A Multi-Stage Battery Remaining Life Prediction Method Based on NTM and Statistical Methods

Abstract

This invention discloses a multi-stage battery remaining life prediction method based on NTM and statistical methods, which can predict the remaining battery life in multiple stages. The method includes the following steps: (1) Processing battery capacity data to extract the health factor HI; (2) Inputting the processed data into a Long Short-Term Memory Neural Network (LSTM) based on step (1), and segmenting the data according to the degradation rate; (3) Obtaining the degradation function for each stage using Bayesian statistical methods based on step (2), and introducing it into a Neural Turing Machine (NTM) model for parameter updates; (4) Inputting the data into the Neural Turing Machine (NTM) for prediction based on step (3); (5) Battery life prediction example analysis: Validation was performed using NASA's public dataset and compared with other methods.

Description

Technical fields:

[0001] This invention relates to a multi-stage battery remaining life prediction method based on NTM and statistical methods, which belongs to the fields of battery health management and civil aviation transportation safety research. Background technology:

[0002] In the civil aviation sector, lithium batteries are being used more extensively due to their advantages such as high energy density, low self-discharge rate, long lifespan, and environmental friendliness. In airports, lithium batteries are often used as backup power for electrical equipment. Currently, more advanced battery technologies are not yet mature enough, and lithium batteries will continue to dominate for a considerable period. To ensure the safe use of batteries, accurately estimating their Remaining Useful Life (RUL) is crucial for aircraft flight safety.

[0003] Currently, methods for predicting remaining useful life (RUL) can be broadly classified into three categories: model-based methods, data-driven methods, and hybrid methods.

[0004] Model-based prediction methods focus on underlying physical failure mechanisms. If the degradation process is well-designed, then physics-based methods can often yield fairly accurate estimates.

[0005] Due to issues such as short-term capacity regeneration and noise interference, the failure mechanism of lithium batteries is highly complex. Compared with prediction methods based on physical models, data-driven prediction methods do not need to consider complex electrochemical characteristics. Instead, they extract health factors (voltage, current, resistance, capacity, etc.) that can reflect the degradation trend from raw data, and use intelligent algorithms to learn and complete the RUL prediction.

[0006] Data-driven approaches uncover patterns in data and establish statistical models that approximate the degradation patterns. Predicting the time or cycle at which batteries reach their failure threshold has become a major focus in the field of RUL (Range Limiting) prediction in recent years.

[0007] Lithium batteries in various devices may undergo irreversible electrochemical reactions due to factors such as operating time, environment, and usage methods. In airport systems with extremely high requirements for operational efficiency and safety, the resulting losses can be incalculable. Therefore, the reliability of lithium battery systems is crucial. Accurately and stably predicting the remaining lifespan of lithium-ion batteries can effectively avoid the irreversible effects of lithium-ion battery failure, providing a safety guarantee for the long-term operation of airport power supply systems. Thus, predicting the remaining battery lifespan has practical application value. Summary of the Invention:

[0008] This invention provides a multi-stage battery remaining life prediction method based on NTM and statistical methods to address the problems existing in the prior art. The original battery contains a large amount of information such as charging and discharging current, voltage, and temperature. Therefore, the charging and discharging current and temperature are extracted and standardized as health factors. Then, LSTM mutation analysis is used to find mutation points based on the changes in degradation rate, dividing the data into multiple stages and mapping them to the original battery capacity. This multi-stage capacity data is input into an NTM network, and the parameters of each stage are updated through Bayesian statistical analysis. Finally, the NTM model is used for prediction.

[0009] The technical solution adopted in this invention is: a multi-stage battery remaining life prediction method based on NTM and statistical methods, characterized by the following steps:

[0010] (1) Process the battery capacity data and extract the health factor HI;

[0011] (2) Based on step (1), the processed data is input into a long short-term memory neural network (LSTM) and the data is segmented according to the degradation rate;

[0012] (3) Based on step (2), Bayesian statistical methods are used to obtain the degradation functions of each stage and introduce them into the Neural Turing Machine (NTM) model to update the parameters;

[0013] (4) Based on step (3), the data is input into the Neural Turing Machine (NTM) for prediction;

[0014] (5) Battery life prediction case analysis: Validation was performed using NASA’s public dataset and compared with other methods.

