Battery system health monitoring system

Abstract

A battery system health monitoring system. A monitoring system (120) for a battery system (110) comprising one or more battery cells (110). The monitoring system (120) comprises one or more sensors (125) for measuring a characteristic associated with the battery system (110) during a charging or discharging operation, and a processing device (130) communicatively coupled to the one or more sensors (125). The processing device (130) comprises a pre-processor (135) configured to receive measurement data from the at least one sensor (125), and a neural network (140) configured to receive processed data from the pre-processor (135). Based on the received measurement data, the pre-processor (135) determines a normalized rate of change, with respect to one of the times, of a first measured characteristic or a second measured characteristic associated with the battery system (110) measured during the charging or discharging operation, where the second measured characteristic differs from the first measured characteristic. The neural network (140) is configured to use the normalized determined rate of change from the pre-processor (135) as an input to determine a state of health (SOH) of the battery system (110).

Description

[Technical Field]

[0001] The present disclosure relates to the fields of battery technology, energy cells, and battery monitoring systems and solutions for measuring the state of health (SoH) of a battery system. Specifically, embodiments of the present disclosure relate to monitoring systems, processing devices, methods, and computer-readable media for determining the state of health of a battery system. [Background technology]

[0002] Battery systems are used in a wide range of modern power applications. For example, battery systems are used to power electric vehicles and in commercial applications such as industrial power applications, transportation, energy storage, and powering modern electronic devices. Given the relatively large power demands in such applications, battery systems often include multiple battery cells coupled together to achieve the required power output. Battery cells can be coupled together to form a battery pack, and a battery system can include one or more battery packs.

[0003] It is common to monitor the voltage and current of battery cells in a battery pack to determine the operating state of the battery cells. Typically, the battery system is connected to a battery management system that receives signals indicative of the current operating state of the battery cells and ensures that the battery system operates within its safe operating limits (i.e., the combination of voltage, current, and temperature conditions under which the battery system can be expected to operate without self-damage). For further details, the interested reader is referred to the following Wikipedia website: https: / / en.wikipedia.org / wiki / Battery_management_system

[0004] One aspect of a battery system that is highly desirable to monitor is the battery system's State of Health, or SoH. The SoH of a battery system is the battery system's State of Health at the time of manufacture. charge Aging battery systems compared to capacity charge It can be defined as a measure of the percentage of capacity. Typically, the State of Health (SOH) is 100% at the time of manufacture and decreases with time and use. Monitoring the SoH of a battery system provides visibility into the degradation of the battery system since manufacture and can be used to effectively operate and maintain the battery system. However, unlike temperature, voltage, or current, the SoH cannot be measured directly. Rather, the SoH must be determined based on the behavior of the battery system during charge and discharge operations.

[0005] Conventional methods for determining the SoH of a battery system typically involve performing a full low-rate (low C-rate) charge / discharge cycle on the battery system. charge The capacitance is measured and then charge capacity of the battery system at the time of initial manufacture charge However, such an approach is not always feasible or viable, in part because some battery systems are never fully charged or discharged, which can adversely affect the battery system's lifespan or interfere with normal operation in battery-powered applications.

[0006] Direct measurement of SoH is often impractical and other methods have been proposed.

[0007] One proposed approach is to design a model that exhibits behavior corresponding to that of the battery system. Such models are typically implemented using software, although some very simple models can be calculated using dedicated hardware. The model can be an equivalent circuit model, where: Charge retention capacityis a state variable that is varied to reduce the error in the predicted open circuit voltage, for example, by a Kalman loop.

[0008] However, using model-based methods, the remaining charge of the battery system, often referred to as the battery system's state of charge, or SoC, can be calculated. charge Capacity, full battery charge Calculated from other state variables such as the ratio to capacity and internal resistance Charge retention capacity Furthermore, it is difficult to separate the charge As a physical model, where capacity is determined by the physical properties of the cell and is allowed to vary, accurate modeling requires knowledge of many different cell parameters as well as other contributing factors, such as available lithium that is lost from the cell with use, most of which cannot be measured directly; therefore, model-based methods are of limited utility.

[0009] A different approach is disclosed in CN110824364A, which proposes using a deep neural network to perform SoH prediction. A deep neural network has multiple hidden layers between the input and output layers and is capable of modeling complex nonlinear relationships with fewer data units than a similarly performing shallow neural network that includes fewer hidden layers (typically less than 10) between the input and output layers.

[0010] However, deep neural networks such as those described in CN110824364A require intensive processing power, making such approaches unsuitable for real-world implementation because the processing requirements necessary to implement deep neural networks exceed the processing power of chips in embedded systems that may be used to provide in-situ monitoring of battery systems, such as battery systems in electric vehicles.

[0011] While advances in cloud computing power mean that some of the processing activities for implementing deep neural networks can be performed in the cloud, implementing deep neural networks remotely presents other challenges. In particular, in the case of SoH prediction, remote processing of deep neural network data requires the battery monitoring system to be constantly connected to the remote processing system and must include additional hardware to enable the system to securely transmit data in real time. These requirements mean that practical SoH prediction based on deep neural networks is limited to laboratory environments where these challenges can be circumvented.

