Distributed sensor data collection for optimized, centralized training of machine learning models

Through data priority sorting technology and automatic encoder evaluation, the transmission and storage of sensor data are optimized, and the problems of data transmission and battery life in machine learning model training are solved, model accuracy and battery life are improved, and more efficient concentration is achieved train.

CN119032367APending Publication Date: 2024-11-26EATON INTELLIGENT POWER LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202280094828.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2022-04-14
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

In the training of machine learning models, data transmission and battery life problems are difficult to take into account, especially in joint learning of edge devices, network bandwidth and battery life are limited, affecting the accuracy and practicality of the model.

Method used

Using data priority sorting technology, by identifying the characteristics and system constraints of sensor data, using an automatic encoder to evaluate data similarity, optimize data transmission and storage, ensure that models are trained in centralized repositories, and prioritization strategies are adjusted to improve battery Lifespan and model accuracy.

Benefits of technology

It effectively improves data transmission efficiency and battery life, while improving the accuracy of machine learning models, solves the dual challenges of data transmission and battery life in edge devices, and optimizes the effect of centralized training.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119032367A_ABST
    Figure CN119032367A_ABST
Patent Text Reader

Abstract

A method and system for optimizing sensor data collection for centralized training of machine learning models is disclosed. The method comprises the following steps: identifying a data priority ordering technology; a data prioritization technique that collects sensor data and implements identification to prioritize the data for storage and transmission to a centralized repository; transmitting the data to a centralized storage library; training a machine learning model on a centralized repository using the transmitted data to predict or identify events, such as faults in vehicles and other machines; evaluating a data prioritization technique to create an output using forward pass of the autoencoder; and determining an adjustment to the data prioritization technique based on the output.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method and system for optimizing sensor data collection for centralized training of machine learning models Background Art

[0002] Machine learning models can be trained to predict or identify events, such as failures in vehicles and other machinery. The learning process is typically data intensive. This typically requires collecting data from a large number of machines and operating states (e.g., different failure modes).

[0003] One approach is to transmit (e.g., via Bluetooth or WiFi) sensor data from edge devices (e.g., from sensor modules on a machine) for training on the cloud. For more efficient power and bandwidth consumption, established techniques can be used to compress the data. However, signals such as Bluetooth or WiFi require a lot of current, which can affect battery life. In addition, network connection constraints may limit the amount of data that can be transferred.

[0004] The alternative approach is called federated learning. This allows training to occur on local “edge” devices without transferring training data to the cloud. However, this requires a lot of bandwidth to repeatedly transfer intermediate values ​​of neural network weight layers between devices during training, and this also significantly reduces battery life. This negative impact on battery life limits the adoption of federated learning in real-world applications. In addition, in applications where the local data distribution between agents is significantly different, federated learning may perform poorly compared to centralized training. For example, Ozdayi et al., “Improving Accuracy of Federated Learning in Non-IID Settings”, Cornell University, 2020, describe a 10-17% drop in accuracy in some applications. Therefore, transferring data to a central repository (such as cloud storage for model training) has advantages in terms of model accuracy.

[0005] The present invention discloses the use of centralized model training in order to exploit the accuracy of the resulting model (compared to the federated learning approach) and to optimize for improved data selection and battery life. Summary of the invention

[0006] According to a first aspect of the present invention, there is provided a method for optimizing sensor data collection for centralized training of machine learning models, comprising: identifying a data prioritization technique; collecting sensor data and implementing the identified data prioritization technique to prioritize the data for storage and transmission to a centralized repository; transmitting the data to the centralized repository; using the transmitted data to train a machine learning model on the centralized repository to predict or identify events, such as failures in vehicles and other machinery; evaluating the data prioritization technique using a forward pass of an autoencoder to create an output; and determining adjustments to the data prioritization technique based on the output.

