A fire-preventing method for an electric bicycle charging pile

Abstract

This invention provides an active fire prevention method for electric bicycle charging piles, relating to the fields of machine learning and charging device technology. The method extracts underlying electrical features to obtain the distortion degree of the charging waveform and infers the attenuation status indicator of potential hazards based on a health prediction network. Simultaneously, it acquires low-frequency fluctuations from adjacent nodes and local high-frequency ripple, performs spatial differential and frequency domain mapping, extracts the spatial common-mode deviation, and differentially compandspans the high-frequency ripple to generate an electrochemical state compensation coefficient. Based on this, it dynamically adjusts the fusion ratio of distortion degree and attenuation indicator, and calculates an adaptive safety judgment threshold. When the feature offset exceeds the limit, physical isolation is triggered; when the common-mode deviation diverges, a derating conservative power-off protection strategy is triggered. Through multi-dimensional feature temporal state alignment and dynamic threshold judgment, it effectively overcomes the concealment of cell thermal runaway precursors and false positives under strong disturbances, achieving a closed-loop blocking mechanism in heterogeneous charging environments.

Description

Technical Field

[0001] This invention relates to the field of machine learning and charging device technology, specifically to an active protection method for preventing fires in electric bicycle charging stations. Background Technology

[0002] With the expansion of edge nodes in new power systems, electric bicycle charging networks are gradually evolving towards high-density, complex outdoor operating conditions. The deep integration of charging devices and machine learning has become an industry trend for improving the safety management of charging networks. In centralized charging scenarios where deep battery management system (BMS) communication is generally lacking, the traditional paradigm of relying on macroscopic electrical characteristics for safety status assessment is facing severe challenges. Existing technologies still have significant limitations in addressing the concealment and hysteresis of cell thermal runaway precursors. For example, patent application CN121200843A proposes a charging pile protection method based on dynamic temperature estimation, which generates a dynamic threshold for temperature rise rate by matching current charging data with a vehicle model identification database.

[0003] Although the scheme introduces a dynamic mechanism, it mainly faces the following problems: the electrical distortion characteristics caused by the micro-internal short circuit in the early stage of cell thermal runaway are extremely weak. Under outdoor conditions, they are easily masked by global common-mode interference such as the diurnal load fluctuations of the power grid and the switching noise of adjacent equipment. Existing technologies lack spatiotemporal decoupling methods for global background noise and micro-polarization characteristics, making it difficult to accurately extract weak fire warning signals. Furthermore, due to the uneven health status of heterogeneous batteries, if the protection threshold is set too high, the hidden danger signal will directly penetrate the defense line. If the threshold is set too low, it is easy to cause defense line oscillation and false positives when the external power grid fluctuates. Existing technologies are unable to achieve an adaptive balance between sensitivity and robustness under strong disturbance environments. Summary of the Invention

[0004] The purpose of this invention is to provide an active fire prevention method for electric bicycle charging stations. This method uses real-time waveform distortion, health status derived from historical cycles, global low-frequency grid fluctuations, and local high-frequency ripple response as collaborative input sources. By constructing spatial difference and frequency domain mapping logic, it actively removes global common-mode noise, generates compensation coefficients characterizing the battery's microscopic electrochemical polarization, and then performs weight injection and dynamic allocation on the business characteristics of each dimension. This addresses the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An active fire prevention method for electric bicycle charging stations includes the following steps: S1: Obtain the underlying real-time electrical environment feature stream and perform local sequence cleaning operations to extract the charging waveform distortion degree; input the historical charging cycle data stream into the preset health prediction network to deduce the hidden danger decay status indicator; S2: Obtain the background grid fluctuation sequence of adjacent nodes representing adjacent non-local charging nodes, and obtain the high-frequency ripple sequence of local charging circuit representing the underlying power supply circuit of local charging port. S3: Using the background power grid fluctuation sequence of the adjacent nodes and the high-frequency ripple sequence of the local charging circuit as input, execute spatial differential and frequency domain mapping logic, and perform differential companding operation on the high-frequency ripple sequence of the local charging circuit based on the global background noise deviation of the transformer area extracted from the background power grid fluctuation sequence of the adjacent nodes, so as to generate electrochemical state compensation coefficients for characterizing the electrochemical microphysical state. S4: Based on the electrochemical state compensation coefficient, the original multi-dimensional feature joint calculation logic is weighted and compensated. The fusion ratio of the charging waveform distortion degree and the hidden danger attenuation state indicator is dynamically adjusted according to the electrochemical state compensation coefficient to generate an adaptive safety judgment threshold. S5: Perform high-frequency joint comparison and verification. If it is determined that the distortion of the charging waveform extracted in real time exceeds the adaptive safety judgment threshold, generate a collaborative control command to instruct the terminal node on the power supply circuit side to perform physical power flow isolation action, and trigger the derating conservative power-off protection strategy when the global background noise deviation of the transformer area exceeds the preset divergence boundary.

[0006] Compared with the prior art, the beneficial effects of the present invention are: The core approach involves acquiring low-frequency waveforms from adjacent nodes to extract the global background noise deviation of the distribution area, and performing differential companding on the local high-frequency switching ripple. Utilizing multi-dimensional feature temporal state alignment and spatial differential mechanisms, the aliased global background grid harmonic interference in the local high-frequency ripple is inversely canceled. This method effectively overcomes the technical obstacle of environmental noise masking the precursors of thermal runaway, enhances the robustness of low-level feature extraction, and achieves high-purity spatiotemporal decoupling analysis of the battery's micro-polarization state.

