Transfer learning in machine learning models

By creating synthetic data points in the target domain and using optimal transmission techniques with smooth regularization, the problem of data sparsity in the manufacturing environment is solved, achieving efficient data distribution alignment and model generalization, and improving the performance of automated testing equipment.

CN123125609APending Publication Date: 2026-11-17ANALOG DEVICES INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202580028466.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2024-04-18
Filing Date
2025-04-18
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In manufacturing environments, acquiring high-quality data for training machine learning models is difficult and costly, especially for sparse failure modes or rare defect categories, where existing technologies struggle to effectively align data distributions and generalize models.

Method used

By employing the optimal transport technique of smoothing regularization, synthetic data points are created in the target domain to fill data gaps, and the data distribution of the source and target domains is aligned using sliced ​​Wasserstein optimal transport, thereby improving the model's generalization ability in the target domain.

Benefits of technology

It improves the robustness and stability of transfer learning, reduces the dependence on large-scale labeled data in the target domain, lowers training costs, and expands the reusability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123125609A_ABST
    Figure CN123125609A_ABST
Patent Text Reader

Abstract

This paper describes systems and techniques for transfer learning to achieve interoperability between machine learning models. Data in the target domain undergoes a transfer process that moves the data to the source domain. This transfer process includes creating synthetic data based on a portion of the target domain data and constraints. A transformation is applied to the combination of the synthetic data and the original data, which moves the combination to the source domain. The ML model trained on the source domain is then invoked on the data resulting from this transformation.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Priority requirements

[0002] This patent application claims priority to U.S. Provisional Application Serial No. 63 / 635,865, filed April 18, 2024, entitled “DOMAINADAPTATION VIA REGULARIZED OPTIMAL TRANSPORT”, pursuant to 35 USC § 119, the entire contents of which are incorporated herein by reference. Technical Field

[0003] The implementation scheme described in this article involves automated testing equipment as a whole, and more specifically, transfer learning in machine learning models. Background Technology

[0004] Automated Test Equipment (ATE) in manufacturing can include computer-controlled systems designed to test or verify the functionality, performance, or safety of a product. ATE systems can be integrated into manufacturing workflows to perform electrical, mechanical, or functional tests with minimal human intervention. ATE can include end-of-line (EOL) testing, which verifies that a product meets required specifications after the final stage of assembly, or can categorize aspects of a product's performance. EOL testing can include measurements such as voltage, current, impedance, or signal integrity, depending on the product type. For example, in battery production, ATE systems can be used to test parameters such as state of charge, capacity, internal resistance, or thermal behavior to ensure compliance with performance and safety standards. ATE has been used in a variety of products, such as printed circuit boards, power modules, or sensors, where test coverage and repeatability help maintain quality and reliability at high manufacturing volumes.

[0005] Artificial intelligence (AI) refers to a broad set of computational methods that enable machines to perform tasks that typically require human intelligence, such as perception, decision-making, or pattern recognition. AI generally includes machine learning (ML), where ML models are trained on large datasets to identify patterns or make predictions. The training process typically involves optimizing model parameters (e.g., synaptic weights) to minimize the error between predicted outputs and actual results, and often uses techniques such as gradient descent, in supervised or unsupervised learning contexts. For example, ML models can be trained on labeled data to perform classification or regression tasks, or on unlabeled data to reveal underlying structures (e.g., patterns). In automated testing applications, ML models can enhance the capabilities of traditional ATE (Automatic Test Equipment) by enabling predictive maintenance, anomaly detection, or adaptive testing strategies. For example, ML models can analyze historical test data to identify patterns indicating potential defects or predict equipment failures, leading to more efficient test scheduling or reduced downtime. Attached Figure Description

[0006] In accompanying drawings that are not necessarily drawn to scale, similar reference numerals may describe similar parts in different views. Similar reference numerals with different letter suffixes may indicate different instances of similar parts. The accompanying drawings generally illustrate the various embodiments discussed in this document by way of example rather than limitation.

[0007] Figure 1 This is a block diagram of an example environment based on an implementation scheme, which includes a system for transfer learning in a machine learning model.

[0008] Figure 2 An example of a transfer from the target domain to the source domain according to one implementation scheme is shown.

[0009] Figure 3 An example of smoothing sparse data in a data domain is illustrated according to one implementation scheme.

[0010] Figure 4 An example of machine learning according to one implementation is illustrated, which is used to generate a transfer from the latent data space embedding of target domain data to the latent data space embedding of target domain data.

[0011] Figure 5 An example based on one implementation is shown, in which a machine learning model trained on source domain data is used to infer target domain data.

[0012] Figure 6 An example according to one implementation is illustrated, in which a machine learning model trained on source domain data transformed to the target domain using a transport mapping is used to infer the target domain data.

[0013] Figure 7 A flowchart illustrating an implementation method is provided, which serves as an example of a transfer learning approach in a machine learning model.

[0014] Figure 8 This is a block diagram illustrating an example machine on which one or more implementations can be carried out. Detailed Implementation

[0015] Typically, to produce effective ML models, training relies on large amounts of high-quality data (e.g., accurate, well-distributed across the problem domain) to achieve accurate and generalizable performance. However, acquiring such data in manufacturing environments presents significant challenges. Data acquisition can be expensive or time-consuming, especially when it involves generating experimental test cases or capturing measurements across multiple product variants. Furthermore, data can be sparse for certain failure modes or rare defect categories, limiting the ability of ML models to be trained to correctly handle (e.g., accurately classify, predict, etc.) these scenarios. In many cases, labeled data requires expert annotation, which further increases costs and reduces scalability.

[0016] Transfer learning provides a mechanism to address these challenges by adapting knowledge from a data-rich source domain to a data-sparse target domain. This approach is particularly useful when direct data acquisition from the target domain is limited or impractical. Techniques such as optimal transfers can be used to align the distributions of source and target data, enabling more efficient transfer of learned representations. Optimal transfer (OT) is a mathematical framework that finds the most efficient way to transform one probability distribution into another by minimizing the transfer cost. In the context of ML, OT can be used to align data distributions across different domains. This alignment helps models trained on the source domain generalize better to target domains with different characteristics. For example, sliced ​​Wasserstein optimal transfers provide a computationally efficient way to learn transformation mappings between domains, enabling the reuse of existing model architectures and reducing reliance on large-scale labeled data in the target domain. This capability enhances the practicality of deploying ML-driven improvements in automated testing in variable production scenarios.

[0017] While sliced ​​Wasserstein optimal transfer achieves efficient domain fitting, it can sometimes be too powerful when mapping a broader distribution of source domain data to sparse data points in the target domain, resulting in a mapping that fails to reflect meaningful structural relationships within the data. This problem can be particularly relevant when processing sensor data, where the measurement space in the target domain may be only partially observed due to limitations imposed by permissible test configurations or operational constraints. In such cases, the target distribution may be underrepresented or incomplete after applying sliced ​​Wasserstein optimal transfer.

