Cloud-based Tina R329 incremental learning and automatic model deployment method

Abstract

The invention discloses a cloud-based Tina R329 incremental learning and automatic model deployment method, which relates to the field of cloud computing, machine learning and embedded systems and comprises the technical steps of data acquisition and standardization, data rebalance, model pre-training, incremental learning, model quantification, model deployment and the like. Finally, the optimized model is automatically pushed to local equipment through wireless communication, and deployment of the latest model is completed in real time in the local equipment without manual intervention. According to the invention, powerful computing resources of the cloud platform are fully utilized, and the environment data collected by the embedded equipment in real time are efficiently analyzed and processed, so that the equipment can quickly respond and adapt to the environment change, and the intelligence and automation levels of the system are remarkably improved. The method is particularly suitable for application scenes which need to frequently update the model and depend on real-time feedback, such as the fields of intelligent hardware, Internet of Things equipment and edge computing, and has wide application prospect and commercial value.

Description

Technical Field

[0001] The present invention relates to the technical fields of cloud computing, machine learning, embedded systems and computer vision algorithms, and in particular to a cloud-based Tina R329 incremental learning and automatic model deployment method. Background Art

[0002] Traditional batch learning methods require retraining the model on the entire data set, and a large amount of data must be processed each time training. This not only makes the training process very time-consuming, but also consumes a lot of computing resources and increases computing overhead. Since each model update relies on full training, it cannot adapt to changes in new data in real time, especially in a dynamic environment where data is constantly changing, it is difficult to achieve timely updates and rapid responses. Therefore, batch learning cannot meet rapidly changing real-time needs, and it is difficult to cope with the high requirements for flexibility in large-scale data and complex scenarios.

[0003] Traditional embedded model deployment usually relies on manual intervention to complete model training, verification, and deployment. The entire process is cumbersome and error-prone, and the inefficiency of manual operations directly affects the speed and quality of deployment. Due to the complexity of manual updates and deployment, rapid iteration and real-time deployment of models are often not guaranteed. In addition, manual intervention can easily lead to version control confusion. When multiple versions run in parallel, they may cause model incompatibility, conflict, or performance degradation, thereby affecting the stability and accuracy of the system. Especially in scenarios with high-frequency updates or large changes in requirements, untimely and incorrect manual deployment may lead to serious performance degradation and even endanger the normal operation of the system. Summary of the invention

[0004] The purpose of the present invention is to provide a cloud-based Tina R329 incremental learning and automatic model deployment method, which aims to improve the intelligence level and adaptability of embedded devices in the Internet of Things (IoT) environment through efficient data processing and model optimization technology; the method combines incremental learning and automatic model update mechanism to enable the device to continuously optimize the model without reprocessing the full amount of data; at the same time, through the collaborative work of the cloud and local devices, it is ensured that the device can efficiently and flexibly adapt to changes in the environment and needs.

[0005] To achieve the above object, the present invention provides a cloud-based Tina R329 incremental learning and automatic model deployment method, comprising the following steps:

[0006] S1, data collection and standardization: upload the surrounding environment test data obtained by Tina R329 from the sensor group to the cloud;

[0007] S2, data rebalancing: rebalance the data obtained in S1, increase minority class samples, and improve the proportion of minority class samples;

[0008] S3, Model pre-training: Whenever new data is uploaded, the cloud system decides whether re-training is needed based on the model version;

[0009] S4, incremental learning: fine-tune and optimize the data in S2 to avoid training from scratch and quickly adjust the model to adapt to new data features;

[0010] S5, model quantization: converting model parameters into a lower precision form to meet the operating requirements of embedded hardware;

[0011] S6. Model deployment: The optimized model is automatically pushed to the local device via wireless communication, and the local device deploys the latest model in real time.

[0012] Preferably, S1 specifically comprises the following steps:

[0013] S11, the data is read by the acquisition sensor group;

[0014] S12. Configure the Wi-Fi module of Tina R238 and connect to the network;

[0015] S13, formatting the collected data into JSON format;

[0016] S14. Use the MQTT protocol to upload the data in JSON format to the cloud server through the specified topic;

[0017] S15, standardization, standardize the data according to formula (1):

[0018]

[0019] (1) In the formula, x′ is the value to be standardized, μ is the mean of the feature, and σ is the standard deviation.

