AI glasses data high-speed read-write method and system based on a storage chip

Abstract

The application relates to the technical field of AI glasses, and discloses an AI glasses data high-speed reading and writing method and system based on a storage chip, the method comprising the following steps: acquiring a data stream of an AI glasses sensing unit; performing preliminary classification on the data stream to generate a preliminary classification label; extracting a corresponding access frequency and judging a trend to obtain a dynamic change index; performing sequence analysis based on the index to generate a behavior switching prediction vector; adjusting a priority according to the prediction vector to construct a storage queue; verifying a path cache of the storage queue to determine a preliminary cache requirement; reordering the data stream according to the preliminary cache requirement to generate a reading and writing path sequence; and finally completing cache mapping and fast channel allocation based on the path sequence to form a complete storage reading and writing strategy. The method can dynamically adapt to the time effectiveness difference of data.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of AI glasses, and in particular to an AI glasses data high-speed reading and writing method and system based on a storage chip. BACKGROUND

[0002] With the continuous development of human-computer interaction technology, intelligent wearable devices gradually become an important entrance connecting users and the digital world. Among them, AI glasses integrating computer vision, voice recognition and real-time feedback functions are gradually becoming key products in the intelligent terminal market. AI glasses have the advantages of convenient wearing and natural interaction, and are widely used in navigation travel, remote collaboration, auxiliary identification, entertainment and socialization and other scenes.

[0003] At present, for the data processing problem of AI glasses, a common technical path is to use a "fixed priority + periodic update" reading and writing scheduling mechanism. In this mechanism, the initial priority of various data is set, for example, video images are set as high priority, positioning information is set as medium priority, and ordinary message notification is set as low priority. However, due to the static priority setting, there are problems such as uneven allocation of buffer resources, channel blocking and access conflict in a high-concurrency environment, which is difficult to meet the processing efficiency requirements in complex data scenarios.

[0004] In summary, due to the static priority setting, the prior art cannot dynamically adapt to the data timeliness difference. SUMMARY

[0005] The application provides an AI glasses data high-speed reading and writing method and system based on a storage chip, to dynamically adapt to the data timeliness difference.

[0006] In a first aspect, to solve the above technical problems, the application provides an AI glasses data high-speed reading and writing method based on a storage chip, comprising:

[0007] obtaining a data stream of an AI glasses perception unit;

[0008] performing preliminary classification according to the data stream to obtain a preliminary classification label;

[0009] performing access frequency extraction and trend judgment according to the preliminary classification label to obtain a dynamic change index;

[0010] performing sequence analysis according to the dynamic change index to obtain a behavior switching prediction vector;

[0011] performing priority adjustment calculation according to the behavior switching prediction vector to obtain a storage queue;

[0012] performing path cache verification processing according to the storage queue to determine the preliminary cache requirement of the path sequence;

[0013] According to the preliminary cache requirement, data stream reordering processing is performed to obtain a read-write path sequence;

[0014] According to the read-write path sequence, cache mapping and fast channel allocation are performed to obtain a storage read-write strategy.

[0015] Preferably, the preliminary classification according to the data stream obtains a preliminary classification label, including:

[0016] The data stream includes an image frame sequence and a user interaction instruction;

[0017] According to the image frame sequence, inter-frame pixel difference calculation and K-means clustering are performed to obtain a dynamic degree classification result of the image;

[0018] According to the user interaction instruction, keyword matching is performed to obtain an instruction timeliness sensitivity level;

[0019] According to the dynamic degree classification result and the instruction timeliness sensitivity level, a priority label fusion is performed to obtain a preliminary classification label.

[0020] Preferably, the access frequency extraction and trend judgment according to the preliminary classification label obtain a dynamic change indicator, including:

[0021] The access frequency of the preliminary classification label corresponding period is extracted to form an access frequency data set;

[0022] According to the access frequency data set and the preliminary classification label, a trend calculation is performed to obtain a behavior switching trend value;

[0023] The behavior switching trend value is compared with a preset user behavior pattern library for similarity comparison and indicator construction to obtain a dynamic change indicator.

[0024] Preferably, the sequence analysis according to the dynamic change indicator obtains a behavior switching prediction vector, including:

[0025] According to the dynamic change indicator, sliding window segmentation sampling is performed to obtain sequence data;

[0026] The sequence data is subjected to time series processing to obtain a behavior switching prediction vector.

[0027] Preferably, the priority adjustment calculation according to the behavior switching prediction vector obtains a storage queue, including:

[0028] The target behavior class with the maximum probability in the behavior switching prediction vector is extracted;

[0029] Based on the target behavior class, data matching and scoring are performed to obtain a priority score;

[0030] Based on the priority score, a priority rearrangement process is performed to obtain a storage queue.

[0031] Preferably, the path cache verification process according to the storage queue is performed to determine the preliminary cache requirement of the path sequence, including:

[0032] The geographic coordinates and timestamp data are obtained, and it is judged whether the commuting scenario activation state is triggered through a preset commuting scenario rule library;

[0033] If the commuting scenario activation state is triggered, the road type and traffic flow information are obtained, and the priority mapping value is determined in combination with the behavior switching prediction vector;

[0034] Based on the priority mapping value, the storage queue is verified to generate an associated path indication;

[0035] According to the associated path indication, the cache space and time efficiency parameters are calculated to obtain the preliminary cache requirement of the path sequence.

[0036] Preferably, the data stream reordering process is performed according to the preliminary cache requirement to obtain a read-write path sequence, including:

[0037] Based on the preliminary cache requirement, a target data stream is extracted;

[0038] The time efficiency sensitivity of the target data stream is calculated to obtain a time efficiency sensitivity, and the target data stream is reordered according to the time efficiency sensitivity to obtain a reordered data stream sequence;

[0039] According to the reordered data stream sequence, path identification generation and node matching are performed to obtain a read-write path sequence.

[0040] Preferably, the cache mapping and fast channel allocation are performed according to the read-write path sequence to obtain a storage read-write strategy, including:

[0041] According to the read-write path sequence, hash mapping and concurrency identification processing are performed to obtain a distributed cache mapping table;

[0042] According to the distributed cache mapping table, dynamic remapping and buffer capacity configuration are performed to obtain a buffer allocation scheme;

[0043] According to the buffer allocation scheme, queue shunting and cache expansion decision are performed to obtain a storage read-write strategy.

[0044] In a second aspect, the present application provides an AI glasses data high-speed read-write system based on a storage chip, including:

[0045] A data acquisition module is configured to acquire a data stream of an AI glasses perception unit.

[0046] A preliminary classification module is configured to perform preliminary classification according to the data stream to obtain a preliminary classification label.

[0047] A change index module is configured to perform access frequency extraction and trend judgment according to the preliminary classification label to obtain a dynamic change index.

