Charging pile data compression method based on multi-modal fusion and adaptive dynamic coding
By using four-dimensional data synchronization alignment and weighted similarity measurement, combined with derivative features and hierarchical dictionaries, the problem of high data storage and transmission costs for charging piles is solved, achieving efficient data compression and improved system scalability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-04-14
AI Technical Summary
Existing technologies cannot effectively mine the modal correlation characteristics of multi-dimensional time-series data of charging piles, resulting in high data storage and transmission costs, and difficulty in adapting to the dynamic characteristics of different charging stages and vehicle models. Linear transformation methods introduce distortion, and sliding window methods have large computational overhead.
We employ four-dimensional data synchronization alignment and weighted similarity measurement, combined with voltage and current derivative features to identify charging stages, and use a hierarchical dictionary and greedy clustering to extract frequent patterns. We also optimize and compress data through edge-cloud collaboration.
It achieves precise elimination of data redundancy, adapts to the dynamic characteristics of the charging process and the differences in vehicle models, reduces edge computing and storage overhead, and improves system scalability.
Smart Images
Figure CN121864103A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electric vehicle technology, specifically relating to a charging pile data compression method based on multimodal fusion and adaptive dynamic coding. Background Technology
[0002] As a key supporting infrastructure for ensuring the continuous operation of the industry, charging piles continuously generate massive amounts of multi-dimensional time-series data during long-term operation. The storage and transmission costs of this type of data remain high, and how to achieve efficient compression of this time-series data, thereby reducing the resource consumption in data storage and transmission, has become one of the key technical bottlenecks restricting the intelligent management and efficient operation of charging networks.
[0003] However, existing technologies have the following shortcomings: they do not deeply explore the correlation and complementarity between modes such as voltage and current for single-type data or simple parallel multi-channel data; key parameters (such as window length and similarity threshold) are mostly statically preset, which cannot adapt to the dynamic data characteristics of different charging stages such as constant current and constant voltage, and are also difficult to cope with the differences in different vehicle models and battery health states; linear transformation methods (such as PCA) cannot characterize data nonlinearity and local quasi-repetitive patterns; fixed model methods (such as fixed measurement matrix) are prone to introducing uncontrollable distortion when pursuing high compression rates; sliding window methods need to maintain a large amount of historical data, resulting in significant computational and storage pressure at the edge of the charging pile.
[0004] Therefore, a new method is urgently needed. Summary of the Invention
[0005] The purpose of this invention is to provide a charging pile data compression method based on multimodal fusion and adaptive dynamic coding. This method uses four-dimensional data synchronization alignment and weighted similarity measurement to deeply fuse multimodal association information to accurately eliminate redundancy; it identifies charging stages based on voltage and current derivative features and coordinates the switching of core parameters to adapt to dynamic characteristics and vehicle differences; it relies on a hierarchical dictionary and greedy clustering to extract frequent patterns, taking into account both general reuse and transient capture to reduce reconstruction distortion; and it reduces edge-side overhead and improves system scalability through edge-cloud collaborative optimization.
[0006] To achieve the above objectives, this invention provides a charging pile data compression method based on multimodal fusion and adaptive dynamic coding, comprising the following steps: S1. Collect four-dimensional multimodal time series data of charging piles, perform timestamp alignment and outlier removal, and obtain preprocessed four-dimensional time series data; S2. Calculate derivative features based on the preprocessed four-dimensional time series data in S1, identify constant current, constant voltage and transition stages, and dynamically configure window length, similarity threshold, frequent pattern frequency threshold and modal weight coefficient. S3. Using the window length configured in S2, perform sliding window segmentation on the preprocessed four-dimensional time series data obtained in S1 with a fixed step size to generate subsequence fragments; S4. Based on the subsequence fragments obtained in S3, calculate the relative differences of each modality, obtain the comprehensive similarity score by configuring the modality weight coefficients, compare the comprehensive similarity score with the similarity threshold configured in S2, determine the similarity of data segments and mark the unmatched residual data; obtain the similarity judgment result and the original value of the residual data. S5. Using a greedy clustering algorithm, based on the similarity judgment results of S4, the subsequence segments in S3 that meet the similarity conditions are merged. Combined with the frequency threshold configured in S2, frequent patterns are filtered and representative pattern sequences are generated. A hierarchical dictionary containing a basic dictionary, a vehicle-specific dictionary, and a conversational temporary dictionary is constructed. At the same time, new pattern features are extracted. S6. Using the hierarchical dictionary constructed in S5, tokenization replacement is performed on the successfully matched subsequence segments in S3, retaining the original values of the residual data marked in S4, and generating a hybrid compressed sequence. S7: The edge side performs real-time compression based on S1-S6, and uploads the hybrid compressed sequence obtained by S6 and the new pattern features extracted by S5 to the cloud; the cloud aggregates data to optimize and update the basic dictionary and the vehicle-specific dictionary and generates a differential update package, which is then distributed to each edge node to complete the dictionary closed-loop update. During data reconstruction, tokens in the mixed compressed sequence are restored into a multimodal data sequence based on the updated hierarchical dictionary, and then concatenated with the residual data in chronological order to recover the original data.
