A method for heterogeneous domain adaptation of multi-sensor data

Abstract

The present invention provides a method for heterogeneous domain adaptation of multi-sensor data. It maps the data to a shared latent space by learning multiple transformations, each domain has an independent transformation, in which the domains are completely consistent, and jointly optimizes multiple transformations in a unified framework through an effective formula. In order to learn robust transformations, a new landmark selection algorithm is proposed to re-weight the samples, that is, increase the weight of the hub samples and reduce the weight of the outliers, which can solve the problems of a large number of various sensor data information in the prior art, such as unlabeled source domain data and different feature dimensions.

Description

Technical Field

[0001] The present invention relates to the technical field of engineering reliability analysis, and in particular to a method for heterogeneous domain adaptation of multi-sensor data. Background Art

[0002] Currently, unsupervised heterogeneous domain adaptation in the most common generalization setting in real-world applications has not been fully explored. Existing methods are either limited to special cases or require labeled target samples for training.

[0003] As for the thermal runaway monitoring of new energy batteries, the large amount of various sensor data information obtained has problems such as unlabeled source domain data and different feature dimensions, which are the problems faced by unsupervised heterogeneous field adaptation. Summary of the invention

[0004] In view of the above problems in the prior art, the present invention proposes a method for heterogeneous domain adaptation of multi-sensor data, aiming to solve the problems of a large amount of multiple sensor data information having unlabeled source domain data and different feature dimensions.

[0005] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A method for heterogeneous domain adaptation of multi-sensor data comprises the following steps:

[0007] S1, use sensors to monitor thermal runaway of new energy batteries and obtain the source sample set matrix X of the collected data S and its labels, and obtain the corresponding target sample set matrix X t , number of sample neighbors, subspace dimension, penalty parameter λ, penalty parameter β and penalty parameter γ; where X represents X S and X t ;

[0008] S2, use the linear kernel to project X, get the core matrix K, initialize G = I, I represents the feature matrix;

[0009] S3, construct the graph structure and obtain the Laplace coefficient matrix L of the graph;

[0010] S4, calculate the maximum mean difference MMD matrix M;

[0011] S5, by solving The projection matrix P is optimized by the eigendecomposition problem in , where Λ is a Lagrange multiplier, is a diagonal matrix, H = I n -(1 / n)1 n , 1 n is an n×n matrix with all elements equal to 1, In Represents the n×n feature matrix I;

[0012] S6, by

[0013] S7, Update X t The label and Laplace coefficient matrix L of the graph;

[0014] S8, update the weight vector α;

[0015] S9, repeat S4–S8 until convergence or maximum iteration;

[0016] S10, output X t The label information of K is t and K s Projected to the learned subspace, that is and Then according to the output X t The label information is coded using support vector machine (SVM) for Z in the subspace. t To classify, Z s is used as a reference; where K t is the target core matrix, K s is the source sample core matrix, p i represents the i-th row of P.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects:

[0018] 1. It only involves simple and fast integer operations, and its time complexity and space complexity are lower than the matrix operations and floating-point operations in existing methods;

[0019] 2. It is applicable to any sample dimension. Different types of sensors can be added or subtracted according to the actual scene needs, and there is no need to mark the sensor data for target sample training;

[0020] 3. It is compatible with sensors of different types and data accuracy, can handle heterogeneous or homogeneous features, independent feature dimensions, multiple fields, and makes HDA a step further towards unsupervised learning (traditional dimensionality reduction methods may lose important information);

[0021] 4. By applying a new method of landmark selection and sample restoring weights, we can learn projections that are more suitable for different sensor datasets, and different projections have the same purpose. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic diagram showing label consistency preservation and sample weighting of the present invention;

[0023] Figure 2A flow chart of the present invention for performing data heterogeneous domain adaptation is shown. DETAILED DESCRIPTION

[0024] The technical solution of the present invention is further described below in conjunction with several embodiments and drawings. It should be noted that the embodiments and terms described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on it.