[0012] Another approach is presented in DE 10 2015 016 987 A1, which involves detecting the terminal voltage of the battery cell during the charging or discharging process and determining the SoC or charge This involves mathematically deriving the terminal voltage based on measurements and identifying the height of the peak value from the derived data set. However, this approach also has several drawbacks. The disclosed system: charge Alternatively, they rely on providing an accurate model for deriving the SoC value and require memory to store previously derived data sets for later comparison with later data sets. Furthermore, the approach discussed in DE 10 2015 016 987 A1 only provides information about the relative degree of degradation of the battery system, rather than an absolute measure of SoH.

[0013] In view of the above, there is a need for an alternative approach to monitoring the SoH of a battery system.

[0014] It is an aim of at least some embodiments of the present disclosure to address one or more shortcomings in prior art systems. Summary of the Invention

[0015] One aspect of the present disclosure provides a monitoring system for a battery system including one or more battery cells. The monitoring system includes one or more sensors for measuring characteristics associated with the battery system during charging or discharging operations, and a processing device communicatively coupled to the one or more sensors. The processing device includes a pre-processor configured to receive measurement data from the at least one sensor, and a neural network configured to receive processed data from the pre-processor. Based on the received measurement data, the pre-processor: A normalized rate of change of a first measured characteristic relative to one of a second measured characteristic associated with the battery system, measured over time or during a charge or discharge operation. where the second measured characteristic is different from the first measured characteristic. The neural network is configured to use the determined normalized rate of change from the pre-processor as an input and determine a state of health (SOH) of the battery system.

[0016] Another aspect of the present disclosure provides a processing device for use in a battery monitoring system, the processing device comprising: a pre-processor configured to receive measurement data from at least one sensor; and a neural network configured to receive processed data from the pre-processor. Based on the received measurement data, the pre-processor: A normalized rate of change of a first measured characteristic relative to one of a second measured characteristic associated with the battery system, measured over time or during a charge or discharge operation. where the second measured characteristic is different from the first measured characteristic. The neural network is configured to use the determined normalized rate of change from the pre-processor as an input and determine a state of health (SOH) of the battery system.

[0017] Another aspect of the present disclosure provides a method for determining a state of health (SOH) of a battery system, according to which measurement data indicative of one or more characteristics associated with the battery system are received during a charge or discharge operation. Based on the received measurement data, A normalized rate of change of a first measured characteristic relative to one of a second measured characteristic associated with the battery system, measured over time or during a charge or discharge operation. is determined, where the second measured characteristic differs from the first measured characteristic. The determined normalized rate of change is then used as an input to a neural network to determine a state of health (SOH) of the battery system.

[0018] A further aspect of the present disclosure provides a non-transitory computer-readable medium for configuring a processing device or a monitoring system.

[0019] Other features of the present disclosure will become apparent from the dependent claims and the following description. Further features of the present disclosure will become apparent from the following description of exemplary embodiments, with reference to the accompanying drawings.

[0020] Further aspects and embodiments will become apparent from the following description and accompanying drawings. [Brief explanation of the drawings]

[0021] [Figure 1] 1 is a schematic diagram of a battery system and a monitoring system for determining the health of the battery system according to an embodiment of the present disclosure. [Figure 2] 1A and 1B are exemplary graphs illustrating how the rate of change of charge relative to voltage (dQ / dV) varies with the health of an exemplary battery system. [Figure 3] 1A and 1B are exemplary graphs illustrating how the normalized rate of change of charge relative to voltage (dQ / dV) varies with the health of an exemplary battery system. [Figure 4] 2 is a process flow diagram illustrating a method for determining the health of a battery system using the monitoring system of FIG. 1. [Figure 5] 1A and 1B are exemplary graphs of how the rate of change of voltage with respect to charge (dV / dQ), measured over a charge / discharge cycle, varies with the health of an exemplary battery system. [Figure 6] 1A and 1B are exemplary graphs of how the rate of change of voltage versus state of charge (dV / dSoC) varies with the health of an exemplary battery system, as measured over a charge / discharge cycle. [Figure 7]1 is an exemplary graph of the rate of change of voltage versus state of charge (dV / dSoC) measured over a charge / discharge cycle and the state of health of a battery system charged at several different charge rates. [Figure 8] 1 is a process flow diagram illustrating a method for training a neural network to determine the health of a battery system based on measured characteristic derivatives according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0022] FIG. 1 is a schematic diagram of a battery system 110 and a monitoring system 120 for determining the SoH of the battery system 110 according to an embodiment of the present disclosure.

[0023] Battery system 110 may include a single battery cell. Alternatively, in other embodiments, battery system 110 may be a battery pack including multiple battery cells. In embodiments where battery system 110 includes a single battery cell, monitoring system 120 is configured to determine the SoH of that battery cell. In other embodiments where battery system 110 includes multiple battery cells, monitoring system 120 may be configured to determine the collective SoH of the multiple battery cells.

[0024] The monitoring system 120 includes one or more sensors 125 for measuring characteristics of the battery system 110 during charging and / or discharging operations. The sensors 125 may include one or more sensors configured to measure the voltage and / or current of the battery system 110 during charging and / or discharging operations. In some embodiments, the sensors 125 may also include one or more sensors configured to measure the temperature of the battery system 110.

[0025] Monitoring system 120 further comprises a processing device 130 configured to receive signals and / or data from sensors 125 representing measured characteristics of battery system 110 (e.g., signals and / or data indicative of voltage, current, temperature, etc.). In some embodiments, processing device 130 may be configured to timestamp the received sensor data to facilitate time series analysis of the battery system characteristics.