[0007] Preferably, identifying data prioritization techniques includes: assembling a sample data set from sensor data; characterizing the data set by identifying characteristics and sample frequencies of the data set; characterizing system constraints to determine percentages of data to store, transmit, and discard, wherein the system constraints include at least one of available memory, remaining battery power, and identified network connections; identifying potential data prioritization techniques using the characterized data set and the characterized system constraints; and identifying the best prioritization technique.

[0008] Preferably, an identified data prioritization technique is implemented to prioritize data for storage and transmission to a centralized repository, including: using a distance metric to evaluate the similarity of sensor data collected at different time periods to determine sensor data to be prioritized for storage and transmission; and based on the evaluation, sensor data determined to represent an underrepresented operating state is prioritized for storage and transmission; wherein the underrepresented operating state is an occurrence rate of the sensor data in the model training data relative to the frequency of other detected operating states falls below a threshold, wherein the threshold is one of a fixed threshold, an adaptive threshold, or an order-based threshold.

[0009] Preferably, using the distance metric to evaluate the similarity of sensor data collected at different time periods comprises: training an autoencoder to encode and reproduce the sensor data; and measuring the result using the metric.

[0010] Preferably, the distance metric comprises:

[0011]

[0012] Where A is the input vector, A i are the elements of vector A, B is the output vector, B i are the elements of the output vector, and n is the rank of vectors A and B.

[0013] Preferably, training the machine learning model on the centralized repository using the transmitted data includes: checking the accuracy of the machine learning model against a predefined benchmark.

[0014] Preferably, evaluating the data prioritization technique using a forward pass of the autoencoder includes: performing a single forward pass of the autoencoder for each sample of collected sensor data, wherein the single forward pass includes: multiplying the sensor data input by a weight matrix specified in a first layer of the model; applying a transformation to the multiplied sensor data; applying operations specified in subsequent layers of the autoencoder to intermediate autoencoder layer outputs to create an output; measuring a distance between the input and output; and adjusting the prioritization technique by retraining the prioritization technique if the distance exceeds a threshold, and maintaining the prioritization technique if the distance does not exceed the threshold.

[0015] According to a second aspect of the present invention, a system of one or more sensors and a centralized repository is provided for optimizing sensor data collection for centralized training of a machine learning model, the system being configured to: identify a data prioritization technique using one or more sensors; collect sensor data using one or more sensors and implement the identified data prioritization technique to prioritize data to be stored using local storage on one or more sensors and transmit to a centralized repository; transmit the data to the centralized repository if a network connection is available; train a machine learning model on the data in the centralized repository; evaluate the data prioritization technique using a forward pass of an autoencoder to create an output; and determine adjustments to the data prioritization technique based on the output.

[0016] Preferably, identifying data prioritization techniques using one or more sensors includes: assembling a sample data set from sensor data using one or more sensors; characterizing a data set by identifying characteristics and sample frequencies of the data set using one or more sensors; characterizing system constraints using one or more sensors to determine percentages of data to store, transmit, and discard, wherein the system constraints include at least one of available memory, remaining battery power, and identified network connections; identifying potential data prioritization techniques using the one or more sensors using the characterized data set and the characterized system constraints; and identifying an optimal prioritization technique using the one or more sensors.

[0017] Preferably, a data prioritization technique for identification is implemented using one or more sensors to prioritize data for storage and transmission to a centralized repository, including: using a distance metric to evaluate the similarity of sensor data collected at different time periods to determine sensor data to be prioritized for storage and transmission; and based on the evaluation, sensor data determined to represent an underrepresented operating state is prioritized for storage and transmission; wherein the underrepresented operating state is an occurrence rate of the sensor data in the model training data relative to the frequency of other detected operating states falls below a threshold, wherein the threshold is one of a fixed threshold, an adaptive threshold, or an order-based threshold.

[0018] Preferably, using the distance metric to evaluate the similarity of sensor data collected at different time periods comprises: training an autoencoder to encode and reproduce the sensor data; and measuring the result using the metric.