[0007] This invention employs a core approach: extracting the distortion degree of the charging waveform and the indicator of potential degradation status, and dynamically adjusting the fusion ratio of the two based on the generated electrochemical state compensation coefficient. Using the compensation coefficient, which characterizes the real-time microscopic physical state, as a penalty or compensation factor, joint calculations are performed on historical health degradation prior data and real-time waveform distortion posterior data to reconstruct the adaptive safety judgment threshold. This ensures that the system maintains high sensitivity to weak distortions in the early stages of thermal runaway under heterogeneous charging conditions, while effectively avoiding line-of-sight oscillations and misjudgment failures caused by external power grid fluctuations.

[0008] By performing high-frequency joint comparison and verification, a physical isolation command is issued when the critical value is exceeded, and a system-level degradation protection mechanism is triggered when the spatial common-mode variance crosses the divergence boundary. A dual adaptive defense closed loop covering "interception" and "degradation" is established. When extreme electromagnetic pollution causes the high-frequency detection chain to fail, this method can proactively cut off the compensation weights of erroneous features and reconstruct the conservative defense. This mechanism effectively addresses the uncertainty risks in complex community centralized charging environments, providing high operational continuity and system-level electrical safety safeguards for active fire protection systems. Attached Figure Description

[0009] Figure 1 This is a technical roadmap for an active fire prevention method for electric bicycle charging stations.

[0010] Figure 2 This is an architecture diagram of an active fire prevention method for electric bicycle charging stations.

[0011] Figure 3 This is a flowchart illustrating the technical principle of the present invention.

[0012] Figure 4 This is a schematic diagram illustrating the numerical verification of the adaptive safety determination threshold based on the electrochemical state compensation coefficient. Detailed Implementation

[0013] 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.

[0014] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various elements, but unless otherwise stated, these elements are not limited by these terms. These terms are used only to distinguish one element from another.

[0015] Example 1: Please see Figures 1 to 4 The present invention provides a technical solution: An active fire prevention method for electric bicycle charging stations includes the following steps: S1: Obtain the underlying real-time electrical environment feature stream and perform local sequence cleaning operations to extract the charging waveform distortion degree; input the historical charging cycle data stream into the preset health prediction network to deduce the hazard decay status indicator.

[0016] S2: Obtain the background grid fluctuation sequence of adjacent nodes representing adjacent non-local charging nodes, and obtain the high-frequency ripple sequence of local charging circuit representing the underlying power supply circuit of local charging port.

[0017] S3: Using the background power grid fluctuation sequence of the adjacent nodes and the high-frequency ripple sequence of the local charging circuit as input, perform spatial differential and frequency domain mapping logic, and perform differential companding operation on the high-frequency ripple sequence of the local charging circuit based on the global background noise deviation of the transformer area extracted from the background power grid fluctuation sequence of the adjacent nodes, so as to generate electrochemical state compensation coefficients for characterizing the electrochemical microphysical state.

[0018] S4: Based on the electrochemical state compensation coefficient, the original multi-dimensional feature joint calculation logic is weighted and compensated. The fusion ratio of the charging waveform distortion degree and the hidden danger attenuation state indicator is dynamically adjusted according to the electrochemical state compensation coefficient to generate an adaptive safety judgment threshold.

[0019] S5: Perform high-frequency joint comparison and verification. If it is determined that the distortion of the charging waveform extracted in real time exceeds the adaptive safety judgment threshold, generate a collaborative control command to instruct the terminal node on the power supply circuit side to perform physical power flow isolation action, and trigger the derating conservative power-off protection strategy when the global background noise deviation of the transformer area exceeds the preset divergence boundary.

[0020] Further, the underlying real-time electrical environment feature stream is acquired, and a local sequence cleaning operation is performed to extract the charging waveform distortion, including: Extract the real-time charging voltage and current parameters from the underlying real-time electrical environment feature stream; The real-time charging voltage and current parameters are extracted sequentially using a sliding time window, and the information entropy of the nonlinear differential sequence within the sliding time window is calculated. The information entropy of the nonlinear difference sequence after removing the macro trend term is determined as the distortion degree of the charging waveform; Historical charging cycle data streams are input into a pre-set health prediction network to deduce the status indicators of potential hazard decay, including: The historical charge-discharge cycle parameters of the associated business identifier are mapped to a time-series decay vector, and the potential decay status identifier, which characterizes the internal aging tendency of the battery, is output based on the health prediction network.

[0021] Further, the acquisition of the adjacent node background grid fluctuation sequence characterizing the adjacent non-local charging node includes: receiving the bus electrical parameter sequence of other charging nodes in the same distribution area that are in a physically unloaded or constant current operating state; Obtain the high-frequency ripple sequence of the local charging circuit, which characterizes the underlying power supply circuit of the local charging port, including: Obtain the intrinsic AC current ripple response sequence excited by the inherent pulse width modulation switching frequency of the AC / DC power conversion stage and superimposed on the output power supply network.

[0022] Further specifying, using the background grid fluctuation sequence of the adjacent nodes and the high-frequency ripple sequence of the local charging circuit as input, spatial difference and frequency domain mapping logic is executed to generate electrochemical state compensation coefficients, including: The temporal energy envelope features of the background power grid fluctuation sequence of the adjacent nodes are extracted, and the global background noise deviation of the transformer area, which characterizes the intensity of global environmental interference, is generated through an asynchronous alignment mechanism. Using the global background noise deviation of the transformer area as a dynamic constraint, differential companding is performed on the gain parameter of the high-frequency ripple sequence of the local charging circuit to cancel the background grid harmonic interference mixed in the high-frequency ripple sequence of the local charging circuit, thereby obtaining a high-purity differential mode ripple sequence. Perform discrete frequency domain phase angle demodulation operation to extract the frequency domain phase angle offset features between the high-purity differential mode ripple sequence and the original excitation pulse width modulation signal; The anti-interference warning sensitivity deviation is calculated in real time, and the frequency domain phase angle offset feature is used to update the anti-interference warning sensitivity deviation. The electrochemical state compensation coefficient is output based on the convergence state of the anti-interference warning sensitivity deviation. The anti-interference warning sensitivity deviation is configured as a detection resolution gain index characterizing the fused purified waveform for weak electrochemical polarization impedance.

