Device and method for predicting supercapacitor life under high current overload conditions

Abstract

The present application provides a device and method for predicting the life of a supercapacitor under high current overload conditions. The method includes: constructing a high current test platform; based on the test platform, performing charge and discharge tests on the supercapacitor module, and recording the test parameters of the supercapacitor module; according to the test parameters, constructing a data set including a supercapacitor life variation function and a supercapacitor cell capacitance distribution function, wherein the life variation function is used to characterize the relationship between the total internal resistance and total capacity of the supercapacitor and the number of working times, and the capacitance distribution function is used to characterize the voltage distribution characteristics of the supercapacitor cell; collecting the operating data of the supercapacitor module, the operating data includes the total internal resistance value, the total capacity value, and the current value; based on the life variation function and the capacitance distribution function in the data set, combined with the operating data, the remaining life times of the supercapacitor module are calculated. The present application realizes the accurate prediction of the life of the supercapacitor system under high current overload conditions, and improves the prediction accuracy and reliability.

Description

Technical Field

[0001] The present application relates to the technical field of supercapacitors, and in particular to a device and method for predicting the life of a supercapacitor under high current overload conditions. Background Art

[0002] As a highly efficient energy storage device, supercapacitors are widely used in power electronics, new energy vehicles and other fields. In practical applications, supercapacitors often need to work under high power density requirements, in which case they will face large current overload conditions (current overload is more than 10 times the rated current).

[0003] At present, common supercapacitor life prediction methods mainly include empirical model methods based on test data and online prediction methods based on monitoring data. However, the existing prediction methods are mainly aimed at rated charge and discharge current conditions, and predictions are made by collecting data under rated conditions and combining empirical formulas under rated conditions. This method has the following technical problems: First, the existing empirical formula is not applicable to high current overload conditions; second, the existing method is mainly aimed at supercapacitor monomers, and lacks predictions at the supercapacitor module and energy storage system levels; finally, because the performance degradation mechanism of supercapacitors under high current overload conditions is more complicated, the prediction accuracy and reliability of the existing methods are difficult to meet actual needs. Summary of the invention

[0004] In view of this, the present application provides a device and method for predicting the life of a supercapacitor under high current overload conditions, which solves the problem of low accuracy and poor reliability in the prior art for predicting the life of a supercapacitor under high current overload conditions.

[0005] The present application provides a method for predicting the life of a supercapacitor under a high current overload condition, comprising:

[0006] Constructing a high current test platform, the test platform includes a supercapacitor module, and a charging power supply, a relay, a counter, a charge and discharge switch, a resistor, a shunt, an oscilloscope and a data acquisition instrument electrically connected to the supercapacitor module, wherein the supercapacitor module is composed of a plurality of supercapacitor monomers connected in series and parallel;

[0007] Based on the test platform, the supercapacitor module is charged and discharged, and the test parameters of the supercapacitor module are recorded, wherein the test parameters include effective current value, total internal resistance, total capacity, single cell voltage and working times;

[0008] According to the test parameters, construct a data set including a supercapacitor life variation function and a supercapacitor single capacitance distribution function, wherein the life variation function is used to characterize the relationship between the total internal resistance and the total capacity of the supercapacitor and the number of operations, and the capacitance distribution function is used to characterize the voltage distribution characteristics of the supercapacitor single.

[0009] Collecting operation data of the supercapacitor module, wherein the operation data includes total internal resistance value, total capacity value, and current value;

[0010] Based on the life variation function and the capacitance distribution function in the data set and in combination with the operation data, the remaining life times of the supercapacitor module are calculated.

[0011] Optionally, according to the test parameters, a data set including a supercapacitor life variation function and a supercapacitor single capacitance distribution function is constructed, including:

[0012] The total internal resistance and total capacity data in the test parameters are grouped according to a preset working number interval, and a life variation function is constructed based on the grouped first test parameter;

[0013] The single cell voltage data in the test parameter are grouped according to a preset working number interval, and a supercapacitor single cell charging voltage variation function is obtained based on the grouped second test parameter;

[0014] Taking the cell voltage variation function as input, using a normal distribution fitting algorithm, and outputting a supercapacitor cell pressure normal distribution function under different working times;

[0015] The corresponding relationship between the life variation function and the normal distribution function under different current impact conditions is integrated into an empirical data set.

[0016] Optionally, the operating data of the supercapacitor module is collected in real time by a monitoring system, including:

[0017] Setting up a voltage detection circuit and a current detection circuit to collect voltage and current data of the supercapacitor module;

[0018] Based on the collected voltage and current data, calculating the total internal resistance of the supercapacitor module;

[0019] The constant current charging and discharging method is used to calculate the total capacity value of the supercapacitor module according to the ratio of the voltage change to the current;

[0020] The total internal resistance value, total capacity value and current value are sampled and stored at a preset frequency, and sent to a data processing unit through a remote transmission channel.

[0021] Optionally, based on the life variation function and the capacitance distribution function in the data set and in combination with the operation data, calculating the remaining life times of the supercapacitor module includes:

[0022] Constructing a Graph Spring neural ODE model, representing the supercapacitor module as a dynamic graph structure, wherein each supercapacitor monomer is used as a node in the graph, and the series-parallel relationship between the nodes is used as an edge, and the Graph Spring neural ODE model includes a graph convolution layer, a spring physics layer, a time series ordinary differential equation layer, and a prediction layer;

[0023] Inputting the data set into the Graph Spring neural ODE model for training to obtain a trained model;

[0024] A channel coordination strategy is used to process the operating data, establish a timing association between different sensor channels, and realize a multi-dimensional analysis of electrical characteristics, thermodynamic characteristics, and life characteristics;

[0025] The processed operating data is input into the trained Graph Spring neural ODE model, and the remaining life times of the supercapacitor module are output by combining the clustering algorithm and the minimum variance algorithm.

[0026] Optionally, build a Graph Spring neural ODE model, including:

[0027] In the graph convolution layer, the initial feature vector of the node is input, and ChebNet is used to perform spatial convolution to extract the local features of the node, and a 64-dimensional node embedding vector is output;

[0028] Input the node embedding vector into the spring physical layer, map it to the physical space, define the spring force between nodes based on the attributes of the edge, and simulate the interaction between capacitors;

[0029] Inputting the state vector of the physical space into the time-series ordinary differential equation layer, constructing a neural ordinary differential equation to describe the dynamic evolution of the system, and using an adaptive step-size solver to perform numerical solution;

[0030] The solution result is input into the prediction layer, and the predicted life span and health status of each node are output through the multi-layer perceptron decoder.

[0031] After building the Graph Spring neural ODE model, it also includes:

[0032] Design a multi-objective loss function based on single-objective prediction loss, system prediction loss, and physical consistency loss;

[0033] Dividing the data set into a training set and a validation set, and using the training set for supervised learning in a pre-training stage;

[0034] The pre-trained model parameters are used as initial values, and physical consistency loss is added to optimize them during the physical constraint training phase;

[0035] A sliding time window method is used to obtain the latest operating data, and the training set and model parameters are continuously updated during the online learning phase.

[0036] Optionally, processing the operation data using a channel coordination strategy includes:

[0037] Represent the data of each sensor channel as a sequence of observations with timestamps and construct a multi-channel time series representation;

[0038] Based on the multi-channel time series, a dynamic time warping algorithm is used to perform time alignment and output a time-varying correlation matrix between channels;

[0039] Apply wavelet transform to each channel data after time alignment to obtain multi-scale feature representation;

[0040] Based on the multi-scale features, an attention mechanism is used to calculate the weight coefficients between channels and establish long-term dependencies;

[0041] Based on the weight coefficients and long-term dependencies, key channels and impact paths that have a significant impact on life prediction are identified.

[0042] Optionally, feed the processed run data into the trained Graph Spring neural ODE model, including:

[0043] Construct a multi-subspace matrix including electrical properties, thermodynamic properties and life characteristics, corresponding to different physical properties;

[0044] The optimization objective function is designed based on the nuclear norm, and the rank of each subspace matrix is ​​constrained to promote a low-rank structure;

[0045] The alternating direction multiplier method is used to iteratively update each subspace matrix to achieve matrix recovery;

[0046] Convert the features in the recovered subspace matrix into node attributes of the graph structure;

[0047] The remaining life is calculated based on the comprehensive information of each subspace, and the confidence interval of the prediction result is output at the same time.

[0048] Optionally, after outputting the remaining life times of the supercapacitor module, the method further includes:

[0049] Based on the analysis results of the channel coordination strategy, a channel coordination relationship diagram is generated to show the evolution process of the timing characteristics;

[0050] According to the characteristic changes in the evolution process, a fault characteristic library including electrical characteristics, thermodynamic characteristics and life characteristics is constructed;

[0051] Based on the fault feature library, a multi-dimensional feature clustering method is used to implement fault mode classification;

[0052] Design and implement a threshold-based hierarchical early warning strategy based on the severity of the failure mode;

[0053] Combine the remaining life prediction results and failure mode classification to evaluate the optimal maintenance timing and generate maintenance decision recommendations.