[0026] In some embodiments, processing device 130 may be coupled to sensors 125. In other embodiments, a communications link exists between processing device 130 and sensors 125, allowing battery system sensor data to be relayed to processing device 130.

[0027] The complexity of the neural network required to derive a measure of SoH from the measured characteristics of the battery system can be reduced by preprocessing the measured characteristics. charge Applicants have recognized that the cost savings can be significantly reduced by processing normalized derivative data (i.e., normalized rates of change data) instead of raw measurements such as voltage, temperature, etc. One such simplification is the ability to derive SoH measurements using shallow neural networks rather than previously proposed deep neural networks. This reduces the computational burden on the processing device 130 to determine the SoH of the battery system 110, making it feasible to determine SoH measurements with limited processing power, such as may be available in embedded systems. Providing a shallow neural network also reduces the likelihood that the neural network will overfit the training data as it is trained, which can lead to errors during validation.

[0028] Moreover, applicants have realized that SoH measurements can be made based on such determined derivative data for a limited portion of charge / discharge cycles without significant loss of accuracy. Thus, by preprocessing data representative of measured characteristics to determine and normalize derivative data of the rate of change of such characteristics, SoH measurements can be determined for a wide range of battery systems, including battery systems in which the battery cells or battery packs are never fully charged or discharged.

[0029] To that end, in this embodiment, the processing device 130 comprises a pre-processor 135 configured to determine and normalize the rate of change of data of the measured characteristics received from the sensor 125, and a neural network 140 trained to determine the SoH from the measured battery system characteristics being processed by the pre-processor 135.

[0030] Normalizing the data rate of change of the measured properties and feeding such normalized data rate of change to an appropriately trained neural network 140 has several advantages over systems in which the neural network processes raw data: Utilizing normalized data rate of change has the potential to improve prediction accuracy, further reduce the size of the neural network needed to determine the SoH, reduce the training data requirements for training the neural network, and aid the trained system in being able to derive SoH for a wider range of battery cells and cell types.

[0031] 2A, 2B, and 3A and 3B illustrate why it can be advantageous to use normalized data change rates rather than raw data change rates.

[0032] Figures 2A and 2B show the voltage and current profiles of two 18650 LFP chemistry cells. Charge relative to3A and 3B are a pair of graphs showing how the rate of change (dQ / dV) of for the corresponding voltage Normalized charge 10 is a corresponding pair of graphs showing how the rate of change of dQ / dV is determined to change.

[0033] As will be appreciated, the graphs of FIGS. 2A and 2B are plotted against the voltage, which corresponds to the SoH of the battery. Charge relative to However, the rate of change of voltage (dQ / dV) against Normalized charge The graphs of FIGS. 3A and 3B show that the rate of change of dQ / dV exhibits a higher level of consistency across cells.

[0034] For example, FIGS. 3A and 3B show the voltage against Normalized charge The graph clearly shows how the peak in the rate of change of voltage (dQ / dV) moves to the right as the battery ages. In contrast, the voltage against As it is charge For the rate of change of (dQ / dV), a pattern exists, but is less clear.

[0035] Preprocessing the data change rates and feeding the normalized data change rates to the neural network reduces the complexity of the required neural network and the level of training data needed to train the neural network to derive a health measure, because normalizing such data removes the variation in peak values ​​(e.g., the variation from 7.2 to 6.5 in the peak values ​​shown in Figures 2A and 2B) and instead presents the neural network with the task of deriving an SoH value based solely on horizontal shift, rather than horizontal shift and amplitude change.

[0036] Furthermore, in addition to reducing the complexity of the neural network required to derive the SoH value, applicants have discovered that the fact that battery systems exhibit similar parabolic derivative characteristics comparable to the trained data set results in a relatively consistent pattern of variation in normalized data rate of change across multiple cell types, despite the precise relationship between SoH and normalized characteristic derivatives for different battery system variants.

[0037] Simply put, normalization allows for improved pattern recognition compared to processing based on absolute derivative values. Because absolute derivative values ​​vary significantly from one cell variation to another, a model trained using absolute derivative values ​​is therefore limited to deriving the SoH for a single type of cell variation. However, because most information about cell degradation is derived from the location of peaks rather than the absolute values ​​of such peaks, it has been found that training based on normalized derivative data can yield satisfactory results without significant loss of accuracy.

[0038] Therefore, for all the above reasons, it has been concluded that in addition to processing raw sensor measurements to derive data rate of change, it is advantageous to then normalize or rescale such data before feeding it to a neural network to convert it into a determined measurement of SoH; this not only reduces the complexity of the neural network required, but also reduces the training data requirements for training such a network, as change patterns in the normalized data rate of change can be interpolated and compared across multiple chemical cell types.

[0039] The described approach is therefore suitable for generating SoH prediction models for new or unknown cell variants. When adapting a model for use with a new or unknown cell type, an existing general prediction model can be first adapted to accommodate the new cell's specifications—e.g., thresholds for voltage, capacity, and current to capacity—with appropriate preprocessing changes. The required adaptations can be identified based on data available from the cell's specification sheet. While the performance of such a model may be sufficiently accurate for some applications, the accuracy of such a model can subsequently be improved by further modifying the system based on cell-specific test data and / or test data from similar cells. Advantageously, such an approach allows SoH models to be available during the early stages of cell availability when developing new cells and new systems.

