi-source data fusion and the large model in the field of fire fighting, accurate diagnosis and automatic disposal of a fire event are realized, the problem of response delay caused by the fact that a traditional fire fighting system depends on manual research and judgment is solved, and emergency efficiency and safety are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fire safety, and more specifically, to a method for constructing a fire safety intelligent agent based on a large model. Background Art

[0002] In recent years, large language models (LLMs), represented by generative pre-trained models (such as GPT series, BERT, etc.), have demonstrated breakthrough capabilities in natural language processing, multimodal understanding, and complex reasoning tasks. These models achieve in-depth modeling of semantics, logic, and context associations through massive data training and a parameter scale of hundreds of billions. Developers can utilize the powerful reasoning ability of large models through prompt engineering for functions such as search, analysis, and decision-making. With the rapid development of large models, the application of intelligent agents (AI Agents) has become an effective means of implementation. As software entities with environmental perception, autonomous decision-making, and dynamic interaction capabilities, their core goal is to automate the execution of complex tasks through continuous learning and reasoning. Traditional intelligent agent technologies are limited by the expressive power of rule engines or small-scale models and are difficult to meet the processing requirements of highly dynamic and multi-source heterogeneous data in industrial scenarios. In contrast, intelligent agents based on large models (LLM-based Agents) significantly enhance the flexibility of task planning, anomaly diagnosis, and cross-modal decision-making by integrating the semantic understanding, logical reasoning, and generation capabilities of large language models. For example, large model-driven intelligent agents can analyze implicit patterns in sensor data streams, generate interpretable control instructions, and even collaborate with human operators through multi-round conversations to optimize strategies. Currently, the academic and industrial communities are conducting in-depth research on the modular architecture of intelligent agents (such as ReAct, AutoGPT), memory mechanisms, and lightweight deployment, promoting their implementation in fields such as industrial control and equipment operation and maintenance.

[0003] The Industrial Internet of Things (Industrial IoT, IIoT) constructs a real-time data acquisition and response network covering the entire production chain by connecting sensors, controllers, and execution devices. In critical fields such as fire safety, the system needs to continuously process high-frequency data streams (up to the TB / day level) from multi-modal devices such as smoke detectors, temperature sensors, and cameras, and trigger emergency mechanisms (such as sprinkler system activation and evacuation route control) based on low-latency reasoning (usually requiring a response in milliseconds to seconds). Research work (see Zhang, C., et al. "Edge Intelligence: Architectures, Challenges, and Applications." IEEE IoT Journal, 2020.) indicates that in a fire scenario, a local intelligent agent can dynamically integrate sensor data, historical event libraries, and emergency plan libraries, quickly identify the evolution trend of a fire through multi-modal reasoning, and generate optimal control instructions that take into account safety regulations and the real-time environment.

[0004] Based on existing research work, there is an urgent need to invent a method for constructing a fire safety intelligent agent to solve the problems that traditional Internet of Things technologies rely on a single type of data source and use a threshold + rule strategy for fire warning, requiring targeted parameter adjustment and having low accuracy in complex scenarios. Summary of the Invention

[0005] The present invention provides a method for constructing a fire safety intelligent agent based on a large model. On the basis of traditional fire diagnosis methods, the LLM large model is applied to integrate multi-source data for reasoning, and fire intelligent diagnosis and disposal are realized through multi-modal data fusion and dynamic analysis, improving the intelligent level of the fire protection system.

[0006] According to one aspect of the present invention, there is provided a method for constructing a fire safety intelligent agent based on a large model, including the following steps:

[0007] Step 1, obtain sensor data, video stream data, and historical data of fire protection, and use Concat to convert the above data into a unified four-dimensional spatio-temporal tensor to obtain fused data;

[0008] Step 2, for the fused data X, divide it according to different modalities to obtain sub-tensors X of M modalities (m) , where m = 1, 2,..., m, and calculate the covariance matrix C of each modal sub-tensor (m) for capturing the correlation between internal features of the m-th modality;

