Edge device, cloud server, battery data transmission system

Abstract

The application discloses an edge device, a cloud server and a battery data transmission system, and relates to the technical field of data transmission.The edge device comprises an acquisition module, a lightweight small model module, an updating module and a low-frequency reporting module.The acquisition module is used to acquire battery data at a first preset frequency.The lightweight small model module is in communication connection with the acquisition module, and is used to compress the battery data acquired by the acquisition module for a preset number of times to obtain a compressed file when the number of times of acquiring battery data by the acquisition module reaches the preset number.The updating module is in communication connection with the acquisition module and the lightweight small model module, and is used to update the lightweight small model module according to the battery data.The low-frequency reporting module is in communication connection with the lightweight small model module, and is used to send the compressed file to the cloud server.Thus, the battery data can be transmitted to the cloud server while taking into account low data volume and high accuracy.

Description

Technical Field

[0001] This invention relates to the field of data transmission technology, and in particular to an edge device, cloud server, and battery data transmission system. Background Technology

[0002] In related technologies, to meet the functional safety (ISO 26262) and national standards requirements for SOC ±3% and SOH ±5% accuracy, advanced cloud-based algorithms must rely on high-frequency battery cell voltage and temperature sequences with a time interval of ≤1 second. However, if the edge device reports battery data to the cloud completely at a time interval of ≤1 second, the daily traffic of a single 160-cell battery pack reaches 300-400MB, and the annual traffic exceeds 100GB. The communication costs of 4G / 5G / IoT cards are high, and the bandwidth of the edge device simply cannot bear the load. If the frequency of reporting battery data to the cloud is directly reduced to a time interval of 5-30 seconds, and only statistical values ​​or extreme values ​​are sent, high-frequency dynamic details are severely lost, causing the cloud-based SOH estimation error to deteriorate from ±2.6% to over ±8%, and the sensitivity of early warning of thermal runaway to decrease significantly, which cannot meet the actual engineering requirements. Summary of the Invention

[0003] This invention aims to at least partially address one of the technical problems in related technologies. Therefore, a first objective of this invention is to provide an edge device that achieves data transmission that balances low data volume with high precision.

[0004] The second objective of this invention is to provide a cloud server.

[0005] The third objective of this invention is to provide a battery data transmission system.

[0006] To achieve the above objectives, a first aspect of the present invention provides an edge device, comprising: an acquisition module configured to acquire battery data at a first preset frequency; a lightweight mini-model module communicatively connected to the acquisition module, the lightweight mini-model module being configured to compress the battery data acquired by the acquisition module in the most recent preset number of acquisitions to obtain a compressed file when the number of times the acquisition module acquires the battery data reaches a preset number; an update module communicatively connected to the acquisition module and the lightweight mini-model module, the update module being configured to update the lightweight mini-model module according to the battery data; and a low-frequency reporting module communicatively connected to the lightweight mini-model module, the low-frequency reporting module being configured to send the compressed file to a cloud server.

[0007] In addition, the edge device according to embodiments of the present invention may also have the following additional technical features: According to one embodiment of the present invention, the lightweight mini-model module includes a lightweight mini-model, the lightweight mini-model includes a calibration head, and the update module is specifically used to: periodically update the calibration head according to a first target battery data and a second target battery data at a second preset frequency; wherein, the first target battery data is the battery data acquired by the acquisition module within a first target time period, the second target battery data is the battery data acquired by the acquisition module within a second target time period, the end time of the second target time period is earlier than the start time of the first target time period, the time difference between the start time of the second target time period and the start time of the first target time period is less than a preset time difference threshold, and for each update, the first target time period corresponding to the update is a time period with the update start time as the end time and a duration of a preset duration.

[0008] According to one embodiment of the present invention, there are multiple second target time periods, the duration of the second target time period is equal to the duration of the first target time period, and for any two second target time periods, the end time of one second target time period is earlier than or equal to the start time of the other second target time period.

[0009] According to an embodiment of the present invention, the updating module is further configured to: encode the first target battery data to obtain a first feature vector, and for each second target time period, encode the second target battery data corresponding to the second target time period to obtain a second feature vector, and obtain a first loss based on the first feature vector and the second feature vector, and update the calibration head based on the first loss.

[0010] According to one embodiment of the present invention, the updating module is further configured to: input the first target battery data and the second target battery data into the lightweight small model module, obtain the battery data extreme value prediction value within the first target time period output by the lightweight small model module based on the first target battery data and the second target battery data, obtain a second loss based on the battery data extreme value prediction value and the first target battery data, and update the calibration head based on the second loss.

[0011] According to one embodiment of the present invention, the edge device further includes: a cloud receiving module, which is communicatively connected to the lightweight small model module. The cloud receiving module is used to receive small model parameters sent by the cloud server and update the lightweight small model module according to the small model parameters.

[0012] According to one embodiment of the present invention, the battery data includes at least one of battery voltage and battery temperature.