[0019] Preferably, the metrics include:

[0020]

[0021] Where A is the input vector, A i are the elements of vector A, B is the output vector, B i are the elements of the output vector, and n is the rank of vectors A and B.

[0022] Preferably, training the machine learning model on the data in the centralized repository using the transmitted data includes: checking the accuracy of the machine learning model against a predefined benchmark.

[0023] Preferably, evaluating the data prioritization technique using a forward pass of the autoencoder includes: performing a single forward pass of the autoencoder for each sample of collected sensor data, wherein the single forward pass includes: multiplying the sensor data input by a weight matrix specified in a first layer of the model; applying a transformation to the multiplied sensor data; applying operations specified in subsequent layers of the autoencoder to intermediate autoencoder layer outputs to create an output; measuring a distance between the input and output; and adjusting the prioritization technique by retraining the prioritization technique if the distance exceeds a threshold, and maintaining the prioritization technique if the distance does not exceed the threshold. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Embodiments of the invention will now be described, by way of example only, with reference to the accompanying drawings, in which:

[0025] Figure 1 A method according to a first aspect of the invention is described;

[0026] Figure 2Describes preferred methods for identifying data prioritization techniques;

[0027] Figure 3 Describes the preferred method for evaluating data prioritization techniques using a forward pass of an autoencoder;

[0028] Figure 4A and Figure 4B a flow chart depicting the method of the present invention; and

[0029] Figure 5 Exemplary embodiments of the present invention are depicted. DETAILED DESCRIPTION

[0030] refer to Figure 1 , which depicts a method 100 according to the first aspect of the invention. Figure 1 Method 100 includes steps 110 to 160. Step 110 includes identifying a data prioritization technique. Step 120 includes collecting sensor data and implementing the identified data prioritization technique to prioritize the data for storage and transmission to a centralized repository (e.g., a cloud). Step 130 includes transmitting the data to the centralized repository. Step 140 includes training a machine learning model on the centralized repository using the transmitted data to predict or identify events, such as failures in vehicles and other machinery. Step 150 includes evaluating the data prioritization technique using a forward pass of an autoencoder to create an output; and step 160 includes determining adjustments to the data prioritization technique based on the output.

[0031] refer to Figure 2 , which depicts a method 200 for identifying data prioritization techniques. Figure 2 The method 200 includes steps 210 to 250. Step 210 includes assembling a sample data set from sensor data. Step 220 includes characterizing the data set by identifying characteristics of the data set and sample frequencies. Step 230 includes characterizing system constraints to determine percentages of data to store, transmit, and discard, wherein the system constraints include at least one of available memory, remaining battery power, and identified network connections. Step 240 includes identifying potential data prioritization techniques using the characterized data set and the characterized system constraints; and step 250 includes identifying an optimal prioritization technique.

[0032] refer to Figure 3, which depicts a method 300 of a preferred method for evaluating a data prioritization technique using a forward pass of an autoencoder, wherein the autoencoder performs steps 310 to 350. Step 310 includes multiplying the sensor data input by a weight matrix specified in a first layer of the model. Step 320 includes applying a transformation to the multiplied sensor data. Step 330 includes applying an operation specified in a subsequent layer of the autoencoder to an intermediate autoencoder layer output to create an output. Step 340 includes measuring a distance between the input and the output. Step 350 includes determining whether the distance exceeds a threshold, then adjusting the prioritization technique by retraining the prioritization technique, and if the distance does not exceed the threshold, maintaining the prioritization technique.

[0033] refer to FIG. 4A to FIG. 4B , which depict a flow chart of a method 400 of the present invention. The method comprises steps 402 to 454.

[0034] exist Figure 4A In steps 402 to 412 of the method for identifying a data prioritization technique is described. In step 402, the data prioritization technique is initialized, and a sample data set is assembled in step 404. Step 406 includes characterizing the data set, including characterizing a plurality of variables and sample frequencies. Step 408 includes characterizing system constraints, wherein the characterized system constraints include memory, battery power, and connectivity. In step 410, a search space is specified for the prioritization technique, and in step 412, the best prioritization technique is identified.

