Data-driven battery thermal fault detection and positioning method

Abstract

According to the data-driven battery thermal fault detection and positioning method provided by the invention, high-efficiency fault detection and positioning can be realized on the premise that accurate partial differential equation description and boundary conditions about a battery thermal process do not need to be established. According to the method, the space-time modeling of the ISOMAP algorithm and the advantages of MICA process monitoring are organically combined, so that the accuracy of battery thermal fault detection can be remarkably improved. And on the basis, the complete coverage of the fault positioning range on the battery system can be realized by utilizing the construction of the space contribution diagram. The method is completely based on data driving, and has good adaptability and reliability for batteries and systems of different specifications and sizes.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery fault detection, and in particular relates to a data-driven battery thermal fault detection and positioning method. Background Art

[0002] As the capacity of lithium-ion batteries used in scenarios such as new energy vehicles and new energy storage equipment and the number of cells they contain continue to increase, the probability of internal battery inconsistencies and lithium plating also increases, making thermal runaway failures and related safety risks difficult to avoid. At present, the detection methods for thermal runaway failures in this field mainly include: sensor-based, lumped model-based, and distributed parameter modeling-based methods, etc. However, these methods still have some defects that affect the accuracy and real-time performance of detection. For example, detection methods based on sensors or lumped models often experience delays; methods based on distributed parameter modeling have a high dependence on the accuracy of the partial differential equation description and boundary conditions used, and most of them are difficult to meet the needs of engineering applications at this stage. Therefore, how to overcome the shortcomings of existing technologies and achieve more accurate and efficient detection and positioning of thermal runaway failures in lithium batteries is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0003] In view of this, and to address the technical problems existing in this field, the present invention provides a data-driven battery thermal fault detection and location method, which specifically includes the following steps:

[0004] Step 1: Perform spatiotemporal decoupling on the battery temperature data collected by the sensor. Then, use the ISOMAP algorithm to reduce the data dimension based on Euclidean distance, converting the spatiotemporal temperature data into a linear combination of time coefficients and discrete spatial basis functions.

[0005] Step 2: Based on the eigenvalue sorting and screening, the linear combination obtained in step 1 is converted into a finite dimension, and the radial basis function with the Gaussian function as the kernel function is used to fit the discrete space basis function to obtain the form of the continuous space basis function;

[0006] Step 3: Use the MICA algorithm to process the time coefficient and obtain the statistics for fault detection, and then set the corresponding fault time domain detection threshold for the statistics;

[0007] Step 4: Use the battery temperature data during normal battery operation to perform the above steps to complete offline training, and obtain stable discrete and continuous space basis functions and fault time domain detection thresholds;

[0008] Step 5: Collect battery temperature data from the newly acquired sensor, obtain the corresponding time coefficient of the discrete space basis function based on offline training, calculate the corresponding statistics online, and compare them with the detection threshold to achieve fault detection;

[0009] Step 6: Use the statistics corresponding to the detected fault samples to establish a battery fault space contribution map, and realize fault location according to the contribution size corresponding to different positions in the contribution map.

[0010] Furthermore, in step 1, the spatiotemporal data collected by the sensor Where n1 and n2 correspond to the number of sensors in the x and y directions respectively. The number of sensors is n = n1 × n2, n t Corresponding to the measurement time dimension, t represents time;

[0011] It is decoupled into the following infinite-dimensional space-time separation form through Fourier transform:

[0012]

[0013] Where, is the time coefficient, is the i-th discrete basis function, which is specifically expressed as a linear combination of spatiotemporal data:

[0014]

[0015] in, is the low-dimensional embedding of the original space, T(x i1 ,y i2 ,t k ) is T(x i1 ,y i2 ,t)’s k-th column vector;

[0016] Define the temperature spatiotemporal data set at time p as T(:,t q ) is T(:,t p ) is one of its K nearest neighbors, and the Euclidean distance between them is d x (p,q), then for each spacetime point k=1,…,n t , calculate T(:,t p ) and T(:,t q ) is the shortest approximate geodesic distance between:

[0017] d G (p,q)=min(d G (p,q),d G (p,k)+d G (k,q)),k=1,…n t

[0018] This creates a matrix Used to represent the shortest approximate geodesic distance between all pairs of points in high-dimensional space;