[0023] Further defining the method, the fusion ratio between the charging waveform distortion and the hazard attenuation status indicator is dynamically adjusted based on the electrochemical state compensation coefficient to generate an adaptive safety judgment threshold, including: The electrochemical state compensation coefficient is used as a multiplication penalty factor and injected into the environmental parameter weighting equation in the multidimensional feature joint calculation logic. In response to the impedance degradation trend indicated by the electrochemical state compensation coefficient, the compensation margin weight of the environmental parameter is reduced proportionally, and the basic risk weight of the hazard attenuation state indicator is simultaneously amplified. The adaptive safety judgment threshold value after narrowing the boundary is calculated and output. Further defining the scope, a coordinated control command is generated to instruct the terminal nodes on the power supply circuit side to perform physical power flow isolation actions, and a derating conservative power-off protection strategy is triggered when the global background noise deviation of the transformer area exceeds a preset divergence boundary, including: When the coordinated control command is issued, the first control branch generates a pulse blocking signal to drive the physical contactor to disconnect the power flow, and the second logic branch synchronously captures the distortion characteristics of the charging waveform at the moment of breaking the critical value and resets the long-term baseline parameters of the health prediction network. When the statistical variance of the global background noise deviation of the substation area continuously exceeds the preset divergence boundary, the output weight of the electrochemical state compensation coefficient is forcibly set to zero, and the adaptive safety judgment threshold is reconstructed into a single-node conservative defense limit value that is completely dependent on the local feature cleaning data.

[0024] Further defining the update of the anti-interference warning sensitivity deviation using the frequency domain phase angle offset feature includes: Extract the frequency domain phase angle offset features within the current operation cycle, and obtain the smoothed reference phase angle feature quantity within the historical time window; Calculate the time difference gradient value between the frequency domain phase offset feature and the reference phase feature; The time difference gradient value is mapped to a dynamic update step size parameter between zero and one, wherein the dynamic update step size parameter has a non-linear positive correlation with the time difference gradient value. The current weight of the frequency domain phase offset feature is calculated using the dynamic update step size parameter, and the historical forgetting weight is calculated using the difference between the numerical value and the dynamic update step size parameter. The first product is obtained by multiplying the current weight by the frequency domain phase offset feature, and the second product is obtained by multiplying the historical forgetting weight by the anti-interference warning sensitivity deviation of the previous cycle. The first product is added to the second product to iteratively update the anti-interference warning sensitivity deviation in the current operation cycle.

[0025] Further specifying, the electrochemical state compensation coefficient is used as a multiplication penalty factor, injected into the environmental parameter weighting equation in the multidimensional feature joint calculation logic, and the compensation margin weight of the environmental parameters is reduced proportionally, while the basic risk weight of the hazard attenuation state indicator is simultaneously amplified, including: The electrochemical state compensation coefficient is input into a preset nonlinear smoothing mapping logic, and the output is a penalty multiplier between the lower limit value of zero and the upper limit value of one. Obtain the initially allocated compensation margin weight in the environmental parameter weighting equation, and multiply the compensation margin weight by the penalty multiplier to obtain the stripping weight attenuation amount; Subtract the stripping weight attenuation from the compensation margin weight to obtain the actual environmental compensation weight after shrinkage; The initial allocation of the basic risk weight is obtained, and the stripping weight decay is added to the basic risk weight as a zero-sum transfer compensation term to obtain the expanded actual risk decision weight; wherein, the sum of the values ​​of the actual environmental compensation weight and the actual risk decision weight is equal to the sum of the compensation margin weight and the initial allocation value of the basic risk weight. A normalized parameter fusion matrix is ​​constructed using the actual environmental compensation weight and the actual risk decision weight to generate the adaptive safety judgment threshold.

[0026] Further defining the frequency domain phase shift characteristic, it characterizes the phase hysteresis of the high-frequency pure differential mode ripple sequence relative to the original excitation signal in the frequency domain, reflecting the dynamic increase in the microscopic electrochemical polarization impedance within the battery cell. This characteristic is obtained by extracting the high-frequency ripple in the underlying power supply circuit and performing discrete frequency domain phase angle demodulation. In detecting the microscopic physical state of battery cells, traditional DC impedance testing requires interrupting the charging process and applying external excitation equipment, which cannot meet the needs of real-time active protection. This embodiment utilizes the AC / DC power conversion circuit built into the charging pile as a free "high-frequency excitation probe." By analyzing the response delay of high-frequency current ripple, it achieves non-destructive, online targeted mapping of the precursors to thermal runaway characterized by increased internal polarization impedance of the battery. When performing this feature extraction, the underlying power conversion hardware environment it relies on meets the following attributes: the power supply circuit has a pulse width modulation switching frequency representing the PWM excitation source, its frequency drift rate does not exceed pm0.5% of the nominal value, and the control base has the synchronous trigger interrupt pin of the PWM control logic open to provide a precise phase start reference zero point.