[0054] The embodiment of the present application also provides a device for predicting the life of a supercapacitor under a large current overload condition, comprising:

[0055] A test platform construction unit, used to construct a high-current test platform, the test platform comprising a supercapacitor module, and a charging power supply, a relay, a counter, a charge and discharge switch, a resistor, a shunt, an oscilloscope and a data acquisition instrument electrically connected to the supercapacitor module, wherein the supercapacitor module is composed of a plurality of supercapacitor monomers connected in series and parallel;

[0056] A data acquisition unit, for performing a charge and discharge test on the supercapacitor module based on the test platform, recording the test parameters of the supercapacitor module, wherein the test parameters include effective current value, total internal resistance, total capacity, single cell voltage and working times; and collecting operation data of the supercapacitor module, wherein the operation data includes total internal resistance value, total capacity value and current value;

[0057] A data processing unit, configured to construct a data set including a supercapacitor life variation function and a supercapacitor cell capacitance distribution function according to the test parameters, wherein the life variation function is used to characterize the relationship between the total internal resistance and the total capacity of the supercapacitor and the number of operations, and the capacitance distribution function is used to characterize the supercapacitor cell voltage distribution characteristics;

[0058] A prediction unit is used to calculate the remaining life times of the supercapacitor module based on the life variation function and the capacitance distribution function in the data set in combination with the operation data.

[0059] An embodiment of the present application also provides a computer device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned method for predicting the life of a supercapacitor under high current overload conditions.

[0060] An embodiment of the present application also provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable a computer to execute the above-mentioned method for predicting the life of a supercapacitor under a large current overload condition.

[0061] An embodiment of the present application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the method for predicting the life of a supercapacitor under the above-mentioned high current overload condition.

[0062] This application has the following technical effects:

[0063] By building a special high-current test platform and data acquisition system, accurate measurement of supercapacitor performance parameters under high-current overload conditions is achieved;

[0064] The Graph Spring neural ODE model combined with the channel coordination strategy improves the accuracy and reliability of life prediction;

[0065] Based on multi-subspace matrix decomposition and fault feature library, system-level health status assessment and early warning are realized;

[0066] Through the online learning mechanism, the prediction model can be continuously optimized and adapted to different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for describing the embodiments are briefly introduced below.

[0068] Figure 1 A flow chart of a method for predicting the life of a supercapacitor under high current overload conditions provided in an embodiment of the present application;

[0069] Figure 2 A schematic diagram of the structure of a high current test platform provided in an embodiment of the present application;

[0070] Figure 3 A schematic diagram of a process for constructing an experience data set based on test data in an embodiment of the present application;

[0071] Figure 4 This is a schematic diagram of the structure of the Graph Spring neural ODE model in the embodiment of this application;

[0072] Figure 5 A schematic diagram showing the visual display of the prediction results in the embodiment of the present application;

[0073] Figure 6 This is a structural block diagram of a supercapacitor life prediction device under high current overload conditions provided in an embodiment of the present application.

[0074] It should be noted that these drawings are only used to illustrate the basic concept of the present application and should not be understood as limiting the scope of protection of the present application. Without departing from the spirit and scope of the present application, those skilled in the art may make various equivalent modifications or substitutions to these drawings, and these modifications or substitutions should be included in the scope of protection of the present application. DETAILED DESCRIPTION

[0075] like Figure 1 As shown, the embodiment of the present application provides a method for predicting the life of a supercapacitor under a large current overload condition, comprising the following steps:

[0076] S1: construct a high current test platform, the test platform includes a supercapacitor module, and a charging power supply, a relay, a counter, a charge and discharge switch, a resistor, a shunt, an oscilloscope and a data acquisition instrument electrically connected to the supercapacitor module, wherein the supercapacitor module is composed of a plurality of supercapacitor monomers connected in series and parallel;

[0077] like Figure 2 As shown, the high current test platform is mainly used to simulate the working state of supercapacitors under high current overload conditions and obtain their performance parameters.

[0078] First of all, the core component of the test platform is the supercapacitor module, which is composed of multiple supercapacitor monomers connected in series and parallel. The series connection of supercapacitor monomers is used to increase the operating voltage of the system, while the parallel connection is used to increase the total capacity of the system to meet the requirements of high current charging and discharging.

[0079] Figure 2 The high current test platform includes:

[0080] Device under test: Supercapacitor module, consisting of multiple cells connected in series and parallel.

[0081] Charging power supply: Select a suitable constant current power supply for charging as needed and set the maximum charging voltage.

[0082] Relay: used to open and close the charging and discharging switches to achieve automated testing.

[0083] Counter: used to control the charge and discharge time of the relay.

[0084] Charge and discharge switch: used to connect the charge and discharge circuit.

[0085] Resistance: Select the appropriate resistance value according to the required overcurrent. The calculation formula is R = U / Imax-R internal resistance.

[0086] Shunt: used to measure the current during charging and discharging. The conversion is done as follows: I = Ushunt / Rshunt.

[0087] Oscilloscope: Collect the voltage across the shunt and calculate the discharge current.

[0088] Data acquisition instrument: used to collect voltage information of all supercapacitor cells.

[0089] In one embodiment, in the charging circuit, the charging power source is connected to the supercapacitor module via a charging switch.

[0090] Among them, the charging power supply adopts a constant current source, the output current can be adjusted according to needs, and the maximum charging voltage value can be set to prevent the supercapacitor from overcharging.

[0091] The relay is installed in the charging circuit to control the on and off of the charging circuit.

[0092] The counter is connected to the relay, and the opening and closing time of the relay is controlled through the timing function to realize the automatic control of the charging and discharging process.

[0093] The discharge circuit is composed of a discharge switch, a resistor and a shunt in series.

[0094] The resistance value of the resistor needs to be selected according to the expected overload current, and its calculation formula is:

[0095] R = U / Imax-R 内阻

[0096] Where U is the rated voltage of the supercapacitor, Imax is the expected maximum discharge current, and R 内阻 is the internal resistance of the supercapacitor module.

[0097] The shunt is connected in series in the discharge circuit to measure the discharge current.

[0098] The oscilloscope is connected in parallel with the shunt, and the actual discharge current is calculated by measuring the voltage value across the shunt. The calculation formula is:

[0099] I = U 分流器 / R 分流器

[0100] Among them U 分流器 is the voltage across the shunt, R 分流器 is the resistance of the shunt.

[0101] In order to realize the status monitoring of each cell of the supercapacitor module, the data acquisition instrument is connected in parallel with each supercapacitor cell through a sampling line.

[0102] The data acquisition instrument is mainly responsible for collecting the voltage information of each monomer. The sampling frequency can be adjusted according to the test requirements and is usually set to 1Hz-10Hz.

[0103] In addition, the data acquisition instrument is also equipped with a temperature sensor to monitor the temperature changes of the supercapacitor module during the charging and discharging process.

[0104] In actual applications, in order to ensure the safety and reliability of the test, the test platform also sets up multiple protection mechanisms. For example, an overvoltage protection circuit is set in the charging circuit, and the charging circuit is automatically disconnected when the supercapacitor voltage exceeds the set value; an overcurrent protection circuit is set in the discharge circuit, and the discharge circuit is cut off when the discharge current exceeds the safety threshold; at the same time, a temperature protection circuit is also set, and the test is stopped and an alarm is issued when the module temperature is too high.

[0105] It should be noted that each component of the test platform uses high-precision devices to ensure the accuracy of the test data.

[0106] For example, the voltage measurement accuracy is better than 0.1%, the current measurement accuracy is better than 0.5%, and the temperature measurement accuracy is better than ±1°C. All measurement data are collected and transmitted to the host computer through the data acquisition instrument to achieve real-time monitoring and storage of data.

[0107] S2: Based on the test platform, perform a charge and discharge test on the supercapacitor module, and record the test parameters of the supercapacitor module, wherein the test parameters include effective current value, total internal resistance, total capacity, single cell voltage and working times;

[0108] After completing the construction of the test platform, it is necessary to conduct a systematic charge and discharge test on the supercapacitor module to obtain its performance parameters under high current overload conditions. The specific test process is as follows:

[0109] First, set the test parameters and control strategy. Set the charging current, maximum charging voltage, discharge resistance value, and number of cycles through the host computer.

[0110] The charging current is generally set to 0.5-1 times the rated current of the supercapacitor to ensure the safety of the charging process; the discharge resistor must be selected to ensure that the discharge current reaches more than 10 times the rated current to simulate the actual overload condition. At the same time, the timing parameters of the counter are set to control the time interval between each charge and discharge.

[0111] During the charging stage, the counter controls the relay to close the charging switch, and the charging power supply provides constant current charging to the supercapacitor module. During this process, the data acquisition instrument records the voltage value of each supercapacitor cell in real time. When the voltage of any cell reaches its rated value, or the total voltage of the module reaches the set maximum charging voltage, the control system immediately disconnects the charging circuit and completes the charging process.