[0048] A prediction vector module is configured to perform sequence analysis according to the dynamic change index to obtain a behavior switching prediction vector.

[0049] A queue storage module is configured to perform priority adjustment calculation according to the behavior switching prediction vector to obtain a storage queue.

[0050] A cache requirement module is configured to perform path cache verification processing according to the storage queue to determine a preliminary cache requirement of a path sequence.

[0051] A path sequence module is configured to perform data stream reordering processing according to the preliminary cache requirement to obtain a read-write path sequence.

[0052] A read-write strategy module is configured to perform cache mapping and fast channel allocation according to the read-write path sequence to obtain a storage read-write strategy.

[0053] In a third aspect, the present application further provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the AI glasses data high-speed read-write method based on a storage chip according to any one of the above.

[0054] In a fourth aspect, the present application further provides a computer readable storage medium, comprising a stored computer program, wherein the computer program controls a device where the computer readable storage medium is located to execute the AI glasses data high-speed read-write method based on a storage chip according to any one of the above when the computer program is running.

[0055] Compared with the prior art, the present application has the following beneficial effects:

[0056] (1) The present application performs inter-frame pixel difference calculation on image frame sequences, classifies the dynamic degree of images by combining a K-means clustering method, extracts keywords and time effectiveness level based on user interaction instructions, and fuses to generate a preliminary classification label. This operation effectively improves the differentiated recognition ability of the system for multi-source data, and provides accurate basis for subsequent priority calculation.

[0057] (2) The application extracts the access frequency in the corresponding period based on the preliminary label, constructs a behavior switching trend vector in combination with a user behavior mode library, and further generates a behavior switching prediction vector through sequence analysis in a sliding window manner. This process realizes the advance judgment of user behavior dynamics, provides prior reference for storage resource scheduling, and reduces sudden read-write conflicts.

[0058] (3) The application calculates a priority mapping value in path cache verification in combination with traffic information and a prediction vector, verifies and filters the storage queue, and accurately extracts the path cache requirement. On this basis, target data flow reordering and path identification generation are performed, ensuring the rationality of read-write path response and the continuity of the channel.

[0059] (4) The application generates the binding relationship between the path and the cache pool address by using a hash mapping method, monitors the load state of each region in real time, triggers a remapping mechanism when the load of a certain region exceeds the preset load threshold, and configures the buffer capacity according to the actual write rate and data packet size. By configuring different buffer capacities under different loads, stable read-write and fast response of the system under high concurrency can be ensured. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 is a flowchart of the AI glasses data high-speed read-write method based on a storage chip provided by the first embodiment of the application;

[0061] Figure 2 is a structural diagram of the AI glasses data high-speed read-write system based on a storage chip provided by the second embodiment of the application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0063] With reference to Figure 1 , the first embodiment of the application provides an AI glasses data high-speed read-write method based on a storage chip, including the following steps:

[0064] S11, acquiring the data stream of the AI glasses perception unit;

[0065] S12, performing preliminary classification according to the data stream to obtain preliminary classification labels;

[0066] S13, performing access frequency extraction and trend judgment according to the preliminary classification labels to obtain dynamic change indicators;

[0067] S14, performing sequence analysis according to the dynamic change index to obtain a behavior switching prediction vector;

[0068] S15, performing priority adjustment calculation according to the behavior switching prediction vector to obtain a storage queue;

[0069] S16, performing path cache verification processing according to the storage queue to determine a preliminary cache requirement of a path sequence;

[0070] S17, performing data stream reordering processing according to the preliminary cache requirement to obtain a read-write path sequence;

[0071] S18, performing cache mapping and fast channel allocation according to the read-write path sequence to obtain a storage read-write strategy.

[0072] In step S11, a data stream of an AI glasses perception unit is obtained;

[0073] It is worth noting that the data stream is composed of two parts, an image frame sequence and a user interaction instruction. The image frame sequence is a picture taken by a front camera, with a frame rate of 30 frames per second and a resolution of 1280x720. The interaction instruction mainly includes voice commands and gesture instructions, which are input through a microphone array and an inertial sensor.

[0074] In step S12, a preliminary classification is performed according to the data stream to obtain a preliminary classification label, including:

[0075] The data stream includes an image frame sequence and a user interaction instruction;

[0076] According to the image frame sequence, inter-frame pixel difference calculation and K-means clustering are performed to obtain a dynamic degree classification result of the image;

[0077] According to the user interaction instruction, keyword matching is performed to obtain an instruction timeliness sensitivity level;

[0078] According to the dynamic degree classification result and the instruction timeliness sensitivity level, a priority label fusion is performed to obtain a preliminary classification label.

[0079] It is worth noting that for the image data part, the system first reads the image frame sequence frame by frame in time sequence, and performs difference operation on the pixel values between each frame and its previous frame. In the specific operation, the image frame is converted into a gray matrix, and the gray value of each pixel point ranges from 0 to 255. The system performs absolute value difference calculation on the pixel points at the same position of the two frames, and counts the number of pixels with a difference value greater than a set deviation threshold. The set deviation threshold is 20. , if the difference of a pixel in two adjacent frames is greater than 25, it is considered as a changed pixel. The number of these changed pixels is divided by the total number of pixels in the image to obtain the inter-frame change rate of the current frame. For example, if a frame has a resolution of 1280x720, a total of 921600 pixels, and there are 150000 pixels that meet the change condition, the change rate is about 0.163, i.e. 16.3%. The threshold value of the deviation is derived from a continuous image sequence of at least 500 frames. The full set of pixel differences between adjacent frames is constructed, the quantile analysis is performed on the set, and the 85% quantile point (i.e. P85) is selected as the value of the threshold value of the deviation. .

[0080] All inter-frame change rates will be collected to form a one-dimensional feature vector sequence, and then the K-means clustering method is used for classification. The value of K is set to 3, which aims to divide the image frames into three categories: static, medium dynamic, and high dynamic, corresponding to three visual states of low, moderate, and severe change rates. Here, the value of the number of clusters is 3, which is based on the statistical analysis of the inter-frame change rates of more than 1000 typical commuting and indoor and outdoor scene image sequences. In this sample set, the system calculates the inter-frame change rate of each frame and constructs a histogram of the overall distribution. According to the frequency density aggregation of the change rate distribution, the system defines the frames with a change rate lower than 0.20 as static, the frames with a change rate between 0.20 and 0.6 as medium dynamic, and the frames with a change rate higher than 0.6 as high dynamic. This division is based on the peak positions of the distribution of the three types of frames in the 95% confidence interval, which are located at 0.20, 0.40, and 0.6, respectively, with clear boundary layers. The three clusters can stably cover more than 95% of the sample frames in the image dynamic feature distribution.

