Systems and methods for monitoring gaseous analytes

Abstract

The present disclosure relates to systems and methods for monitoring and classifying gases released in a closed system having a gas source using a gas sensor that has been pre-trained a priori to distinguish off-gas events (OGE) or thermal run-off events (TRE) from non-OGE interfering gas releases. Pre-training utilizes one of machine learning (ML) or deep learning (DL) algorithms to pre-train the gas sensor to detect multiple known gas analytes to generate sensor signals with their own unique characteristics, extract features from the sensor signal to distinguish from non-OGE types of interfering gas releases, and establish a decision boundary or estimated probability of false positive releases of OGE or TRE. The established decision boundary or probability distribution is implemented as a candidate model for field deployment to classify the released gas to distinguish whether OGE or TRE has occurred.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to and the benefit of U.S. Provisional Patent Application Serial No. 63 / 040,260, filed on June 17, 2020, entitled “Systems And Methods For Monitoring A Gas Analyte.” This application also references U.S. Application No. 15 / 637,381, filed on June 29, 2017, and issued on December 29, 2020 as U.S. Patent No. 10,877,011 B2, which claims the benefit of U.S. Provisional Application No. 62 / 356,111, filed on June 29, 2016, entitled “SYSTEMS AND METHODS FOR ANALYTE DETECTION AND CONTROL,” and U.S. Provisional Application No. 62 / 454,516, filed on February 3, 2017, entitled “SYSTEMS INCLUDING AN ENERGY STORAGE ENCLOSURE AND MONITORING THEREOF,” the contents of which are incorporated herein by reference in their entireties. Technical Field

[0003] The present disclosure generally relates to systems and methods for monitoring and classifying gases released in a closed system having a gas source by means of a gas sensor that is a priori pre-trained to distinguish off-gas events (OGE) or thermal run off events (TRE) from non-OGE interfering gas releases. Background Art

[0004] Batteries are electronic devices that can store high-density electrical energy. As with any battery, thermal runaway events (TRE) can occur during discharge and charging. For example, thermal runaway can be caused by a short circuit within the battery (e.g., a cell), improper battery use, physical abuse, manufacturing defects, or exposure of the battery to extreme external temperatures. Thermal runaway occurs when the battery's internal reaction rate increases to the point where more heat can be generated than can be recovered, causing the internal reaction rate and heat generation to increase further.

[0005] The impact of a thermal runaway condition can depend on the battery type. For example, in flooded electrolyte batteries such as lead-acid batteries, a thermal runaway condition can cause the electrolyte to evaporate, leading to the escape of hazardous electrolyte gases (also known as an off-gassing event (OGE)) into the surrounding environment. In sealed batteries such as lithium-ion batteries, which can be used in devices such as electric vehicles, laptops, and cell phones, a thermal runaway condition can cause swelling, which can cause the sealed battery to explode and release hazardous electrolyte gases into the surrounding environment or create a fire hazard. Summary of the Invention

[0006] The present disclosure relates to systems and methods for monitoring and classifying gases released from a closed system having a gas source using a gas sensor that has been pre-trained a priori to distinguish an exhaust gas event (OGE) or a thermal runaway event (TRE) from non-OGE interfering gas releases. The pre-training utilizes one of a machine learning (ML) or deep learning (DL) algorithm to pre-train the gas sensor to detect multiple known gas analytes to generate sensor signals with their own unique characteristics, extract features from the sensor signal to distinguish from non-OGE interfering gas releases, and establish a decision boundary or estimated probability of false positive releases of OGE or TRE. The established decision boundary or estimated probability can be implemented as a candidate model for field deployment to classify the released gas as one or both of OGE and TRE to distinguish from non-OGE interfering gas releases.

[0007] A method for monitoring and classifying gas released in a closed system having a gas source may include the following steps: monitoring the gas source for releasing gaseous analytes by at least one gas sensor having one or more sensing electrodes, wherein, prior to initial field deployment of the sensor, the at least one gas sensor has been a priori pre-trained using one of a machine learning (ML) or deep learning (DL) algorithm to distinguish the released gaseous analytes from non-OGE interfering gas releases and classify the release as an event including one or both of the following: an off-gas event (OGE) or a thermal runaway event (TRE).

[0008] In an example, a priori pre-training of at least one gas sensor using an ML or DL ​​algorithm to classify released gas analytes may include at least the following steps: (1) training the at least one gas sensor to detect each of a plurality of known gas analytes over a duration of time through each of one or more sensing electrodes in the at least one gas sensor to generate a corresponding sensor signal representing a unique characteristic of each of the plurality of known gas analytes; (2) pre-processing the respective sensor signals generated during the time period to extract a corresponding plurality of features for each of the plurality of known gas analytes; (3) processing the extracted corresponding plurality of features to establish a decision boundary for false positive releases for one or both of OGE and TRE, and to establish a corresponding decision boundary for each of the remaining non-OGE types of interfering gas releases; and (4) storing the decision boundaries established in the pre-trained neural network of the ML or DL ​​algorithm providing estimated probability in a memory as one or more candidate models for post-field deployment of the sensor to distinguish gas analytes released from a gas source from non-OGE interfering gas releases as one or both of OGE or TRE.