[0112] Then it enters the discharge phase, and the counter controls the relay to switch to the discharge circuit. The supercapacitor module discharges with a large current through the discharge resistor, and the shunt and oscilloscope monitor the change of the discharge current in real time. The effective value of the current is obtained by calculating the root mean square of the current during the discharge process.

[0113] In order to accurately obtain the total internal resistance of the supercapacitor module, the current step method is used for measurement.

[0114] Specifically, at the moment of discharge start, the voltage drop value ΔV and the corresponding current change value ΔI are recorded, and the total internal resistance R = ΔV / ΔI. Considering the measurement accuracy, each measurement is repeated 3 times and the average value is taken as the final result.

[0115] The total capacity is measured by constant current discharge method. Record the voltage change curve during the discharge process, according to the formula:

[0116] C = I × Δt / ΔV

[0117] Calculate the capacity value, where I is the discharge current, Δt is the discharge time, and ΔV is the voltage change. It should be noted that in order to ensure measurement accuracy, data in the voltage linear change range should be selected for calculation.

[0118] Throughout the test, the data acquisition system continuously records the following parameters:

[0119] The real-time voltage of each supercapacitor cell, with a sampling frequency of 1Hz;

[0120] The total voltage and discharge current of the module are sampled at 10kHz to capture transient characteristics;

[0121] Temperature data of key points, including monomer surface temperature and ambient temperature;

[0122] The number of charge and discharge cycles is the basic parameter for life evaluation.

[0123] Multiple protection measures are set during the test: when any single cell voltage exceeds 1.1 times the rated value, the module temperature exceeds 65°C, or the discharge current exceeds 1.2 times the set value, the system automatically stops the test and alarms. All test data is uploaded to the database in real time and automatically backed up to ensure data integrity and reliability.

[0124] In the long-term cycle test, the supercapacitor module is fully tested every fixed number of times (such as 1,000 times), including detailed measurements of parameters such as internal resistance and capacity, to track its performance degradation trend. The test continues until the supercapacitor module reaches the end-of-life condition: the capacity drops to 80% of the initial value or the internal resistance increases to 200% of the initial value.

[0125] S3: constructing a data set including a supercapacitor life variation function and a supercapacitor cell capacitance distribution function according to the test parameters, wherein the life variation function is used to characterize the relationship between the total internal resistance and the total capacity of the supercapacitor and the number of operations, and the capacitance distribution function is used to characterize the voltage distribution characteristics of the supercapacitor cell;

[0126] In order to build a reliable data set to support subsequent life prediction, it is necessary to systematically process and analyze the raw data obtained from the test. The specific process of data processing and function construction is described in detail below.

[0127] like Figure 3 As shown, according to the test parameters, a data set including a supercapacitor life variation function and a supercapacitor single capacitance distribution function is constructed, including:

[0128] S3.1: Grouping the total internal resistance and total capacity data in the test parameters according to a preset working number interval, and constructing a life variation function based on the grouped first test parameter.

[0129] In the processing of total internal resistance and total capacity data, the working number interval is first set to 500 times, and the test data is divided into multiple interval groups according to the working number.

[0130] For each interval group, the statistical characteristics of the total internal resistance and total capacity in the interval are calculated, including the mean, standard deviation and coefficient of variation. Based on these statistical characteristics, the least squares method is used to fit the life variation function.

[0131] Specifically, the variation function of the total internal resistance adopts an exponential model:

[0132] R(n) = (1 + αn^β)

[0133] in is the initial internal resistance, n is the number of working times, α and β are unknown coefficients;

[0134] The change function of total capacity adopts the power function model:

[0135] C(n) = (1 - γn^δ)

[0136] in is the initial capacity, γ and δ are unknown coefficients.

[0137] The optimal values ​​of these model parameters are determined by minimizing the residual sum of squares.

[0138] S3.2: Grouping the cell voltage data in the test parameters according to a preset working number interval, and obtaining a supercapacitor cell charging voltage variation function based on the grouped second test parameters.

[0139] For the processing of cell voltage data, the data is also grouped according to the interval of 500 working times. In each group, the voltage values ​​of all supercapacitor cells at the end of charging are collected to form a voltage data matrix. Through the polynomial regression method, the relationship between the change of the cell charging voltage and the working times is established, that is:

[0140] V(n) = + n+ n² + n³

[0141] Where V(n) is the charging voltage, n is the number of operations, ~ is the fitting coefficient. This function can reflect the gradual change trend of the monomer voltage with the number of cycles.

[0142] S3.3: Taking the cell voltage variation function as input, using a normal distribution fitting algorithm, and outputting a supercapacitor cell pressure normal distribution function under different working times.

[0143] In order to accurately describe the distribution characteristics of the pressure-bearing capacity of supercapacitor cells, the normal distribution fitting algorithm is introduced based on the cell voltage change function. The specific steps are as follows:

[0144] First, for each working number point, the actual voltage values ​​of all monomers are collected, and their mean μ(n) and standard deviation σ(n) are calculated.

[0145] Then, the maximum likelihood estimation method is applied to fit the normal distribution function:

[0146] f(V,n) = (1 / (2πσ²(n)))exp(-(V-μ(n))² / (2σ²(n)))

[0147] Where V is the voltage value and n is the number of operations. The chi-square test is used to verify the fitting effect and ensure the accuracy of the distribution function.

[0148] S3.4: Integrate the corresponding relationship between the life variation function and the normal distribution function under different current impact conditions into an empirical data set.

[0149] It is necessary to integrate the corresponding relationships of the above different functions under various current shock conditions into a complete empirical data set.

[0150] First, the test data is divided into multiple levels according to the current impact intensity (such as 5 times the rated current, 10 times the rated current, etc.). For each current level, a data structure containing the following contents is established:

[0151] The parameter matrix of the life variation function includes the α and β values ​​of the internal resistance function and the γ and δ values ​​of the capacity function;

[0152] The coefficient matrix of the cell voltage change function, including the polynomial fit ~ value;

[0153] The parameter matrix of the normal distribution function, including the changing rules of the mean function μ(n) and the standard deviation function σ(n);

[0154] The correlation coefficient matrix between parameters is used to characterize the correlation between different feature quantities.

[0155] The entire data set is stored in a hierarchical structure to facilitate quick retrieval and call during subsequent model training.

[0156] At the same time, to ensure data quality, an outlier detection mechanism was set up: data points that deviate more than 3 times the standard deviation are marked and corrected through interpolation methods. In addition, a data update mechanism was established to allow dynamic updating of relevant function parameters when new test data is obtained, so that the data set can be continuously optimized and improved.

[0157] In one embodiment, the data processing flow may also be as follows:

[0158] 1) According to the changes of the total internal resistance R and the total capacity C detected after N operations, the life change function f1 (R, C, N, Irms) is constructed.

[0159] 2) According to the V1-Vn of each supercapacitor detected after N times of operation, the charging voltage change fv (V, N, Irms) of the supercapacitor is obtained, and the normal distribution function of the supercapacitor pressure under different working times is fitted:

[0160] M(E,δ) = M(f2(Cs,N,Irms),f3(Cs,N,Irms))

[0161] 3) Obtain basic data of the system through 1) and 2).

[0162] By repeating 1)-3), the life transformation function f1 and the capacitance distribution function M of the supercapacitor under different current impacts are obtained, and an empirical data set is constructed.

[0163] S4: collecting operation data of the supercapacitor module, wherein the operation data includes total internal resistance value, total capacity value, and current value;

[0164] In the actual operation process, a complete monitoring system needs to be established to collect the operation data of the supercapacitor module in real time. The following details the construction of the monitoring system and the specific implementation plan of data collection.

[0165] S4 includes:

[0166] S4.1: Set up a voltage detection circuit and a current detection circuit to collect voltage and current data of the supercapacitor module.

[0167] The voltage detection circuit is built with a high-precision differential amplifier to eliminate common-mode interference.

[0168] Specifically, a differential amplifier AD8221 is connected in parallel across each supercapacitor cell, and its common-mode rejection ratio is greater than 100dB and the bandwidth is 1MHz.

[0169] In order to adapt to different voltage levels, a programmable gain circuit is designed, and automatic range adjustment is achieved through the digital potentiometer AD8400. At the same time, an RC low-pass filter is added to the signal conditioning circuit, and the cut-off frequency is set to 1kHz to suppress high-frequency interference. All voltage signals are digitally converted by a 16-bit ADC (AD7606), and the sampling rate can reach 200kSPS.

[0170] S4.2: Calculate the total internal resistance of the supercapacitor module based on the collected voltage and current data.

[0171] The current detection circuit consists of a Hall current sensor LEM LA100-P and a signal conditioning circuit.

[0172] Among them, the Hall sensor has a range of ±100A, an accuracy of ±0.5%FS, and a bandwidth of 200kHz, which can accurately capture the transient characteristics during large current impact.

[0173] The signal conditioning circuit includes a second-order Butterworth filter cascaded with an operational amplifier to further improve the signal-to-noise ratio. The current signal is also converted into a digital quantity through a 16-bit ADC, and the sampling rate is synchronized with the voltage signal.