[0027] The frequency domain phase shift characteristic is defined as the difference in phase angle between the actual ripple response and the ideal excitation, calculated as follows: ;in Frequency domain phase shift characteristics to characterize the degree of polarization hysteresis; The pulse width modulation switching frequency is inherent in the power conversion stage. In this embodiment, its preferred calibration range is [20kHz, 50kHz]. This frequency band is located in the sensitive boundary region in the battery electrochemical impedance spectrum that characterizes the charge transfer impedance and the solid electrolyte interface membrane impedance. The phase angle of the high-purity differential mode ripple obtained at the target frequency; The reference starting phase angle of the original excitation pulse width modulation signal is mapped and calibrated as the reference starting phase zero point in the local hardware synchronization logic.

[0028] The frequency domain phase offset characteristic calculation logic is represented as follows: the rising edge trigger timestamp of the original excitation pulse width modulation signal is obtained through the synchronous interrupt interface, and it is mapped and locked as the reference starting phase zero point; a high-purity differential mode ripple sequence within a fixed time window is extracted, and a preset Hanning window weighted multiplication operation is performed on the sequence to suppress the spectral leakage phenomenon in the subsequent frequency domain transformation process; a Fast Fourier Transform (FFT) operation is performed on the windowed sequence to generate a complex spectral matrix; in a preferred embodiment, the Fast Fourier Transform (FFT) operation is manifested as follows: the local microprocessor calls its internally integrated digital signal processing (DSP) hardware acceleration core, and uses the windowed discrete time sequence stored in the memory space as the physical input source, and sends it to the multiply-accumulate unit logic array of the DSP hardware acceleration core. The DSP hardware acceleration core performs hardware-level multiplication, addition, and level shifting operations on the physical input source based on radix-2 time decimation, using a rotation factor constant table pre-stored in read-only memory (ROM) according to the clock cycle. The DSP hardware acceleration core writes the high and low level state groups representing the complex result of the operation into a block overlay into a continuous memory address space. The set of level data stored in this continuous memory address space forms a complex spectrum matrix.

[0029] Based on the pulse width modulation switching frequency, the system performs calculations to locate the corresponding target frequency point. The mathematical positioning formula is defined as follows: ;in This is the one-dimensional integer index value of the target frequency point location to be calculated in the complex spectrum matrix; The obtained pulse width modulation switching frequency; The total number of transformation points configured for the preceding Fast Fourier Transform (preferably 1024 in this embodiment); The preset global sampling frequency parameter (preferred in this embodiment is 100kHz) is used when performing digital quantization on the underlying ADC channel. To perform the rounding mathematical operator, during the dynamic calculation flow, the pulse width modulation switching frequency and the total number of transformation points are obtained. These are multiplied to obtain an intermediate frequency product. The intermediate frequency product is then divided by the global sampling frequency parameter to obtain a floating-point mapping value. This floating-point mapping value is rounded to the nearest integer, and the output integer value is used as the index of the target frequency position. Based on this index, target complex data is extracted from the complex spectrum matrix. For the complex data within the target frequency position, an arctangent operation is performed to extract its imaginary and real parts, generating the high-purity differential mode ripple phase angle at that frequency. A subtraction operation is then performed to subtract the reference starting phase zero from the high-purity differential mode ripple phase angle, and the output absolute difference is determined as the frequency domain phase shift characteristic within the current calculation cycle.

[0030] The time-difference gradient value characterizes the abrupt change rate within the phase angle shift feature time window in the frequency domain, and is used to capture the nonlinear polarization deterioration trend in the early stages of thermal runaway. This is achieved by retrieving the frequency domain phase angle shift feature of the current computation cycle. The difference is obtained by subtracting the smoothed baseline phase angle feature value within the historical time window, and the absolute value of this difference is used as the gradient value.

[0031] Specific implementation instructions for time difference gradient values: In this embodiment, high-frequency phase angle characteristics are easily affected by transient interference from charging cable swaying or slight fluctuations in plug contact resistance under actual operating conditions. Directly using instantaneous phase angle offset characteristics for comparison can easily trigger frequent malfunctions of the protection system. By introducing a time-series dimension, the acceleration of its evolution trend is extracted. The evolution trend of the time difference gradient value is defined as follows:

[0032] in This represents the time difference gradient value; The frequency domain phase shift feature extracted for the current operation cycle; is the phase angle feature in the j-th sampling period of the historical cache queue; M is the smoothing depth parameter of the historical time window, which is preferably calibrated to [10, 20] sampling periods in this embodiment.

[0033] The time difference gradient calculation logic is as follows: retrieve the newly generated frequency domain phase angle offset feature in the current operation cycle and enter it into the operation register; extract the historical phase angle feature values ​​from the local memory queue for multiple consecutive cycles, and perform an arithmetic mean calculation on all values ​​in the queue to generate a smoothed reference phase angle feature value; perform a subtraction operation, subtracting the reference phase angle feature value from the frequency domain phase angle offset feature of the current operation cycle to obtain the trend offset; perform an absolute value extraction operation on the trend offset, and determine the abrupt change amplitude after removing the positive and negative sign attributes as the final time difference gradient value, which serves as the direct basis for subsequent sensitivity adjustment.

[0034] The dynamically updated step size parameter is a dynamically adjustable coefficient between a lower limit of 0 and an upper limit of 1, which determines the system's "confidence level" in accepting newly emerging electrical distortion features. By inputting the aforementioned time difference gradient value into a preset nonlinear positive correlation activation function, when the gradient value surges, the output step size parameter approaches the upper limit of 1.