[0007] Preferably, in S1, the four-dimensional time-series data is the battery pack voltage. Charging current State of charge and battery pack temperature The data format is .
[0008] Preferably, in S2, the derivative characteristics are calculated using the central difference method, and the formula is: ; ; In the formula, for The battery pack voltage values collected at all times; for The battery pack voltage values collected at all times; for The charging current value collected at all times; for The charging current value collected at all times; The sampling interval is... ; During the charging phase, the derivative threshold is used. determination: and This is the constant current stage; and This is the constant pressure stage; The others are transitional phases.
[0009] Preferably, in S2, the modal weighting coefficients are configured as follows: Constant current stage: ; Constant pressure stage: ; Transition phase: .
[0010] Preferably, in S3, the fixed step size of the sliding window is 1 / 2 of the window length, and the first step is... The formula for calculating each data segment is: ; In the formula, The start time of valid data; The first result obtained after dividing the sliding window A multimodal data segment; This is the sequence number of the data segment; This represents the movement step size of the sliding window; The length of the sliding window.
[0011] Preferably, in S4, the formula for calculating the relative difference between single modes is: ; ; ; ; In the formula, The relative differences under voltage modes; The relative differences under current modes; The relative differences under the charged state modes; The relative differences under temperature modes; For data segment The first in Each voltage value; For data segment The first in Each voltage value; For data segment The first in One current value; For data segment The first in One current value; For data segment The first in One SOC value; The table contains data segments. The first in One SOC value; For data segment The first in A temperature value; For data segment The first in A temperature value.
[0012] Preferably, in S4, the formula for calculating the comprehensive similarity score is: ; In the formula, These are the weighting coefficients for the voltage modes; These are the weighting coefficients for the current modes; These are the weighting coefficients for the SOC mode; These are the weighting coefficients for the temperature modes; when The similarity of data segments is determined at that time.
[0013] Preferably, in S5, greedy clustering uses Euclidean distance for determination; frequent clustering requires satisfying the requirement of a certain number of data segments within a group. Frequency threshold, a representative pattern sequence is represented as: ; In the formula, A standardized representative pattern sequence; This is the average sequence of voltage samples across all data segments within this mode group; This is the average sequence of current sample values for all data segments within this mode group; This is the sequence of average state-of-charge samples for all data segments within this mode group; This is the mean sequence of temperature samples for all data segments within this mode group.
[0014] Therefore, the present invention employs the above-mentioned charging pile data compression method based on multimodal fusion and adaptive dynamic coding. Compared with the prior art, the technical solution of the present invention has the following beneficial effects: (1) By using a four-dimensional data synchronization alignment and weighted similarity measurement model, the correlation information of voltage, current, SOC and temperature is deeply integrated to more accurately eliminate data redundancy; (2) Based on the characteristics of voltage and current derivatives, the charging stage is identified in real time, and parameters such as window length, similarity threshold, and modal weight are switched in coordination to adapt to the dynamic characteristics of the charging process and the differences in vehicle models; (3) By using a hierarchical dictionary and greedy clustering to extract frequent patterns, we can ensure the reuse of general patterns and capture specific transient patterns, thereby reducing reconstruction distortion. (4) Lightweight real-time compression is performed on the edge side, and the dictionary is aggregated and optimized in the cloud and distributed differentially to solve the problem of high storage and computing overhead on the edge side and improve the scalability of the system.
[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0016] Figure 1 This is a flowchart of an embodiment of the charging pile data compression method based on multimodal fusion and adaptive dynamic coding of the present invention; Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.
