An end-cloud collaborative edge inference synchronization method

Abstract

The application discloses an end-cloud cooperative edge inference synchronization method, belongs to the field of edge computing and end-cloud cooperative inference, and aims at solving the problems of inference result lag, model version inconsistency, key event missing transmission and repeated transmission bandwidth occupation of an edge device in network jitter, bandwidth limitation, link interruption or multi-device concurrent access. The application realizes the technical effects of real-time response to on-site events and consistent traceability of cloud edge states through local inference event encapsulation, priority-based synchronization scheduling, semantic fingerprint deduplication and version vector confirmation replay.

Description

Technical Field

[0001] This invention relates to the field of edge computing and edge-cloud collaborative reasoning, and in particular to an edge-cloud collaborative edge reasoning synchronization method. Background Technology

[0002] With the development of edge computing scenarios such as vehicle networking, industrial control, and smart gateways, edge devices are increasingly undertaking real-time inference tasks for field data and need to synchronize model versions, inference results, and device status to the cloud or edge cloud so that the cloud can maintain state copies, review inference results, and distribute model patches.

[0003] Existing synchronization methods typically employ fixed-period uploads, batch aggregation uploads, or centralized cloud processing. Under conditions of network jitter, bandwidth constraints, link interruptions, or concurrent access by multiple devices, these methods can easily lead to critical inference events queuing, ordinary repetitive events continuously consuming link resources, and cloud-based confirmation status not being promptly relayed to edge devices.

[0004] Meanwhile, there may be differences between the local model version of the edge device and the confirmed version in the cloud. If there is a lack of a collaborative mechanism for version vectors, event content summaries and log confirmation levels, it will be difficult to merge the cloud device status copies in a timely manner, and there may be omissions or duplicate retransmissions after the connection is restored.

[0005] Therefore, a method for edge-cloud collaborative inference synchronization is needed to address the shortcomings of existing technologies. Summary of the Invention

[0006] One objective of this invention is to propose an edge-to-cloud collaborative inference synchronization method. Addressing the problems of inference result lag, model version inconsistency, missed transmission of critical events, and duplicate transmission in existing technologies under unstable links and multi-device concurrency conditions, this invention proposes a technical solution that encapsulates local inference results into inference synchronization events and determines the synchronization strategy based on event severity, information freshness, uncertainty, version differences, link quality, and cloud confirmation status. This invention achieves the technical effects of prioritizing the transmission of critical events, suppressing ordinary duplicate events, supplementing transmission according to confirmation gaps after connection loss, and maintaining consistent cloud-edge status.

[0007] This invention provides an edge-cloud collaborative inference synchronization method, comprising:

[0008] S1. Edge devices collect on-site data and call local inference models to generate inference results, confidence values, uncertainty scores, event types, event severity, and event timestamps;

[0009] S2. Encapsulate the inference result, confidence value, uncertainty score, event type, event severity, model version identifier, input summary, content summary, event timestamp, device state vector, and link state into an inference synchronization event. The content summary is generated from the inference result and the device state vector.

[0010] S3. The synchronization module calculates the synchronization priority and determines the synchronization strategy based on the event severity, information freshness, uncertainty score, model version difference, link quality index and cloud confirmation status determined by the saved cloud confirmation vector.

[0011] S4. Upload inference synchronization events that are uploaded immediately or in summary, and merge, compress and deduplicate inference synchronization events that are delayed in merging or retained locally in the local event log.

[0012] S5. The cloud updates the device status copy according to the model version vector, the content summary and the status merging rules, returns the cloud confirmation vector, and returns the cloud review result or model patch when the conditions are met, so that the edge device updates the saved cloud confirmation vector and the model version vector.

[0013] S6. The edge device generates a retransmission list based on the local event log, model version vector, and saved cloud confirmation vector, and replays the retransmission list according to the synchronization priority when the link is restored or a confirmation gap occurs.

[0014] Optionally, S1 includes:

[0015] The field data is organized into input data frames according to the collection time window, and the input data frames are input into the local inference model to obtain the category probability distribution and regression output;

[0016] The probability of the first value in the category probability distribution is taken as the confidence value;

[0017] The uncertainty score is generated based on the degree of dispersion of the category probability distribution, the confidence value, and the deviation of the input data frame from the historical input summary set.

[0018] The event type and the event severity are determined based on the inference results and the preset event mapping table.

[0019] Optionally, S2 includes:

[0020] The field data is organized into input data frames according to the acquisition time window, and a summary operation is performed on the sampling window identifier, feature field identifier, and preprocessing version identifier of the input data frames to obtain the input summary;

[0021] Perform a summary operation on the category identifier, regression output, event type, event severity, and fields in the device state vector in the inference result to obtain the content summary;

[0022] The link state is composed of available bandwidth, round-trip time, packet loss rate, and connection status;

[0023] The edge device assigns a device identifier and a log sequence number, and writes the device identifier, the log sequence number, the event identifier, and the information described in step S2 into the inference synchronization event.

[0024] Optionally, S3 includes:

[0025] The information freshness is obtained by the time difference between the event timestamp and the synchronization scheduling time through a preset monotonically decreasing mapping table;

[0026] The model version difference is obtained by the version sequence difference between the model version identifier and the confirmed model version identifier of the corresponding device in the saved cloud confirmation vector;

[0027] The link status includes available bandwidth, round-trip time, packet loss rate, and connection status. The link quality index is obtained by normalizing and fusing the available bandwidth, round-trip time, packet loss rate, and connection status.

[0028] The cloud confirmation status includes unconfirmed, confirmed, and conflicting confirmation, and is determined by the confirmed log sequence number, confirmed model version identifier, and conflict flag in the saved cloud confirmation vector;

[0029] Furthermore, a current semantic fingerprint is generated based on the content summary, event type, event severity, and device state vector of the current inference synchronization event. The current semantic fingerprint is then compared with the historical semantic fingerprints recorded in the local event log to obtain the semantic fingerprint similarity.

[0030] The first hit is output in the following order: disconnection judgment, critical event judgment, link transmittable judgment, duplicate event judgment, and default judgment.

[0031] When the link status is disconnected and the event severity is less than the second severity threshold, a local retention policy is output.

