SEO keyword intelligent analysis system based on edge AI chip

Abstract

The invention relates to the technical field of information retrieval, in particular to an intelligent SEO keyword analysis system based on an edge AI chip, which comprises three core modules: a data acquisition module acquires user behaviors, webpage contents and equipment environment data in real time; an AI chip is arranged in the edge calculation module, data is processed on local equipment, user preference characteristics are generated, and optimized query is generated in combination with search keywords; according to the method, terminal equipment is sunk through AI computing power, a closed-loop system with real-time collection, local intelligent processing and personalized result enhancement is constructed, and the long-term contradiction between personalized service and user privacy protection is solved. Sensitive data is locally processed without being uploaded to the cloud, so that the risk of privacy leakage is avoided from the source; the edge AI chip realizes real-time search personalization with extremely low delay and high precision, and the intelligent level of a search engine and the satisfaction degree of a user are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of information retrieval technology, and more specifically, to an intelligent SEO keyword analysis system based on an edge AI chip. Background Technology

[0002] With the explosive growth of internet information, search engines have become the core entry point for users to obtain information. Search engine optimization (SEO) strategies and personalized search results are crucial for improving user experience. Currently, personalized search services are mainly implemented based on cloud servers. The typical approach is that cloud servers continuously collect a large amount of data such as users' search history, click behavior, and geographical location, build user profiles through complex algorithms, and use these profiles to optimize queries or reorder search results when users initiate new searches.

[0003] However, this cloud-based centralized processing model has several inherent drawbacks: First, there is a high risk of privacy leaks: users' sensitive behavioral data and preference models are stored on cloud servers, which are at risk of being accessed, leaked, or misused without authorization, making users hesitant to use personalized features.

[0004] Second, there is network latency: every user interaction needs to be uploaded to the cloud for processing and feedback. When the network conditions are poor, it will cause a noticeable delay, affecting the smoothness and real-time performance of the search experience.

[0005] Third, personalized feedback is delayed: cloud-based model updates are often not real-time, and cannot immediately learn from the user's browsing behavior and adjust the strategy for the next search, resulting in personalized recommendations that are not accurate or timely enough.

[0006] In addition, while some existing local browser plugins can record user behavior, they lack powerful local computing capabilities, making it impossible to perform complex real-time artificial intelligence analysis and modeling. Their functions are relatively limited, making it difficult to achieve deep and truly personalized real-time search.

[0007] In view of this, an intelligent SEO keyword analysis system based on edge AI chips is proposed. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent SEO keyword analysis system based on edge AI chips, so as to provide a highly personalized search experience while fundamentally protecting user privacy and reducing latency.

[0009] To address the aforementioned technical problems, this invention provides an intelligent SEO keyword analysis system based on an edge AI chip, comprising: The data acquisition module is used to collect behavioral data, webpage content data, and device environment data generated during user interaction with the search engine in real time. The edge computing module, with a built-in AI chip, is connected to the data acquisition module and is used to process and analyze the acquired data in real time on the local device, generate user preference features, and generate optimized search queries based on the user's input search keywords and user preference features. The search results processing module, connected to the edge computing module, is used to send optimized search queries to the search engine and personalize the search results returned by the search engine.

[0010] As a further improvement to this technical solution, the data acquisition module includes: The user behavior collection unit captures user clicks, page dwell time, scrolling actions, and mouse movements on the search results page through browser plugins. The web page content acquisition unit obtains the title text, body content, and meta tag information of the page visited by the user through client-side document object model parsing technology; The environmental data acquisition unit obtains the latitude and longitude coordinates and system timestamp of the user's device through the device application programming interface, and obtains search trend data of the user's area through the network interface.

[0011] As a further improvement to this technical solution, the edge computing module includes: The data preprocessing unit is used to clean, normalize, and extract features from the collected raw data. The user preference modeling unit operates based on an AI chip. It constructs a user preference model through real-time incremental learning using an online learning algorithm and outputs a user preference feature vector through the user preference model. The user preference model is a multi-layer neural network model. The query optimization unit, upon detecting a user's submitted search request, fuses the user's input search keywords with the user's preference feature vector to calculate and output an optimized search query containing user preference information.