[0040] Returning to FIG. 1 , the preprocessor 135 for processing the sensor data and converting such sensor data into a normalized data rate of change can take the form of, but is not limited to, a microprocessor, an embedded processor, or the like, or can be incorporated in a system-on-chip (SoC). In some embodiments, the preprocessor 135 may comprise fixed-function digital logic circuitry. Furthermore, according to some embodiments, the preprocessor 135 may be derived from a family of processors manufactured by Intel®, AMD®, Qualcomm®, Apple®, NVIDIA®, or the like. The preprocessor 135 may also be based on an ARM architecture, a mobile processor, a graphics processing unit, or the like. The disclosed embodiments are not limited to a particular type of preprocessor.

[0041] In one embodiment, neural network 140 is a feed-forward neural network with multiple processing layers trained in a manner described below to determine the SoH of the monitored battery system 110 and to determine the SoH from a particular normalized battery system characteristic derivative.

[0042] Preferably, neural network 140 is a shallow feed-forward neural network with a limited number of nodes and layers. In one embodiment, neural network 140 may include an input layer, a hidden layer, and an output layer. In some embodiments, neural network 140 may include multiple hidden layers. In some embodiments, neural network 140 may include between four and six hidden layers. As discussed above, the reliability of the SoH measurements derived by the shallow neural network is increased by pre-processing the data, performed by pre-processor 135, which results in a reduction of input variability in the input data and converts such input values ​​into derived SoH measurements, reducing the complexity of the task of neural network 140 and making the task more suitable for the capabilities of the shallow neural network.

[0043] In some embodiments, the processing device 130 may be configured to sequentially fire each of multiple layers of the neural network 140 to determine the SoH of the battery system 110. Firing the layers of the neural network 140 in this manner further reduces the processing resources required to determine the SoH by a factor equal to the number of layers in the neural network 140. This, combined with providing a preprocessor 135 that derives a normalized data rate of change from measurements generated by the sensors 125, further expedites the SoH determination by the monitoring system 120, which may have limited processing power. Furthermore, this can be achieved without jeopardizing the accuracy of the SoH determination. This sequential firing of the layers of the neural network 140 can be initiated by software on the processing device 130. Advantageously, the SOH determination of a battery system using the disclosed techniques can be performed in a single run of the neural network 140 using data collected for a given charge / discharge cycle.

[0044] In some embodiments, the processing device 130 may further comprise a post-processing means configured to remove noise or unexpected values ​​from the output of the neural network. In one embodiment, the post-processing means may comprise at least one of a low-pass filter or a smoothing filter.

[0045] The processing device 130 can report the determined battery system SoH, as well as, optionally, any additional measured characteristics of the battery system, to a battery management system (BMS) (not shown). The BMS can compare the performance of multiple battery systems (or battery cells). If the performance of a battery system is found to differ from other battery systems by more than a predetermined amount, this may indicate a battery system fault and prompt the BMS to raise a fault flag for the identified battery system. The predetermined amount may initially be set to a default binding established at the time of manufacture. The predetermined amount may be selected to correspond to behavior learned from a previous battery system. In some embodiments, the BMS can transmit the collected data to the cloud on a continuous / regular / irregular schedule. Identified differences in battery system performance can be used to optimize the operation and / or maintenance of the battery system.

[0046] FIG. 4 is a process flow diagram illustrating a method 200 for determining the health of the battery system 110 using the monitoring system 120 shown in FIG.

[0047] As an initial step 210, the battery system 110 undergoes at least a partial charge / discharge cycle, during which the sensors 125, in step 220, collect measurement data indicative of characteristics associated with the battery system 110 as it is charged / discharged.

[0048] In one embodiment, the collected measurement data may include current and voltage data. In some embodiments, the temperature of the battery system 110 may also be measured.

[0049] In an embodiment, measurement data can be collected by the sensor 125 over a subrange of a complete charge / discharge cycle of the battery system 110. For example, the measurement data can be collected over a maximum SoC of the battery system. chargeThe measurement data can be collected between 20 and 80% of capacity. In other embodiments, the measurement data can be collected during one of the following charge / discharge ranges: 30 to 70% SoC (or 70 to 30% SoC), 40 to 60% SoC (or 60 to 40% SoC). In one embodiment, the range of charge / discharge operation over which the measurement data is collected is one of 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or 100% change in SoC. While the described methods can provide sufficient SoH measurements using a smaller range of charge / discharge data, the accuracy and effectiveness of these methods for variations in battery systems (e.g., battery systems with different designs or cell chemistries) can be improved by extending the range of charge / discharge operation over which the measurement data is collected. In some embodiments, the charge / discharge range can be the range that naturally occurs in the application of the battery system. In other embodiments, the sub-ranges can be determined based on the properties of the battery system to avoid overcharging or over-discharging the battery system and extend the life of the battery system. Embodiments of the present disclosure are not limited to a single charge / discharge range. In some embodiments, the monitoring system 120 can determine different charge / discharge ranges for different charge / discharge cycles.

[0050] In step 220, the pre-processor 135 receives measurement data from the sensors 125 indicative of a characteristic associated with the battery system 110. The pre-processor then determines, in step 230, a rate of change of another characteristic of the battery system 110 over time or during a charge / discharge operation based on the received measurement data.