[0009] In another embodiment, a system for monitoring and classifying gas released in a closed system having a gas source may include a housing having a gas source, at least one gas sensor having one or more sensing electrodes, the gas sensor being deployed to monitor the gas source for releasing a gaseous analyte, wherein at least one gas sensor has been a priori pre-trained using one of a machine learning (ML) or deep learning (DL) algorithm prior to deployment, the ML or DL ​​algorithm being stored as program code in a memory for execution by a processor to detect the released gaseous analyte and classify it into events including one or more of the following: an off-gas event (OGE), an interfering gas release event, and a thermal runaway event (TRE).

[0010] In an example, at least one gas sensor is pre-trained a priori using an ML or DL ​​algorithm to classify released gas analytes such that a processor pre-trains the at least one gas sensor prior to initial field deployment of the sensor to: (1) detect each of a plurality of known gas analytes over a duration of time via each of one or more sensing electrodes in the at least one gas sensor to generate a corresponding sensor signal representing a unique characteristic of each of the plurality of known gas analytes; (2) pre-process the respective sensor signals generated during the time period to extract a corresponding plurality of features for each of the plurality of known gas analytes; (3) process the extracted corresponding plurality of features to establish a decision boundary for false positive releases for one or both of OGE and TRE and to establish a corresponding decision boundary for each remaining non-OGE type of interfering gas from the plurality of known gas analytes; and (4) store the decision boundaries established in the ML or DL ​​algorithm in a memory as one or more candidate models for post-field deployment of the sensor to classify gas analytes released from a gas source into one or more of OGE, interfering gas events, and TRE.

[0011] The disclosed system for monitoring and classifying gas released from a closed system having a gas source using a gas sensor can be implemented as a non-transitory memory to store machine-readable instructions, wherein the gas sensor has been pre-trained a priori to distinguish an off-gas event (OGE) or a thermal runaway event (TRE) from a non-OGE interfering gas release. A processor can access the non-transitory memory and execute the machine-readable instructions on the machine to perform steps including: monitoring a gas source for releasing a gaseous analyte using at least one gas sensor having one or more sensing electrodes, wherein the at least one gas sensor has been pre-trained a priori using a machine learning (ML) or deep learning (DL) algorithm prior to initial field deployment of the sensor to distinguish the released gaseous analyte from a non-OGE interfering gas release and classify the released gaseous analyte as an event including one or both of the following: an off-gas event (OGE) or a thermal runaway event (TRE).

[0012] In an example, the machine-readable instructions may utilize an ML or DL ​​algorithm to a priori pre-train at least one gas sensor to classify released gas analytes, which may include at least the following steps:

[0013] (1) training at least one gas sensor to detect each of a plurality of known gas analytes through each of one or more sensing electrodes of the at least one gas sensor over a time duration to generate a corresponding sensor signal representing a unique characteristic of each of the plurality of known gas analytes; (2) preprocessing each sensor signal generated during the time duration to extract a corresponding plurality of features for each of the plurality of known gas analytes; (3) processing the extracted corresponding plurality of features to establish a decision boundary for false positive releases for one or both of OGE and TRE, and to establish a corresponding decision boundary for each of the remaining non-OGE types of interfering gas releases; and (4) storing the established decision boundaries in the ML or DL ​​algorithm in a memory as one or more candidate models for a field deployment of the sensor to distinguish gas analytes released from a gas source from non-OGE interfering gas releases and classify them as one or both of OGE and TRE. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 An example of a system 100 for monitoring exhaust gas events (OGE) is depicted.

[0015] Figure 2 Depicted are examples of equivalent circuit models of gas sensors that simulate monitoring and detection of events including off-gas events (OGE), thermal runaway events (TRE), and interfering (non-OGE) gas events.

[0016] Figure 3A and Figure 3B Two separate exhaust gas events are shown, which are depicted as Figure 2 The equivalent circuit model of the gas sensor shown extracts impedance and capacitance changes as ML features.

[0017] Figure 4 An example of OGE of a single-electrode gas sensor from an OGE gas source is depicted.

[0018] Figure 5 Depicted are examples of false positive detections of interfering gas sources from a single-electrode gas sensor using existing algorithms.

[0019] Figure 6 An example of OGE of at least one electrode gas sensor from an OGE gas source is depicted.

[0020] Figure 7 An example of an interfering gas event for at least one electrode gas sensor from a non-OGE gas source is depicted.

[0021] Figure 8 Depicts an example of the machine learning (ML) classifier design process.

[0022] Figure 9 Depicted is an example of a decision boundary separating real OGE from fake OGE, established by features extracted from a trained ML algorithm.

[0023] Figure 10 Depicted are examples of true and false exhaust events and other gas selectivity within the decision boundary trained using the ML algorithm.

[0024] Figure 11A and Figure 11B Depicted is an exemplary gas sensor pre-training flow chart applying a machine learning (ML) or deep learning (DL) algorithm.

[0025] Figure 12 Describes the implementation of convolutional neural networks in deep learning (DL) training.

[0026] Figure 13 Describes the implementation of Long Short-Term Memory (LSTM) neural networks in DL training.

[0027] Figure 14 Depicted a DL framework for time series classification.