[0174] S4.3 uses a constant current charge and discharge method to calculate the total capacity value of the supercapacitor module based on the ratio of voltage change to current.

[0175] The total internal resistance is calculated using the online impedance analysis method.

[0176] The system records the voltage and current values ​​before and after the current mutation, and calculates the transient internal resistance through ΔV / ΔI. To improve the calculation accuracy, the sliding window method is used to process the data: each calculation window contains 100 sampling points, and the window overlap rate is 50%. The calculation results of each window are weighted averaged, and the weight coefficient is proportional to the signal-to-noise ratio of the signal.

[0177] At the same time, an adaptive threshold algorithm is implemented, which can automatically adjust the judgment thresholds of voltage and current changes according to the operating conditions.

[0178] The measurement of total capacity is based on the principle of constant current charging and discharging. The system controls the charging current to be constant at 0.5 times the rated value and records the voltage rise curve during the charging process. The actual capacity value is obtained by integral calculation:

[0179] C = I× dt / ΔV

[0180] Where I is the charging current, dt is the time unit, and ΔV is the voltage change. In order to reduce the measurement error, the nonlinear segment of the voltage curve is excluded from the calculation, and only the data in the linear interval (usually 0.4V~0.8V) is used. In addition, the system also implements a temperature compensation algorithm, which can automatically correct the capacity calculation results according to the ambient temperature.

[0181] S4.4: The total internal resistance value, the total capacity value and the current value are sampled and stored at a preset frequency, and sent to a data processing unit via a remote transmission channel.

[0182] Data storage and transmission adopt a layered architecture design.

[0183] At the field level, a 32-bit ARM microcontroller (STM32F407) is used as the main controller, equipped with 32MB Flash for local data cache. The data sampling frequency can be dynamically adjusted according to demand:

[0184] Normal working mode: voltage and current data are sampled at 10Hz frequency;

[0185] Overload monitoring mode: sampling frequency increased to 1kHz;

[0186] Fault recording mode: sampling frequency reaches 10kHz, and pre-trigger data is saved at the same time.

[0187] The remote transmission channel adopts a dual-channel redundant design: the main channel uses industrial Ethernet, and the backup channel uses 4G wireless communication.

[0188] Data transmission uses the MQTT protocol and TLS encryption to ensure transmission security. The transmitted data includes:

[0189] Real-time operation data: including voltage, current, internal resistance, capacity and other parameters;

[0190] Status flag: indicates the equipment operation status and alarm information;

[0191] Diagnostic information: Contains data quality indicators and system self-test results.

[0192] On the data processing unit side, a data integrity verification mechanism is implemented.

[0193] CRC verification is used to ensure that data transmission is correct, and data packets that fail verification are automatically requested to be retransmitted. At the same time, a data backup mechanism is established to archive historical data regularly and generate data quality reports to provide a reliable data basis for subsequent life prediction.

[0194] S5: Based on the life variation function and the capacitance distribution function in the data set and in combination with the operation data, the remaining life times of the supercapacitor module are calculated.

[0195] like Figure 4 As shown, S5 includes:

[0196] S5.1: Construct a Graph Spring neural ODE model to represent the supercapacitor module as a dynamic graph structure.

[0197] Among them, each supercapacitor cell is used as a node in the graph, and the series-parallel relationship between the nodes is used as an edge.

[0198] The Graph Spring neural ODE model includes a graph convolution layer, a spring physics layer, a temporal ordinary differential equation layer and a prediction layer.

[0199] Among them, building the Graph Spring neural ODE model includes:

[0200] S5.1.1: In the graph convolution layer, the initial feature vector of the node is input, spatial convolution is performed using ChebNet, local features of the node are extracted, and a 64-dimensional node embedding vector is output.

[0201] The implementation of the graph convolution layer adopts the improved ChebNet structure. The specific steps are as follows:

[0202] For each supercapacitor single node vi, an initial feature vector xi is constructed, which includes key parameters such as voltage Vi, internal resistance Ri, capacity Ci, temperature Ti, etc. After these features are standardized, an initial feature matrix X∈R^(N×F) is formed, where N is the number of nodes and F is the feature dimension.

[0203] Then, construct the Laplacian matrix of the graph L = D - A, where D is the degree matrix and A is the adjacency matrix. Normalize L to:

[0204] = I - D^(-1 / 2)AD^(-1 / 2)

[0205] Next, use the K-order Chebyshev polynomial expansion to perform spatial convolution:

[0206] H = ∑Tk( )XWk

[0207] Where Tk is the k-order Chebyshev polynomial, Wk is the learnable weight matrix, k=0-K. In order to enhance the expressiveness of the model, a multi-head attention mechanism is introduced:

[0208] αij = softmax(LeakyReLU( [WHi||WHj]))

[0209] Where a is the attention vector. The final output is a 64-dimensional node embedding vector

[0210] hi = σ(∑j∈Ni αij Whj)

[0211] σ is the ReLU activation function.

[0212] S5.1.2: Input the node embedding vector into the spring physical layer, map it to the physical space, define the spring force between nodes based on the attributes of the edges, and simulate the interaction between capacitors.

[0213] The spring physics layer maps the graph structure to the physical space and simulates the interaction between supercapacitor cells.

[0214] First, define the state vector s = [p, v] in the physical space, where p represents the position and v represents the velocity. For the connected node pair (i, j), define the spring potential energy based on the attributes of the edge:

[0215] Eij = (1 / 2)kij(||pi - pj|| - )²

[0216] where kij is the spring stiffness, determined by the edge weight, is the natural length. Calculate the spring force between the nodes:

[0217] Fij = -kij(||pi - pj|| - )(pi - pj) / ||pi - pj||

[0218] Considering the system damping, the velocity-related damping force is introduced: Di = -γvi, where γ is the damping coefficient. The total force on the node is:

[0219] Fi = ∑j∈Ni(Fij) + Di

[0220] In order to ensure numerical stability, an upper limit constraint on the force is introduced:

[0221] ||Fi|| ≤ Fmax

[0222] S5.1.3: Input the state vector of the physical space into the time-series ordinary differential equation layer, construct a neural ordinary differential equation to describe the dynamic evolution of the system, and use an adaptive step-size solver to perform numerical solution.

[0223] The time-series ordinary differential equation layer uses the neural ODE framework to describe the dynamic evolution of the system.

[0224] First, construct the state transfer function f(s, t, θ) and parameterize it using a multi-layer perceptron:

[0225] f = MLP([s||t],θ)

[0226] Define the ordinary differential equation:

[0227] ds / dt = f(s,t,θ)

[0228] The Dormand-Prince (DOPRI5) solver is selected for numerical solution, which has an adaptive step size feature:

[0229] Δt = min(ηtol / ε, Δtmax)

[0230] Where ε is the local truncation error and ηtol is the allowable error. In order to improve the computational efficiency, a parallel solution strategy is implemented: batch data is sent to the solver in parallel. At the same time, the gradient is calculated:

[0231] L / θ = - ( a / θ f(s,t,θ)dt

[0232] where a is the accompanying variable.

[0233] S5.1.4: Input the solution result into the prediction layer, and output the predicted life span and health status of each node through the multi-layer perceptron decoder.

[0234] The prediction layer uses a decoder structure to map the state of the physical space to the prediction space.

[0235] The decoder consists of a multi-layer perceptron:

[0236] First hidden layer: h1 = ReLU(W1s + b1), dimension is 128;

[0237] Second hidden layer: h2 = ReLU(W2h1 + b2), dimension is 64;

[0238] The output layer is divided into two branches:

[0239] Life prediction branch: L = σ(WLh2 + bL)

[0240] Healthy state branch: H = softmax(WHh2 + bH)

[0241] Where σ is the Sigmoid activation function, which ensures that the predicted lifespan value is positive. During training, dropout is used to prevent overfitting: h' = Dropout(h, p=0.3). The loss function consists of two parts:

[0242] Life prediction loss: Ll = MSE(L, L_true)

[0243] Health state classification loss: Lh = CrossEntropy(H, H_true)

[0244] Total loss: L = αLl + βLh

[0245] where α and β are balance coefficients.

[0246] In order to improve the robustness of the prediction, uncertainty estimation is also implemented in the prediction layer:

[0247] Use Monte Carlo dropout to generate multiple prediction samples;

[0248] Calculate the mean and variance of the forecast;

[0249] Output confidence interval: [μ-2σ, μ+2σ];

[0250] S5.2: inputting the data set into the Graph Spring neural ODE model for training to obtain a trained model;

[0251] The construction of the Graph Spring neural ODE model is a complex process. The following is a detailed description of its architecture design and training method.

[0252] First, the supercapacitor module is represented as a dynamic graph structure G = (V, E).

[0253] Among them, the vertex set V contains all supercapacitor cells, each vertex vi has a characteristic vector xi, which contains the voltage, internal resistance, capacity and other attributes of the cell; the edge set E represents the connection relationship between the cells, and each edge eij has a weight wij, which is used to characterize the strength of the series-parallel relationship.

