A multi-component gas precision detection system for energy storage devices

Abstract

The application discloses a kind of multi-component gas precision detection systems of energy storage device, belong to optical gas sensing and energy storage safety monitoring technical field.The device includes multi-wavelength step optical comb generation unit, sensing probe, photoelectric detection and time domain demultiplexing unit and concentration inversion unit, the multi-wavelength step optical comb generation unit is connected the light input end of the sensing probe, the light output end of the sensing probe is connected the input end of the photoelectric detection and time domain demultiplexing unit, the output end of the photoelectric detection and time domain demultiplexing unit is connected the concentration inversion unit.The application is synergized by time-domain step optical pulse and spatially separated sensitive film, enhances the selective response of multi-component gas, suppresses cross interference;Using multi-dimensional response vector extraction and improved kernel auto-encoder blind source separation model, nonlinear mixed signal can be decoupled under a small amount of calibration samples, improve the accuracy and practicality of multi-component gas concentration synchronous detection.

Description

Technical Field

[0001] This invention relates to the field of optical gas sensing and energy storage safety monitoring technology, specifically to a multi-component gas precision detection system for energy storage devices. Background Technology

[0002] Secondary energy storage devices are prone to thermal runaway under abuse conditions such as overcharging and overheating, releasing various characteristic gases such as hydrogen, carbon monoxide, and ethylene. Accurate real-time detection of these gas components is crucial for early warning and safety protection. Existing multi-component optical detection schemes, if employing multiple independent sensors, result in complex system configurations; if using broadband light sources combined with spectral analysis, issues such as cross-sensitivity between channels and nonlinear response can easily arise, affecting detection selectivity in multi-component mixed environments. Fiber optic evanescent field sensing technology has attracted attention due to its high sensitivity and intrinsic safety; however, when simultaneously distinguishing multiple gases on a single fiber resonant cavity, the refractive index changes caused by different gases can easily interfere with each other, and their mixed responses often exhibit nonlinearity, increasing the difficulty of accurately separating concentrations. At the signal processing level, traditional blind source separation methods are mostly based on linear mixing assumptions, making it difficult to effectively handle the nonlinear mappings introduced by sensitive thin-film adsorption and optical resonance processes. While using depth autoencoders for blind source separation shows potential, without appropriate statistical independence constraints, latent variables are unlikely to converge to the true source signal, and calibration typically relies on a large number of labeled samples, increasing the cost of practical applications. Competitive adsorption among sensitive films and fluctuations in ambient temperature and humidity mean that the actual mixing process, while based on the linearity described by the sensitivity matrix, is still superimposed with complex nonlinear interferences, making it difficult to completely decouple using a fixed physical model. Therefore, there is an urgent need for a multi-component gas detection scheme that can effectively suppress cross-interference, adapt to nonlinear mixing, and is easy to implement. Summary of the Invention

[0003] The purpose of this invention is to provide a multi-component gas precision detection system for energy storage devices. By synergistically combining time-domain gradient optical pulses and spatially separated sensitive films, the selective response of multi-component gases is enhanced, and cross-interference is suppressed. By utilizing multi-dimensional response vector extraction and an improved nucleated autoencoder blind source separation model, nonlinear mixed signals can be decoupled with a small number of calibration samples, thereby improving the accuracy and practicality of synchronous detection of multi-component gas concentrations.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A multi-component gas precision detection system for an energy storage device includes: a multi-wavelength gradient optical comb generation unit, a sensing probe, a photoelectric detection and time-domain demultiplexing unit, and a concentration inversion unit. The multi-wavelength gradient optical comb generation unit is connected to the optical input end of the sensing probe, the optical output end of the sensing probe is connected to the input end of the photoelectric detection and time-domain demultiplexing unit, and the output end of the photoelectric detection and time-domain demultiplexing unit is connected to the concentration inversion unit. The multi-wavelength gradient optical comb generation unit is used to generate a set of optical pulse sequences that are arranged sequentially in the time domain and do not overlap with each other in time. The optical pulse sequence contains at least two types of optical pulses with different center wavelengths, and each center wavelength is aligned with the optimal response wavelength of the corresponding sensitive area in the sensing probe. The sensing probe includes an optical fiber resonant cavity, on which at least two sensitive regions are respectively arranged at different positions along the light propagation direction. Each sensitive region has selective response capability to different analytes. The optical pulse sequence enters the optical fiber resonant cavity and interacts with each sensitive region in the form of an evanescent wave. The resonant characteristics of each optical pulse change due to the change in the concentration of the corresponding analyte, thereby outputting a transmitted light signal carrying information about each analyte. The photoelectric detection and time-domain demultiplexing unit is used to convert the transmitted light signal into an electrical signal and perform analog-to-digital conversion. It uses the time-domain position information of each light pulse to perform channel separation on the converted digital signal, extracts the peak characteristics of each pulse subsequence corresponding to different center wavelengths, and calculates the resonant wavelength shift and extinction ratio change caused by each component under test based on the peak characteristics. Then, it combines the resonant wavelength shift and extinction ratio change into a multidimensional response vector. The concentration inversion unit is equipped with a pre-trained nonlinear blind source separation model, which is configured to receive the multidimensional response vector and output real-time estimates of the concentrations of each analyte through decoupling mapping.

[0005] Furthermore, the multi-wavelength cascade optical comb generation unit includes a femtosecond fiber laser, a dispersive fiber, an optical amplifier, and a microring resonator array connected in sequence. The microring resonator array is composed of multiple microring resonator cavities with different free spectral ranges, used to slice the dispersion-broadened broadband chirped pulses into multiple narrowband pulses and output the optical pulse sequence in the time domain.

[0006] Furthermore, the fiber optic resonant cavity of the sensing probe is a micro-nano fiber optic resonant cavity, which includes a conical waist region with a waist diameter on the order of μm. Bragg gratings or fiber optic ring mirrors are provided at both ends of the conical waist region to form a resonant cavity. At least two sensitive regions are sensitive thin film arrays integrated at different longitudinal positions along the surface of the conical waist region by self-assembly or inkjet printing.

[0007] Furthermore, the sensitive thin film array has selective adsorption and response capabilities for hydrogen, carbon monoxide and ethylene respectively, and each sensitive thin film is spatially separated from the others, so that the resonant wavelength shift and extinction ratio change caused by different analytes are assigned to the corresponding light pulses in the time domain.