[0035] The specific implementation details for dynamically updating the step size parameter are as follows: To remain "unresponsive" to stable data but instantly "alert" to sudden anomalies, this step converts the time difference gradient value, which has physical dimensions, into a logical control ratio. Specifically, an exponential empirical mapping mechanism is employed to map heterogeneous physical quantities to dimensionless numerical values, achieving adaptive nonlinear scheduling of the system's learning rate. The internal mapping function for dynamically updating the step size parameter is defined as follows: ;in To dynamically update the step size parameter; The input is the time difference gradient value; In this embodiment, the sensitivity smoothing constant, which characterizes the radicality of the system, is preferably quantized to a value between [0.2, 0.5].

[0036] The calculation logic for dynamically updating the step size parameter is described as follows: Obtain the time difference gradient value with phase angle difference physical dimensions output from the previous step, and retrieve the preset sensitivity smoothing constant. The sensitivity smoothing constant is configured as a dimension conversion coefficient, its function being to eliminate the physical dimensions attached to the time difference gradient value, forcibly mapping heterogeneous physical quantities to a unified dimensionless pure value, ensuring the absolute validity of subsequent nonlinear exponential mapping operations in mathematical and physical logic. Perform a multiplication operation on the dimensional time difference gradient value and the sensitivity smoothing constant as the dimension conversion coefficient. After canceling the dimensions through coefficient conversion, take the negative of the product result to obtain the dimensionless intermediate negative power pure value. Using the natural constant as the base and the aforementioned dimensionless intermediate negative power pure value as the exponent, perform a power operation to obtain the dynamic convergence factor, and determine the difference between the output values ​​located in the absolute interval between 0 and 1 as the dynamically updated step size parameter.

[0037] The penalty multiplier is a core adjustment factor used to unify multidimensional physical parameters to a standard dimensionless interval, achieving dimensional consistency. By obtaining the generated electrochemical state compensation coefficient, it is input into the sigmoid nonlinear smoothing mapping logic to force the physical compensation value of the infinite domain to be mapped to the [0,1] interval.

[0038] The weight decay, in a zero-sum game mechanism, represents the absolute value of the weight that needs to be removed from the external environment tolerance and transferred to the internal risk decision-making. It is obtained by multiplying the initially allocated environmental parameter compensation margin weight in the basic business logic with the aforementioned penalty multiplier.

[0039] Joint implementation description of the penalty multiplier: This extension introduces a Sigmoid normalization mapping mechanism as the transformation logic, aiming to completely eliminate the dimensions of electrochemical physics and ensure that all parameters involved in subsequent defense reconstruction are calculated within the dimensionless domain of [0,1]. The mapping formula for the penalty multiplier and the formula for calculating the attenuation are as follows:

[0040]

[0041] Where P is a dimensionless penalty multiplier; The input is the electrochemical state compensation coefficient; is the preset threshold for the inflection point of electrochemical deterioration center (preferably set empirically to a characteristic value that deviates from the normal internal resistance baseline by 15%); k is the steepness coefficient of the mapping curve; This represents the final weight decay amount. Compensation margin weights for initially assigned environmental parameters.

[0042] In this embodiment, the kurtosis coefficient is defined as a dimensionless constant within the interval [8, 15]. If the value is below this lower limit, the mapping curve will tend to be flat, causing the system to react sluggishly to polarization degradation in the critical state and unable to output the penalty multiplier in a timely and sufficient manner; if the value is above this upper limit, the curve will degenerate into an extremely narrow step response similar to the hard-switching characteristics of a diode, which is prone to causing violent oscillations in the system's defense weights when there are slight environmental fluctuations. This preferred interval ensures that a smooth and gradual penalty transfer path with a sufficient warning gradient can be provided in the early stages of thermal runaway.

[0043] The calculation logic of the penalty multiplier is as follows: Obtain the generated electrochemical state compensation coefficient and subtract the preset deterioration center inflection point threshold to obtain the polarization overscaling; multiply the polarization overscaling with the preset steepness coefficient and take the opposite, then use it as an exponent in the power operation with the natural constant as the base; then add the result of the power operation to the value to obtain the denominator; divide the value by the above denominator to force the infinitely divergent electrochemical physical quantity to compress and expand and output it as a pure value between the lower limit 0 and the upper limit 1, which is established as the penalty multiplier; retrieve the compensation margin weight of the initially allocated environmental parameters from the system read-only memory area; multiply the initial compensation margin weight and the penalty multiplier, and the output absolute product value is determined as the stripping weight attenuation; deduct the attenuation as a subtrahend from the original environmental weight and simultaneously add it as an addend to the hidden danger basic risk weight, completing the zero-sum reconstruction of the multidimensional parameter weights with absolute closed loop and dimension conservation.

[0044] For step S1, the underlying real-time electrical environment feature stream is acquired, and a local sequence cleaning operation is performed to extract the charging waveform distortion degree; the historical charging cycle data stream is input into a pre-set health prediction network to deduce the hazard attenuation status indicator. The specific implementation logic includes: For the extracted low-level real-time electrical environment feature stream, the local microprocessor uses a sliding time window to extract the real-time charging voltage and current parameters sequentially, and calculates the information entropy of the nonlinear differential sequence within the sliding time window. In this embodiment, to eliminate the interference of gradual fluctuations in the overall grid voltage on the micro-waveform analysis, a first-order differential or moving average filtering algorithm is invoked to perform baseline drift filtering. The information entropy of the nonlinear differential sequence after removing macro-trend terms is determined as the charging waveform distortion degree, ensuring that the extracted features possess absolute micro-purity.