[0025] The present invention provides a method for heterogeneous domain adaptation of multi-sensor data, which comprises the following steps:

[0026] S1, use sensors to monitor thermal runaway of new energy batteries and obtain the source sample set matrix X of the collected data S and its labels, and obtain the corresponding target sample set matrix X t , number of sample neighbors, subspace dimension, penalty parameter λ, penalty parameter β and penalty parameter γ; where X represents X S and X t ;

[0027] S2, use the linear kernel to project X, get the core matrix K, initialize G = I, I represents the feature matrix;

[0028] S3, construct the graph structure and obtain the Laplace coefficient matrix L of the graph;

[0029] S4, calculate the maximum mean difference MMD matrix M;

[0030] S5, by solving The projection matrix P is optimized by the eigendecomposition problem in , where Λ is a Lagrange multiplier, is a diagonal matrix, H = I n -(1 / n)1 n , 1 n It is an n×n matrix with all elements equal to 1, and In represents the n×n characteristic matrix I;

[0031] S6, by

[0032] S7, Update X t The label and Laplace coefficient matrix L of the graph;

[0033] S8, update the weight vector α;

[0034] S9, repeat S4–S8 until convergence or maximum iteration;

[0035] S10, output X t The label information of K is t and K s Projected to the learned subspace, that is and Then according to the output X t The label information is coded using support vector machine (SVM) for Z in the subspace. t To classify, Z s is used as a reference; where K t is the target core matrix, K s is the source sample core matrix, p i represents the i-th row of P.

[0036] Furthermore, in step S2, a domain D consists of three parts: a sensor data feature space X, its probability distribution P(X), and a label set Y, where X∈X ο , use the subscript s to denote the source domain and t to denote the target domain, respectively map the source domain and the target domain to a latent space, merge the two projections into one variable and learn them simultaneously to ensure that the two projections map the two domains to a true shared space rather than two spaces with only the same dimensions. From a local point of view, different domains are processed independently by different projections. To achieve this, we define the data matrix X and the projection matrix A as follows:

[0037]

[0038] Furthermore, in step S3, it is assumed that the label information of the target domain is known and the data samples are visualized; in order to achieve structural consistency of the sensor data, instances with the same label are kept in a compact cluster, that is, a sample tends to have the same label as its nearest neighbors; therefore, in the distribution matching process, the sensor data samples from the same category are kept close together, and z is used. j To represent x j In the distribution matching new feature space, the new feature representation is:

[0039]

[0040] Among them, W is a symmetric adjacency matrix, W ij Characterize the like relationship between samples, which can be calculated by various standards such as Gaussian similarity of Euclidean distance, local neighborhood relationship and prior class information in supervised learning algorithm. In the present invention, we use W to characterize the sample relationship and apply cosine similarity to obtain W, as follows:

[0041]

[0042] Among them, N k (x i ) is x i The k nearest neighbors of x ibelong to the same category. Specifically, for the i-th row and j-th column of W, if the sample x j is x i is one of the k nearest neighbors of x i have the same label, then W ij will be set to the cosine distance between the two samples, otherwise W ij will be set to 0; W stores the sample relationship in the original feature space, and z is the new feature representation. Therefore, the adjacent relationship of samples in the new feature space can be retained to alleviate negative migration; You can get:

[0043]

[0044] Where L = DW is the Laplace of the graph, and D is a diagonal matrix whose diagonal elements are calculated as the sum of the i-th row of W, that is,

[0045] Further, in step S4, MMD is used to estimate the distance between distributions because it has the advantage of being non-parametric. Generally speaking, and The MMD between can be calculated as follows:

[0046]

[0047] where φ s (·) and φ t (·) are the feature maps of the source domain and the target domain respectively, n s and n t is the number of samples in the source domain and the target domain. Since sensor data of different categories have heterogeneous features with independent dimensions, they cannot be directly compared. We adopt the method of learning the latent subspace shared by the two domains. In order to handle arbitrary sample dimensions, we learn two trans-A s and A t , one for each domain, mapping the two domains to the latent subspace, so that:

[0048]

[0049] Since step S2 has explained that the two projections are merged into one variable and transferred together and processed locally independently, the above formula can be rewritten as

[0050]

[0051] Where M is the MMD matrix, which can be calculated as

[0052]

[0053] In addition, in step S4, in order to match the first-order and higher-order statistics, the distribution matching is preferably performed in the regenerated Hilbert space RKHS rather than in the original space; considering the kernel mapping ψ:x→ψ(x), ψ(X)=[ψ(x 1 ),...,ψ(x n )], and the kernel matrix We use A = ψ(X)P to kernelize the principal component analysis PCA, thus Can be rewritten as

[0054] Where H is a general RKHS; the linear kernel is used in this invention because it is simple, effective and parameter-free; it should be noted that P = [P s ;P t ] is used as a transformation matrix for kernelized PCA. From the perspective of subspace learning, P also plays the role of projecting K into a shared subspace.

[0055] Further, in step S5, from Figure 1 From the explanation, we can know that landmark selection is reasonable and effective for battery sensor data transfer learning. In fact, landmark selection is sample selection; considering the duality between samples and features, it is necessary not only to rebalance samples but also to select features; since we have learned multiple projections to map high-dimensional features to a shared latent subspace, low-dimensional features may be independent. Generally speaking, each dimension of the subspace is different; if we cannot wisely select the multiple features we are facing, then we may have all the features but still do not know which one is really important for adaptation. Therefore, we introduce the norm l 2,1 Criteria to projection matrix ‖P‖ 2.1 , which can produce row sparsity regularization and help us select shared features.

[0056] Further, in step S5, it is desired to s With X t To achieve knowledge transfer, we do not want to lose the data attributes of the target domain; to avoid information loss, we introduce a PCA-like regularization term to retain the energy of the original target signal, that is, Thus, the equivalent minimization problem can be solved by algebraic steps:

[0057]

[0058] in is a diagonal matrix defined as: if x i ∈X t ,but otherwise

[0059] Then, multiple transformations are jointly learned, one for each domain, to find a shared latent subspace while maintaining the consistency of the sensor data structure, selecting features and weighting samples; the objective function can be expressed as follows:

[0060]

[0061]

[0062] Among them, λ>0, β>0 and γ>0 are penalty parameters, H=I n -(1 / n)1 n and 1 n is an n×n matrix with all elements equal to 1; introduce the constraint This is to avoid the solution of non-sensor data scenarios; according to the constraint optimization theory, we introduce a Lagrangian multiplier Λ and obtain the Lagrangian function of the above formula as follows:

[0063]

[0064] Then take the derivative of this formula with respect to P and set the derivative to 0, and we can get This is how P is optimized.

[0065] Furthermore, in step S6, it should be noted that ‖P‖ 2.1 It is not smooth, so the present invention calculates its sub-gradient G, which is a diagonal matrix whose i-th diagonal element is equal to

[0066]

[0067] where p i Denotes the i-th row of P. By computing the d smallest eigenvectors, we can solve the learning problem of multiple transformations in series. In the new representation Under this condition, the data distribution between domains is shortened, and it can be Train a classifier on Classify and update G.

[0068] Furthermore, in step S9, the algorithm of the present invention discovers the geometric essence of the landmarks and cleverly avoids various computational problems caused by a large number of matrix operations. What needs to be paid attention to is the sample relationship rather than the sample features. For the classification involving weights, a weight vector α∈R is further introduced. ns , and its constraints are as follows:

[0069] α i ∈[0,1],

[0070] And 1 / n s Initialize (each sample has the same weight) and use the vertex v during optimization i The degree of which corresponds to x∈X s , Update α to The weight vector α can be used to train a weighted support vector machine (SVM) model on the source domain.