[0020] Preferably, S2 specifically includes the following steps:

[0021] S21. For the training data, use the synthetic minority oversampling technique SMOTE to identify the samples belonging to the minority class from the training data set and determine the sample categories that need to be oversampled;

[0022] S22. For each minority sample, calculate the Euclidean distance between the minority sample and other samples, and select its K nearest neighbor samples;

[0023] S23. For each minority class sample x i , randomly select a neighbor x from its K nearest neighbors j, and use formula (2) to generate new samples:

[0024] x new =x i +λ·(x j -x i ) (2);

[0025] (2) In the formula, λ is a random number with a value range of [0, 1], which is used to control the new sample in x i and x j The greater the value of λ, the greater the distance between the generated new samples and x. i The further;

[0026] S24, repeatedly generating new samples until the expected number of synthetic samples is reached;

[0027] S25. Synthesize the newly synthesized sample data with the original data set to construct a new balanced data set.

[0028] Preferably, S3 specifically includes the following steps:

[0029] S31. Check whether the current model version meets the business requirements. If so, retraining will not be triggered.

[0030] S32. If the answer to step S31 is no, the current model version does not meet the business requirements. The cloud server starts model pre-training. The training uses the full data set, 80% as the training set and 20% as the test set. A prediction model is generated through standard machine learning methods.

[0031] Preferably, S4 specifically comprises the following steps:

[0032] S41, model loading, loading the latest version of the model currently trained on the cloud;

[0033] S42, incremental update, using incremental gradient descent, each time taking small batches of data from the new data to update the model weights until all the data are processed. In incremental gradient descent, the formula (3) for each parameter update is:

[0034]

[0035] (3) where θ represents the model weight, η is the learning rate used to control the step size of each update, and J(θ; x (i) ,y (i) ) is the loss function, is the gradient of the loss function with respect to the model parameters;

[0036] S43, convergence detection, judging whether the model has reached the expected convergence state;

[0037] S44, model evaluation, after each update, generates the prediction recall and precision of the prediction model as well as the F1 score.

[0038] Preferably, S5 specifically includes the following steps:

[0039] S51. Model conversion: use TensorFlow Lite Converter to convert the original TensorFlow model into a TensorFlow Lite model.

[0040] S52, model quantization, use TensorFlow Lit to quantize the floating-point model into an 8-bit integer model. The quantization formula is shown in (4):

[0041] r = S × (qz) (4);

[0042] (4) where r is the real value, q is the B-bit integer quantization representation, and S is the scaling factor, which is defined as:

[0043]

[0044] S determines the mapping ratio from floating point to integer, z is the zero point, which represents the offset referenced in the quantization process to map the floating point range;

[0045] S53. Test the accuracy and performance of the model using TensorFlow Lite Interpreter to ensure that the conversion did not introduce errors.

[0046] S54. Save the model and name it model_v_x.tflite;

[0047] S55. Generate a model version number, save it as version.json, and record the version number, release date, and accuracy of the current model;

[0048] S56, file integrity, generates the national secret SM3 hash check code of the model file model_v_x.tflite, and saves it in the version.json file.

[0049] Preferably, S6 specifically includes the following steps:

[0050] S61, Tina R329 start SSH remote connection, configure SSH service, modify / etc / ssh / sshd_config file, set PermitRootLogin and PasswordAuthentication to yes, and then restart SSH service using systemctl restartssh;

[0051] S62. Transfer the model from the cloud to Tina R329. Use SCP to transfer the local model_v_x.tflite model file to the / root / models / directory of Tina R329.

[0052] S63. Compare the version numbers and accuracy in the version.json file of the current version and the new version on Tina R329, and verify the integrity of the transferred file by verifying the consistency of the SM3 hash check codes of the local and cloud models. If they are consistent, choose to update the model. If they are inconsistent, the model will not be updated.