[0013] An edge device according to an embodiment of the present invention includes: an acquisition module, configured to acquire battery data at a first preset frequency; a lightweight mini-model module, communicatively connected to the acquisition module, configured to compress the battery data acquired by the acquisition module in the most recent preset number of acquisitions to obtain a compressed file when the number of times the acquisition module acquires battery data reaches a preset number; an update module, communicatively connected to the acquisition module and the lightweight mini-model module, configured to update the lightweight mini-model module according to the battery data; and a low-frequency reporting module, communicatively connected to the lightweight mini-model module, configured to send the compressed file to a cloud server. Thus, it is possible to achieve both low data volume and high accuracy in transmitting battery data to the cloud server.

[0014] To achieve the above objectives, a second aspect of the present invention provides a cloud server, comprising: a receiving module, which is communicatively connected to a large model module, the receiving module being used to receive compressed files and send the compressed files received at multiple times to the large model module, wherein the compressed files are files sent by the aforementioned edge device; and a large model module, which is used to perform restoration processing on the compressed files to obtain battery data.

[0015] In addition, the cloud server according to embodiments of the present invention may also have the following additional technical features: According to one embodiment of the present invention, the receiving module is further configured to: generate a receiving time of the compressed file when the compressed file is received; the large model module is further configured to: perform large model distillation training at a third preset frequency based on the compressed file whose receiving time is within a third target time period to obtain small model parameters, wherein the end time of the third target time period is equal to the start time of the large model distillation training; the cloud server further includes: a sending module, the sending module being communicatively connected to the large model module, the sending module being configured to send the small model parameters to the edge device.

[0016] According to one embodiment of the present invention, the duration of the third target time period is equal to the reciprocal of the third preset frequency.

[0017] According to an embodiment of the present invention, the cloud server can decompress the compressed file sent by the edge device in the above embodiment to obtain battery data, thereby realizing battery data reception.

[0018] To achieve the above objectives, a third aspect of the present invention provides a battery data transmission system, including the aforementioned edge device and the aforementioned cloud server, wherein the edge device is communicatively connected to the cloud server.

[0019] The battery data transmission system according to embodiments of the present invention can achieve data transmission that balances low data volume and high precision.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] Figure 1 This is a structural block diagram of an edge device according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the operation of an edge device according to an example of the present invention; Figure 3 This is a structural block diagram of a cloud server according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the operation of a cloud server, as an example of the present invention. Figure 5 This is a structural block diagram of the battery data transmission system according to an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the operation of a battery data transmission system according to an example of the present invention. Detailed Implementation

[0022] The edge device, cloud server, and battery data transmission system of embodiments of the present invention are described below with reference to the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described with reference to the accompanying drawings are exemplary and should not be construed as limiting the present invention.

[0023] Figure 1 This is a structural block diagram of an edge device according to an embodiment of the present invention.

[0024] like Figure 1 As shown, the edge device 100 includes: an acquisition module 101, which is used to acquire battery data at a first preset frequency; a lightweight mini-model module 102, which is communicatively connected to the acquisition module 101 and is used to compress the battery data acquired by the acquisition module 101 in the most recent preset number of acquisitions to obtain a compressed file when the number of times the acquisition module 101 acquires battery data reaches a preset number; an update module 103, which is communicatively connected to the acquisition module 101 and the lightweight mini-model module 102 and is used to update the lightweight mini-model module 102 according to the battery data; and a low-frequency reporting module 104, which is communicatively connected to the lightweight mini-model module 102 and is used to send the compressed file to a cloud server.

[0025] The aforementioned battery data includes at least one of battery voltage and battery temperature.

[0026] The aforementioned edge device 100 can be a device with a dual-core heterogeneous architecture consisting of two processor cores, and the two processor cores can be an A core and an R core, respectively. The aforementioned acquisition module 101 can be set in the R core, and the aforementioned lightweight small model module 102, update module 103, and low-frequency reporting module 104 can be set in the A core.

[0027] When the acquisition module 101 in the R core acquires battery data, it can send the acquired battery data to the A core via a bus or other possible means, that is, to the lightweight mini-model module 102 and the update module 103. The lightweight mini-model module 102 can determine whether the number of times the acquisition module 101 has acquired battery data has reached a preset number, and if so, compress the battery data acquired by the acquisition module 101 in the most recent preset number of acquisitions. The update module 103 can update the lightweight mini-model module 102 according to the received battery data based on a pre-set update rule.

[0028] The aforementioned first preset frequency can be set such that the time difference between two consecutive times the acquisition module 101 acquires battery data is less than or equal to 1 second.

[0029] The aforementioned lightweight small model module 102 includes a lightweight small model, which may include an input layer, a backbone encoder, a time pooling layer, and a calibration head.

[0030] The above-mentioned compression process for battery data to obtain a compressed file involves performing feature extraction on the battery data to obtain feature vectors, then obtaining a feature file based on the feature vectors, and finally using the feature file as the compressed file.

