Deep learning-based battery health state and residual life collaborative prediction method

Abstract

The invention discloses a battery health state and residual life collaborative prediction method based on deep learning, and relates to the technical field of battery health prediction. The method comprises the following steps: firstly, determining a current abrupt change time sequence segment, and performing second-order differential calculation on a corresponding voltage transient response segment to generate a voltage recovery time sequence track, thereby realizing time domain decoupling of reversible polarization attenuation and irreversible aging; an irreversible aging time sequence segment is determined through the time sequence coincidence degree, the battery degradation inflection point is analyzed and accurately recognized by combining the voltage residual error change consistency and the voltage trend coherence, and the problems that reversible polarization fluctuation is misjudged to be irreversible aging, and degradation inflection point recognition is deviated are avoided; and finally, on the basis of the decoupled aging characteristics, voltage residual related parameters are extracted, SOH and RUL predicted values are calculated and corrected, it is ensured that the SOH and RUL predicted values achieve cooperative prediction on the basis of the same set of irreversible aging rules, and the problem that the cooperative prediction precision is low due to the fact that an existing method cannot be decoupled, the SOH is misjudged, and inflection point recognition is wrong is solved.

Description

Technical Field

[0001] This invention relates to the field of battery health prediction technology, specifically a method for co-predicting battery health status and remaining life based on deep learning. Background Technology

[0002] In practical scenarios of co-predicting battery SOH and RUL, the battery voltage response signal is a superposition of ohmic voltage drop transients, reversible polarization relaxation, and irreversible aging steady state. Sampling noise or operating condition fluctuations cause significant interference, and the time-domain boundary between polarization relaxation and aging steady state is blurred. Existing technologies have the following defects: First, when processing battery voltage time-series signals, existing co-prediction methods often directly extract features from the original voltage segments. They neither decouple the reversible polarization relaxation process and the irreversible aging steady state component in the voltage signal in the time domain, nor accurately locate the start and end boundaries of polarization relaxation through differential operations. They cannot remove interference signals such as ohmic voltage drop transients and sampling noise, resulting in the polarization decay and aging signal being coupled in the extracted features. Then, when identifying battery degradation inflection points and calculating SOH and RUL, the existing solution does not make trend judgments based on the pure aging characteristics after decoupling. Instead, it directly calculates the aging rate and residual ratio based on the coupled signal. This can easily misjudge the instantaneous fluctuations of reversible polarization as the decay trend of irreversible aging, causing the estimated SOH value to deviate from the true health state. At the same time, the polarization interference causes a time shift in the identification of degradation inflection points.

[0003] Therefore, there is an urgent need for a method that can decouple reversible polarization and irreversible aging signals through second-order differential operations, accurately locate the polarization relaxation boundary and battery degradation inflection point, and calculate the battery health status and remaining life based on pure aging characteristics to jointly predict the battery health status and remaining life, so as to solve the above technical problems and improve the accuracy and reliability of the joint prediction of SOH and RUL. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a deep learning-based method for co-predicting battery health status and remaining life. This method solves the problem that existing co-prediction methods cannot decouple reversible polarization decay and irreversible aging, are prone to misjudging SOH and leading to incorrect identification of degradation inflection points, ultimately resulting in reduced accuracy of SOH and RUL co-prediction.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: a method for co-predicting battery health status and remaining life based on deep learning, comprising the following steps: performing time-series first-order differential calculation and current change rate calculation on current time-series segments to determine current abrupt change time-series segments.

[0006] Perform time-series second-order differential calculations on the voltage transient response segments corresponding to the current abrupt change time series segments to generate the voltage recovery time series trajectory.

[0007] The timing overlap between the starting and ending points of the voltage recovery timing trajectory is calculated, and the irreversible aging timing segment is determined.

[0008] The consistency of voltage residual changes and the coherence of voltage trends were analyzed for irreversible aging time segments to determine the inflection point of battery degradation.

[0009] The irreversible aging time series segment after the battery degradation inflection point is processed by time series feature encoding. After extracting the instantaneous change in voltage residual and the cumulative amount of voltage residual, the SOH prediction value is calculated and corrected.

[0010] The irreversible aging rate after the battery degradation inflection point is calculated based on the voltage residual, and the RUL prediction value is determined and corrected based on the irreversible aging rate.

[0011] Compared with existing technologies, this invention has the following advantages: First, it determines the current transient timing segment and performs second-order differential calculation on the corresponding voltage transient response segment to generate the voltage recovery timing trajectory, thus achieving time-domain decoupling of reversible polarization decay and irreversible aging. Then, it determines the irreversible aging timing segment through timing overlap and accurately identifies the battery degradation inflection point by combining the consistency of voltage residual change and the coherence of voltage trend analysis, avoiding the problems of misjudging reversible polarization fluctuations as irreversible aging and inflection point identification deviation. Finally, based on the decoupled aging characteristics, it extracts voltage residual related parameters to calculate and correct the SOH and RUL prediction values, ensuring that the two achieve collaborative prediction based on the same set of irreversible aging laws, solving the problem of low collaborative prediction accuracy caused by the inability to decouple, misjudgment of SOH, and incorrect inflection point identification in existing methods. Attached Figure Description

[0012] Figure 1 This is a flowchart of the deep learning-based method for co-predicting battery health status and remaining lifespan according to the present invention.