[0018] Example 1 like Figure 1 As shown, this embodiment provides a charging pile data compression method based on multimodal fusion and adaptive dynamic coding. It should be understood that the specific parameters, models and protocols mentioned in this embodiment are only examples to help those skilled in the art understand the present invention, and are not intended to limit the present invention.
[0019] The charging pile data compression method based on multimodal fusion and adaptive dynamic coding of the present invention includes the following steps: S1. Synchronously collect and record multimodal charging data during a single charging process at a fixed sampling frequency; and perform timestamp alignment on the data, removing zero-value abnormal data from the first and last frames; to obtain preprocessed four-dimensional time-series data. In this step, the multimodal charging data includes: battery pack voltage. Charging current State of charge and battery pack temperature The preprocessed four-dimensional time series data is represented as follows: ; Timestamp alignment is achieved through GPS-synchronized timestamps, satisfying... Outlier removal requires setting a zero-value threshold. Eliminate those that meet the requirements Zero-value data of the first and last frames; S2 receives the preprocessed four-dimensional time-series data transmitted by S1, and calculates the first derivatives of voltage and current using the central difference method to characterize the dynamic trend of data changes. The formula is: ; ; In the formula, for The battery pack voltage values collected at all times; for The battery pack voltage values collected at all times; for The charging current value collected at all times; for The charging current value collected at all times; The sampling interval is... The first and last frames are supplemented with calculations using the one-sided difference method. Based on the derivative characteristics of voltage and current, the charging stage (constant current, constant voltage, and transition stage) is identified in real time, and the corresponding parameters (window length) are dynamically configured. Similarity threshold Frequent pattern frequency threshold Weighting coefficients During the charging phase, the derivative threshold is used. determination: Constant current stage: And the voltage rises, configuration , , , ; Constant pressure stage: And the current drops, configuration , , , ; Transition Phase: Configuration , , , ; S3, Receive the preprocessed four-dimensional timing data from S1 and the corresponding stage window length transmitted by S2. ,by To maintain a fixed step size, sliding segmentation is performed on the preprocessed four-dimensional time series data to obtain overlapping fixed-length subsequence fragments; In this step, the overlapping fixed-length subsequence segments are Given a fixed-length subsequence segment, the formula is: ; In the formula, The start time of valid data; The first result obtained after dividing the sliding window A multimodal data segment; This is the sequence number of the data segment; This represents the movement step size of the sliding window; The length of the sliding window; S4 receives the fixed-length subsequence segment from S3 and the parameters passed by S2; for any two subsequences Calculate the relative differences between each mode ( The formula is: ; ; ; ; In the formula, The relative differences under voltage modes; The relative differences under current modes; The relative differences under the charged state modes; The relative differences under temperature modes; For data segment The first in Each voltage value; For data segment The first in Each voltage value; For data segment The first in One current value; For data segment The first in One current value; For data segment The first in One SOC value; The table contains data segments. The first in One SOC value; For data segment The first in A temperature value; For data segment The first in A temperature value; Through the weighting coefficients of S2 Calculate the overall similarity score The formula is: ; In the formula, These are the weighting coefficients for the voltage modes; These are the weighting coefficients for the current modes; These are the weighting coefficients for the SOC mode; These are the weighting coefficients for the temperature modes; like Then determine the data segment and Similarity, obtain a set of similar subsequences; S5: Receive all subsequence segments from S3, the set of similar subsequences from S4, and the frequency threshold of frequent patterns transmitted from S2; use a greedy clustering algorithm based on Euclidean distance to group the data segments in all data segments of S3 that satisfy the similarity condition of S4 into the same pattern group. Count the number of data segments contained in each mode group Only when Only when the number of such patterns does the pattern group become a "frequent pattern" and the infrequent pattern groups that do not reach the threshold are removed. For each retained frequent pattern group, calculate the average value of all data segments within the group at each data point, and generate a standardized representative pattern sequence, represented as: ; In the formula, A standardized representative pattern sequence; This is the average sequence of voltage samples across all data segments within this mode group; This is the average sequence of current sample values for all data segments within this mode group; This is the sequence of average state-of-charge samples for all data segments within this mode group; This is the mean sequence of temperature samples across all data segments within this mode group; Construct a hierarchical dictionary consisting of a basic dictionary, a vehicle-specific dictionary, and a conversational ad-hoc dictionary. Specifically: Basic Dictionary : Stores a universal high-frequency mode that spans all charging processes and adapts to multiple vehicle models, supporting global data compression and reuse; Vehicle-Specific Dictionary After identifying the vehicle model based on charging characteristics, a unique pattern that is strongly correlated with the battery characteristics of that vehicle model is stored to improve the data compression accuracy of specific vehicle models. Temporary Conversation Dictionary Dynamically created during a single charge, specifically capturing the transient patterns unique to this charge, adapting to the sudden data characteristics during the charging process; S6. Receive the preprocessed four-dimensional time series data from S1, the hierarchical dictionary from S5, and the residual data points from S4. Traverse all data segments in the preprocessed four-dimensional time series data in chronological order and match and verify them with the patterns in the dictionary in turn. If the data segment and the pattern in the dictionary meet the similarity condition, then replace it with the corresponding short token: Basic Dictionary The pattern is identified by “B + three digits” (such as “B001”, “B002”). Vehicle-Specific Dictionary The pattern is identified by “V + three digits” (such as “V001”, “V002”). Temporary Conversation Dictionary The pattern is identified by “S + three digits” (such as “S001”, “S002”), and the token is stored using 8-bit binary encoding (corresponding to the index mapping of the character format).