[0015] This invention offers the following advantages: It utilizes a Long Short-Term Memory (LSTM) neural network to divide the battery into multi-stage signals containing degradation characteristics, introduces Bayesian statistical methods for parameter updates, and combines this with a Neural Turing Machine (NTM) for fault prediction. This invention incorporates statistical analysis into the NTM prediction model to achieve model interpretability. The statistical analysis employs three types of interpretability indicators: 1) statistical analysis of abrupt change points; 2) statistical analysis of the NTM; and 3) statistical analysis of the predictions. Therefore, this invention can perform multi-stage prediction of the remaining battery life, ultimately reducing maintenance costs and ensuring the safe operation of airport power equipment, thus possessing practical application value. Attached image description:

[0016] Figure 1 This is a flowchart of the algorithm of the present invention.

[0017] Figure 2 The segmentation results for battery B0005.

[0018] Figure 3 The prediction result is for battery number B0005.

[0019] Figure 4 The prediction result is for battery number B0006.

[0020] Figure 5 Performance comparison of the algorithm of this invention with other algorithms. Detailed implementation method:

[0021] The invention will now be further described with reference to the accompanying drawings.

[0022] This invention provides a multi-stage battery remaining life prediction method based on NTM and statistical methods, comprising the following steps:

[0023] Step (1): Process the battery capacity data and extract the health factor HI.

[0024] Step (1) includes the following:

[0025] The battery capacity data contains health factors (HI) such as current, voltage, and temperature at each time point. Therefore, the original data needs to be processed to calculate these HI.

[0026] standardization:

[0027] γ i =2(x i -x min ) / (x max -x min )-1 (1)

[0028] Where, x max and x min Represents the maximum and minimum values ​​of feature X; γ i To standardize the data, the battery capacity is calculated based on the standardized data, using the following formula:

[0029]

[0030] Where, γ i (k) is the i-th degradation parameter of stage k, and D(k) is the capacity after data fusion in stage k.

[0031] Step (2): Based on step (1), the processed data is input into a long short-term memory neural network (LSTM) and the data is segmented according to the degradation rate.

[0032] In step (2), to improve the accuracy of the prediction results, LSTM is used to calculate the points in the capacity data where the degradation rate changes beyond a threshold, thereby segmenting the data.t =[x t-m ,x t-m+1 ,...,x t-i ,x t [x] represents the degradation data from time tm to time t, where x represents the ratio of data with decreasing capacity to data with increasing capacity. rate The LSTM used for change point detection is a class classifier that determines the optimal parameter θ1 in equation (9) to minimize the number of time points. The change point determined in equation (4) is denoted as T. cp Defined as output y t Exceeding the threshold.

[0033]

[0034] Step (3): Based on step (2), Bayesian statistical methods are used to obtain the degradation functions of each stage and they are introduced into the Neural Turing Machine (NTM) model for parameter updates.

[0035] Step (3) includes the following:

[0036] A novel Bayesian update strategy based on stage-specific tasks is proposed. It first obtains a batch of samples from the current stage data and contextual memory. Battery capacity is equivalent to a single-node Bayesian network consisting of variables X. X also contains values ​​r and Ω. x ={x1,x2,...,x r}, then the single-node Bayesian network has parameters r and θ to be estimated. i =P(X=x) i ), i = 1, 2, ... r. Using Let θ1, θ2, ..., θ be used to represent vectors. r ,in If D = (D1, D2, ..., D...) m ), satisfying X = x i The number of samples is m i Then we obtain the likelihood function:

[0037]

[0038] Where {m i Sufficient statistics for |i=1,2,...,r}. Calculate the conjugate distribution of the multinomial likelihood function for each stage. Assume that regarding... The prior knowledge contains n virtual data samples, i.e., the assumption is... The prior distribution is D = (a1, a2, ..., a r ),Right now:

[0039]

[0040] in

[0041] Substituting the two formulas above into Bayes' theorem, we get... The posterior distribution satisfies:

[0042]

[0043] To obtain the Bayesian estimation results for each stage, each component θ is required. i The loss using the temporary model and the SGD optimizer is as follows:

[0044]

[0045] Where β represents the learning rate, θ base Here, k is the parameter, and k is the current instance β. n The number of times it is updated.

