Self-adaptive intelligent navigation system based on depth information and online incremental learning

Abstract

The invention relates to the technical field of intelligent navigation, in particular to a self-adaptive intelligent navigation system based on depth information and online incremental learning, which comprises a multi-mode self-adaptive perception and fusion module, an uncertainty-driven online safety learning engine and swarm intelligent update. According to the method, advantage complementation and robustness improvement of perception capability are realized, catastrophic forgetting is effectively resisted, safe and stable lifelong learning is realized, encryption is carried out during transmission, Gaussian noise subjected to sensitivity calibration is added to model updating before a local model is uploaded, and meanwhile, a MomentsAccount mechanism is used for accurately accumulating and controlling long-term privacy consumption.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent navigation, in particular to an adaptive intelligent navigation system based on depth information and online incremental learning. BACKGROUND

[0002] With the advancement of artificial intelligence and sensor technology, intelligent navigation systems (such as autonomous driving, robot navigation, and assistive blind devices) are increasingly widely used in complex environments. Existing intelligent navigation systems, such as the intelligent navigation system disclosed in the invention patent with publication number CN104185775B and the safety operation intelligent navigation system disclosed in the invention patent with publication number CN119146974B, can all achieve intelligent navigation of vehicles.

[0003] However, in use, it is found that the robustness of the existing intelligent navigation system in complex environments is poor, and the existing navigation system usually relies on pre-trained static models. Once these models are deployed, their performance is fixed and cannot continuously learn and self-optimize from new or uncertain scenarios encountered in actual use, making the adaptability of the models poor. In addition, traditional centralized learning mode requires uploading all user data to the cloud for model training, which involves a large amount of sensitive individual location and behavior data in navigation applications, posing a serious risk of user privacy leakage. Therefore, there is an urgent need for an adaptive intelligent navigation system based on depth information and online incremental learning to improve the above problems. SUMMARY

[0004] To solve the above technical problems, the present application provides an adaptive intelligent navigation system based on depth information and online incremental learning, which realizes the complementary advantage of perception ability and the improvement of robustness, effectively counteracts catastrophic forgetting, realizes safe and stable lifelong learning, encrypts during transmission, and adds Gaussian noise calibrated by sensitivity to model updates before uploading the local model. At the same time, the MomentsAccountant mechanism is used to accurately accumulate and control long-term privacy consumption.

[0005] The adaptive intelligent navigation system based on depth information and online incremental learning of the present application comprises: Multi-modal adaptive perception and fusion module: robust and comprehensive environmental information is obtained in complex environments to provide accurate data support for navigation decision-making; Uncertainty-driven online safety learning engine: in actual use, continuously learn and self-optimize for new or predicted uncertain scenarios while ensuring the stability and safety of the learning process; Group intelligence update: integrate the data experience of all user devices through federated learning to realize model co-evolution while strictly protecting user privacy.

[0006] Preferably, the multi-modal adaptive perception and fusion module comprises: Heterogeneous sensor array: adopt the complementary combination of depth sensor and high-speed dynamic vision module, the depth sensor can provide absolute and accurate depth information, help the system accurately perceive the distance and spatial position of the surrounding objects, and the high-speed dynamic vision module has the ability of dynamic information capture with microsecond level delay, which can effectively deal with complex situations such as high-speed motion and dramatic changes in light, and ensure that the system can obtain effective visual information in real time in dynamic environment; Adaptive weighted fusion algorithm: combine the accurate depth information of the depth sensor and the high-speed dynamic information of the high-speed dynamic vision module to make up for their respective shortcomings and form a comprehensive and robust perception of the environment; Fusion formula-weighted complementary filtering: for a certain point in the scene, the final estimated depth value Can be expressed as: Wherein: is the depth value directly measured by the depth sensor; is the depth change calculated by the event stream; and are adaptive weights, in static or slow scenes, higher, in high-speed motion or dramatic changes in light, rise; Lightweight multi-task neural network: Input: fused depth map, event stream cumulative map; Network structure: lightweight convolutional neural network or visual Transformer Output task: 1), semantic segmentation: classify each pixel; Loss function: cross-entropy loss is usually used; Wherein is the cross-entropy loss, represents the th pixel, is the total number of pixels, is the number of categories, is the true label, is the predicted probability; 2), obstacle detection and distance estimation: identify the bounding box of the obstacle and calculate the accurate distance from the user; 3), passable area detection.