[0081] The one-dimensional feature vector sequence of the inter-frame change rates of all image frames is executed by K-means clustering, and the number of clusters is set to 3. In the initial stage, the system randomly selects three center values, and then iteratively performs frame assignment and center update operations. Each frame is assigned to the nearest center according to its change rate, and the average change rate of the frames in each cluster is recalculated as the new center after classification. The iteration continues until the center change is less than 0.01 or reaches 10 times. After clustering, the center values are sorted from small to large, corresponding to static (<0.20), medium dynamic (0.20-0.6), and high dynamic (>0.6) respectively. Each frame of image is assigned a corresponding dynamic degree classification result accordingly.

[0082] Subsequently, the user interaction instruction is required to extract the voice transcription text or gesture code, and keyword matching is performed in the local instruction library. The local instruction library is constructed based on more than 100,000 real user instruction data collected in the past 6 months, and through statistical analysis of the average response time delay, trigger priority and context dependence of various instructions in the user behavior chain, 30 types of high-frequency instructions are extracted and assigned fixed time-sensitive grades, with a numerical range of 0 to 1. For example, the avoidance setting is 1.00, the emergency parking is 0.95, the immediate avoidance is 0.92, which is classified as a high time-sensitive class, the weather query is 0.40, the music playing is 0.35, and the store recommendation is 0.30, which is classified as a low time-sensitive class. If the user issues the instruction "Is there a car in front of me?", the system recognizes the keyword "car in front of me", hits the safe obstacle avoidance instruction, and the corresponding sensitivity level is 0.95. This value is then used as the instruction time-sensitive level.

[0083] Subsequently, the dynamic degree classification result of the image and the instruction time-sensitive level are correspondingly bound according to the time stamp, and the two are written as a set of structured data into the label cache area as the preliminary classification label.

[0084] In step S13, according to the preliminary classification label, the access frequency extraction and trend judgment are performed to obtain a dynamic change index, including:

[0085] The access frequency of the time period corresponding to the preliminary classification label is extracted to constitute an access frequency data set;

[0086] According to the access frequency data set and the preliminary classification label, a trend calculation is performed to obtain a behavior switching trend value;

[0087] The behavior switching trend value is compared with the preset user behavior mode library to obtain a dynamic change index.

[0088] It is worth noting that the function category involved in the preliminary classification label is compared with the time stamp of the historical user behavior log, and the number of calls of the function in the corresponding hour period in the past 7 days is extracted. For each function, the system respectively counts the number of accesses in the same hour period of each day, and calculates the 7-day average access frequency of the hour period, and then writes into the access frequency data table in turn. For example, for the navigation function appearing in the label, if the number of calls during 7:00 to 8:00 every day is 13, 12, 14, 11, 12, 13, and 14, the average access frequency of the hour period is 12.7 times per hour.

[0089] Subsequently, the access frequency data table is normalized. Taking the navigation function and the security reminder function as examples, if the average access frequency of the navigation function is 14 times per hour and the security reminder function is 9 times per hour, and the system sets the upper limit of access frequency to 20 times, then the normalized access frequency values ​​are 0.70 and 0.45 respectively.

[0090] Subsequently, trend calculation is performed. If the inter-frame change rate of the current image frame is 0.8, a weighted average is calculated using the normalized access frequency values ​​of each function item to obtain the behavior switching trend value. The specific calculation formula is as follows: in, To switch trend values ​​for behavior, The number of functional items involved in the current frame. To indicate the first The weight of each function This represents the global inter-frame change rate of the current image frame. For the first The normalized access frequency value for each function. The weight value for each function is calculated by counting the total number of accesses for all functions within the given time period, then dividing the access count for each function by the total access count to obtain its weight value. For example, the inter-frame change rate during the time period from 7 AM to 8 AM. =0.8, involving navigation ( =0.525, =0.70), Safety Reminder ( =0.35, =0.45) and taking a picture ( =0.125, =0.25) Three functions, then the behavior switching trend value is calculated as =0.8×(0.525×0.70)+(0.35×0.45)+(0.125×0.25)=0.445.

[0091] It should be noted that the design of the behavior switching trend value formula is based on the coupling effect of data dynamism and access frequency. The global inter-frame change rate reflects the intensity of environmental changes, the access frequency reflects the density of user behavior, and the multiplication operation strengthens the priority of high dynamic and high frequency data.

[0092] After obtaining the behavior switching trend value, it needs to be compared with a preset user behavior pattern library. For example, if the user behavior pattern library shows that the trend values ​​for the 7 days within this time period are 0.47, 0.45, 0.50, 0.46, 0.44, 0.48, and 0.46, then the calculated stable trend mean for this period is 0.466, and the standard deviation is 0.019. If the trend value of the current frame is 0.48, then the deviation... The ratio of deviation to standard deviation is... If the ratio is greater than or equal to 2, it is determined that the behavior deviates. Then, the trend value of the current frame, the corresponding timestamp, and the deviation judgment result of the historical mode are combined into a dynamic change index.

[0093] And the deviation of the current trend value and the historical stable trend value exceeds 2 times the standard deviation of the trend value of the period, that is, it is considered that the behavior deviates. The value is based on the normal distribution characteristics. If the mean value of the trend value sample is , and the standard deviation is , about 95.45% of the samples will fall within the interval Therefore, if the trend value deviates from the interval, it is an abnormal situation with a probability of less than 4.55%.

[0094] The user behavior mode library is constructed based on the user's continuous 7-day behavior log. The system divides the time period into 1 hour units, extracts the behavior trend value of each hour, and records it as a trend value sequence. For each hour period, the system calculates the mean value And the standard deviation Write to the mode library. And update the data in the library every 24 hours, replace it with the statistical results in the latest 7-day rolling window.

[0095] In step S14, according to the dynamic change index, sequence analysis is performed to obtain a behavior switching prediction vector, including:

[0096] According to the dynamic change index, sliding window segmentation sampling is performed to obtain sequence data;

[0097] The sequence data is subjected to time series processing to obtain a behavior switching prediction vector.

[0098] It is worth noting that the behavior time sequence is constructed according to the dynamic change index extracted in the previous step. Specifically, the user behavior data is sliced and processed in time sequence with a sliding window length of 30 seconds, and the window is pushed forward every 10 seconds. In each time window, the user's behavior frequency is counted, such as in a 30-second window, the user performs browsing goods, switching interface, clicking details, and other operations a total of 9 times, then the behavior frequency corresponding to the window is 9. Then record three core elements for each piece of retained sequence data, the timestamp of the behavior, the type code of the specific behavior, and the corresponding session identifier.