[0008] Furthermore, the photoelectric detection and time-domain demultiplexing unit includes a photodetector, a transimpedance amplifier, an analog-to-digital converter, and a digital signal processor connected in sequence. The analog-to-digital converter samples under the control of the frame synchronization signal. The digital signal processor has time-domain gated demultiplexing logic to generate a rectangular time window corresponding to the nominal position and width of each optical pulse, and to extract each pulse subsequence from the full frame sampled data.

[0009] Furthermore, the digital signal processor is configured to extract the peak time and peak amplitude of each pulse subsequence with subsampling precision, and to convert the extracted peak time and peak amplitude into the resonant wavelength offset and extinction ratio change respectively using a pre-calibrated sensitivity matrix.

[0010] Furthermore, the nonlinear blind source separation model in the concentration inversion unit is an improved kernelized autoencoder network, which includes a stochastic Fourier feature kernel mapping layer, at least two fully connected hidden layers and a linear output layer, and is solidified and deployed after two training stages: unsupervised pre-training and weakly supervised calibration.

[0011] Furthermore, the unsupervised pre-training stage utilizes unlabeled multidimensional response vector data to train the system with reconstruction loss and Hilbert-Schmidt independence criterion regularization term as optimization objectives, thereby making the latent variable components tend to be statistically independent; the weakly supervised calibration stage uses standard gas samples to determine the correspondence between each latent variable component and the concentration of the analyte component, as well as the linear correction coefficient.

[0012] Furthermore, the device also includes a controller, which is electrically connected to the multi-wavelength cascade optical comb generation unit, the sensing probe, and the photoelectric detection and time-domain demultiplexing unit, respectively, for generating and distributing frame synchronization signals to synchronize the generation of the optical pulse sequence, the temperature control of the optical fiber resonant cavity, and the sampling triggering of the analog-to-digital converter.

[0013] In summary, the present invention has the following beneficial technical effects: A multi-wavelength gradient optical comb generation unit generates temporally separated pulse sequences with different center wavelengths. These sequences, combined with a spatially separated selective sensitive thin-film array on the sensing probe, utilize frequency gradient multiplexing and the spatial overlap difference of the evanescent field to enhance the response specificity of each gas channel and help suppress cross-interference. The photoelectric detection and temporal demultiplexing unit employs temporal-gated demultiplexing and peak feature extraction to synchronously acquire the resonant wavelength shift and extinction ratio changes caused by each component from a single-path transmitted composite optical signal, forming a multidimensional response vector and improving information extraction efficiency. The concentration inversion unit utilizes an improved kernelized autoencoder nonlinear blind source separation model. Introducing a random Fourier feature kernel mapping and a regularization term based on the kernel Hilbert-Schmidt independence criterion, this model can decouple the multidimensional response vector into statistically independent gas concentration estimates even in the absence of extensive calibration data. This reduces the model's dependence on labeled samples and can adapt to the nonlinear mixing relationship between the sensing unit and the gas. The device employs all-fiber sensing probes and synchronous control, has a compact structure, and has the potential to be deployed in situ inside energy storage devices, which helps to improve the reliability and response speed of early warning of thermal runaway. Attached Figure Description

[0014] Figure 1 This is a block diagram of the device of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0016] like Figure 1 As shown, the present invention provides a multi-component gas precision detection system for an energy storage device, comprising: a multi-wavelength gradient optical comb generation unit, a sensing probe, a photoelectric detection and time-domain demultiplexing unit, and a concentration inversion unit. The multi-wavelength gradient optical comb generation unit is connected to the optical input terminal of the sensing probe, the optical output terminal of the sensing probe is connected to the input terminal of the photoelectric detection and time-domain demultiplexing unit, and the output terminal of the photoelectric detection and time-domain demultiplexing unit is connected to the concentration inversion unit. The multi-wavelength gradient optical comb generation unit is used to generate a set of optical pulse sequences that are arranged sequentially in the time domain and do not overlap with each other in time. The optical pulse sequence contains at least two types of optical pulses with different center wavelengths, and each center wavelength is aligned with the optimal response wavelength of the corresponding sensitive area in the sensing probe. The sensing probe includes an optical fiber resonant cavity, on which at least two sensitive regions are respectively arranged at different positions along the light propagation direction. Each sensitive region has selective response capability to different analytes. The optical pulse sequence enters the optical fiber resonant cavity and interacts with each sensitive region in the form of an evanescent wave. The resonant characteristics of each optical pulse change due to the change in the concentration of the corresponding analyte, thereby outputting a transmitted light signal carrying information about each analyte. The photoelectric detection and time-domain demultiplexing unit is used to convert the transmitted light signal into an electrical signal and perform analog-to-digital conversion. It uses the time-domain position information of each light pulse to perform channel separation on the converted digital signal, extracts the peak characteristics of each pulse subsequence corresponding to different center wavelengths, and calculates the resonant wavelength shift and extinction ratio change caused by each component under test based on the peak characteristics. Then, it combines the resonant wavelength shift and extinction ratio change into a multidimensional response vector. The concentration inversion unit is equipped with a pre-trained nonlinear blind source separation model, which is configured to receive the multidimensional response vector and output real-time estimates of the concentrations of each analyte through decoupling mapping.

[0017] Next, each component of the device will be described in more detail: 1. Multi-wavelength stepped optical comb generation unit This unit generates a sequence of optical pulses arranged at strict equal intervals in the time domain, with each pulse's center wavelength aligned with the characteristic absorption lines of different gases. Its core comprises a passively mode-locked femtosecond fiber laser, outputting an ultrashort optical pulse train with a repetition frequency on the order of MHz. This pulse train is first linearly chirped broadened through a precisely truncated dispersive fiber, separating the different wavelength components in time to form a broadband pulse whose instantaneous frequency changes approximately linearly with time. The broadened optical signal is then boosted by an optical amplifier and injected into a microring resonator array. This array consists of multiple cascaded or parallel microring resonators with different free spectral ranges. The resonant peak of each microring is precisely tuned to a different target wavelength. The Lorentz transfer function at the download port slices the spectrum of the input chirped pulse into multiple narrowband pulses, thus outputting sequentially arranged sub-comb pulses in the time domain. The center frequency of each pulse corresponds to the operating wavelength required for subsequent sensing. The time-domain interval and pulse width of each sub-comb pulse are jointly determined by the micro-ring parameters and chirp rate, and are locked by the controller through a thermo-optical tuning signal applied to the micro-ring. The resulting multi-wavelength cascaded optical comb is then adjusted by a polarization controller and coupled into a single-mode transmission fiber, which is then transmitted to the sensing probe.