[0009] Step 3, based on the covariance matrix C (m) , obtain the subspace U of the m-th modality and the feature data X m through the principal component analysis method pca, obtain a globally optimal subspace on the Grassmann manifold to minimize the weighted sum of the distances between each modal subspace and the global subspace;

[0010] Step 4, calculate the variance σ of each principal component in X pca (i = 1, 2,.., n, where n is the original dimension of X i 2 ) and calculate the variance contribution rate r of each principal component pca ; Based on the variance contribution rate r i , accumulate the variance contribution rates in sequence until the cumulative variance contribution rate exceeds the threshold T, record the number k of principal components included at this time, and select the principal components that contain most of the information; i

[0011] Step 5, analyze the data after fusion in Step 1, clarify the actual physical meaning represented by each principal component, determine the correlation by viewing the element values of the eigenvectors corresponding to X pca , and construct indexes for different types of knowledge based on the fire knowledge base; sensor

[0012] Step 6, construct a prompt template according to different fire diagnosis scenarios and requirements, corresponding to the classification of the fire knowledge base in Step 5;

[0013] Step 7, based on the prompt template combined with multi-source data and the knowledge base, establish a false alarm database to verify the output results of the prompt template, and use the cosine similarity algorithm to calculate the cosine similarity between the input and output and the input and output of historical false alarm cases to obtain the similarity degree between the two.

[0014] On the basis of the above scheme, preferably, the fusion formula in Step 1 is as follows:

[0015]

[0016] where T represents the time step, S represents the number of sensors, M represents the size of the feature map after video space dimensionality reduction, C' represents the number of channels after video compression, L represents the historical time series window, D sensor represents the sensor data of the fire department, D video represents the video stream data of the fire department, D history represents the historical data of the fire department.

[0017] On the basis of the above scheme, preferably, the covariance matrix C (m) calculation formula:

[0018] C (m) = Cov(X (m) );

[0019] X (m)is the sub-tensor of the m-th modality, where m = 1, 2, …, M.

[0020] Based on the above solution, preferably, the calculation formula for minimizing the weighted sum of the distances between each modal subspace and the global subspace in step 3 is as follows:

[0021]

[0022] where U is the globally optimal subspace, and U m is the subspace of the m-th modality, and d(U, U m ) represents the distance between the global subspace U and the m-th modal subspace U m , and α m is the signal-to-noise ratio weight of the m-th modality.

[0023] Based on the above solution, preferably, the calculation formula for the variance contribution rate is as follows:

[0024]

[0025] where r i represents the proportion of the information contained in the i-th principal component in the overall data; σ i 2 represents the variance of the i-th principal component in X pca , and i = 1, 2, …, n, where n is the original dimension of X pca .

[0026] Based on the above solution, preferably, the prompt templates are divided into three categories: templates for inferring the cause of a fire, templates for determining the fire level, and templates for giving treatment measures.

[0027] A method for constructing a fire safety intelligent agent based on a large model according to the present invention uses a method of dynamic weighted tensor fusion analysis to fuse and model sensor data, video stream data, and historical data as the inference data source of the large model; combines the fused data with a fire knowledge base to form the input prompt of the large model in the form of a template; finally, analyzes and verifies the inference results of the large model and corresponds them to the platform capabilities in the form of function call to automatically execute treatment measures.

[0028] A method for constructing a fire safety intelligent agent based on a large model according to the present invention realizes accurate diagnosis and automatic disposal of fire incidents through the deep combination of multi-source data fusion and a large model in the fire field, solves the problem of response delay caused by the traditional fire protection system relying on manual judgment, and significantly improves the emergency efficiency and safety. In addition, the intelligent design of the present invention can effectively reduce the false alarm rate, reduce unnecessary resource waste, and provide strong technical support for modern urban fire safety. Description of the Drawings

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings. In the attached

[0030] In the figure:

[0031] Figure 1 is a flowchart of the method for constructing a fire safety intelligent agent based on a large model of the present invention; Specific Embodiments

[0032] The following will further describe in detail the specific embodiments of the present invention in conjunction with the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0033] It should be understood that when used in this specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections.