[0031] The above lightweight small model is illustrated below with a specific embodiment.

[0032] In this specific embodiment, the aforementioned lightweight small model includes an input layer, a backbone encoder, a Temporal Pooling layer, and a calibration head.

[0033] The first preset frequency is set such that the time difference between two consecutive times the acquisition module 101 acquires battery data is 0.5s.

[0034] The battery data mentioned above is a complete original message containing all 160 cell voltages, 160 cell temperatures, and 92 key scalar fields. That is, the acquisition module 101 acquires a complete original message containing all 160 cell voltages, 160 cell temperatures, and 92 key scalar fields every 0.5 seconds.

[0035] The preset number of times can be set to 20.

[0036] The input layer mentioned above may include a sliding window, through which the battery data acquired by the acquisition module 101 in the last 20 times can be obtained.

[0037] The dimensions of the input layer mentioned above are 20×412.

[0038] The aforementioned backbone encoder comprises three 1D convolutional layers and two TTM (Tiny Time-Mixer) layers. The three 1D convolutional layers reduce the dimensionality of the input data in the input layer in the order of 412→256→128→64. The two TTM layers have four mixing heads, and the hidden layers in the feedforward network have a dimension of 128.

[0039] The aforementioned time pooling layer is a concatenation of global averaging and max pooling, and the output data of the time pooling layer is set to 128 dimensions.

[0040] The processing of the data output from the temporal pooling layer by the aforementioned calibration head can include first upscaling the 128-dimensional input from the temporal pooling layer to 256 dimensions through a fully connected layer with GeLU activation, then normalizing it using L2 normalization, and finally compressing it into an 8-bit integer. This results in an output 256-dimensional quantized feature vector (256 bytes in 8-bit format, 128 bytes in 4-bit highly compressed format). Furthermore, the fully connected layer filters the information input to the calibration head based on importance, and the output data is in the required format.

[0041] Therefore, by setting up a lightweight small model module 102, the battery data acquired by the acquisition module 101 a preset number of times is compressed to obtain a compressed file, which is then sent to the cloud server. The update module 103 is set to continuously update the lightweight small model module 102 based on the battery data, thereby enabling the transmission of comprehensive and accurate battery data with a low data volume.

[0042] In some embodiments of the present invention, the update module 103 is specifically used to: periodically update the calibration head according to the first target battery data and the second target battery data at a second preset frequency; wherein, the first target battery data is the battery data acquired by the acquisition module 101 within a first target time period, the second target battery data is the battery data acquired by the acquisition module 101 within a second target time period, the number of second target time periods is multiple, the duration of the second target time period is equal to the duration of the first target time period, the end time of the second target time period is earlier than the start time of the first target time period, the time difference between the start time of the second target time period and the start time of the first target time period is less than a preset time difference threshold, for any two second target time periods, the end time of one second target time period is earlier than or equal to the start time of the other second target time period, and for each update, the first target time period corresponding to the update is a time period with the start time of the update as the end time and a duration of a preset duration.

[0043] The following description uses a specific example.

[0044] In this specific embodiment, the second preset frequency can be set such that the update module 103 updates the calibration head once every 10 seconds, the duration of the first target time period and the second target time period is set to 10 seconds, the preset time difference threshold is set to 1 hour, and the number of the second target time periods is 8.

[0045] At this time, the update module 103 will acquire battery data within the past 10 seconds every 10 seconds as the first target battery data, and acquire battery data within 8 random 10-second intervals within the past hour as the second target battery data. Then, the calibration head of the lightweight small model will be trained based on the first target battery data and the second target battery data to update the calibration head.

[0046] The first target battery data mentioned above is used as a positive sample when training the calibration head, and the second target battery data mentioned above is used as a negative sample when training the calibration head.

[0047] Taking the time difference between two consecutive battery data acquisitions by the acquisition module 101 as an example, where the time difference is 0.5s.

[0048] Positive samples: The current 10-second window being processed (≈20 high-frequency data points of 0.5s) is treated as a whole as a positive example. This means "This is the current real battery state, and I want the lightweight small model to remember its characteristic appearance".

[0049] Negative samples: Randomly select 8 different 10-second windows from the historical 1-hour cache. This means "These are states from other times in the past, which are different from the present. I want to push their features away with the lightweight small model."

[0050] In other words, positive samples represent 'now' and negative samples represent 'not now'. The training objective is to make the feature vectors bring the current state closer to themselves and push away historical states in the embedding space, thereby achieving strong temporal discriminative power. This is particularly useful for BMS: battery aging / different operating conditions can cause voltage and temperature pattern drift. If the features do not distinguish between time, cloud reconstruction can easily confuse old and new states, increasing the error.