[0051] In one embodiment, the pre-processor 135 determines the following for the battery system 110 based on the measurement data received from the sensor 125: charge In the above embodiment, the received measurement data may include current and voltage measurement data associated with the battery system 110 spanning a period of charging and discharging operations.

[0052] 5A and 5B show the results of measurements taken during charge-discharge cycles. for electric charge 5A and 5B are exemplary graphs illustrating how the rate of change of voltage (dV / dQ) varies with the health of an exemplary battery system. The rate at which a battery system is charged or discharged is known as the C-rate. The C-rate of a charge / discharge operation can be determined based on the charge / discharge current passing through the battery system divided by the theoretical current draw at which the battery system delivers its nominal rated capacity in one hour. FIG. 5A shows the dV / dQ measured on the battery system 110 during a charge operation at a C-rate of 3.6, while FIG. 5B shows the dV / dQ measured on the battery system 110 during a charge operation at a C-rate of 4.8. As can be seen in the diagrams, a clear trend is discernible as the battery system 110 degrades from 100% to 85% SoH. The illustrated graphs can be used to identify where a battery system of unknown SoH fits within the graph to predict the SoH value of the battery system, for example, by interpolation / extrapolation. The x-axis of the diagram is derived by coulomb counting across a range starting from the open circuit voltage at Vmin using current measurement data from the battery system 110. charge In the disclosed method, the calculated SoH results obtained using dV / dQ analysis have proven to be highly accurate, ranging from 0.35Ah to 0.70Ah. charge Using the absolute dV / dQ relationship determined from measurement data collected across a range, we achieve an RMSE of less than 1%. Thus, unlike the prior art, the disclosed solution can accurately determine SoH at higher C rates using measurement data from partial charge / discharge cycles.

[0053] In another embodiment, pre-processor 135 is configured to determine the rate of change of voltage over SoC (dV / dSoC) of battery system 110 based on measurement data received from sensor 125. In such an embodiment, the received measurement data may include current and voltage data associated with battery system 110 spanning the duration of a charge / discharge operation. Pre-processor 135 may be configured to determine the SoC from the received measurement data using any suitable method known in the art.

[0054] Figures 6A and 6B show the state of charge measured during charge-discharge cycles. against 6A and 6B are exemplary graphs illustrating how the rate of change of voltage (dV / dSoC) varies with the state of health of an exemplary battery system. FIG. 6A shows the dV / dSoC measured on the battery system 110 during a charging operation at a C rate of 3.6, while FIG. 6B shows the dV / dQ measured on the battery system 110 during a charging operation at a C rate of 4.8. The x-axis of the graph represents the SOC of the battery system 110, calculated as the percentage of available A Ah relative to the maximum Q Ah achieved at the end of each charging cycle. Using the dV / dQ method described above, the graph shows a discernible trend as the battery system 110 degrades from 100% to 85% SoH. SoH simulations using dV / dSoC in the described manner achieved an RMSE of less than 1% using measurement data collected across a charging range of 25-80% SoC.

[0055] In many applications, it is desirable to vary the C-rate of charge / discharge operations based on the SoC of the battery system. For example, in some cases, the battery system 110 may be charged at a high C-rate when the SoC is below 50%, followed by a slow C-rate charge when the SoC is above 80%. When the charge rate transitions from the high C-rate to the slow C-rate, a voltage discontinuity, or step response, can be observed across the battery system 110 due to the internal resistance and charging characteristics of the battery system chemistry. This results in a discontinuity in the dV / dSOC measurement, as shown in FIG. 7. Unlike prior art SoH techniques, the disclosed method, combining differential analysis of battery system characteristics with trained neural network pattern recognition, has proven effective in determining the SoH of a battery system from measurement data collected across charge / discharge cycles involving multiple different charge rates.

[0056] Figure 7 shows the state of charge measured during the charge-discharge cycle. against 1 is an exemplary graph of the rate of change of voltage (dV / dSoC) and the resulting change in state of health of a battery system charged at multiple different charge rates. The graph shows the determined dV / dSoC for a battery system 110 charged to 40% SoC at C rate 6, followed by charging to 80% SoC at C rate 4. Using an additional processing step to mask absolute dV / dSOC values ​​below a fixed threshold (which may be based on the battery system's configuration) enables SoH determination using measurement data that includes multiple charge rates during a single charge cycle. Thus, further processing of the resulting battery system characteristic differential data ensures that SoH determination is not distorted by interference caused by swapping between charge rates.

[0057] In another embodiment, pre-processor 135 is configured to determine a rate of change including at least one of dV / dt, dV / dT, dQ / dT, dQ / dSoC, dQ / dt, and dSoC / dT. For embodiments in which the rate of change of a battery system characteristic with respect to time is determined, pre-processor 135 may first timestamp the received measurement data to analyze the battery system characteristic with respect to the time of the charge / discharge operation. Each battery system characteristic differential has its own advantages and disadvantages associated with it.

[0058] For example, while dV / dt has been found to provide a high degree of SoH accuracy for a single battery system design, dV / dt does not perform as well as other derivatives (dV / dQ and dV / dSoC) when used with variations in battery systems (e.g., dV / dt trained on cell chemistry A may not provide as accurate SoH predictions as untrained cell chemistry B). Advantageously, using dV / dt also yields a simpler implementation, since raw, time-stamped voltage measurements can be sampled directly without further data manipulation. The characteristic derivatives for a particular battery system used in the disclosed methods can therefore be selected based on the particular application.