[0045] The system extracts real-time charging voltage and current parameters from the underlying real-time electrical environment feature stream. This extraction is specifically performed by a local microprocessor (MCU) configured within the charging pile, coupled to an analog-to-digital converter (ADC) peripheral module connected to the power supply circuit. The local microprocessor sends a high-frequency sampling trigger level command to the ADC peripheral module, driving it to digitally quantize the analog electrical signals in the physical power supply circuit according to a set sampling period. The quantized discrete digital quantities are then stored in a pre-allocated direct memory access (DMA) cache via a serial bus, thus forming real-time charging voltage and current parameters within the storage medium. The set sampling period is defined as the time interval between two level capture operations performed by the ADC peripheral module. In this embodiment, the preferred value is 0.01 milliseconds (corresponding to a physical global sampling frequency of 100kHz to match the sampling requirements of the subsequent digital signal processing hardware acceleration core). In actual engineering adaptation, the preferred range is configured within the [0.005, 0.02] millisecond interval.

[0046] The real-time charging voltage and current parameters are sequentially extracted using a sliding time window with a preset step size. For the data points within this sliding time window, the nonlinear differential sequence information entropy, which reflects the degree of system disorder, is calculated. A high-pass filter is used to remove macroscopic trend terms (such as the voltage trend that naturally increases with the charging depth) from the nonlinear differential sequence information entropy. The pure entropy value after removal is determined as the charging waveform distortion degree. The historical charge and discharge cycle parameters (including the historical highest temperature and the maximum charge and discharge rate) associated with the current business identifier are retrieved. The above parameter set is mapped into a time-series decay vector, which is input into the health prediction network. The output is a hidden decay status identifier that characterizes the internal aging tendency of the battery.

[0047] Regarding the specific implementation of charging waveform distortion, traditional charging piles rely solely on passive overload or short-circuit protection, which is insufficient to capture the subtle distortions in the early stages of thermal runaway in complex residential areas. Simply setting static voltage or current amplitude thresholds is prone to failure and is easily penetrated by grid noise. This method extracts dynamic parameters characterizing the degree of AC / DC power conversion disorder from the nonlinear feature dimension of the waveform sequence. During physical data acquisition, the sensing hardware layer upon which the real-time electrical environment characteristic flow depends is acquired, satisfying the following attributes: the acquisition link is configured with a signal acquisition medium having a high-frequency sampling rate of no less than 10kHz and a full-range dynamic measurement error of no more than 0.1% pm, ensuring lossless and distortion-free digitization of microsecond-level transient electrical glitches. When extracting charging waveform distortion, its nonlinear differential sequence information entropy evaluation benchmark is defined as... Where H is the theoretical information entropy value of the sequence under the current analysis window; K1 is the total number of effective discrete data samples intercepted within the sliding time window, which is marked as [500, 1000] continuous sampling points in this embodiment. Setting this range can completely cover at least two power frequency AC fundamental wave cycles, thereby achieving the optimal balance between preventing low frequency cutoff effects and avoiding excessive smoothing to mask transient changes. This is the empirical probability distribution identifier for the occurrence of the difference amplitude interval i within the entire time window.

[0048] The logic for determining the distortion of the charging waveform is as follows: extract the real-time charging voltage and current parameters from the real-time electrical environment characteristic flow; establish a sliding time window with the above-mentioned preset step size range on the time axis, and extract the discrete voltage and current values ​​within the window in the order of time sequence; for the discrete voltage and current values ​​at adjacent time points, perform a difference operation by subtracting the value of the previous time point from the value of the next time point to generate a nonlinear difference sequence; divide the values ​​into several equidistant intervals based on the maximum and minimum amplitudes, count the frequency of the numerical points falling into each corresponding interval in the nonlinear difference sequence, and divide the frequency by the total number of valid discrete data samples K1 within the sliding time window to obtain the probability distribution value of each corresponding interval; For each interval's probability distribution value, a multiplication operation is performed between the probability distribution value and its base-2 logarithm to obtain the intermediate entropy sub-item for each interval. The intermediate entropy sub-items of all intervals are summed, and the summation result is inversely calculated to obtain the nonlinear differential sequence information entropy within the current time window. In this embodiment, to eliminate the interference of the naturally occurring gradual voltage trend that appears as the battery's charging depth increases, a preset digital high-pass filtering algorithm (preferably with a cutoff frequency of 0.5Hz) is introduced to filter out DC or low-frequency drift components below this cutoff frequency in the nonlinear differential sequence information entropy. The purified output value after filtering is then determined as the final indicator of abnormal charging waveform distortion.

[0049] In this step, the digital high-pass filtering algorithm is specifically configured as a second-order Butterworth digital high-pass filter. Its detailed execution logic is as follows: First, the preset filter coefficients are obtained. These coefficients are based on a 0.5Hz cutoff frequency and a preset sampling rate (corresponding to the aforementioned 100kHz). They are calculated using a standard second-order Butterworth transfer function and conventional digital signal processing discretization methods such as bilinear transformation, resulting in a unique set of constants (e.g., feedforward coefficient b and feedback coefficient a). This filter exhibits the flattest amplitude-frequency characteristics within the passband, minimizing artificial distortion of the micro-amplitude distortion characteristics required for information entropy calculation. Second, the filter's differential iterative equation is constructed. The current real-time entropy sample value is combined with the real-time entropy input samples from the previous two time steps, along with the preliminary filtered output samples from the previous two time steps, and a discrete weighted combination operation is performed according to preset coefficients to obtain the preliminary filtered sequence. Third, a zero-phase bidirectional filtering operation is performed to eliminate the influence of the inherent nonlinear phase lag of the second-order filter on subsequent phase angle detection. Finally, the preliminary filtered sequence is reversed on the time axis and filtered again using the same filter. The output result is then reversed again to restore the original result. By canceling out the phase shift through two mirror-image filtering actions, the phase information of the signal in the time domain is preserved while filtering out low-frequency drift terms, thus providing an absolutely true physical reference for microsecond-level frequency domain phase shift feature extraction.