[0032] When the link status is disconnected and the event severity is not less than the second severity threshold, output the local retention policy and write the retransmission priority flag;

[0033] When the link status is not disconnected, and the event severity is not less than the first severity threshold, or the uncertainty score is not less than the uncertainty threshold, or the model version difference is not less than the version difference threshold, the immediate upload strategy is output.

[0034] When the aforementioned strategy is not hit, and the link status is not disconnected, the link quality index is less than the link transmittable threshold, and the event severity is less than the first severity threshold, the summary upload strategy is output.

[0035] When the aforementioned strategy is not hit, and the semantic fingerprint similarity is not less than the merging similarity threshold, and the historical event corresponding to the historical semantic fingerprint similar to the current semantic fingerprint is within the confirmation retention period, a delayed merging strategy is output.

[0036] If the aforementioned strategy is not hit and the link status is not disconnected, output the digest upload strategy.

[0037] Optionally, S4 includes:

[0038] The semantic fingerprint deduplication structure performs content hashing on the content summary, the event type, the event severity, and the device state vector to obtain a content hash value, and performs vectorized summarization on the device state vector in the state window to obtain a semantic summary;

[0039] The status window consists of the device status vector, event timestamp, and log sequence number within a preset collection time window under the same device identifier;

[0040] A semantic fingerprint is generated based on the content hash value and the semantic digest;

[0041] The retention period is defined as the log range in which the log sequence number in the local event log is not greater than the confirmed log sequence number in the saved cloud confirmation vector and does not exceed the preset retention length.

[0042] When the similarity between the semantic fingerprint and the historical semantic fingerprint within the confirmation retention period is not less than the merged similarity threshold, the repeated uploading of the corresponding reasoning synchronization event is suppressed.

[0043] Furthermore, the local event log is an append log written in ascending order of log sequence number. The append log records device identifier, log sequence number, event identifier, model version identifier, content summary, content hash value, semantic fingerprint, event type, event severity, event timestamp, synchronization priority, synchronization strategy, upload status, and confirmation status.

[0044] A semantic fingerprint cluster is a set of inference synchronization events whose similarity to the current semantic fingerprint is not less than the merged similarity threshold.

[0045] For inference synchronization events that are determined to be subject to the delayed merging strategy, multiple state windows of the same event type and belonging to the same semantic fingerprint cluster are merged into a window summary.

[0046] For inference synchronization events that are determined to be subject to the summary upload policy, the upload device identifier, log sequence number, event identifier, content summary, content hash value, semantic summary, event severity, and event timestamp are included.

[0047] Optionally, S5 includes:

[0048] The model version vector records the local model version identifier, the cloud-confirmed model version identifier, and the patch version identifier for each edge device;

[0049] The content hash value is generated from the content digest, event type, event severity, and device state vector of the inference synchronization event, and the semantic fingerprint is generated from the vectorized digest of the content hash value and the device state vector;

[0050] After receiving the inference synchronization event carrying the device identifier and log sequence number, the cloud performs orderly merging according to the device identifier, event timestamp and log sequence number, and retains the state increment after the confirmation level based on the confirmed log sequence number in the cloud confirmation vector currently recorded in the cloud.

[0051] The cloud-based confirmation vector records the confirmed log sequence number, confirmed model version identifier, conflict flag, and conflict event identifier for each edge device.

[0052] Furthermore, after receiving the model patch, the edge device verifies the model patch based on the target version identifier, patch content summary, and applicable device status conditions carried by the model patch.

[0053] After the verification is passed, the model patch is written into the candidate model area, and the patch version identifier corresponding to the candidate model area is written into the model version vector.

[0054] If the verification fails, a model patch conflict event is generated and written to the local event log.

[0055] Optionally, S6 includes:

[0056] The replay structure reads the local event log, the model version vector, and the saved cloud confirmation vector. It adds inference synchronization events with log numbers greater than the confirmed log numbers in the saved cloud confirmation vector, inference synchronization events with event identifiers consistent with the conflict event identifiers, and inference synchronization events with retransmission priority flags to the candidate retransmission set.

[0057] The retransmission priority flag is a flag that is written to the local event log when the connection is lost and the event severity is not less than the second severity threshold. The event identifier is an identifier recorded in the local event log along with the inference synchronization event.

[0058] Candidate retransmission events are sorted according to the synchronization priority, wherein the retransmission priority flag, conflict flag, model version difference, and event severity are used as the sorting inputs for the synchronization priority;

[0059] Read the hash values ​​of historical events whose confirmation status is confirmed from the local event log, and remove duplicate text from the retransmission list for candidate retransmission events with the same content hash value.

[0060] The beneficial effects of this invention are:

[0061] 1. By generating inference synchronization events at the edge that include inference results, confidence values, uncertainty scores, model version identifiers, input summaries, device state vectors, and link states, real-time inference output and synchronous scheduling input form a unified event carrier, which can reduce information gaps between on-site response and cloud synchronization.

[0062] 2. By using the synchronization priority calculator to prioritize the transmission of critical events, model inconsistency events, and uncertain events based on event severity, information freshness, uncertainty score, model version difference, link quality indicators, and cloud confirmation status output, the system can prioritize the transmission of critical events, model inconsistency events, and uncertain events when link resources are limited.

[0063] 3. By using semantic fingerprint deduplication structure and version vector confirmation replay structure, duplicate inference results and state windows can be suppressed. After connection loss is restored, a retransmission list is generated based on local event logs, model version vectors and confirmation vectors, thereby reducing the risk of missing and duplicate transmission of critical events. Attached Figure Description

[0064] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0065] Figure 1 This is a flowchart of an edge inference synchronization method for end-to-cloud collaboration.

[0066] Figure 2 This is a flowchart of step S3 of the present invention. Detailed Implementation

[0067] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0068] refer to Figures 1-2 An edge-cloud collaborative inference synchronization method includes:

[0069] S1. Edge devices collect on-site data and call local inference models to generate inference results, confidence values, uncertainty scores, event types, event severity, and event timestamps;

[0070] S2. Encapsulate the inference result, confidence value, uncertainty score, event type, event severity, model version identifier, input summary, content summary, event timestamp, device state vector, and link state into an inference synchronization event. The content summary is generated from the inference result and the device state vector.