[0046] Then combine the parameter θ i The final model parameters θ are obtained. i -θ base Let represent the gradient change value. The evolution of the gradient is controlled by a monotonic scheduling strategy applied through a dynamic controller λ, which dynamically changes the parameters towards the average direction. The update rules are as follows:

[0047] θ new =θ base +λ(θ i -θ base (10)

[0048]

[0049] Step (4): Based on step (3), the data is input into the Neural Turing Machine (NTM) for prediction.

[0050] Step (4) includes the following:

[0051] In NTM, the memory library M∈R l×s It consists of l memory locations, i.e., M = [m1, m2, ... m l ] where each position is a vector with feature s, i.e., m i ∈R l×s NTM updates a hidden state h at each time step. t ∈R l×s This serves as a summary of the measurement vectors received so far and their interactions over time. Then, the hidden state sequence H = {h} is utilized. i |i∈N + ,i≤t l} serves as a feature of T extracted automatically. NTM performs three types of operations in sequence: write, read, and hidden state update.

[0052] To determine how much memory should be modified, an attention mechanism is used to calculate the weight of each memory location. The procedure is as follows:

[0053] α t =δ p1 (v a tanh([c t ,h t-1 ])W ha +b a (12)

[0054]

[0055] Formulas (12) and (13) can capture the complex relationships between multi-sensor data, α t ={a t,1 ,...,a t,i ,...,a t,l} represents the attention weights, satisfying ∑ i a t,i =1. W ha ∈R (s+s)×l v a ∈R and b a ∈R 1×l All parameters are trainable. Unlike previous NTM-based models, a dropout operator has been added. Equations (12) and (13) can alleviate overfitting and improve its generality. Then, it applies a parameter to each memory slot m. i The following updates will be made:

[0056] m i =a t,1 c t +(1-a t,i )m i for i = 1, 2, ... l (14)

[0057] Balance memory using attention mechanisms t The new information that should be retained in the memory. After the memory bank is updated, the NTM reads the new data from it to update the hidden state. For write operations, the attention computation is regularized by the dropout operator. First, as in equations (12) and (13), in c t and h t-1 The attention weights are calculated. Then, the weighted average of the vectors currently contained in memory is calculated using the attention weights to obtain a read vector r. t At time t, the memory library writes data from the currently measured value o. tNew information obtained and previously hidden state t Update. In this way, it stores new knowledge from the input while retaining previously captured information. The information to be written to the memory bank M is the content vector c. t ∈R l×s The calculation is as follows:

[0058] c t =δ p1 (σ([σ t ,h t-1 W hc +b c (15)

[0059] Where δ p1 It is the exit operator with a probability of δ. p1 o t ∈R l×f h represents the current sensor measurement vector. t-1 ∈R l×s Indicates the previous hidden state. W hc ∈R (f+s)×s and W hc ∈R (f+s)×s These are trainable parameters.

[0060] In hidden state updates, knowledge gathered from the updated memory and previous hidden states is used as contextual memory. The hidden state is updated based on a Bayesian parameter update method, incorporating the current input metric. Degenerate features at each stage are stored as follows:

[0061] h t =σ(o t W oh +r t W rh +h t-1 W hh +b h (16)

[0062] Among them, W oh ∈R f×s W rh ∈R t×s W rh ∈R s×s and b h ∈R s These are all trainable parameters. The hidden state sequence h1, h2, ... h N Grouped in a matrix H∈R t1×s In the process, features are automatically extracted from the input time series.

[0063] After extracting features using NTM, a simple decoder is used to learn the mapping from these features to RUL values:

[0064] d1=δ p2 (σ(HW dh +b dh (17)

[0065]

[0066] in and For trainable parameters, δ p2 This is an sigmoid function. As the output of this step, we obtain... It represents a sequence of predicted RUL values, and then outputs the prediction curve and error in sequence.