[0013] Figure 2 This is a flowchart illustrating the determination of the battery degradation inflection point in the deep learning-based battery health status and remaining lifespan co-prediction method of the present invention.

[0014] Figure 3 This is a flowchart illustrating the process of determining and correcting the predicted state of health (SOH) value in the deep learning-based collaborative prediction method for battery health status and remaining lifespan of this invention. Detailed Implementation

[0015] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. Please refer to the accompanying drawings. Figure 1 The present invention provides a technical solution: a method for co-predicting battery health status and remaining life based on deep learning, comprising the following steps: S1, performing time-series first-order differential calculation and current change rate calculation on current time-series segments to determine current abrupt change time-series segments.

[0016] Considering that the original battery charging and discharging current signal contains a large number of invalid signals with smooth fluctuations and slow drifts, as well as pseudo-abrupt signals caused by sampling spikes and electromagnetic interference, the first-order derivative of the time sequence can distinguish between smooth current changes and abrupt current jumps, filter out the current baseline drift caused by slow changes in SOC, and further eliminate the interference of single-point sampling noise based on the current change rate calculation based on the sliding window, and lock in the real current abrupt change with a continuous trend. Therefore, the specific process of determining the current abrupt change time sequence segment is as follows: S101, calculate the difference between the current sampling values ​​of adjacent sampling times within the current time sequence segment, and use the ratio of the difference to the corresponding sampling time interval as the current first-order derivative value to form the current first-order derivative sequence.

[0017] S102. Traverse the first-order derivative sequence of the current using a preset sliding window, and take the average value of all the first-order derivative values ​​of the current in each sliding window to obtain the current change rate corresponding to each sliding window, forming a current change rate sequence.

[0018] S103. If the absolute value of the first derivative of all currents within a certain sliding window is greater than the non-zero threshold, and the variance of the corresponding current change rate and the current change rate of the adjacent sliding window before and after it is greater than the judgment threshold, it is judged as a valid abrupt change window.

[0019] S104. Define N consecutive valid mutation windows as current mutation time segments. N is a positive integer not less than 2.

[0020] The physical meaning of the first-order differential value of the current is the instantaneous change of the current within a unit sampling interval. The larger the absolute value, the steeper the instantaneous jump of the current at that moment, and the closer it is to the ideal step signal. The positive or negative sign of the value corresponds to the direction of the current change, that is, the switching of current increase or decrease, charging or discharging.

[0021] The absolute values ​​of the first derivatives of the current are all greater than the non-zero threshold, indicating that there is no steady current or noise fluctuation near zero within the sliding window, and there is a continuous current change throughout the entire process. The non-zero threshold ranges from 0.5 A / s to 5 A / s.

[0022] The rate of change of current is the average of the first derivatives of all currents within the sliding window. Its magnitude reflects the overall intensity and continuous trend of current change within the corresponding sliding window. The larger the absolute value, the more continuous and stable large jumps there are in the current within the window, rather than scattered instantaneous noise fluctuations.

[0023] When the variance of the rate of change of current within the sliding window and the rate of change of current corresponding to the two adjacent sliding windows exceeds the judgment threshold, it indicates that the current change trend of that sliding window differs statistically significantly from the stable intervals before and after it, representing a genuine abrupt change in operating conditions rather than continuous current fluctuations. The judgment threshold ranges from 0.1 to 1.

[0024] The reason for determining the effective mutation window first is that a single-point current jump or a sliding window that does not meet the conditions is likely a pseudo-mutation caused by sampling interference or electromagnetic noise, which cannot produce a complete voltage relaxation response that can be used for analysis. Only by locking the effective mutation window through two conditions can all invalid signals be filtered out with the smallest analysis unit, and then a complete current mutation time series segment can be obtained by splicing continuous effective mutation windows.

[0025] In this embodiment, firstly, the current transient timing segment is locked, which can capture the entire process of battery polarization from instantaneous generation to gradual relaxation recovery, providing an effective time-domain analysis window for decoupling reversible polarization decay and irreversible aging. Secondly, the effective transient window locked by the first derivative and rate of change ensures that the initial conditions and the severity of each current transient are consistent, avoiding misjudging differences in polarization response as changes in the degree of irreversible aging. Finally, all invalid signals that are not transient are filtered out, ensuring that the voltage transient response segment analyzed subsequently only contains the polarization relaxation signal caused by the current step and the residual signal caused by irreversible aging, without interference from other baseline drift or slow fluctuations.

[0026] Among them, the voltage transient response segment is a segment of the voltage timing sampling sequence that is synchronized with the time axis of the current change timing segment.

[0027] S2. Perform time-series second-order differential calculation on the voltage transient response segment corresponding to the current sudden change time sequence segment to generate the voltage recovery time sequence trajectory.

[0028] Since the voltage response after a sudden change in battery current is a composite signal of the superposition of ohmic voltage drop transient, reversible polarization relaxation, and irreversible aging steady state, the original voltage sequence changes gradually and cannot distinguish the time domain boundaries of each electrochemical process. The simple threshold method is easily affected by sampling noise and operating condition fluctuations and cannot lock the pure polarization relaxation process. Therefore, the voltage recovery time sequence trajectory is determined by the time-series second-order differential progressive calculation method. The specific process is as follows: S201, calculate the difference between the voltage sampling values ​​of adjacent sampling times within the voltage transient response segment, and take the ratio of the difference to the corresponding sampling time interval as the voltage first-order differential value to form the voltage first-order differential sequence.