[0035] exist Figure 4A In steps 414 to 426 of , a method of collecting sensor data and implementing an identified data prioritization technique to prioritize the data for storage and transmission to a centralized repository is described. In step 414, the data collection and prioritization steps are initialized. In step 416, sensor data is collected; in step 418, the prioritization technique is implemented; and in step 420, the priority data is stored locally. In step 422, it is determined whether the battery charge is sufficient for data transmission.

[0036] If it is determined in step 422 that the battery charge is sufficient for data transmission, the method proceeds to step 424. In step 424, it is determined whether a suitable connection (e.g., a suitable WiFi connection) exists. If it is determined in step 424 that a suitable connection does not exist, the method proceeds to step 428, as described below. If it is determined that a suitable connection exists, the method proceeds to step 426, where the data is transmitted to the cloud (i.e., a centralized repository), and the method proceeds to step 432.

[0037] If it is determined in step 422 that the battery charge is insufficient for data transfer, the method proceeds to step 428. In step 428, it is determined whether the data storage has reached a threshold. If it is determined in step 428 that the data storage has not reached a threshold, the method returns to step 416 and sensor data is collected, and the method continues from step 416. When it is determined in step 428 that the data storage has reached a threshold, the method proceeds to 430, wherein in step 430, a prioritization algorithm (i.e., a prioritization technique) is implemented to prioritize new data over old data, and the method returns to step 416. The process continues iteratively until it is determined in step 424 that a suitable connection exists.

[0038] exist Figure 4B In steps 432 to 442 of , a method is described for training a machine learning model on a centralized repository using the data transmitted in step 426 to predict or identify events (e.g., failures in vehicles and other machinery).

[0039] In step 432, a method for training a model on the cloud is initialized. In step 434, the basis of the neural network is frozen. In step 436, a subset of the network layers of the neural network are fine-tuned, and in step 438, the accuracy is checked against a benchmark. In an example, the fine-tuning of step 436 includes making changes to the neural network to take into account the newly available data, and also includes changing the weights of the later layers during the training process, while the weights of the base layer remain unchanged during the training process. In step 440, it is determined whether the model meets the benchmark. If it is determined in step 440 that the model does not meet the benchmark, the method proceeds to step 444. If it is determined in step 440 that the model meets the benchmark, the method proceeds to step 442. In step 442, the updated model is deployed to the device (i.e., the sensor device that transmitted the data to the cloud in step 426).

[0040] exist Figure 4B In steps 444 to 454 of , a method of evaluating a data prioritization technique and determining adjustments to the data prioritization technique is shown.

[0041] In step 444, a method for re-evaluating the data prioritization technique is initialized. In step 446, a forward pass of the autoencoder is performed to create an output, and the distance between the input (i.e., sensor data) and the output is measured in step 448. In step 450, it is determined whether the measured distance of step 448 exceeds a threshold. If the distance does not exceed the threshold, the method proceeds to step 454, and the method ends. If the distance exceeds the threshold, the method proceeds to step 452 and the prioritization technique is adjusted. After the adjustment, the method proceeds to step 454.

[0042] refer to Figure 5 , which depicts an exemplary embodiment of a system 500 of the present invention. In the system 500, a plurality of vehicles 510, 520, 530, and 540 are depicted. Each of the plurality of vehicles 510, 520, 530, and 540 transmits sensor data (sensor data 1, 2, 3, and 4, respectively) to a central repository 550.

[0043] It will be appreciated that the above described embodiments of the first and second aspects of the invention are given by way of example only and that various modifications may be made to the embodiments without departing from the scope of the invention as defined in the appended claims.