[0018] 2. Sensor probe The sensing probe is an all-fiber structure, its main body being a micro / nano fiber resonant cavity tapered to the μm scale. This micro / nano fiber includes a tapered waist region with a diameter of only a few μm, at both ends of which are formed by Bragg gratings or fiber ring mirrors to create a high-quality-factor resonant cavity. The optical field within the cavity penetrates to a certain depth outside the tapered waist surface in the form of an evanescent wave. On the tapered waist surface of the micro / nano fiber, three spatially separated arrays of sensitive thin films are integrated through layer-by-layer self-assembly or inkjet printing, exhibiting selective adsorption and response capabilities for hydrogen, carbon monoxide, and ethylene, respectively. The positions of each sensitive film strictly correspond to different longitudinal regions on the tapered waist, allowing the equivalent refractive index of the evanescent wave at different spatial locations within the resonant cavity to be independently modulated by their respective target gases. When each sub-pulse of the multi-wavelength gradient optical comb passes through the resonant cavity, because the cavity resonant frequency is pre-aligned with the center frequency of the corresponding sub-comb through temperature tuning, the resonant wavelength shift and extinction ratio change caused by each gas are respectively converted into peak delay and peak amplitude changes of the corresponding sub-pulse. The three-channel response exhibits no crosstalk in the time domain due to the different pulse arrival times. A semiconductor cooling temperature control module and negative feedback circuit are externally mounted on the sensing probe to ensure a constant operating temperature of the micro / nano fiber optic resonant cavity. The output of the sensing probe is also connected to the subsequent detection unit via a single-mode transmission fiber, forming a transmissive sensing optical path.

[0019] 3. Photoelectric detection and time-domain demultiplexing unit This unit receives the composite transmitted light signal carrying three-channel gas information transmitted back from the sensing probe and converts it into a multidimensional response vector. The internal configuration is as follows: First, a broadband photodetector is set up with a response band covering all sub-comb wavelengths and a bandwidth of no less than 500MHz to linearly convert the incident light pulse power into photocurrent. Following this is a low-noise transimpedance amplifier, which converts the weak photocurrent into a voltage signal and incorporates a bandpass filter network to suppress out-of-band noise. The analog voltage signal output from the transimpedance amplifier is fed into a high-speed analog-to-digital converter (ADC) with a sampling rate of no less than 1 GSa / s and a resolution of 14 bits. Data acquisition for each scan cycle is initiated by a frame synchronization signal from the controller as an external trigger source. The digital sequence output from the high-speed ADC is fed into a field-programmable gate array (FPGA). The FPGA internally incorporates time-domain gated demultiplexing logic, generating rectangular time windows using preset nominal pulse positions and widths for each channel. Single-channel pulse sub-sequences corresponding to the three gases are extracted in parallel from the full-frame sampling sequence. Subsequently, the FPGA's built-in interpolation and parabolic fitting modules extract the peak time and peak amplitude of each sub-sequence with sub-sampling precision. Then, based on a pre-calibrated sensitivity matrix, the resonant wavelength shift and extinction ratio change for that cycle are calculated in real time. The calculated resonant wavelength shift and extinction ratio change values ​​of each channel are sent to the digital signal processor. The digital signal processor assembles the six parameters into a six-dimensional response vector in a fixed order of hydrogen, carbon monoxide, and ethylene, and sends it to the concentration inversion unit through a high-speed data bus.

[0020] 4. Concentration Inversion Unit The concentration inversion unit is an embedded industrial computer or workstation equipped with a graphics processing accelerator, internally running a trained modified kernelized autoencoder nonlinear blind source separation model. This unit receives a real-time six-dimensional response vector stream from the photoelectric detection and temporal demultiplexing unit via a data interface. In the preprocessing module, the original vector is first subjected to sliding window denoising and normalization to eliminate slow-varying drift and amplitude scale differences. The normalized data is then fed into the encoder network embedded in the hardware. The encoder front end includes a stochastic Fourier feature kernel mapping layer, mapping the six-dimensional input to a high-dimensional approximate kernel space. This is then compressed into a three-dimensional latent vector through two fully connected hidden layers and a linear output layer. This latent vector is ensured to be non-negative by a ReLU activation, and then passed through a linear correction layer calibrated by a small number of labeled gas samples, ultimately outputting real-time estimates of hydrogen, carbon monoxide, and ethylene concentrations simultaneously. Before deployment, the unit underwent two phases of training: first, unsupervised blind pre-training was performed using a large amount of unlabeled online data, employing reconstruction loss and Hilbert-Schmidt independence criterion regularization to make the components of the latent variables tend to be statistically independent; then, weakly supervised calibration was performed using a very small number of standard gas samples to determine channel correspondences and linear scaling. During online inference, only encoder forward computation and correction need to be performed to complete high-precision multi-component concentration inversion within each scan cycle.

[0021] 5. Controller and Auxiliary Module As the timing and control center of the entire device, the controller generates frame synchronization signals and distributes them to the micro-ring tuning driver in the multi-wavelength tiered optical comb generation unit, the temperature control module in the sensing probe, and the high-speed analog-to-digital converter in the photoelectric detection and time-domain demultiplexing unit, ensuring strict synchronization between optical comb generation, sensing, and signal acquisition. In addition, the controller is also responsible for monitoring the laser status, cooler temperature, and each power rail, enabling fault protection and health management of the device.

[0022] The above components are interconnected through single-mode transmission optical fibers and shielded cables, forming a complete measurement link from optical comb generation, selective optical sensing, photoelectric time-domain demultiplexing to intelligent concentration inversion.

[0023] Next, for ease of understanding, the present invention also provides a method for accurate detection of multi-component gases in an energy storage device implemented using the above-mentioned apparatus, specifically including the following steps: Step 1: Frequency-sequential multiplexed optical signal injection; Step 2: Selective resonant coupling of multi-component gases; Step 3: Photoelectric detection, time-domain demultiplexing, and multidimensional response vector extraction; Step 4: Concentration inversion based on nonlinear blind source separation using an improved nuclear autoencoder.

[0024] Next, the above steps will be explained in detail.