[0071] S3. The synchronization module calculates the synchronization priority and determines the synchronization strategy based on the event severity, information freshness, uncertainty score, model version difference, link quality index and cloud confirmation status determined by the saved cloud confirmation vector.

[0072] S4. Upload inference synchronization events that are uploaded immediately or in summary, and merge, compress and deduplicate inference synchronization events that are delayed in merging or retained locally in the local event log.

[0073] S5. The cloud updates the device status copy according to the model version vector, the content summary and the status merging rules, returns the cloud confirmation vector, and returns the cloud review result or model patch when the conditions are met, so that the edge device updates the saved cloud confirmation vector and the model version vector.

[0074] S6. The edge device generates a retransmission list based on the local event log, model version vector, and saved cloud confirmation vector, and replays the retransmission list according to the synchronization priority when the link is restored or a confirmation gap occurs.

[0075] In this specific embodiment, S1 includes:

[0076] Edge devices perform frame-based processing of field data using a fixed acquisition time window, with the acquisition time window length set to [value missing]. Window step size set to This ensures that each input data frame corresponds to a continuous sequence of field data and has a unique acquisition start time as an event timestamp;

[0077] Each input data frame is denoted as ,in This is a multidimensional sensor data matrix arranged in chronological order. Each row of the matrix corresponds to a feature vector at a sampling time. The fields of the feature vector include acceleration, angular velocity, temperature, and voltage. (The last part, "edge device...", appears to be incomplete and requires further context.) Perform a preprocessing process consistent with the local inference model. The preprocessing process includes linear normalization by field and missing value imputation. Linear normalization uses the mean vector and standard deviation vector embedded in the model file. Missing value imputation adopts a forward preservation strategy to ensure that the input dimension is constant.

[0078] The local inference model adopts a multi-task neural network structure and fixes the model parameters. The shared feature extraction backbone network consists of two one-dimensional convolutional layers and one fully connected layer. The first one-dimensional convolutional layer has 32 output channels, a kernel size of 3, and a stride of 1. The second one-dimensional convolutional layer has 64 output channels, a kernel size of 3, and a stride of 1. A max pooling layer is inserted between the two convolutional layers with a kernel size of 2. The features after convolution and pooling are flattened and input into a fully connected layer with a hidden layer dimension of 128 and the ReLU activation function is used.

[0079] The classification header uses Softmax output. Category probability distribution of event categories ,in Each component corresponds to normal, slight abnormality, general abnormality, severe abnormality, and critical abnormality, respectively. The regression head outputs a regression output vector. and The two components represent the temperature rise amplitude and the vibration peak value, respectively;

[0080] Edge devices will distribute category probabilities The probability with the highest median value is used as the confidence score. And the category corresponding to the highest probability is taken as the reasoning result;

[0081] Edge devices generate uncertainty scores based on the dispersion of the category probability distribution, confidence values, and the deviation of the input data frame from the historical input summary set. The degree of dispersion Shannon entropy using class probability distribution and according to Normalization yields the deviation. The input summary is obtained by calculating the normalized Hamming distance between the input summary of the current input data frame and each input summary in the historical input summary set, and taking the minimum distance. The input summary is a binary summary obtained by performing a fixed-length hash operation on the sampling window identifier, feature field identifier, and preprocessing version identifier, and the historical input summary set is the set of the most recent input summaries from the edge device. A set of summaries maintained in log sequence number order on each input data frame;

[0082] The uncertainty score is obtained by linear weighting and satisfies The calculation formula is as follows:

[0083] ;

[0084] in Represents the uncertainty score. Indicates the degree of dispersion. Indicates the confidence level value. Indicates the deviation amount. The weights represent the degree of dispersion and take values ​​of 1. This represents the weight of the confidence term and its value is... This indicates the deviation weight and has a value of 0.2;

[0085] The edge device determines the event type and severity based on the inference results and a preset event mapping table. The event mapping table maps the category identifier corresponding to the inference result to an event type field and sets the event severity field to an integer level from 1 to 5. Specifically, the event severity is set to 5 for a critical anomaly, 4 for a severe anomaly, 3 for a general anomaly, 2 for a minor anomaly, and 1 for a normal result. Furthermore, the regression output vector... When the temperature rise amplitude component is not less than 15 or the vibration peak component is not less than 3.0, the event severity will be increased by one level based on the above, with the upper limit capped at 5, thereby generating an inference output result on the edge device side that includes inference results, confidence value, uncertainty score, event type, event severity and event timestamp.

[0086] In this specific embodiment, S2 includes:

[0087] The edge device obtains the inference result and confidence value after completing step S1. Uncertainty score After specifying the event type, event severity, and event timestamp, the input data frame to be used for this inference will be determined. Corresponding data collection start time With the length of the acquisition time window Combined to form a sampling window identifier ,in It is a monotonically increasing timestamp in milliseconds and is consistent with the event timestamp in step S1. The window length is fixed at 200ms;

[0088] Edge devices solidify the set of feature fields contained in the input data frame into feature field identifiers. ,in Using fixed-length bitmap encoding, with each bit representing whether acceleration, angular velocity, temperature, and voltage are involved in the input in the order of fields, the edge device solidifies the preprocessing process executed in step S1 into a preprocessing version identifier. ,in This is a 32-bit unsigned integer written to the header of the local inference model file and read when the model is loaded;

[0089] Edge devices and Perform a digest operation to obtain the input digest. The input summary is calculated using the following deterministic method:

[0090] ;

[0091] in This represents the input digest. SHA256(•) represents the function that performs a SHA-256 hash operation on the input byte sequence and outputs a 256-bit hash value. The operator represents concatenation according to a fixed byte order, and the concatenation order is... back back , This indicates a truncation function that extracts the first 128 bits from the most significant bit of a 256-bit hash value as the output. Indicates the sampling window identifier and is coded as 64-bit unsigned integer and The byte sequence of a 32-bit unsigned integer concatenated in little-endian order. This indicates that the feature field is identified and encoded as a 32-bit bitmap byte sequence. A sequence of bytes representing the preprocessed version identifier and encoded as a 32-bit unsigned integer;

[0092] The edge device simultaneously processes the category identifier and regression output vector in the inference results. Event type, event severity, and device state vector Perform a digest operation on the fields to obtain the content digest. The category identifier is a 16-bit unsigned integer with a value ranging from 0 to 4, and the regression output vector is... Each component is first converted to a 16-bit signed integer according to quantization rules to ensure consistent summary results across ends. The quantization rule is to multiply each component by 100, round it, and then truncate it to [ ]. The event type is an enumeration value that corresponds one-to-one with the category identifier and is encoded as an 8-bit unsigned integer. The event severity is an integer level from 1 to 5 and is encoded as an 8-bit unsigned integer.