[0019] The ISOMAP algorithm is executed to obtain a low-dimensional embedding by preserving the shortest geodesic distance between high-dimensional spatiotemporal data points. Its optimization objective is expressed as:

[0020]

[0021] Among them, d N (p,q) are two points in low-dimensional space and γ(:,t p ) and γ(:,t q ), the shortest path distance between is the shortest path distance matrix embedded in low dimension; on this basis, the above optimization objective is transformed into the following minimum optimization problem:

[0022]

[0023] Among them, τ(D G )=-HAH / 2, Represented as {A pq =D Gpq 2}, is the centering matrix, H=I-1 / n t ee T , I is the identity matrix, L represents the matrix 2 norm;

[0024] Define λ i is the matrix τ(D G )’s ith eigenvalue, V i k is the kth component of the i-th eigenvector, and the low-dimensional embedding is calculated

[0025]

[0026] Leveraging low-dimensional embeddings The discrete space basis function can be obtained

[0027] Time coefficient a i (t) is calculated by the following inner product method:

[0028]

[0029] Furthermore, in step 2, the matrix τ(D G ) are sorted in descending order by their eigenvalues The first N of them are selected for dimensionality reduction; N is selected to satisfy η>η0.99, which is calculated by the following formula:

[0030]

[0031] The discrete space basis functions are optimized based on the following objectives to capture their lost spatial information:

[0032]

[0033] stψ i (x,y)∈C r (x,y),i=1,...,N,r=0,1,2...

[0034] Among them, ψ i (x, y) is the i-th continuous space basis function to be designed, C r (x,y) represents the set of functions with continuous derivatives of order r in the x and y directions;

[0035] The continuous space basis function ψ is obtained by fitting the radial basis function with Gaussian function as the kernel function i (x,y), combined with the selected N eigenvalues, the following finite-dimensional battery continuous space temperature estimation can be obtained:

[0036]

[0037] Furthermore, when executing the MICA algorithm in step 3, Considered as a mixed signal, specifically expressed as a mixing matrix With independent component matrix The product of:

[0038] A=BY

[0039] By rearranging:

[0040]

[0041] in, is the demixing matrix, which is also the main estimation target of the MICA algorithm. is the estimated independent component matrix;

[0042] The MICA algorithm process includes:

[0043] First, A is whitened based on principal component analysis to eliminate cross-correlation and the following score matrix is ​​calculated:

[0044] S=U T A

[0045] in, is the covariance matrix E{AA T}=UΛU T , where Λ is a diagonal matrix consisting of N eigenvalues;

[0046] Normalize the score matrix to get the whitened data matrix of A

[0047] Z=Λ -1 / 2 S=Λ -1 / 2 U T A=QA

[0048] Where Q = Λ -1 / 2 U T , E{ZZ T}=I;

[0049] Assign the same variance to each independent component as the corresponding principal component:

[0050]

[0051] in, is the variance matrix, E{CC T}=D=diag(λ1,λ2,…λ m ), m is determined by discarding elements corresponding to low eigenvalues;

[0052] Pair Matrix Normalize and get:

[0053]

[0054] in,

[0055] Using FastICA algorithm to optimize matrix C n , we get the following dominant independent components:

[0056]

[0057] Among them, the demixing matrix Mixing matrix B = UΛ 1 / 2 C n D -1 / 2 , WB=I;

[0058] On this basis, the following Hotelling T are defined respectively 2 Statistics and the squared process error (SPE) statistic are used for fault detection:

[0059]

[0060] in, for The t-th column vector of

[0061]

[0062] in, represents the square of the vector 2 norm, then It can be calculated by the following formula:

[0063]

[0064] in, is the t-th column vector of A;

[0065] Use kernel density estimation to estimate T 2 The confidence bounds of the SPE statistic are used as the corresponding fault time domain detection thresholds.

[0066] Furthermore, in step 5, the newly acquired battery temperature data The fault detection process specifically includes:

[0067] First, using the same discrete space basis function Calculate the new time coefficient matrix

[0068]

[0069] Compute new independent components:

[0070]

[0071] Calculate the corresponding T 2 Statistics:

[0072]

[0073] in, for The t-th column vector of ;

[0074] Calculate the new SPE statistic:

[0075]

[0076] in, Calculated by the following formula:

[0077]

[0078] in, A new The t-th column vector of ;

[0079] The calculated T 2 new and SPE new Compare with the corresponding confidence limits to complete fault detection.