[0053] S64. Perform real-time reasoning and analysis on the data collected by the sensor group locally, quickly process the input data from the sensor group, process the data through the model, and the generated binary classification results will drive the system to make corresponding decisions and actions.

[0054] Therefore, the present invention adopts the above-mentioned cloud-based Tina R329 incremental learning and automatic model deployment method, and the beneficial effects are as follows:

[0055] (1) Compared with the traditional data processing method, the SMOTE algorithm of the present invention is an improved algorithm based on the random oversampling algorithm. It can analyze the minority class samples and add them to the data set by artificially synthesizing new samples based on the minority class samples. By increasing the data set of minority class samples, the minority class samples and majority class samples reach a certain ratio, thus achieving the sample rebalancing effect.

[0056] (2) The present invention provides intelligent adaptive capabilities for embedded devices by combining incremental learning with automatic model update technology, which can achieve real-time optimization in environmental changes and completely free from human intervention. Through efficient incremental updates and model quantization, it effectively reduces computing and storage resource consumption, ensuring that the device always has high performance and high reliability, while significantly reducing operating costs.

[0057] (3) The system provided by the present invention supports fully automated model deployment, has fast real-time response capabilities, and achieves a stable and efficient workflow through a flexible and secure cloud and device collaboration mechanism; its excellent scalability and compatibility enable the system to be widely used in smart homes, industrial automation, precision agriculture and other fields; when completing complex training tasks in the cloud, the system fully utilizes the advantages of low-power hardware, enabling embedded devices to maintain high performance and stability for a long time, thereby ensuring continuous functionality.

[0058] (4) In addition, the present invention uses the national secret SM3 algorithm to verify the transmitted files. By calculating the SM3 hash value of the file and comparing it with the received hash value, it can effectively verify whether the file has been tampered with during the transmission process and ensure its integrity; this verification mechanism can not only prevent the file from being tampered with or damaged during the transmission process, but also ensure the consistency and correctness of the data each time the model is updated and deployed, thereby further improving the security and reliability of the system.

[0059] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a step diagram of an embodiment of a cloud-based Tina R329 incremental learning and automatic model deployment method of the present invention;

[0061] Figure 2 It is a flow chart for realizing a fire alarm binary classification task according to an embodiment of a cloud-based Tina R329 incremental learning and automatic model deployment method of the present invention.

[0062] Reference numerals

[0063] 1. Environmental data acquisition subsystem; 2. Environmental data receiving and transmission subsystem; 21. Environmental data receiving module; 22. Environmental data transmission module; 3. Cloud database; 4. Model generation and verification subsystem; 41. Data preprocessing module; 42. Model generation module; 43. Data integrity verification module; 5. Model deployment subsystem; 6. Local reasoning subsystem; 61. Data loading module; 62. Data preprocessing module; 63. Data calculation and analysis module; 7. Comparison subsystem; 8. Abnormal indicator data processing subsystem; 9. Fire alarm display subsystem. DETAILED DESCRIPTION

[0064] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0065] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0066] The present invention uses an incremental learning method based on real-time data to enable the model to continuously optimize performance, reduce computing overhead, and improve the flexibility and continuous adaptability of the system. Automatic model updates are achieved with the help of a cloud computing platform, and the optimization and push of the model can be completed without human intervention, effectively reducing the computing pressure of local devices. The cloud analyzes and processes the uploaded data, and optimizes and pushes the model to the device side, so that the smart terminal can respond to environmental changes in real time and maintain efficient operation. Therefore, the "device-cloud" collaborative mode is particularly suitable for embedded scenarios such as smart homes that require long-term stable operation. Through real-time data interaction and dynamic adjustment, the system can continuously optimize performance, thereby significantly improving efficiency, enhancing autonomy, and achieving higher flexibility and intelligence.