[0067] Statistical analysis:

[0068] In the widely studied problem of lifespan prediction, various deep learning algorithms, such as Long Short-Term Neural Networks (LSTN), Convolutional Neural Networks (CNN), and Graph Neural Networks (Graph Neural Networks), are extensively used in prediction models, significantly improving their predictive performance. However, current research on lifespan prediction models focuses on enhancing predictive performance, with insufficient emphasis on model interpretability. Most interpretability-based prediction models are primarily based on ex-post interpretability, meaning the prediction process is opaque, and users lack understanding of the model's internal mechanisms, resulting in low interpretability. Therefore, this invention introduces a statistical distribution model into the NTM prediction model to achieve ex-ante interpretability. The statistical distribution employs three types of interpretability indices.

[0069] 1) Statistical analysis of mutation points: The LSTM used for mutation point detection is a binary classifier with two categories. A cell is either in a normal operating phase or a degradation process phase; otherwise, it is labeled 0 (normal operation) or 1 (degradation process). Clustering is used, with Dunn index, Davidson-Bolding index, H-score, and silhouette coefficient as evaluation indicators to determine the optimal number and location of segmentation points for battery capacity.

[0070] A. Dunn Index (DVI) Formula

[0071]

[0072] Formula meaning: The ratio of the shortest distance between any two clusters to the maximum intra-cluster distance of any cluster.

[0073] Significance of the results: A larger DVI value indicates that the clustering results are more compact within the same cluster and more widely separated between different clusters. In other words, the greater the inter-cluster distance, the smaller the intra-cluster distance.

[0074] B. Davidson-Burden Index (DB)

[0075]

[0076] The numerator represents the sum of the average distances from all points within the cluster to the centroid of the cluster, and the denominator d(c i ,c j ) represents the distance between the centroids of the two classes.

[0077] Formula meaning: DB is the maximum value of the ratio of the sum of the average intra-class distances between any two categories to the distance between the centroids of the two clusters.

[0078] Significance of the results: A smaller DB value indicates that the clustering results are more compact within the same cluster and more widely separated between different clusters. In other words, the smaller the intra-cluster distance, the larger the inter-cluster distance.

[0079] C.CH Index (CH)

[0080]

[0081] in, tr(W) represents the trace of the distance difference matrix between clusters; tr(W) represents the trace of the distance difference matrix within a cluster. It is the mean of the entire dataset; z j It is the j-th cluster ω z The mean of N; N represents the number of clusters; K represents the current cluster.

[0082] Formula meaning: The ratio of data separation (the sum of squared distances between the centers of each cluster and the center of the dataset) to compactness (the distances between each point in a cluster and the cluster center).

[0083] Significance of the results: The larger the CH index value, the more compact the clusters themselves and the more dispersed the clusters are, which means the clustering results are better.

[0084] D. Profile coefficient (SI)

[0085]

[0086] Among them, z j b(i) represents the average distance from pattern i to every pattern in all the same clusters; b(i) represents the average distance from pattern i to every pattern in the closest other cluster.

[0087] Formula meaning: It measures the similarity between patterns in the same cluster and patterns in other nearest clusters.

[0088] Significance of the result: A higher SI value indicates that i matches its own cluster better and matches other clusters worse. That is, it is dense within a cluster and sparse between clusters.

[0089] 2) Statistical analysis of NTM

[0090] A. Bayesian fusion probability

[0091] The probability after Bayesian fusion is expressed as:

[0092]

[0093] The expected value is calculated based on the different distributions at each stage, ultimately yielding the components θ. i The complete Bayesian estimate.

[0094] The formulas for the three degenerate distribution functions are as follows:

[0095] B. Gaussian normal distribution formula:

[0096]

[0097] C. Second-order sine sum distribution:

[0098] f(x)=a1·sin(b1·x+c1)+a2·sin(b2·x+c2)

[0099] D. Fourier distribution formula:

[0100] f(x)=a0+a1 cos(x·w)+b1 sin(x·w)+a2 cos(2x·w)+b2 sin(2x·w)

[0101] ER 2

[0102]

[0103] R 2 Used to determine the correlation between the distribution function and the actual capacity value, when R 2 When R is 0, it means that the two are absolutely uncorrelated. 2 A value of 1 indicates that the two are absolutely uncorrelated. Therefore, R 2 The closer the value is to 1, the better the distribution function reflects the distribution characteristics of battery capacity data.