[0080] Furthermore, in step 6, for samples where the detected fault detection statistic is greater than the corresponding confidence limit and

[0081]

[0082] Calculate separately and The average position contribution map of is used as the battery failure space contribution map:

[0083]

[0084] Among them, ψ(x,y)=[ψ1(x,y),ψ2(x,y),…,ψ N (x,y)] T , ||·|| represents the Euclidean norm, |·| represents the absolute value, M and L are and The cardinality, Indicates that in the sample set T on 2 Continuous space contribution plot of a statistic, Indicates that in the sample set Continuous space contribution plot of the upper SPE statistic;

[0085] According to the spatial position with the largest contribution value in the spatial contribution graph and The spatial location of the fault can be located.

[0086] The data-driven battery thermal fault detection and location method provided by the present invention achieves efficient fault detection and location without the need to establish precise partial differential equation descriptions and boundary conditions for battery thermal processes. By organically combining the spatiotemporal modeling of the ISOMAP algorithm with the advantages of MICA process monitoring, the accuracy of battery thermal fault detection can be significantly improved. Furthermore, by constructing a spatial contribution graph, the fault location range can be fully covered for the battery system. The method of the present invention is completely data-driven and has good adaptability and reliability for batteries and systems of different specifications and sizes. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 This is a diagram of the thermal fault detection and location framework of the method provided in the present invention.

[0088] Figure 2 is the continuous space basis function ψ i (x,y) schematic diagram.

[0089] Figure 3 This is a schematic diagram of ISOMAP space mapping.

[0090] Figure 4 Schematic diagram of the ISOMAP spatiotemporal modeling process.

[0091] Figure 5 Schematic diagram of the kernel density estimation probability density function.

[0092] Figure 6 Schematic diagram of the thermal runaway simulation model of the battery system in an example of the present invention.

[0093] Figure 7 This is the result of thermal runaway fault detection in the battery system.

[0094] Figure 8 This is the result of locating the thermal runaway fault in the battery system. DETAILED DESCRIPTION

[0095] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0096] The data-driven battery thermal fault detection and location method provided by the present invention is as follows: Figure 1 As shown, the specific steps include:

[0097] Step 1: Perform spatiotemporal decoupling on the battery temperature data collected by the sensor. Then, use the ISOMAP algorithm to reduce the data dimension based on Euclidean distance, converting the spatiotemporal temperature data into a linear combination of time coefficients and discrete spatial basis functions.

[0098] Step 2: Based on the eigenvalue sorting and screening, the linear combination obtained in step 1 is converted into a finite dimension, and the radial basis function with the Gaussian function as the kernel function is used to fit the discrete space basis function to obtain the form of the continuous space basis function;

[0099] Step 3: Use the MICA algorithm to process the time coefficient and obtain the statistics for fault detection, and then set the corresponding fault time domain detection threshold for the statistics;

[0100] Step 4: Use the battery temperature data during normal battery operation to perform the above steps to complete offline training, and obtain stable discrete and continuous space basis functions and fault time domain detection thresholds;

[0101] Step 5: Collect battery temperature data from the newly acquired sensor, obtain the corresponding time coefficient of the discrete space basis function based on offline training, calculate the corresponding statistics online, and compare them with the fault detection threshold to achieve fault detection;

[0102] Step 6: Use the statistics corresponding to the detected fault samples to establish a battery fault space contribution map, and realize fault location according to the contribution size corresponding to different positions in the contribution map.

[0103] In a preferred embodiment of the present invention, in step 1, the spatiotemporal data collected by the sensor Where n1 and n2 correspond to the number of sensors in the x and y directions respectively. The number of sensors is n = n1 × n2, n t Corresponding to the measurement time dimension, t represents time;

[0104] It is decoupled into the following infinite-dimensional space-time separation form through Fourier transform:

[0105]

[0106] Where, is the time coefficient, is the i-th discrete basis function, which is specifically expressed as a linear combination of spatiotemporal data:

[0107]

[0108] in, is the low-dimensional embedding of the original space, for The k-th column vector of ;

[0109] Define the temperature collection collected at time p T(:,t p ) The neighborhood graph consists of the K points with the closest Euclidean distance. If T(:,t q ) is T(:,t p ) is one of the K nearest neighbors, then T(:,t p ) and T(:,t q ) are connected, define T(:,t p ) and T(:,t q ) The Euclidean distance is d x (p,q), otherwise the two points are not connected.