[0028] Figure 15 Depicts examples of convolutional neural network architectures for executing DL algorithms. DETAILED DESCRIPTION

[0029] The present disclosure generally relates to systems and methods for monitoring a closed system having a gas source (e.g., a battery) for any of the following: off-gas events (OGE), thermal runaway events (TRE), and interfering (non-OGE) gas events using at least one gas sensor, wherein the at least one gas sensor has been pre-trained a priori (i.e., pre-trained in the factory) prior to initial deployment in the field, such that the deployed gas sensor does not require further training in the field and does not require the use of a reference gas sensor to detect gas release events.

[0030] Batteries may gradually degrade over their useful life, which may result in reduced capacity, cycle life, and safety. Degraded batteries may release gases, which may be referred to as "off-gassing." In one example, off-gassing may be released by the battery during cycling conditions (e.g., charge and discharge cycles). One or more causes of battery degradation may include improper battery use, physical abuse, manufacturing defects, exposure of the battery to extreme external temperatures, overcharging, and the like.

[0031] The systems and methods described herein can detect off-gas events (OGE) during cycling conditions and provide early warnings of thermal runaway event (TRE) conditions. In one example, the early warnings can include logic signal outputs, audible alarms, visual alarms, fire suppression, and communications with other systems and users. Off-gassing detected during cycling conditions can be interpreted as a warning that the battery is at risk of thermal runaway. By providing early warnings, fires, explosions, and injuries caused in response to thermal runaway conditions can be greatly mitigated. In addition, the systems and methods described herein can be configured to monitor any type of battery for off-gas conditions. Therefore, the systems and methods described herein can be used to monitor lithium-ion batteries and lead-acid batteries. In broader applications, the systems and methods described herein can be applicable to any closed system with a gas source for detecting gas leaks of flammable or toxic gases, such as nuclear reactor environments, oil and gas well drilling platforms, coal and gas generators, etc.

[0032] The terms "offgas," "released gas," and "gaseous analyte" are used interchangeably herein and refer to gaseous byproducts of chemical reactions of a gas source (e.g., a battery). Exhaust gases (i.e., "released gas" and "gaseous analyte") may include electrolyte gases, such as volatile electrolyte solvents, volatile components of the battery's electrolyte mixture, and the like. Volatile electrolyte or offgas analyte types may include at least the following flammable or toxic gases: lithium-ion battery exhaust, dimethyl carbonate, diethyl carbonate, ethyl methyl carbonate, ethylene carbonate, propylene carbonate, vinylene carbonate, carbon dioxide, carbon monoxide, hydrocarbons, methane, ethane, ethylene, propylene, propane, benzene, toluene, hydrogen, oxygen, nitrogen oxides, volatile organic compounds, toxic gases, hydrogen chloride, hydrogen fluoride, hydrogen sulfide, sulfur oxides, ammonia, and chlorine, and the like. Additionally, the terms "electrode" and "pad" may be used interchangeably to refer to a conductive terminal.

[0033] In addition, the systems and methods described herein can be configured with multiple battery housings. Thus, the systems and methods described herein can be used to monitor gaseous analytes ("off-gases") released by one or more batteries located within a battery housing. The term "battery housing" as used herein refers to any housing that can partially encapsulate one or more batteries. In an example, the battery housing can include ventilated and non-ventilated battery housings. A ventilated battery housing can include a ventilation system that can include an air inlet and an exhaust port. In a further example, the battery housing can include a battery transport container.

[0034] In addition, the term "processor" as used herein may refer to any device capable of executing machine-readable instructions, such as a computer, a controller, an integrated circuit (IC), a microchip, or any other device capable of implementing logic. The term "memory" as used herein may refer to non-transitory computer storage media, such as volatile memory (e.g., random access memory), non-volatile memory (e.g., hard drive, solid-state drive, flash memory, etc.), or a combination thereof.

[0035] Figure 1 An example of a system 100 for monitoring exhaust gas events (OGE) is shown. The system 100 includes at least one gas source, such as a battery 102, and at least one gas sensor 104, which is configured to monitor the release of a gaseous analyte 102a from the battery 102. The battery 102 may be a lithium-ion battery having a housing (i.e., enclosed in a protective shell). In embodiments, the gas sensor 104 may be a semiconductor-type gas sensor or any suitable gas sensor that detects the release of a gaseous analyte 102a (e.g., a chemical vapor) from the battery 102.

[0036] Unlike most other systems, the deployed gas sensor 104 eliminates the need to use a separate reference sensor in the system 100 to calculate a moving average from the real-time sensor signal 104a to detect off-gas events (OGEs) in the battery 102. In an implementation, the gas sensor 104 can be a sensor having one or more sensing electrodes 104b and has been pre-trained a priori (e.g., during manufacturing) using one of a machine learning (ML) or deep learning (DL) algorithm (program code) stored in the memory 108 for execution by the processor 106 to enable the gas sensor 104 to detect and classify any released gas analyte 102a in real time as one or more of the following events: off-gas events (OGEs), thermal runaway events (TREs), and interfering gas release events (i.e., non-OGEs). In an example, the gas sensor 104, the processor 106, and the memory 108 can be an integrated chip 101, such as an ASIC semiconductor chip. In other examples, the gas sensor 104 , the processor 106 , and the memory 108 may each be discrete components electrically connected by wiring harnesses or mounted on a printed circuit board (PCB).