[0104] F.SSE

[0105]

[0106] SSE is used to represent the sum of squares of the errors between the distribution function and the corresponding points of the battery capacity data. The closer the SSE is to 0, the better the distribution function reflects the degradation characteristics of the original data.

[0107] 3) Statistical analysis of prediction

[0108] Three key indicators were introduced: monotonicity, trend, and prediction error.

[0109] A. Monotonicity is represented by the average absolute difference between the positive and negative derivatives, which can characterize the potential positive or negative trend of HI. The larger the monotonicity value, the fewer the battery capacity recovery phases and the smaller the data fluctuation.

[0110]

[0111] B. Trend: Parameters used to characterize capacity have the same basic shape. The larger the trend value, the higher the similarity of capacity degradation characteristics.

[0112]

[0113] tendency = 1 - t i

[0114] Both indicators range from 0 to 1. The closer the indicator value is to 1, the better the indicator performs. If the value is close to 0, it means that the indicator has not reached the ideal performance.

[0115] C. The root mean square error (RMSE) is the square root of the ratio of the sum of the squares of the deviations of all observed values ​​from the true value to the number of observations. It measures the deviation between the true value and the observed values. Its calculation formula is as follows:

[0116]

[0117] in Let y be the predicted data obtained at time i. i Let be the actual data at time i, and n be the number of data points in the dataset. The closer the RMSE value is to 1, the better the fit and prediction results.

[0118] Step 5: Battery life prediction case analysis: The method was validated using NASA's public dataset and compared with other methods, showing that the method of the present invention has better performance.

[0119] The battery data was obtained from a publicly available dataset provided by NASA. This invention uses NASA's publicly available lithium-ion battery dataset. Two batteries from the first group were selected as the research subjects, numbered B0005 and B0006. The test temperature was 24 degrees Celsius. Charging was performed at a constant current of 1.5A until the battery voltage reached 4.2V, then charged at a constant voltage until the charging current dropped to 20mA. Discharging was performed at a constant current of 2A until the voltages of batteries 5 and 6 dropped to 2.7V and 2.5V respectively. Batteries reached the end of their lifespan when their rated capacity decreased to 70% of their initial capacity; that is, when the rated capacity of both batteries decreased from 2Ahr to 1.4Ahr, their lifespan was considered over.

[0120] (1) Data processing

[0121] First, the battery health factors are calculated using the formula in step two, thereby determining the actual battery capacity. A correlation analysis is then performed between each health factor and the battery capacity, and the results are shown in Table 1.

[0122] Table 1. Correlation analysis results of various battery health factors

[0123]

[0124] (2) Multi-stage division

[0125] 1) Through cluster analysis, the Dunn index, Davidson-Boldt index, H-score, and silhouette coefficient were calculated for different partitioning stages to determine the number of mutation points. LSTM was then used to analyze the locations of these mutation points. The results of the mutation point count are shown in Table 2, and the final partitioning results are as follows: Figure 2 , Figure 3 As shown:

[0126] Table 2 Evaluation Results of the Number of Mutation Points in B0005

[0127]

[0128] 2) Therefore, the most accurate stage division occurs when the final clustering result of the four batteries is 3, meaning they are divided into three stages. Based on the LSTM analysis results, the battery capacity is divided into different stages, and the degradation rate of each stage is shown in Table 3:

[0129] Table 3 Degradation rate of each battery at different stages

[0130]

[0131]

[0132] (3) Multi-stage fault prediction

[0133] 1) Based on the data segmentation results obtained in the previous step, analyze the degradation characteristics of each stage and obtain the degradation functions of each stage as second-order Gaussian function, second-order sine function and second-order Fourier function.

[0134] The specific formula is as follows:

[0135] B0005 Phase 1:

[0136]

[0137] B0005 Phase Two:

[0138] f(x)=5.974sin(0.0351x+0.9483)+4.449sin(0.04062x+3.97)

[0139] B0005 Phase 3:

[0140] f(x)=1.354+0.04559cos(3.595)+0.02673sin(3.595x)+0.01552cos(7.19x)-6.145×10 -4 sin(7.19x).