[0034] To make the drawings concise, only the parts related to the present invention are schematically shown in each drawing, and they do not represent the actual structure of the product. In addition, to make the drawings concise and easy to understand, in some drawings, components with the same structure or function are only schematically shown one of them, or only one of them is marked. In this article, "one" not only means "only this one", but also means "more than one" situation.

[0035] It should also be further understood that the term "and / or" used in the description of this application and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0036] In the embodiments shown in the drawings, the indication of directions (such as up, down, left, right, front and back) is used to explain that the structures and movements of various components of the present invention are not absolute but relative. When these components are in the positions shown in the drawings, these explanations are appropriate. If the description of the positions of these components changes, the indication of these directions also changes accordingly.

[0037] In addition, in the description of this application, the terms "first", "second", etc. are only used for distinguishing descriptions, and cannot be understood as indicating or implying relative importance.

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will describe the specific embodiments of the present invention with reference to the accompanying drawings. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings, and other embodiments can be obtained.

[0039] Please refer to Figure 1 , a method for constructing a fire safety intelligent body based on a large model of the present invention specifically includes the following steps:

[0040] 1. Use the dynamic weighted tensor fusion analysis method for multi-modal data (DWTF-PCA) to obtain the inference data source. The present invention realizes the efficient fusion of sensor data, video stream data, and historical data through tensor preprocessing, multi-scale PCA, and data dimensionality reduction. The specific method is as follows:

[0041] First, unify the spatio-temporal structure of different modal data to solve the problem of dimension and format mismatch. The sensor data, video stream data, and historical data collected by the fire protection system are converted into a unified four-dimensional spatio-temporal tensor through Concat to achieve the alignment of multi-modal data in the time, space, and feature dimensions. The processing method is as follows:

[0042]

[0043] Among them, T represents the time step (in seconds), S represents the number of sensors (such as the number of smoke sensors, temperature sensors, humidity sensors, etc.), M represents the size of the feature map after video space dimensionality reduction, usually 8*8, C' represents the number of channels after video compression, and L represents the historical time series window, such as 24 hours.

[0044] Then, perform multi-scale tensor PCA analysis on the tensor data aligned in the time, space, and feature dimensions. For the fused data X, it is divided according to different modalities to obtain sub-tensors X of M modalities (m) , m = 1, 2,..., m. Then calculate the covariance matrix C (m) . The calculation formula is:

[0045] C (m) = Cov(X (m) )

[0046] The covariance matrix C (m) is used to capture the correlation between internal features of the m-th modality. For example, in the sensor data modality, it reflects the co-variation relationship between different sensors.

[0047] Finally, find a globally optimal subspace on the Grassmann manifold such that the weighted sum of the distances between each modal subspace and the global subspace is minimized. The calculation formula is as follows:

[0048]

[0049] U is the globally optimal subspace to be found, comprehensively reflecting the main features of all modal data. U m is the subspace of the m-th modality, which can be obtained from C through the principal component analysis (PCA) method (m) . d(U, U m ) represents the distance between the global subspace U and the m-th modal subspace U m , and usually the Grassmann distance can be used to measure it. α m is the signal-to-noise ratio weight of the m-th modality. The higher the signal-to-noise ratio of a modality, the greater its weight and the greater its contribution to subspace alignment.

[0050] Then, the feature data X after multi-scale PCA analysis pca is reduced in dimension using the method of variance contribution rate. The specific operation method is to calculate the variance σ pca of each principal component in X i 2 (i = 1, 2,.., n, where n is the original dimension of X pca ), and then calculate the variance contribution rate r i of each principal component. The calculation formula is as follows:

[0051]

[0052] The variance contribution rate r i represents the proportion of the information contained in the i-th principal component in the overall data. The larger the variance, the more information the principal component contains.

[0053] Usually, a cumulative variance contribution rate threshold T is set. For example, the cumulative variance contribution rate threshold T takes a value of 0.8 - 0.95. The principal components are sorted in descending order of variance contribution rate, and the variance contribution rates are accumulated in turn until the cumulative variance contribution rate exceeds the threshold T. Record the number k of principal components included at this time.