[0099] Subsequently, the pre-trained LSTM neural network is used to process the obtained sequence data in time sequence to obtain a behavior switching prediction vector. The behavior switching prediction vector is a one-dimensional vector composed of multiple probability values, which is used to represent the operations that the user may perform after the current behavior sequence and the corresponding occurrence probabilities. Each dimension corresponds to a behavior type, and the sum of all values is 1, reflecting the confidence degree of the next step behavior of the user. For example, in the shopping scenario, the system outputs the prediction vector [0.62, 0.20, 0.13, 0.05] based on the input behavior sequence, which respectively represents the probabilities of continuing browsing, adding to the shopping cart, immediate purchase, and exiting the page as 62%, 20%, 13%, and 5%.

[0100] It should be noted that the LSTM neural network adopts a typical structure, the input layer dimension is sequence length x feature number (features include behavior type, time interval, etc.), the hidden layer is 2 layers, each layer has 128 units, and the tanh activation function is used; the output layer is a softmax classifier; during training, the learning rate is optimized through grid search, and the initial value is 0.001.

[0101] The LSTM neural network training process is as follows. First, a structured training sample set is constructed. Based on the user's continuous 7-day historical behavior data, the training sample is constructed. Starting from each dynamic change index, the behavior sequence is divided into windows of 30 seconds and steps of 10 seconds in chronological order. Each time step contains three fields: behavior type number, behavior interval time (seconds), and session identification number, which are embedded and normalized after embedding and normalization, and used as LSTM input vectors. For each input sequence in the training sample, the target output is set as the target behavior that is most likely to occur within 60 seconds after the sequence, such as adding to the shopping cart, exiting the interface, etc., which is encoded as one-hot form as a supervised label. For example, if the real behavior after the sliding window is to exit the interface, the target output is [0, 0, 1, 0] (assuming there are four types of behavior).

[0102] After the training sample is prepared, the training method is batch feeding of samples (default batch size is 64), the cross-entropy loss function is used to calculate the error between the current prediction result and the true label, and the BPTT algorithm is used to perform back propagation on the LSTM network. The Adam optimizer is used to adjust the learning rate adaptively, and the default initial learning rate is 0.001. Each round of training is called an epoch, and the system defaults to 100 rounds of training, and evaluates the accuracy, loss value, etc. on the validation set after every 5 epochs.

[0103] To determine whether the network converges, the system sets the following three types of convergence conditions: first, when the accuracy on the validation set continuously improves by less than 0.1% for 5 rounds, the training is terminated in advance; second, if the validation set loss has no downward trend for 10 consecutive rounds, the optimal parameters are retained and the training is interrupted; third, the training is automatically ended after reaching the maximum epoch number.

[0104] Finally, the system retains the weight parameter set with the smallest loss as the prediction model for the inference stage.

[0105] In step S15, the behavior switching prediction vector is switched according to the priority adjustment calculation to obtain a storage queue, including:

[0106] Extracting the target behavior category with the highest probability in the behavior switching prediction vector;

[0107] Based on the target behavior category, data matching and scoring are performed to obtain a high-priority weight value;

[0108] Based on the high-priority weight value, priority rearrangement processing is performed to obtain a storage queue.

[0109] It is worth noting that if the input behavior switching prediction vector is [0.62, 0.20, 0.13, 0.05], it represents the probabilities of continuing browsing, adding to the shopping cart, immediately purchasing, and exiting the page as 62%, 20%, 13%, and 5%, respectively. Among them, 0.62 corresponds to continuing browsing, so the current target behavior is set to the user's possible continued browsing in the current store.

[0110] Next, according to the continued browsing, the data content associated with the target behavior is retrieved in the local storage, including product image data, price comparison information in the history record, frequently called product introduction or explanation documents by the user. Then, the access frequency of the user to each type of data within the past 24 hours is counted. Assuming that the user has accessed three types of data in the past day, the total access frequency of the price information of a certain brand of beverage is 12 times, the product label information is 4 times, and the navigation route data is 1 time, with a total access frequency of 17 times. The system calculates the access proportion of each type of data accordingly. And this access proportion is used as a priority score, and after sorting from high to low, it is written into the storage queue, and the sorting result determines the position of the data in the storage queue.

[0111] In step S16, according to the storage queue, path cache verification processing is performed to determine the preliminary cache requirement of the path sequence, including:

[0112] Obtaining geographic coordinates and timestamp data, and determining whether the commuting scenario activation state is triggered through a pre-set commuting scenario rule library;

[0113] If the commuting scenario activation state is triggered, the road type and traffic flow information are obtained, and the priority mapping value is determined in combination with the behavior switching prediction vector;

[0114] Based on the priority mapping value, the storage queue is verified to generate an associated path indication;

[0115] According to the associated path indication, the cache space and time parameter are calculated to obtain the preliminary cache demand of the path sequence.

[0116] It is worth noting that the collected positioning components are integrated and installed in the right temple of the AI glasses. The components are composed of a small multi-mode GNSS chip and a three-mode communication module, which can simultaneously receive GPS satellite signals, cellular base station location information, and WiFi hotspot information. The system is set to automatically collect position data every 2 seconds, with a coordinate accuracy control within 3 meters. The collected time information is stored in UTC time format with a precision of milliseconds, which is used to ensure the continuity of the tracking of the position information. Then, according to the time point and coordinate point, the local preset commuting scenario rule library is called for judgment. The rule table sets the time range of the morning and evening peak of weekdays, for example, the early peak is from 7:30 to 9:30, and defines the typical coordinate pairs of residential areas and office areas, for example, from the location of latitude 39.8976, longitude 116.3856 to latitude 39.9156, longitude 116.4123. If the moving time is within the above commuting time period, the commuting scenario activation state is triggered. The commuting scenario rule library automatically generates rules through historical trajectory clustering (such as the DBSCAN algorithm), and is updated every 24 hours.

[0117] Once the commuting scenario activation state is triggered, the user's current road type information and traffic flow information are extracted. The road types include four categories: main road, secondary road, branch road, and pedestrian road. The corresponding priority coefficients are 0.5, 0.3, 0.1, and 0.1, respectively. The priority coefficients are based on the user's historical commuting path samples, and the approximate frequency ratio of main road, secondary road, branch road, and pedestrian road is 5:3:1:1. The traffic flow information is based on historical traffic flow and user behavior data analysis. The system divides the road traffic state into three levels: smooth, slow, and congested. When the number of vehicles on this section is between 40 and 80 per minute, and the average vehicle distance is 15 to 30 meters, the system marks it as smooth, and assigns a traffic coefficient of 0.6. When the number of vehicles exceeds 80 and the vehicle distance is shortened to 5 to 15 meters, it is marked as slow, and the coefficient is set to 0.3. If the number of vehicles is less than 40 and the vehicle distance is less than 5 meters, it is determined to be congested, and the traffic coefficient is set to 0.1.