[0044] For example, in use, Figure 1 Step 110 may involve assembling a sample data set, such as Figure 2 As described in step 210 of . A sample data set may be assembled to develop initial characteristics of training data (i.e., sensor data). The training data may come from one or more devices including sensors. For example, during a predictive maintenance task (e.g., vehicle maintenance), three-phase voltages and currents may be collected for selection of devices and operating modes (e.g., high power mode, low power mode, fault mode).

[0045] Step 110 may further involve characterizing the data set from the sensor data by analyzing a plurality of variables and sample frequencies, such as with respect to Figure 2 As described in step 220. For example, the analyzed characteristics may include data volume, data maximum value, data minimum value, data mean value, data standard deviation and autocorrelation. In the case where the data set includes multiple predictor variables, a correlation matrix may be used to characterize the relationship between the multiple predictor variables.

[0046] Step 110 may further include characterizing system constraints, such as Figure 2 As described in step 230 of . For example, system constraints may include memory capacity, persistent storage, battery power, and connectivity capabilities. Characterizing the system constraints may determine the percentage of sensor data to be stored, transmitted, and discarded. For example, a system with poor network connectivity and / or with low data storage capacity may be configured to discard a large amount of sensor data. Based on the system constraints, a data storage target may be defined.

[0047] Step 110 may further include identifying a data prioritization technique by specifying a search space for the prioritization technique, such as Figure 2As described in step 240 of . Potential methods for data prioritization can be specified. For example, an autoencoder (further described below with respect to step 120) can be used for data prioritization. The number of layers of the autoencoder can be determined by the performance constraints of the sensor device (also referred to herein as the local device) (i.e., the system constraints and the characteristics of the above-mentioned data set). By analyzing the performance constraints, the search space of the prioritization technique can be determined.

[0048] Step 110 may further include identifying an optimal prioritization technique by systematically exploring the search space, such as Figure 2 As described in step 250 of . For example, this may involve training an autoencoder neural network to encode and reproduce the input data. The optimal architecture of the autoencoder can be determined by its ability to minimize the distance between the input (i.e., the input data) and the output (i.e., the output from the autoencoder). This is further described below with respect to step 120.

[0049] Figure 1 Step 120 may further include collecting and characterizing sensor data on the local edge device. For example, sensor data may be collected from any device such as a hydraulic pump, a diesel engine, or any other device.

[0050] Step 120 may further include implementing the prioritization technique determined in step 110 above. The prioritization technique may include an algorithm that uses a distance metric to evaluate the similarity of data collected over different time periods to determine which data will be prioritized for data retention (i.e., storage). For example, data that has been determined to be more unique or corresponds to new or underrepresented operating states (e.g., new failure modes) may be determined to be more valuable and may be prioritized for data transmission.

[0051] The uniqueness of the data can be analyzed using a variety of methods. For example, an initial (first pass) prioritization test that can be utilized is to use summary statistics of the distribution of the sensor data. A second (or alternative) prioritization test can involve the use of an autoencoder (as described above with respect to step 110). An autoencoder is an artificial neural network that is trained to encode and then reproduce training data. The ability of an autoencoder to accurately reproduce the input data is a measure of the novelty of the data, which allows for prioritization as described above. To assess the similarity between the autoencoder input and output, a metric such as the cosine distance can be calculated:

[0052]

[0053] Where A is the input vector, A i are the elements of vector A, B is the output vector, B iare the elements of the output vector, and n is the rank of vectors A and B.

[0054] A larger value of the cosine distance equation above indicates that the data is more unique when compared to a lower value from the cosine distance equation, and therefore more likely to improve the accuracy of the predictive model. In other words, a higher cosine distance provides a higher priority score for the data.

[0055] Step 120 may further include locally storing prioritized data based on available storage and the determined priority of the new data. In addition, step 120 may involve determining whether the battery charge is sufficient for data transmission. If the battery charge is sufficient, it is determined whether a network connection is available for transmitting the data. If a network connection is available, the data can be transmitted to a central repository (e.g., a cloud) for model training. If the network connection is not available (or insufficient) and / or if the battery charge is insufficient for data transmission, the method further includes determining whether the data storage has reached a specified threshold. If the threshold is met, at least some of the data is discarded. New data can be prioritized for storage relative to old data.