[0051] To achieve this effect, the update module 103 is further configured to: encode the first target battery data to obtain a first feature vector, and for each second target time period, encode the second target battery data corresponding to the second target time period to obtain a second feature vector, and obtain a first loss based on the first feature vector and the second feature vector, and update the calibration head based on the first loss.

[0052] Specifically, the positive samples can be encoded to obtain a 256-dimensional feature vector e_pos, and the eight negative samples can be encoded separately to obtain e_neg_1~e_neg_8.

[0053] The update module 103 obtains the first loss using the following formula: L=-log[exp(sim(e_pos,e_pos) / τ) / (exp(sim(e_pos,e_pos) / τ)+Σexp(sim(e_pos,e_neg_i) / τ))], Where L is the first loss, e_pos is the first feature vector, e_neg_i is the i-th second feature vector, i is a positive integer greater than 0 and less than or equal to N, N is the number of second feature vectors, τ is the temperature parameter, and sim() refers to the cosine similarity.

[0054] Therefore, we can derive the first loss, which is the NT-Xent contrast loss. Specifically, based on the aforementioned positive and negative sample pairs, it narrows the similarity of positive sample features and widens the similarity of negative samples. The goal is to teach the calibration head to 'learn the unique characteristics of the current battery state and avoid confusion with the past'. This solves the problem of distribution drift during long-term battery operation.

[0055] Furthermore, the update module 103 is also used to: input the first target battery data and the second target battery data into the lightweight small model module 102, obtain the battery data extreme value prediction value within the first target time period output by the lightweight small model module 102 based on the first target battery data and the second target battery data, obtain the second loss based on the battery data extreme value prediction value and the first target battery data, and update the calibration head based on the second loss.

[0056] The second loss obtained above based on the predicted extreme value of battery data and the first target battery data can be obtained by obtaining the actual extreme value of battery data within the first target time period based on the first target battery data, and then obtaining the second loss based on the actual extreme value of battery data and the predicted extreme value of battery data.

[0057] Furthermore, the lightweight small model module 102 can be configured to output battery data extreme value prediction values ​​corresponding to the first target battery data and the second target battery data. Taking the number of the second target time periods as 8 as an example, the lightweight small model module 102 needs to output one battery data extreme value prediction value corresponding to one first target time period and eight battery data extreme value prediction values ​​corresponding to eight second target time periods. Then, based on the nine battery data extreme value prediction values ​​and the actual battery data extreme values, the mean square error is obtained, which is the average of the squares of the differences between the predicted value and the true value. This mean square error is then used as the second error.

[0058] Therefore, by training the calibration head using the aforementioned first and second errors, the first error enables the calibration head to have strong temporal discrimination capabilities, while the second error allows the calibration head to retain the most critical anomaly / boundary information for the cloud algorithm. This allows the lightweight small model, fine-tuned daily, to generate compressed files that can still support the reconstruction of high-frequency sequences with millivolt-level accuracy from large models on the cloud server, even with extremely low bandwidth (1.38KB / 10s). Moreover, by updating only the calibration head in the lightweight small model, only fine-tuning of the lightweight small model is needed to ensure that the data output by the lightweight small model can meet the requirements for sending comprehensive and accurate battery data under low data volume conditions.

[0059] The following example will illustrate this point.

[0060] See in this example. Figure 2 The lightweight small model is 310KB in total, with a 45K calibration head that can be adjusted, and the remaining 265K data is frozen.

[0061] Figure 2 The 1D-Conv in the above refers to the one-dimensional convolutional layer. Figure 2 The phrase "Avg+Max concatenation -> 128 dimensions" means that the time pooling layer is a concatenation of global average and max pooling, and the dimension of the output data of the time pooling layer is set to 128 dimensions. Figure 2 The adaptive calibration head mentioned above is the same as the one described in the text. Figure 2 The FC+GeLU 128->256 in the text refers to the calibration head upscaling the input 128-dimensional data to 256 dimensions through a fully connected layer with GeLU activation. Figure 2The L2 normalization + 8-bit quantization in the calibration head refers to normalizing the data that has been upgraded to 256 dimensions using L2 normalization and finally compressing it into an 8-bit integer. Figure 2 The "Freeze" and "Freeze 265K" parameters indicate that the components indicated by the corresponding arrows are not fine-tuned based on battery data. Figure 2 The phrase "only fine-tuning the 45K parameter" means that the component indicated by the corresponding arrow needs to be fine-tuned based on the battery data.

[0062] As can be seen, the lightweight small model in the lightweight small model module 102 processes the battery data obtained by the acquisition module 10120 times to obtain a 256-dimensional feature vector.

[0063] After obtaining the 256-dimensional feature vector, the lightweight small model can generate a 1380-byte compressed file based on this 256-dimensional feature vector. The structure of the compressed file can be seen in Table 1 below.

[0064] Table 1

[0065] Optionally, the structure of the compressed file can be adjusted according to actual needs, that is, the compressed file can be less than or equal to 1380 bytes.