[0007] Preferably, the depth sensor is a ToF camera, which directly acquires the depth information of the scene by emitting modulated near-infrared light waves toward the target and calculating the time or phase difference that the light waves take from emission to reception by the sensor.

[0008] Preferably, the phase-based ToF ranging principle is as follows: The phase change of the modulated light wave as it propagates through the air With flight time and distance The relationship is: in: It is the modulation frequency; It's the speed of light; It is the distance to be measured; Therefore, by measuring the phase difference The distance can be calculated: d =4 π c ⋅Δ ϕ .

[0009] Preferably, the high-speed dynamic vision module is an event camera, which uses a bionic human retina. Each pixel works independently, and when the brightness change of a pixel exceeds a set threshold, the pixel asynchronously outputs an event containing the pixel coordinates. x , y ), timestamp t and polarity p .

[0010] Preferably, the event triggering condition is: An event in pixels and time It is triggered if and only if the logarithmic change in brightness satisfies: in: It is the brightness at the current moment; It is the brightness of the last time the event was triggered; It is the contrast threshold.

[0011] Preferably, the uncertainty-driven online security learning engine includes: Uncertainty monitoring and triggering: Method: Monte Carlo Dropout; During reasoning, for the same input conduct After one random forward propagation, the set of probability distributions is obtained. ; Quantification of uncertainty: Prediction entropy: measures overall uncertainty; in It is predictive entropy, which measures overall uncertainty. The average predicted probability is obtained by averaging the results of T predictions. The total number of categories is a constant. Cognitive uncertainty: a measure of the uncertainty inherent in the model itself; Triggering condition: When When the cognitive uncertainty threshold is reached, the sample is determined to be a "valuable new sample," triggering the online learning process. in It is cognitive uncertainty (mutual information). It is the threshold of cognitive uncertainty; Fine-tuning to combat catastrophic forgetting: Continuous learning based on knowledge distillation: fine-tuning the model with new samples while preventing catastrophic forgetting of old knowledge; process: 1) Fixed Teacher Model: The current model on the device is used as the teacher model, and its parameters are frozen; 2) Data preparation: New samples that will trigger learning Compared to an old sample randomly sampled from the local cache Forming a micro-batch, For output data; 3) Student Model Update: Initialize a student model identical to the teacher model and update it using the following loss function: It is the overall loss function, which consists of two weighted parts: cross-entropy loss. and knowledge distillation loss ; The balancing parameter is usually set to a smaller value to prioritize the retention of older knowledge. Representing the teacher model, Representing the student model; The true label of a new sample is a vector, usually encoded using one-hot encoding; Standard cross-entropy loss ensures that the student model learns new knowledge; This represents the value (0 or 1) of the actual label in category cc. : Category index, summation of all classes; The student model's predicted probability distribution for new samples is a vector, where... This indicates that the student model predicts the category to which a new sample belongs. The probability of; Knowledge distillation loss ensures that the student model retains the old knowledge learned from the teacher model; The teacher model's predicted probability distribution for old samples is a vector, where... This indicates that the teacher model predicts the category to which the old sample belongs. The probability of; The student model's predicted probability distribution for old samples is a vector, where... This indicates that the student model predicts the category to which the old sample belongs. The probability of; Temperature parameter is a constant used in knowledge distillation to determine the softening probability distribution; higher values ​​indicate a higher probability distribution. This makes the probability distribution smoother, which helps the student model learn "tacit knowledge" from the teacher model. (The formula...) Used to scale the K-divergence to balance the gradient; KL divergence is used to measure the difference between two probability distributions. Model replacement: After fine-tuning, replace the original online model with the student model.