[0093] Device state vector This data is collected synchronously by the edge device at the time of packaging and arranged in a fixed field order. The fixed field order is: CPU utilization, available memory, available storage, battery voltage, device internal temperature, and local inference queue length. CPU utilization is an integer percentage from 0 to 100 encoded as an 8-bit unsigned integer; available memory is in MB and encoded as a 32-bit unsigned integer; available storage is in MB and encoded as a 32-bit unsigned integer; battery voltage is in mV and encoded as a 16-bit unsigned integer; device internal temperature is represented as an integer multiplied by 10 in degrees Celsius and encoded as a 16-bit signed integer; and the local inference queue length is an integer from 0 to 65535 encoded as a 16-bit unsigned integer. (Content Summary) The content digest is calculated by using the same SHA-256 and truncation rules as the input digest and concatenating the byte sequences of the above fields in a fixed order, thereby satisfying the constraint that the content digest is generated from the inference result and the device state vector;

[0094] Edge devices encapsulate link state into link state vectors. ,in From available bandwidth Round-trip delay Packet loss rate and connection status composition, The result is obtained by dividing the number of valid bytes sent in the last 5 seconds by the time, expressed in kbps, and encoded as a 32-bit unsigned integer. The ACK round-trip time of the most recent 10 acknowledged data packets is calculated by taking the arithmetic mean and expressed in milliseconds, encoded as a 32-bit unsigned integer. The result is obtained by dividing the number of retransmitted packets in the last 5 seconds by the number of sent packets, and is represented as a fractional integer and encoded as a 16-bit unsigned integer. The connection state is enumerated and takes values ​​of 0 for disconnection and 1 for connection, and is encoded as an 8-bit unsigned integer.

[0095] Edge devices generate and persist device identifiers upon initial startup. ,in The log sequence number is a 128-bit random number stored as a hexadecimal string to ensure it remains unchanged after a restart. Edge devices maintain a monotonically increasing log sequence number. ,in It is a 64-bit unsigned integer that increments from 1, and is incremented by one for each inference synchronization event encapsulated. The edge device is based on and Generate event identifiers ,in To byte sequence and The identifier is obtained by concatenating the byte sequence in a fixed order, performing SHA-256, and truncating the first 128 bits.

[0096] Edge devices will use inference results and confidence values Uncertainty score Event type, event severity, model version identifier Input summary Content Summary Event timestamp Device state vector and link state vector Encapsulate it as an inference synchronization event and identify the device. Log sequence number Event Identifier The model version identifier is written together with the above fields into the event header of the inference synchronization event. The model version field is a 32-bit unsigned integer derived from the header of the local inference model file. The inference synchronization event adopts a binary serialization format with a fixed-length field header and a variable-length field body, and is encoded in the order of field name, field length and field value to ensure consistent parsing between the end and the cloud. After encapsulation, the inference synchronization event is appended to the local event log and enters the scheduling queue of the synchronization module.

[0097] In this specific embodiment, S3 includes:

[0098] The synchronization module retrieves an inference synchronization event from the scheduling queue at a fixed scheduling period of 100ms and reads its event severity and event timestamp. Uncertainty score Model version identifier Link state vector And the device identifier in the cloud-saved confirmation vector. Corresponding records ( ),in Available bandwidth in kbps The round-trip time is expressed in milliseconds (ms). The packet loss rate is expressed as a percentage of ten thousand (%). Enumerate the connection states and take 0 for disconnected and connected states. This is the sequence number of the log that has been confirmed in the cloud. This is a model version identifier that has been confirmed in the cloud. For conflict flags, take 0 if there is no conflict, and take 0 if there is a conflict. A marker for conflict events tagged in the cloud;

[0099] The synchronization module reads the current scheduling time. And calculate the time difference Then, through a preset monotonically decreasing mapping table, Mapped to information freshness score , among which when hour ,when hour ,when hour ,when hour ,when hour ;

[0100] The synchronization module maps event severity to severity scores using a fixed mapping table. When the severity of the event is 1 When the severity of the incident is 2 When the severity of the incident is 3 When the severity of the incident is 4 When the severity of the incident is 5 ;

[0101] The synchronization module will score the uncertainty. Mapped to uncertainty score ,in Pick And limited to 0 to 100;

[0102] Synchronization module calculates model version difference As a model version identifier Confirmed model version identifier in the cloud version number difference and when hour Take 0, and Mapped to version difference score ,in hour hour hour ;

[0103] Synchronization module based on link state vector Calculate link quality metrics Map it to link quality score Among them, Normalization with a maximum limit of 5000kbps is performed to obtain bandwidth sub-fractions. The delay sub-fraction is obtained by performing inverse normalization with a maximum duration of 500ms. Performing reverse normalization with a maximum value of 1000 yields the score for dropping buns. Directly used as the connective sub-score, then weighted. Weighted fusion is performed to obtain and take And limited to 0 to 100;

[0104] The synchronization module consists of Determine the cloud confirmation status and map it to a confirmation status score. , among which when And the event identifier of the current inference synchronization event is And satisfy The cloud confirmation status is conflict confirmation and ,when And the log sequence number of the current inference synchronization event. Not greater than The cloud confirmation status is "confirmed" and In all other cases, the cloud confirmation status is "unconfirmed" and ;

[0105] The synchronization module inputs the above scores into the synchronization priority calculator to obtain the synchronization priority. The calculation formula is as follows:

[0106] ;

[0107] in Indicates synchronization priority. Indicates the severity score. This indicates the score for the freshness of information. Indicates the uncertainty score. This indicates the version difference score. Indicates the link quality score. Indicates the confirmation status score;

[0108] Before calculating the synchronization strategy, the synchronization module generates the current semantic fingerprint. It also compares the similarity with historical semantic fingerprints recorded in the local event log, where... Generated using deterministic computation methods:

[0109] Content summary Event type, event severity, and device state vector The quantized byte sequence is concatenated in a fixed order, then SHA-256 is performed, and the first 128 bits are truncated as... Device state vector The quantization rules are consistent with those in step S2 to ensure that the same state generates the same semantic fingerprint;

[0110] Semantic fingerprint similarity is calculated using bit-level similarity and is taken as the percentage of bits in two 128-bit semantic fingerprints that have the same value at the same bit position.