[0066] The aforementioned fourth-order statistical moments include the mean, variance, skewness, and kurtosis of all 412 original channels (160 individual cell voltages, 160 individual cell temperatures, and 92 key scalars). This supplements the precise global distribution information that might be lost in the eigenvectors, particularly providing a mathematical description of the distribution tails (extreme values / abnormal clusters). This enables large models on cloud servers to accurately capture the statistical regularities and long-term evolution characteristics of battery states during time-series reconstruction, avoiding overall distribution drift or loss of anomalous information in the reconstructed sequence.

[0067] The aforementioned mutation cell mask is a 160-bit bitmap, with each bit corresponding to a single cell. A value of 1 indicates that the cell has undergone a mutation within the current time window (voltage change > 50mV or temperature change > 2℃ and lasting ≥ 3 frames). Using this mask, the cloud-based large model can selectively enhance the attention weight and reconstruction accuracy of abnormal cells during the reconstruction process, achieving adaptive fusion of millivolt-level reconstruction under normal operating conditions and higher accuracy under abnormal operating conditions, while seamlessly integrating with the mechanism for retroactively sending original messages during abnormal operations.

[0068] The sub-version + CRC32 field is used to identify the sub-iteration version of the compressed file format, supporting future feature expansions (such as quantization bit width adjustment or adding statistical items) without changing the total length; the CRC32 field provides 32-bit cyclic redundancy check for the entire compressed file, ensuring the integrity and tamper-proof capability of low-frequency reported data during transmission, meeting the ISO 26262 functional safety and industrial-grade reliability requirements.

[0069] The aforementioned reserved extension area is for future expansion space, supporting the addition of fields in subsequent iterations (such as additional statistical features, security signatures, operating condition tags, or encrypted information), maintaining a fixed length of compressed files, and facilitating communication protocol stack compatibility and long-term system evolution.

[0070] Furthermore, an online fine-tuning process is set up, meaning that after the lightweight small model module 102 obtains the compressed file based on the battery data, the update module 103 still needs to continue caching the battery data.

[0071] Furthermore, the update module 103 can be configured to cache the most recent 20 data entries after the acquisition module 101 performs 20 acquisitions, which are the battery data obtained by the acquisition module 101 in the most recent 20 acquisitions.

[0072] Furthermore, after acquiring battery data, the update module 103 also needs to maintain the normalized statistics (mean and variance) of the 412 channels for the most recent hour in real time, and perform Z-score normalization on the input.

[0073] Furthermore, using the current 10-second window as the positive sample, eight negative samples are randomly sampled from the historical 1-hour cache.

[0074] Afterwards, the update module 103 can keep the trunk encoder frozen and only unfreeze the 45K calibration head.

[0075] Then, the NT-Xent+ weakly supervised MSE loss is obtained, which is the first and second loss mentioned above.

[0076] After obtaining the first and second losses, the AdamW optimizer (learning rate 1e-4), which has extremely short training time and extremely low RAM peak, can be used to obtain the latest compressed file based on the first and second losses, and then the calibration head can be updated based on the latest compressed file.

[0077] In some embodiments of the present invention, the edge device 100 further includes: a cloud receiving module, which is communicatively connected to the lightweight small model module 102. The cloud receiving module is used to receive small model parameters sent by the cloud server and update the lightweight small model module 102 according to the small model parameters.

[0078] In some embodiments of the present invention, after receiving the small model parameters, the cloud receiving module first performs a preliminary verification of the small model parameters, and then sends the verified small model parameters to the R core, which verifies the small model parameters. After the R core verifies the small model parameters, the cloud receiving module updates the lightweight small model module 102 according to the small model parameters.

[0079] The R core verifies the small model parameters. It can first perform signature verification on the R core, and then write the small model parameters to the B partition allocated to the B core on the Flash storage chip after the verification is passed, thereby updating the dual backup boot flag and triggering a reset.

[0080] After power-on, the bootloader embedded in the R core verifies the signature of the B partition. If the verification result is valid, it jumps to the B partition to run the new version, that is, the small model parameter verification passes. If the verification result is invalid, it rolls back, that is, the small model parameter verification fails.

[0081] Furthermore, the A core can also report the update results to the cloud server. If the cloud server is connected to multiple edge devices 100, and if the report result of a certain edge device 100 is a rollback heartbeat, the cloud server can retry updating the lightweight small model in the edge device 100 after a preset time (such as the next day).

[0082] In some embodiments of the present invention, when the low-frequency reporting module 104 sends the compressed file to the cloud server, it can also obtain the rate of change of battery data. If the rate of change of battery data is greater than a preset rate of change threshold within a certain period of time, the battery data within that period of time is sent to the cloud server.

[0083] For example, assuming that based on battery data, it is known that the voltage change of any single cell is greater than 50mV or the temperature change is greater than 2℃ within ≤1s and lasts for more than 3 frames, the low-frequency reporting is immediately interrupted and a complete original message (approximately 2.19KB) is forcibly sent within ≤1s to ensure zero loss under extreme operating conditions.