[0059] The determined rate of change is then rescaled / normalized in step 240. Rescaling and / or normalization can be performed on the rate of change using any suitable method known in the art. As already mentioned above, rescaling / normalizing the determined rate of change improves the validity of the SoH determination from measurement data associated with battery system variations.

[0060] 6A and 6B, the range for the determined dV / dSoC may be rescaled to a range of 1 to -1, e.g., using 32 bits, to represent 1 to -1 instead of the raw value, to facilitate easier weighting and support higher resolution. Additionally, less accurate measurement data from sensor 125 may be offset by reducing the sampling interval of the measurement data used to determine dV / dSOC. In one embodiment, a sampling interval of 0.5 to 5% SOC may be used.

[0061] In step 250, the processed (i.e., rescaled / normalized) rate of change is used as input to the neural network 140, which uses pattern matching analysis on the received, normalized rate of change to output, in step 260, the SoH for the battery system 110.

[0062] In some embodiments, in addition to the determined normalized rate of change, the true average current measured by the battery system over the duration of the charge / discharge operation can also be used as an input to the neural network 140. Including the average current can improve the accuracy of the SoH and its adaptability to variations in the battery system. This is because normalized derivatives that independently include dV, dt, dT, and dQ do not provide an indication of the rate of change for a given step. While an average of the rotational current can be used in some embodiments, this can introduce errors if the charge / discharge rate changes during the charge / discharge cycle. Including the average current in a machine learning model based on the normalized derivative value can also offset any loss of accuracy resulting from the normalization step and prevent the neural network from deriving information from the absolute value rather than the normalized derivative value.

[0063] In some embodiments, in addition to the determined normalized rate of change, the temperature of the battery system over the duration of the charge / discharge operation can also be used as an input to neural network 140. In one embodiment, the true average temperature of the battery system over the duration of the charge / discharge operation can also be used as an input to neural network 140. Including temperature can also improve the accuracy of the SoH by correcting for changes in battery system behavior that result from temperature differences. Specifically, temperature variations affect both the internal resistance of the battery system and the degree to which a charge can be drawn from the battery system.

[0064] In some embodiments, post-processing can be performed on the output of the neural network 140 to remove outliers and unexpected noise spikes. In some embodiments, values ​​outside of expected parameters are removed. In one embodiment, where the trough of the SoH curve is outside the expected 40-60% SoC, a post-processing step can be implemented to remove this data from the curve, as shown in FIG. 6B. Post-processing the SoH output in this manner can offset the loss of accuracy that occurs from using measurement data collected across a partial charge / discharge range. It will be appreciated that the post-processing step can also perform standard deviation analysis, threshold levels, and comparison to historical / training data at identification, as well as outlier / noise removal from the determined data set.

[0065] The determined SoH of the battery system 110 can then be used to modify the effective operation of the battery system 110 (e.g., modify the charge / discharge operation to maintain the life of the battery system 110) and to determine when replacement of the battery system 110 (or one or more of its components) is necessary.

[0066] 8 is a process flow diagram illustrating a method 600 for training a neural network to determine the health of a battery system based on normalized characteristic derivatives measured according to an embodiment of the present disclosure. In step 610, training data is obtained that includes measurement data indicative of characteristics of the battery system associated with the battery system during charging and discharging operations. The training data can include rates of change expected in field applications.

[0067] In one embodiment, the neural network 140 may be trained using normalized battery system characteristics sampled from the MIT dataset. The training data set is available in the public domain - MIT - https: / / data.matr.io / 1 / .

[0068] Using the obtained training data, in step 620, the rate of change of a characteristic of the battery system, either with respect to time or with respect to another characteristic of the battery system, observed over charge and discharge operations is determined. The determined rate of change is then rescaled / normalized in step 630 using known techniques to standardize the training process and improve the effectiveness of neural network 140 across battery system variations. Specifically, as discussed above, rescaling trains neural network 140 to focus on the determined rate of change patterns and reduces the use of large values ​​in the implementation, thereby enabling a higher resolution, fixed-point conversion. In some embodiments, a filtering step may be implemented to remove unexpected values / outliers, as described above.

[0069] In step 640, the determined normalized rate of change is used to train neural network 140 and weight and set parameters for processing the rate of change to determine the SoH of the battery system. In some embodiments, the training parameters can be set using Bayesian regularization backpropagation.

[0070] The neural network 140 can have a configuration size based on application requirements. In one embodiment, the configuration of the base network may be [18 15 10 10 5 3]. In one embodiment, the neural network 140 utilizes the Tanh function to determine the State of Health of the battery system. Activation / Transfer Function: Tanh—allows rapid convergence with a limited number of nodes, although other known activation functions may be used. Vanishing gradients reduce the risk of exponential errors in untrained datasets. Alternative activation functions used in regression problems, such as the rectified linear unit (ReLU) function, require complex neural networks with an order of magnitude larger number of nodes to match the accuracy of predictions obtained from Tanh. Similar issues are expected with leaky ReLU and switch activation functions. The limited range of Tanh (-1 to 1) allows for very accurate conversion in fixed-point operations.