[0111] The filtering scope of historical semantic fingerprints is limited to historical events within the confirmation retention period. In this embodiment, the confirmation retention period is defined as the period in which the confirmation status in the local event log is "confirmed" and the log sequence number is within the specified range. A set of historical events within the interval;

[0112] The synchronization module outputs its initial hit result in the following order: disconnection determination, critical event determination, link transmitlable determination, duplicate event determination, and default determination to determine the synchronization strategy. The disconnection determination is based on the connection status. For input and when And the severity of the event is less than the second severity threshold. Output the local retention policy and set the upload status to not uploaded when... And the severity of the incident is no less than Output the local retention policy and write a retransmission priority flag for the event in the local event log; critical event determination is performed in... It takes effect when the event severity is not less than the first severity threshold. or uncertainty score Not less than the uncertainty threshold Or model version differences Not less than the version difference threshold Output the immediate upload policy and place the event in the immediate upload queue;

[0113] Link transmittability determination is made after the critical event determination is not met and Effective at the time and when the link quality index Less than the link's transmittable threshold And the severity of the event is less than Output the summary upload strategy and place the event in the summary upload queue;

[0114] The duplicate event determination is made when the aforementioned determination is not met and Effective when the semantic fingerprint similarity is not less than the merged similarity threshold. Furthermore, when similar historical semantic fingerprints correspond to historical events that are within the confirmation retention period, a delayed merging strategy is output and the event is marked as a mergeable event;

[0115] The default judgment is that the aforementioned judgment is not met and Output summary upload strategy in real time;

[0116] The synchronization module will synchronize priorities. The synchronization policy is written into the scheduling field of the inference synchronization event and the synchronization priority and synchronization policy fields in the local event log are updated for use in steps S4 and S6.

[0117] In this specific embodiment, S4 includes:

[0118] The edge device has already written the synchronization policy and synchronization priority for step S3. The closed-loop process of inference synchronization event execution upload and local log processing is implemented, where the local event log is arranged according to log sequence number. The incrementally written append log is persistently stored in the non-volatile storage medium of the edge device. Each record in the append log contains a device identifier. Log sequence number Event identifier Model version identifier Content Summary Content hash value Semantic fingerprint Event type, event severity, event timestamp Synchronization priority Synchronization strategy, upload status and confirmation status, where the upload status can be none, uploaded, or text-suppressed upload, and the confirmation status can be none, confirmed, or conflict confirmation.

[0119] The semantic fingerprint deduplication structure calculates the content hash value before uploading or merging each inference synchronization event. Content hash value By summarizing the content Event type, event severity, and device state vector The quantized byte sequence is concatenated in a fixed order, then SHA-256 is performed, and the first 128 bits are truncated to obtain the device state vector. The field order is consistent with step S2 and the quantization encoding rule is consistent with step S2 to ensure that duplicate events on the end side obtain consistent content hash values;

[0120] Semantic fingerprint deduplication structure construction state window and calculation of semantic summary The status window is limited to the same device identifier. Below is the event timestamp The window ends at the time specified in the original text. The length of the window is the area containing the device. Append the device state vector already written in the log within the interval Corresponding event timestamp and corresponding log sequence number Semantic summarization A fixed-dimensional state statistical vector is formed by calculating the mean and maximum value of each device status field within the status window. This state statistical vector is then quantized into a fixed-length byte sequence according to field order, and SHA-256 is performed, followed by truncating the first 64 bits to obtain the semantic digest. It is sensitive to state changes within the state window and the calculation results are deterministic.

[0121] Semantic fingerprint Based on content hash value With semantic summarization The combined generation and appending of the log is calculated using the following formula:

[0122] ;

[0123] in Represents semantic fingerprint, This function performs a SHA-256 hash operation on the input byte sequence and outputs a 256-bit hash value. This indicates a truncation function that extracts the first 128 bits from the most significant bit of a 256-bit hash value as the output. The operator represents concatenation according to a fixed byte order, and the concatenation order is... back , This represents the content hash value and is a 128-bit binary sequence. It represents a semantic digest and is a 64-bit binary sequence;

[0124] The semantic fingerprint deduplication structure determines the confirmation retention period in the local event log and performs deduplication. The confirmation retention period is defined as the log sequence number not being greater than the cloud-confirmed log sequence number corresponding to the device in the saved cloud confirmation vector. And located in the interval The historical data set is used to calculate semantic fingerprint similarity at the bit level, taking the percentage of bits in two 128-bit semantic fingerprints that have the same value at the same bit position. The deduplication criterion is the current semantic fingerprint. The similarity with any historical semantic fingerprint within the retention period is not less than the merged similarity threshold. ;

[0125] When the synchronization strategy for the inference synchronization event is immediate upload or summary upload, the upload module will perform the upload according to the synchronization strategy if the deduplication check fails. Immediate upload uses the complete set of upload fields, including the inference result and confidence value. Uncertainty score Model version identifier Input summary Content Summary Device state vector Link state vector Event timestamp Equipment identification Log sequence number Event Identifier The summary upload uses a set of summary upload fields and only includes the device identifier. Log sequence number Event identifier Content Summary Content hash value Semantic summary Event severity and event timestamp And after successful transmission, the upload status of the record in the append log will be set to uploaded;

[0126] When the deduplication determination is successful, the semantic fingerprint deduplication structure suppresses the repeated text upload of the inference synchronization event and switches the sending load of the upload module to the digest upload field set. At the same time, the upload status of the record in the append log is set to text suppression upload and the synchronization strategy field is kept unchanged so that S6 can still perform re-upload sorting based on the synchronization strategy and synchronization priority.