[0029] S202. Calculate the difference between the first-order voltage derivative values ​​at adjacent sampling times within the first-order voltage derivative sequence, and use the ratio of the difference to the corresponding sampling time interval as the second-order voltage derivative value to form the second-order voltage derivative sequence.

[0030] S203. Based on the second-order differential voltage sequence, locate the voltage transient jump inflection point and polarization relaxation termination point corresponding to the voltage transient response segment, and extract the relaxation time sequence interval from the voltage transient jump inflection point to the polarization relaxation termination point.

[0031] S204. Combining the values ​​of the relaxation time interval with the reference voltage, perform time sequence feature restoration and baseline alignment to generate a voltage recovery time sequence trajectory.

[0032] The physical meaning of the first derivative of voltage is the instantaneous rate of change of battery voltage within a unit sampling interval. The larger the absolute value, the faster the voltage jump or recovery speed. The positive or negative sign of the value corresponds to the direction of voltage rise or fall. The ohmic voltage drop jump at the moment of current change will cause the absolute value of the first derivative to reach its peak value. During polarization relaxation, the absolute value of the first derivative gradually decays as the voltage recovers and approaches 0 after relaxation.

[0033] The physical meaning of the second derivative of voltage is the instantaneous rate of change of voltage, that is, the acceleration of voltage change. The larger the absolute value, the more drastic the change trend of voltage has changed. The positive or negative sign of the value corresponds to the acceleration or deceleration of the change trend. The voltage transient jump inflection point corresponds to the extreme point of the second derivative, and the polarization relaxation termination point corresponds to the stable range where the second derivative continuously approaches 0.

[0034] The relaxation time series is a complete time domain interval that starts at the voltage transient jump inflection point and ends at the polarization relaxation termination point. It is the analysis interval for the reversible polarization relaxation behavior of the battery. The starting voltage transient jump inflection point is the critical moment when the ohmic voltage drop transient process caused by the current change completely ends and the reversible polarization relaxation process officially starts. The ending polarization relaxation termination point is the critical moment when all reversible polarization processes, such as double-layer capacitance and electrochemical polarization, are completely restored and the voltage enters a steady state dominated by irreversible aging.

[0035] The voltage change within the relaxation time interval is entirely dominated by the relaxation recovery behavior of the battery's reversible polarization, excluding the transient jumps in ohmic voltage drop in the early stage and the irreversible aging steady-state voltage in the later stage. The time domain length, voltage change amplitude, recovery rate, and other characteristics of the interval can directly and quantitatively characterize the degree of degradation of the battery's reversible polarization, and serve as the time domain boundary for distinguishing between reversible polarization and irreversible aging.

[0036] This embodiment first uses first-order differentiation to strip away the rate characteristics of voltage change, then uses second-order differentiation to locate the start and end inflection points of the polarization relaxation process, eliminating ohmic transient segments and stable aging segments unrelated to polarization relaxation, and finally generating a voltage recovery time-series trajectory through feature restoration and baseline alignment. This is a pure time-domain signal containing only reversible polarization relaxation behavior, providing a precise and repeatable analysis carrier for decoupling reversible polarization and irreversible aging.

[0037] When a battery experiences a sudden current change, a transient jump in ohmic voltage drop immediately occurs. This process causes a very strong peak in the second derivative of the voltage. This peak is the critical point at which the ohmic transient ends and reversible polarization relaxation officially begins. The complete termination of reversible polarization relaxation in a battery is marked by the voltage entering a sustained stable state. At this point, the rate of voltage change is almost zero, and the corresponding second derivative of the voltage will also continuously approach zero without irregular fluctuations. Therefore, the process of locating the voltage transient jump inflection point and polarization relaxation termination point corresponding to the voltage transient response segment based on the second derivative of the voltage sequence is as follows: using the sampling time aligned with the time axis of the current sudden change sequence segment as the time reference, a preset inflection point detection sliding window is used to traverse the second derivative of the voltage sequence.

[0038] Calculate the maximum absolute value of the second-order derivative of the voltage within the sliding window at each inflection point.

[0039] When the absolute maximum value exceeds the preset transient threshold, and the time difference between the sampling time corresponding to the absolute maximum value and the start time of the current transient timing segment is within the preset transient response time window, the sampling time is marked as the voltage transient inflection point. The transient threshold ranges from 0.01V / s² to 0.5V / s², and the transient response time window is the interval from 0ms to 100ms after the start time of the current transient timing segment.

[0040] In actual measurements, sampling spikes and electromagnetic interference from the wiring harness can generate spurious second-order differential spikes. The amplitude of these spikes is usually very small, so the maximum absolute value exceeds the jump detection threshold. Therefore, these noise interferences are directly filtered out, retaining only the true voltage jump signal. Meanwhile, the ohmic transient of the battery is an instantaneous response triggered by a sudden current change. It will always occur within a very short time after the current change and cannot occur independently of the current change. Therefore, a constraint is added on the time difference within the transient response time window.

[0041] Within the time interval following the voltage transient inflection point, a preset relaxation termination detection sliding window is used to traverse the second-order differential voltage sequence.