[0018] To address the problem of sliced ​​Wasserstein optimal transfers, smoothing regularization can be applied to the target domain data to expand the observed data points by intelligently distributing or "smoothing" the data points in the target domain to unobserved regions (e.g., gaps) within the domain. This technique enables a more uniform data distribution coverage in the target domain without additional measurements before applying sliced ​​Wasserstein optimal transfers—or other types of optimal transfers or transits (e.g., transformations) between the target and source domains. This improves the robustness and stability of the learned transformation mappings and enhances the reliability of transfer learning in automated testing applications. Further details and examples are described below.

[0019] Figure 1 This is a block diagram of an example environment according to one implementation, which includes a system 105 for transfer learning in an ML model 155. System 105 includes a processing circuitry system 110, a storage device 120 (e.g., a storage device that retains data after power failure, such as a hard disk drive, solid-state drive, etc.), and a memory 115. Memory 115 is typically used to maintain operational state information of system 105, which is typically discarded between power-on and power-off cycles or restarts. Both memory 115 and storage device 120 are in the form of computer-readable media. The processing circuitry system 110—or software residing in memory 115 or storage device 120 and executing on the processing circuitry system 110—configures system 105 to perform various operations during runtime (e.g., during operation 135).

[0020] The system may include sensor 125, be coupled to sensor 125 (e.g., via an interface), or receive data originating from sensor 125. In the illustrated scenario, sensor 125 may be part of an ATE component that measures product 130. In one example, the measurement may be one of a set of measurements of product 130. As described above, in the context of manufacturing, such measurements can be used to verify compliance with requirements (e.g., manufacturing tolerances), classify products (e.g., classify them into one of several performance categories), or otherwise assist the manufacturing process. While the examples shown and some of the examples discussed below are given in the context of manufacturing, this technique is generally applicable to improving the transfer process for applying an ML model 155 trained on a source domain to a target domain.

[0021] At a higher level, techniques for improving transfer learning in ML models involve smoothing the target domain data to improve the results of the transfer function. Smoothing creates synthetic data based on data points in the target domain data (e.g., anchored by these data points). This synthetic data provides additional elements for the transfer process, thereby improving the results of the transfer process and thus improving the performance of ML model 155 on the transferred target domain data. Therefore, processing circuitry 110 is configured to access a medium—such as memory 115, storage device 120, buffers in sensor 125, etc. (e.g., via an interface)—that includes data in the target domain (e.g., sampled data 140).

[0022] In this paper, the term "target domain" refers to the data domain that is defined as the target for use with the ML model 155, while the "source domain" is the data domain on which the ML model 155 is trained. These relationships typically arise when the target domain has insufficient data points to effectively train the model. Thus, in one example, the data in the target domain is sparse. Generally, sparsity refers to the ratio of present data to missing data in a dataset. This ratio can be estimated by the absence of data within a window of data in one dimension of the target domain. For example, if that dimension is human height and there are no values ​​between five and six feet, then that window of data in the height dimension is missing data because it is expected but missing. In one example, "sparse" means that the data within the target domain includes gaps exceeding a threshold (e.g., windows of missing information) (e.g., gaps with a certain amount of value within one dimension).

[0023] The processing circuitry 110 is configured to perform a transfer process on data from the target domain to move the data toward the source domain and create input data 150. The direction of the movement can be either (e.g., from the target domain to the source domain, or from the source domain to the target domain), but it is generally more efficient to transfer the target domain data (e.g., transform, translate, etc.) to the source domain because the target domain typically contains less data. The result is referred to as "input data" because it will be the data that the ML model 155 operates on, or the data used as input to that ML model.

[0024] The transfer process described herein involves one or more sub-processes. In one example, the transfer process includes a processing circuitry 110 configured to create synthetic data based on a portion of data from a target domain (e.g., sampled data 140) and constraints. For example, the synthetic data may include copies of data points from sampled data 140, shifted along one dimension of the target domain such that, for example, if a data point has a value of 1, the copy could have a value of 2. The constraint restricts the boundaries of the synthetic data. For example, the constraint may be based on the original data points and may limit how much variation the synthetic data can have. Using the preceding example, for a set of synthetic data points, sequentially increasing integers may be added to the data point values ​​until the constraint is reached. Thus, if the constraint is 5 and the original data point is 8, the synthetic data could include values ​​at 9, 10, 11, 12, and 13. These simplified examples are only intended to illustrate how data points in the target domain, together with constraints, are used to generate (e.g., define) synthetic data points.

[0025] Synthetic data is designed to provide a more complete dataset for applications on which to transfer data. Therefore, in one example, the data in the target domain is sparse, and synthetic data fills a portion of the gaps in the source data. In one example, this portion of data is located at the edge of the gap. In this example, creating synthetic data based on this portion involves smoothing the data to fill the gap. In this case, smoothing means creating synthetic data in such a way that synthetic data points with higher concentrations are closer to the data points in the target domain (e.g., sampled data 140), and the further away the synthetic data points are, the more dispersed they become. Using the integer example above, synthetic data points can be added in this dimension at double the interval for each iteration. For example, if the data point is 1 and the step size starts at 0.25, the first synthetic data point is at 1.25, the next at 1.75 (e.g., 0.25 doubled to 0.5 and added to 1.25), the third at 2.25, and so on, until the constraint is met. Smoothing provides a synthetic data concentration centered on the obtained measurements, and the further the synthetic data points are from the known data points in the target domain data, the less impact the synthetic data points have.

[0026] In one example, the constraint is the distribution of values. In another example, this distribution is a Gaussian distribution. Such distributions typically reflect characteristics of the source domain, or can be chosen to reflect characteristics of the source domain. This allows for obtaining a better target domain dataset to determine the transfer to the source domain. In one example, the distribution is based on a dimension of the source domain. This example allows the distribution of values ​​in the source domain to influence how much smoothing is applied to the synthetic data in the target domain. In one example, creating synthetic data based on constraints involves maintaining this distribution when creating new values ​​in a dimension of the synthetic data. In another example, the constraint is a configurable parameter obtained from a configuration file or user interface. This last example illustrates that the distribution is a parameter of the synthetic generation process. Here, a configuration file or user interface can be used to allow the user to choose which distribution (e.g., type or parameter) to apply as a constraint.

[0027] The processing circuitry 110 is configured to combine synthesized data with data from the target domain (e.g., sampled data 140) to create intermediate data 145. This combination can be incorporated into the creation of the synthesized data (e.g., added to intermediate data 145 each time a synthesized data point is created, such as during buffer filling during creation), or it can be completely separate and then added to the sampled data 140 to create intermediate data 145. Furthermore, the data can be stored in different ways. For example, the synthesized data can be added directly to a data structure that already contains the sampled data 140, or a completely new data structure (e.g., a file, a database, etc.) can be used.