[0050] Specific implementation instructions for identifying the degradation status of potential hazards: This embodiment addresses the inconsistent cell quality in outdoor electric bicycle battery packs and the inability of charging stations to obtain accurate BMS deep status reports due to interface limitations. Ignoring the long-term aging of the battery and directly applying a uniform safety monitoring threshold can easily lead to frequent false alarms for new batteries and missed alarms for older batteries causing fires. This embodiment constructs a method to infer the internal chemical state using external charging cycle characterization. Specifically, the feature extraction execution action employs a health prediction network built on a Long Short-Term Memory (LSTM) artificial neural network architecture. The forget gate and input gate mechanisms within the LSTM effectively overcome the problem of decaying correlations in historical long-cycle data, uncovering the cumulative damage caused by increased cell polarization resistance due to historical operating conditions such as deep discharge, overcharging, or high-temperature operation.

[0051] Before performing the prediction simulation, the historical charging cycle data stream input to the health prediction network meets the following data cleaning prerequisites: all non-same-source business parameters have been Z-score normalized to eliminate the physical dimension gap between "temperature parameters" and "voltage parameters", and the sampling missing segments in the data sequence caused by equipment disconnection are repaired by median nearest neighbor interpolation. The sequence length is uniformly resampled into a fixed multidimensional feature tensor matrix.

[0052] When performing Z-score normalization, the following statistical benchmark is established: the "benchmark feature table" preset in the system memory is retrieved. This table is obtained by offline statistical fitting of at least two thousand sets of standard charging cycle data covering different battery aging levels (SOH from 100% to 70%). See Table 0 below for details. Table 0: Examples of battery degradation extracted from offline fitting in the benchmark feature map table

[0053] Extract the baseline mean of voltage parameters from the baseline feature table. Standard deviation from the benchmark and the reference average value of current parameters Standard deviation from the benchmark In this embodiment, the average voltage reference value The optimal quantization value is calibrated to 3.7V, with a reference standard deviation. Calibration was performed at 0.45V; offline fitting calibration was conducted based on historical charging datasets, with the current parameter's baseline mean value. The preferred quantization value is configured as 2.0A, with a baseline standard deviation. Configured to 1.5A; acquire real-time charging voltage samples and perform subtraction operation to subtract the baseline average value. Obtain the centralized residual; divide the centralized residual by the benchmark standard deviation. This outputs dimensionless normalized voltage characteristics; the same subtraction of the mean and division by the standard deviation are performed on the current parameter to obtain dimensionless normalized current characteristics. This embodiment eliminates the physical dimension deviation caused by the difference in gain between different batches of sensor hardware, ensuring that the feature vectors entering the subsequent health prediction network have absolute logical consistency and temporal stability.

[0054] The calculation logic for the hazard attenuation status indicator is as follows: retrieve the historical charging and discharging cycle parameters associated with the current charging interface service identifier, specifically extracting parameters including the historical highest single charging isothermal extreme value, constant current stage duration, and static voltage drop amplitude after charging termination; concatenate the above-mentioned normalized and cleaned parameters according to the time sequence of charging batches, and map them to construct a time-series attenuation vector; use the time-series attenuation vector as input stimulus and inject it into the sequence input layer of the pre-trained offline health prediction network (in this embodiment, a long short-term memory artificial neural network with an LSTM architecture); it should be noted that for non-homogeneous parameters such as the highest charging isothermal extreme value, constant current stage duration, and static voltage drop amplitude, the normalization process also relies on a pre-set benchmark feature table, which contains the global benchmark mean and standard deviation of each corresponding parameter obtained through offline large-sample statistics, and the calculation logic is consistent with the Z-score processing of voltage / current.

[0055] For the health prediction network, it should be noted that: before inputting the time-series decay vector into the health prediction network, this pre-configured health prediction network has undergone an offline training phase. The historical charging training dataset stored in the cloud server configuration file is retrieved for model training; the historical charging training dataset meets the following data characteristics and calibration conditions: its feature input samples are extracted from historical operational parameters of batteries from the same batch that have undergone at least 500 standard charge-discharge cycles in a controlled temperature chamber; its corresponding supervision label is configured to obtain the absolute value of the true battery health (SOH) decay percentage through a capacity calibration experiment using constant current discharge testing every 50 cycles.