[0056] In practice, Figure 1 Step 130 may further include compressing the data to be transmitted using known methods prior to transmission. For example, for periodic signals, the data can be converted from the time domain to the frequency domain using a fast Fourier transform (or some variation thereof). The frequency domain data can then be transmitted to a central repository where it can be used to recreate the underlying time domain signal using an algorithm such as an inverse fast Fourier transform. The recreated signal can then be used as input for training a machine learning model. A conventional approach to data augmentation is to use all available real data in a training dataset and generate additional data by changing certain values ​​of the data from their "true" values ​​within a predetermined probability distribution. For example, the scale, color, or rotation angle of an image is changed to provide additional input to the machine learning model. This can improve the generalization of the machine learning model to real-world settings.

[0057] Figure 1 Steps 150 and 160 may further include performing a single forward pass of the autoencoder, wherein a single forward pass is performed for each sample of the most recently collected data. Performing the single forward pass may involve multiplying the input by the weights associated with the first layer of the model, applying a transformation (e.g., a rectified linear unit), and implementing each subsequent layer of the autoencoder network. The method may further include measuring a distance between the input and output, and if it is determined that the distance exceeds a threshold, adjusting the prioritization technique by retraining (i.e., fine-tuning) using the new data.

[0058] about Figure 4B, and in particular with respect to step 432, which includes freezing the basis of the neural network, which may not be necessary if the prediction model has not yet been trained. In step 436, fine-tuning may further include training the neural network based on a user-defined search space, such as using hyperparameters of the neural network that may be defined using a grid search. The model weights may be updated based on a process of backpropagation and gradient descent. Alternatively, a variant of gradient descent and backpropagation, such as by using an adaptive moment estimator or root mean square propagation. For example, the model weights may be updated using the following method: θ 1 =θ 0 -α▽J(θ),

[0059] Among them, α is the learning rate; θ 1 is the updated weight; θ 0 is the weight before update; and ▽J(θ) is the weight difference.

Claims

1. A method for optimizing sensor data collection for centralized training of a machine learning model, comprising: Identify data prioritization techniques; collecting sensor data and implementing the identified data prioritization techniques to prioritize said data for storage and transmission to a centralized repository; transmitting the data to the centralized repository; Using the transmitted data to train the machine learning model on the centralized repository to predict or identify events, such as failures in vehicles and other machinery; evaluating the data prioritization technique using a forward pass of the autoencoder to create an output; as well as Adjustments to the data prioritization technique are determined based on the output.

2. The method according to claim 1, wherein: Identify the data prioritization techniques including: assembling a sample data set from the sensor data; characterizing the data set by identifying characteristics and sample frequencies of the data set; characterizing system constraints to determine percentages of data to store, transmit, and discard, wherein the system constraints include at least one of available memory, remaining battery power, and an identified network connection; identifying potential data prioritization techniques using the characterized data sets and the characterized system constraints; and Identify the best prioritization techniques.

3. The method according to claim 1, wherein: Implement the identified data prioritization techniques to prioritize said data for storage and transmission to a centralized repository, including: evaluating similarities of the sensor data collected at different time periods using a distance metric to determine the sensor data to be prioritized for storage and transmission; and based on the evaluation, the sensor data determined to represent an underrepresented operating state is prioritized for storage and transmission; The under-represented operating state is when the occurrence rate of sensor data in the model training data relative to the frequency of other detected operating states falls below a threshold, wherein the threshold is one of a fixed threshold, an adaptive threshold, or a sequence-based threshold.

4. The method according to claim 3, wherein: Distance metrics used to evaluate the similarity of sensor data collected at different time periods include: The autoencoder is trained to encode and reproduce the sensor data; and a metric is used to measure the results.