[0025] In step 1, the frequency-sequentially multiplexed optical signal injection is specifically as follows: The purpose of step 1 is to generate a periodic cycle, consisting of... The center frequencies are respectively A sequence of narrowband optical pulses is synthesized and injected into a single-mode transmission fiber as a detection signal to sequentially excite the sensing probe to selectively detect different target gases within one cycle. The core of this unit has two optional implementation methods: one is to use a passively mode-locked femtosecond fiber laser to generate broadband pulses, which are then broadened by a dispersive fiber and sliced ​​by a micro-ring array (as described in the device description); the other is to use a tunable laser to generate swept-frequency continuous light, which sequentially sweeps through the resonant frequencies of the micro-rings to generate time-domain pulses. For ease of description, this invention will use the swept-frequency continuous light implementation as an example for mathematical description; its pulse generation mechanism and time-frequency mapping relationship are equivalent to the spectral slicing method. In the spectral slicing method, the broadband pulse output from the femtosecond laser is linearly chirped broadened by a dispersive fiber, separating the different wavelength components in the time domain. The narrowband filtering effect of each download port of the microring resonator array slices the spectrum of the chirped pulse into multiple narrowband pulses. Their time domain positions are determined by the corresponding times of each microring resonator frequency in the chirped pulse time-frequency mapping. Therefore, a sequence of optical pulses without temporal overlap can be output sequentially in the time domain. This step is entirely completed by the multi-wavelength gradient optical comb generation unit driven by the controller. The multi-wavelength gradient optical comb generation unit contains a tunable laser (in the spectral slicing method, a passively mode-locked femtosecond fiber laser combined with a dispersive fiber), and a... The microring resonator array, composed of cascaded microring resonators, is described in detail through its specific generation steps: Step 101: The laser generates continuous light with a linear frequency chirp. The arbitrary waveform generator inside the controller generates a periodic linear sawtooth voltage signal. Its repetition period is the scan period. The voltage signal is applied to the current drive port of the laser. The instantaneous frequency of the laser output light... It exhibits a linear relationship with the driving voltage within a preset range. (Setting) The range of variation within a period is The corresponding range of output light frequency variation is This range is called a free spectral range and must cover all target resonant peak frequencies. And with an appropriate margin. Therefore, the instantaneous frequency of the continuous light output by the laser can be described by the following formula: ; In the formula, The linear frequency chirp rate provided for a laser is defined as: ; in, The starting frequency, For the termination frequency, This is the sweep period. The swept continuous light has an extremely narrow instantaneous linewidth, typically less than [value missing]. The output is achieved through a single-mode polarization-maintaining fiber optic cable to ensure coherence during subsequent resonant cavity excitation.

[0026] Step 102: The micro-ring resonator array converts the swept light into spatiotemporally separated sub-optical comb pulses. A swept-frequency continuous-wave optical coupling is introduced into the common input waveguide of the microring resonator array. This microring resonator array consists of... The system consists of several micro-ring resonators arranged in a cascaded manner. All micro-ring resonators share the same input waveguide, but each has its own independent download port. For the first... Each micro-ring resonant cavity, through precise design and thermo-optical tuning, is set to have a specific resonant frequency. ,That The resonant bandwidth is denoted as When the frequency of the sweeping laser... Slowly sweep across When the resonance condition is met, the light field will be significantly enhanced within the micro-ring resonant cavity, and a light pulse will be output through the pull-down port.

[0027] No. The normalized transmission spectrum function of each download port is approximately Lorentz-shaped, i.e.: ; in, Represents the instantaneous input light frequency. This is the resonant center frequency of the microring. For bandwidth.

[0028] Because the frequency of the input light changes linearly with time, i.e. The change in light intensity output from the download port of the microring resonator over time This will reproduce the Lorentzian time mapping described above, outputting a brief light pulse. Pulse occurrence time satisfy It is determined by the following formula: ; The power full width at half maximum (FWHM) of the optical pulse The value is determined by both the sweep rate and the bandwidth of the microring resonator, and is as follows: ; The controller uses a thermo-optical tuner to set the resonant frequency of each microring. Precisely locked at the preset value, and ensure This ensures that the output pulse strictly follows the time domain. The sequence appears periodically. In a specific implementation, it is set... ,correspond , and Three target gases, each with a frequency value , , Frequency sweep range , Scan cycle The chirping rate can be calculated from this. Pulse delay , , ,when At that time, pulse width This forms a clear and distinguishable time-division multiplexing sequence.

[0029] These pulses output from each download port are combined by an on-chip combiner to form a central pulse carrying... A sequence of light synthesized in a sequence of independent frequency components. The photoelectric field of the entire sequence can be uniformly described as follows: ; In the formula, For the first The slowly varying amplitude of the optical comb pulse Its initial phase, The normalized rectangular function is used here to characterize the presence or absence of the pulse envelope; its specific shape is determined by... The time mapping is precisely determined, and a rectangular gate is used here as a schematic diagram. This sequence of synthesized light is precisely coupled into the core of a single-mode transmission fiber and transmitted along the fiber to the sensing probe.

[0030] Step 103: Generation of control and synchronization signals The controller outputs a linear sawtooth wave drive signal. At the same time, it also generates a frame synchronization signal that is strictly synchronized with the scan cycle. This signal is a signal in At a certain time (at the beginning of each scan cycle), a rising edge transistor-to-transistor logic level pulse is generated, with a repetition frequency of... In this invention, .Should The signal is transmitted to the photoelectric detection and time-domain demultiplexing unit via a coaxial cable, serving as the reference clock for time-domain gated demultiplexing in the subsequent step 3, thereby establishing a strict timing relationship between the light source module and the detection unit.

[0031] In step 2, the selective resonant coupling of the multi-component gas is specifically as follows: Step 2 is performed by a sensing probe. This probe consists of a micro / nano fiber resonator located at the end of a single-mode transmission fiber and an array of sensitive thin films modified on its tapered waist surface. The sequence of light injected in step 1 is then used to synthesize the light. After transmission via single-mode fiber, the signal is sequentially enhanced with the micro / nano fiber resonant cavity in this step, and selectively coupled with different functional layers of the sensitive thin-film array at different frequencies to generate evanescent field coupling. This converts the target gas molecule information present at that moment into detectable changes in resonant characteristic parameters. This process is then broken down into several sub-steps, specifically including: Step 201: The sequence of synthesized optical light is coupled into the micro / nano fiber resonant cavity and excited to resonate. Ejected from the core of a single-mode transmission optical fiber First, it enters the micro / nano fiber resonant cavity. This micro / nano fiber resonant cavity is a Fabry-Perot cavity composed of two Bragg grating mirrors, with a uniform diameter, denoted as the waist region of a micro / nano fiber formed by stretching an adiabatic fused cone in the middle. The length is the cavity length of the resonant cavity. Free spectral range Depend on Effective refractive index of the cone waist region The decision, their relationship is as follows: ; In the formula, The speed of light in a vacuum. The quality factor of this micro / nano fiber optic resonator. Extremely high, exhibiting a series of discrete narrow-band resonance peaks in the transmission spectrum, with linewidths... With quality factor The relationship satisfies: ; In the implementation, set , ,but The three preset frequency steps in step 1... , , Through micro / nano fabrication and stress tuning, they have been precisely aligned with the center frequencies of three adjacent resonant peaks in the transmission spectrum of the micro / nano fiber resonator. Therefore, when the first of the sequenced synthesized light... Each optical comb pulse at the center frequency When the fiber enters the micro / nano fiber resonant cavity, the cavity resonance condition is satisfied: ; In the formula The value is an integer. This pulse generates constructive interference within the cavity, forming a standing wave field. The circulating optical power within the cavity is much greater than the input power, achieving optical pre-enhancement of weak signals. Because the frequencies of each channel do not overlap, within a pulse duration... Within this range, only the frequency component is amplified, thus forming a tiered selection mechanism where "one frequency corresponds to one independent resonant channel".