[0141] 2) Train the NTM according to step four, and then input the multi-stage data obtained in the previous step into the NTM for prediction. The prediction results for the two batteries are as follows: Figure 3 As shown in Table 4, the calculation results of the key indicators are shown in Table 4.

[0142] Table 4 Statistical Indicators of Prediction Results for Different Stages of Each Battery

[0143]

[0144] Table 4 shows that batteries B0005 and B0006 exhibited the best monotonicity in the second stage and the worst in the third stage, indicating that their performance was more stable in the second stage. Trend results show that the similarity of capacity degradation characteristics for battery B0005 across all three stages is above 0.86, while the similarity for battery B0006 across all stages is above 0.83. This demonstrates that the prediction method used in this patented model effectively reflects the degradation characteristics of the original data.

[0145] (4) Comparison of prediction errors between this patented model and three other methods

[0146] To demonstrate the superiority of this patented model, the prediction results are compared with those of other methods, such as... Figure 5 As shown in the figure. The results show that after prediction by the patented model, in the case of multi-stage prediction, the average RMSE of B0005 per stage is reduced by 1.87% compared with Bi-LSTM; the average RMSE of B0006 per stage is reduced by 1.25% compared with Bi-LSTM. Because the degradation data of the previous stage in the multi-stage prediction obtained by statistical process analysis does not contribute much to the prediction of the degradation data of the next stage, combining NTM with statistical process analysis can more effectively predict battery life.

[0147] This patented model improves upon existing traditional NTM models by employing statistical analysis methods. It proposes a mutation point detection method based on LSTM and cluster analysis, which can divide a complex full-lifecycle battery into multiple data segments. Based on this, lifetime prediction is performed using NTM according to the degradation characteristics of each stage. Furthermore, statistical analysis methods are introduced to support the model's pre-interpretability.

[0148] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A multi-stage battery remaining life prediction method based on NTM and statistical methods, characterized in that: Includes the following steps: (1) Process the battery capacity data and extract the health factor HI; (2) Based on step (1), the processed data is input into a long short-term memory neural network (LSTM) and the data is segmented according to the degradation rate; (3) Based on step (2), Bayesian statistical methods are used to obtain the degradation functions of each stage and introduce them into the Neural Turing Machine (NTM) model to update the parameters; (4) Based on step (3), the data is input into the Neural Turing Machine (NTM) for prediction; (5) Battery life prediction case study: Validation was performed using NASA's public dataset and compared with other methods; Step (3) includes the following: First, obtain a batch of samples from the current stage data and contextual memory. The battery capacity is equivalent to a single-node Bayesian network composed of variables X, where X contains values ​​r and Ω. x ={x1,x2,...,x r }, then the single-node Bayesian network has parameters r and θ to be estimated. i =P(X=x) i ), i = 1, 2, ..., r, using Let θ1, θ2, ..., θ be used to represent vectors. r ,in If D = (D1, D2, ..., D...) m ), satisfying X = x i The number of samples is m i Then we obtain the likelihood function: Where {m i For each stage of the multinomial likelihood function, calculate the conjugate distribution of the multinomial likelihood function for each stage, assuming that for |i=1,2,...,r}, sufficient statistics are available. The prior knowledge contains n virtual data samples, i.e., the assumption is... The prior distribution is D = (a1, a2, ..., a r ),Right now: in Substituting the two formulas above into Bayes' theorem, we get... The posterior distribution satisfies: To obtain the Bayesian estimation results for each stage, each component θ is required. i The loss using the temporary model and the SGD optimizer is as follows: Where β represents the learning rate, θ base Here, k is the parameter, and k is the current instance β. n The number of updates; Then combine the parameter θ i The final model parameters θ are obtained. i -θ base The gradient change value is represented by λ, and its evolution is controlled by a monotonic scheduling strategy applied through a dynamic controller λ. This strategy dynamically changes the parameters in the direction of the average, and the update rule is as follows: i new =θ base +λ(θ i -θ base ) (10) 。 2. The multi-stage battery remaining life prediction method based on NTM and statistical methods as described in claim 1, characterized in that: Step (1) includes the following: The battery capacity data includes current, voltage, temperature, and health factor HI at each time point. The original data is processed to calculate HI. standardization: γ i =2(x i -x min ) / (x max -x min )-1 (1) Where, x max and x min Represents the maximum and minimum values ​​of feature X; γ i To standardize the data, the battery capacity is calculated based on the standardized data, using the following formula: Where, γ i (k) is the i-th degradation parameter of stage k, and D(k) is the capacity after data fusion in stage k.