[0012] Federal average algorithm: process: 1) Server broadcast: The central server broadcasts the current global model. Distribute to all eligible online devices; 2) Local training: for each participating device Use local private data Fine-tuning the global model yields the local model. The goal of local training is to minimize the local loss. ; 3) Model Upload: The device will upload the updated model parameters. Upload to the server after encryption; 4) Federated average aggregation: The server collects at least The model is updated for each device, and then a weighted average is calculated: in It is the original model update for the k-th device. This is a model update after adding noise. It is Gaussian noise with a mean of 0 and a variance of . The Gaussian distribution is given, where I is the identity matrix, representing the independent addition of noise to each model parameter. It is the sensitivity of model updates. It is a noise scale; Privacy Budget: Use Differential privacy is used to rigorously quantify the degree of privacy protection. A privacy budget is a measure of the risk of privacy breaches. The smaller the size, the stronger the privacy protection. It represents the probability of failure, allowing the privacy protection mechanism to fail with a very small probability. - Differential privacy indicates that The probability may not satisfy -Differential privacy; The system sets the privacy budget for each aggregation step to be [specified]. This is a very strict level of protection. The total privacy consumption is accumulated through the MomentsAccountant mechanism to ensure that privacy will not be exhausted during long-term training.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. A perception array is constructed using two specific heterogeneous sensors: a ToF depth camera and an event camera. Through an adaptive weighted fusion algorithm, the weights of depth information and event information are dynamically adjusted according to the static / slow motion vs. high-speed motion / drastic changes in lighting of the scene, so as to achieve complementary advantages and improved robustness of perception capabilities. 2. The triggering mechanism of the uncertainty-driven online secure learning engine adopts the Monte Carlo Dropout method to perform multiple inferences and calculates cognitive uncertainty as a quantitative indicator. When the threshold is exceeded, learning is automatically triggered, which makes learning happen "on demand", efficient and intelligent. The learning mechanism adopts an incremental learning method based on knowledge distillation. The key is to use micro-batches of "new samples + old samples" and optimize the loss of new knowledge learning and old knowledge distillation by balancing the parameter α, effectively combating catastrophic forgetting and achieving safe and stable lifelong learning. 3. In the specific application scenario of the navigation system, a collaborative architecture of terminal-edge-cloud is constructed, which integrates federated learning and differential privacy technology. A complete federated learning process is defined, which includes model broadcasting, local fine-tuning, encrypted upload, and weighted aggregation, and is carried out at a fixed period. Not only is the transmission encrypted, but Gaussian noise with sensitivity calibration is added to the model update before the local model is uploaded. Strict (ε,δ)-differential privacy (such as ε=0.1) is used for quantitative protection. At the same time, the MomentsAccountant mechanism is used to accurately accumulate and manage long-term privacy consumption. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the ToF camera of this invention. Figure 2 This is a schematic diagram of the shooting structure of the event camera of the present invention; Figure 3 This is a schematic diagram of the present invention in real-time use. Detailed Implementation

[0015] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. The present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0016] Example like Figures 1 to 3 As shown, an adaptive intelligent navigation system based on deep information and online incremental learning includes: Multimodal adaptive perception and fusion module: Acquires robust and comprehensive environmental information in complex environments, providing accurate data support for navigation decisions; Uncertainty-driven online security learning engine: In practical use, it continuously learns and optimizes itself for newly encountered or predicted uncertain scenarios, while ensuring the stability and security of the learning process; Swarm intelligence update: By integrating data and experience from all user devices through federated learning, the model can evolve collaboratively while strictly protecting user privacy.