[0042] When, within a predetermined number of consecutive relaxation termination detection sliding windows, the absolute values ​​of all second-order voltage derivatives are less than a predetermined relaxation termination threshold, and the variance of the corresponding second-order voltage derivatives within the relaxation termination detection sliding window is less than a predetermined stability determination threshold, a predetermined number of consecutive relaxation termination detection sliding windows are locked. The predetermined number is 3-5, the relaxation termination threshold ranges from 0.0001V / s² to 0.001V / s², and the stability determination threshold ranges from 1e-8 to 1e-6.

[0043] The starting sampling time of the first relaxation termination detection sliding window of a continuously preset number of relaxation termination detection sliding windows is marked as the polarization relaxation termination point.

[0044] If we only look at single-point values ​​or a single relaxation termination detection sliding window, it's easy to misjudge the drop in single-point values ​​caused by sampling noise or the occasional stability of values ​​in a single relaxation termination detection sliding window as polarization relaxation termination. This ultimately leads to an incomplete extraction of the relaxation interval, incorrectly calculating irreversible aging residuals as unrecovered reversible polarization signals. Therefore, ensuring that a continuously preset number of windows meet the conditions verifies that this stable state is continuously stable and not accidental; all second-order derivative absolute values ​​being less than the relaxation termination threshold confirms numerically that the voltage no longer has a significant upward or downward trend, and the dynamic process of polarization relaxation has stopped; and variance being less than the stability judgment threshold confirms from the perspective of fluctuation characteristics that the signal within the window does not have small oscillations and is a true steady state, rather than a fluctuating signal that just happens to fall within the threshold.

[0045] Furthermore, the process of restoring the timing features and aligning the baseline to generate the voltage recovery timing trajectory is as follows: extract the values ​​of the corresponding relaxation timing intervals from the second-order differential voltage sequence to obtain the second-order differential sequence of the relaxation interval.

[0046] For the second-order differential sequence in the relaxation interval, perform point-by-point first-order numerical integration according to the step size of the original sampling time sequence to obtain the restored first-order differential sequence of the voltage in the relaxation interval.

[0047] The voltage amplitude reconstruction sequence in the relaxation interval is obtained by performing point-by-point second-order numerical integration on the first-order differential reconstruction sequence of the voltage in the relaxation interval according to the step size of the original sampling time sequence.

[0048] The average value of the voltage samples within a preset time period before the start of the current change time segment in the preprocessed voltage time sequence is taken as the reference voltage.

[0049] Using a predetermined reference voltage as a unified alignment reference, the deviation between the final steady-state amplitude of the voltage amplitude recovery sequence in the relaxation interval and the reference voltage is calculated. The voltage amplitude of all time-series sampling points in the sequence is synchronously superimposed with this deviation value to complete the overall linear baseline translation correction and generate a voltage recovery time-series trajectory that is completely synchronized with the original sampling time sequence.

[0050] In this embodiment, the first-order numerical integral restores the second-order differential sequence to the instantaneous recovery rate sequence of polarization relaxation. The second-order numerical integral further restores it to the voltage amplitude sequence containing only the dynamic changes of polarization relaxation, preserving the time-domain dynamic characteristics of polarization recovery within the relaxation interval, while completely eliminating all DC and slowly varying interference components.

[0051] S3. Calculate the timing overlap between the starting and ending points of the voltage recovery timing trajectory and determine the irreversible aging timing segment.

[0052] Since the reversible polarization of the battery can be completely recovered through the relaxation process, ideally the voltage after the polarization relaxation ends should completely coincide with the reference voltage before the polarization starts. However, irreversible aging is caused by permanent structural changes such as loss of active materials and thickening of the SEI film inside the battery, which will cause the voltage to fail to return to the reference voltage after the polarization relaxation ends, forming a permanent residual. Therefore, the process of determining the irreversible aging time sequence segment by the time sequence coincidence is as follows: S301, calculate the absolute value of the difference between the voltage corresponding to the starting point and the ending point of the trajectory and the reference voltage, respectively, to obtain the starting point deviation value and the ending point deviation value.

[0053] S302. Based on the starting point deviation value and the ending point deviation value, calculate the temporal overlap between the starting point and the ending point of the trajectory: ,in, This refers to the degree of temporal overlap. This is the endpoint deviation value. This is the starting deviation value. for and The absolute value of the difference, in physical terms, represents the magnitude of the change in battery voltage relative to the reference voltage during the entire process of a single polarization relaxation. This is the reference voltage.

[0054] Timing overlap quantifies the degree to which the start and end points of a single voltage recovery trajectory coincide with the reference voltage. Essentially, it characterizes the stability of the battery's reversible polarization relaxation process and the degree of irreversible aging accumulation. A timing overlap closer to 1 indicates... The closer the voltage approaches 0, the more consistent the offsets of the start and end points of the relaxation process relative to the reference voltage. This indicates that the reversible polarization relaxation process of the battery is stable and repeatable, with no additional permanent voltage offset before and after polarization relaxation, and no obvious irreversible aging of the battery, placing it within the normal operating range.

[0055] S303. If the timing overlap is lower than the aging judgment threshold, the timing interval corresponding to the voltage recovery timing trajectory is determined to be the irreversible aging candidate interval.

[0056] S304. The time series range covered by the continuously distributed irreversible aging candidate intervals is labeled as an irreversible aging time series segment.