[0067] Tina R329 collects data from the sensor group in real time and uploads it to the cloud server. After data preprocessing (including standardization and sample rebalancing) in the cloud, it trains a binary classification model based on the incremental learning method to gradually improve its prediction capabilities. The trained model undergoes format conversion and performance tuning to adapt to the embedded hardware environment and ensure efficient operation. The optimized model is automatically pushed to the Tina R329 local device via wireless communication to achieve continuous updates and iterations to maintain optimal performance. The combination of cloud computing and incremental learning significantly improves the intelligence level of the system, enabling it to respond and adapt to environmental changes in real time, fully meeting the long-term and efficient operation requirements of smart devices in complex scenarios.

[0068] A cloud-based Tina R329 incremental learning and automatic model deployment method includes the following steps:

[0069] S1. Data collection and standardization: Upload the ambient environment test data obtained by Tina R329 from the sensor group to the cloud, which specifically includes the following steps:

[0070] S11, the data is read by the acquisition sensor group;

[0071] S12. Configure the Wi-Fi module of Tina R238 and connect to the network;

[0072] S13, formatting the collected data into JSON format;

[0073] S14. Use the MQTT protocol to upload the data in JSON format to the cloud server through the specified topic;

[0074] S15, standardization, standardize the data according to formula (1):

[0075]

[0076] (1) In the formula, x′ is the value to be standardized, μ is the mean of the feature, and σ is the standard deviation.

[0077] S2, data rebalancing: rebalance the data obtained in S1, increase minority class samples, and improve the proportion of minority class samples, specifically including the following steps:

[0078] S21. For the training data, use the synthetic minority oversampling technique SMOTE (Synthetic Minority Oversampling Technique) to identify the samples belonging to the minority class from the training data set and determine the sample categories that need to be oversampled;

[0079] S22. For each minority sample, calculate the Euclidean distance between the minority sample and other samples, and select its K nearest neighbor samples;

[0080] S23. For each minority class sample x i , randomly select a neighbor x from its K nearest neighbors j , and use formula (2) to generate new samples:

[0081] x new =x i +λ·(x j -x i ) (2);

[0082] (2) In the formula, λ is a random number with a value range of [0, 1], which is used to control the new sample in x i and x j The greater the value of λ, the greater the distance between the generated new samples and x. i The further;

[0083] S24, repeatedly generating new samples until the expected number of synthetic samples is reached;

[0084] S25. Synthesize the newly synthesized sample data with the original data set to construct a new balanced data set.

[0085] S3, Model pre-training: Whenever new data is uploaded, the cloud system will decide whether re-training is needed based on the model version. The specific steps include:

[0086] S31. Check whether the current model version meets the business requirements. If so, retraining will not be triggered.

[0087] S32. If step S31 is no, the current model version does not meet the business requirements. The cloud server starts model pre-training. The training uses the full data set, 80% as the training set and 20% as the test set. A prediction model is generated through standard machine learning methods.

[0088] S4, incremental learning: fine-tune and optimize the data in S2 to avoid training from scratch and quickly adjust the model to adapt to new data features. Specifically, it includes the following steps:

[0089] S41, model loading, loading the latest version of the model currently trained on the cloud;

[0090] S42, incremental update, using incremental gradient descent, each time taking small batches of data from the new data to update the model weights until all the data are processed. In incremental gradient descent, the formula (3) for each parameter update is:

[0091]

[0092] (3) where θ represents the model weight, η is the learning rate used to control the step size of each update, and J(θ; x (i) ,y (i) ) is the loss function, is the gradient of the loss function with respect to the model parameters;

[0093] S43, convergence detection, judging whether the model has reached the expected convergence state;

[0094] S44, model evaluation, after each update, generates the prediction recall and precision of the prediction model as well as the F1 score.

[0095] S5, model quantization: convert the model parameters into a lower precision form to meet the operating requirements of the embedded hardware, including the following steps:

[0096] S51. Model conversion: use TensorFlow Lite Converter to convert the original TensorFlow model into a TensorFlow Lite model (.tflite format);

[0097] S52, model quantization, use TensorFlow Lit to quantize the floating-point model into an 8-bit integer model. The quantization formula is shown in (4):

[0098] r = S × (qz) (4);

[0099] (4) In the formula, r is a real value (usually a floating point number, such as float32), q is a B-bit integer (such as unit8, unit32, etc.) quantized representation, and S is a scaling factor (floating), which is defined as:

[0100]

[0101] S determines the mapping ratio from floating point to integer, and z is the zero point (integer), which represents the offset referenced during the quantization process to map the floating point range.