[0254] For the series connection, the edge weight is set to 1; for the parallel connection, the edge weight is set to 0.5. This graph structure design can naturally reflect the topological characteristics of the supercapacitor module.

[0255] The graph convolution layer uses an improved ChebNet structure to extract local features of nodes. The specific implementation is:

[0256] Expand the input feature matrix X∈R^(N×F) through Chebyshev polynomials: H = ∑Tk( )XWk

[0257] in is the normalized Laplace matrix, Tk is the k-th order Chebyshev polynomial, and Wk is the learnable weight matrix.

[0258] Design residual connection structure: Y = H + σ(XWskip)

[0259] Where σ is the ReLU activation function and Wskip is the weight matrix of the skip connection.

[0260] Use attention mechanism to enhance feature extraction:

[0261] α = softmax( [WHi||WHj])

[0262] Where a is the attention vector and || represents the feature concatenation operation.

[0263] The spring physics layer maps the graph structure to the physical space to simulate the interaction between supercapacitor cells. Specifically, it includes:

[0264] Define the spring potential energy function:

[0265] E = ∑kij(||xi - xj|| - )²

[0266] where kij is the spring constant, is the equilibrium length.

[0267] Calculate the forces between nodes:

[0268] F = - E

[0269] The direction of the force is along the connecting edges between nodes.

[0270] Introduce the damping term:

[0271] D = -γv

[0272] where γ is the damping coefficient and v is the nodal velocity.

[0273] The time-series ordinary differential equation layer uses the neural ODE framework to describe the dynamic evolution of the system:

[0274] Define the state vector s(t), which contains position and velocity information:

[0275] ds / dt = f(s(t), t, θ)

[0276] where f is a vector field parameterized by a neural network and θ is a learnable parameter.

[0277] Solve the ODE using an adaptive step-size solver such as the Dormand-Prince method:

[0278] s( ) = s( ) + f(s(t), t, θ)dt

[0279] Implementing adaptive error control:

[0280] Set the relative error threshold ε, and automatically reduce the step size when the local truncation error exceeds ε.

[0281] The prediction layer uses a multi-layer perceptron structure to map the state of the physical space to the prediction space:

[0282] Designing the encoder-decoder architecture:

[0283] encoder: h = MLP_e(s)

[0284] decoder: y = MLP_d(h)

[0285] Introduce batch normalization and Dropout in the hidden layer: h' = Dropout(BatchNorm(h))

[0286] The output layer uses the Sigmoid activation function to ensure that the predicted value is within a reasonable range.

[0287] Model training adopts a phased strategy:

[0288] Pre-training phase:

[0289] Use historical data sets for supervised learning; the loss function includes prediction error terms and physical consistency terms; use the Adam optimizer with an initial learning rate of 0.001; and use cosine annealing to adjust the learning rate.

[0290] Fine-tuning phase:

[0291] Introduce contrastive learning loss to enhance the feature representation ability of the model; use difficult sample mining strategy to focus on optimizing samples with poor prediction effects; use gradient clipping to prevent gradient explosion.

[0292] Online update phase:

[0293] Use a sliding window strategy to continuously receive new data; use an incremental learning method to update model parameters; maintain an experience replay buffer to prevent catastrophic forgetting; hyperparameter selection during training: batch size: 64; number of training rounds: 200; early stopping strategy: stop if the validation set loss does not improve for 10 consecutive rounds; Dropout rate: 0.3; gradient clipping threshold: 1.0.

[0294] S5.3: Use a channel coordination strategy to process the operating data, establish a timing association between different sensor channels, and achieve multi-dimensional analysis of electrical characteristics, thermodynamic characteristics, and life characteristics.

[0295] Wherein, the operation data is processed by adopting a channel coordination strategy, including:

[0296] S5.3.1: Represent the data of each sensor channel as a sequence of time-stamped observations and construct a multi-channel time series representation.

[0297] First, a multi-channel time series representation is constructed.

[0298] The data of each sensor channel is expressed as: Xi(t) = {(ti,j, vi,j) | j = 1,2,...,Ni}, where i represents the channel number, ti,j is the timestamp, and vi,j is the observed value. Specifically, it includes the following channels:

[0299] Voltage channel: V(t) = {(tv,j, vv,j)}

[0300] Current channel: I(t) = {(ti,j, vi,j)}

[0301] Temperature channel: T(t) = {(tt,j, vt,j)}

[0302] Internal resistance channel: R(t) = {(tr,j, vr,j)}

[0303] Capacity channel: C(t) = {(tc,j, vc,j)}

[0304] To process data with different sampling frequencies, an adaptive resampling strategy is implemented: downsampling is used for fast-changing signals and interpolation is used for slow-changing signals to make the sampling interval of all channels uniform.

[0305] S5.3.2: Based on the multi-channel time series, a dynamic time warping algorithm is used to perform time alignment and output a time-varying correlation matrix between channels.

[0306] The implementation of the dynamic time warping (DTW) algorithm adopts the improved FastDTW method. The specific steps are as follows:

[0307] Construct the distance matrix: D(i,j) = ||Xi - Xj| + λ|ti - tj|, where λ is the time penalty factor.

[0308] Define path constraints: |i - j| ≤ δ, δ is the window size;

[0309] Dynamic programming solution: γ(i,j) = d(i,j) + min{γ(i-1,j), γ(i,j-1), γ(i-1,j-1)}

[0310] Extract alignment path: P = {(ik,jk) | k = 1,...,K}

[0311] Based on the alignment results, the time-varying correlation matrix is ​​calculated: R(t) = [ρij(t)]n×n, where ρij(t) is the correlation coefficient between channels i and j at time t.

[0312] S5.3.3: Apply wavelet transform to each channel data after time alignment to obtain multi-scale feature representation.

[0313] Wavelet transform uses the continuous wavelet transform (CWT) method to extract multi-scale features:

[0314] Select the mother wavelet function:

[0315] Using Morlet wavelet: ψ(t) = π^(-1 / 4)e^(-t² / 2)e^(i t)

[0316] Calculate the wavelet coefficients: W(a,b) = (1 / √a) x(t)ψ ((tb) / a)dt, where a is the scale parameter and b is the translation parameter.

[0317] Construct the time spectrum: S(f,t) = |W(f,t)|²

[0318] Extract feature vector: F = [ , , ..., En], where Ei is the energy characteristic of different frequency bands.

[0319] S5.3.4: Based on the multi-scale features, the attention mechanism is used to calculate the weight coefficients between channels and establish long-term dependencies.

[0320] The implementation of the attention mechanism is based on a multi-head self-attention structure:

[0321] Define the query, key, and value matrix:

[0322] Q = XWQ, K = XWK, V = XWV

[0323] Calculate the attention score:

[0324] A = softmax(Q / √dk)V

[0325] Multi-head fusion:

[0326] MultiHead(X) = Concat(head , ..., headh)WO

[0327] Temporal Attention: Using Positional Encoding to Enhance Temporal Information:

[0328] PE(pos,2i) = sin(pos / 10000^(2i / d))

[0329] PE(pos,2i+1) = cos(pos / 10000^(2i / d))

[0330] S5.3.5: Based on the weight coefficients and long-term dependencies, identify the key channels and impact paths that have a significant impact on life prediction.

[0331] Long-term dependencies are established using a layered architecture:

[0332] Local dependency layer: uses sliding windows to capture short-term patterns; window size W = 100, step size S = 10;

[0333] Global dependency layer: build a global attention graph and extract key node paths;

[0334] Timing dependency layer: Establish state transition matrix and calculate transition probability.

[0335] The identification of critical channels and impact paths is based on the following steps:

[0336] Channel importance assessment:

[0337] Calculate the mean attention weight: i = (1 / T)∑w_i(t)

[0338] Evaluate channel contribution: ci = || i|

[0339] Impact Path Analysis:

[0340] Construct causal graph G = (V,E)

[0341] The edge weight eij represents the influence strength. The minimum spanning tree algorithm is used to extract the main path.

[0342] Sensitivity analysis:

[0343] Calculate the input perturbation response, evaluate the coupling strength between channels, and quantify the sensitivity of the prediction to each channel.

[0344] The final output results include:

[0345] Key channel ranking table: contains the importance score and ranking of each channel.

[0346] Impact Path Diagram: Visualize the relationship between channels.

[0347] Time-varying characteristics report: describes the dynamic changes in channel importance.

[0348] S5.4: Input the processed operating data into the trained Graph Spring neural ODE model, combine the clustering algorithm and the minimum variance algorithm, and output the remaining life of the supercapacitor module.

[0349] Specifically, the processed running data is input into the trained Graph Spring neural ODE model, including:

[0350] S5.4.1: Construct a multi-subspace matrix including electrical properties, thermodynamic properties, and lifetime properties, corresponding to different physical properties.