[0084] In other words, it can be configured that after the acquisition module 101 acquires battery data each time, the low-frequency reporting module 104 compares the currently acquired battery data with the battery data acquired by the acquisition module 101 1 second ago. If the voltage change is >50mV or the temperature change is >2℃, the battery data within this 1 second will be sent to the cloud server.

[0085] In summary, the edge device 100 of this embodiment includes: an acquisition module 101, which is used to acquire battery data at a first preset frequency; a lightweight mini-model module 102, which is communicatively connected to the acquisition module 101, and is used to compress the battery data acquired by the acquisition module 101 in the most recent preset number of acquisitions to obtain a compressed file when the number of times the acquisition module 101 acquires battery data reaches a preset number; an update module 103, which is communicatively connected to the acquisition module 101 and the lightweight mini-model module 102, and is used to update the lightweight mini-model module 102 according to the battery data; and a low-frequency reporting module 104, which is communicatively connected to the lightweight mini-model module 102, and is used to send the compressed file to a cloud server. Therefore, by setting up a lightweight small model module 102, the battery data acquired by the acquisition module 101 a preset number of times is compressed to obtain a compressed file, which is then sent to the cloud server. The update module 103 is set to continuously update the lightweight small model module 102 based on the battery data, thereby enabling the transmission of battery data that meets the requirements under the premise of low data volume.

[0086] Furthermore, this invention proposes a cloud server.

[0087] Figure 3 This is a structural block diagram of a cloud server according to an embodiment of the present invention.

[0088] like Figure 3 As shown, the cloud server 200 includes: a receiving module 201, which is communicatively connected to a large model module 202. The receiving module 201 is used to receive compressed files and send the compressed files received at multiple times to the large model module 202. The compressed files are files sent by the edge device 100 as described above. The large model module 202 is used to perform restoration processing on the compressed files to obtain battery data.

[0089] The above-mentioned compressed files received at multiple times are sent to the large model module 202 together. For each compressed file received, multiple compressed files whose receiving time is closest to the current time are selected from the historically received compressed files, so that the compressed file received at the current time is sent together with the historically received compressed files to the large model module 202.

[0090] Optionally, when the number of received compressed files reaches a preset number, the receiving module 201 may send the preset number of compressed files to the large model module 202 together.

[0091] The range of values ​​for the compressed file sent to the large model module 202 at one time can be set to 3~10.

[0092] Therefore, by setting the receiving module 201 to send the compressed files to the large model module 202 when the number of received compressed files reaches a preset threshold, and the large model module 202 to obtain battery data based on the compressed files, the significant temporal inertia and thermo / electrochemical coupling effect of the dynamic evolution of battery cell voltage and temperature can be avoided. A single 10-second compressed file can only characterize local transient information, resulting in insufficient accuracy of the acquired battery data. Setting the large model module 202 to send multiple consecutive compressed files can provide the model with sufficient cross-cycle context, enabling it to accurately capture long-term temporal patterns such as voltage polarization, thermal diffusion, and aging trends, thereby achieving high-precision reconstruction.

[0093] Furthermore, when the receiving module 201 receives the compressed file, it can also obtain the current low-frequency scalar in the compressed file. The current low-frequency scalar refers to a value in the battery data that does not change much, such as a part of the above 92 key scalar fields, including total power, insulation resistance, charging and discharging status, and alarm flags. That is, the receiving module 201 can extract the current low-frequency scalar from the key scalar fields in the compressed file, and then send the current low-frequency scalar along with the compressed file to the large model module 202.

[0094] In some embodiments of the present invention, the receiving module 201 is further configured to: generate the receiving time of the compressed file when the compressed file is received; the large model module 202 is further configured to: perform large model distillation training at a third preset frequency based on the compressed file whose receiving time is within a third target time period to obtain small model parameters, wherein the end time of the third target time period is equal to the start time of the large model distillation training; the cloud server 200 further includes: a sending module, which is communicatively connected to the large model module 202, and is configured to send the small model parameters to the edge device 100.

[0095] The aforementioned third preset frequency can be set such that the time interval between two large model distillation training operations performed by the large model module 202 is 24 hours, that is, the large model module 202 performs large model distillation training once every 24 hours.

[0096] Furthermore, the duration of the third target time period can be set to be equal to the reciprocal of the third preset frequency. In this case, assuming that the large model module 202 performs large model distillation training once every 24 hours, the data used by the large model module 202 each time it performs large model distillation training is the compressed file received by the receiving module 201 in the past 24 hours.

[0097] As an example, it can be set to upload approximately 1024 high-frequency samples (including real 0.5s label data) of 10s each day (the default value in this article) to the cloud based on edge devices at midnight. This will be used to perform a complete epoch distillation training on the large model in the large model module 202, generating the latest 310KB small model parameters. The sending module will then securely send the small model parameters to the edge device 100 via OTA.