[0017] The multimodal adaptive sensing and fusion module includes: Heterogeneous sensor array: It adopts a complementary combination of depth sensor and high-speed dynamic vision module. The depth sensor can provide absolute and accurate depth information to help the system accurately perceive the distance and spatial position of surrounding objects. The high-speed dynamic vision module has the ability to capture dynamic information with microsecond-level latency, which can effectively cope with complex situations such as high-speed motion and drastic changes in lighting, ensuring that the system can acquire effective visual information in real time in dynamic environments. Adaptive weighted fusion algorithm: Combines the precise depth information from the depth sensor and the high-speed dynamic information from the high-speed dynamic vision module to make up for their respective shortcomings and form a comprehensive and robust perception of the environment. Fusion Formula - Weighted Complementary Filtering: For a given point in the scene, the final estimated depth value... It can be represented as: in: It is the depth value directly measured by the depth sensor; It is the depth change calculated from the event flow (estimated through techniques such as optical flow and motion structure recovery). and It uses adaptive weights, especially in static or slow-moving scenarios. Higher, especially during high-speed movement or drastic changes in light. Increase; Lightweight multi-task neural networks: Inputs: fused depth map, event flow cumulative map; Network architecture: Lightweight convolutional neural networks or visual Transformers Output task: 1) Semantic segmentation: classifying each pixel; Loss function: Cross-entropy loss is typically used; in It is cross-entropy loss. It is the total number of pixels. It is the number of categories. It's a real label. It is a predicted probability; 2) Obstacle detection and distance estimation: Identify the bounding boxes of obstacles and calculate their precise distance from the user; 3) Detection of passable areas.

[0018] The depth sensor is a ToF camera, which directly acquires the depth information of the scene by emitting modulated near-infrared light waves toward the target and calculating the time or phase difference that the light waves take from emission to reception by the sensor.

[0019] The phase-based ToF ranging principle: The phase change of the modulated light wave as it propagates through the air With flight time and distance The relationship is: in: It is the modulation frequency; It's the speed of light; It is the distance to be measured; Therefore, by measuring the phase difference The distance can be calculated: d =4 π c ⋅Δ ϕ .

[0020] The high-speed dynamic vision module is an event camera that uses a bionic human retina. Each pixel works independently, and when the brightness change of a pixel exceeds a set threshold, the pixel asynchronously outputs an event containing the pixel coordinates. x , y ), timestamp t (Microsecond level) and polarity p (Indicates brightening or darkening).

[0021] The event triggering conditions are as follows: An event in pixels and time It is triggered if and only if the logarithmic change in brightness satisfies: in: It is the brightness at the current moment; It is the brightness of the last time the event was triggered; It is the contrast threshold.

[0022] The uncertainty-driven online security learning engine includes: Uncertainty monitoring and triggering: Method: Monte Carlo Dropout; During reasoning, for the same input conduct After one random forward propagation, the set of probability distributions is obtained. ; Quantification of uncertainty: Prediction entropy: measures overall uncertainty; in It is predictive entropy, which measures overall uncertainty. The average predicted probability is obtained by averaging the results of T predictions. The total number of categories is a constant. Cognitive uncertainty: a measure of the uncertainty inherent in the model itself; Triggering condition: When When the cognitive uncertainty threshold is reached, the sample is determined to be a "valuable new sample," triggering the online learning process. in It is cognitive uncertainty (mutual information). It is the threshold of cognitive uncertainty; Fine-tuning to combat catastrophic forgetting: Continuous learning based on knowledge distillation: fine-tuning the model with new samples while preventing catastrophic forgetting of old knowledge; process: 1) Fixed Teacher Model: The current model on the device is used as the teacher model, and its parameters are frozen; 2) Data preparation: New samples that will trigger learning Compared to an old sample randomly sampled from the local cache Forming a micro-batch, For output data; 3) Student Model Update: Initialize a student model identical to the teacher model and update it using the following loss function: It is the overall loss function, which consists of two weighted parts: cross-entropy loss. and knowledge distillation loss ; The balancing parameter is usually set to a smaller value to prioritize the retention of older knowledge. Representing the teacher model, Representing the student model; The true label of a new sample is a vector, usually encoded using one-hot encoding; Standard cross-entropy loss ensures that the student model learns new knowledge; This represents the value (0 or 1) of the actual label in category cc. : Category index, summation of all classes; The student model's predicted probability distribution for new samples is a vector, where... This indicates that the student model predicts the category to which a new sample belongs. The probability of; Knowledge distillation loss ensures that the student model retains the old knowledge learned from the teacher model; The teacher model's predicted probability distribution for old samples is a vector, where... This indicates that the teacher model predicts the category to which the old sample belongs. The probability of; The student model's predicted probability distribution for old samples is a vector, where... This indicates that the student model predicts the category to which the old sample belongs. The probability of; Temperature parameter is a constant used in knowledge distillation to determine the softening probability distribution; higher values ​​indicate a higher probability distribution. This makes the probability distribution smoother, which helps the student model learn "tacit knowledge" from the teacher model. (The formula...) Used to scale the K-divergence to balance the gradient; KL divergence is used to measure the difference between two probability distributions. Model replacement: After fine-tuning, replace the original online model with the student model.