[0351] First, construct a multi-subspace matrix structure. Each subspace corresponds to a type of physical properties, including:

[0352] Electrical characteristics subspace matrix Me∈R^(n× ):

[0353] Voltage characteristics: nominal voltage, voltage fluctuation, voltage drop rate;

[0354] Current characteristics: charge and discharge current, ripple factor;

[0355] Internal resistance characteristics: equivalent series resistance, impedance spectrum;

[0356] Capacity characteristics: rated capacity, actual capacity, capacity attenuation rate;

[0357] Thermodynamic property subspace matrix Mt∈R^(n× ):

[0358] Temperature characteristics: surface temperature, internal temperature, temperature gradient;

[0359] Heat flow characteristics: heat flux, thermal diffusivity;

[0360] Heat dissipation characteristics: thermal resistance, cooling efficiency;

[0361] Environmental characteristics: ambient temperature and humidity;

[0362] The life characteristic subspace matrix Ml∈R^(n× ):

[0363] Cycle characteristics: number of cycles, depth;

[0364] Degradation characteristics: capacity retention rate and efficiency decrease;

[0365] Health characteristics: status assessment indicators, failure probability;

[0366] Usage characteristics: working hours, load conditions;

[0367] S5.4.2: Design an optimization objective function based on the nuclear norm and constrain the rank of each subspace matrix to promote a low-rank structure.

[0368] Design of optimization objective function based on nuclear norm:

[0369] Main optimization goal: min ∑αi||Mi|| + β||W ⊙ (M - X)||²F, where ||·|| represents the nuclear norm, ⊙ represents the Hadamard product;

[0370] Constraints:

[0371] Non-negative constraint: Mi ≥ 0

[0372] Rank constraint: rank(Mi) ≤ ri

[0373] Consistency constraint: ||MiM ||F ≤ εij

[0374] Regularization term: Ω(M) = || M| + ||M||²F is used to maintain spatial smoothness and prevent overfitting.

[0375] S5.4.3: Use the alternating direction multiplication method to iteratively update each subspace matrix to achieve matrix recovery.

[0376] The iterative update process of the alternating direction method of multipliers (ADMM):

[0377] Variable decomposition: Introducing auxiliary variables Z and dual variables Y

[0378] L(M,Z,Y) = f(M) + g(Z) +<Y,M-Z> + (ρ / 2)||MZ||²F

[0379] Update steps:

[0380] M step: Mk+1 = argmin L(M,Zk,Yk)

[0381] Z step: Zk+1 = argmin L(Mk+1,Z,Yk)

[0382] Y step: Yk+1 = Yk + ρ(Mk+1 - Zk+1)

[0383] Convergence criterion:

[0384] Original residual: ||Mk+1 - Zk+1||F ≤ ε

[0385] Dual residual: ρ||Zk+1 - Zk||F ≤ ε

[0386] S5.4.4 convert the features in the recovered subspace matrix into node attributes of the graph structure;

[0387] Convert the recovered matrix features into graph structure attributes:

[0388] Node attribute mapping:

[0389] Electrical properties: vi.e = Me[i,:]

[0390] Thermodynamic properties: vi.t = Mt[i,:]

[0391] Life attribute: vi.l = Ml[i,:]

[0392] Edge attribute construction:

[0393] Physical connection: eij.p = f(vi,vj)

[0394] Feature correlation: eij.c = sim(vi,vj)

[0395] Graph embedding generation: hi = GNN([vi.e; vi.t; vi.l])

[0396] S5.4.5: Calculate the remaining life based on the comprehensive information of each subspace and output the confidence interval of the prediction result; life prediction based on comprehensive information.

[0397] Feature Fusion:

[0398] Linear fusion: F = Me + Mt + M

[0399] Nonlinear fusion: F = MLP([Me||Mt||Ml])

[0400] Prediction Model:

[0401] Point forecast: L = f(F)

[0402] Distribution prediction: p(L|F) = N(μ(F), σ²(F))

[0403] Confidence interval estimates:

[0404] Interval calculation: [L - kσ, L + kσ]

[0405] Credibility score: s = exp(-||F - ||² / 2τ²)

[0406] The final output results include:

[0407] Point prediction values: remaining number of cycles, expected usage time.

[0408] Confidence interval: 95% confidence interval, prediction error range.

[0409] Reliability indicators: prediction credibility, uncertainty assessment.

[0410] In summary, in S3-4.4, the multi-dimensional analysis of operating data and life prediction is a complex system engineering. The specific solutions for data processing and prediction are described in detail below.

[0411] The channel coordination strategy must first solve the synchronization problem of multi-source heterogeneous data. The system uses a timestamp alignment method to unify the data of different sensor channels to the same time base:

[0412] Electrical characteristic data: including voltage, current, internal resistance, etc., with a sampling frequency of 1kHz;

[0413] Thermodynamic property data: including temperature, heat flow, etc., with a sampling frequency of 10Hz;

[0414] Life characteristic data: including number of cycles, capacity, etc., with a sampling frequency of 1Hz.

[0415] The alignment of data of different frequencies is achieved through interpolation algorithms. Cubic spline interpolation is used for rapidly changing electrical data, and linear interpolation is used for slowly changing temperature data.

[0416] Establish a multi-dimensional feature analysis framework:

[0417] Electrical characteristics analysis:

[0418] Calculate the instantaneous power P(t) = V(t)×I(t)

[0419] Extract voltage fluctuation rate δV = σV / μV

[0420] Evaluate the internal resistance change trend ΔR / Δt

[0421] Thermodynamic properties analysis:

[0422] Construct temperature distribution model T(x,y,z,t)

[0423] Calculate thermal stress σth = αE(T- )

[0424] Estimated heat loss Ploss = I²R + CT(dT / dt)

[0425] Life characteristics analysis:

[0426] Tracking capacity decay rate dC / dN

[0427] Evaluation performance degradation acceleration factor k = f(T,I,DOD)

[0428] Establish stress-life relationship SN curve;

[0429] In order to realize the correlation analysis between characteristics, a multivariate time series analysis method is used:

[0430] Compute the cross-correlation function between features:

[0431] Rxy(τ) = E[X(t)Y(t+τ)]

[0432] Extract the main influencing factors: Apply principal component analysis to reduce the dimension and retain the features that explain 90% of the variance;

[0433] Construct Granger causal network: analyze the strength of causal relationship between different characteristics;

[0434] The processed data is input into the trained Graph Spring neural ODE model. The specific prediction process is as follows:

[0435] Data preprocessing stage:

[0436] Normalization: Z = (X-μ) / σ;

[0437] De-noising: wavelet transform combined with soft threshold method;

[0438] Feature engineering: construct time window features;

[0439] Cluster analysis stage:

[0440] The improved DBSCAN algorithm is used for working condition clustering;

[0441] The density parameter ε is determined adaptively: ε = f(local density);

[0442] The minimum number of points MinPts is set based on statistical characteristics;

[0443] Minimum variance algorithm implementation:

[0444] Construct the objective function: J = ∑(yi - i)² + λ∑(Δ i)²

[0445] Introduce smoothness constraints: limit the sudden changes in prediction results;

[0446] Use quadratic programming to find the optimal solution;

[0447] Specific steps for life expectancy prediction:

[0448] Model online prediction:

[0449] Input the current state vector s(t); calculate the state evolution through the ODE solver; output the multi-step prediction results;

[0450] Ensemble prediction results:

[0451] Select similar historical cases based on clustering results;

[0452] Calculate the weighted prediction value: L = ∑wiLi;

[0453] The weight wi is proportional to the similarity;

[0454] Confidence interval estimates:

[0455] Use the Bootstrap method to generate multiple prediction samples; calculate the 95% confidence interval;

[0456] Assess forecast uncertainty;

[0457] The remaining lifespan outputted in the end contains three key indicators:

[0458] Point prediction value: the most likely number of remaining cycles;

[0459] Interval forecast: a range estimate that takes into account uncertainty;

[0460] Credibility score: confidence level based on historical prediction accuracy;

[0461] The system also implements a dynamic update mechanism for prediction results:

[0462] Sliding window update:

[0463] Window length W = 1000 cycles; step size S = 100 cycles;

[0464] Prediction Error Feedback:

[0465] Calculate the forecast deviation e(t) = y(t) - (t)

[0466] Update model parameters θ' = θ - η θe(t)

[0467] Adaptive Adjustment:

[0468] Dynamically adjust the prediction frequency based on the prediction accuracy; abnormal working conditions trigger instant prediction updates;

[0469] like Figure 5 As shown, after outputting the remaining life times of the supercapacitor module, it also includes:

[0470] A1, based on the analysis results of the channel coordination strategy, a channel coordination relationship diagram is generated to show the evolution process of the timing characteristics.