[0098] This further ensures the accuracy of the battery data sent by the edge device 100.

[0099] In some embodiments of the present invention, after the large model module 202 obtains battery data from the compressed file, if the receiving module 201 receives battery data sent by the edge device 100 to the cloud server 200 because the rate of change of the battery data is greater than a preset rate of change threshold, the large model module 202 uses the battery data received by the receiving module 201 to overwrite the corresponding data in the battery data obtained from the compressed file.

[0100] The following explanation uses the example shown in Table 2. In Table 2, taking 10 seconds as an example, the acquisition module 101 in the edge device 100 performed 20 battery data acquisitions during those 10 seconds.

[0101] Table 2

[0102] The aforementioned original message refers to the battery data sent by the edge device 100 to the cloud server 200 because the rate of change of battery data exceeds a preset rate of change threshold. The phrase "serious sudden change" in Table 2 above indicates that the edge device 100 sent 0.5 seconds of battery data to the cloud server 200 because the rate of change of battery data exceeded the preset rate of change threshold.

[0103] The following is combined with Figure 4 The specific examples shown will illustrate this.

[0104] Specifically, first, the cloud server 200 receives the current 10-second low-frequency data packet (1380 bytes, including compressed files), then retrieves 3 to 10 consecutive historical compressed files (30 to 100 seconds of context) from the cache, and then extracts the current low-frequency scalar. Based on the current low-frequency scalar and the compressed files, a time-series input tensor containing a feature sequence and a scalar sequence is constructed. The feature sequence is a sequence composed of multiple compressed files, and the scalar sequence is a sequence composed of the current low-frequency scalars obtained from different compressed files.

[0105] Furthermore, the temporal input tensor is input into the large model in the large model module 202 to achieve forward inference. This large model is a hybrid large model of TFT (Temporal Fusion Transformer) + FNO (Fourier Neural Operator). The aforementioned TFT module captures long- and short-term temporal dependencies and inter-unit attention relationships, while the aforementioned FNO module efficiently models the periodic propagation characteristics of voltage / temperature in the frequency domain. The large model module 202 also includes a decoder, which outputs the complete high-frequency full sequence (160 individual unit voltages + 160 individual unit temperatures) of the time interval corresponding to the compressed file received by the large model module 202, with a duration of ≤1s.

[0106] The large model input consists of all 0.5s high-frequency sequences within the corresponding time interval, i.e., 160 voltage channels and 160 temperature channels updated at a frequency of 0.5s. The difference mentioned above corresponds to the time interval of the compressed file input to the large model.

[0107] If the interval contains original messages that were resent due to an anomaly, the battery data in the real original message will overwrite the original data restored by the large model with the corresponding timestamp, and then write it into the time series library for use by algorithms such as SOH / SOP / RUL.

[0108] In summary, the cloud server 200 of this embodiment can obtain accurate battery data based on the compressed file sent by the edge device 100 of the above embodiment.

[0109] Furthermore, this invention proposes a battery data transmission system.

[0110] Figure 5 This is a structural block diagram of the battery data transmission system according to an embodiment of the present invention.

[0111] like Figure 5 As shown, the battery data transmission system 10 includes the edge device 100 of the above embodiment and the cloud server 200 of the above embodiment, and the edge device 100 is communicatively connected to the cloud server 200.

[0112] The following is combined with Figure 6 The specific embodiments shown will be described in detail.

[0113] The edge device 100 at the edge has a heterogeneous structure of A core + R core, and the cloud includes a cluster of cloud servers 200. There are multiple edge devices 100. For example, for battery packs located in different locations, a corresponding edge device 100 can be set up. Each edge device 100 can be a device in the battery management system that has the function of sending battery data to the cloud server 200.

[0114] Edge device 100 communicates with cloud server 200 via a network, which can be 4G / 5G / NB-IoT. Cloud server 200 updates the lightweight mini-model in edge device 100 daily according to ASIL-D level, i.e., meeting ISO 26262 ASIL-D.

[0115] Specifically, the edge device 100 acquires raw battery data at high frequency. This raw battery data includes 160 channels of voltage, temperature, and scalar data, and the interval between two acquisitions is less than or equal to 1 second.

[0116] Subsequently, the lightweight mini-model in edge device 100 generates a 1380-byte compressed file based on the battery data. This lightweight mini-model is less than or equal to 500KB.

[0117] In addition, the edge device 100 also needs to perform anomaly detection to determine whether the rate of change of battery data is greater than a preset rate of change threshold.

[0118] Furthermore, the low-frequency reporting module 104 in the edge device 100 sends a compressed file to the cloud server 200 every 5 to 30 seconds.

[0119] After receiving the compressed file, the cloud server cluster of 200 caches the compressed file and then sends the compressed file received at multiple times to a large model with 180M parameters to achieve high-frequency reconstruction at the millivolt level in 0.5s, thereby obtaining the battery data. This completes the reception of battery data.