[0023] The swarm intelligence update includes: Federal average algorithm: process: 1) Server broadcast: The central server broadcasts the current global model. Distribute to all eligible online devices; 2) Local training: for each participating device Use local private data Fine-tuning the global model yields the local model. The goal of local training is to minimize the local loss. ; 3) Model Upload: The device will upload the updated model parameters. Upload to the server after encryption; 4) Federated average aggregation: The server collects at least The model is updated for each device, and then a weighted average is calculated: in It is the original model update for the k-th device. This is a model update after adding noise. It is Gaussian noise with a mean of 0 and a variance of . The Gaussian distribution is given, where I is the identity matrix, representing the independent addition of noise to each model parameter. It is the sensitivity of model updates. It is a noise scale; Privacy Budget: Use Differential privacy is used to rigorously quantify the degree of privacy protection. A privacy budget is a measure of the risk of privacy breaches. The smaller the size, the stronger the privacy protection. It represents the probability of failure, allowing the privacy protection mechanism to fail with a very small probability. - Differential privacy indicates that The probability may not satisfy -Differential privacy; The system sets the privacy budget for each aggregation step to be [specified]. This is a very strict level of protection. The total privacy consumption is accumulated through the MomentsAccountant mechanism to ensure that privacy will not be exhausted during long-term training.

[0024] The main functions achieved by this invention are: 1. A perception array is constructed using two specific heterogeneous sensors: a ToF depth camera and an event camera. Through an adaptive weighted fusion algorithm, the weights of depth information and event information are dynamically adjusted according to the static / slow motion vs. high-speed motion / drastic changes in lighting of the scene, so as to achieve complementary advantages and improved robustness of perception capabilities. 2. The triggering mechanism of the uncertainty-driven online secure learning engine adopts the Monte Carlo Dropout method to perform multiple inferences and calculates cognitive uncertainty as a quantitative indicator. When the threshold is exceeded, learning is automatically triggered, which makes learning happen "on demand", efficient and intelligent. The learning mechanism adopts an incremental learning method based on knowledge distillation. The key is to use micro-batches of "new samples + old samples" and optimize the loss of new knowledge learning and old knowledge distillation by balancing the parameter α, effectively combating catastrophic forgetting and achieving safe and stable lifelong learning. 3. In the specific application scenario of the navigation system, a collaborative architecture of terminal-edge-cloud is constructed, which integrates federated learning and differential privacy technology. A complete federated learning process is defined, which includes model broadcasting, local fine-tuning, encrypted upload, and weighted aggregation, and is carried out at a fixed period. Not only is the transmission encrypted, but Gaussian noise with sensitivity calibration is added to the model update before the local model is uploaded. Strict (ε,δ)-differential privacy (such as ε=0.1) is used for quantitative protection. At the same time, the MomentsAccountant mechanism is used to accurately accumulate and manage long-term privacy consumption.