[0054] Through the above method, the principal components containing most of the information can be selected, thus achieving dimensionality reduction. For example, when T = 0.9, the selected combination of principal components can retain 90% of the information of the original data.

[0055] 2. The inference ability of the large model depends on the quality of the prompt. The above steps are to fuse multi-source data into unified tensor data and preserve the characteristics of the original data as much as possible. Then, it is further necessary to generate a fire diagnosis prompt template in combination with the fire knowledge base. The main steps include data preparation and preprocessing, template selection and customization, knowledge base association and information filling.

[0056] (2.1) Data preparation and preprocessing. Analyze the data fused in step 1 to clarify the actual physical meaning represented by each principal component. For example, principal component 1 mainly reflects the comprehensive information of smoke concentration and temperature, and principal component 2 is related to humidity and ventilation conditions, etc. Determine by viewing the element values of the eigenvectors in X pca The larger the absolute value of the element value, the greater the contribution of the corresponding original variable to this principal component. Select the variable with the largest contribution value as the physical meaning variable of the principal component, that is, determine the physical meaning variable of the principal component. At the same time, organize the fire knowledge base (collected from the Internet or enterprise data centers) into three major categories: fire type judgment rules, fire level classification standards, and disposal measures for different fire scenarios. Establish indexes for different types of knowledge based on these three major categories.

[0057] (2.2) Prompt template customization. Construct a series of prompt templates according to different fire diagnosis scenarios and requirements. Corresponding to the classification of the fire knowledge base, the prompt templates are also divided into three categories: inferring the cause of the fire, determining the fire level, and giving treatment measures. Specific examples are as follows:

[0058] Scenario 1: Judging whether a fire has occurred

[0059] Template: "Based on the fusion analysis data, the value of the [physical meaning] index represented by principal component [abnormal principal component number] has reached [specific value], which has exceeded the normal threshold. Please combine the fire knowledge base to judge whether a fire has occurred."

[0060] Scenario 2: Determining the fire level

[0061] Template: "The fusion analysis data shows that the [physical meaning 1] of principal component [abnormal principal component number 1] is [specific value 1], and the [physical meaning 2] of principal component [abnormal principal component number 2] is [specific value 2], both of which are in an abnormal state. Please determine the current fire level based on the fire knowledge base."

[0062] Scenario 3: Giving disposal suggestions

[0063] Template: "It is known that the [physical meaning] reflected by principal component [abnormal principal component number] after fusion analysis is abnormal (the value is [specific value]), and a [fire type] fire has been determined to have occurred. Please combine the fire knowledge base to give corresponding disposal suggestions."

[0064] Select a suitable template from the template library according to the key exception information and specific diagnosis objectives. For example, if only one principal component exception is detected and the objective is to determine whether a fire has occurred, select the template for Scenario 1.

[0065] (2.3) Knowledge base association and information filling. Knowledge base matching: According to the physical meaning of the abnormal principal component, search for relevant knowledge entries in the fire protection knowledge base. For example, if the abnormal principal component represents too high smoke concentration, search for the rules regarding the relationship between smoke concentration and fire in the knowledge base; Information filling: Fill the extracted key exception information (abnormal principal component number, specific value, physical meaning, etc.) and the relevant knowledge matched from the knowledge base into the selected prompt template to generate a complete prompt template.

[0066] 3. Large model inference and result verification. Based on the high-quality prompt template combining multi-source data and the knowledge base obtained in the previous two steps, use it as the input to the large model to obtain the inference result.

[0067] Since the inference result of the large model has a certain degree of uncertainty, therefore, it is necessary to establish a false alarm database to verify its result. Use the cosine similarity algorithm to calculate the cosine similarity between the input data and output result inferred according to the method of the present invention and historical false alarm cases, so as to obtain the similarity degree between the two. If the similarity exceeds a certain threshold (a preset hyperparameter, such as 80%), it indicates that the current output is similar to the characteristics of historical false alarm cases, and there may be a false alarm and it is filtered. At the same time, if this false alarm situation appears for the first time, the operation and maintenance personnel will add it to the false alarm database to become one of the verification sets for the next result.