[0028] The processing circuitry 110 is configured to apply a transformation to the intermediate data 145 to move the intermediate data to the source domain. This transformation is a "transfer" in the transmission process. Figure 2 An exemplary illustration of a transfer from one domain to another is provided. This transformation includes translation, rotation, scaling, or other operations to “fit” data points in the target domain to the source domain. In one example, this transformation is created through a training process to map data points from the target domain to corresponding data points in the target domain. In one example, Slice Wasserstein Optimal Transfer (SWOT) is used as a loss function during the training process to obtain the optimal transfer from the target domain to the source domain. In one example, this transformation is represented by a transfer map. In one example, a transfer map is a function that defines how probability quality is moved or redistributed from the source probability distribution to the target probability distribution in a way that minimizes transfer costs (e.g., the Wasserstein distance in the context of Slice Wasserstein Optimal Transfer).

[0029] Although the foregoing example does not specify the form of the data (e.g., target data points or sampled data 140, intermediate data 145, etc.), several possible forms exist. For example, data sampled from sensor 125 may be cleaned, normalized, or otherwise preprocessed. In one example, the data is represented in a latent data space.

[0030] Modern machine learning techniques typically operate on latent data spaces. A latent data space is generally a representation (e.g., a compressed representation) of the input data that captures underlying features or patterns within it. Latent data spaces are usually not directly observable but are learned by the model during training to more compactly represent complex data distributions. Latent data spaces are often created by neural network architectures such as autoencoders, variational autoencoders (VAEs), or generative adversarial networks (GANs). In these models, the encoder typically maps high-dimensional input data to a lower-dimensional latent space, where similar inputs tend to have similar latent representations. Latent data spaces enable tasks such as clustering, interpolation, generation, or anomaly detection. A decoder or generator can then reconstruct or synthesize data from the latent representation. Latent data spaces enable models to generalize from observed examples, facilitate efficient computation, or support interpretability by revealing structure within the data. Thus, in one example, processing circuitry 110 is configured to invoke the encoder on source or intermediate data 145 to produce an embedding (e.g., a vector representation) of input data 150 into the latent data space. In one example, the transfer process accepts the embedding as input (e.g., the embedding is performed on sampled data 140 before generating synthetic data). In one example, ML model 155 is configured to perform inference on the latent data space to produce results.

[0031] Processing circuitry 110 is configured to invoke ML model 155 on input data 150 to obtain result 160. As described above, ML model 155 is trained or partially trained on data from the source domain to produce output. The aforementioned technique enables the use of ML model 155 to perform efficient inference (or other ML output) on data from the target domain without training on the target domain. Since model training is expensive, this type of transfer learning allows a model trained once to be applied to many target domains. By using the aforementioned synthetic data generation (e.g., smoothing), target domains with sparse data can be adapted to use ML model 155, further extending the reusability of ML model 155 and avoiding unnecessary training or latency.

[0032] The techniques described herein can be used in industries producing different configurations of the same product. In one example, the source domain is a first set of measurements for a first product, and the target domain is a second set of measurements for a second product. In one example, the first and second products are different products of the same type. In one example, this same type is a battery. Here, in one example, the first product has a different chemical composition or a different form factor than the second product. In one example, the first and second sets of measurements include corresponding electrochemical impedance spectroscopy (EIS) data, voltage, current, or state of charge (SoC) data for the first and second products. In one example, the result is the core temperature of the battery. Therefore, consider the source domain from EIS data of a battery that has been cycled at many cycles and temperatures, and the target domain from EIS data of a battery that has been cycled at fewer cycles and temperatures and has a different form factor and chemical composition. The task can be to estimate the battery core temperature from the EIS data using an ML model trained on the source domain. Consider data obtained from early-life electrical data of lithium-ion batteries, which includes EIS as well as voltage, current, and SoC data. The source domain can correspond to a 100 ampere-hour (Ah) lithium-ion (Li-ion) nickel manganese cobalt aluminum oxide (NMCA) pouch cell, and the target domain can correspond to a 65Ah lithium-ion nickel manganese cobalt oxide (NMC) pouch cell with different EIS characteristics.

[0033] Figure 2 An example of a transfer from a target domain to a source domain according to one implementation is illustrated. The original state on the left includes source domain 205 and target domain 210. Label 215 shows the relationship between square and circular data points in target domain 210. Transfer 220 (e.g., transfer mapping) is applied to the data points in target domain 210 to transfer these data points to source domain 205. The rightmost portion includes label 215 in source domain 205 to illustrate that transfer 220 preserves the relative relationships between square and circular data points from target domain 210 in source domain 205.

[0034] Figure 3 An example of smoothing sparse data in a data domain is illustrated according to one implementation scheme. In transmission (e.g., as...) Figure 2A potential problem in the example is that sparse data can cause distortion of the transmitted data points. To address this, the data in the sparse domain can be smoothed (e.g., using synthetic data representing the gradient starting from existing data points) to fill the sparse domain before computation and transmission. As illustrated, consider a sparse domain 320 with sparse data points 305, 310, and 315. If the distance between data points 305 and 310 is as expected, it is clear that there is missing data between data points 310 and 315. This gap can be defined in different ways, including using a predefined threshold.

[0035] Figure 3 The bottom section illustrates applying distribution constraints to synthetic data 325 based on data point 310 to help fill gaps. Although not illustrated, similar constraints on distributions in another direction could use data point 315 as an anchor to further help fill gaps. The type of distribution used as a constraint can depend on the expected value of the sparse domain 320, measurements from the source domain, or other factors. Exemplary distributions may include Gaussian, Bernoulli, binomial, Poisson, exponential, gamma, beta, or Dirichlet distributions, etc. The Gaussian distribution, also known as the normal distribution, is defined by its mean and variance and is symmetric about the mean; it is commonly used to model continuous data with a central tendency. The Bernoulli distribution models binary outcomes and is parameterized by a single probability of success. The binomial distribution extends the Bernoulli distribution to a fixed number of independent trials with the same probability of success. The Poisson distribution models count data or the number of rare events occurring within a fixed interval. The exponential distribution models time between independent events occurring at a constant average rate. The gamma distribution generalizes the exponential distribution to model waiting times for multiple events and is characterized by shape and rate parameters. The beta distribution models random variables constrained to the interval [0, 1] and is commonly used for Bayesian inference. The Dirichlet distribution generalizes the beta distribution to multivariate settings and is used to model probability vectors over a finite number of classes.