[0037] The pre-trained gas sensor 104 may store the ML or DL ​​algorithm 108 a as a candidate model in the memory 108 to distinguish the sensor signal 104 a detected by the gas sensor 104 as one of OGE and TRE from non-OGE interfering gas events without requiring a reference gas sensor and without requiring retraining the gas sensor 104 after field deployment.

[0038] The machine learning and training of the algorithm 108a step can be performed a priori in the factory during the manufacturing process, or offline at any time prior to the physical commissioning or installation of the sensor 104 in the system 100. Once the sensor 104 is commissioned in the system 100, no real-time adjustments are required. Alternatively, however, in another option, the ML or DL ​​algorithm 108a can be retrained or updated by the sensor 104 to learn new encounters with other gas analytes that have not been pre-trained or listed in the database. The goal of such pre-training using the ML or DL ​​algorithm is not only to detect OGE, but also to be able to identify other gas sources detected by the sensor 104, thereby eliminating the need for a reference sensor.

[0039] In an example, performing a priori pre-training on at least one gas sensor 104 using an ML or DL ​​algorithm 108a may include at least the following steps: (1) training the at least one gas sensor to detect each of a plurality of known gas analytes (i.e., training gases) via each of one or more sensing electrodes of the at least one gas sensor 104 for a duration of time to generate a corresponding sensor signal 104a representing a unique characteristic of each of the plurality of known gas analytes used to establish a database; and (2) pre-processing the generated respective sensor signals 104a during the time period to extract a corresponding plurality of features (e.g., impedance and capacitance, see Figure 3A 、 Figures 3B to 7 ); (3) processing the corresponding multiple features extracted to establish a decision boundary for false positive release for one or both of OGE and TRE (see Figure 9 and Figure 10 ), and establish corresponding decision boundaries for each remaining non-OGE type of interfering gas release; and (4) storing the decision boundaries established in the ML or DL ​​algorithm (108a) in the memory 108 as one or more candidate models for field deployment of the sensor to distinguish the gas analyte 102a released by the gas source (e.g., battery) from the non-OGE interfering gas release as one or both of OGE or TRE. An output signal 110, such as a warning alarm or logic signal, can be sent for warning or for display on a screen to take preventive measures to avoid accidents or damage to the system 100.

[0040] In an example, the gas source can be a rechargeable lithium-ion battery system or an electrical energy storage system 102, wherein the gas analytes released in the OGE or TRE can be a combination of at least the following flammable or toxic gases: lithium-ion battery exhaust, dimethyl carbonate, diethyl carbonate, ethyl methyl carbonate, ethylene carbonate, propylene carbonate, vinylene carbonate, carbon dioxide, carbon monoxide, hydrocarbons, methane, ethane, ethylene, propylene, propane, benzene, toluene, hydrogen, oxygen, nitrogen oxides, volatile organic compounds, toxic gases, hydrogen chloride, hydrogen fluoride, hydrogen sulfide, sulfur oxides, ammonia and chlorine, etc.

[0041] In practice, ML or DL ​​learning methods during pre-training can be based on Figure 2 An equivalent circuit model 200 of the sensor 104 is shown. For illustration, Figure 2 The equivalent circuit model of a semiconductor gas sensor can be represented. However, other gas sensors can be represented using other equivalent circuit models or other transfer functions. The examples of gas sensor types and equivalent circuit models shown are non-limiting.

[0042] Figure 2 , the equivalent impedance circuit model 200 in FIG. 1 illustrates that the corresponding gas sensor signal generated by each of the one or more sensing electrodes 104 b in the at least one gas sensor 104 may include an impedance value based on a first parallel resistor Rn and capacitor CPEvr pair 204 connected in series with a second parallel resistor Rn and capacitor CPEs pair 206. The first parallel resistor and capacitor pair 204 in the equivalent circuit model 200 may simulate the dynamics of the at least one gas sensor response 104 a that may be exposed to different combinations of released analyte gases 102 a. The impedance equivalent circuit model 200 of the sensor 104 also includes a contact resistance Rc connected in series with the first parallel resistor Rn and capacitor CPEvr pair 204.

[0043] Given the output impedance R of the gas sensor 104 with a fixed input voltage V in the Laplace domain, a corresponding input / output transfer function can be developed for the equivalent circuit model 200, which can be expressed as:

[0044]

[0045] Where Rc is the contact resistance, Rn and Rs are the resistances of the circuit model, and CPEvr and CPEs are the capacitances of the gain in the circuit model.