[0127] When the synchronization strategy for inference synchronization events is delayed merging, the local merging structure maintains aggregate entries using event type and semantic fingerprint cluster as keys. The semantic fingerprint cluster is defined as the sequence of events related to the current semantic fingerprint. The similarity is not less than The inference synchronization event set is limited to the same event type. The aggregated entries store the window start timestamp, window end timestamp, cumulative event count, maximum event severity, and cumulative sum and maximum value of the state window statistical vector within the cluster. This merges multiple state windows within the same semantic fingerprint cluster into a window summary and generates a window summary record only when the difference between the window end timestamp and the window start timestamp reaches 10s or the cumulative event count reaches 50. The record is written back to the append log and the upload status is set to not uploaded to enter subsequent scheduling.

[0128] When the synchronization policy for inference synchronization events is local retention, the edge device will only store the event... Synchronization priority The synchronization strategy, upload status, and confirmation status are written to the append log without performing network transmission. This allows delayed merging or locally retained inference synchronization events to be merged, compressed, and deduplicated in the local event log, providing traceable log evidence for S5's cloud confirmation and step S6's confirmation replay.

[0129] In this specific embodiment, S5 includes:

[0130] The cloud synchronization service receives inference synchronization events uploaded by edge devices and parses them to obtain the device identifier. Log sequence number Event identifier Model version identifier Content Summary Device state vector Event timestamp and the content hash value carried by the edge side with the event With semantic fingerprints ,in For persistent and unique identifiers of edge devices, This is a monotonically increasing append log sequence number on the device side. For the reason and The generated event is a unique identifier. This serves as the model version identifier for the current local inference model on the edge device. This is a content summary generated from the inference results and the device state vector, where s is the device state vector encoded in a fixed field order. For event timestamps;

[0131] The cloud for each Maintain cloud-based confirmation vector records With model version vector record ,in middle This represents the highest consecutive log sequence number water level that the cloud has confirmed for this device. This is an identifier for the model version that has been confirmed in the cloud. For conflict flags, take 0 if there is no conflict, and take 0 if there is a conflict. For identifying conflict events, middle This is the identifier of the local model version most recently observed from this device event in the cloud. This is an identifier for a patch version that has been distributed from the cloud and is awaiting confirmation and activation by the device.

[0132] Before being stored in the database, the cloud performs a consistency check on each inference synchronization event to ensure idempotent merging. The check includes a deterministic summary and quantization rule pair based on the same criteria as in step S4. (Event type, event severity) Recalculate the content hash and match it with the event's carried hash. Compare, and correct ( (Device status window semantic summary) recalculate semantic fingerprint and combine it with the event-carried semantic fingerprint. When any comparison fails, the event is marked as a verification failure and placed in the cloud exception queue without being elevated. ;

[0133] The cloud processes the verification pass event by ( The sorting keys are used to perform ordered merging and write to the receive buffer of each device. The sorting rule is to first sort by... Ascending order, then press Ascending order, and when the received events satisfy... The state increment is discarded directly and only the received statistics are recorded to avoid duplicate merging;

[0134] Cloud-based update based on conflict determination rules The conflict determination rule is that if the same [something] already exists in the cloud... Same as Historical events and historical events With the current event If they are not the same, then Set to 1 and Set as the current event At the same time, current events and historical events are written into the conflict table for subsequent review and processing;

[0135] Cloud-based rule updates are based on confirmed water levels. The water level advance rule is confirmed to be from The process begins by searching the receive buffer for consecutive event sequences that have passed verification and have not been marked as conflict acknowledgments. These sequences are then dequeued sequentially and merged until a sequence number gap or conflict acknowledgment event is encountered. The largest consecutive sequence number successfully merged is then written back. ;

[0136] The cloud maintains a copy of the device state during state merging. The set of device status replica fields and the edge-side device status vector The field set is consistent and includes CPU utilization, available memory, available storage, battery voltage, device internal temperature, and local inference queue length. The state merging rule uses the last write of a field to overwrite the previous write, based on the event timestamp. As a basis for version comparison, that is, if and only if the current event The cloud field value is overwritten only when the value of the field corresponding to the current event is not less than the most recent update timestamp already recorded in the field, and the most recent update timestamp of the field is updated synchronously to ensure that the merge result is certain and traceable;

[0137] The cloud updates the model version vector record after the merge is complete. The updated rules will be Set to be carried in the latest event that has been received, verified, and is participating in the merging. and will Set as The latest confirmed event carries the following: ;

[0138] The cloud returns a cloud acknowledgment vector record for each edge device it responds to. This enables edge devices to update their saved cloud-confirmed vectors and returns a model patch when patch delivery conditions are met. These patch delivery conditions are defined by the conditions recorded in the cloud. Smaller than the target version identifier currently released in the cloud And the cloud has already generated from arrive Patch version identifier The corresponding patch file;

[0139] Model patches include target version identifiers Patch version identifier Patch Summary Applicable equipment status conditions ,in To obtain the digest by performing SHA-256 on the byte sequence of the patch file and truncating the first 128 bits, For the device state vector The set of field threshold constraints is defined in this embodiment as CPU utilization not greater than 80%, available memory not less than 256MB, and local inference queue length not greater than 10.

[0140] After receiving the model patch, the edge device performs a verification process and updates the model version vector. The verification process includes verifying the target version identifier. Compared with the locally saved cloud confirmation vector Consistent, recalculate the patch summary for the received patch file and match it. Consistent with and read the device state vector at the time of reception. Verify each field to ensure it meets the applicable device status conditions. When all the above checks pass, the patch file is written to the candidate model area and the patch version corresponding to the candidate model area is identified. Write the model version vector maintained locally on the edge device. When any verification fails, a model patch conflict event is generated and written as an inference synchronization event to the local event log for subsequent synchronization.

[0141] In this specific embodiment, S6 includes:

[0142] Edge devices will monitor connection status in the network monitoring thread. The confirmation replay structure is triggered the instant the connection is switched from 0 to 1, and is repeatedly triggered at a fixed period of 1 second in the connected state to handle new confirmations and conflicting confirmations in the cloud.