[0120] This enables the transmission of battery data from the edge to the cloud. By using the aforementioned edge-compression and cloud-reconstruction method, traffic compression of over 12 times (up to 17 times in actual testing) can be achieved, with the time interval between the cloud algorithm accuracy and the full set of adjacent battery data being ≤1 second and the reporting difference less than 0.2%. This solves the problem of the trade-off between traffic cost and algorithm accuracy.

[0121] After receiving the battery data, the cloud can use the received battery data to implement advanced algorithm clusters, including SOH / SOP / RUL / thermal runaway early warning.

[0122] In addition, the cloud needs to send OTA parameters daily, that is, send small model parameters to each edge device 100 so that each edge device 100 can update the small model daily.

[0123] In summary, the battery data transmission system 10 of this embodiment can achieve a balance between traffic cost and algorithm accuracy.

[0124] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein can be considered as a ordered list of executable instructions for implementing logical functions, which can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0125] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0126] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0127] In the description of this specification, the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and should not be construed as limiting the present invention.

[0128] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0129] In this specification, unless otherwise stated, the terms "installation," "connection," "joining," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly defined. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0130] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0131] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An edge device, characterized in that, include: The acquisition module is used to acquire battery data at a first preset frequency; A lightweight small model module is communicatively connected to the acquisition module. The lightweight small model module is used to compress the battery data acquired by the acquisition module in the most recent preset number of times when the number of times the acquisition module acquires the battery data reaches a preset number of times, so as to obtain a compressed file. An update module is communicatively connected to the acquisition module and the lightweight small model module, and the update module is used to update the lightweight small model module according to the battery data; A low-frequency reporting module is communicatively connected to the lightweight small model module, and the low-frequency reporting module is used to send the compressed file to the cloud server.

2. The edge device according to claim 1, characterized in that, The lightweight small model module includes a lightweight small model, which in turn includes a calibration head. The update module is specifically used for: The calibration head is periodically updated at a second preset frequency based on first target battery data and second target battery data. The first target battery data is the battery data acquired by the acquisition module within a first target time period, and the second target battery data is the battery data acquired by the acquisition module within a second target time period. The end time of the second target time period is earlier than the start time of the first target time period, and the time difference between the start time of the second target time period and the start time of the first target time period is less than a preset time difference threshold. For each update, the corresponding first target time period is a time period with the update start time as the end time and a preset duration.

3. The edge device according to claim 2, characterized in that, There are multiple second target time periods. The duration of each second target time period is equal to the duration of the first target time period. For any two second target time periods, the end time of one second target time period is earlier than or equal to the start time of the other second target time period.

4. The edge device according to claim 3, characterized in that, The update module is also used for: The first target battery data is encoded to obtain a first feature vector. For each second target time period, the second target battery data corresponding to the second target time period is encoded to obtain a second feature vector. A first loss is obtained based on the first feature vector and the second feature vector. The calibration head is updated based on the first loss.

5. The edge device according to claim 3, characterized in that, The update module is also used for: The first target battery data and the second target battery data are input into the lightweight small model module, and the extreme value prediction of battery data within the first target time period is obtained by the lightweight small model module based on the first target battery data and the second target battery data. The second loss is obtained according to the extreme value prediction of battery data and the first target battery data, and the calibration head is updated according to the second loss.

6. The edge device according to claim 1, characterized in that, The edge device also includes: A cloud receiving module is communicatively connected to the lightweight small model module. The cloud receiving module is used to receive the small model parameters sent by the cloud server and update the lightweight small model module according to the small model parameters.

7. The edge device according to any one of claims 1-6, characterized in that, The battery data includes at least one of battery voltage and battery temperature.

8. A cloud server, characterized in that, include: A receiving module, which is communicatively connected to a large model module, is used to receive compressed files and send the compressed files received at multiple times to the large model module, wherein the compressed files are files sent by the edge device according to any one of claims 1-7; The large model module is used to restore the compressed file to obtain battery data.

9. The cloud server according to claim 8, characterized in that, The receiving module is further configured to: Upon receiving the compressed file, the receiving time of the compressed file is generated; The large model module is also used for: Large model distillation training is performed on the compressed file at a third preset frequency based on the receiving time being within the third target time period to obtain small model parameters, wherein the end time of the third target time period is equal to the start time of the large model distillation training. The cloud server also includes: A sending module is communicatively connected to the large model module, and the sending module is used to send the small model parameters to the edge device.

10. The cloud server according to claim 9, characterized in that, The duration of the third target time period is equal to the reciprocal of the third preset frequency.

11. A battery data transmission system, characterized in that, It includes an edge device according to any one of claims 1-7, and a cloud server according to any one of claims 8-10, wherein the edge device is communicatively connected to the cloud server.

Images