[0012] As a further improvement to this technical solution, the neural network model implemented by the user preference modeling unit achieves user preference learning in the following way: Extract the correlation features between page content and search queries from the user's historical click records; Analyze the distribution patterns of user dwell time on different types of web pages; Learn users' preferences for website domains or content sources with a historical visit frequency higher than a preset threshold; Each dimension of the user preference feature vector corresponds to a specific preference category.

[0013] As a further improvement to this technical solution, the query optimization unit generates an optimized search query through the following steps: Perform lexical analysis and semantic parsing on the search keywords entered by the user; Calculate the semantic similarity between the user's input search keywords and each preference dimension in the user preference feature vector, and select the preference dimension with the highest similarity score; Combine the semantic tags corresponding to the preference dimensions with the original search keywords to generate expanded queries; Multiple candidate optimized queries are generated and sorted and selected based on a user preference model.

[0014] As a further improvement to this technical solution, the search result processing module includes: The query proxy unit is used to intercept the original query request sent by the user to the search engine, and after obtaining the user's authorization, replace the original query request with an optimized search query and then send it to the search engine. The result reordering unit is used to calculate the relevance of search results returned by the search engine based on user preference features. The visualization enhancement unit is used to highlight search results that match user preferences in the user interface.

[0015] As a further improvement to this technical solution, the result reordering unit processes search results in the following manner: Calculate the matching score between each search result and the user's preference features; The search results list is reordered based on the match score; Preserve the integrity of the original search results, only adjust the display order of the search results.

[0016] As a further improvement to this technical solution, it also includes: The cloud-edge collaborative update module is connected to the edge computing module and is used to realize the synchronization of model parameters and knowledge base updates of the user preference model between the edge computing module and the cloud server. The cloud-edge collaborative update module periodically receives anonymized user preference model parameters from the edge computing module and encrypts and transmits the model parameters to the cloud server. The cloud-edge collaborative update module also receives updated model parameters and global knowledge data from the cloud server and delivers them to the edge computing module for use. The model parameters include neural network weights and bias terms.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This SEO keyword intelligent analysis system based on edge AI chips builds a closed-loop system that integrates real-time data collection, local intelligent processing, and personalized result enhancement by pushing AI computing power down to user terminal devices (edge ​​side), fundamentally solving the long-standing contradiction between "personalized services" and "user privacy protection".

[0018] 2. In this SEO keyword intelligent analysis system based on edge AI chips, the system completes the processing of all sensitive data and user preference modeling on the local device without uploading the original behavioral data to the cloud, thus eliminating the risk of privacy leakage at the source. At the same time, with the powerful computing power of edge AI chips, it achieves real-time personalized search with extremely low latency and high accuracy, significantly improving the intelligence level of the search engine and user satisfaction. Attached Figure Description