[0032] Step 202: Generation of resonant-enhanced evanescent field and its gradient interaction with composite gas-sensitive thin film In the resonant state, the optical field within the cone waist region of the micro / nano fiber resonant cavity exists as a Fabry-Perot standing wave, with a significant portion of its energy distributed outside the fiber surface as an evanescent field, penetrating to a depth of [insert depth here]. From the operating wavelength and the refractive index of the cone waist region and the refractive index of the external medium The decision is made jointly, and its expression is: ; The evanescent field strongly overlaps with the array of sensitive thin films directly deposited on the cone waist surface.

[0033] The sensitive thin-film array is constructed into a three-layer heterostructure using a layer-by-layer self-assembly technique, consisting of the following layers from the inside out: First layer: Palladium-doped zinc oxide nanofilm, thickness ,right It exhibits specific adsorption and response; Second layer: Indium oxide nanofilm with a specific crystalline phase, thickness... ,right Sensitive; Third layer: Polydopamine-functionalized carbon nanotube membrane, thickness ,right It has a strong affinity.

[0034] Each gas-sensitive thin film material exhibits differentiated light absorption cross-sections and refractive index variations at specific frequencies. Specifically, the frequency-thin film layer correspondence designed in step 1 is as follows: Located near the absorption peak of the palladium-doped zinc oxide nanofilm, Located in the absorption enhancement region of the indium oxide nanofilm, It is located at the sensitive wavelength of the polydopamine-functionalized carbon nanotube film. Therefore, when the resonant frequency is When a pulsed evanescent field is applied to a sensitive thin-film array, the evanescent field energy is selectively and maximally coupled to the corresponding sensitive layer, while the interference response contributed by other non-target layers is physically suppressed because their absorption cross-section at that frequency is extremely small. This physical mechanism is the core of "frequency gradient selective sensitization".

[0035] Step 203: Gas molecule adsorption causes a change in the complex refractive index of the sensitive layer When target gas molecules are present in the test environment, the gas molecules diffuse through the loose structure of the polydopamine-functionalized carbon nanotube membrane, reaching their respective sensitive layers and undergoing physical or chemical adsorption. Taking the interaction between molecules and palladium-doped zinc oxide nanofilms as an example, palladium nanoparticle catalysis The hydrogen atoms dissociate into hydrogen atoms, which diffuse into the zinc oxide lattice, altering the free carrier concentration of the thin film and thus affecting its complex refractive index. real part and the virtual part All have changed. Similarly, Molecules react on the surface of indium oxide, affecting the width of its surface depletion layer and altering its molecular structure. and ; The adsorption of molecules into polydopamine-functionalized carbon nanotube films causes changes in the material's polarizability, altering... and .

[0036] For the A target gas, its concentration The change in the real part of the refractive index of the corresponding sensitive thin film layer Changes in the imaginary part The relationship between them, in the low-concentration linear response region, can be expressed as: ; ; In the formula, and The first The sensitive thin film for the first layer The sensitivity coefficients for the real and imaginary parts of the refractive index of a gas are determined through a pre-calibration experiment.

[0037] Step 204: Gas response-induced shift in the resonant characteristics of micro / nano fiber optic resonator The change in the complex refractive index of the sensitive thin film is fed back to the micro / nano fiber resonator through evanescent field coupling, causing the corresponding resonant frequency of the cavity to change. The change in the resonance condition manifests specifically in two detectable characteristic parameters: the shift in the resonant wavelength. Changes in the extinction ratio of the resonance peak .

[0038] The drift in resonant wavelength is mainly due to the real part of the refractive index of the sensitive thin film. The change in is triggered, and the two satisfy the following relationship: ; In the formula, This is the center wavelength of the channel in a vacuum. For the fleeting scene and the first The overlap factor of the first thin film is defined as the overlap factor of the second thin film. The ratio of evanescent field power within the thin film region to the total evanescent field power. Due to film thickness and penetration depth Calculated using the following formula: ; in, The radial coordinates are the distances from the surface of the cone waist. For the first The interface location between the thin film and the previous layer; the denominator of the above formula is the whole-space integral of the evanescent field from the cone waist surface to infinity; evanescent field intensity. Follow Decays exponentially, with a decay constant of: By designing the thickness and refractive index distribution of each layer, it is possible to achieve... , and For their respective frequencies It has a significant value, while its value is smaller for other frequencies, further enhancing selectivity.

[0039] The change in the extinction ratio of the resonance peak is due to the imaginary part of the refractive index. Dominant, representing the additional intracavity loss caused by molecular absorption. Change in extinction ratio. and The relationship is approximated as: ; In the formula, The loss conversion factor is related to the cavity's intrinsic quality factor, coupling efficiency, and... The correlation can be determined experimentally.

[0040] Step 205: Constructing a gas-sensitive multidimensional response vector Over time, across multiple scan cycles, in Each time the channel is excited, the response speed of the micro / nano fiber resonant cavity is much faster than the change in gas concentration, allowing the measurement system to obtain the quasi-steady-state response of the gas at that moment. Therefore, for each complete gradient scan cycle... The transmitted light output from the micro / nano fiber resonator sequentially carries all of... The gas information from each channel is integrated. This transmitted light, containing multiple information streams, is used as the output of the sensing probe and will proceed to step 3 for photoelectric detection and demultiplexing.

[0041] At the signal processing level, regarding the first Each channel has a tiered design, and its resonant characteristic parameters are defined. and These elements together constitute the response element of the channel. All of them... The combination of responses from each channel can form a dimensional response vector , corresponding to the One scan cycle: ; The vector This is the observation signal vector required for concentration inversion in the subsequent step 4. The direct source. In step 3, the vector is demodulated and extracted from the aliased time-domain signal after photoelectric conversion, and then input into the concentration inversion model.