3. The multi-stage battery remaining life prediction method based on NTM and statistical methods as described in claim 1, characterized in that: In step (2), LSTM is used to calculate the points in the capacity data where the degradation rate changes exceed the threshold, thereby segmenting the data; X t =[x t-m ,x t-m+1 ,...,x t-i ,x t [x] represents the degradation data from time tm to time t, where x represents the ratio of data with decreasing capacity to data with increasing capacity. rate This indicates that the LSTM used for change point detection is a class classifier, and the optimal parameter θ1 in formula (3) is determined to minimize the number of time points. The change point determined in formula (4) is denoted as T. cp Defined as output y t Exceeding the threshold 。 4. The multi-stage battery remaining life prediction method based on NTM and statistical methods as described in claim 1, characterized in that: Step (4) includes the following: In NTM, the memory library M∈R l×s It consists of l memory locations, i.e., M = [m1, m2, ... m l ], where each position is a vector with feature s, i.e., m i ∈R l×s NTM updates a hidden state h at each time step. t ∈R l×s This serves as a summary of the measurement vectors received to date and their interactions over time; Then, using the hidden state sequence H = {h} i |i∈N + ,i≤t l As features of T extracted automatically, NTM performs three types of operations in sequence: write, read, and hidden state update; To determine how much memory should be modified, an attention mechanism is used to calculate the weight of each memory location, as follows: α t =δ p1 (v a tanh([c t ,h t-1 ])W ha +b a ) (12) Formulas (12) and (13) can capture the complex relationships between multi-sensor data, α t ={a t,1 ,...,a t,i ,...,a t,l } represents the attention weights, satisfying ∑ i a t,i =1; W ha ∈R (s+s)×l v a ∈R and b a ∈R 1×l All of these are trainable parameters; Then, for each memory slot m i The following updates will be made: m i =a t,1 c t +(1-a t,i )m i for i=1,2,...l (14) Balance memory using attention mechanisms t The new information that should be retained in the memory is updated. After the memory bank is updated, NTM reads the new data from it to update the hidden state. For write operations, the attention calculation is regularized by the dropout operator. First, as in equations (12) and (13), in c t and h t-1 The attention weights are calculated first; then, the weighted average of the vectors currently contained in memory is calculated using the attention weights to obtain a read vector r. t At time t, the memory library writes data from the currently measured value o. t New information obtained and previously hidden state t renew; In this way, new knowledge from the input is stored while previously captured information is retained, along with the information to be written to the memory bank M, i.e., the content vector c. t ∈R l×s The calculation is as follows: ct=δ p1 (σ([σ t ,h t-1 ]W hc +b c )) (15) Where δ p1 It is the exit operator with a probability of δ. p1 o t ∈R l×f h represents the current sensor measurement vector. t-1 ∈R l×s W represents the previous hidden state. hc ∈R (f+s)×s and W hc ∈R (f+s)×s These are trainable parameters; In the hidden state update, the knowledge collected from the updated memory and the previous hidden state is used as the context memory. Combined with the current input metric, the hidden state is updated based on the Bayesian parameter update method. The degenerate features at each stage are stored as follows: h t =σ(o t W oh +r t W rh +h t-1 W hh +b h ) (16) Among them, W oh ∈R f×s W rh ∈R t×s W rh ∈R s×s and b h ∈R s These are all trainable parameters, and the sequence of hidden states h1, h2, ... h N Grouped in a matrix In the process, features are automatically extracted from the input time series; After extracting features using NTM, a simple decoder is used to learn the mapping from these features to RUL values: d1=δ p2 (s(HW dh +b dh )) (17) in and For trainable parameters, δ p2 As an sigmoid function, as the output of this step, we obtain... This represents a sequence of predicted RUL values.

Images