[0019] Figure 1 This is an overall system block diagram of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0021] Currently, existing search engine personalization services rely entirely on cloud servers to process user data. The typical approach is that cloud servers continuously collect a large amount of data such as users' search history, click behavior, and geographical location, build user profiles through complex algorithms, and use these profiles to optimize queries or reorder search results when users initiate new searches. However, this approach also has problems such as high risk of privacy leakage, network latency, and delayed personalized feedback. In view of this, please refer to Figure 1 As shown, the purpose of this invention is to provide an intelligent SEO keyword analysis system based on an edge AI chip. This system provides a three-layer distributed solution, the core of which lies in pushing computing power down to user terminal devices, including: The data acquisition module is used to collect real-time behavioral data, webpage content data, and device environment data generated during user interactions with the search engine; for example: On Windows / Mac OS PCs, this is achieved through the Chrome browser extension and a local background daemon. The browser extension is responsible for capturing user behavior within the browser, while the daemon is responsible for calling the operating system API to obtain device environment data. After collecting the data, the data is serialized into Protocol Buffers format through inter-process communication (IPC) and written to a shared memory area for the edge computing module to read quickly. The edge computing module, with a built-in AI chip, connects to the data acquisition module and is used to process and analyze the acquired data in real time on a local device, generate user preference features, and generate optimized search queries based on the user's input search keywords and user preference features; for example: The NPU (Neural Processing Unit) is driven by directly calling the low-level interface provided by the hardware manufacturer through a locally compiled C++ dynamic link library. The library contains a monitoring thread that continuously polls the shared memory area mentioned above. Once new data is detected, the NPU driver is immediately invoked to load the data into the NPU's memory for model inference calculation. The entire process is completed locally without a network connection. The search result processing module, connected to the edge computing module, is used to send optimized search queries to the search engine and personalize the search results returned by the search engine; for example: This is achieved through the collaborative work of a local proxy server and a browser extension. The local proxy server, written in Go, listens on a local port (e.g., port 8080). The browser extension modifies the browser's proxy settings (or configures a PAC script) to redirect all requests destined for preset search engine domains (e.g., www.google.com, www.bing.com) to the local port (e.g., 127.0.0.1:8080). After intercepting a search request, the local proxy server communicates with the edge computing module to obtain optimized query terms, replaces the parameters in the original request, and then forwards the request to the actual search engine. Subsequently, it parses and personalizes the page content returned by the search engine before returning it to the browser for rendering and display. Through the aforementioned technical means, a complete and efficient localized personalized search closed loop is constructed. All raw data processing involving user privacy is completed within the user's device. High-speed communication is achieved through shared memory and local loopback network, eliminating the risk of data leakage at the source and effectively protecting user privacy. At the same time, by utilizing the dedicated computing power of the local NPU, the latency of personalized processing is reduced, realizing seamless and real-time personalized search, and improving user search efficiency and satisfaction. Example

[0022] To build an accurate user preference model, it is necessary to collect comprehensive, accurate, and low-power multi-dimensional raw data on local devices. This includes how users interact with search results, what content they view, and their environment. Simultaneously, it is crucial to ensure that the data collection process does not significantly impact device performance or browser stability. Therefore, the data collection module includes: The user behavior collection unit captures user clicks, page dwell time, scrolling actions, and mouse movements on the search results page through a browser plugin; specifically: This unit is implemented through a browser extension. The extension uses a page navigation completion event listener to detect when the search results page has finished loading. Then, it injects a content script into the page using the chrome.scripting.execute Script method. This injected script collects user behavior by listening to events such as page clicks (recording the XPath path and timestamp of the event.target element), mouse movements (sampling mouse coordinates 10 times per second), and scrolling (recording the value of performance.now() and calculating the difference when the page finishes loading and when the user leaves the page). All collected data is sent to the extension's background script for temporary storage and processing. The webpage content acquisition unit uses client-side document object model parsing technology to obtain the title text, body content, and meta tag information of the pages accessed by the user; specifically: This unit is also integrated into the browser extension. When a user clicks on a search result and navigates to the target webpage, the extension detects that the target webpage has finished loading and injects a special content script into the target webpage. This script obtains the page title (by calling document.title) and meta description information (by document.query Selector('meta[name="description"]')?.content). For the main content, the script uses a special content extraction library (with a lightweight, modified copy of the Readability.js library embedded) to automatically analyze the page structure and identify and extract the core article content. The environmental data acquisition unit obtains the latitude and longitude coordinates and system timestamp of the user's device through the device application programming interface (API), and acquires search trend data of the user's location area through the network interface; specifically: The implementation of this unit is platform-dependent. On mobile platforms, latitude and longitude coordinates are obtained by calling the location service interface (requesting GPS or network location updates); the system timestamp is obtained through the system interface. Regional search trend data is obtained by initiating network requests to the application programming interface provided by the search engine to obtain popular keyword data; on PC platforms, location information may require manual input by the user or approximate location via IP address. Through the above technical means, a complete and working local data acquisition solution is provided. By making full use of the capabilities of browser extensions and operating system interfaces, it accurately captures multi-dimensional and fine-grained data necessary for building user preference models. All operations are completed locally, laying a solid data foundation for subsequent real-time analysis, while firmly protecting the privacy of user data. Example 3: Because the collected raw data is unstructured and messy, it cannot be directly used for machine learning models. Furthermore, on terminal devices with limited computing resources, memory, and power consumption, efficiently performing data cleaning, transformation, and running a complex user preference model for real-time inference and query optimization presents a significant technical challenge. Therefore, the edge computing module includes: The data preprocessing unit is used to clean, normalize, and extract features from the collected raw data; specifically: After reading the raw data from shared memory, this unit executes the following standardized process: Data cleaning: Use regular expressions to match and remove useless characters from the text; remove invalid data according to business rules.