[0057] Irreversible aging of batteries is a gradual, cumulative, and irreversible process. Once it enters the irreversible aging stage, the time-series overlap of all subsequent voltage recovery time-series trajectories will remain below the aging judgment threshold. There will be no situation where a single interval has low overlap and subsequent intervals return to normal. First, by identifying irreversible aging candidate intervals through single time-series interval judgment, the first step of initial screening is completed, filtering out all normal intervals that meet the overlap standard and have no aging risk. Then, the second step of verification is completed through continuously distributed candidate intervals. Only continuously low overlap intervals will be identified as the final irreversible aging time-series segments.

[0058] S4. Perform voltage residual change consistency and voltage trend coherence analysis on irreversible aging time segments to determine the battery degradation inflection point.

[0059] It should be noted that, as Figure 2 As shown, the specific process is as follows: S401, extract the voltage residuals of each voltage recovery time sequence trajectory within the irreversible aging time sequence segment, and arrange them in chronological order to generate an irreversible aging residual time sequence. The voltage residual is the absolute value of the difference between the endpoint steady-state amplitude and the reference voltage at the endpoint of the trajectory.

[0060] S402. Use a fixed-step, non-overlapping sliding window to traverse the irreversible aging residual time series to obtain several consecutive time series analysis windows. Calculate the consistency of voltage residual changes and the coherence of voltage trends for each time series analysis window.

[0061] S403. Calculate the statistical median of consistency and the statistical median of continuity based on the consistency of voltage residual changes and the continuity of voltage trends, respectively, using the full time series analysis window.

[0062] S404. Select time series analysis windows with voltage residual change consistency not lower than the consistency statistical median and voltage trend continuity not lower than the continuity statistical median, and mark them as valid trend windows.

[0063] S405. Calculate the aging change rate of each effective trend window in chronological order. The aging change rate is the ratio of the difference between the first and last voltage residuals within the effective trend window to the duration of the effective trend window.

[0064] S406. When the absolute value of the difference between the aging change rates of two adjacent effective trend windows exceeds twice the standard deviation of the aging change rate of the full effective trend window, the boundary moment of the two adjacent effective trend windows is marked as the battery degradation inflection point.

[0065] The battery degradation inflection point is the critical moment when irreversible aging of the battery changes from a steady and slow decline to a rapid and accelerated deterioration. It marks the point in time when the internal performance damage of the battery enters an irreversible accelerated stage and the aging characteristics undergo fundamental changes.

[0066] The formula for calculating the consistency of voltage residual variation is as follows: ,in, For the consistency of voltage residual change, This refers to the number of times that the direction of change of adjacent voltage residuals within the time series analysis window is the same as the overall trend direction of the time series analysis window. This represents the total number of voltage residual samples within the time-series analysis window. This represents the total number of changes in the voltage residual within the time series analysis window.

[0067] The direction of change in adjacent voltage residuals is determined by the sign of the difference between the subsequent and preceding voltage residuals; a positive difference indicates a positive trend, and a negative difference indicates a negative trend. The overall trend direction is determined by the sign of the difference between the last and first voltage residuals in the time series analysis window; a positive difference indicates a positive trend, and a negative difference indicates a negative trend. Furthermore, since the battery is in an aging state, the overall trend is positive; a negative result indicates an anomaly.

[0068] The consistency of voltage residual change refers to the degree to which the adjacent change directions of voltage residual within the time series analysis window are consistent with the overall aging trend direction. The higher the value, the more stable the aging change direction is and the less it fluctuates back and forth. The lower the value, the more chaotic the residual change direction is and the more seriously the aging trend is disturbed.

[0069] The formula for calculating voltage trend consistency is: .

[0070] .

[0071] .

[0072] .

[0073] in, This is the sequence number of the voltage residual. For the first The ideal residuals corresponding to each sequence number , and These are the first and second items in the time series analysis window, respectively. The and the first Voltage residual, The total deviation between reality and ideal. For the maximum theoretical deviation, This represents the consistency of voltage trend.

[0074] Voltage trend consistency is a quantitative indicator that measures the degree to which the actual voltage residual sequence matches the ideal uniform aging trend within the time series analysis window. The closer the value is to 1, the more likely the actual residual sequence coincides with the ideal uniform aging line, indicating a very smooth and uniform aging process without obvious jumps or disturbances, which is a sign of stable battery aging.

[0075] In this embodiment, the consistency of voltage residual change can filter out non-steady-state windows with chaotic change directions, and the consistency of voltage trend can eliminate abnormal fluctuation windows that deviate from the uniform aging trend. The effective trend windows selected by the two combined have the characteristics of uniform direction and smooth trend of residual change. The aging change rate calculated based on these effective trend windows can truly characterize the aging rate of the battery at different stages. Then, by using the abrupt change (more than 2 standard deviations) of the aging change rate of adjacent effective trend windows to calibrate the degradation inflection point, the critical moment when the battery transitions from stable aging to accelerated aging can be accurately captured.

[0076] S5. Perform time-series feature encoding on the irreversible aging time series segment after the battery degradation inflection point, extract the instantaneous change in voltage residual and the cumulative amount of voltage residual, calculate the SOH prediction value and make corrections.

[0077] like Figure 3 As shown, the specific process is as follows: S501, extract the voltage residuals corresponding to all voltage recovery time-series trajectories within the irreversible aging time-series segment after the battery degradation inflection point, and perform time-series feature encoding with equal time steps. After encoding, the data is input into a lightweight deep learning time-series model. The encoding dimension matches the original sampling frequency to achieve standardized extraction of residual time-series features. After encoding, the instantaneous change in voltage residuals and the cumulative amount of voltage residuals after the inflection point are extracted simultaneously.