[0046] The change in impedance values ​​(i.e., resistance and capacitance) from the electrical components (Rc, Rn, Rs, CPEvr, CPEs) caused by each of the plurality of known gas analytes 102a in the circuit model 200 over time (see Figure 3A 、 Figures 3B to 7 ) can be transmitted as a sensor signal 104a for training, representing a unique characteristic or property of a known gas analyte detected. The ML or DL ​​learning method simulates the monitoring and detection of the gas analyte 102a detected to be released from the gas source (i.e., the battery 102), thereby classifying the gas release event as one or a combination of an off-gas event (OGE), a thermal runaway event (TRE), and an interfering (non-OGE) gas event. The known gas analyte can be a gas mentioned in the OGE or TRE used for training, including but not limited to at least the following flammable or toxic gases: lithium-ion battery off-gas, dimethyl carbonate, diethyl carbonate, ethyl methyl carbonate, ethylene carbonate, propylene carbonate, vinylene carbonate, carbon dioxide, carbon monoxide, hydrocarbons, methane, ethane, ethylene, propylene, propane, benzene, toluene, hydrogen, oxygen, nitrogen oxides, volatile organic compounds, toxic gases, hydrogen chloride, hydrogen fluoride, hydrogen sulfide, sulfur oxides, ammonia, and chlorine, etc.

[0047] In an example, a mathematical representation using estimation techniques such as least squares and gradient algorithms may be used to extract equivalent resistance and capacitance in the equivalent model 200 of the gas sensor 104. In practice, pre-training the at least one gas sensor 104 to detect the release of the gaseous analyte 102a by an ML or DL ​​algorithm may include distinguishing between changes in sensor impedance due to environmental perturbations caused by one or more of: temperature changes, relative humidity changes, and other gases in the environment that affect the oxygen partial pressure causing a false positive to be reported.

[0048] In an example, the ML algorithm is pre-trained on at least one gas sensor 104 (see Figure 8 ) to extract the corresponding multiple features of each of the multiple known gas analytes may include utilizing any one or a combination of the following features: moving average calculation, Bollinger band, minimum electrode impedance, maximum impedance change rate, maximum impedance recovery rate of each of at least one electrode 104b on at least one gas sensor 104, principal component analysis (PCA), linear discriminant analysis, wherein the DL algorithm pre-trained on the at least one gas sensor to extract the corresponding multiple features of each of the multiple known gas analytes is included in the hidden layer of the neural network (see Figure 11A 、 Figure 11B and Figure 12 ).

[0049] In another example, pre-training the ML or DL ​​algorithm on at least one gas sensor to establish a decision boundary for false positive releases of OGE or TRE and a corresponding decision boundary for each remaining type of non-OGE interfering gas release may include evaluating the generated sensor signal 104a using any of the following determination methods: support vector machine, discriminant analysis or nearest neighbor algorithm, naive Bayes and neural neighborhood, linear regression, generalized linear model (GLM), support vector regression, Gaussian process regression (GPR), ensemble method, decision tree and DL neural network, the DL neural network including at least one of the following: Convolutional Neural Network (CNN) Figure 12 ) and Long Short Term Memory (LSTM) networks ( Figure 11A 、 Figure 11B ).

[0050] Figure 3A and Figure 3B Two separate exhaust gas events from the received sensor signal 104a are shown, these events being depicted as Figure 2 The equivalent circuit model of the gas sensor 104 shown in FIG extracts the change of impedance and capacitance with time duration as ML features (e.g., Figure 3A From area 320A to area 330A, Figure 3B These data can be used as training parameters for ML and DL machine learning algorithms to establish decision boundaries (see Figure 9 、 Figure 10 ).

[0051] Figure 4 An example of OGE for a single electrode gas sensor 104 from an OGE gas source is depicted. In the example, the sensor signal 104a parameter may include at least a moving average of changes in impedance value over a period of time that may result in an alarm being generated in the output signal 110. Furthermore, Figure 5 Depicted are examples of false positive detections of interfering gas sources (non-OGE) from a single-electrode gas sensor using existing algorithms.

[0052] Figure 6 An example of OGE of at least one electrode gas sensor from an OGE gas source is depicted. Figure 7Depicted is an example of an interfering gas event from at least one electrode gas sensor that originates from a non-OGE gas source. The unique response of each electrode in the gas sensor provides the ML learning algorithm with the ability to distinguish between OGE and common interfering gases in energy storage facilities.

[0053] The ML algorithm has unique selectivity, allowing for the identification of OGEs and identifying which interfering gas is present to enable other diagnostic capabilities. In contrast, increasing the number of electrodes 104b in the gas sensor 104 actually provides more unique responses, which can be unique characteristics or attributes (e.g., fingerprints) to help identify the detected gas analyte. Similarly, features extracted from one or more sensing electrodes can establish one or more dimensional dynamic responses while more accurately establishing false positive decision boundaries for OGE, TRE, and non-OGE type interfering gas events, thereby improving the reliability and accuracy of the trained candidate model (by reducing the probability of error) in identifying the type of gas analyte and classifying the event.

[0054] Figure 8 Depicted is an example of a machine learning (ML) classification design process or how to develop an ML algorithm. Pre-training (supervised learning) methods can combine many signal features (from the gas sensor 104) using heuristics and physics-based impedance information during the data pre-processing step of the algorithm development phase. It may also include environmental measurements such as temperature, pressure, and relative humidity included in the sensor suite. For example, signal features can include moving average, Bollinger Bands, minimum electrode impedance, maximum impedance change rate, maximum impedance recovery rate per electrode, principal component analysis (PCA), and linear discriminant analysis. In addition, pre-training (supervised learning) of gas sensors to distinguish OGE or TRE from non-OGE can also include, but are not limited to, classification techniques such as support vector machines, discriminant analysis or nearest neighbor algorithms, basic statistics of the distribution of time series values ​​(e.g., location, diffusion, Gaussianity, outlier properties), linear correlation (e.g., autocorrelation, power spectrum characteristics), stationarity (e.g., sliding window measurements, prediction error), information theory, and entropy / complexity, etc.