[0143] The confirmation replay structure first reads the corresponding record for the device from the saved cloud confirmation vector. ),in This indicates that the log sequence number level has been confirmed in the cloud. This indicates that the model version identifier has been confirmed in the cloud. The conflict flag indicates whether there is a conflict or not. A value of 0 is used if there is a conflict. Indicates a conflict event identifier;

[0144] After confirming the replay structure, the local event log is read and sorted by device identifier. The filter yields an append log stream for the device, where each record contains a log sequence number. Event identifier Model version identifier Content hash value Synchronization priority Event severity field, retransmission priority flag Upload status and confirmation status, with priority flag for re-upload. A binary flag, set to 0 or 1, is written to the local event log when the connection is lost in step S3 and the event severity is not less than the second severity threshold.

[0145] The confirmation replay structure adds inference synchronization events that satisfy any of the following conditions to the candidate replay set: condition one is... And the status is not confirmed, condition two is and Condition three is Furthermore, the confirmation status is not confirmed, thus simultaneously covering the retransmission of confirmation gaps, replay of conflict events, and priority retransmission of critical disconnection events;

[0146] The replay structure is used to calculate the model version difference for each event in the candidate replay set. The calculation formula is as follows:

[0147] ;

[0148] in Indicates the difference between model versions. This indicates the model version identifier carried by the candidate retransmission event. This indicates that the confirmed model version identifier has been saved in the cloud confirmation vector. This operation represents taking the larger of two input values.

[0149] The replay structure confirms that the candidate replay set is generated into a replay list and a deterministic sort is performed. The sorting rule adopts a multi-key descending order, and when the keys are equal, the log number is used. Ascending stable sorting, with key priority from high to low as the priority flag for supplementary transmission. Conflict signs Conflict matching conditions Model version differences Event severity field, synchronization priority This ensures that critical disconnection events and conflict events are replayed first after the link is restored, and that events with inconsistent model versions are replayed before ordinary events.

[0150] Before generating the final replay list, the confirmed replay structure performs duplicate text deletion to suppress unnecessary bandwidth usage. This process involves reading historical events with a confirmed status from the local event log and extracting their content hash values. Form a confirmed hash set, then iterate through the supplementary transmission list and hash the content. Candidate retransmission events appearing in the confirmed hash set switch their retransmission payload from full text payload to summary payload and mark them as text deletion in the retransmission list entry, thereby avoiding duplicate text being sent up again without deleting the retransmission entry itself and preserving the cloud side's ability to confirm the continuity of log sequence numbers.

[0151] The confirmation replay structure writes each entry into the upload queue in the order of the retransmission list during replay and waits for a new cloud confirmation vector from the cloud to update the saved cloud confirmation vector. Promotion or conflict mark When clearing, the confirmation status of the corresponding confirmed entry is updated to confirmed and removed from the pending upload status. When the upload fails, the entry is reinserted into the tail of the upload queue at a fixed retry interval of 500ms, and its sorting key remains unchanged to ensure that critical events are not occupied by low-priority events for a long time.

[0152] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0153] This invention connects local real-time inference on edge devices, synchronous priority calculation, semantic fingerprint deduplication, cloud state merging, and confirmation replay into a continuous processing chain, enabling on-site events to be responded to at the edge without waiting for centralized cloud processing. At the same time, it maintains traceable consistency of cloud and edge states through confirmation vectors and model version vectors.

[0154] This invention sets up an event synchronization priority calculator, a semantic fingerprint deduplication structure, and a version vector confirmation replay structure within the synchronization module. This makes the synchronization strategy not only dependent on a fixed period or a single link state, but also combined with the criticality of the event, the freshness of the information, uncertainty, version differences, and cloud confirmation gaps. This makes it more suitable for edge inference synchronization scenarios with unstable networks and concurrent access from multiple devices.

Claims

1. A method for synchronizing edge inference in a cloud-edge collaborative manner, characterized in that, include: S1. Edge devices collect on-site data and call local inference models to generate inference results, confidence values, uncertainty scores, event types, event severity, and event timestamps; S2. Encapsulate the inference result, confidence value, uncertainty score, event type, event severity, model version identifier, input summary, content summary, event timestamp, device state vector, and link state into an inference synchronization event. The content summary is generated from the inference result and the device state vector. S3. The synchronization module calculates the synchronization priority and determines the synchronization strategy based on event severity, information freshness, uncertainty score, model version difference, link quality index and cloud confirmation status determined by the saved cloud confirmation vector. S4. Upload inference synchronization events that are uploaded immediately or in summary, and merge, compress and deduplicate inference synchronization events that are delayed in merging or retained locally in the local event log. S5. The cloud updates the device status copy according to the model version vector, content summary and status merging rules, returns the cloud confirmation vector, and returns the cloud review result or model patch when the conditions are met, so that the edge device updates the saved cloud confirmation vector and model version vector. S6. The edge device generates a retransmission list based on the local event log, model version vector, and saved cloud confirmation vector, and replays the retransmission list according to the synchronization priority when the link is restored or a confirmation gap occurs.

2. The edge inference synchronization method for end-to-cloud collaboration according to claim 1, characterized in that, S1 includes: The field data is organized into input data frames according to the collection time window, and the input data frames are input into the local inference model to obtain the category probability distribution and regression output; The probability of the first value in the category probability distribution is taken as the confidence value; The uncertainty score is generated based on the degree of dispersion of the category probability distribution, the confidence value, and the deviation of the input data frame from the historical input summary set. The event type and the event severity are determined based on the inference results and the preset event mapping table.

3. The edge inference synchronization method for end-to-cloud collaboration according to claim 1, characterized in that, S2 includes: The field data is organized into input data frames according to the acquisition time window, and a digest operation is performed on the sampling window identifier, feature field identifier, and preprocessing version identifier of the input data frames to obtain the input digest. A digest operation is performed on the category identifier, regression output, event type, event severity, and fields in the device state vector in the inference result to obtain the content digest. The available bandwidth, round-trip latency, packet loss rate, and connection status are combined to form the link status. The edge device is assigned a device identifier and log sequence number, and the device identifier, the log sequence number, the event identifier, and the information in step S2 are written into the inference synchronization event.