[0118] Subsequently, the fusion calculation is combined with the prediction vector to determine the priority mapping value. Assuming that the current vector is [0.68, 0.17, 0.10, 0.05], which respectively corresponds to the four types of behaviors of continuing browsing, path selection, positioning viewing and exiting navigation, the current maximum value is 0.68, and the system sets the target behavior as continuing browsing navigation path accordingly. Assuming that the current user is located on the main road, the priority coefficient is (0.5), the traffic state of the road section is slow, the traffic coefficient is (0.3), the behavior switching prediction vector is [0.62, 0.20, 0.13, 0.05], and the maximum value is 0.62 (corresponding to continuing commuting browsing). Then the priority mapping value is equal to the weighted average of the road coefficient (0.5), the traffic coefficient (0.3), the behavior prediction value (0.62) and the corresponding weight, wherein the weight coefficient can be set according to the historical data (such as the weight of the road coefficient is 0.4, the coefficients of the traffic coefficient and the behavior prediction value are 0.3 and 0.3 respectively, and the total sum is ≤1), and the exemplary 0.4x0.5+0.3x0.3+0.3x0.62=0.476; The priority mapping value is normalized to the interval [0, 1] when calculating each coefficient (road priority coefficient, traffic coefficient, behavior prediction value), ensuring the balance of the values.

[0119] If the calculated priority mapping value is greater than the set threshold value 0.4, the storage queue is verified to generate an associated path indication. Specifically, all navigation data tags in the current storage queue are traversed, and for each navigation data, its target position point and channel identification information are extracted and matched with a preset keyword table, and the keywords include highway entrance, subway station and BRT channel. If the matching is successful, it is marked as an associated path indication. And the associated path indication includes path starting point coordinates, end point coordinates and expected travel time.

[0120] And the path starting point coordinates and end point coordinates are obtained by querying In the mapping record acquisition in the geographic information database, the system acquires the longitude and latitude coordinate points corresponding to the matched unique coding index. The estimated travel time is obtained by dividing the distance of the road section by the average travel speed. The average travel speed of the main road is 40 km / h in the smooth state, and the average travel speed of the branch road is 15 km / h in the congested state. If there is a commuting path composed of multiple road sections, the system will calculate the time of each section and then sum up the total estimated travel time. For example, a commuting path is composed of three continuous road sections, which are the main road, the secondary road, and the branch road. The first section is the main road with a length of 1.5 km and a current state of smooth. The system calculates the travel time as 1.5 divided by 40 multiplied by 60, which is about 2.25 minutes. The second section is the secondary road with a length of 0.8 km and a current state of slow. The average speed corresponding to the slow state is 25 km / h, and the calculated travel time is about 1.92 minutes. The third section is the branch road with a length of 0.6 km and a current state of congestion. The average speed corresponding to the congested state is 15 km / h, and the calculated travel time is 2.4 minutes. The system adds up the travel times of the three sections one by one to obtain the total estimated travel time of the entire commuting path, which is 6.57 minutes.

[0121] Subsequently, the cache space and the time-to-live parameter are calculated. Specifically, the data volume estimation is calculated based on the length of the path section and the types of information contained. For example, each 200-meter path section contains the starting coordinate, the ending coordinate, one turning instruction, road type description, and real-time traffic information, which requires about 64 KB of storage space. If a commuting path is 2.5 km long, it can be divided into 13 sections, each of which occupies 64 KB, and the total data volume is about 832 KB. If the path contains image maps, voice instructions, and other multimedia content, the expanded data volume of each section can reach 128 KB, and the total volume can exceed 1.5 MB.

[0122] Secondly, the access frequency statistics are calculated based on the user's operation records on the path-related data, such as navigation updates, route adjustments, and voice broadcasts. The cumulative number of calls per unit time is calculated. When a user performs more than 50 data call operations on a certain route within 5 minutes, the average access frequency is 10 times per minute.

[0123] The system sets the allocation standard for the cache space as follows: if the data volume exceeds 2 MB or the access frequency exceeds 10 times per minute, the system allocates at least 128 MB of initial cache space for the path to ensure that the data can be quickly called and not frequently cleaned; if the data volume is below 1 MB and the access frequency is low, 32 MB is allocated as the default cache space.

[0124] As for the time parameter, the system sets two key values, data validity period and refresh cycle, according to the data type and path characteristics. For example, for a high-frequency commuting path with an average access frequency of more than 10 times, the default data validity period is 30 minutes, meaning that the cached content is valid within this time period; if the user is still in the commuting state, the path will be automatically updated every 5 minutes. The average access frequency of more than 10 times / minute is set as the threshold for determining a high-frequency commuting path. This threshold is collected from the commuting navigation data of 2000 AI glasses users within a week. The access frequency of users in the commuting scenario approximately follows a normal distribution, with a mean of about 6 times / minute and a standard deviation of about 2 times / minute. According to the statistical law of normal distribution, about 95% of the access behavior is distributed between 2 and 10 times / minute, and the access frequency of more than 10 times / minute is outside the 2 standard deviations, belonging to the high-density operation interval in a statistical sense.

[0125] Finally, the system writes the cache space size, data survival time, update frequency, and coordinate information of each path into the preliminary cache demand. If the preliminary cache demand cannot be met (e.g., insufficient space), the priority recalculation is triggered, and the storage queue is adjusted in step S15. At the same time, a timeout mechanism is set to prevent a dead loop.

[0126] In step S17, data stream reordering processing is performed according to the preliminary cache demand to obtain a read-write path sequence, including:

[0127] Based on the preliminary cache demand, the target data stream is extracted;

[0128] The time-sensitive sensitivity of the target data stream is calculated, and the target data stream is reordered according to the time-sensitive sensitivity to obtain a sorted data stream sequence;

[0129] According to the sorted data stream sequence, path identification generation and node matching are performed to obtain a read-write path sequence.

[0130] It is worth noting that by comparing the path identification code with the coordinate information in the cache demand, the navigation, traffic, voice broadcast, and other related data items corresponding to the path in the storage queue are filtered out; combined with the access frequency field, they are extracted as the target data stream for this round of reordering processing.

[0131] Then, the time-sensitive sensitivity of each data stream is calculated to determine the real-time value and reading priority of its content. Specifically, the system first obtains the generation timestamp of each data and performs difference calculation with the current system time to obtain the time delay of the data , which is in minutes. Then, according to the type to which the data belongs, a basic sensitivity weight coefficient , the coefficient reflects the importance of the data at the time of generation. Set as follows, navigation path data is 0.9, traffic event data is 0.8, voice broadcast data is 0.6, and general prompt information is 0.5.