[0036] use Figure 2 and Figure 3The example illustrated below illustrates how target domain data is smoothed using a Gaussian distribution, and then sliced ​​Wasserstein optimal transfer is learned to transfer the target domain to the source domain for inference by a model trained on the source domain. This example also discusses fine-tuning the model using smoothed target domain data (e.g., few-shot learning) to enable the model to infer more directly from data in the target domain. As mentioned above, the effectiveness of ML often depends primarily on the quantity and quality of data available for training. However, in many real-world scenarios, acquiring such data can be prohibitively expensive, time-consuming, or impractical, limiting the development of accurate ML-based inference models. Transfer learning addresses this constraint by adapting a model trained in a data-rich source domain to a data-sparse target domain.

[0037] In sensor-based monitoring systems, inference techniques can be deployed in diverse environments, each with varying data characteristics. Applying transfer learning in such settings reduces the need for independent data acquisition or model training (e.g., retraining) in each context. This reduction lowers deployment costs and accelerates development. The following example illustrates how transfer learning can be used to estimate battery core temperature using electrochemical impedance spectroscopy (EIS) measurements, for instance, when limited data is available for a target domain with a different battery structure than the source domain. In one example, this method has been used to address the problem of estimating battery core temperature using EIS data where sufficient training data exists for one battery type, but limited data is available for batteries with different physical and chemical properties.

[0038] Transfer learning for estimation tasks can be formulated as follows. Let This represents the distributions corresponding to the source and target domains, respectively. and The observation vector extracted from the data. Let... From The inferred properties are assumed to have a distribution. Identical in both domains. Definition and the corresponding joint distribution in the source and target domains and Here, it is assumed that sufficient source domain data is available to regress the data. Training to It achieves acceptable accuracy, but only a limited amount of target domain data exists. The goal of the transfer learning process is to adapt... From the sample Estimate y. In this example, a two-step approach is used. First, learn the transformation. To take source domain samples Map to the target domain. Secondly, Retrain using both the transformed source data and the available target domain data.

[0039] In order to learn transformations It employs an optimal transmission (OT) framework. OT search will improve quality from... Mapped to Transmission plan To minimize the expected squared Euclidean distance:

[0040]

[0041] The constraint is as follows: and yes The marginal distribution.

[0042] When both distributions are Gaussian, OT produces a linear mapping:

[0043]

[0044]

[0045] in and denoted as the mean and covariance of the source and target distributions, respectively.

[0046] If the data is better modeled as a Gaussian mixture, the OT solution can be obtained by computing the pairwise linear transport mapping between the components of the source mixture and the target mixture. To expand. Samples from the source can be used with these transformations to increase probability. Mapped to the target, this probability is determined by solving the discrete OT problem:

[0047]

[0048]

[0049] in and It is a mixed weight, and It is the Wasserstein distance between the i-th and j-th Gaussian components in the source and target domains. This method can be called Linear Optimal Transport Mixing (MOLT). In one example, it can be used for... Additional constraints are applied to control the flexibility of the MOLT mapping. The probabilities represented are in a smooth form, where synthetic data points are generated during transport computation. Here, the generation of synthetic data can be considered "transient" because they do not need to be recorded or added to a data structure that holds the target domain data, but exist to directly affect the mapping.

[0050] Figure 4An example of machine learning according to one implementation is illustrated, which is used to generate a transfer from latent data space embeddings of target domain data to latent data space embeddings of source domain data. As illustrated, source domain data 405 is encoded (e.g., compressed by an encoder) to create embedded source data 410. Similarly, target domain data 415 is encoded (e.g., using the same encoder used to encode source domain data 405) to create embedded target data. An optimal transfer process (e.g., a process as described above) is then applied to learn a transfer mapping 425 from embedded target data 420 to embedded source data 410, or vice versa. Operations on the latent data space can reduce components because the latent data space is typically a lower-dimensional space than the original data space (e.g., the source domain).

[0051] The directionality of the transfer mapping 425 can be used for various purposes. For example, mapping the source domain to the target domain allows source domain data to supplement target domain data for training or retraining an ML model on the target domain. This example is illustrated in model fine-tuning 430. Here, source domain data 435 is transferred to target domain 440 and combined with target domain data 445 to fine-tune the inference model 450. This approach can be useful, for example, when using additional target domain data over time to continue optimizing the model specifically for the target domain.

[0052] Conversely, mapping the target domain to the source domain allows the ML model to operate on the transmitted target domain data as if it were source domain data to provide inference. This approach can be useful when, for example, the same model will be applied (e.g., infrequently) to different target domains. In such cases, when the target domain data can be simply transmitted to the source domain to obtain results from the ML model, the additional work of fine-tuning the model for each domain is unlikely to be worth the time or expense.

[0053] Figure 5 An example is illustrated using an ML learning model trained on source domain data to infer target domain data 505 according to one implementation. Target domain data 505 (e.g., ATE test data) may be obtained (e.g., retrieved, received, delivered, etc.) by an ML pipeline (e.g., implemented in an on-chip accelerator 525). In the illustrated example, an encoder is invoked on the target domain data 505 to create embedded target data 510. A transfer mapping (e.g., including smoothing of the embedded target data 510 as described above) transfers the embedded target data 510 to the source domain to create an embedded representation 515 operable by the ML model (e.g., accepted as input) to produce result 520. This configuration utilizes an on-chip model trained on the source domain for the target domain.

[0054] Figure 6An example according to one implementation is illustrated, in which a machine learning model trained on source domain data transformed to the target domain using a transport mapping is used to infer target domain data. In this implementation, the ML model is retrained for the target domain using data from the source domain that has been transported to the target domain. The retrained model is exemplified as follows. Here, the target data can be directly processed by the retrained model. Therefore, target domain data 605 is obtained and encoded into the latent data space to create embedded target data 610. The retrained model then operates directly on the target data embedding to produce result 620.

[0055] Figure 7 A flowchart illustrating an example of a method 700 for transfer learning in a machine learning model according to one embodiment is provided. The operation of method 700 is performed by computer hardware, such as the computer hardware described above or below (e.g., a processing circuit system).

[0056] At operation 705, access is made to the medium containing data from the target domain. In one example, this data is sparse in the target domain. Here, "sparse" means that the data includes gaps (e.g., missing information) within the target domain that exceed a threshold.

[0057] At operation 710, a transfer procedure is performed on the data to move the data to the source domain and create input data.

[0058] The transfer process includes operations 715 to 725. At operation 715, synthetic data is created based on a portion of the data and constraints. In one example, where the data in the target domain is sparse, the synthetic data fills a portion of the gaps in the source data. In one example, this portion of data is located at the edge of the gap. In this example, creating synthetic data based on this portion of data involves smoothing the portion of data to fill the gap.