[0068] 3. Automatic disposal execution. Sensing and execution are the basic capabilities of the intelligent agent. The inference output of the large model not only contains the early warning information and disposal methods for fire safety, but also determines the atomic capabilities of the Internet of Things platform to be called (provided to the large model in the form of function call). For example, if it is determined to be a first-level fire, automatically start the sprinkler system, close the ventilation equipment, and simultaneously send an alarm message to the fire department (Internet services such as phone / sms, etc.); if it is a second-level fire, give priority to notifying the property management personnel to conduct on-site verification. The system supports multiple communication protocols and can be seamlessly connected to various fire protection devices to ensure the rapid execution of disposal instructions. In addition, the system will also be connected to Internet services such as sms, phone, map, etc.

[0069] Among them, the execution process of automatic disposal is formalized as:

[0070] E = g(A, D fusion );

[0071] Among them, E represents the set of actions to be executed, including specific actions such as starting the sprinkler system, closing the ventilation equipment, sending alarm messages, etc. A represents the alarm event, covering key information such as the event location and level, which are the key basis for triggering the disposal actions. Dfusion is the fusion data (feature data filtered by the large model), which integrates environmental feature data, sensor data, equipment status data, etc. Through multi-source data fusion, it provides more comprehensive and accurate information support for the generation of processing rules. g represents the disposal rule function generated based on the alarm and fusion data. This function determines the final set of actions to be executed according to different alarm levels and the actual situation reflected by the fusion data through the conditions preset in the scenario, realizing the precise execution of automated disposal.

[0072] A method for constructing a fire safety intelligent agent based on a large model according to the present invention uses the method of dynamic weighted tensor fusion analysis to fuse and model sensor data, video stream data, and historical data as the inference data source of the large model; combines the fused data with the fire knowledge base to form the input prompt of the large model in the form of a template; finally, analyzes and verifies the inference result of the large model, and corresponds it to the platform capabilities in the way of function call to automatically execute the disposal means.

[0073] A method for constructing a fire safety intelligent agent based on a large model according to the present invention realizes the precise diagnosis and automated disposal of fire events through the deep combination of multi-source data fusion and the large model in the fire field, solves the problem of response delay caused by the traditional fire protection system relying on manual judgment, and significantly improves the emergency efficiency and safety. In addition, the intelligent design of the present invention can effectively reduce the false alarm rate, reduce unnecessary resource waste, and provide strong technical support for the fire safety of modern cities.

[0074] In order to further elaborate the technical solution of the present invention in detail, the following will take the fire protection scenario of commercial buildings as an example for detailed description:

[0075] First, the data input stage. Assume the following data is generated according to the Internet of Things platform:

[0076] 1) The smoke sensor detects in real time that the smoke concentration in the corridor on the 3rd floor has risen to 15% LEL of the smoke data, and the temperature sensor obtains the temperature data showing 50°C and continuously rising;

[0077] 2) The video stream analysis module identifies and obtains the images of flashing red light and smoke diffusion in this area.

[0078] 2. Diagnostic enhancement prompt generation. Generate the final prompt based on the data and prompt template generated above. For example, "Currently, the sensor on the 3rd floor shows [Smoke: 15% LEL, Temperature: 50°C], and a red light is detected at the video frame [coordinates (x1, y1)]. Please combine the knowledge base to determine whether it is an electrical fire and output disposal suggestions."

[0079] 3. Multimodal analysis and decision-making. The large model analyzes the sensor data and video features, and infers that the similarity with the knowledge base case "Initial fire caused by electrical short circuit" reaches 92%. It is determined as "Electrical fire (Level 1)", and the disposal rule is "Cut off the power + Activate the sprinkler". And generate an alarm event based on this result:

[0080] {Event ID: 001, Location: 3rd floor distribution room, Level: Level 1, Suggestion: Immediately cut off the power and activate the sprinkler}.