[0132] To simulate the decrease in importance of data over time, the system sets a maximum valid time threshold , indicating how long the data of this type will be considered sensitive. For example, navigation data is set = 60 minutes, traffic event data is set = 15 minutes, voice broadcast data is set = 5 minutes, and general prompt information is set = 2 minutes. Finally, the time-sensitive sensitivity can be calculated according to the following linear decay formula, . Assuming that a traffic event data was generated 15 minutes ago, with a base weight of 0.9 and a maximum valid time of 60 minutes, ; the time-sensitive sensitivity model can choose linear or exponential decay according to the data type, and the default linear model is used for easy calculation, and those skilled in the art can replace it with a more complex model.

[0133] The base weight values of navigation path data, traffic event data, voice broadcast data, and general prompt information (0.9, 0.8, 0.6, and 0.5, respectively) are set based on long-term user behavior data statistics, through probability distribution analysis and normal distribution confidence interval method. First, the system collects the call frequency, trigger time, and user interaction response of different types of data based on the commuting use records of more than 2000 AI glasses users in a continuous week. Taking navigation path data as an example, the mean of its call frequency is about 12 times per minute, and the standard deviation is about 3 times. More than 95% of the navigation data calls fall within the range of 6 to 18 times per minute, indicating that it is the type of information with the highest user interaction density. According to the upper quantile calculation of normal distribution, its sensitivity is significantly higher than that of other categories, so the weight is set to 0.9. Traffic event data is triggered on average every 10-15 minutes, with a concentrated fluctuation range and a confidence interval covering a frequency of more than 80%, so it is set to the second highest weight of 0.8. Voice broadcast data has a certain timeliness, but has high delay tolerance and moderate information intensity, so it is assigned a value of 0.6. The average call frequency of general prompt information is less than 2 times per minute, which is non-core information, and the weight is the lowest, which is 0.5. After the calculation is completed, the system sorts all target data streams in order of sensitivity value from high to low to obtain the sorted data stream sequence.

[0134] On the basis of the sorting, the system reads the data at the head of the queue, and judges whether the sensitivity exceeds the preset sensitivity threshold. The sensitivity threshold is set to 0.75. For example, if the sensitivity of the data at the head is 0.91, the system confirms that it is a data item that needs to be read immediately.

[0135] At this time, the system generates the read-write path identifier corresponding to the data. The identifier content adopts hexadecimal encoding, which is composed of data type, emergency level and belonging storage partition number. Taking A7F2E8 as an example, A7 indicates that the data is accident information, F2 indicates that its emergency level is level 2, and E8 corresponds to the real-time storage node number. Subsequently, the system performs node matching according to the path identifier. Specifically, the associated storage node address and access frequency record are found in the preset path allocation table.

[0136] The sensitivity threshold setting is based on the statistical time-sensitive value of more than 100,000 data in the last thirty days. It is found that the overall distribution is approximately normal, with a mean of 0.65 and a standard deviation of 0.05. Therefore, by using the mean plus twice the standard deviation method, 0.75 is set as the sensitivity threshold.

[0137] The preset path allocation table is automatically generated by the system during long-term operation. The system continuously records the access behavior of users to various data (including navigation path, real-time traffic, voice broadcast, etc.) in the background, and writes the corresponding path identifier, storage node address and access frequency of each access into the log database. Subsequently, the system archives and arranges the data with cumulative access times exceeding 100 times and still having call records in the last 7 days, and generates formal path allocation table items. The table is sorted by access frequency from high to low, and is automatically updated at the end of each day, eliminating long-time non-accessed items and supplementing new high-frequency path records.

[0138] In step S18, according to the read-write path sequence, cache mapping and fast channel allocation are performed to obtain a storage read-write strategy, including:

[0139] According to the read-write path sequence, hash mapping and concurrent identification processing are performed to obtain a distributed cache mapping table;

[0140] According to the distributed cache mapping table, dynamic remapping and buffer capacity configuration are performed to obtain a buffer allocation scheme;

[0141] According to the buffer allocation scheme, queue shunting and cache expansion decision are performed to obtain a storage read-write strategy.

[0142] It is worth mentioning that first, based on the read-write path sequence generated in the previous step, the identification code, access frequency and corresponding storage location field of each path are extracted in turn. Then each path identification is regarded as a string composed of six hexadecimal numbers, for example A7F2E8. First, the system disassembles the string by bit and converts it into decimal numbers, A7 is 167, F2 is 242, and E8 is 232. Next, add these three values in turn and multiply by a preset disturbance coefficient (such as 37, the disturbance coefficient is a prime number to enhance the uniformity of hash distribution), which is used to enhance the discreteness of address distribution. The specific calculation is (167+242+232)×37=23717. Then, the system takes the result modulo the number of cache pools, assuming that the cache pool is divided into 8 regions, then 23717 mod 8 = 5, finally the path should be mapped to the cache pool number 5 region. Then all path identifications are mapped in parallel. At the same time, in order to avoid hash conflicts (i.e. different paths are mapped to the same cache pool address), the system uses the open address method to handle conflicts, that is, if a target region is full, then sequentially find the next free region as a replacement, and update its final address in the hash table; the hash conflict processing adopts the combination of open address method and chain method, when all cache regions are full, trigger the emergency expansion mechanism to allocate new regions from the standby pool. The mapping structure is saved in the form of key-value pair, the key is the path identification (such as "A7F2E8"), the value is the corresponding cache pool number (such as "5"), and the access frequency field is also attached.

[0143] When the system detects that multiple paths are frequently accessed at the same time, and their access frequency exceeds the concurrency threshold (for example, 20 times per minute), it is marked as concurrent path, and independent address index entries are established for these paths in the mapping table. At this time, the mapping table contains multiple cache region entries, forming a distributed cache mapping table. The setting of the concurrency threshold is based on the statistical analysis results of the long-term running data of the system. The system collected about 2000 navigation path access records in the test stage, and conducted distribution statistics on the access frequency of each path in unit time. The results show that the access frequency of most paths is concentrated between 5 to 15 times per minute, and the distribution form approximately conforms to the normal distribution, with a mean of about 9.8 times per minute and a standard deviation of about 4.7. According to the statistical law, the access frequency exceeding twice the standard deviation of the mean belongs to the high access interval, that is, 9.8+2×4.7=19.2, so the system sets 20 times per minute as the concurrency threshold.

[0144] After obtaining the distributed cache mapping table, the running load state of each cache pool area is continuously monitored. Mainly including the number of write request times per unit time; secondly, the proportion of the used storage capacity of the current area; and finally, the number of data packets queuing to enter the area. Once it is detected that the load rate of a certain cache area reaches or exceeds the set load threshold of 80%, the dynamic remapping of the area is performed.