[0110] For any two points T(:,tp) and T(:,tq) in high-dimensional space, the geodesic distance between the two points can be expressed as Figure 2 The dashed or solid line approximation shown is shown, and it can be seen that the solid line is shorter than the dashed line. If the solid line is the shortest of all paths, it is chosen as the approximation of the true geodesic distance.

[0111] For each space-time point k=1,…,n t , through the following formula T(:,t p ) and T(:,tq ) is the shortest approximate geodesic distance between:

[0112] d G (p,q)=min(d G (p,q),d G (p,k)+d G (k,q)),k=1,…n t

[0113] This creates a matrix Used to represent the shortest approximate geodesic distance between all pairs of points in high-dimensional space;

[0114] The ISOMAP algorithm is executed to obtain a low-dimensional embedding by preserving the shortest geodesic distance between high-dimensional spatiotemporal data points. Its optimization objective is expressed as:

[0115]

[0116] Among them, d N (p,q) are two points in low-dimensional space and γ(:,t p ) and γ(:,t q ), the shortest path distance between is the shortest path distance matrix embedded in low dimension; on this basis, the above optimization objective is transformed into the following minimum optimization problem:

[0117]

[0118] Among them, τ(D G )=-HAH / 2, Represented as {A pq =D Gpq 2}, is the centering matrix, H=I-1 / n t ee T , I is the identity matrix, L represents the matrix 2 norm;

[0119] Define λ i is the matrix τ(D G )’s ith eigenvalue, V i k is the kth component of the i-th eigenvector, and the low-dimensional embedding is calculated

[0120]

[0121] Leveraging low-dimensional embeddings The discrete space basis function can be obtained

[0122] Time coefficient a i (t) is calculated by the following inner product method:

[0123]

[0124] In a preferred embodiment of the present invention, in step 2, the matrix τ(D G ) are sorted in descending order by their eigenvalues The first N of them are selected for dimensionality reduction; N is selected to satisfy η>η0.99, which is calculated by the following formula:

[0125]

[0126] ISOMAP from battery temperature spatiotemporal data Extracted basis functions It is discrete in space. exist Figure 3 The discrete space basis functions are optimized based on the following objectives to capture their lost spatial information:

[0127]

[0128] stψ i (x,y)∈C r (x,y),i=1,...,N,r=0,1,2...

[0129] Among them, ψ i (x, y) is the i-th continuous space basis function to be designed, C r (x,y) represents the set of functions with continuous derivatives of order r in the x and y directions;

[0130] The continuous space basis function ψ is obtained by fitting the radial basis function with Gaussian function as the kernel function i (x,y), combined with the selected N eigenvalues, the following finite-dimensional battery continuous space temperature estimation can be obtained:

[0131]

[0132] Figure 4 The figure shows the entire process of spatiotemporal modeling of battery thermal processes based on the ISOMAP algorithm, which mainly includes the spatiotemporal separation of battery temperature, the construction of continuous space basis functions, and the spatiotemporal synthesis of finite dimensions.

[0133] In a preferred embodiment of the present invention, when executing the MICA algorithm in step 3, Considered as a mixed signal, specifically expressed as a mixing matrix With independent component matrix The product of:

[0134] A=BY

[0135] By rearranging:

[0136]

[0137] in, is the demixing matrix, which is also the main estimation target of the MICA algorithm. is the estimated independent component matrix;

[0138] The MICA algorithm process includes:

[0139] First, A is whitened based on principal component analysis to eliminate cross-correlation and the following score matrix is ​​calculated:

[0140] S=U T A

[0141] in, is the covariance matrix E{AA T}=UΛU T , where Λ is a diagonal matrix consisting of N eigenvalues;

[0142] Normalize the score matrix to get the whitened data matrix of A

[0143] Z=Λ -1 / 2 S=Λ -1 / 2 U T A=QA

[0144] Where Q = Λ -1 / 2 U T , E{ZZ T}=I;

[0145] Assign the same variance to each independent component as the corresponding principal component:

[0146]

[0147] in, is the variance matrix, E{CC T}=D=diag(λ1,λ2,…λ m ), m is determined by discarding elements corresponding to low eigenvalues;

[0148] Pair Matrix Normalize and get:

[0149]

[0150] in,

[0151] Using FastICA algorithm to optimize matrix C n , we get the following dominant independent components:

[0152]

[0153] Among them, the demixing matrix Mixing matrix B = UΛ 1 / 2 C n D -1 / 2 , WB=I;

[0154] On this basis, the following Hotelling T are defined respectively 2 Statistics and the squared process error (SPE) statistic are used for fault detection:

[0155]

[0156] in, for The t-th column vector of

[0157]

[0158] in, represents the square of the vector 2 norm, then It can be calculated by the following formula:

[0159]

[0160] in, is the t-th column vector of A;

[0161] Use kernel density estimation to estimate T 2 The confidence limit of the SPE statistic is used as the corresponding fault time domain detection threshold. Kernel density estimation is a non-parametric method that is suitable for estimating the probability density function of a univariate random process. i (i=1,2,…,l); its probability density function can be derived as:

[0162]

[0163] Where l and ω represent the number of samples and bandwidth, respectively. The kernel function K(x) selects a Gaussian kernel and uses the Silverman method to set the bandwidth ω:

[0164]

[0165] Among them, σ is the standard deviation of the data, and IQR is the interquartile range of the data.

[0166] Figure 5 Schematic diagram of kernel density estimation, where x∈{T 2 , SPE}. f(x) is the estimated PDF. It is worth noting that the points that occupy 99% of the area of ​​f(x) serve as the confidence limits of the monitoring statistic.

[0167] In a preferred embodiment of the present invention, in step 5, the newly acquired battery temperature data The fault detection process specifically includes:

[0168] First, using the same discrete space basis function Calculate the new time coefficient matrix

[0169]

[0170] Compute new independent components:

[0171]

[0172] Calculate the corresponding T 2 Statistics:

[0173]

[0174] in, for The t-th column vector of ;

[0175] Calculate the new SPE statistic:

[0176]

[0177] in, Calculated by the following formula:

[0178]

[0179] in, A new The t-th column vector of ;

[0180] The calculated T 2 new and SPE new Compare with the corresponding confidence limits to complete fault detection.

[0181] In a preferred embodiment of the present invention, in step 6, for samples whose detected fault detection statistics are greater than the corresponding confidence limit, and

[0182]

[0183] Calculate separately and The average position contribution map of is used as the battery failure space contribution map:

[0184]

[0185] Among them, ψ(x,y)=[ψ1(x,y),ψ2(x,y),…,ψ N (x,y)] T , ||·|| represents the Euclidean norm, |·| represents the absolute value, M and L are and The cardinality, Indicates that in the sample set T on 2 Continuous space contribution plot of a statistic, Indicates that in the sample set Continuous space contribution plot of the upper SPE statistic;

[0186] According to the spatial position with the largest contribution value in the spatial contribution graph and The spatial location of the fault can be located.

[0187] In an example based on the present invention, Figure 6 The battery pack thermal runaway propagation model shown in the figure is simulated. In the battery pack, 35 21700 batteries are connected in a form of 5 parallel and 7 series. Figure 6 In the battery pack composed of 35 cells shown in (a), temperature probes are provided for the temperatures of 12 battery cells, and the battery cells provided with the temperature probes are deeper.

[0188] The battery pack was charged from its initial state of charge at a 2C charge rate for 10 minutes. The initial battery pack temperature was set to an ambient temperature of 25°C. During the charging process, thermal runaway was actively triggered in battery #18. This triggered a transient heat source, releasing a total of 22 kJ of heat. When this heat dissipated, causing the maximum temperature of the remaining batteries to exceed 80°C, thermal runaway was passively triggered in the remaining batteries, causing the runaway to spread throughout the battery pack. Figure 6 (b) shows the temperature distribution of the battery pack at 350s when thermal runaway is triggered in battery #18. It can be seen that the thermal runaway spreads around battery #18.

[0189] Because cylindrical batteries have high thermal conductivity along their height, the temperature is assumed to be uniform along the height of the cell. Therefore, the thermal process of a cylindrical battery pack can be considered a two-dimensional thermal model. Simulated temperature data from battery pack charging without thermal runaway triggering is used for training to obtain confidence limits for fault detection, and simulated data from thermal runaway faults is used to validate the fault detection and localization method.