4. The edge inference synchronization method for end-to-cloud collaboration according to claim 1, characterized in that, S3 includes: the information freshness is obtained by the time difference between the event timestamp and the synchronization scheduling time through a preset monotonically decreasing mapping table; the model version difference is obtained by the version sequence number difference between the model version identifier and the confirmed model version identifier of the corresponding device in the saved cloud confirmation vector; the link status includes available bandwidth, round-trip time, packet loss rate and connection status, and the link quality index is obtained by normalized fusion of the available bandwidth, the round-trip time, the packet loss rate and the connection status; the cloud confirmation status includes unconfirmed, confirmed and conflict confirmed, and is determined by the confirmed log sequence number, the confirmed model version identifier and the conflict flag in the saved cloud confirmation vector.

5. The edge inference synchronization method for end-to-cloud collaboration according to claim 1, characterized in that, S3 includes: generating a current semantic fingerprint based on the content summary, event type, event severity, and device state vector of the current inference synchronization event; comparing the current semantic fingerprint with historical semantic fingerprints recorded in the local event log to obtain semantic fingerprint similarity; outputting the first hit in the order of disconnection determination, critical event determination, link transmittable determination, duplicate event determination, and default determination; outputting a local retention policy when the link status is disconnected and the event severity is less than a second severity threshold; outputting a local retention policy and writing a retransmission priority flag when the link status is disconnected and the event severity is not less than the second severity threshold; and outputting a local retention policy and writing a retransmission priority flag when the link status is not disconnected and the event severity is not less than the second severity threshold. If the event severity is not less than the first severity threshold, or the uncertainty score is not less than the uncertainty threshold, or the model version difference is not less than the version difference threshold, an immediate upload strategy is output. If the aforementioned strategy is not hit, and the link status is not disconnected, the link quality index is less than the link transmittable threshold, and the event severity is less than the first severity threshold, a summary upload strategy is output. If the aforementioned strategy is not hit, and the semantic fingerprint similarity is not less than the merge similarity threshold, and the historical event corresponding to the historical semantic fingerprint similar to the current semantic fingerprint is within the confirmation retention period, a delayed merging strategy is output. If the aforementioned strategy is not hit and the link status is not disconnected, a summary upload strategy is output.

6. The edge inference synchronization method for end-to-cloud collaboration according to claim 1, characterized in that, S5 includes: The model version vector records the local model version identifier, the cloud-confirmed model version identifier, and the patch version identifier for each edge device; the content hash value is generated from the content summary, event type, event severity, and device state vector of the inference synchronization event; and the semantic fingerprint is generated from the vectorized summary of the content hash value and the device state vector. After receiving the inference synchronization event carrying the device identifier and log sequence number, the cloud performs orderly merging according to the device identifier, event timestamp, and log sequence number, and retains the state increment after the confirmation level based on the confirmed log sequence number in the cloud confirmation vector currently recorded in the cloud; the cloud confirmation vector records the confirmed log sequence number, confirmed model version identifier, conflict flag, and conflict event identifier of each edge device.

7. The edge inference synchronization method for end-to-end cloud collaboration according to claim 6, characterized in that, S6 includes: The replay structure reads the local event log, the model version vector, and the saved cloud confirmation vector. It adds inference synchronization events with log numbers greater than the confirmed log numbers in the saved cloud confirmation vector, inference synchronization events with event identifiers consistent with the conflict event identifiers, and inference synchronization events with retransmission priority flags to the candidate retransmission set. The retransmission priority flag is a flag that is written to the local event log when the connection is lost and the event severity is not less than the second severity threshold. The event identifier is an identifier recorded in the local event log along with the inference synchronization event. Candidate retransmission events are sorted according to the synchronization priority, wherein the retransmission priority flag, conflict flag, model version difference, and event severity are used as sorting inputs for the synchronization priority. The hash values ​​of the content of historical events with a confirmed status are read from the local event log, and duplicate text is deleted from the retransmission list for candidate retransmission events with the same content hash value.

8. The edge inference synchronization method for end-to-cloud collaboration according to claim 5, characterized in that, S4 includes: a semantic fingerprint deduplication structure performs content hashing on the content summary, the event type, the event severity, and the device state vector to obtain a content hash value, and performs vectorized summarization on the device state vector in the state window to obtain a semantic summary; the state window consists of the device state vector, event timestamp, and log sequence number within a preset collection time window under the same device identifier; a semantic fingerprint is generated based on the content hash value and the semantic summary; the retention period is a log interval in the local event log where the log sequence number is not greater than the confirmed log sequence number in the saved cloud confirmation vector and does not exceed a preset retention length; when the similarity between the semantic fingerprint and the historical semantic fingerprint within the retention period is not less than the merged similarity threshold, the duplicate text upload of the corresponding inference synchronization event is suppressed.

9. The edge inference synchronization method for end-to-cloud collaboration according to claim 8, characterized in that, S4 includes: The local event log is an append log written in ascending order of log sequence number. The append log records device identifier, log sequence number, event identifier, model version identifier, content summary, content hash value, semantic fingerprint, event type, event severity, event timestamp, synchronization priority, synchronization strategy, upload status, and confirmation status. The semantic fingerprint cluster is a set of inference synchronization events whose similarity to the current semantic fingerprint is not less than the merging similarity threshold. For inference synchronization events determined to be under the delayed merging strategy, multiple status windows of the same event type and belonging to the same semantic fingerprint cluster are merged into a window summary. For inference synchronization events determined to be under the summary upload strategy, the device identifier, log sequence number, event identifier, content summary, content hash value, semantic summary, event severity, and event timestamp are uploaded.

10. The edge inference synchronization method for end-to-end cloud collaboration according to claim 6, characterized in that, Step S5 is followed by: after receiving the model patch, the edge device verifies the model patch according to the target version identifier, patch content summary and applicable device status conditions carried by the model patch; if the verification is successful, the model patch is written into the candidate model area and the patch version identifier corresponding to the candidate model area is written into the model version vector; if the verification fails, a model patch conflict event is generated and written into the local event log.

Images