[0145] The dynamic remapping specifically traverses the real-time load state of all cache pool areas, compares the occupancy rate, and selects the area with the smallest load rate. Then, the path identifier corresponding to the current data item to be written is removed from the original mapping relationship, and new address binding information is written in the hash mapping table to ensure that subsequent read and write operations of the data can correctly access the newly allocated area. For example, if path A7F2E8 is originally mapped to cache area 2, but the current load of the area is 85%, while the load of area 6 is only 45%, the system will rewrite the path to area 6 to avoid conflicts. At the same time, the buffer capacity of each cache area needs to be configured or adjusted. The buffer capacity here refers to the memory space allocated to each cache pool area for temporarily storing data packets, and the capacity size directly affects the data throughput capacity and the buffer capacity of the area to deal with burst requests. The basis for configuring the buffer capacity is the average data packet size of each path (such as 200KB per packet), the maximum number of requests per unit time (such as 120 packets per second), and the average residence time of data in the area (such as 10 seconds to complete the storage). For example, if a path is expected to write 100 200KB data packets per second, and requires to complete the storage and empty within 10 seconds, the area needs to be configured with at least 100 packets / second x 200KB / packet x 10 seconds = 200000KB ≈ 195.3MB of buffer capacity. In a typical case, if the system writes 120 data packets per second, each packet is about 200KB, the system needs to reserve at least 120 packets / second x 200KB / packet x 1 second = 24000KB ≈ 24MB of cache capacity. The system will reserve an additional 10%-20% of redundant space based on the above calculation value and the historical congestion risk level of the current area to cope with traffic fluctuations.

[0146] Subsequently, queue shunting processing is performed according to the traffic condition of the data read-write request. Specifically, the system first counts the total number of current write requests, and establishes a to-be-processed queue in chronological order of data generation. The shunting standard here is based on the time of data arrival and the request frequency. The system sets a maximum request number threshold for each sub-queue, with 500 requests as the upper limit. If the number of requests in a sub-queue exceeds this value, the new data is automatically allocated to the next idle sub-queue. 500 requests as the maximum request number threshold for the sub-queue is a statistical result obtained by analyzing the actual sampling of AI glasses navigation data during peak hours. The request frequency approximately follows a normal distribution, with a mean of about 320 requests per minute and a standard deviation of about 90. According to experience, the mean plus twice the standard deviation, i.e. 320 + 2 x 90 = 500, is set as the safety upper limit under high concurrency.

[0147] After shunting is completed, the system continuously monitors the space usage of each buffer. If the usage rate of any buffer approaches saturation, for example, reaches 90% or above, the system considers it as triggering cache expansion. At this time, the system allocates an idle area from the standby cache pool, the size of which is estimated according to the average data packet size and the current write rate. Taking a typical case as an example, if the system writes 120 data packets per second, each packet being about 200KB, the system reserves at least 24MB of cache capacity to support the buffering needs for about 1 second.

[0148] The system sets the buffer usage rate reaching 90% as the triggering threshold for cache expansion, which is derived from the statistical analysis of the cache logs of the navigation system running continuously for 7 days. The real-time usage rate of each cache area is sampled to form a sample set, and a normal distribution model is fitted. Experimental data shows that the mean of the buffer usage rate is concentrated around 65%, with a standard deviation of about 12%. According to the distribution law of normal distribution, about 99.7% of the data usage rate is concentrated within three standard deviations of the mean, i.e. not more than 101%. Therefore, the system sets 90% as the warning limit from the saturation state.

[0149] The cache expansion process includes three operations, one is to demarcate a new cache address and register it in the cache management table; two is to transfer part of the to-be-written data from the original buffer to the new expanded area, which follows the first-in-first-out principle to maintain data sequence consistency; three is to update the path mapping relationship so that subsequent write operations can correctly correspond to the new address. The expansion process is automatically performed in the background without interrupting the front-end navigation function.

[0150] After the expansion is completed, the system records the overall read-write strategy to generate a triple consisting of path address, cache area and corresponding shunting channel, as the final output of the storage read-write strategy.

[0151] Finally, through the two steps of queue shunting and cache expansion, the system realizes stable bearing of sudden large-scale navigation data write requests, and realizes dynamic adaptation to data timeliness differences, thereby improving the response capability of the navigation device during the rush hour period, and avoiding data delay or loss caused by cache congestion, thereby forming a stable and continuous storage read-write strategy.

[0152] It should be noted that in steps S12, S16 and S18, all thresholds and parameters (including the deviation threshold Δ=25, the priority mapping value threshold 0.4, the cache space allocation threshold, the load threshold 80%, and the expansion threshold 90%) are set based on large-scale user data statistical analysis; Specifically, the actual use data of no less than 5000 AI glasses users for 30 consecutive days is collected, covering multiple scenes such as commuting, indoor, outdoor, etc., and the parameter value is determined through normal distribution test and confidence interval calculation (confidence level 95%); For example, the inter-frame pixel difference threshold Δ=25 is derived from the quantile analysis of the pixel difference value of 100,000 image sequences, taking the 85% quantile point; The cache expansion threshold 90% is based on the normal distribution fitting of the buffer usage rate data (mean 65%, standard deviation 12%), which ensures that 99.7% of the use cases are covered; Those skilled in the art can adjust the parameters according to the actual data distribution.

[0153] Referring to Figure 2 The second embodiment of the application provides an AI glasses data high-speed read-write system based on a storage chip, comprising:

[0154] A data acquisition module for acquiring a data stream of an AI glasses perception unit;

[0155] A preliminary classification module for performing preliminary classification according to the data stream to obtain a preliminary classification label;

[0156] A change index module for performing access frequency extraction and trend judgment according to the preliminary classification label to obtain a dynamic change index;

[0157] A prediction vector module for performing sequence analysis according to the dynamic change index to obtain a behavior switching prediction vector;

[0158] A queue storage module for performing priority adjustment calculation according to the behavior switching prediction vector to obtain a storage queue;

[0159] A cache demand module for performing path cache verification processing according to the storage queue to determine the preliminary cache demand of the path sequence;

[0160] A path sequence module for performing data stream reordering processing according to the preliminary cache demand to obtain a read-write path sequence;

[0161] The read-write strategy module is configured to perform cache mapping and fast channel allocation according to the read-write path sequence, and obtain a storage read-write strategy.

[0162] It should be noted that the AI glasses data high-speed read-write system based on a storage chip provided by the embodiment of the present application is used to execute all flow steps of the AI glasses data high-speed read-write method based on a storage chip provided by the above embodiment, and the working principles and beneficial effects of the two are one-to-one correspondence, thus not being repeated.

[0163] The embodiment of the present application further provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an AI glasses data high-speed read-write program based on a storage chip. The processor executes the computer program to implement the steps in each of the above AI glasses data high-speed read-write methods based on a storage chip, such as the step S11 shown in the above embodiment. Figure 1 The processor executes the computer program to implement the functions of each module / unit in each of the above system embodiments, such as the prediction vector module.