[0023] Data normalization: For numerical features, the minimum-maximum normalization method is used to linearly transform them to a fixed interval.

[0024] Feature extraction: For text content, a pre-trained lightweight neural network model is used to convert the text string into a fixed-length feature vector, which effectively represents the semantic information of the text; The user preference modeling unit operates based on an AI chip. It constructs a user preference model through real-time incremental learning using an online learning algorithm and outputs a user preference feature vector. The user preference model is a multi-layer neural network model; specifically: This unit is a multilayer perceptron model, defined as follows: an input layer (768 dimensions), a first fully connected layer (256 neurons, using the Re LU activation function), a dropout layer (dropout rate of 0.5), and an output layer (10 neurons, using the Softmax activation function). The model is built using a mobile machine learning framework and deployed on an NPU for inference via dedicated hardware acceleration. Training is conducted online using momentum stochastic gradient descent as the optimizer, with a learning rate of 0.01 and a momentum parameter of 0.9. The loss function is classification cross-entropy, and the model weights are updated after processing N (e.g., N=100) new sample batches to achieve continuous incremental learning. The query optimization unit, upon detecting a user's search request, fuses the user's input search keywords with the user's preference feature vector to calculate and output an optimized search query containing user preference information; specifically: This unit is a Python script that runs on the device's CPU. It exchanges data with the preference modeling unit in the NPU via shared memory. When a user submits a search request, it immediately initiates the following process: First, it uses a word segmentation library (such as jieba.lcut(query)) to segment and tag the user's original keywords, retaining nouns (n) and verbs (v) as the core stems. Then, it reads the latest user preference feature vector (e.g., 10-dimensional) output by the user preference modeling unit from the shared memory. Finally, it generates a set of candidate optimized search queries through a predefined fusion function (e.g., concatenating the word vectors of the core stems with the preference feature vectors and then inputting them into a small linear classifier), and selects the one with the highest confidence as the final output. Through the aforementioned technical means, using dedicated AI hardware acceleration, model lightweighting technology (quantization), and efficient online learning algorithms, complex neural network models can be successfully run on resource-constrained terminal devices in a low-power, high-performance manner. This enables large-scale data processing and real-time model inference that could traditionally only be completed in the cloud, providing core technical support for truly real-time, adaptive personalized search.