[0042] Step 3, photoelectric detection, time-domain demultiplexing, and multidimensional response vector extraction, specifically includes: Step 3 is performed by the photoelectric detection and time-domain demultiplexing unit. This unit receives the composite transmitted light signal from the output of the sensing probe and transmits it back via a single-mode transmission fiber. Simultaneously, it receives frame synchronization signals from controller 103. Its task is to alias the data in the time domain during each scan cycle. The optical response of each channel is completely converted into a digital electrical signal, and then the resonant wavelength shift of each target gas is extracted through time-domain demultiplexing and pulse parametric demodulation. and extinction ratio change Finally, it is assembled into a multidimensional response vector. As the input for concentration inversion, we will now introduce it step by step: Step 301: Photoelectric conversion and low-noise amplification of composite transmitted light signals Returning from single-mode transmission fiber The optical signal is a series of periodic, high-repetition-rate, narrow pulses, optically represented as a sequence of stepped pulses output from step 2, modulated by a micro / nano fiber optic resonator. This optical signal first enters a broadband photodetector within the photodetector and time-domain demultiplexing unit. The photodetector uses an indium gallium arsenide avalanche photodiode, which has a coverage... Sufficient bandwidth for each frequency tier and higher than The responsivity. The weak photocurrent obtained by photoelectric conversion is approximately equal to the instantaneous incident light power. They are directly proportional, and their expression is: ; in Photodetector responsivity (unit) The photocurrent is then fed into an adjacent low-noise transimpedance amplifier, which converts the current signal into a voltage signal. The gain-transimpedance ratio is... , obtain voltage signal The transimpedance amplifier internally incorporates an adjustable bandpass filter stage, with its high-pass cutoff frequency set to... This is used to suppress ambient light and low-frequency drift; the low-pass cutoff frequency is set to... To ensure that at least the width is retained without distortion. The pulse waveform. The signal is the input for subsequent analog-to-digital conversion.

[0043] Step 302: High-speed analog-to-digital conversion and frame synchronization trigger acquisition voltage signal The data is fed into a high-speed analog-to-digital converter. The resolution of the analog-to-digital converter is... Sampling rate To satisfy the contract The pulse requires precise sampling (at least 5 sampling points within one pulse cycle). The analog-to-digital converter's trigger acquisition mode is set to external trigger, with its trigger input connected to the frame synchronization signal issued by the controller. At the start of each scan cycle, The rising edge of the signal triggers the analog-to-digital converter to sample a complete frame of data. The sampling time window length is set to [value missing]. ,in To protect margin and ensure full-cycle signal capture. In the first... In one scan cycle, the analog-to-digital converter outputs a signal of length [missing value]. Discrete-time series of points This sequence completely records all of the events within that period. The composite waveform of an optical comb pulse passing through a micro / nano fiber resonant cavity.

[0044] Step 303: Time-domain gated demultiplexing and single-channel pulse signal extraction The captured full-frame sequence The pulses are fed into a field-programmable gate array (FPGA) in real time and first undergo time-domain gated demultiplexing. The nominal occurrence time of each channel pulse is pre-calibrated to obtain the pulse's nominal occurrence time. and pulse width And convert it to the number of starting points at the level of sampling points. and End Point Thus for the first Each channel defines a rectangular time window. Field-Programmable Gate Array (FPGA) pairs Apply parallel window capture to generate Subsequences: ; in Because the pulses are spaced sufficiently large in the time domain (e.g., , , Intervals reached far greater than This window capture does not introduce crosstalk. This step achieves complete demultiplexing of aliased time-division multiplexed signals into independent signals in the digital domain. Single-channel pulse signal.

[0045] Step 304: Single-pulse time-frequency domain parametric demodulation to obtain and

[0046] Single pulse sequence for each channel Two key features need to be extracted: the time of pulse peak occurrence relative to the nominal time. offset and the relative change of the pulse peak voltage relative to the reference value These characteristic quantities are related to those in step 2. and It has a defined mapping relationship and needs to be transformed through the following physical model.

[0047] (1) Transmission model and impulse response of micro / nano fiber resonator Micro / nano fiber resonator at frequency The transmittance function at a certain point can be written as: ; in The resonant center frequency, The full width at half maximum (FWHM) of the resonance peak. This represents the minimum transmittance at the resonance point (directly reflecting the extinction ratio). When the... After the sensitive layer responds to the target gas, Offset to ,in , The initial resonant frequency in the absence of gas, tuned to the sequence frequency. Alignment, that is .at the same time Become ,in and Linear correlation ( Apparent extinction ratio is defined as (This is directly related to changes in emissivity).

[0048] Optical power envelope of input sub-comb pulse The Lorentz mapping of the micro-ring download port in step 1 determines this, namely: ; This utilizes instantaneous frequency mapping. , The peak power of the incident pulse. The output power of this pulse after passing through the micro / nano fiber optic resonant cavity is the product: ; Will Substituting the values, we can see that the output pulse waveform is simultaneously subjected to two Lorentz linear modulations: one is the pulse base shape determined by the micro-loop download port, with the center at... ,width Secondly, the narrowband transmission spectrum of the micro / nano fiber optic resonator, centered in... ,width Due to the micro / nano fiber resonator... Typically much smaller (For example ,Compare The output pulse shape is mainly characterized by a "comb-like" truncation of the bottom pulse by the transmission window of the micro / nano fiber resonator, macroscopically appearing as a narrow pulse, the peak time and peak intensity of which are related to... and Extremely sensitive.

[0049] (2) Peak delay and Relationship right Perform approximate analysis: when When the value is small, the peak time of the output pulse is approximately the peak position of the product of the two Lorentz functions. This can be established by differentiation. and Linear relationship: ; because Although the sensitivity is somewhat reduced, reliable measurements are still possible. The final resonant wavelength shift can be calculated from the time delay shift: ; The coefficients were determined through preliminary calibration. That is, from the measurement Reverse push .

[0050] (3) Relative change in peak amplitude and Relationship Output pulse peak voltage Simultaneously affected and Impact. Expanding near the nominal alignment point and combining it with the pulse shape, the relative amplitude change can be obtained: ; When extracted from the latency information Then, the solution can be found. Thus obtain A pre-calibrated sensitivity matrix in a field-programmable gate array. To achieve decoupling operations: ; in The cross-compensation factor is obtained from theoretical calculations or calibration experiments.