[0071] In this invention, we learn multiple transformations, one (independent) transformation for each domain, mapping the data to a shared latent space where the domains are completely consistent. Multiple transformations are jointly optimized in a unified framework through an efficient formula. In addition, in order to learn robust transformations, a new landmark selection algorithm is proposed to reweight the samples, i.e., increase the weight of hub samples and reduce the weight of outliers. This landmark selection is based on graph optimization and focuses on the geographic metric relationships of samples rather than sample features. In this invention, we use bold lowercase letters to represent vectors and bold uppercase letters to represent matrices. In particular, I represents the feature matrix. For a matrix M, its norm l 2,1 -norm (used for feature selection) is defined as We also use the Frobenius norm where δ i (M) is the first singular value of the matrix M. A sample is represented by a vector. In addition, using X s ,Y s represents the source sample / label, X t ,y t represents the target sample / label, X represents X S and X t , L represents the Laplace coefficient of the graph, M represents the MMD matrix, P represents the projection matrix, K represents the core matrix, and λ, β, γ represent penalty parameters.

[0072] The embodiments described above provide a detailed description of the technical solutions of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements or similar substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for heterogeneous domain adaptation of multi-sensor data, characterized in that: It includes the following steps: S1, use sensors to monitor thermal runaway of new energy batteries and obtain the source sample set matrix X of the collected data S and its labels, and obtain the corresponding target sample set matrix X t , number of sample neighbors, subspace dimension, penalty parameter λ, penalty parameter β and penalty parameter γ; where, S2, use the linear kernel projection X to obtain the core matrix K, initialize the sub-gradient G = I of the matrix P, I represents the feature matrix; S3, construct the graph structure and obtain the Laplace coefficient matrix L of the graph; S4, calculate the maximum mean difference MMD matrix M; S5, by solving The projection matrix P is optimized by the eigendecomposition problem in , where Λ is a Lagrange multiplier, is a diagonal matrix, H = I n -(1 / n)1 n , 1 n is an n×n matrix with all elements equal to 1, I n Represents the n×n feature matrix I; S6, by S7, Update X t The label and Laplace coefficient matrix L of the graph; S8, update the weight vector α; S9, repeat S4–S8 until convergence or maximum iteration; S10, output X t The label information of K is t and K s Projected to the learned subspace, that is and Then according to the output X t The label information is coded using support vector machine (SVM) for Z in the subspace. t To classify, Z s is used as a reference; where K t is the target core matrix, K s is the source sample core matrix, p i represents the i-th row of P.