[0025] Example 4: Since simple keyword matching cannot deeply understand users' true interests and preferences, the core problem to be solved is how to enable machine learning models to automatically learn deep, semantic-level preference patterns from user behavior, thereby generating a digital profile (feature vector) that accurately represents user interests. Therefore, the neural network model implemented by the user preference modeling unit achieves user preference learning in the following way: Extract the correlation features between page content and search queries from the user's historical click records; Analyze the distribution patterns of user dwell time on different types of web pages; Learn users' preferences for website domains or content sources with a historical visit frequency higher than a preset threshold; Each dimension of the user preference feature vector corresponds to a specific preference category. For example, neural network models perform deep learning through various mechanisms, specifically: Associative Feature Learning: During model training, the system automatically constructs training sample pairs. For example, (user search term, page title clicked by the user) is used as a positive sample pair. The model employs a dual-tower structure (Siamese Network), with one tower processing search terms and the other processing page titles and summaries. By maximizing the cosine similarity between the two vectors in the latent space of the positive sample pair, the model learns to map semantically similar searches and clicks to similar vector spaces. For example, if a user searches for "apple" multiple times and then clicks on a page about "iPhone," the model will automatically learn a strong semantic association between "apple" and "iPhone," rather than an association with "fruit." Temporal distribution pattern learning: The model inputs the time t spent by users on each type of webpage as an important continuous numerical feature into the fully connected layer. The model automatically learns the weight of this feature during training. For example, the model might discover a strong positive correlation between prolonged stays on technical documentation pages and high output scores in the "programming" preference dimension, thus giving this feature greater weight in predictions. Source Preference Learning: During data preprocessing, the system parses URL strings into their top-level domains (e.g., extracting github.com from https: / / github.com / path / to / repo). This domain string is then used as a categorical feature and input into an embedding layer of the model. This embedding layer maps each domain name to a low-dimensional, dense floating-point vector (e.g., 8-dimensional). During training, the model adjusts the parameters of the embedding layer and its subsequent networks so that when the input features include a vector representing github.com, the value of the "technology" or "programming" preference dimension in the final output vector increases significantly. Preference Vector Output: The output layer of this neural network model is precisely configured with 10 neurons, each corresponding to a predefined preference category (e.g., neuron 0: technology, neuron 1: finance, neuron 2: sports, ..., neuron 9: short videos). The activation value of each neuron (after Softmax transformation) represents the instantaneous preference strength of the user for that category predicted by the model. These 10 output values ​​together form a 10-dimensional user preference feature vector, which is a highly condensed and interpretable digital profile of user interests. Through the aforementioned techniques, the model can go beyond superficial keyword matching, extracting abstract and deep-seated interests and preferences from users' behavioral sequences. The generated preference feature vector is a precise and efficient semantic representation, providing a powerful and reliable basis for subsequent query optimization and search result re-ranking, greatly improving the accuracy of personalization.

[0026] Example 5: Since user-inputted search keywords are often short and vague, the goal of this implementation is to transform the internally abstracted, numerical user preference vector into specific and effective query terms that the search engine can understand and process, thereby significantly improving the relevance of search results on the first query. Therefore, the query optimization unit generates an optimized search query through the following steps: Step 1, Lexical Analysis and Semantic Parsing: Lexical analysis and semantic parsing are performed on the user-input search keywords. Specifically, the locally integrated jieba word segmentation library is used to process the original user-input keywords; the jieba.lcut_for_search(query) function is called to perform word segmentation in search engine mode, obtaining a keyword list. Subsequently, jieba.posseg.lcut(query) is called to perform part-of-speech tagging, retaining words with core parts of speech such as nouns (n) and verbs (v) as core stems representing the user's search intent. Step 2, Preference Dimension Matching: Calculate the semantic similarity between the user's input search keywords and each preference dimension in the user's preference feature vector, and select the preference dimension with the highest similarity score. Specifically, convert one or more core stems obtained in Step 1 and 10 preference category labels (such as "technology", "programming", "sports") into word vectors respectively. A pre-trained 300-dimensional Word2Vec word vector model downloaded locally can be used here. Calculate the cosine similarity between each core stem vector and each category label vector, using the following formula: Then, all similarity scores are sorted, and the preference category label with the highest score is selected (for example, entering "Python" has the highest similarity to the "programming" label). Step 3, Query Expansion: Combine the semantic tags corresponding to the preference dimensions with the original search keywords to generate expanded queries. Specifically, using a rule-based strategy, combine the matched semantic tags corresponding to the preference dimensions with the original search keywords to generate new, more informative expanded queries. Combination strategies include, but are not limited to: direct appending (e.g., "Python" + "programming" = "Python programming"), adding site restrictions (e.g., "Python" + "site:github.com" = "Pythonsite:github.com"), or adding file types (e.g., "Python" + "filetype:pdf"). Step 4, Ranking and Selection: Multiple candidate optimization queries are generated and ranked based on a user preference model. Specifically, the system generates 3-5 candidate optimization queries according to preset rules. Then, each candidate query is simulated as a "virtual" search request and input into the local user preference model for rapid inference. The model outputs a predicted value representing the probability that a user will click on the search result returned by that candidate query. Finally, the candidate query with the highest predicted click probability is selected as the optimized search query to be sent to the search engine. Through the above steps, the user's short-term search intent (keywords) and long-term interest preferences (preference vectors) can be effectively integrated at the semantic level, generating search queries with richer information and clearer intent. This can guide the search engine to return a more relevant list of results that better meets the user's expectations from the source.