[0055] Table 1 below provides some examples of different features extracted from the sensor signal 104a in each of the plurality of electrodes 104b that can be used in the ML or DL ​​algorithm 108a for a priori training and for establishing a decision boundary to classify the gas analyte 102a as one of an OGE, a TRE, or a non-OGE interfering gas event (e.g., Figure 8 and Figure 11A 、 Figure 11B Machine learning is a supervised learning technique that can utilize linear regression, nearest neighbor, support vector regression, and neural networks in training to build candidate models 854 for use when deploying gas sensors 104 in the field.

[0056] Table 1

[0057] Feature 1 Feature 2 Feature 3 Feature 4 Feature 5 Feature 6 Gas type 0.774691 0.701516 0.674988 0.866383 -0.03044 -0.03406 Interference gas 1 0.829714 0.7555 0.838982 0.919975 -0.07514 -0.07582 Interference gas 1 0.768456 0.697034 0.675138 0.87619 -0.10458 -0.11013 Interference gas 1 0.920379 0.875693 0.958897 0.964187 -0.02719 -0.03223 Interference gas 1 0.885411 0.821586 0.952449 0.961552 -0.01485 -0.01715 Interference gas 2 0.827432 0.7544 0.833974 0.918782 -0.06566 -0.06751 Interference gas 2 0.737762 0.672099 0.610817 0.843494 -0.14042 -0.13649 Interference gas 2 0.778001 0.705386 0.698497 0.883416 -0.09282 -0.11494 Interference gas 2 0.738887 0.67214 0.616065 0.849154 -0.11429 -0.11681 Interference gas 3 0.78249 0.675069 0.699502 0.885526 -0.09325 -0.11944 Interference gas 3 0.818389 0.713922 0.814425 0.91466 -0.07889 -0.11787 Interference gas 3 0.774843 0.670419 0.692568 0.882074 -0.13414 -0.1787 Interference gas 3 0.932275 0.864262 0.964597 0.978366 -0.02777 -0.04665 OGE 0.892163 0.806155 0.952128 0.969763 -0.01669 -0.02967 OGE 0.845022 0.743398 0.870444 0.934548 -0.06587 -0.08874 OGE 0.732432 0.63537 0.610214 0.839241 -0.14826 -0.20349 OGE

[0058] Table 1 shows some examples of features extracted from each of at least one electrode for each gas type mentioned in the previous section. The converted data is listed in Table 1 and has various features extracted from the original data, which are used to create candidate models using an optimization process that searches for model parameters (including using the fitted data) and evaluates the model parameters using test data that was not used for fitting, and adjusts the candidate model until the best performance is achieved. Feature extraction in deep learning can be incorporated into neural networks (see Figure 11A 、 Figure 11B 、 Figure 12 、 Figure 13 、 Figure 14 A and Figure 14 B) and does not need to be performed before classification.

[0059] In another example, as part of the pre-training process, features extracted from known gas analytes can be used to train an ML or DL ​​algorithm to fully quantify the approximate percentage (%) or parts per million (ppm) of each identified gas analyte in the detected gas analyte composition, which may help classify one or a combination of OGE, TRE, and non-OGE as interfering gas releases from a gas source. Table 2 below illustrates some examples of detected gas analyte compositions.

[0060] Table 2

[0061]

[0062] Figure 9 An example of a decision boundary 930 separating true OGE 910 from false OGE 920 is depicted, established by features extracted using the trained ML algorithm 108a over a historical duration. Within region 910 means an OGE has been detected, while within region 920 means a non-OGE. In practice, based on time series data, decision boundary 930 will be high-dimensional (i.e., much higher than three dimensions), making it nearly impossible to identify a false OGE by using the feature extraction algorithm 108a. Figure 9 and Figure 10 The examples shown in depict all decision boundary graphs.

[0063] Figure 10 An example of selectivity of true OGE 1010 and false OGE and other gases (1021-1029) in a decision boundary 1030 trained with an ML algorithm 108a is depicted. The ML or DL ​​algorithm 108a may include a selectivity algorithm that uses Figure 8The same idea of ​​training the decision boundary is similar to the process or technique shown in Figure 8 A combination of several dimensions corresponding to other extracted features in the figure can be shown to illustrate different decision boundaries for each possible interfering gas that has been detected. The idea behind using these two techniques (i.e., composite dimensions) is to make it easier to detect false positives and to provide some diagnostic information to the user by identifying the gas being detected. The different areas (1021-1029) illustrate how different types of interfering gases can be detected by the pre-trained gas sensor 104. Figure 10 A true OGE region 1010 is shown along with several false positive (non-OGE) regions (1021-1029), including a decision boundary that enables an ML or DL ​​algorithm to identify false positive interfering gases.