[0078] S502. Calculate the ratio of the cumulative voltage residual after the inflection point to the theoretical total increase in voltage residual from the degradation inflection point to complete failure, thus obtaining the residual aging percentage. The value of the theoretical total increase in voltage residual is determined by the battery model and failure threshold, and is a fixed endpoint reference value.

[0079] S503. Subtract the product of the residual aging ratio and the adaptation coefficient from 100% to obtain the SOH predicted value. Then, perform a linear weighted sum with the SOH model predicted value output by the lightweight deep learning time series model to obtain the initial SOH predicted value. For example, the weight of the SOH predicted value is 0.4, and the weight of the SOH model predicted value is 0.6.

[0080] S504. Multiply the initial SOH prediction value by 1 and the difference between the fluctuation coefficient and the value to obtain the final SOH prediction value. The fluctuation coefficient is the ratio of the standard deviation to the average value of the instantaneous change in voltage residual. The larger the value, the more drastic the fluctuation in the aging rate, the more unstable the aging process, and the lower the reliability of the battery performance.

[0081] The fit factor is a calibration factor that matches the battery type, ranging from 0.8 to 1.2. It is calibrated through aging tests on batteries of the same type and is used to correct the nonlinear relationship between voltage residual and battery capacity decay. The closer the initial SOH prediction value is to 100%, the better the battery's health; the lower the value, the more severe the aging.

[0082] Among them, the instantaneous change in voltage residual is the difference in voltage residual between adjacent time steps in the coding sequence. The larger the absolute value, the more violent the instantaneous fluctuation of the aging rate per unit time, and the more unstable the aging process.

[0083] The cumulative voltage residual after the inflection point is the sum of all voltage residuals in the encoded sequence. The cumulative voltage residual after the inflection point is monotonically increasing with the battery capacity decay. The larger the value, the more severe the aging. The lower the SOH, the more severe the total degree of irreversible aging after the inflection point and the worse the battery health.

[0084] The closer the residual aging percentage is to 1, the closer the current aging process is to complete failure, and the smaller the remaining healthy space.

[0085] It should be noted that the input of the lightweight deep learning time series model is the sequence of voltage residuals corresponding to all voltage recovery time series trajectories within the irreversible aging time series segment after the battery degradation inflection point, encoded by time series features with equal time steps. Before inputting the encoded sequence into the model, it is necessary to perform uniform preprocessing. First, the average length of the irreversible aging time series segment in the same type of battery aging experiment is used as the reference length to complete the sequence alignment. If the length is shorter than the reference length, zeros are added to the end. If the length is longer than the reference length, the first half is truncated. Then, the extreme values ​​of the encoded sequence in the training samples of the same type of battery are used as the upper and lower limits to uniformly map the aligned sequence values ​​to the interval between 0 and 1, ensuring that the length and numerical range of the input data completely match the model requirements.

[0086] The lightweight deep learning temporal model adopts a lightweight long short-term memory network structure, which consists of only three layers: an input layer, a single hidden layer, and an output layer. The dimension of the input layer is the same as the length of the aligned encoding sequence, and it is used to receive the preprocessed voltage residual encoding sequence. The hidden layer has 32 neurons and uses the ReLU activation function, retaining only the core temporal memory and aging feature extraction functions of the long short-term memory network. The output layer has only 1 neuron and no additional activation function.

[0087] The model training uses full-lifecycle aging experimental data of the same type of battery. First, a training sample set is constructed. Each sample contains a preprocessed voltage residual encoding sequence after the corresponding battery degradation inflection point, as well as the actual SOH value obtained by battery capacity testing at the corresponding time series. Then, the sample set is divided into training set, validation set and test set in a ratio of 7:2:1. The Adam optimizer is used in the training process. The initial learning rate is set to 0.001. Every 100 training rounds, the learning rate is reduced to 0.1 times the original value. The mean squared error is used as the loss function and the batch size is set to 16. When the validation set loss does not decrease significantly for 5 consecutive rounds, training is stopped and the model parameters are saved. If the validation set loss continues to decrease, training is stopped after 200 rounds to avoid model overfitting.

[0088] When generating SOH model predictions after training, the encoded sequence to be predicted is first preprocessed according to the same rules as the training process. Then, the preprocessed sequence is input into the model, passed through the input layer to the hidden layer to extract time-series features related to irreversible battery aging. Finally, the corresponding SOH model predictions are directly output through the output layer. The output results range from 0% to 100%, and can be directly linearly weighted and summed with the SOH predictions calculated by the residual aging ratio to obtain the initial SOH predictions.

[0089] S6. Calculate the irreversible aging rate after the battery degradation inflection point based on the voltage residual, determine the RUL prediction value based on the irreversible aging rate and make corrections.

[0090] It should be noted that the specific process of determining and correcting the RUL prediction value is as follows: S601, extract the voltage residual of the time series corresponding to the irreversible aging time series segment after the battery degradation inflection point, and generate the voltage residual time series sequence after the inflection point. The start time of the voltage residual time series after the inflection point is the sampling time corresponding to the battery degradation inflection point, and the end time is the current sampling time corresponding to the prediction.

[0091] S602. Divide the difference between the voltage residual at the end time and the voltage residual at the beginning time in the voltage residual time sequence after the inflection point by the total time sequence length corresponding to the voltage residual time sequence after the inflection point to obtain the irreversible aging rate.