[0027] Example 6: Considering how to seamlessly take over the search process without disrupting users' existing search habits and ensuring informed consent, the optimized query is sent to the search engine, the returned results are personalized and enhanced, and finally presented to the user in an intuitive form. Therefore, the search result processing module includes: The query proxy unit is used to intercept the original query request sent by the user to the search engine, and after obtaining the user's authorization, replace the original query request with an optimized search query before sending it to the search engine; specifically: This unit primarily utilizes browser extensions for implementation. The extension leverages the chrome.declarativeNetRequest API (Manifest V3) to intercept network requests. The extension pre-declares the search engine URL patterns to be intercepted (e.g., *: / / www.google.com / search?*) in its rule list. When a request is intercepted, the extension's background script triggers a pop-up page clearly explaining the "Smart Search Optimization" feature and requesting authorization. Only after the user clicks "Agree" to authorize will the extension dynamically modify the request URL using the chrome.declarativeNetRequest.updateDynamicRules API, replacing the original query parameters with the optimized ones, and then allow the request to proceed. The result reordering unit is used to calculate the relevance of search results returned by the search engine based on user preference features. Specifically, this unit is integrated into a local proxy server. After receiving the HTML response returned by the search engine, the proxy server uses the Go language's goquery library (similar to jQuery) to parse the HTML document. It iterates through and extracts each search result item (div.g is the container selector for the Google search results page) by calling goquery.NewDocumentFromResponse(resp) and doc.Find("div.g").Each(...). For each result item, it further extracts its title (h3 tag text), URL (href attribute of the a tag), and summary (div.st tag text). The visualization enhancement unit is used to highlight search results that match user preferences in the user interface; specifically: This unit follows the results reordering unit. For each extracted search result item, its title and summary text are fed into a local user preference model for relevance scoring. Then, all search result items are sorted in descending order based on the scoring results. Subsequently, goquery is used to manipulate the DOM, first clearing the contents of the original search result container #rso, and then re-inserting the result items in the new order. Simultaneously, a custom CSS class, such as highlight-result, is added to the div.g elements of the top three most relevant result items; and inline CSS styles are inserted into the relevant sections of the returned HTML document. <style>.highlight-result { border-left: 3px solid #4CAF50; padding-left: 10px;}< / style> This allows for a green highlighting of the optimal results; Through the aforementioned technical means, while fully respecting users' right to know and right to choose, the entire process of query optimization and result enhancement is completed automatically and non-intrusively; ultimately, the most valuable information is presented to users in a more intuitive and efficient manner, significantly improving users' information acquisition efficiency and search satisfaction.