[0059] In one example, the constraint is the distribution of values. In one example, the distribution is a Gaussian distribution. In one example, the distribution is based on one dimension of the source domain. In one example, creating synthetic data based on constraints involves maintaining the distribution while creating new values ​​in one dimension of the synthetic data. In one example, the constraint is a configurable parameter obtained from configuration or the user interface.

[0060] In one example, the operation of method 700 may include invoking an encoder on source or intermediate data to produce an embedding of the input data into a latent data space. In one example, the transport process accepts this embedding as input. In one example, the ML model used in operation 730 is configured to perform inference on the latent data space to produce a result.

[0061] At operation 720, the synthetic data is combined with this data to create intermediate data.

[0062] At operation 725, a transformation is applied to the intermediate data to move it to the source domain. In one example, this transformation is created during the training process to map data points from the target domain to their corresponding data points in the target domain. In one example, Sliced ​​Wasserstein Transfer (SWT) is used as the loss function during the training process to obtain the optimal transfer from the target domain to the source domain.

[0063] At operation 730, a machine learning model is invoked on the input data to obtain the result. This machine learning model is trained on data from the source domain to produce the output.

[0064] In one example, the source domain is a first set of measurements for the first product, and the target domain is a second set of measurements for the second product. In one example, the first and second products are different products of the same type. In one example, the same type is a battery, and the first product has a different chemical composition or a different shape factor than the second product. In one example, the first and second sets of measurements include corresponding electrochemical impedance spectroscopy (EIS) data, voltage, current, or state of charge data for the first and second products. In one example, the result is the core temperature of the battery.

[0065] Figure 8A block diagram of an example machine 800 is illustrated, on which any one or more techniques (e.g., methods) discussed herein may be implemented. As described herein, the example may include, or be operated by, logic or multiple components or mechanisms in machine 800. A circuit system (e.g., a processing circuit system) is a collection of circuits implemented in a tangible entity of machine 800, including hardware (e.g., simple circuits, gates, logic, etc.). The composition of the members of a circuit system may change flexibly over time. A circuit system includes members that can perform a specified operation individually or in combination during operation. In one example, the hardware of the circuit system may be immutably designed to perform a specific operation (e.g., hardwired). In one example, the hardware of the circuit system may include physically connected components (e.g., execution units, transistors, simple circuits, etc.) including physically modified machine-readable media (e.g., by magnetic means, electrical means, changing the movable placement of particles with invariant mass, etc.) to encode instructions for a specific operation. When connecting physical components, the fundamental electrical properties of the hardware components are changed, for example, from an insulator to a conductor, or vice versa. Instructions enable embedded hardware (e.g., an execution unit or loading mechanism) to create members of a circuit system within the hardware via variable connections to perform portions of a specific operation during operation. Thus, in one example, a machine-readable medium element is either part of the circuit system or communicatively coupled to other components of the circuit system when the device is operating. In one example, any of the physical components can be used in more than one member of more than one circuit system. For example, during operation, an execution unit may be used at one point in time in a first circuit of a first circuit system and reused at a different time by a second circuit in the first circuit system or by a third circuit in the second circuit system. Additional examples of these components of machine 800 are given below.

[0066] In alternative implementations, machine 800 may operate as a standalone device or be connected (e.g., networked) to other machines. In a networked deployment, machine 800 may operate as a server machine, a client machine, or both in a server-client network environment. In one example, machine 800 may act as a peer-to-peer (P2P) or other distributed network environment. Machine 800 may be a personal computer (PC), tablet PC, set-top box (STB), personal digital assistant (PDA), mobile phone, network appliance, network router, switch, or bridge, or any machine capable of executing instructions (sequentially or otherwise) specifying actions to be taken by that machine. Furthermore, while only a single machine is illustrated, the term "machine" should also be understood to include any collection of machines that individually or jointly execute a set (or more) of instructions to perform any or more of the methods discussed herein, such as cloud computing, Software as a Service (SaaS), and other computer cluster configurations.

[0067] Machine (e.g., computer system) 800 may include hardware processor 802 (e.g., central processing unit (CPU), graphics processing unit (GPU), hardware processor core, or any combination thereof), main memory 804, static memory (e.g., memory or storage device for firmware, microcode, basic input / output (BIOS), unified extensible firmware interface (UEFI), etc.) 806, and mass storage device 808 (e.g., hard disk drive, tape drive, flash memory device, or other block device), some or all of which may communicate with each other via interconnect (e.g., bus) 830. Machine 800 may further include display unit 810, alphanumeric input device 812 (e.g., keyboard), and user interface (UI) navigation device 814 (e.g., mouse). In one example, display unit 810, input device 812, and UI navigation device 814 may be a touchscreen display. Machine 800 may additionally include a storage device (e.g., a drive unit) 808, a signal generation device 818 (e.g., a speaker), a network interface device 820, and one or more sensors 816, such as a Global Positioning System (GPS) sensor, a compass, an accelerometer, or other sensors. Machine 800 may include an output controller 828, such as a serial (e.g., Universal Serial Bus (USB)), parallel, or other wired or wireless (e.g., infrared (IR), near field communication (NFC), etc.) connection, for communicating or controlling one or more peripheral devices (e.g., a printer, a card reader, etc.).

[0068] The registers of processor 802, main memory 804, static memory 806, or mass storage device 808 may be or include machine-readable medium 822, on which one or more sets of data structures or instructions 824 (e.g., software) embodying any one or more of the technologies or functions described herein or used by such technologies or functions are stored. During execution of the instructions 824 by machine 800, the instructions may also reside wholly or at least partially in any register of processor 802, main memory 804, static memory 806, or mass storage device 808. In one example, one or any combination of hardware processor 802, main memory 804, static memory 806, or mass storage device 808 may constitute machine-readable medium 822. While machine-readable medium 822 is illustrated as a single medium, the term "machine-readable medium" may include a single medium or multiple media (e.g., a centralized or distributed database and / or associated caches and servers) configured to store one or more instructions 824.

[0069] The term "machine-readable medium" can include any medium capable of storing, encoding, or carrying instructions for execution by machine 800 and enabling machine 800 to perform any one or more of the technologies disclosed herein, or any medium capable of storing, encoding, or carrying data structures used by or associated with such instructions. Examples of non-limiting machine-readable media can include solid-state memory, optical media, magnetic media, and signals (e.g., radio frequency signals, other photon-based signals, sound signals, etc.). In one example, a non-transitory machine-readable medium includes a machine-readable medium having multiple particles with invariant (e.g., rest) masses, and thus being a combination of matter. Therefore, a non-transitory machine-readable medium is a machine-readable medium that does not include transiently propagating signals. Specific examples of non-transitory machine-readable media can include: non-volatile memory, such as semiconductor memory devices (e.g., electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM)) and flash memory devices; disks, such as internal hard disks and removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks.