[0064] Figure 11A and Figure 11B An exemplary flow chart for pre-training a gas sensor 104 using a machine learning (ML) algorithm using multiple known gas analytes is depicted. For example, in step 1102, for each known gas analyte, raw signals 104a, such as resistance and capacitance, can be generated from multiple electrodes 104b of the gas sensor 104 and sent to a processor for feature extraction (e.g., changes in impedance or transfer function over time) in step 1104. The feature extraction step 1104 can include a time-frequency transform (e.g., a discrete cosine transform (DCT) or a discrete Fourier transform (DFT)) to transform the time-domain analog signal into a frequency-domain signal. In step 1106, the extracted features can be organized. In step 1108, a machine learning (ML) algorithm is applied to the organized data to build a candidate model 1110 (e.g., a multidimensional decision boundary construction). By repeating steps 1102 to 1110 for the remaining plurality of known gas analytes, the candidate model 1110 may be updated (see step 1111 ) to build a database or construct a composite decision boundary graph to complete the training of the ML algorithm 108 a stored in the memory 108 for execution by the processor 106 . Figure 11B Shown Figure 11A Detailed information is provided for the ML learning step 1108 in FIG. More specifically, step 1108 can be accomplished by repeating training steps 1108a through 1108n. Each of the training steps (e.g., 1108a) can include sequentially performing the following operations: convolution, rectified linear unit (ReLU), and pooling. When used in conjunction with the multi-electrode gas sensor 104 to perform gas analyte classification, the deployed model 1112 is a field-ready ML algorithm.

[0065] Likewise, these desired classifications can be achieved using deep learning (DL) algorithms, such as Figure 11A 、 Figure 11B As shown in , using pre-trained convolutional neural networks (e.g., convolutional neural networks CNN and long short-term memory (LSTM)) and automatic signal feature extraction, as Figure 12 and Figure 13 shown.

[0066] Figure 14 An example of a DL framework for time series classification is shown. Figure 15 Depicts an example of a convolutional neural network architecture that uses the extracted features to perform a DL algorithm pre-trained. A deep learning (DL) algorithm can consist of multiple layers that implement nonlinear functions (see Figure 14 、 Figure 15 ). Each layer can take input from the output of the previous layer and apply a nonlinear transformation to compute its output (i.e., sequential pipelining). These nonlinear transformations can be determined by trainable parameters during the fitting process. Some deep learning architectures that can be implemented include convolutional neural networks (see Figure 11B ), start time and echo state network.

[0067] While the above specific examples have been illustrated and described herein, it will be appreciated that various other changes and modifications may be made without departing from the spirit and scope of the claimed subject matter. Furthermore, while various aspects of the claimed subject matter have been described herein, these aspects need not be used in combination. Accordingly, the appended claims are intended to cover all such changes and modifications as come within the scope of the claimed subject matter.

Claims

1. A method for monitoring a gaseous analyte, wherein: The method comprises: A gas source for releasing a gaseous analyte is monitored by at least one gas sensor having one or more sensing electrodes, wherein the at least one gas sensor has been a priori pre-trained using one of a machine learning (ML) or deep learning (DL) algorithm prior to initial field deployment of the sensor to distinguish the released gaseous analyte from non-OGE interfering gas releases as events comprising one or both of the following: an exhaust gas event (OGE) or a thermal runaway event (TRE), and The step of pre-training the at least one gas sensor using the ML or DL ​​algorithm to classify the released gas analytes includes: training the at least one gas sensor to detect each of a plurality of known gas analytes over a duration of time via each of the one or more sensing electrodes of the at least one gas sensor to generate a respective sensor signal representative of a unique characteristic of each of the plurality of known gas analytes; pre-processing the respective sensor signals generated during the duration to extract a respective plurality of features for each of the plurality of known gas analytes; Processing the extracted features to establish a decision boundary for false positive releases for one or both of the OGE or TRE, and establishing a corresponding decision boundary for each of the remaining non-OGE interfering gas releases; and The decision boundary established in the ML or DL ​​algorithm is stored in a memory as one or more candidate models for post-field deployment of sensors to distinguish the gas analytes released by the gas source from the non-OGE interfering gas release and classify them as one or both of the OGE or TRE.

2. The method according to claim 1, wherein The gas source includes a rechargeable lithium-ion battery system or an electrical energy storage system, wherein the OGE or TRE includes detection of the release of at least any one or combination of the following flammable or toxic gases: dimethyl carbonate, diethyl carbonate, ethyl methyl carbonate, ethylene carbonate, propylene carbonate, vinylene carbonate, carbon dioxide, carbon monoxide, methane, ethane, ethylene, propylene, propane, benzene, toluene, hydrogen, oxygen, nitrogen oxides, hydrogen chloride, hydrogen fluoride, hydrogen sulfide, sulfur oxides, ammonia and chlorine.

3. The method according to claim 1, wherein The respective gas sensor signal generated by each of the one or more sensing electrodes of the at least one gas sensor includes an impedance value based on an impedance equivalent circuit model having a first parallel resistor and capacitor pair connected in series with a second parallel resistor and capacitor pair.