[0092] S603. Divide the difference between the battery voltage residual failure threshold and the current voltage residual by the irreversible aging rate to obtain the initial RUL prediction value.

[0093] S604. Multiply the initial RUL prediction value by 1 and the difference between the volatility coefficient to obtain the final RUL prediction value.

[0094] The post-inflection point voltage residual time series is a set of voltage residuals arranged in time after the degradation inflection point, reflecting the temporal changes in aging after the inflection point. The voltage residual at the termination moment is the aging residual at the last moment, and the voltage residual at the beginning moment is the initial aging residual corresponding to the inflection point. The larger the difference between the two and the shorter the total time series length, the greater the irreversible aging rate. The irreversible aging rate is the increment of the voltage residual per unit time. The larger the value, the faster the battery ages and the shorter the remaining life. The voltage residual failure threshold is the critical residual value at which the battery reaches the failure state, and it is a fixed calibration value. The smaller the difference between the current voltage residual and the failure threshold, the closer the battery is to failure. The initial RUL prediction value is the uncorrected remaining life estimate. The smaller the value, the shorter the remaining life. The fluctuation coefficient is used to characterize aging stability. The larger the value, the more severe the aging fluctuation. The final RUL prediction value is the result after stability correction, which is closer to the actual remaining life of the battery.

[0095] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0096] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0097] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0098] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0099] Finally, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based method for co-predicting battery health status and remaining life, characterized in that, Includes the following steps: Perform time-series first-order differential calculation and current change rate calculation on the current time series segment to determine the current abrupt change time series segment; Perform time-series second-order differential calculations on the voltage transient response segments corresponding to the current abrupt change time series segments to generate voltage recovery time series trajectories; The timing overlap between the start and end points of the voltage recovery timing trajectory is calculated, and the irreversible aging timing segment is determined. Analysis of voltage residual change consistency and voltage trend coherence was performed on irreversible aging time segments to determine the inflection point of battery degradation. The irreversible aging time series segment after the battery degradation inflection point is processed by time series feature encoding. After extracting the instantaneous change and cumulative voltage residual, the SOH prediction value is calculated and corrected. The irreversible aging rate after the battery degradation inflection point is calculated based on the voltage residual, and the RUL prediction value is determined and corrected based on the irreversible aging rate.

2. The method for co-predicting battery health status and remaining life based on deep learning according to claim 1, characterized in that, The process of determining the timing segment of the current abrupt change is as follows: The difference between the current sampled values ​​at adjacent sampling times within the current time sequence segment is calculated, and the ratio of the difference to the corresponding sampling time interval is used as the first derivative value of the current, forming a first derivative sequence of the current. By traversing the first-order derivative sequence of current through a preset sliding window, and taking the average value of all first-order derivative values ​​of current in each sliding window, the current change rate corresponding to each sliding window is obtained, forming a current change rate sequence. If the absolute value of the first derivative of all currents within a certain sliding window is greater than the non-zero threshold, and the variance of the corresponding current change rate and the current change rate of the adjacent sliding window before and after it is greater than the judgment threshold, it is judged as a valid abrupt change window. N consecutive valid mutation windows are labeled as current mutation time series segments.

3. The method for co-predicting battery health status and remaining life based on deep learning according to claim 1, characterized in that, The process of generating the voltage recovery timing trace is as follows: The voltage sample values ​​at adjacent sampling times within the voltage transient response segment are calculated by difference. The ratio of the difference to the corresponding sampling time interval is used as the first-order differential value of the voltage, forming a first-order differential sequence of voltage. The difference between the first-order voltage derivative values ​​at adjacent sampling times within the first-order voltage derivative sequence is calculated, and the ratio of the difference to the corresponding sampling time interval is taken as the second-order voltage derivative value, thus forming the second-order voltage derivative sequence. Based on the voltage second-order differential sequence, the voltage transient jump inflection point and polarization relaxation termination point corresponding to the voltage transient response segment are located, and the relaxation time sequence interval from the voltage transient jump inflection point to the polarization relaxation termination point is extracted. By combining the values ​​of the relaxation time series interval with the reference voltage, time series characteristics are restored and baseline is aligned to generate a voltage recovery time series trajectory.

4. The method for co-predicting battery health status and remaining life based on deep learning according to claim 3, characterized in that, The process of locating the voltage transient jump inflection point and polarization relaxation termination point corresponding to the voltage transient response segment based on the second-order differential voltage sequence is as follows: Using the sampling time aligned with the time axis of the current change time sequence segment as the time reference, a preset inflection point detection sliding window is used to traverse the second-order differential voltage sequence; Calculate the maximum absolute value of the second derivative of the voltage within the sliding window at each inflection point; When the maximum absolute value exceeds the preset transition judgment threshold, and the time difference between the sampling time corresponding to the maximum absolute value and the start time of the current change timing segment is within the preset transient response time window, the sampling time is marked as the voltage transient transition inflection point. Within the time interval following the voltage transient inflection point, a preset relaxation termination detection sliding window is used to traverse the second-order differential voltage sequence. When the absolute value of all second-order voltage derivatives is less than the preset relaxation termination threshold within a consecutive preset number of relaxation termination detection sliding windows, and the variance of the second-order voltage derivatives within the corresponding relaxation termination detection sliding window is less than the preset stability determination threshold, the consecutive preset number of relaxation termination detection sliding windows are locked. The starting sampling time of the first relaxation termination detection sliding window of a continuously preset number of relaxation termination detection sliding windows is marked as the polarization relaxation termination point.