[0070] In one example, information stored on or otherwise provided on a machine-readable medium 822 may represent instructions 824, such as instructions 824 themselves or a format from which instructions 824 can be derived. Such a format from which instructions 824 can be derived may include source code, encoded instructions (e.g., in compressed or encrypted form), packaged instructions (e.g., divided into multiple packages), etc. Information representing instructions 824 in the machine-readable medium 822 may be processed by a processing circuitry system into instructions to perform any of the operations discussed herein. For example, deriving instructions 824 from information (e.g., processed by a processing circuitry system) may include: compiling the information (e.g., from source code, object code, etc.), interpreting, loading, organizing (e.g., dynamically or statically linking), encoding, decoding, encrypting, decrypting, packaging, unpacking, or otherwise manipulating the information into instructions 824.

[0071] In one example, deriving instruction 824 may include assembling, compiling, or interpreting information (e.g., by a processing circuit system) to create instruction 824 from some intermediate or preprocessed format provided by machine-readable medium 822. When provided in multiple parts, information can be combined, unpacked, and modified to create instruction 824. For example, information may be in multiple compressed source code packages (or object code, or binary executable code, etc.) on one or more remote servers. The source code packages may be encrypted when transmitted over a network and may be decrypted, decompressed, assembled (e.g., linked) if necessary, and compiled or interpreted on the local machine (e.g., compiled or interpreted into a library, standalone executable, etc.) and executed by the local machine.

[0072] Instruction 824 can be further transmitted or received on communication network 826 via a transmission medium through network interface device 820 using any of a variety of transmission protocols (e.g., Frame Relay, Internet Protocol (IP), Transmission Control Protocol (TCP), User Datagram Protocol (UDP), Hypertext Transfer Protocol (HTTP), etc.). Exemplary communication networks may include local area networks (LANs), wide area networks (WANs), packet data networks (e.g., the Internet), LoRa / LoRaWAN or satellite communication networks, mobile phone networks (e.g., cellular networks, such as cellular networks compliant with 3G, 4G LTE / LTE-A or 5G standards), conventional telephone (POTS) networks, and wireless data networks (e.g., IEEE 802.11 family of standards (known as Wi-Fi®), IEEE 802.15.4 family of standards, peer-to-peer (P2P) networks, etc.). In one example, network interface device 820 may include one or more physical jacks (e.g., Ethernet, coaxial, or telephone jacks) or one or more antennas for connecting to communication network 826. In one example, network interface device 820 may include multiple antennas for wireless communication using at least one of single-input multiple-output (SIMO), multiple-input multiple-output (MIMO), or multiple-input single-output (MISO) technology. The term "transmission medium" should be understood to include any intangible medium capable of storing, encoding, or carrying instructions executable by machine 800, and includes digital or analog communication signals or other intangible media used to facilitate communication of such software. The transmission medium is a machine-readable medium.

[0073] Additional notes and examples

[0074] Example 1 is an apparatus for transfer learning in a machine learning (ML) model, the apparatus comprising: an interface configured to access a medium including data in a target domain; a memory including instructions; and a processing circuitry configured by the instructions, in operation, to: perform a transfer process on data to move the data to a source domain and create input data, the transfer process including: creating synthetic data based on a portion of the data and constraints; combining the synthetic data with the data to create intermediate data; applying a transformation to the intermediate data to move the intermediate data to the source domain; and invoking an ML model on the input data to obtain a result, the ML model being trained on data from the source domain to produce an output.

[0075] In Example 2, the subject matter of Example 1 is used, wherein the data is sparse in the target domain and is non-uniformly sampled in the target domain label space, including gaps exceeding a threshold.

[0076] In Example 3, the subject matter of Example 2 is used, wherein synthetic data fills a portion of the gaps in the target domain.

[0077] In Example 4, the subject matter of Example 3 is used, wherein the data portion is located at the edge of the gap, and wherein, in order to create synthetic data based on the data portion, the processing circuitry is configured to smooth the data portion to fill the gap.

[0078] In Example 5, the subject matter of any one of Examples 1 to 4, wherein the constraint is the distribution of values.

[0079] In Example 6, the subject matter of Example 5 is followed, wherein, in order to create synthetic data based on constraints, the processing circuitry is configured to maintain the distribution when creating new values ​​in one dimension of the synthetic data.

[0080] In Example 7, the subject matter of any one of Examples 1 to 6, wherein the constraints are configurable parameters obtained from the configuration or user interface.

[0081] In Example 8, the subject content of any one of Examples 1 to 7, wherein the transformation is created through a training process to map data points from the target domain to corresponding data points in the target domain.

[0082] In Example 9, the subject matter of Example 8 is repeated, wherein Slice Wasserstein Transport (SWT) is used as a loss function during the training process to obtain the optimal transport from the target domain to the source domain.

[0083] In Embodiment 10, the subject matter of any one of Embodiments 1 to 9 is included, wherein the processing circuitry is configured to invoke an encoder on data or intermediate data to generate an embedding of input data into a potential data space, wherein the transmission process accepts the embedding as input.

[0084] In Example 11, the subject matter of Example 10 is followed, wherein the ML model is configured to perform inference on the potential data space to produce results.

[0085] In Example 12, the subject matter of any one of Examples 1 to 11 is used, wherein the source domain is a first set of measurements for a first product and the target domain is a second set of measurements for a second product.

[0086] In Example 13, the subject matter of Example 12 is the same, wherein the first product and the second product are different products of the same type.

[0087] In Example 14, the subject matter of Example 13 is the same type, wherein the first product has a different chemical composition or a different shape factor than the second product.

[0088] In Example 15, the subject matter of Example 14 is as follows, wherein the first set of measurements and the second set of measurements include corresponding electrochemical impedance spectroscopy (EIS) data, voltage, current or state of charge data for the first product and the second product.

[0089] In Example 16, the subject matter of Example 15 is discussed, where the result is the core temperature of the battery.

[0090] Example 17 is a method for transfer learning in a machine learning (ML) model, the method comprising: accessing a medium including data in a target domain; performing a transfer process on the data to move the data to a source domain and create input data, the transfer process comprising: creating synthetic data based on a portion of the data and constraints; combining the synthetic data with the data to create intermediate data; applying a transformation to the intermediate data to move the intermediate data to the source domain; and calling an ML model on the input data to obtain a result, the ML model being trained on the data from the source domain to produce an output.

[0091] In Example 18, the subject matter of Example 17 is used, wherein the data is sparse in the target domain and is non-uniformly sampled in the target domain label space, including gaps exceeding a threshold.

[0092] In Example 19, the subject matter of Example 18 is used, wherein synthetic data fills a portion of the gaps in the target domain.

[0093] In Example 20, the subject matter of Example 19 is used, wherein the data portion is located at the edge of the gap, and wherein creating synthetic data based on the data portion includes smoothing the data portion to fill the gap.