[0071] The results of the trained neural network can then be validated and tested against battery system data using techniques known in the art in step 650. In some embodiments, validation of the neural network 140 can be performed on a portion (e.g., 15%) of the entire MIT dataset. After the testing and validation step, the neural network 140 can be optimized in step 660 by adjusting weights and / or parameters to minimize error in the battery system SoH determination.

[0072] The implemented algorithm, trained on dV / dQ values, was proven to predict the SoH of a lithium iron phosphate (LFP) cell (A123 APR18650M1A 1.1 Ah) with an RMSE of less than 1% using calculated absolute dV / dQ values ​​obtained from Q Ah values ​​ranging from 0.35 Ah to 0.70 Ah. The algorithm, retrained on the same MIT dataset using rescaled dV / dSoC v SoC data, enabled accurate SoH predictions for another LFP cell (LithiumWerks APR18650M1B 1.2 Ah), with prediction errors of less than 3% RMSE. Rescaled dV / dQ or dV / dSoC provides a more general solution because it is more accurate across all types of battery systems.

[0073] Advantageously, a model generated in the above manner can derive a measurement of SoH without needing to know the battery's previous state, as required in some prior art systems, such as DE 10 2015 016 987 A1. This alleviates the requirement for the system to include non-volatile memory to store such state data or other state variables necessary to update the model. Historical data is not used in training. Rather, normalized data, along with average current values ​​and / or temperature, if needed, are directly linked to the SoH value.

[0074] In one embodiment, neural network training can occur in the cloud. The trained neural network 140 can be transferred to an embedded device and function without relying on the cloud. While training the neural network 140 in this manner eliminates the requirement for continuous connectivity to the cloud to make SoH decisions, in certain embodiments, a cloud connection can be utilized to assist in SoH decisions.

[0075] While the above embodiments describe techniques for determining a battery system SOH defined relative to capacity, it will be understood that such techniques apply equally to other definitions of SOH known in the art. Specifically, in some embodiments, SOH may be defined as one of the following: - Power health can be defined as the ratio of the current maximum power that a cell can deliver during a given time to the maximum power that the cell could deliver at the time of manufacture; The health of fast charging capability can be defined as the ratio of the current maximum charging rate to the maximum charging rate of a new cell; and The health of the cell with respect to its internal resistance can be defined as the ratio of the internal resistance that has increased with age to the internal resistance at the time of manufacture of the cell.

[0076] The descriptions of various embodiments of the present disclosure have been presented for illustrative purposes and are not intended to be exclusive or limiting to the scope of the present disclosure. Many modifications and variations of the disclosed embodiments will be apparent to those skilled in the art from the embodiments disclosed herein without departing from the scope of the present disclosure. The terms used herein to disclose the embodiments of the present disclosure were selected to best express the principles of the embodiments, practical applications or technical improvements to technology found in the market, or to enable those skilled in the art to understand the embodiments disclosed herein.

[0077] It is understood that certain features of the present disclosure, which are for clarity described in the context of separate embodiments, can also be provided in combination in a single embodiment. Conversely, various features of the present disclosure that are, for brevity, described in the context of a single embodiment, can also be provided separately or in any suitable subcombination, or as suitable in any other, described embodiment of the present disclosure. Certain features described in the context of various embodiments should not be considered essential features of those embodiments, unless the embodiment is inoperable without those elements.