4. The method according to claim 3, wherein: The first parallel resistor and capacitor pair in the equivalent circuit model simulates the dynamics of the response of the at least one gas sensor when exposed to a combination of different released analyte gases.

5. The method according to claim 4, wherein The impedance equivalent circuit model of the sensor further includes connecting a contact resistance in series with the first parallel resistor and capacitor pair.

6. The method according to claim 1, wherein Detecting the release of gaseous analyte from the gas source by the at least one gas sensor that has been pre-trained by the ML or DL ​​algorithm eliminates the use of a reference sensor.

7. The method according to claim 1, wherein Pre-training the at least one gas sensor by the ML or DL ​​algorithm to detect the release of the gaseous analyte further includes distinguishing changes in sensor impedance due to environmental perturbations caused by one or more of: temperature changes, relative humidity changes, and other gases that affect the oxygen partial pressure in the environment resulting in reporting a false positive.

8. The method according to claim 1, wherein The ML algorithm pre-training the at least one gas sensor to extract the corresponding multiple features of each of the multiple known gas analytes includes utilizing any one or a combination of the following features: minimum electrode impedance, maximum impedance change rate, and maximum impedance recovery rate of each of at least one electrode on the at least one gas sensor, wherein the DL algorithm pre-training the at least one gas sensor to extract the corresponding multiple features of each of the multiple known gas analytes is included in the hidden layer of the neural network.

9. The method according to claim 1, wherein The ML or DL ​​algorithm pre-trains the at least one gas sensor to establish a decision boundary for false positive releases of the OGE or the TRE and a corresponding decision boundary for each remaining type of non-OGE interfering gas release, including evaluating the generated sensor signal using any one of the following determination methods: support vector machine, discriminant analysis or nearest neighbor algorithm, naive Bayes and neural neighborhood, linear regression, generalized linear model GLM, support vector regression, Gaussian process regression GPR, integration method, decision tree and DL neural network, the DL neural network including at least one of the following: convolutional neural network CNN, Inception time architecture, echo state network and long short-term memory LSTM network.

10. A system for monitoring a gaseous analyte, wherein: The system comprises: an enclosure having an air supply; and At least one gas sensor having one or more sensing electrodes, the at least one gas sensor being deployed to monitor a gas source for releasing a gaseous analyte, wherein the at least one gas sensor has been a priori pre-trained prior to deployment using one of a machine learning (ML) or deep learning (DL) algorithm, the ML or DL ​​algorithm being stored as program code in a memory for execution by a processor to distinguish the released gaseous analyte from non-OGE interfering gas release detection and classify the released gaseous analyte as an event comprising one or both of the following: an exhaust gas event (OGE) or a thermal runaway event (TRE), wherein the at least one gas sensor is pre-trained a priori using the ML or DL ​​algorithm to classify the released gas analytes, such that the processor pre-trains the at least one gas sensor prior to initial field deployment of the sensor to: detecting each of a plurality of known gas analytes over a duration of time via each of the one or more sensing electrodes of the at least one gas sensor to generate a respective sensor signal representative of a unique characteristic of each of the plurality of known gas analytes; pre-processing the respective sensor signals generated during the duration to extract a respective plurality of features for each of the plurality of known gas analytes; Processing the extracted features to establish a decision boundary for false positive releases for one or both of the OGE and TRE, and establishing a corresponding decision boundary for each of the remaining non-OGE interfering gas releases; and The decision boundary established in the ML or DL ​​algorithm is stored in a memory as one or more candidate models for post-field deployment of sensors to distinguish the gas analytes released by the gas source from the non-OGE interfering gas release and classify them as one or both of the OGE or the TRE.

11. The system according to claim 10, wherein: The respective gas sensor signal generated by each of the one or more sensing electrodes in the at least one gas sensor includes an impedance value based on an impedance equivalent circuit model having a first parallel resistor and capacitor pair connected in series with a second parallel resistor and capacitor pair, and the first parallel resistor and capacitor pair in the equivalent circuit model simulates the dynamics of the response of the at least one gas sensor when exposed to different released gases.

12. The system according to claim 11, wherein Detecting the release of gaseous analyte from the gas source by the at least one gas sensor that has been pre-trained by the ML or DL ​​algorithm eliminates the use of a reference sensor.

13. The system according to claim 10, wherein: Pre-training the ML or DL ​​algorithm on the at least one gas sensor to extract a corresponding plurality of features for each of the plurality of known gas analytes includes utilizing any one or a combination of: minimum electrode impedance, maximum impedance change rate, and maximum impedance recovery rate for each of at least one electrode on the at least one gas sensor.

14. The system according to claim 10, wherein: The ML or DL ​​algorithm pre-trains the at least one gas sensor to establish a decision boundary for false positive releases of the OGE or the TRE and a corresponding decision boundary for each remaining type of non-OGE interfering gas release, including evaluating the generated sensor signal using any one of the following determination methods: support vector machine, discriminant analysis or nearest neighbor algorithm, naive Bayes and neural neighborhood, linear regression, generalized linear model GLM, support vector regression, Gaussian regression process GPR, integration method, decision tree and DL neural network, the DL neural network including at least one of the following: convolutional neural network CNN, Inception time architecture, echo state network and long short-term memory LSTM network.

Images