[0051] (4) Implementation of the actual extraction algorithm In practical implementation, the field-programmable gate array pairs sub-sequences Perform the following processing: Cubic spline interpolation is used to transform discrete sampling points into continuous-time signals in order to accurately find the peak position and peak amplitude.

[0052] By using parabolic fitting to sample several points near the peak, the peak time within the sub-sampling interval is obtained. and peak amplitude .

[0053] calculate , ,in This is a reference value used to utilize background cycles or previous baseline updates.

[0054] Will and Substituting the above decoupling matrix, output the first period of this cycle. Each channel and .

[0055] Step 305: Multidimensional Response Vector Assembly and Data Interface In the All scan cycles are completed. After pulse demodulation of each channel, the digital signal processor in the photoelectric detection and time-domain demultiplexing unit collects the demodulated values ​​of all channels and assembles them into a 2N-dimensional response vector according to a predetermined order. : ; The vector is transmitted in real time to the concentration inversion unit in step 4 via a high-speed data bus (such as PCIe).

[0056] In step 4, the concentration inversion based on the nonlinear blind source separation of the improved nuclear autoencoder is specifically as follows: Step 4 is performed by the concentration inversion unit. This unit receives the real-time multidimensional response vector stream from the photoelectric detection and time-domain demultiplexing unit. (In this invention) (i.e., 6-dimensional), its task is to decouple these mixed observations contaminated by cross-sensitivity and nonlinear effects, and finally output the results for the three target gases ( )concentration Reliable estimation. This step establishes and trains an improved kernelized autoencoder nonlinear blind source separation model. This model introduces an explicit stochastic Fourier feature kernel mapping layer and regularization based on the kernel Hilbert-Schmidt independence criterion (HSIC) into the traditional autoencoder framework, overcoming the problem of insufficient separation accuracy of existing methods in multi-component cross-response and weakly nonlinear scenarios. It achieves blind separation with minimal calibration and highly robust concentration inversion. The following steps will be described in detail: Step 401: Problem Modeling and Data Preprocessing The high-dimensional response vector obtained in step 3 In this context, each element can be a nonlinear function of multiple gas concentrations. Extending this further, we can assume the existence of an unknown nonlinear mixing mapping. And an additive noise term , so that: ; Here Let be the source concentration vector to be determined. Due to the physical interactions between sensitive films (such as competitive adsorption and non-superposition of refractive indices) and the nonlinear response of the micro / nano resonant cavity, It cannot be simply decomposed into linear aliasing. Traditional blind source separation (such as independent component analysis, ICA) relies on the linear assumption, which fails in this scenario; while when ordinary autoencoders are directly used for blind separation, without appropriate statistical constraints, the correspondence between latent variables and real sources cannot be guaranteed.

[0057] Therefore, an inverse mapping model is constructed. , making Approaching (At most, there is uncertainty in scale and ranking). Model training is divided into two stages: unsupervised pre-training uses a large number of unlabeled mixed samples to learn decoupled representations, and weakly supervised calibration uses a very small number of labeled samples to correct scale and ranking.

[0058] In the preprocessing stage, the original response vector sequence is first averaged using a sliding window to suppress high-frequency noise, and then normalization is performed: ; in These are the mean vector and covariance matrix estimated based on long-term monitoring historical data, respectively. (Standardized) As network input.

[0059] Step 402: Improve the kernelized autoencoder model structure This model is generated by the encoder. and decoder Composition, with a low-dimensional latent layer between the two. Its key improvement lies in the introduction of an explicit random Fourier feature kernel mapping in the encoder input layer, thereby directly constructing an approximate regenerating kernel Hilbert space in the input space, giving the network a strong expressive ability for complex nonlinear mixtures, while maintaining the convexity advantage of the training process.

[0060] Input layer (kernel mapping layer): for normalized vectors This layer maps it to Wei (take) The kernel feature space of ) ; in Randomly sampled from a Gaussian distribution and fixed (kernel width) (This is a hyperparameter, set to the median distance of the input data). This mapping approximates a Gaussian kernel. It provides non-linear dimensionality enhancement capabilities.

[0061] Encoder body: In Two dense fully connected layers are stacked on top: ; ; Finally, the latent vector is output through a linear transformation: ; Decoder: Mirror-symmetric structure, from Refactoring Two hidden layers, the output layer uses identity activation, and the dimension is reduced to [value missing]. : ; The set of network weights is denoted as This structure extends the traditional autoencoder to the kernel ridge regression framework through a kernel mapping layer, enhancing its ability to capture patterns caused by nonlinear distortions due to cross-sensitivity.

[0062] Step 403: Training Strategy and Independence Reinforcement Regularization To make latent variables Each component accurately corresponds to its respective gas concentration and is statistically independent of each other—that is, nonlinear blind source separation is achieved. Based on the reconstruction loss, we introduce the kernel Hilbert-Schmidt independence criterion as a decoupling regularizer.

[0063] Reconstruction loss: Mean squared error is used to ensure... Carry sufficient input information: ; HSIC Independence Constraint: For a batch of latent variable sample matrices We compute the HSIC unbiased estimate between any two latent components. The and the first Taking one component as an example, define two kernel matrices. , element is ,in Select a scale-adjusted Gaussian kernel. The standardized HSIC value can then be estimated using the following formula: ; The total independence penalty is the sum of all non-repeating pairs: ; Minimizing this condition forces different latent variables to be uncorrelated in the reproducing kernel Hilbert space, thereby ensuring statistical independence under the Gaussian kernel.

[0064] Non-negativity and amplitude constraints: due to gas concentration We The output is activated using ReLU to embed the prior; at the same time, to avoid excessive scaling, a very small amount is added. Norm encourages signal to maintain amplitude: ; In practice, the encoder output is obtained by passing it through ReLU. Then the decoder can reconstruct the data, without explicitly writing it into the loss statement. However, it can be used as an aid in the early stages of training.

[0065] Total loss function: ; in To balance the hyperparameters, they are set according to the trade-off curve between reconstruction and independence.

[0066] Training process: A two-stage strategy is adopted.

[0067] Unsupervised blind pre-training: Collecting a large amount of data generated during long-term operation in step 3 (Without concentration labels) constitute the dataset Using mini-batch gradient descent, batch size... The optimizer is Initial learning rate Iterate for 200 epochs. During this phase, The weights are linearly increased from 0 to the target value to avoid early latent variable collapse. After pre-training, the latent variables... The components have shown a good trend of statistical independence, but the order and absolute scale have not yet been determined.