5. The method for co-predicting battery health status and remaining life based on deep learning according to claim 4, characterized in that, The process of performing timing feature restoration and baseline alignment to generate voltage recovery timing trajectories is as follows: Extract the values ​​of the corresponding relaxation time interval from the second-order differential voltage sequence to obtain the second-order differential sequence of the relaxation interval; For the second-order differential sequence in the relaxation interval, perform point-by-point first-order numerical integration according to the step size of the original sampling time sequence to obtain the restored first-order differential sequence of the voltage in the relaxation interval. The voltage amplitude reconstruction sequence in the relaxation interval is obtained by performing point-by-point second-order numerical integration on the first-order differential reconstruction sequence of the voltage in the relaxation interval according to the step size of the original sampling time sequence. Using the reference voltage as the alignment reference, the voltage amplitude restoration sequence in the relaxation interval is baseline shifted and corrected to generate a voltage recovery time sequence trajectory that is completely synchronized with the original sampling time sequence.

6. The method for co-predicting battery health status and remaining life based on deep learning according to claim 1, characterized in that, The process of calculating the time overlap between the starting and ending points of the voltage recovery timeline and determining the irreversible aging timeline segment is as follows: Calculate the absolute value of the difference between the voltage at the starting point and the end point of the trajectory and the reference voltage, respectively, to obtain the starting point deviation value and the end point deviation value; Based on the starting point deviation value and the ending point deviation value, calculate the temporal overlap between the starting point and the ending point of the trajectory: ,in, This refers to the degree of temporal overlap. This is the endpoint deviation value. This is the starting deviation value. for and The absolute value of the difference The reference voltage; If the timing overlap is lower than the aging judgment threshold, the timing interval corresponding to the voltage recovery timing trajectory is determined to be the irreversible aging candidate interval. The time series range covered by the continuously distributed irreversible aging candidate intervals is labeled as an irreversible aging time series segment.

7. The method for co-predicting battery health status and remaining life based on deep learning according to claim 1, characterized in that, The process of analyzing the consistency of voltage residual changes and the coherence of voltage trends in irreversible aging time segments to determine the inflection point of battery degradation is as follows: The voltage residuals of each voltage recovery time sequence within the irreversible aging time sequence segment are extracted and arranged in chronological order to generate an irreversible aging residual time sequence. A fixed-step, non-overlapping sliding window is used to traverse the irreversible aging residual time series to obtain several consecutive time series analysis windows. The consistency of voltage residual changes and the coherence of voltage trends are calculated for each time series analysis window. Based on the full time series analysis window, the statistical median values ​​of consistency and continuity of voltage residual change and voltage trend coherence were calculated respectively. Time series analysis windows that have a voltage residual change consistency not lower than the consistency statistical median and a voltage trend continuity not lower than the continuity statistical median are selected and marked as valid trend windows. The aging rate of change for each effective trend window is calculated sequentially over time. When the absolute value of the difference between the aging change rates of two adjacent effective trend windows exceeds twice the standard deviation of the aging change rate of the full effective trend window, the boundary moment between the two adjacent effective trend windows is marked as the battery degradation inflection point.

8. The method for co-predicting battery health status and remaining life based on deep learning according to claim 1, characterized in that, The process of calculating and correcting the predicted SOH value is as follows: Extract the voltage residuals corresponding to all voltage recovery time series trajectories within the irreversible aging time series segment after the battery degradation inflection point, and perform time series feature encoding with equal time steps. The encoded features are then input into a lightweight deep learning time series model. Simultaneously, extract the instantaneous change in voltage residuals and the cumulative amount of voltage residuals after the inflection point. The ratio of the cumulative voltage residual after the inflection point to the theoretical total increase in voltage residual from the degradation inflection point to the complete failure state is calculated to obtain the residual aging ratio. Subtract the product of the residual aging ratio and the adaptation coefficient from 100% to obtain the SOH prediction value, and then perform a linear weighted sum with the SOH model prediction value output by the lightweight deep learning time series model to obtain the initial SOH prediction value. The final SOH prediction value is obtained by multiplying the initial SOH prediction value by 1 and the difference between the fluctuation coefficient and the initial SOH prediction value.

9. The method for co-predicting battery health status and remaining life based on deep learning according to claim 1, characterized in that, The process of determining and correcting the RUL prediction value is as follows: Extract the voltage residuals of the irreversible aging time series after the battery degradation inflection point, and generate the voltage residual time series sequence after the inflection point; The irreversible aging rate is obtained by dividing the difference between the voltage residual at the end time and the voltage residual at the beginning time in the voltage residual time series after the inflection point by the total time series length corresponding to the voltage residual time series after the inflection point. The initial RUL prediction value is obtained by dividing the difference between the battery voltage residual failure threshold and the current voltage residual by the irreversible aging rate. The final RUL prediction is obtained by multiplying the initial RUL prediction value by 1 and the difference between the volatility coefficient and the initial RUL prediction value.

10. The method for co-predicting battery health status and remaining life based on deep learning according to any one of claims 8-9, characterized in that, The fluctuation coefficient is the ratio of the standard deviation to the average value of the instantaneous change in voltage residual.

Images