5. The method according to claim 4, wherein: The metrics include: Where A is the input vector, A i are the elements of vector A, B is the output vector, B i are the elements of the output vector, and n is the rank of vectors A and B.

6. The method according to claim 1, wherein: Training the machine learning model on the centralized repository using the transmitted data includes: Check the accuracy of the machine learning model against a predefined benchmark.

7. The method according to claim 1, wherein: A forward pass using the autoencoder to evaluate the data prioritization technique involves: Performing a single forward pass of the autoencoder for each sample of the collected sensor data, wherein the single forward pass comprises: Multiplying the sensor data input by the weight matrix specified in the first layer of the model; Applying transformations to the multiplied sensor data; applying operations specified in subsequent layers of the autoencoder to intermediate autoencoder layer outputs to create outputs; measuring a distance between the input and the output; and If the distance exceeds a threshold, the prioritization technique is adjusted by retraining the prioritization technique, and if the distance does not exceed the threshold, the prioritization technique is maintained.

8. A system of one or more sensors and a centralized repository for optimizing sensor data collection for centralized training of a machine learning model, the system being configured to: identifying a data prioritization technique using the one or more sensors; collecting sensor data using the one or more sensors and implementing the identified data prioritization technique to prioritize the data to be stored using local storage on the one or more sensors and transmitted to the centralized repository; If a network connection is available, transmitting the data to the centralized repository; training the machine learning model on the data in the centralized repository; evaluating the data prioritization technique using a forward pass of the autoencoder to create an output; as well as Adjustments to the data prioritization technique are determined based on the output.

9. The system according to claim 8, wherein: Identifying the data prioritization technique using the one or more sensors includes: assembling a sample data set from the sensor data using the one or more sensors; characterizing the data set using the one or more sensors by identifying characteristics and sample frequencies of the data set; characterizing system constraints using the one or more sensors to determine percentages of data to store, transmit, and discard, wherein the system constraints include at least one of available memory, remaining battery power, and an identified network connection; using the one or more sensors, identifying potential data prioritization techniques using the characterized data sets and the characterized system constraints; and An optimal prioritization technique is identified using the one or more sensors.

10. The system according to claim 8, wherein: Implementing the identified data prioritization technique using the one or more sensors to prioritize the data for storage and transmission to a centralized repository, including: evaluating similarities of the sensor data collected at different time periods using a distance metric to determine the sensor data to be prioritized for storage and transmission; and based on the evaluation, the sensor data determined to represent an underrepresented operating state is prioritized for storage and transmission; The under-represented operating state is when the occurrence rate of sensor data in the model training data relative to the frequency of other detected operating states falls below a threshold, wherein the threshold is one of a fixed threshold, an adaptive threshold, or a sequence-based threshold.

11. The system according to claim 8, wherein: Distance metrics used to evaluate the similarity of sensor data collected at different time periods include: The autoencoder is trained to encode and reproduce the sensor data; and a metric is used to measure the results.

12. The system according to claim 11, wherein: The metrics include: Where A is the input vector, A i are the elements of vector A, B is the output vector, B i are the elements of the output vector, and n is the rank of vectors A and B.

13. The system according to claim 8, wherein: Training the machine learning model on the data in the centralized repository using the transmitted data includes: Check the accuracy of the machine learning model against a predefined benchmark.

14. The system according to claim 8, wherein: A forward pass using the autoencoder to evaluate the data prioritization technique involves: Performing a single forward pass of the autoencoder for each sample of the collected sensor data, wherein the single forward pass comprises: Multiplying the sensor data input by the weight matrix specified in the first layer of the model; Applying transformations to the multiplied sensor data; applying operations specified in subsequent layers of the autoencoder to intermediate autoencoder layer outputs to create outputs; measuring a distance between the input and the output; and If the distance exceeds a threshold, the prioritization technique is adjusted by retraining the prioritization technique, and if the distance does not exceed the threshold, the prioritization technique is maintained.