[0102] S53. Test the accuracy and performance of the model using TensorFlow Lite Interpreter to ensure that the conversion did not introduce errors.

[0103] S54. Save the model and name it model_v_x.tflite;

[0104] S55. Generate a model version number, save it as version.json, and record the version number, release date, and accuracy of the current model (e.g., version: 1.2., release_date: 2024-12-12, accuracy; 82.5%).

[0105] S56, file integrity, generates the national secret SM3 hash check code of the model file model_v_x.tflite, and saves it in the version.json file.

[0106] S6, model deployment: The optimized model is automatically pushed to the local device through wireless communication, and the local device deploys the latest model in real time without manual intervention, which specifically includes the following steps:

[0107] S61, Tina R329 start SSH remote connection, configure SSH service, modify / etc / ssh / sshd_config file, set PermitRootLogin and PasswordAuthentication to yes, and then restart SSH service using systemctlrestart ssh;

[0108] S62. Transfer the model from the cloud to Tina R329. Use SCP to transfer the local model_v_x.tflite model file to the / root / models / directory of Tina R329.

[0109] S63. Compare the version numbers and accuracy in the version.json file of the current version and the new version on Tina R329, and verify the integrity of the transferred file by verifying the consistency of the SM3 hash check codes of the local and cloud models. If they are consistent, choose to update the model. If they are inconsistent, the model will not be updated.

[0110] S64. Perform real-time reasoning and analysis on the data collected by the sensor group locally, quickly process the input data from the sensor group, and after processing the data through the model, the generated binary classification results will drive the system to make corresponding decisions and actions, thereby achieving intelligent response and automated control.

[0111] Example

[0112] like Figure 1 As shown, the technical solution of the present invention includes six steps: data preprocessing, data rebalancing, model training, incremental learning, model quantization and model deployment.

[0113] S1. Data preprocessing (data collection and annotation): The network traffic test data set is uploaded to the database after being standardized. This data includes a large amount of audit data generated by the computer network.

[0114] S2, data rebalancing: rebalancing the data obtained after data preprocessing, increasing minority class samples and improving the proportion of minority class samples;

[0115] S3, model training: The cloud system decides whether retraining is needed based on the model version;

[0116] S4, incremental learning: fine-tune and optimize the data in step S2 to avoid training from scratch and quickly adjust the model to adapt to new data features;

[0117] S5, model quantization: converting model parameters into a lower precision form to meet the operating requirements of embedded hardware;

[0118] S6, Model deployment: The optimized model will be automatically pushed to the local device via wireless communication, and the device can deploy the latest model in real time;

[0119] like Figure 2 As shown, the model process of the present invention is divided into the following steps:

[0120] 1. Environmental data acquisition subsystem (1)

[0121] It is composed of a group of external sensors used to collect environmental data in real time, such as temperature, humidity, gas concentration, etc. These sensors will pass the acquired information to the subsequent receiving and transmitting modules.

[0122] 2. Environmental data receiving and transmission subsystem (2)

[0123] Environmental data receiving module (21): responsible for obtaining raw data from the environmental data acquisition subsystem (1) and performing preliminary formatting or packaging processing;

[0124] Environmental data transmission module (22): uploads the received environmental data to the cloud database (3) via a wireless network, and transmits the instructions or models issued by the cloud back to the local computer when necessary.

[0125] 3. Cloud database (3)

[0126] Used to store and manage environment data, model versions, and related business configurations; when new data is uploaded locally or the model needs to be updated, the database will be used for unified management and call.

[0127] 4. Model generation and verification subsystem (4)

[0128] The data prediction processing module (41) performs basic operations such as missing value processing, noise filtering, and feature extraction on the incoming data to ensure the quality of the data used in subsequent training;

[0129] Model generation module (42): performs model training or incremental learning based on preprocessed data, and updates or generates a new classification model;

[0130] Data integrity verification module (43): The consistency and integrity of the data used in the model training process and the final generated model are reviewed. Only after this link is passed can the model be officially released and deployed.