[0081] 4. Automated execution. Execute instructions according to the interfaces provided by the IoT platform and cloud services. For example:

[0082] 1) The IoT platform calls the device control interface:

[0083] {"action": "power_off", "device": "3F_main_switch"}

[0084] {"action": "activate", "device": "3F_sprinkler"}

[0085] Simultaneously send an alarm message to the fire command center: "Level 1 fire in the 3rd floor distribution room. Emergency disposal has been initiated. Requesting support."

[0086] 6. Follow-up processing and optimization. After the event is disposed of, the system automatically records the whole process data of this fire (including sensor data, video stream, etc.), and incorporates it into the historical database for optimizing the training and decision-making capabilities of the large model. At the same time, the system generates an event report for managers to review and analyze.

[0087] Finally, the method of this application is only a preferred implementation, and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for constructing a fire safety agent based on a large model, characterized in that: The following steps are involved: Step 1: Obtain fire sensor data, video stream data, and historical data, and use Concat to convert the above data into a unified four-dimensional space-time tensor to obtain fused data X; Step 2: For the fused data X, divide it according to different modes to obtain M modal sub-tensors X (m) , m = 1, 2, ..., m, calculate the covariance matrix C of each modal subtensor (m) , used to capture the correlation between the internal features of the mth modality; Step 3, based on the covariance matrix C (m) , obtain the m-th mode subspace U through principal component analysis m and feature data X pca , obtain a global optimal subspace on the Grassmann manifold so that the weighted sum of the distances between each modal subspace and the global subspace is minimized; Step 4, calculate X pca The variance σ of each principal component in i 2 , where i = 1, 2, .., n, n is X pca The original dimension of the principal component is used to calculate the variance contribution rate r of each principal component. i ; Based on variance contribution rate r i , accumulate the variance contribution rate in sequence until the cumulative variance contribution rate exceeds the threshold T, record the number of principal components k at this time, and select the principal component containing most of the information; Step 5: Analyze the fused data from step 1 to clarify the actual physical meaning of each principal component. pca The element values ​​of the feature vector in are used to determine the correlation, and the fire protection knowledge base is divided into three categories: fire type judgment rules, fire level classification standards, and disposal measures for different fire scenes, so as to construct an index of different types of knowledge; Step 6: Construct prompt templates according to different fire diagnosis scenarios and requirements, and correspond to the fire protection knowledge base classification in step 5. Step 7: Based on the prompt template that combines multi-source data with the knowledge base, a false alarm database is established to verify the output results of the prompt template. The cosine similarity algorithm is used to calculate the cosine similarity between the input and output and the input and output of historical false alarm cases to obtain the degree of similarity between the two.

2. A method for constructing a fire safety intelligent agent based on a large model as claimed in claim 1, characterized in that: The fusion formula in step 1 is as follows: Among them, T represents the time step, S represents the number of sensors, M represents the feature map size after the video space dimension reduction, C' represents the number of channels after video compression, L represents the historical time series window, and D sensor Represents fire sensor data, D video Represents the video stream data of fire fighting, D history Represents historical data of firefighting.

3. A method for constructing a fire safety intelligent agent based on a large model as claimed in claim 1, characterized in that: The covariance matrix C in step 2 (m) Calculation formula: C (m) =Cov(X (m) ); X (m) is the subtensor of the mth mode, m=1,2,…,m.

4. A method for constructing a fire safety intelligent agent based on a large model as claimed in claim 1, characterized in that: The calculation formula for minimizing the weighted sum of the distances between each modal subspace and the global subspace in step 3 is as follows: Among them, U is the global optimal subspace, U m is the subspace of the mth mode, d(U,U m ) represents the global subspace U and the mth modal subspace U m The distance between m is the signal-to-noise ratio weight of the mth mode.

5. A method for constructing a fire safety intelligent agent based on a large model as claimed in claim 4, characterized in that: The formula for calculating variance contribution rate is as follows: Among them, r i Indicates the proportion of information contained in the i-th principal component in the overall data; σ i 2 Represents X pca The variance of the i-th principal component in X, i = 1, 2, .., n, n is pca The original dimension of .

Images