2. The method for heterogeneous domain adaptation of multi-sensor data according to claim 1, characterized in that: In step S2, when linear kernel projection is used, a domain D consists of three parts: sensor data feature space X, its probability distribution P(X), and a label set Y, where X∈X ° , use the subscript s to represent the source domain, and use t to represent the target domain. Map the source domain and the target domain to a latent space respectively, merge the two projections of their mapping into one variable and learn them simultaneously to ensure that the two projections map the two domains to a real shared space, rather than two spaces with only the same dimension. In order to achieve independent processing of different domains by different projections, define the projection matrix A as follows: A=[A s ;A t ]。 3. The method for heterogeneous domain adaptation of multi-sensor data according to claim 1, characterized in that: In step S3, assuming that the label information of the target domain is known, the data samples are visualized; instances with the same label are kept in a compact cluster, that is, a sample tends to have the same label as its nearest neighbor; sensor data samples from the same category are kept close in the distribution matching process, using z j To represent x j In the distribution matching new feature space, the new feature representation is: Among them, W is a symmetric adjacency matrix, W ij Characterize the like relationship between samples, use W to characterize the sample relationship, and apply cosine similarity to get W, as follows: Among them, N k (x i ) is x i The k nearest neighbors of x i belong to the same category. Specifically, for the i-th row and j-th column of W, if the sample x j is x i One of the k nearest neighbors of x i have the same label, then W ij is set to the cosine distance between two samples, otherwise W ij is set to 0; W stores the sample relationship in the original feature space, and z is the new feature representation. Therefore, the adjacent relationship of samples in the new feature space can be retained to alleviate negative migration; since Z = PTK, we can get: Where L = DW is the Laplace of the graph, and D is a diagonal matrix whose diagonal elements are calculated as the sum of the i-th row of W, that is, 4. The method for heterogeneous domain adaptation of multi-sensor data according to claim 1, characterized in that: In step S4, MMD is used to estimate the distance between the distributions of the two data sets. and The MMD between is calculated as follows: where φ s (·) and φ t (·) are the feature maps of the source domain and the target domain respectively, n s and n t is the number of samples in the source domain and the target domain. Since sensor data of different categories have heterogeneous features with independent dimensions, they cannot be directly compared. We adopt the method of learning the latent subspace shared by the two domains. In order to handle arbitrary sample dimensions, we learn two trans-A s and A t , one for each domain, mapping the two domains to the latent subspace, so that: The two projections are combined into one variable and transferred together and processed locally independently, so the above formula can be rewritten as Where M is the MMD matrix, which can be calculated as 5. The method for heterogeneous domain adaptation of multi-sensor data according to claim 4, characterized in that: In step S4, distribution matching is performed in the regenerated Hilbert space RKHS in order to match the first-order and higher-order statistics; considering the kernel mapping ψ:x→ψ(x), ψ(X)=[ψ(x1),...,ψ(x n )], and the kernel matrix A = ψ(X)P is used to kernelize the principal component analysis PCA, thereby Rewrite as where H is a general RKHS; P = [P s ;P t ] is used as a transformation matrix for kernelized PCA. From the perspective of subspace learning, P plays the role of projecting K into a shared subspace.

6. The method for heterogeneous domain adaptation of multi-sensor data according to claim 1, characterized in that: In step S5, considering the duality between samples and features, it is necessary not only to rebalance the samples but also to select the features; since multiple projections are learned to map high-dimensional features to a shared latent subspace, low-dimensional features may be independent. In order to clarify the features that are really important for adaptation, the norm l is introduced 2,1 Criteria to projection matrix ||P|| 2，1 , to produce row sparsity regularization and help select shared features.

7. The method for heterogeneous domain adaptation of multi-sensor data according to claim 6, characterized in that: In step S5, it is desired to s With X t To achieve knowledge transfer, we do not want to lose the data attributes of the target domain; therefore, in order to avoid information loss, a regularization term is introduced: To preserve the energy of the original target signal, the equivalent minimization problem is achieved through algebraic steps: in is a diagonal matrix defined as: if x i ∈X t ,but otherwise We then jointly learn multiple transformations, one for each domain, to find a shared latent subspace while maintaining the consistency of the sensor data structure, selecting features and weighting samples. The objective function can be expressed as follows: Among them, λ>0, β>0 and γ>0 are penalty parameters, H=I n -(1 / n)1 n and 1 n is an n×n matrix with all elements equal to 1; introduce the constraint This is to avoid the solution of non-sensor data scenarios; according to the constraint optimization theory, a Lagrangian multiplier Λ is introduced, and the Lagrangian function is obtained as follows: Then take the derivative of this formula with respect to P and set the derivative to 0, and we can get This is how P is optimized.

8. The method for heterogeneous domain adaptation of multi-sensor data according to claim 7, characterized in that: Since the projection matrix ||P|| 2,1 is not smooth, so in step S6, its sub-gradient G is calculated, whose i-th diagonal element is equal to where p i represents the i-th row of P. In the new representation Under this condition, the data distribution between domains is shortened. Train a classifier on Classify and update G.

9. The method for heterogeneous domain adaptation of multi-sensor data according to claim 1, characterized in that: In step S8, for the classification involving weights, a weight vector α∈R is further introduced ns , and its constraints are as follows: And 1 / n s Initialize, set each sample to have the same weight, and use vertex v in the optimization process i The degree of which corresponds to x∈X s , Then update α to The weight vector α is used to train a weighted support vector machine (SVM) model on the source domain.

Images