[0020] For residual data points that are not covered by any dictionary pattern, no quantization is performed; their original values are directly retained. Following the rule of replacing matching data segments with tokens and preserving residual data of original values, the data segments are sequentially concatenated to form a hybrid compressed sequence. The format is as follows: ; In the formula, Basic / Model Type Token This is a temporary mode token for the session, followed by the original residual value; S7. Receive the hybrid compressed sequence generated by S6, the hierarchical dictionary from S5, and newly added representative pattern features; the charging pile edge side fully executes the real-time compression process of S1-S6, where the weighting coefficients of S2 are executed according to the preferred configuration; compressed representation. hierarchical dictionary Together, they constitute the final compressed output, which is directly used for edge storage and transmission and interaction with the cloud; The edge side will compress the output. and features of the new modes Uploaded synchronously to the cloud; The cloud aggregates compressed data and pattern features from multiple edge nodes, performs global frequent pattern mining, and optimizes the basic dictionary. Vehicle-specific dictionary Make optimizations and adjustments; The cloud generates a dictionary differential update package (containing only the differences between the old and new dictionaries) and distributes it to each edge node. After receiving the package, the edge node updates its local hierarchical dictionary, completing the closed loop of edge compression, cloud optimization and edge update. When compressed data is required, it is based on the updated hierarchical dictionary. Compress the sequence The token sequence in the data is restored to the corresponding multimodal data sequence, and then concatenated with the residual values in the sequence in chronological order to completely reconstruct the original charging time series data.
[0021] Therefore, this invention adopts the above-mentioned charging pile data compression method based on multimodal fusion and adaptive dynamic coding. This method uses four-dimensional data synchronization alignment and weighted similarity measurement to deeply fuse multimodal association information to accurately eliminate redundancy; it identifies the charging stage based on voltage and current derivative features, and coordinates the switching of core parameters to adapt to dynamic characteristics and vehicle differences; it relies on hierarchical dictionary and greedy clustering to extract frequent patterns, taking into account both general reuse and transient capture to reduce reconstruction distortion; and it reduces edge-side overhead and improves system scalability through edge-cloud collaborative optimization.