[0164] For example, the computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present application. The one or more modules / units can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program in the electronic device.

[0165] The electronic device can be a desktop computer, a notebook, a palm computer, and a smart tablet, etc. The electronic device can include, but is not limited to, a processor, a memory. Those skilled in the art can understand that the above components are only examples of the electronic device, and do not constitute a limitation on the electronic device, and can include more or less components than the above, or combine certain components, or different components, such as the electronic device can also include an input / output device, a network access device, a bus, etc.

[0166] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The processor is a control center of the electronic device, and connects various parts of the electronic device through various interfaces and lines.

[0167] The memory can be used to store the computer program and / or modules, and the processor realizes various functions of the electronic device by running or executing the computer program and / or modules stored in the memory, and calling data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playing function, an image playing function, etc.), etc.; and the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a nonvolatile memory, for example, a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state memory device.

[0168] The modules / units integrated in the electronic device, if realized in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. The computer program can implement the steps of each method embodiment when executed by a processor. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate forms, etc. The computer readable medium can include any entity or system, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium, etc. that can carry the computer program code. It should be noted that the contents included in the computer readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer readable medium does not include electric carrier signals and telecommunication signals.

[0169] It should be noted that the above-described system embodiments are only illustrative, and the units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. In addition, the connection relationship between the modules in the system embodiment provided by the present application indicates that there is a communication connection between them, which can be realized as one or more communication buses or signal lines. Those skilled in the art can understand and implement it without creative labor.

[0170] The above-described specific embodiments further detail the purpose, technical solutions, and beneficial effects of the present application. It should be understood that the above-described specific embodiments are only examples of the present application and are not intended to limit the scope of protection of the present application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A high-speed data read / write method for AI glasses based on a storage chip, characterized in that, include: Acquire the data stream from the sensing unit of the AI ​​glasses; Based on the data stream, preliminary classification is performed to obtain preliminary hierarchical labels; Based on the preliminary classification labels, access frequency is extracted and trend is judged to obtain dynamic change indicators; Based on the dynamic change indicators, sequence analysis is performed to obtain the behavior switching prediction vector; Based on the behavior switching prediction vector, priority adjustment calculations are performed to obtain the storage queue; Based on the storage queue, perform path cache verification processing to determine the initial cache requirements of the path sequence; Based on the initial caching requirements, data stream reordering is performed to obtain a read / write path sequence; Based on the read / write path sequence, cache mapping and fast channel allocation are performed to obtain the storage read / write strategy; The step of extracting access frequency and judging trends based on the preliminary classification tags to obtain dynamic change indicators includes: Extract the access frequency of the time period corresponding to the preliminary classification label to form an access frequency dataset; Based on the access frequency dataset and the preliminary classification labels, trend calculation is performed to obtain behavior switching trend values; The behavior switching trend value is compared with a preset user behavior pattern library for similarity and index construction to obtain dynamic change index; The step of performing path cache verification processing based on the storage queue to determine the initial cache requirements of the path sequence includes: Obtain geographic coordinates and timestamp data, and determine whether to trigger the commuting scenario activation state through a preset commuting scenario rule base; If the commuting scenario is activated, road type and traffic flow information are obtained, and the priority mapping value is determined by combining the behavior switching prediction vector. The storage queue is verified based on the priority mapping value, and an associated path indication is generated. Based on the associated path indication, the cache space and time parameters are calculated to obtain the preliminary cache requirements of the path sequence.

2. The high-speed data read / write method for AI glasses based on a storage chip according to claim 1, characterized in that, The preliminary classification based on the data stream to obtain preliminary hierarchical labels includes: The data stream includes: image frame sequences and user interaction commands; Based on the image frame sequence, inter-frame pixel difference calculation and K-means clustering are performed to obtain the dynamic range classification result of the image; Based on the user interaction instructions, keyword matching is performed to obtain the instruction timeliness sensitivity level; Based on the dynamic level classification results and the instruction timeliness sensitivity level, priority labels are fused to obtain preliminary hierarchical labels.

3. The high-speed data read / write method for AI glasses based on a storage chip according to claim 1, characterized in that, The step of performing sequence analysis based on the dynamic change index to obtain the behavior switching prediction vector includes: Based on the dynamic change index, a sliding window segmented sampling is performed to obtain sequence data; The sequence data is processed temporally to obtain the behavior switching prediction vector.

4. The high-speed data read / write method for AI glasses based on a storage chip according to claim 1, characterized in that, The step of performing priority adjustment calculations based on the behavior switching prediction vector to obtain the storage queue includes: Extract the target behavior category with the highest probability from the behavior switching prediction vector; Based on the target behavior category, data matching and scoring are performed to obtain a priority score; Based on the priority score, priority reordering is performed to obtain the storage queue.

5. The high-speed data read / write method for AI glasses based on a storage chip according to claim 1, characterized in that, The step of reordering the data stream based on the initial caching requirements to obtain the read / write path sequence includes: Based on the initial caching requirements, extract the target data stream; The time sensitivity of the target data stream is calculated to obtain the time sensitivity, and the target data stream is reordered according to the time sensitivity to obtain a sorted data stream sequence. Based on the sorted data stream sequence, path identifiers are generated and nodes are matched to obtain a read / write path sequence.

6. The high-speed data read / write method for AI glasses based on a storage chip according to claim 1, characterized in that, The step of performing cache mapping and fast channel allocation based on the read / write path sequence to obtain the storage read / write strategy includes: Based on the read / write path sequence, perform hash mapping and concurrency identification processing to obtain a distributed cache mapping table; Based on the distributed cache mapping table, dynamic remapping and buffer capacity configuration are performed to obtain a buffer allocation scheme; Based on the buffer allocation scheme, queue splitting and cache expansion decisions are made to obtain the storage read / write strategy.

7. A high-speed data read / write system for AI glasses based on a storage chip, characterized in that, A method for implementing high-speed data read / write of AI glasses based on a memory chip as described in any one of claims 1 to 6, comprising: The data acquisition module is used to acquire the data stream from the AI ​​glasses' sensing unit; The preliminary classification module is used to perform preliminary classification based on the data stream to obtain preliminary classification labels; The change index module is used to extract access frequency and determine trends based on the preliminary classification labels to obtain dynamic change indicators; The prediction vector module is used to perform sequence analysis based on the dynamic change index to obtain the behavior switching prediction vector; The queue storage module is used to perform priority adjustment calculations based on the behavior switching prediction vector to obtain a storage queue; The cache requirement module is used to perform path cache verification processing based on the storage queue to determine the initial cache requirements of the path sequence. The path sequence module is used to reorder the data stream according to the initial caching requirements to obtain the read / write path sequence; The read / write strategy module is used to perform cache mapping and fast channel allocation based on the read / write path sequence to obtain the storage read / write strategy.

Images