[0131] 5. Model deployment subsystem (5)

[0132] When the new model in the cloud is verified, the updated model will be distributed to the Tina R329 device locally through a secure transmission protocol (such as SCP) to replace or incrementally update the existing inference model.

[0133] 6. Local Reasoning Subsystem (6)

[0134] Data loading module (61):

[0135] Responsible for receiving and reading the raw environmental data collected by sensors, or loading the input information required for the model from the cloud and local storage, and preparing data for subsequent processing and reasoning.

[0136] Data preprocessing module (62):

[0137] The loaded data is preprocessed by cleaning, denoising, normalizing, feature extraction, etc., so that the data format and content meet the requirements of model reasoning. At the same time, feature reorganization or encoding conversion can be performed according to the input requirements of different models.

[0138] Data calculation and analysis module (63):

[0139] After preprocessing, the data is inferred in real time or analyzed regularly in combination with the deployed model to obtain classification results, anomaly detection information, etc. The corresponding business logic or alarm mechanism is triggered according to the inference results, ultimately improving the system's ability to respond quickly to environmental changes.

[0140] 7. Comparison subsystem (7)

[0141] Compare the latest inference results with historical data or thresholds. If the difference is within a reasonable range, continue to observe; otherwise, further confirm or process the abnormal situation.

[0142] 8. Abnormal indicator data processing subsystem (8)

[0143] Data analysis and feature screening are performed on abnormal indicators that appear during the reasoning process (such as sudden temperature increase, excessive gas concentration, etc.) to determine whether it is a short-term fluctuation or an abnormal trend, and processing suggestions are output.

[0144] 9. Fire alarm display subsystem (9)

[0145] When the system determines that a serious abnormality or disaster occurs, an alarm signal is sent to this subsystem. This subsystem is responsible for visually displaying the alarm information, sending notifications (such as sound and light alarms, information reporting), etc., to help relevant personnel take timely response measures.

[0146] The present invention collects environmental data from local sensors and uploads it to a cloud database to implement data integrity verification and incremental model training. The updated model is then distributed locally using the model deployment subsystem. It combines data calculation and analysis, reasoning, abnormal indicator processing, and alarm display functions to form a complete cloud-local linkage intelligent detection and response process.

[0147] Therefore, the present invention adopts the above-mentioned cloud-based Tina R329 incremental learning and automatic model deployment method, which not only ensures the high accuracy of Tina R329 in real-time reasoning, but also meets the ever-changing business needs through automated model iteration deployment.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A cloud-based Tina R329 incremental learning and automatic model deployment method, characterized in that: The following steps are involved: S1, data collection and standardization: upload the surrounding environment test data obtained by Tina R329 from the sensor group to the cloud; S2, data rebalancing: rebalance the data obtained in S1, increase minority class samples, and improve the proportion of minority class samples; S3, Model pre-training: Whenever new data is uploaded, the cloud system decides whether re-training is needed based on the model version; S4, incremental learning: fine-tune and optimize the data in S2 to avoid training from scratch and quickly adjust the model to adapt to new data features; S5, model quantization: converting model parameters into a lower precision form to meet the operating requirements of embedded hardware; S6. Model deployment: The optimized model is automatically pushed to the local device via wireless communication, and the local device deploys the latest model in real time.

2. According to a cloud-based Tina R329 incremental learning and automatic model deployment method according to claim 1, it is characterized in that: S1 specifically includes the following steps: S11, the data is read by the acquisition sensor group; S12. Configure the Wi-Fi module of Tina R238 and connect to the network; S13, formatting the collected data into JSON format; S14. Use the MQTT protocol to upload the data in JSON format to the cloud server through the specified topic; S15, standardization, standardize the data according to formula (1): (1) In the formula, x′ is the value to be standardized, μ is the mean of the feature, and σ is the standard deviation.