[0094] In Example 21, the subject matter of any one of Examples 17 to 20 is used, wherein the constraint is the distribution of values.

[0095] In Example 22, the subject matter of Example 21 is described, wherein creating synthetic data based on constraints includes maintaining the distribution when creating new values ​​in one dimension of the synthetic data.

[0096] In Example 23, the subject matter of any one of Examples 17 to 22, wherein the constraints are configurable parameters obtained from the configuration or user interface.

[0097] In Example 24, the subject content of any one of Examples 17 to 23, wherein the transformation is created through a training process to map data points from the target domain to corresponding data points in the target domain.

[0098] In Example 25, the subject matter of Example 24 is carried out, wherein Sliced ​​Wasserstein Transport (SWT) is used as a loss function during the training process to obtain the optimal transport from the target domain to the source domain.

[0099] In Example 26, the subject matter of any one of Examples 17 to 25 includes invoking an encoder on data or intermediate data to generate an embedding of input data into a potential data space, wherein the transmission process accepts the embedding as input.

[0100] In Example 27, the subject matter of Example 26 is followed, wherein the ML model is configured to perform inference on the potential data space to produce results.

[0101] In Example 28, the subject matter of any one of Examples 17 to 27 is used, wherein the source domain is a first set of measurements for a first product and the target domain is a second set of measurements for a second product.

[0102] In Example 29, the subject matter of Example 28 is the same, wherein the first product and the second product are different products of the same type.

[0103] In Example 30, the subject matter of Example 29 is the same type, wherein the first product has a different chemical composition or a different shape factor than the second product.

[0104] In Example 31, the subject matter of Example 30 is as follows, wherein the first set of measurements and the second set of measurements include corresponding electrochemical impedance spectroscopy (EIS) data, voltage, current or state of charge data for the first product and the second product.

[0105] In Example 32, the subject matter of Example 31 is discussed, where the result is the core temperature of the battery.

[0106] Example 33 is a system that includes means for performing the method of any one of Examples 17 to 32.

[0107] Example 34 is a machine-readable medium including instructions that, when executed by a processing circuitry system, cause the processing circuitry system to perform the method of any one of Examples 17 to 32.

[0108] Example 35 is a machine-readable medium including instructions for transfer learning in a machine learning (ML) model, which, when executed by a processing circuitry system, cause the processing circuitry system to perform operations including: accessing the medium containing data in a target domain; performing a transfer process on the data to move the data to a source domain and create input data, the transfer process including: creating synthetic data based on a portion of the data and constraints; combining the synthetic data with the data to create intermediate data; applying a transformation to the intermediate data to move the intermediate data to the source domain; and invoking an ML model on the input data to obtain a result, the ML model being trained on data from the source domain to produce an output.

[0109] In Example 36, the subject matter of Example 35 is used, wherein the data is sparse in the target domain and is non-uniformly sampled in the target domain label space, including gaps exceeding a threshold.

[0110] In Example 37, the subject matter of Example 36 is used, wherein synthetic data fills a portion of the gaps in the target domain.

[0111] In Example 38, the subject matter of Example 37 is used, wherein the data portion is located at the edge of the gap, and wherein creating synthetic data based on the data portion includes smoothing the data portion to fill the gap.

[0112] In Example 39, the subject matter of any one of Examples 35 to 38, wherein the constraint is the distribution of values.

[0113] In Example 40, the subject matter of Example 39 is described, wherein creating synthetic data based on constraints includes maintaining the distribution when creating new values ​​in one dimension of the synthetic data.

[0114] In Example 41, the subject matter of any one of Examples 35 to 40, wherein the constraints are configurable parameters obtained from the configuration or user interface.

[0115] In Example 42, the subject content of any one of Examples 35 to 41, wherein the transformation is created through a training process to map data points from the target domain to corresponding data points in the target domain.

[0116] In Example 43, the subject matter of Example 42 is repeated, wherein Sliced ​​Wasserstein Transport (SWT) is used as a loss function during the training process to obtain the optimal transport from the target domain to the source domain.

[0117] In embodiment 44, the subject matter of any one of embodiments 35 to 43 is included, wherein the operation includes invoking an encoder on data or intermediate data to generate an embedding of input data into a potential data space, wherein the transmission process accepts the embedding as input.

[0118] In Example 45, the subject matter of Example 44 is described, wherein the ML model is configured to perform inference on the potential data space to produce results.

[0119] In Example 46, the subject matter of any one of Examples 35 to 45 is used, wherein the source domain is a first set of measurements for the first product and the target domain is a second set of measurements for the second product.

[0120] In Example 47, the subject matter of Example 46 is the same, wherein the first product and the second product are different products of the same type.

[0121] In Example 48, the subject matter of Example 47 is the same type, wherein the first product has a different chemical composition or a different shape factor than the second product.

[0122] In Example 49, the subject matter of Example 48 is as follows, wherein the first set of measurements and the second set of measurements include corresponding electrochemical impedance spectroscopy (EIS) data, voltage, current, or state of charge data for the first product and the second product.

[0123] In Example 50, the subject matter of Example 49 is discussed, where the result is the core temperature of the battery.

[0124] Example 51 is a system for transfer learning in a machine learning (ML) model, comprising: means for accessing a medium including data in a target domain; means for performing a transfer process on the data to move the data to a source domain and create input data, the transfer process including: creating synthetic data based on a portion of the data and constraints; combining the synthetic data with the data to create intermediate data; and applying a transformation to the intermediate data to move the intermediate data to the source domain; and means for calling an ML model on the input data to obtain a result, the ML model being trained on the data from the source domain to produce an output.

[0125] In Example 52, the subject matter of Example 51 is used, wherein the data is sparse in the target domain and is non-uniformly sampled in the target domain label space, including gaps exceeding a threshold.

[0126] In Example 53, the subject matter of Example 52 is used, wherein synthetic data fills a portion of the gaps in the target domain.

[0127] In embodiment 54, the subject matter of embodiment 53 is as follows, wherein the portion of data is located at the edge of the gap, and wherein the means for creating synthetic data based on the portion of data includes means for smoothing the portion of data to fill the gap.

[0128] In Example 55, the subject matter of any one of Examples 51 to 54, wherein the constraint is the distribution of values.

[0129] In embodiment 56, the subject matter of embodiment 55 is further described, wherein the means for creating synthetic data based on constraints includes means for maintaining the distribution when creating new values ​​in one dimension of the synthetic data.

[0130] In Example 57, the subject matter of any one of Examples 51 to 56, wherein the constraints are configurable parameters obtained from the configuration or user interface.

[0131] In Example 58, the subject content of any one of Examples 51 to 57, wherein the transformation is created through a training process to map data points from the target domain to corresponding data points in the target domain.