[0190] The method of the present invention is used to detect and locate thermal runaway batteries within a battery system. First, offline training is performed based on simulated battery pack charging temperature data without thermal runaway triggering to obtain fault detection confidence limits. This is then used to perform online monitoring of thermal runaway faults and locate the faulty battery.

[0191] Figure 7 The time domain detection results of thermal faults by the method proposed in this invention are shown in FIG. Figure 7 (a) and (b) are T 2 and the test results of SPE statistics. Figure 7 The points on the horizontal axis represent the time when the fault occurs, and the dotted line represents the confidence limit T 2 R and SPE R , the solid line represents T 2 and SPE, the moment when the solid line exceeds the dotted line is considered to be a fault detected. 2 Both the SPE statistic and the FPS statistic can quickly detect a fault after it occurs, but the SPE statistic has certain false detection issues.

[0192] Figure 8 The results of the spatial positioning of thermal faults by the method proposed in this invention are shown in FIG. Figure 8 (a) and (b) are based on T 2 The circles in the figure represent the locations of the batteries that actively trigger thermal runaway, and the triangles represent the fault location locations. It can be seen that the proposed method based on T 2 The fault location successfully falls within the range of the thermal runaway battery, while the fault location based on SPE statistics is significantly far away from the actual location of the thermal runaway battery. Figure 8 Further analysis of the fault detection results presented shows that based on T 2 The fault detection and location based on statistics is more outstanding in overall performance and has better accuracy and reliability than the method based on SPE statistics.

[0193] It should be understood that the size of the serial numbers of the steps in the embodiment of the present invention does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0194] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A data-driven battery thermal fault detection and location method, characterized by: The specific steps include: Step 1: Perform spatiotemporal decoupling on the battery temperature data collected by the sensor. Then, use the ISOMAP algorithm to reduce the data dimension based on Euclidean distance, converting the spatiotemporal temperature data into a linear combination of time coefficients and discrete spatial basis functions. Step 2: Based on the eigenvalue sorting and screening, the linear combination obtained in step 1 is converted into a finite dimension, and the radial basis function with the Gaussian function as the kernel function is used to fit the discrete space basis function to obtain the form of the continuous space basis function; Step 3: Use the MICA algorithm to process the time coefficient and obtain the statistics for fault detection, and then set the corresponding fault time domain detection threshold for the statistics; Step 4: Use the battery temperature data during normal battery operation to perform the above steps to complete offline training, and obtain stable discrete and continuous space basis functions and fault time domain detection thresholds; Step 5: Collect battery temperature data from the newly acquired sensor, obtain the corresponding time coefficient of the discrete space basis function based on offline training, calculate the corresponding statistics online, and compare them with the detection threshold to achieve fault detection; Step 6: Use the statistics corresponding to the detected fault samples to establish a battery fault space contribution map, and realize fault location according to the contribution size corresponding to different positions in the contribution map.

2. The method according to claim 1, wherein: The spatiotemporal data collected by sensors in step 1 Where n1 and n2 correspond to the number of sensors in the x and y directions respectively. The number of sensors is n = n1 × n2, n t Corresponding to the measurement time dimension, t represents time; It is decoupled into the following infinite-dimensional space-time separation form through Fourier transform: Where, is the time coefficient, is the i-th discrete space basis function, which is specifically expressed as a linear combination of spatiotemporal data: in, is the low-dimensional embedding of the original space, for The k-th column vector of ; Define the temperature spatiotemporal data set at time p as T(:,t q ) is T(:,t p ) is one of its K nearest neighbors, and the Euclidean distance between them is d x (p,q), then for each spacetime point k=1,…,n t , calculate T(:,t p ) and T(:,t q ) is the shortest approximate geodesic distance between: d G (p,q)=min(d G (p,q),d G (p,k)+d G (k,q)),k=1,…n t This creates a matrix Used to represent the shortest approximate geodesic distance between all pairs of points in high-dimensional space; The ISOMAP algorithm is executed to obtain a low-dimensional embedding by preserving the shortest geodesic distance between high-dimensional spatiotemporal data points. Its optimization objective is expressed as: Among them, d N (p,q) are two points in low-dimensional space and γ(:,t p ) and γ(:,t q ), the shortest path distance between is the shortest path distance matrix embedded in low dimension; on this basis, the above optimization objective is transformed into the following minimum optimization problem: Among them, τ(D G )=-HAH / 2, Represented as {A pq =D Gpq 2 }, is the centering matrix, H=I-1 / n t ee T , I is the identity matrix, L represents the matrix 2 norm; Define λ i is the matrix τ(D G )’s ith eigenvalue, V i k is the kth component of the i-th eigenvector, and the low-dimensional embedding is calculated Leveraging low-dimensional embeddings The discrete space basis function can be obtained Time coefficient a i (t) is calculated by the following inner product method:

3. The method according to claim 2, wherein: In step 2, firstly transform the matrix τ(D G ) are sorted in descending order by their eigenvalues The first N of them are selected for dimensionality reduction; N is selected to satisfy η>η0.99, which is calculated by the following formula: The discrete space basis functions are optimized based on the following objectives to capture their lost spatial information: s.t.ψ i (x,y)∈C r (x,y),i=1,...,N,r=0,1,2... Among them, ψ i (x, y) is the i-th continuous space basis function to be designed, C r (x,y) represents the set of functions with continuous derivatives of order r in the x and y directions; The continuous space basis function ψ is obtained by fitting the radial basis function with Gaussian function as the kernel function i (x,y), combined with the selected N eigenvalues, the following finite-dimensional battery continuous space temperature estimation can be obtained:

4. The method according to claim 3, wherein: When executing the MICA algorithm in step 3, Considered as a mixed signal, specifically expressed as a mixing matrix With independent component matrix The product of: A=BY By rearranging: in, is the demixing matrix, which is also the main estimation target of the MICA algorithm. is the estimated independent component matrix; The MICA algorithm process includes: First, A is whitened based on principal component analysis to eliminate cross-correlation and the following score matrix is ​​calculated: S=U T A in, is the covariance matrix E{AA T }=UΛU T , where Λ is a diagonal matrix consisting of N eigenvalues; Normalize the score matrix to get the whitened data matrix of A Z=Λ -1 / 2 S=Λ -1 / 2 U T A=QA where Q = Λ -1 / 2 U T , E{ZZ T} = I; Assign the same variance to each independent component as the corresponding principal component: in, is the variance matrix, E{CC T }=D=diag(λ1,λ2,…λ m ), m is determined by discarding elements corresponding to low eigenvalues; Pair Matrix Normalize and get: in, Using FastICA algorithm to optimize matrix C n , we get the following dominant independent components: Among them, the demixing matrix Mixing matrix B = UΛ 1 / 2 C n D -1 / 2 , WB=I; On this basis, the following Hotelling T are defined respectively 2 Statistics and the squared process error (SPE) statistic are used for fault detection: in, for The t-th column vector of in, represents the square of the vector 2 norm, then It can be calculated by the following formula: in, is the t-th column vector of A; Use kernel density estimation to estimate T 2 The confidence bounds of the SPE statistic are used as the corresponding fault time domain detection thresholds.

5. The method according to claim 4, wherein: In step 5, the newly acquired battery temperature data The fault detection process specifically includes: First, using the same discrete space basis function Calculate the new time coefficient matrix Compute new independent components: Calculate the corresponding T 2 Statistics: in, for The t-th column vector of ; Calculate the new SPE statistic: in, Calculated by the following formula: in, A new The t-th column vector of ; The calculated T 2 new and SPE new Compare with the corresponding confidence limits to complete fault detection.

6. The method according to claim 5, wherein: In step 6, for samples where the detected fault detection statistic is greater than the corresponding confidence limit and Calculate separately and The average position contribution map of is used as the battery failure space contribution map: Among them, ψ(x,y)=[ψ1(x,y),ψ2(x,y),…,ψ N (x,y)] T , ||·|| represents the Euclidean norm, |·| represents the absolute value, M and L are and The cardinality, Indicates that in the sample set T on 2 Continuous space contribution plot of a statistic, Indicates that in the sample set Continuous space contribution plot of the upper SPE statistic; According to the spatial position with the largest contribution value in the spatial contribution graph and The spatial location of the fault can be located.

Images