[0025] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An adaptive intelligent navigation system based on deep information and online incremental learning, characterized in that, include: Multimodal adaptive perception and fusion module: Acquires robust and comprehensive environmental information in complex environments, providing accurate data support for navigation decisions; Uncertainty-driven online security learning engine: In practical use, it continuously learns and optimizes itself for newly encountered or predicted uncertain scenarios, while ensuring the stability and security of the learning process; Swarm intelligence update: By integrating data and experience from all user devices through federated learning, the model can evolve collaboratively while strictly protecting user privacy.

2. The adaptive intelligent navigation system based on deep information and online incremental learning as described in claim 1, characterized in that, The multimodal adaptive sensing and fusion module includes: Heterogeneous sensor array: It adopts a complementary combination of depth sensor and high-speed dynamic vision module. The depth sensor can provide absolute and accurate depth information to help the system accurately perceive the distance and spatial position of surrounding objects. The high-speed dynamic vision module has the ability to capture dynamic information with microsecond-level latency, which can effectively cope with complex situations such as high-speed motion and drastic changes in lighting, ensuring that the system can acquire effective visual information in real time in dynamic environments. Adaptive weighted fusion algorithm: Combines the precise depth information from the depth sensor and the high-speed dynamic information from the high-speed dynamic vision module to make up for their respective shortcomings and form a comprehensive and robust perception of the environment. Fusion Formula - Weighted Complementary Filtering: For a given point in the scene, the final estimated depth value... It can be represented as: in: It is the depth value directly measured by the depth sensor; It is the depth change calculated from the event stream; and It uses adaptive weights, especially in static or slow-moving scenarios. Higher, especially during high-speed movement or sudden changes in light. Increase; Lightweight multi-task neural networks: Inputs: fused depth map, event flow cumulative map; Network architecture: Lightweight convolutional neural networks or visual Transformers Output task: 1) Semantic segmentation: classifying each pixel; Loss function: Cross-entropy loss is typically used; in It is cross-entropy loss. It represents the first 1 pixel, It is the total number of pixels. It is the number of categories. It's a real label. It is a predicted probability; 2) Obstacle detection and distance estimation: Identify the bounding boxes of obstacles and calculate their precise distance from the user; 3) Detection of passable areas.

3. The adaptive intelligent navigation system based on depth information and online incremental learning as described in claim 2, wherein the depth sensor is a ToF camera, which directly acquires the depth information of the scene by emitting modulated near-infrared light waves toward the target and calculating the time or phase difference experienced by the light waves from emission to reception by the sensor.

4. The adaptive intelligent navigation system based on deep information and online incremental learning as described in claim 3, characterized in that, The phase-based ToF ranging principle: The phase change of the modulated light wave as it propagates through the air With flight time and distance The relationship is: in: It is the modulation frequency; It's the speed of light; It is the distance to be measured; Therefore, by measuring the phase difference The distance can be calculated: d =4 π c ⋅Δ ϕ 。 5. The adaptive intelligent navigation system based on deep information and online incremental learning as described in claim 2, characterized in that, The high-speed dynamic vision module is an event camera that uses a bionic human retina. Each pixel works independently, and when the brightness change of a pixel exceeds a set threshold, the pixel asynchronously outputs an event containing pixel coordinates (x, y) and a timestamp. t and polarity p .

6. The adaptive intelligent navigation system based on deep information and online incremental learning as described in claim 5, characterized in that, The event triggering conditions are as follows: An event in pixels and time It is triggered if and only if the logarithmic change in brightness satisfies: in: It is the brightness at the current moment; It is the brightness of the last time the event was triggered; It is the contrast threshold.