[0028] Example 7: Considering that when re-ranking search results for personalization, it is necessary to ensure the completeness and diversity of the results and avoid creating "information cocoons" due to excessive personalization, causing users to miss new information that may be important but is temporarily considered irrelevant by the model, the result re-ranking unit processes search results in the following way: Calculate the match score between each search result and the user's preference features; specifically: For each original search result item returned by the search engine (i.e., each div.g element), its title text and summary text are concatenated into a complete string S. String S is then fed into the local user preference model for preprocessing and NPU inference. The model outputs a 10-dimensional preference vector. This vector represents the content attributes of the search result itself; the matching score. By calculating vectors With the current user preference vector The dot product is obtained by calculating the product using the following formula: This dot product value reflects the overall relevance between the search results and the user's personal interests; a higher value indicates a stronger relevance. The search results list is reordered based on the match score; specifically: The system creates an empty list to store the tuple (search_result_dom_element, Iterate through all search result items and calculate the value for each. And store them in a list. Then, use a built-in sorting algorithm (such as Go's `sort.Slice` function) to sort them according to... Sort the list in descending order; The integrity of the original search results is preserved, and only the display order of the search results is adjusted. Specifically, in the final output HTML, the original DOM elements of each search result item (without any content modification) are rendered onto the page in the order they were sorted. No results are deleted, and the HTML content within any result is not modified (the title, links, and summary remain as returned by the search engine), only their arrangement order is changed. Through the above steps, while providing powerful personalized sorting capabilities, the integrity and neutrality of the search engine's returned results are strictly guaranteed. All information seen by users is undisturbed; the system only adjusts priorities based on individual preferences, effectively avoiding the negative effects of "information cocoons" and balancing the needs for personalization and information diversity.

[0029] Example 8: Considering that local models distributed across tens of millions of terminal devices are isolated and unable to learn and evolve from global data, a key challenge is how to enable these models to absorb collective intelligence, continuously optimize performance, and simultaneously resolutely prevent any raw data from leaving user devices to protect privacy. Therefore, the system also includes: The cloud-edge collaborative update module connects to the edge computing module and is used to synchronize model parameters and update the knowledge base of the user preference model between the edge computing module and the cloud server. The cloud-edge collaborative update module periodically receives anonymized user preference model parameters from the edge computing module and encrypts and transmits the model parameters to the cloud server; for example: Edge-side upload: The on-device update service is triggered weekly as a scheduled task (Cron Job). It calls TFLite's Interpreter::GetTensor() and related interfaces to extract the parameters of all trainable layers in the current local user preference model, including the weight matrix W and bias vector b. Subsequently, using the EVP_EncryptInit_ex series of functions from the OpenSSL library, the parameters (W, b) are encrypted using the AES-256-GCM encryption algorithm and a device-unique key generated by the hardware security module (such as TPM) or Secure Enclave at the time of device manufacturing or initial activation. The encrypted data block is uploaded to a specific endpoint on the cloud server (e.g., https: / / api.example.com / model-updates) via an HTTPS POST request (with TLS certificate verification). This request header never contains any information that can directly identify the user (such as IMEI or username). Cloud processing: The cloud server runs the Federated Averaging (FPAvg) algorithm. The server waits until a sufficient number of cryptographic model updates (e.g., 10,000 devices) have been collected before decrypting them in a secure environment. The aggregation process is as follows: first, the weight matrix of all devices is... Add them together, then divide by the total number of devices N to obtain the average global weight. Perform the same operation on the bias vector b. This new... and It integrates the collective knowledge of numerous user devices; The cloud-edge collaborative update module also receives updated model parameters and global knowledge data from the cloud server and delivers them to the edge computing module for use; for example: Cloud-based distribution: The cloud will distribute the aggregated global parameters. and The file is serialized into a single file and digitally signed using a private key in the cloud to ensure its integrity and provenance. The signed file is then published to a Content Delivery Network (CDN). The update service on the device periodically (e.g., daily) queries a fixed URL on the CDN to check for a new version of the model file (by comparing the version number or file hash value). Edge-side update: After downloading the new global model file, the device first verifies the digital signature using the cloud public key pre-built into the system; after successful verification, the device calls the TFLite's Interpreter::SetTensor() or Interpreter::CopyToBuffer() interface with the new global parameters. and This overwrites the corresponding parameters in the local model, thus completing a model update. Through the aforementioned technical means, a secure and privacy-preserving distributed machine learning ecosystem has been established. By absolutely ensuring that the raw data of any individual user never leaves the device, global user knowledge sharing and collaborative evolution of model performance are achieved. This enables each device's local model to continuously and securely become "smarter" and more accurate, achieving the dual goals of privacy protection and improved model performance.