[0068] Weakly supervised calibration: Only a few sets (e.g., 30 sets) of known concentrations are needed. calibration point ( (Fixed encoder parameters, in) Then add a linear correction layer Using this labeled data, the optimal correspondence between channels can be solved simultaneously. (Hungarian algorithm matching) and regression coefficients After calibration, the concentration can be directly output. This process has minimal computational overhead and requires very little standard gas, greatly reducing actual deployment costs.

[0069] Step 404: Online Concentration Inversion Calculation Process After deployment, for each scanning cycle Perform the following steps: Get After undergoing preprocessing consistent with training, the following results were obtained. .

[0070] Calculate latent variables via encoder forward propagation (encoder portion only): ; Apply calibration mapping: ; Output the estimated concentrations of the three gases at the current time. .

[0071] The entire step 4, through this improved nuclear autoencoder nonlinear blind source separation model, achieves accurate and stable inversion from multidimensional response vectors to multi-component concentrations under complex cross-sensitivity and nonlinear backgrounds, and is seamlessly connected with the real-time response vector flow of the preceding step 3, forming a complete measurement-demodulation-inversion signal processing link.

[0072] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0073] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0074] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0075] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0076] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A multi-component gas precision detection system for an energy storage device, characterized in that, include: The system includes a multi-wavelength gradient optical comb generation unit, a sensing probe, a photoelectric detection and time-domain demultiplexing unit, and a concentration inversion unit. The multi-wavelength gradient optical comb generation unit is connected to the optical input terminal of the sensing probe, the optical output terminal of the sensing probe is connected to the input terminal of the photoelectric detection and time-domain demultiplexing unit, and the output terminal of the photoelectric detection and time-domain demultiplexing unit is connected to the concentration inversion unit. The multi-wavelength gradient optical comb generation unit is used to generate a set of optical pulse sequences that are arranged sequentially in the time domain and do not overlap with each other in time. The optical pulse sequence contains at least two types of optical pulses with different center wavelengths, and each center wavelength is aligned with the optimal response wavelength of the corresponding sensitive area in the sensing probe. The sensing probe includes an optical fiber resonant cavity, on which at least two sensitive regions are respectively arranged at different positions along the light propagation direction. Each sensitive region has selective response capability to different analytes. The optical pulse sequence enters the optical fiber resonant cavity and interacts with each sensitive region in the form of an evanescent wave. The resonant characteristics of each optical pulse change due to the change in the concentration of the corresponding analyte, thereby outputting a transmitted light signal carrying information about each analyte. The photoelectric detection and time-domain demultiplexing unit is used to convert the transmitted light signal into an electrical signal and perform analog-to-digital conversion. It uses the time-domain position information of each light pulse to perform channel separation on the converted digital signal, extracts the peak characteristics of each pulse subsequence corresponding to different center wavelengths, and calculates the resonant wavelength shift and extinction ratio change caused by each component under test based on the peak characteristics. Then, it combines the resonant wavelength shift and extinction ratio change into a multidimensional response vector. The concentration inversion unit is equipped with a pre-trained nonlinear blind source separation model, which is configured to receive the multidimensional response vector and output real-time estimates of the concentrations of each analyte through decoupling mapping.

2. The multi-component gas precision detection system for an energy storage device according to claim 1, characterized in that, The multi-wavelength cascade optical comb generation unit includes a femtosecond fiber laser, a dispersive fiber, an optical amplifier, and a microring resonator array connected in sequence. The microring resonator array is composed of multiple microring resonator cavities with different free spectral ranges, which are used to slice the dispersion-broadened broadband chirped pulses into multiple narrowband pulses and output the optical pulse sequence in the time domain.

3. The multi-component gas precision detection system for an energy storage device according to claim 2, characterized in that, The fiber optic resonant cavity of the sensing probe is a micro-nano fiber optic resonant cavity. The micro-nano fiber optic resonant cavity includes a conical waist region with a waist diameter on the order of μm. Bragg gratings or fiber optic ring mirrors are provided at both ends of the conical waist region to form a resonant cavity. At least two sensitive regions are sensitive thin film arrays integrated at different longitudinal positions along the surface of the conical waist region by self-assembly or inkjet printing.

4. The multi-component gas precision detection system for an energy storage device according to claim 3, characterized in that, The sensitive thin film array has selective adsorption and response capabilities for hydrogen, carbon monoxide and ethylene, respectively, and the sensitive films are spatially separated from each other, so that the resonant wavelength shift and extinction ratio change caused by different analytes are assigned to the corresponding light pulses in the time domain.

5. The multi-component gas precision detection system for an energy storage device according to claim 4, characterized in that, The photoelectric detection and time-domain demultiplexing unit includes a photodetector, a transimpedance amplifier, an analog-to-digital converter, and a digital signal processor connected in sequence. The analog-to-digital converter samples under the control of the frame synchronization signal. The digital signal processor has time-domain gated demultiplexing logic to generate a rectangular time window corresponding to the nominal position and width of each optical pulse, and to extract each pulse subsequence from the full frame sampled data.

6. The multi-component gas precision detection system for an energy storage device according to claim 5, characterized in that, The digital signal processor is configured to extract the peak time and peak amplitude of each pulse subsequence with subsampling precision, and to convert the extracted peak time and peak amplitude into the resonant wavelength offset and extinction ratio change, respectively, using a pre-calibrated sensitivity matrix.

7. The multi-component gas precision detection system for an energy storage device according to claim 6, characterized in that, The nonlinear blind source separation model in the concentration inversion unit is an improved kernelized autoencoder network. This kernelized autoencoder network includes a random Fourier feature kernel mapping layer, at least two fully connected hidden layers and a linear output layer, and is deployed after being trained in two stages: unsupervised pre-training and weakly supervised calibration.

8. The multi-component gas precision detection system for an energy storage device according to claim 7, characterized in that, The unsupervised pre-training stage utilizes unlabeled multidimensional response vector data to train the system with reconstruction loss and Hilbert-Schmidt independence criterion regularization term as optimization objectives, making the latent variable components tend to be statistically independent. The weakly supervised calibration stage uses standard gas samples to determine the correspondence between each latent variable component and the concentration of the analyte and the linear correction coefficient.

9. A multi-component gas precision detection system for an energy storage device according to claim 8, characterized in that, The device also includes a controller, which is electrically connected to the multi-wavelength cascade optical comb generation unit, the sensing probe, and the photoelectric detection and time-domain demultiplexing unit, respectively, for generating and distributing frame synchronization signals to synchronize the generation of the optical pulse sequence, the temperature control of the optical fiber resonant cavity, and the sampling triggering of the analog-to-digital converter.

Images