[0132] In Example 59, the subject matter of Example 58 is followed, wherein Slice Wasserstein Transport (SWT) is used as a loss function during the training process to obtain the optimal transport from the target domain to the source domain.

[0133] In embodiment 60, the subject matter of any one of embodiments 51 to 59 includes means for invoking an encoder on data or intermediate data to generate an embedding of input data into a potential data space, wherein the transmission process accepts the embedding as input.

[0134] In Example 61, the subject matter of Example 60 is described, wherein the ML model is configured to perform inference on the potential data space to produce results.

[0135] In Example 62, the subject matter of any one of Examples 51 to 61 is as follows: the source domain is a first set of measurements for the first product, and the target domain is a second set of measurements for the second product.

[0136] In Example 63, the subject matter of Example 62 is the same, wherein the first product and the second product are different products of the same type.

[0137] In Example 64, the subject matter of Example 63 is the same type, wherein the first product has a different chemical composition or a different shape factor than the second product.

[0138] In Example 65, the subject matter of Example 64 is as follows, wherein the first set of measurements and the second set of measurements include corresponding electrochemical impedance spectroscopy (EIS) data, voltage, current, or state of charge data for the first product and the second product.

[0139] In Example 66, the subject matter of Example 65 is discussed, where the result is the core temperature of the battery.

[0140] Example 67 is at least one machine-readable medium comprising instructions that, when executed by a processing circuitry system, cause the processing circuitry system to perform an operation for implementing any one of Examples 1 to 66.

[0141] Example 68 is an apparatus that includes means for implementing any one of Examples 1 to 66.

[0142] Example 69 is a system for implementing any one of Examples 1 to 66.

[0143] Example 70 is a method for implementing any one of Examples 1 to 66.

[0144] The above detailed description includes reference to the accompanying drawings, which form part of this detailed description. The drawings illustrate specific embodiments that can be practiced by way of illustration. These embodiments are also referred to herein as “examples.” Such examples may include elements other than those shown or described. However, the inventors also contemplate examples that provide only those elements shown or described. Furthermore, the inventors contemplate examples that use any combination or arrangement of those elements (or one or more aspects thereof) relative to a particular example (or one or more aspects thereof) or relative to other examples (or one or more aspects thereof) shown or described in the text.

[0145] All publications, patents, and patent documents referenced in this document are incorporated herein by reference in their entirety as if they were individually incorporated. In the event of any inconsistency between the usage in this document and those references so incorporated, the usage in the incorporated references shall be considered supplementary to the usage in this document; in the case of irreconcilable inconsistencies, the usage in this document shall prevail.

[0146] In this document, as is common in patent documents, the terms “a” or “an” are used to include one or more, and are not related to any other instance or use of “at least one” or “one or more”. In this document, unless otherwise specified, the term “or” is used to mean a non-exclusive “or” such that “A or B” includes “A but not B”, “B but not A”, and “A and B”. In the appended claims, the terms “including” and “in which” are used as concise English equivalents to the corresponding terms “comprising” and “wherein”. Furthermore, in the appended claims, the terms “including” and “comprising” are open-ended, meaning that a system, apparatus, article, or process that includes elements other than those listed after such terms in the claims is still considered to fall within the scope of the claims. Additionally, in the appended claims, the terms “first,” “second,” and “third,” etc., are used merely as labels and are not intended to impose numerical requirements on their objects.

[0147] The above description is intended to be illustrative and not restrictive. For example, the examples (or one or more aspects thereof) described above may be used in combination with each other. Other embodiments may be used by those skilled in the art after reading the above description. The abstract is intended to allow the reader to quickly determine the nature of this technical disclosure and is submitted with the understanding that it is not intended to interpret or limit the scope or meaning of the claims. Furthermore, in the detailed description above, various features may be combined to simplify this disclosure. This should not be construed as meaning that any unclaimed disclosed features are necessary for any claim. Rather, the subject matter of the invention may lie in fewer than all features of a particular disclosed embodiment. Therefore, the claims below are hereby incorporated into the detailed description, wherein each claim exists independently as a separate embodiment. Thus, the scope of each embodiment should be determined with reference to the appended claims and the full scope of equivalents enjoyed by such claims.

Claims

1. A method for transfer learning in machine learning (ML) models, the method comprising: The medium through which data in the target domain is accessed; A transmission process is performed on the data to move the data to the source domain and create input data, the transmission process including: Synthetic data is created based on a portion of the data and constraints. The synthetic data is combined with the data to create intermediate data; and Apply a transformation to the intermediate data to move the intermediate data to the source domain; and The input data is used to invoke an ML model to obtain results, the ML model being trained on data from the source domain to produce output.

2. The method according to claim 1, wherein the data is sparse in the target domain, and the data is non-uniformly sampled in the target domain label space, including gaps exceeding a threshold.

3. The method of claim 2, wherein the synthesized data fills a portion of the gap in the target domain.

4. The method of claim 3, wherein the portion of the data is located at the edge of the gap, and wherein creating the synthetic data based on the portion of the data includes smoothing the portion of the data to fill the portion of the gap.

5. The method according to any one of claims 1 to 4, wherein the constraint is the distribution of values.

6. The method of claim 5, wherein creating the synthetic data based on the constraints includes maintaining the distribution when creating new values ​​in one dimension of the synthetic data.

7. The method according to any one of claims 1 to 6, wherein the constraint is a configurable parameter obtained from configuration or user interface.

8. The method according to any one of claims 1 to 7, wherein the transformation is created through a training process to map data points from the target domain to corresponding data points in the target domain.

9. The method of claim 8, wherein the slice Wasserstein transfer (SWT) is used as a loss function in the training process to obtain the optimal transfer from the target domain to the source domain.

10. The method of any one of claims 1 to 9, comprising invoking an encoder on the data or the intermediate data to generate an embedding of the input data into a potential data space, wherein the transfer process accepts the embedding as input.

11. The method of claim 10, wherein the ML model is configured to perform inference on the potential data space to produce the result.

12. The method according to any one of claims 1 to 11, wherein the source domain is a first set of measurements for a first product, and the target domain is a second set of measurements for a second product.

13. The method of claim 12, wherein the first product and the second product are different products of the same type.

14. The method of claim 13, wherein the same type is a battery, and wherein the first product has a different chemical composition or a different shape factor than the second product.

15. The method of claim 14, wherein the first set of measurements and the second set of measurements include corresponding electrochemical impedance spectroscopy (EIS) data, voltage, current, or state of charge data for the first product and the second product.

16. The method of claim 15, wherein the result is the core temperature of the battery.

17. A system comprising means for performing the method of any one of claims 1 to 16.

18. A machine-readable medium comprising instructions that, when executed by a processing circuitry system, cause the processing circuitry system to perform the method of any one of claims 1 to 16.