[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A smart SEO keyword analysis system based on an edge AI chip, characterized in that, include: The data acquisition module is used to collect behavioral data, webpage content data, and device environment data generated during user interaction with the search engine in real time. The edge computing module, with a built-in AI chip, is connected to the data acquisition module and is used to process and analyze the acquired data in real time on the local device, generate user preference features, and generate optimized search queries based on the user's input search keywords and user preference features. The search results processing module, connected to the edge computing module, is used to send optimized search queries to the search engine and personalize the search results returned by the search engine.

2. The SEO keyword intelligent analysis system based on edge AI chip according to claim 1, characterized in that, The data acquisition module includes: The user behavior collection unit captures user clicks, page dwell time, scrolling actions, and mouse movements on the search results page through browser plugins. The web page content acquisition unit obtains the title text, body content, and meta tag information of the page visited by the user through client-side document object model parsing technology; The environmental data acquisition unit obtains the latitude and longitude coordinates and system timestamp of the user's device through the device application programming interface, and obtains search trend data of the user's area through the network interface.

3. The SEO keyword intelligent analysis system based on edge AI chip according to claim 1, characterized in that, The edge computing module includes: The data preprocessing unit is used to clean, normalize, and extract features from the collected raw data. The user preference modeling unit operates based on an AI chip. It constructs a user preference model through real-time incremental learning using an online learning algorithm and outputs a user preference feature vector through the user preference model. The user preference model is a multi-layer neural network model. The query optimization unit, upon detecting a user's submitted search request, fuses the user's input search keywords with the user's preference feature vector to calculate and output an optimized search query containing user preference information.

4. The SEO keyword intelligent analysis system based on edge AI chip according to claim 3, characterized in that: The neural network model implemented by the user preference modeling unit achieves user preference learning in the following way: Extract the correlation features between page content and search queries from the user's historical click records; Analyze the distribution patterns of user dwell time on different types of web pages; Learn users' preferences for website domains or content sources with a historical visit frequency higher than a preset threshold; Each dimension of the user preference feature vector corresponds to a specific preference category.

5. The SEO keyword intelligent analysis system based on edge AI chip according to claim 3, characterized in that, The query optimization unit generates an optimized search query through the following steps: Perform lexical analysis and semantic parsing on the search keywords entered by the user; Calculate the semantic similarity between the user's input search keywords and each preference dimension in the user preference feature vector, and select the preference dimension with the highest similarity score; Combine the semantic tags corresponding to the preference dimensions with the original search keywords to generate expanded queries; Multiple candidate optimized queries are generated and sorted and selected based on a user preference model.

6. The SEO keyword intelligent analysis system based on edge AI chip according to claim 1, characterized in that: The search result processing module includes: The query proxy unit is used to intercept the original query request sent by the user to the search engine, and after obtaining the user's authorization, replace the original query request with an optimized search query and then send it to the search engine. The result reordering unit is used to calculate the relevance of search results returned by the search engine based on user preference features. The visualization enhancement unit is used to highlight search results that match user preferences in the user interface.

7. The SEO keyword intelligent analysis system based on edge AI chip according to claim 6, characterized in that: The result reordering unit processes search results in the following manner: Calculate the matching score between each search result and the user's preference features; The search results list is reordered based on the match score; Preserve the integrity of the original search results, only adjust the display order of the search results.

8. The SEO keyword intelligent analysis system based on edge AI chip according to claim 3, characterized in that, Also includes: The cloud-edge collaborative update module is connected to the edge computing module and is used to realize the synchronization of model parameters and knowledge base updates of the user preference model between the edge computing module and the cloud server. The cloud-edge collaborative update module periodically receives anonymized user preference model parameters from the edge computing module and encrypts and transmits the model parameters to the cloud server. The cloud-edge collaborative update module also receives updated model parameters and global knowledge data from the cloud server and delivers them to the edge computing module for use. The model parameters include neural network weights and bias terms.

Images