[0022] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0023] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A charging pile data compression method based on multimodal fusion and adaptive dynamic coding, characterized in that, Includes the following steps: S1. Collect four-dimensional multimodal time series data of charging piles, perform timestamp alignment and outlier removal, and obtain preprocessed four-dimensional time series data; S2. Calculate derivative features based on the preprocessed four-dimensional time series data in S1, identify constant current, constant voltage and transition stages, and dynamically configure window length, similarity threshold, frequent pattern frequency threshold and modal weight coefficient. S3. Using the window length configured in S2, perform sliding window segmentation on the preprocessed four-dimensional time series data obtained in S1 with a fixed step size to generate subsequence fragments; S4. Based on the subsequence fragments obtained in S3, calculate the relative differences of each modality, obtain the comprehensive similarity score by configuring the modality weight coefficients, compare the comprehensive similarity score with the similarity threshold configured in S2, determine the similarity of data segments and mark the unmatched residual data; obtain the similarity judgment result and the original value of the residual data. S5. Using a greedy clustering algorithm, based on the similarity judgment results of S4, the subsequence segments in S3 that meet the similarity conditions are merged. Combined with the frequency threshold configured in S2, frequent patterns are filtered and representative pattern sequences are generated. A hierarchical dictionary containing a basic dictionary, a vehicle-specific dictionary, and a conversational temporary dictionary is constructed. At the same time, new pattern features are extracted. S6. Using the hierarchical dictionary constructed in S5, tokenization replacement is performed on the successfully matched subsequence segments in S3, retaining the original values of the residual data marked in S4, and generating a hybrid compressed sequence. S7: The edge side performs real-time compression based on S1-S6, and uploads the hybrid compressed sequence obtained by S6 and the new pattern features extracted by S5 to the cloud; the cloud aggregates data to optimize and update the basic dictionary and the vehicle-specific dictionary and generates a differential update package, which is then distributed to each edge node to complete the dictionary closed-loop update. During data reconstruction, tokens in the mixed compressed sequence are restored into a multimodal data sequence based on the updated hierarchical dictionary, and then concatenated with the residual data in chronological order to recover the original data.
2. The charging pile data compression method based on multimodal fusion and adaptive dynamic coding according to claim 1, characterized in that, In S1, the four-dimensional time-series data is the battery pack voltage. Charging current State of charge and battery pack temperature The data format is .
3. The charging pile data compression method based on multimodal fusion and adaptive dynamic coding according to claim 2, characterized in that, In S2, the derivative characteristics are calculated using the central difference method, and the formula is: ； ； In the formula, for The battery pack voltage values collected at all times; for The battery pack voltage values collected at all times; for The charging current value collected at all times; for The charging current value collected at all times; The sampling interval is... ; During the charging phase, the derivative threshold is used. determination: and This is the constant current stage; and This is the constant pressure stage; The others are transitional phases.
4. The charging pile data compression method based on multimodal fusion and adaptive dynamic coding according to claim 3, characterized in that, In S2, the modal weighting coefficients are configured as follows: Constant current stage: ; Constant pressure stage: ; Transition phase: .
5. The charging pile data compression method based on multimodal fusion and adaptive dynamic coding according to claim 4, characterized in that, In S3, the fixed step size of the sliding window is 1 / 2 of the window length. The formula for calculating each data segment is: ； In the formula, The start time of valid data; The first result obtained after dividing the sliding window A multimodal data segment; This is the sequence number of the data segment; This represents the movement step size of the sliding window; The length of the sliding window.
6. The charging pile data compression method based on multimodal fusion and adaptive dynamic coding according to claim 5, characterized in that, In S4, the formula for calculating the relative difference between single modes is: ； ； ； ； In the formula, The relative differences under voltage modes; The relative differences under current modes; The relative differences under the charged state modes; The relative differences under temperature modes; For data segment The first in Each voltage value; For data segment The first in Each voltage value; For data segment The first in One current value; For data segment The first in One current value; For data segment The first in One SOC value; The table is a data segment. The first in One SOC value; For data segment The first in A temperature value; For data segment The first in A temperature value.
7. The charging pile data compression method based on multimodal fusion and adaptive dynamic coding according to claim 6, characterized in that, In S4, the formula for calculating the overall similarity score is: ； In the formula, These are the weighting coefficients for the voltage modes; These are the weighting coefficients for the current modes; These are the weighting coefficients for the SOC mode; These are the weighting coefficients for the temperature modes; when The similarity of data segments is determined at that time.
8. The charging pile data compression method based on multimodal fusion and adaptive dynamic coding according to claim 7, characterized in that, In S5, greedy clustering uses Euclidean distance for determination; frequent clustering requires satisfying the requirement of a certain number of data segments within a group. Frequency threshold, a representative pattern sequence is represented as: ； In the formula, A standardized representative pattern sequence; This is the average sequence of voltage samples across all data segments within this mode group; This is the average sequence of current sample values for all data segments within this mode group; This is the sequence of average state-of-charge samples for all data segments within this mode group; This is the mean sequence of temperature samples for all data segments within this mode group.
9. A computer device, characterized in that, include: A processor configured to be coupled to memory, read and execute instructions and / or program code in the memory to perform the method as described in any one of claims 1-8.
10. A computer-readable medium, characterized in that, The computer-readable medium stores computer program code that, when executed on a computer, causes the computer to perform the method as described in any one of claims 1-8.