[0078] While this disclosure has been described in conjunction with specific embodiments thereof, it is evident that many alternatives, modifications, and variations will be apparent to those skilled in the art. Accordingly, it is intended to embrace all such alternatives, modifications, and variations that fall within the spirit and broad scope of the appended claims. The inventions disclosed herein include the following: [Aspect 1] 1. A monitoring system for determining a state of health, or SOH, of a battery system comprising one or more battery cells, comprising: at least one sensor for measuring at least one characteristic associated with the battery system during a charging or discharging operation; a processing device communicatively coupled to the at least one sensor, receiving measurement data from the at least one sensor; and Based on the received measurement data, a normalized rate of change of a first measured characteristic relative to one of a second measured characteristic associated with the battery system measured over time or during the charge or discharge operation; a pre-processor configured to determine the second measured characteristic, wherein the second measured characteristic is different from the first measured characteristic; the processing device comprising: a neural network configured to determine a state of health (SOH) of the battery system using the determined normalized rate of change; A monitoring system comprising: [Aspect 2] 2. The monitoring system of claim 1, wherein the at least one sensor comprises at least one of a temperature sensor, a voltage sensor, a current sensor, or a combination of the voltage sensor, the current sensor, and / or the temperature sensor. [Aspect 3] The first measured characteristic is voltage, charge 3. The monitoring system of any one of aspects 1 to 2, further comprising one of: a power consumption status; [Aspect 4] The second measured property is charge 4. The monitoring system of any one of aspects 1 to 3, wherein the monitoring information includes one of: a battery status, a charging state, or a temperature. [Aspect 5] A monitoring system according to any one of aspects 1 to 4, wherein the first measured characteristic is voltage and the second measured characteristic is state of charge. [Aspect 6] The first measured characteristic is voltage and the second measured characteristic is charge 4. The monitoring system according to any one of aspects 1 to 3, wherein: [Aspect 7] The monitoring system of any one of aspects 1 to 6, wherein the neural network comprises a plurality of layers, and the processing device sequentially fires each layer of the plurality of layers to determine the health of the battery system. [Aspect 8] A battery system comprising the monitoring system according to any one of aspects 1 to 7. [Aspect 9] A processing device for use in the monitoring system according to any one of aspects 1 to 7, comprising: receiving, by at least one sensor, measurement data indicative of at least one characteristic associated with the battery system during a charging or discharging operation; and Based on the received measurement data, a normalized rate of change of a first measured characteristic relative to one of a second measured characteristic associated with the battery system measured over time or during the charge or discharge operation; a preprocessor configured to determine the second measured characteristic, wherein the second measured characteristic is different from the first measured characteristic; a neural network configured to determine a state of health (SOH) of the battery system using the determined normalized rate of change; The processing device. [Aspect 10] 1. A computer-implemented method for determining a state of health, or SOH, of a battery system, comprising: receiving, by at least one sensor, measurement data indicative of at least one characteristic associated with the battery system during a charging or discharging operation; Based on the received measurement data, a normalized rate of change of a first measured characteristic relative to one of a second measured characteristic associated with the battery system measured over time or during the charge or discharge operation; wherein the second measured characteristic is different from the first measured characteristic; and determining a state of health (SOH) of the battery system using a neural network, using the normalized determined rate of change as input data to the neural network; A method comprising: [Aspect 11] 11. The method of embodiment 10, further comprising normalizing the determined rate of change. [Aspect 12] 12. The method of claim 10 or 11, wherein the neural network comprises a plurality of layers, the method further comprising: firing each layer of the plurality of layers in succession to determine the state of health of the battery system. [Aspect 13] The first measured characteristic includes one of voltage, charge, or state of charge, and the second measured characteristic includes one of: charge , state of charge, or temperature, and preferably the first measured characteristic is voltage and the second measured characteristic is charge or state of charge. [Aspect 14] A computer-readable medium comprising instructions that, when executed by a processor, cause the processor to perform the method of any one of aspects 10-13.

Claims

1. 1. A monitoring system for determining a state of health, or SOH, of a battery system comprising one or more battery cells, comprising: at least one sensor for measuring at least one characteristic associated with the battery system during a charging or discharging operation; a processing device communicatively coupled to the at least one sensor, receiving measurement data from the at least one sensor; and a pre-processor configured to determine, based on the received measurement data, a normalized rate of change of a first measured characteristic relative to a second measured characteristic associated with the battery system measured over time or during the charge or discharge operation, the second measured characteristic being different from the first measured characteristic; and a shallow feedforward neural network configured to determine a state of health (SOH) of the battery system using the determined normalized rate of change; the processing device comprising: A monitoring system comprising:

2. The monitoring system of claim 1 , wherein the at least one sensor comprises at least one of a temperature sensor, a voltage sensor, a current sensor, or a combination of the voltage sensor, the current sensor, and / or the temperature sensor.

3. The monitoring system of claim 1 , wherein the first measured characteristic comprises one of a voltage, a charge, or a state of charge.

4. The monitoring system of claim 1 , wherein the second measured characteristic comprises one of charge, state of charge, or temperature.

5. 2. The monitoring system of claim 1, wherein the first measured characteristic is voltage and the second measured characteristic is state of charge.

6. 2. The monitoring system of claim 1, wherein the first measured characteristic is voltage and the second measured characteristic is charge.

7. 2. The monitoring system of claim 1, wherein the shallow feedforward neural network comprises a plurality of layers, and the processing device sequentially fires each layer of the plurality of layers to determine the state of health of the battery system.

8. A battery system comprising the monitoring system according to any one of claims 1 to 7.

9. A processing device for use in a monitoring system according to any one of claims 1 to 7, comprising: receiving, by at least one sensor, measurement data indicative of at least one characteristic associated with the battery system during a charging or discharging operation; and a pre-processor configured to determine, based on the received measurement data, a normalized rate of change of a first measured characteristic relative to a second measured characteristic associated with the battery system measured over time or during the charge or discharge operation, the second measured characteristic being different from the first measured characteristic; and a shallow feedforward neural network configured to determine a state of health (SOH) of the battery system using the determined normalized rate of change; The processing device.

10. 1. A computer-implemented method for determining a state of health, or SOH, of a battery system, comprising: receiving, by at least one sensor, measurement data indicative of at least one characteristic associated with the battery system during a charging or discharging operation; determining a normalized rate of change of a first measured characteristic relative to a second measured characteristic associated with the battery system measured over time or during the charge or discharge operation based on the received measurement data, the second measured characteristic being different from the first measured characteristic; determining a state of health (SOH) of the battery system using a shallow feedforward neural network, using the normalized determined rate of change as input data to the shallow feedforward neural network; A method comprising:

11. 11. The method of claim 10, wherein the shallow feedforward neural network comprises a plurality of layers, the method further comprising: firing each layer of the plurality of layers in succession to determine the state of health of the battery system.

12. 11. The method of claim 10, wherein the first measured characteristic comprises one of voltage, charge, or state of charge, and the second measured characteristic comprises one of charge, state of charge, or temperature, preferably the first measured characteristic is voltage and the second measured characteristic is charge or state of charge.

13. A computer readable medium comprising instructions which, when executed by a processor, cause the processor to perform the method of any one of claims 10 to 12.

Images