[0471] Assume that a 48V / 165F supercapacitor module is monitored, which contains 16 cells (each cell is 2.7V), and the following key timing characteristics are collected:

[0472] Voltage coordination relationship:

[0473] t=0h: the cell voltage is distributed between 2.65-2.70V, with a range of 0.05V;

[0474] t=1000h: the voltage distribution expands to 2.60-2.72V, with a range of 0.12V;

[0475] t=2000h: Obvious stratification occurs, and the voltage difference between the two groups of cells reaches 0.2V;

[0476] Temperature coordination relationship:

[0477] t=0h: temperature is uniformly distributed at 25±2℃;

[0478] t=1000h: the temperature of the central monomer rises to 30°C, and the edge remains at 26°C;

[0479] t=2000h: The hot spot reaches 35°C, and the temperature gradient is significant;

[0480] Internal resistance coordination relationship:

[0481] t=0h: internal resistance is distributed between 0.8-1.0mΩ;

[0482] t=1000h: the local monomer internal resistance increases to 1.5mΩ;

[0483] t=2000h: high temperature monomer internal resistance reaches 2.0mΩ;

[0484] A2, based on the feature changes during the evolution process, a fault feature library including electrical characteristics, thermodynamic characteristics and life characteristics is constructed.

[0485] For example:

[0486] Electrical characteristics failure:

[0487] Monomer imbalance: voltage deviation> 0.1V for 30 minutes;

[0488] Abnormal internal resistance: internal resistance increases by >50% and continues to rise;

[0489] Capacity attenuation: The single cell capacity is less than 85% of the rated value;

[0490] Thermodynamic properties failure:

[0491] Local overheating: monomer temperature>35℃ or temperature difference>10℃;

[0492] Abnormal heat dissipation: cooling efficiency decreases by >30%;

[0493] Thermal runaway: temperature rise rate>2℃ / min;

[0494] Life characteristic failure:

[0495] Premature aging: Capacity decay rate>1% / 1000 cycles;

[0496] Performance degradation: charge and discharge efficiency <90%;

[0497] Cycle abnormality: charge and discharge depth fluctuation> 20%;

[0498] A3, based on the fault feature library, a multi-dimensional feature clustering method is used to implement fault mode classification.

[0499] Cluster analysis of a batch of 100 supercapacitor modules yielded the following failure modes:

[0500] Normal operating mode (60%):

[0501] Voltage fluctuation <0.05V;

[0502] Temperature <30℃;

[0503] Internal resistance increased by <20%;

[0504] Mild degradation mode (25%):

[0505] Voltage fluctuation 0.05-0.1V;

[0506] Temperature 30-35℃;

[0507] Internal resistance increases by 20-50%;

[0508] Accelerated decay mode (10%):

[0509] Voltage fluctuation 0.1-0.2V;

[0510] Temperature 35-40℃;

[0511] Internal resistance increases by 50-100%;

[0512] Fault warning mode (5%):

[0513] Voltage fluctuation>0.2V;

[0514] Temperature>40℃;

[0515] Internal resistance increases by >100%;

[0516] A4. Design and implement a threshold-based hierarchical early warning strategy based on the severity of the failure mode.

[0517] Example of implementing a layered warning strategy:

[0518] Level 1 warning (prompt attention):

[0519] Trigger condition: Entering mild degradation mode;

[0520] Warning content: It is recommended to strengthen the monitoring frequency;

[0521] Handling suggestions: Record exceptions and track changes;

[0522] Level 2 warning (intervention required):

[0523] Trigger condition: Entering accelerated decay mode;

[0524] Warning content: Warning of significant performance degradation;

[0525] Treatment suggestions: Check the working conditions and adjust the load;

[0526] Level 3 warning (immediate processing):

[0527] Trigger condition: Enter fault warning mode;

[0528] Warning content: Warning of potential safety hazards;

[0529] Treatment suggestions: Stop using and carry out maintenance;

[0530] A5, combining the remaining life prediction results and failure mode classification, evaluates the optimal maintenance timing and generates maintenance decision recommendations.

[0531] Assume that a supercapacitor module has the following conditions after running for 2000 hours:

[0532] Performance status:

[0533] Remaining life prediction: 3000±500 hours;

[0534] Failure mode: accelerated decay mode;

[0535] Health index: 75%;

[0536] Maintenance recommendations:

[0537] Timing selection: It is recommended to complete maintenance within 1000 hours;

[0538] Maintenance content:

[0539] Balance charge and discharge; clean the cooling system; replace high-resistance monomers;

[0540] Priority: Medium urgency;

[0541] Cost-Benefit Analysis:

[0542] Immediate maintenance: downtime loss of 100 man-hours and maintenance cost of 50,000 yuan;

[0543] Deferred maintenance: performance continues to degrade, which may lead to cascading failures;

[0544] Recommended solution: Perform maintenance during the next planned downtime;

[0545] Example of a preventive maintenance plan:

[0546] Short-term plan (within 1 month):

[0547] Perform equalizing charging every week;

[0548] Check temperature distribution daily;

[0549] Record performance trends;

[0550] Medium-term plan (within 3 months):

[0551] Replace aging monomers;

[0552] Optimize cooling system;

[0553] Adjust working parameters;

[0554] Long-term plan (within 6 months):

[0555] Evaluate overall replacement timing;

[0556] Develop upgrade plans;

[0557] Optimize system configuration;

[0558] In addition, after building the Graph Spring neural ODE model, it also includes:

[0559] B1, design a multi-objective loss function based on single prediction loss, system prediction loss and physical consistency loss,

[0560] The design of the multi-objective loss function is based on three key components:

[0561] First, the predicted cell loss reflects the accuracy of the life prediction of each supercapacitor cell:

[0562] Lunit = (1 / N)∑(yi - i)²,

[0563] Where yi is the true lifespan value, i is the predicted value and N is the number of monomers.

[0564] In order to deal with the prediction errors at different life stages, a weighting mechanism is introduced:

[0565] Lw_unit = ∑wiLi

[0566] The weight wi is inversely proportional to the remaining life, making the model pay more attention to the prediction accuracy at the end of life. Secondly, the system prediction loss considers the comprehensive performance of the entire module:

[0567] Lsys = ||Y - | + λ|| Y - |

[0568] The first term is the L2 norm of the prediction error, and the second term is the L1 norm of the gradient difference, which is used to maintain the smoothness of the prediction curve. Finally, the physical consistency loss ensures that the prediction results conform to the laws of physics:

[0569] Lphy = Lkirchhoff + Lenergy + Lthermo

[0570] These include Kirchhoff's law constraints, energy conservation constraints, and thermodynamic constraints.

[0571] The comprehensive loss function is:

[0572] Ltotal = Lunit + Lsys + Lphy

[0573] in , , is an adjustable weight coefficient.

[0574] B2, dividing the data set into a training set and a validation set, and using the training set for supervised learning in the pre-training stage.

[0575] The data set is divided using a stratified sampling strategy: all samples are divided into multiple layers according to life span levels, and samples are randomly selected from each layer. The training set accounts for 80% of the total data, and the validation set accounts for 20%. In order to ensure the temporal continuity of the data, the sampling unit is set to a complete cycle.

[0576] At the same time, data enhancement technology is implemented:

[0577] Time series perturbation: adding Gaussian noise;

[0578] Feature combination: construct composite features;

[0579] Time window sliding: generate overlapping sequence samples;

[0580] Working condition change: simulate different operating conditions;

[0581] The pre-training stage adopts a hierarchical training strategy:

[0582] Stage 1 (Feature Extraction):

[0583] Only the graph convolutional layer and the prediction layer are trained; the Adam optimizer is used with an initial learning rate of 0.001; the batch size is set to 64; the training is performed for 50 epochs and early stopping is performed when the verification loss no longer decreases.

[0584] Phase 2 (Physical Modeling):

[0585] Unfreeze the spring physics layer and ODE layer, reduce the learning rate to 0.0001, increase the batch size to 128, train for 100 epochs, and use cosine annealing to adjust the learning rate.

[0586] B3, taking the pre-trained model parameters as the initial values, and adding physical consistency loss for optimization in the physical constraint training phase.

[0587] The iterative optimization strategy is introduced in the physical constraint training phase:

[0588] Initialization phase:

[0589] Load pre-trained model parameters and set physical constraint weights The initial value is 0.1.

[0590] Progressive training:

[0591] Increase the physical constraint weight every 10 rounds: = min( ×1.2, 1.0), dynamically adjust the weights of other loss items, and monitor the satisfaction of physical constraints.

[0592] Fine-tuning optimization:

[0593] Use L-BFGS optimizer for fine-tuning, adopt gradient clipping to prevent gradient explosion, and verify the degree of compliance with physical laws.

[0594] B4, uses a sliding time window method to obtain the latest operating data and continuously updates the training set and model parameters during the online learning phase.

[0595] Incremental update mechanism is implemented in the online learning phase:

[0596] Time window design:

[0597] The window length W is set to 1000 cycles, the sliding step size S is set to 100 cycles, and the latest N window data are kept.

[0598] Data update strategy:

[0599] The FIFO queue is used to manage training samples, and expired data is cleaned up regularly, retaining key historical samples as experience playback.

[0600] Parameter update method:

[0601] Online updates are performed using stochastic gradient descent.

[0602] Learning rate adaptive adjustment: η = / (1 + kt)

[0603] Set parameter update threshold to prevent over-adjustment.

[0604] In another embodiment, the method for life prediction in the embodiment of the present application further includes:

[0605] Use clustering algorithm and minimum variance algorithm to determine the minimum radius and minimum variance.