3. The cloud-based Tina R329 incremental learning and automatic model deployment method according to claim 2, characterized in that: S2 specifically includes the following steps: S21. For the training data, use the synthetic minority oversampling technique SMOTE to identify the samples belonging to the minority class from the training data set and determine the sample categories that need to be oversampled; S22. For each minority sample, calculate the Euclidean distance between the minority sample and other samples, and select its K nearest neighbor samples; S23. For each minority class sample x i , randomly select a neighbor x from its K nearest neighbors j , and use formula (2) to generate new samples: x new =x i +λ·(x j -x i ) (2); (2) In the formula, λ is a random number with a value range of [0, 1], which is used to control the new sample in x i and x j The greater the value of λ, the greater the distance between the generated new samples and x. i The further; S24, repeatedly generating new samples until the expected number of synthetic samples is reached; S25. Synthesize the newly synthesized sample data with the original data set to construct a new balanced data set.

4. The cloud-based Tina R329 incremental learning and automatic model deployment method according to claim 3, characterized in that: S3 specifically includes the following steps: S31. Check whether the current model version meets the business requirements. If so, retraining will not be triggered. S32. If the answer to step S31 is no, the current model version does not meet the business requirements. The cloud server starts model pre-training. The training uses the full data set, 80% as the training set and 20% as the test set. A prediction model is generated through standard machine learning methods.

5. The cloud-based Tina R329 incremental learning and automatic model deployment method according to claim 4, characterized in that: S4 specifically includes the following steps: S41, model loading, loading the latest version of the model currently trained on the cloud; S42, incremental update, using incremental gradient descent, each time taking small batches of data from the new data to update the model weights until all the data are processed. In incremental gradient descent, the formula (3) for each parameter update is: (3) where θ represents the model weight, η is the learning rate used to control the step size of each update, and J(θ; x (i) ,y (i) ) is the loss function, is the gradient of the loss function with respect to the model parameters; S43, convergence detection, judging whether the model has reached the expected convergence state; S44, model evaluation, after each update, generates the prediction recall and precision of the prediction model as well as the F1 score.

6. The cloud-based Tina R329 incremental learning and automatic model deployment method according to claim 5, characterized in that: S5 specifically includes the following steps: S51. Model conversion: use TensorFlow Lite Converter to convert the original TensorFlow model into a TensorFlow Lite model. S52, model quantization, use TensorFlow Lit to quantize the floating-point model into an 8-bit integer model. The quantization formula is shown in (4): r = S × (qz) (4); (4) where r is the real value, q is the B-bit integer quantization representation, and S is the scaling factor, which is defined as: S determines the mapping ratio from floating point to integer, z is the zero point, which represents the offset referenced in the quantization process to map the floating point range; S53. Test the accuracy and performance of the model using TensorFlow Lite Interpreter to ensure that the conversion did not introduce errors. S54. Save the model and name it model_v_x.tflite; S55. Generate a model version number, save it as version.json, and record the version number, release date, and accuracy of the current model; S56, file integrity, generates the national secret SM3 hash check code of the model file model_v_x.tflite and stores it in the version.json file.

7. The cloud-based Tina R329 incremental learning and automatic model deployment method according to claim 6, characterized in that: S6 specifically includes the following steps: S61, Tina R329 start SSH remote connection, configure SSH service, modify / etc / ssh / sshd_config file, set PermitRootLogin and PasswordAuthentication to yes, and then restart SSH service using systemctl restart ssh; S62. Transfer the model from the cloud to Tina R329. Use SCP to transfer the local model_v_x.tflite model file to the / root / models / directory of Tina R329. S63. Compare the version numbers and accuracy in the version.json file of the current version and the new version on Tina R329, and verify the integrity of the transferred file by verifying the consistency of the SM3 hash check codes of the local and cloud models. If they are consistent, choose to update the model. If they are inconsistent, the model will not be updated. S64. Perform real-time reasoning and analysis on the data collected by the sensor group locally, quickly process the input data from the sensor group, process the data through the model, and the generated binary classification results will drive the system to make corresponding decisions and actions.

Images