7. The adaptive intelligent navigation system based on deep information and online incremental learning as described in claim 1, characterized in that, The uncertainty-driven online security learning engine includes: Uncertainty monitoring and triggering: Method: Monte Carlo Dropout; During reasoning, for the same input conduct After one random forward propagation, the set of probability distributions is obtained. ; Quantification of uncertainty: Prediction entropy: measures overall uncertainty; in It is predictive entropy, which measures overall uncertainty. The average predicted probability is obtained by averaging the results of T predictions. The total number of categories is a constant. Cognitive uncertainty: a measure of the uncertainty inherent in the model itself; Triggering condition: When When the cognitive uncertainty threshold is reached, the sample is determined to be a "valuable new sample", triggering the online learning process; in It is cognitive uncertainty (mutual information). It is the threshold of cognitive uncertainty; Fine-tuning to combat catastrophic forgetting: Continuous learning based on knowledge distillation: fine-tuning the model with new samples while preventing catastrophic forgetting of old knowledge; process: 1) Fixed Teacher Model: The current model on the device is used as the teacher model, and its parameters are frozen; 2) Data preparation: New samples that will trigger learning Compared to an old sample randomly sampled from the local cache Forming a micro-batch, For output data; 3) Student Model Update: Initialize a student model identical to the teacher model and update it using the following loss function: It is the overall loss function, which consists of two weighted parts: cross-entropy loss. and knowledge distillation loss ; The balancing parameter is usually set to a smaller value to prioritize the retention of older knowledge. Representing the teacher model, Representing the student model; The true label of a new sample is a vector, usually encoded using one-hot encoding; Standard cross-entropy loss ensures that the student model learns new knowledge; This represents the value (0 or 1) of the actual label in category cc. : Category index, summation of all classes; The student model's predicted probability distribution for new samples is a vector, where... This indicates that the student model predicts the category to which a new sample belongs. The probability of; Knowledge distillation loss ensures that the student model retains the old knowledge learned from the teacher model; The teacher model's predicted probability distribution for old samples is a vector, where... This indicates that the teacher model predicts the category to which the old sample belongs. The probability of; The student model's predicted probability distribution for old samples is a vector, where... This indicates that the student model predicts the category to which the old sample belongs. The probability of; Temperature parameter is a constant used in knowledge distillation to determine the softening probability distribution; higher values ​​indicate a higher probability distribution. This makes the probability distribution smoother, which helps the student model learn "tacit knowledge" from the teacher model. (The formula is incomplete.) Used to scale the K-divergence to balance the gradient; KL divergence is used to measure the difference between two probability distributions. Model replacement: After fine-tuning, replace the original online model with the student model.

8. The adaptive intelligent navigation system based on deep information and online incremental learning as described in claim 1, characterized in that, The swarm intelligence update includes: Federal average algorithm: process: 1) Server broadcast: The central server broadcasts the current global model. Distribute to all eligible online devices; 2) Local training: for each participating device Use local private data Fine-tuning the global model yields the local model. The goal of local training is to minimize the local loss. ; 3) Model Upload: The device will upload the updated model parameters. Upload to the server after encryption; 4) Federated average aggregation: The server collects at least The model is updated for each device, and then a weighted average is calculated: in It is the original model update for the k-th device. This is a model update after adding noise. It is Gaussian noise with a mean of 0 and a variance of . The Gaussian distribution is given, where I is the identity matrix, representing the independent addition of noise to each model parameter. It is the sensitivity of model updates. It is a noise scale; Privacy Budget: Use Differential privacy is used to rigorously quantify the degree of privacy protection. A privacy budget is a measure of the risk of privacy breaches. The smaller the size, the stronger the privacy protection. It represents the probability of failure, allowing the privacy protection mechanism to fail with a very small probability. - Differential privacy indicates that The probability may not satisfy -Differential privacy; The system sets the privacy budget for each aggregation step to be [specified]. This is a very strict level of protection. The total privacy consumption is accumulated through the MomentsAccountant mechanism to ensure that privacy will not be exhausted during long-term training.

Images