[0606] Through hardware monitoring, obtain the current module information, Rn, Cn, In.

[0607] Start calculating the current state, and obtain the estimated number of times Nused is used under different currents based on the Rn and Cn recorded in the database

[0608] According to the Rn, Cn and In recorded in the database, clustering and minimum variance are used to predict the remaining life of the supercapacitor module under the current current.

[0609] Thereby, the service life of a single module can be obtained.

[0610] For the entire system, whether the life of the entire system has reached the end of its life, and the overall internal resistance and capacitance information.

[0611] If there are not so many sensors, then the life prediction of the entire system can form an overall evaluation plan based on the data, and the function will be fitted by the two functions f1 and M.

[0612] For overall life prediction, it cannot be based solely on the current R and C data, but also needs to monitor healthy usage. Specifically: based on the current R and C, consider the number of uses N.

[0613] The health status prediction is based on the R and C status of a single module, and only the rated discharge conditions are considered. Then the current range distribution can be estimated, and the probability distribution of the current energy storage system can be derived through mathematical deduction. The impact of the range is that the voltage distribution of the capacitor is uneven, which can easily cause a large high and low voltage difference, causing a single over-capacitor to over-voltage, resulting in a risk of damage.

[0614] The calculation method is as follows: the current number is calculated based on the probability distribution, and the probability distribution of occurrence is MM(R, C, I). After obtaining the distribution, the expectation E of the fault is calculated.

[0615] like Figure 6 As shown, the embodiment of the present application also provides a device for predicting the life of a supercapacitor under a large current overload condition, comprising:

[0616] A test platform construction unit, used to build a test platform including a supercapacitor module, and a charging power supply electrically connected to the supercapacitor module, a relay, a counter, a charge and discharge switch, a resistor, a shunt, an oscilloscope, and a data acquisition instrument;

[0617] A data acquisition unit, based on the test platform, performs a charge and discharge test on the supercapacitor module, records the test parameters of the supercapacitor module, the test parameters include current effective value, total internal resistance, total capacity, single cell voltage and working times; and collects the operation data of the supercapacitor module, the operation data includes total internal resistance value, total capacity value and current value;

[0618] A data processing unit, configured to construct a data set including a supercapacitor life variation function and a supercapacitor cell capacitance distribution function according to the test parameters, wherein the life variation function is used to characterize the relationship between the total internal resistance and the total capacity of the supercapacitor and the number of operations, and the capacitance distribution function is used to characterize the supercapacitor cell voltage distribution characteristics;

[0619] A prediction unit is used to calculate the remaining life times of the supercapacitor module based on the life variation function and the capacitance distribution function in the data set in combination with the operation data.

Claims

1. A method for predicting the life of a supercapacitor under high current overload conditions, characterized in that: include: Constructing a high current test platform, the test platform includes a supercapacitor module, and a charging power supply, a relay, a counter, a charge and discharge switch, a resistor, a shunt, an oscilloscope and a data acquisition instrument electrically connected to the supercapacitor module, wherein the supercapacitor module is composed of a plurality of supercapacitor monomers connected in series and parallel; Based on the test platform, the supercapacitor module is charged and discharged, and the test parameters of the supercapacitor module are recorded, wherein the test parameters include effective current value, total internal resistance, total capacity, single cell voltage and working times; According to the test parameters, construct a data set including a supercapacitor life variation function and a supercapacitor single capacitance distribution function, wherein the life variation function is used to characterize the relationship between the total internal resistance and the total capacity of the supercapacitor and the number of operations, and the capacitance distribution function is used to characterize the voltage distribution characteristics of the supercapacitor single. Collecting operation data of the supercapacitor module, wherein the operation data includes total internal resistance value, total capacity value, and current value; Based on the life variation function and the capacitance distribution function in the data set and in combination with the operation data, the remaining life times of the supercapacitor module are calculated; Based on the life variation function and the capacitance distribution function in the data set and in combination with the operation data, the remaining life times of the supercapacitor module are calculated, including: Constructing a Graph Spring neural ODE model, representing the supercapacitor module as a dynamic graph structure, wherein each supercapacitor monomer is used as a node in the graph, and the series-parallel relationship between the nodes is used as an edge, and the Graph Spring neural ODE model includes a graph convolution layer, a spring physics layer, a time series ordinary differential equation layer, and a prediction layer; Inputting the data set into the Graph Spring neural ODE model for training to obtain a trained model; A channel coordination strategy is used to process the operating data, establish a timing association between different sensor channels, and realize a multi-dimensional analysis of electrical characteristics, thermodynamic characteristics, and life characteristics; Input the processed operating data into the trained Graph Spring neural ODE model, combine the clustering algorithm and the minimum variance algorithm, and output the remaining life of the supercapacitor module; Wherein, the operation data is processed using a channel coordination strategy, including: Represent the data of each sensor channel as a sequence of observations with timestamps and construct a multi-channel time series representation; Based on the multi-channel time series, a dynamic time warping algorithm is used to perform time alignment and output a time-varying correlation matrix between channels; Apply wavelet transform to each channel data after time alignment to obtain multi-scale feature representation; Based on the multi-scale features, an attention mechanism is used to calculate the weight coefficients between channels and establish long-term dependencies; Based on the weight coefficients and long-term dependencies, key channels and impact paths that have a significant impact on life prediction are identified.

2. The method according to claim 1, characterized in that According to the test parameters, a data set including a supercapacitor life variation function and a supercapacitor monomer capacitance distribution function is constructed, including: The total internal resistance and total capacity data in the test parameters are grouped according to a preset working number interval, and a life variation function is constructed based on the grouped first test parameter; The single cell voltage data in the test parameters are grouped according to a preset working number interval, and a supercapacitor single cell charging voltage variation function is obtained based on the grouped second test parameters; Taking the monomer charging voltage variation function as input, using a normal distribution fitting algorithm, and outputting a supercapacitor monomer pressure bearing normal distribution function under different working times; The corresponding relationship between the life variation function and the normal distribution function under different current impact conditions is integrated into an empirical data set.

3. The method according to claim 1, characterized in that The operating data of the supercapacitor module is collected in real time through the monitoring system, including: Setting up a voltage detection circuit and a current detection circuit to collect voltage and current data of the supercapacitor module; Based on the collected voltage and current data, calculating the total internal resistance of the supercapacitor module; The constant current charging and discharging method is used to calculate the total capacity value of the supercapacitor module according to the ratio of the voltage change to the current; The total internal resistance value, total capacity value and current value are sampled and stored at a preset frequency, and sent to a data processing unit through a remote transmission channel.

4. The method according to claim 1, characterized in that: Build a Graph Spring neural ODE model, including: In the graph convolution layer, the initial feature vector of the node is input, and ChebNet is used to perform spatial convolution to extract the local features of the node, and a 64-dimensional node embedding vector is output; Input the node embedding vector into the spring physical layer, map it to the physical space, define the spring force between nodes based on the attributes of the edge, and simulate the interaction between capacitors; Inputting the state vector of the physical space into the time-series ordinary differential equation layer, constructing a neural ordinary differential equation to describe the dynamic evolution of the system, and using an adaptive step-size solver to perform numerical solution; The solution result is input into the prediction layer, and the predicted life span and health status of each node are output through the multi-layer perceptron decoder.

5. The method according to claim 1, characterized in that After building the Graph Spring neural ODE model, it also includes: Design a multi-objective loss function based on single-objective prediction loss, system prediction loss, and physical consistency loss; Dividing the data set into a training set and a validation set, and using the training set for supervised learning in a pre-training stage; The pre-trained model parameters are used as initial values, and physical consistency loss is added to optimize them during the physical constraint training phase; A sliding time window method is used to obtain the latest operating data, and the training set and model parameters are continuously updated during the online learning phase.

6. The method according to claim 1, characterized in that Feed the processed run data into the trained GraphSpring neural ODE model, including: Construct a multi-subspace matrix including electrical properties, thermodynamic properties and life characteristics, corresponding to different physical properties; The optimization objective function is designed based on the nuclear norm, and the rank of each subspace matrix is ​​constrained to promote a low-rank structure; The alternating direction multiplier method is used to iteratively update each subspace matrix to achieve matrix recovery; Convert the features in the recovered subspace matrix into node attributes of the graph structure; The remaining life is calculated based on the comprehensive information of each subspace, and the confidence interval of the prediction result is output at the same time.

7. The method according to claim 1, characterized in that After outputting the remaining life times of the supercapacitor module, the method further includes: Based on the analysis results of the channel coordination strategy, a channel coordination relationship diagram is generated to show the evolution process of the timing characteristics; According to the characteristic changes in the evolution process, a fault characteristic library including electrical characteristics, thermodynamic characteristics and life characteristics is constructed; Based on the fault feature library, a multi-dimensional feature clustering method is used to implement fault mode classification; Design and implement a threshold-based hierarchical early warning strategy based on the severity of the failure mode; Combine the remaining life prediction results and failure mode classification to evaluate the optimal maintenance timing and generate maintenance decision recommendations.

Images