[0056] During training, forward propagation is performed to generate predicted decay values, and the mean squared error (MSE) loss function is invoked. The sum of squared residuals between the predicted decay values ​​and the corresponding supervision labels is calculated, and the weight matrix inside the health prediction network is iteratively updated based on the gradient backpropagation mechanism of the loss function (calling the Adam optimizer algorithm) until the loss function converges to an error threshold of less than 0.5%. After training convergence, the extracted historical charging and discharging cycle parameters associated with the current business identifier are mapped to a time-series decay vector, which is then injected into the network as input. The health prediction network performs matrix multiplication and activation mapping based on the fixed weight matrix, generating continuous regression values ​​in its fully connected output layer, and mapping these values ​​to identify the hazard decay status. The health prediction network performs high-dimensional space mapping and nonlinear activation operations on the time-series decay vector based on its internally fixed hidden layer weight matrix and bias terms, generating continuous regression values ​​in the fully connected layer. The standard Sigmoid activation function is used to perform range compression and expansion processing on the continuous regression values, forcibly clamping them to the probability interval [0,1]. The final output value within this interval is determined as the indicator of the potential hazard decay state. It's important to note that the bias term is configured as a long-term baseline parameter to characterize the historical aging of the battery cell at a deeper network level. This bias term also has a corresponding interface allocated in system memory to support resetting the battery during a power flow isolation operation when it is overwritten by catastrophic features. High-dimensional spatial mapping and nonlinear activation operations are performed on the time-series decay vector. The standard Sigmoid activation function is used to perform range companding on the continuous regression value, forcibly clamping it to the probability interval [0,1]. The final output value within this interval is then determined as the indicator of the potential hazard decay state. The closer the absolute value of the potential hazard decay state indicator is to 1, the more severe the current internal aging tendency of the battery is, and the higher the probability of thermal runaway.

[0057] Step S2: Obtain the background grid fluctuation sequence of adjacent nodes representing adjacent non-local charging nodes, and obtain the high-frequency ripple sequence of the local charging circuit representing the underlying power supply circuit of the local charging port. The specific implementation logic is as follows: Cross-node communication is performed by receiving bus electrical parameter sequences fed back from other adjacent charging nodes in the same distribution area that are in a physically unloaded or constant current operating state. These sequences are defined as background power grid fluctuation sequences of adjacent nodes to characterize the background noise of the macroscopic power grid. Specifically, cross-node communication is performed by a broadband power line carrier (HPLC) communication module or twisted-pair transceiver coupled to the current distribution area. In order to ensure the timing alignment of the data at the distribution area level, the network clock synchronization error between nodes is configured to be no more than 5 milliseconds. The detailed extraction and alignment mechanism will be described in subsequent steps.

[0058] The intrinsic AC current ripple response sequence, excited by the pulse width modulation (PWM) switching frequency inherent in the AC / DC power conversion stage and superimposed on the output power supply network, is acquired and defined as the high-frequency ripple sequence of the local charging circuit, serving as a high-frequency excitation probe for detecting battery polarization impedance. The preferred range of the PWM switching frequency is [20kHz, 50kHz]. The high-frequency sequence is acquired through digital quantization using a low-level signal capture medium configured with a high-frequency sampling rate of at least 10kHz.

[0059] Step S3: Using the background grid fluctuation sequence of adjacent nodes and the high-frequency ripple sequence of the local charging circuit as input, perform spatial difference and frequency domain mapping logic to generate electrochemical state compensation coefficients to characterize the electrochemical microscopic physical state. Specific implementation logic: The time-domain energy envelope features of the background grid fluctuation sequences of adjacent nodes are extracted, and the global background noise deviation of the transformer area, representing the intensity of global environmental interference, is generated through an asynchronous alignment mechanism. This global background noise deviation is then used as a dynamic constraint to perform differential companding on the gain parameters of the high-frequency ripple sequence of the local charging circuit, in order to counteract aliased background grid harmonic interference. The mapping model for differential companding is defined as follows:

[0060] in This is the high-purity differential mode ripple sequence output at the current time t after companding; The original high-frequency ripple sequence of the local charging circuit is obtained. This represents the global background noise deviation of the station area extracted earlier. To suppress the sensitivity constant, in this embodiment, its preferred value is set within the range of [0.1, 0.5]. This ensures that when encountering common-mode noise from the distribution area, the system can exponentially and rapidly reduce the gain confidence of the local high-frequency ripple, preventing false high-frequency abrupt changes from causing malfunctions.

[0061] The calculation logic for the high-purity differential mode ripple sequence is as follows: Specifically, the global background noise deviation of the transformer area and the preset suppression sensitivity constant are obtained, and a multiplication operation is performed on the two. The inverse of the obtained product value is taken as the intermediate exponent term. Using the natural constant as the base and the intermediate exponent term as the exponent, a power operation is performed to generate a dynamic attenuation coefficient between 0 and 1. The discrete amplitude of the high-frequency ripple sequence of the local charging circuit is extracted point by point in the time domain. Each discrete amplitude is multiplied with the dynamic attenuation coefficient, and the set of sequences output by the operation is established as the high-purity differential mode ripple sequence.

[0062] Discrete frequency domain phase angle demodulation is performed to extract the frequency domain phase angle offset features between the high-purity differential mode ripple sequence and the original excitation pulse width modulation signal. The specific implementation method has been defined in the previous description, namely: a Fast Fourier Transform (FFT) is performed on the high-purity differential mode ripple sequence after adding a Hanning window to generate a complex spectrum matrix. The target frequency point location index is located according to the fixed ratio between the pulse width modulation switching frequency and the global sampling frequency. The imaginary and real parts of the complex data under the index are extracted and arctangent operation is performed. Then, the reference starting phase zero point is subtracted to obtain the absolute difference value, which is used as the frequency domain phase angle offset feature.

[0063] Extract the frequency domain phase offset features within the current operation cycle, and retrieve the reference phase feature quantity smoothed by arithmetic mean calculation within the historical time window (preferably the past 10 to 20 consecutive sampling cycles) in the cache; calculate the absolute value of the difference between the two to obtain the time difference gradient value; The time difference gradient value is input into a preset nonlinear activation function, and the dynamically updated step size parameter between 0 and 1 is mapped to the output. As described above, the nonlinear activation function is configured as an exponential decay mapping model based on the natural constant base. The specific logic is as follows: the time difference gradient value is multiplied by the dimensionless sensitivity smoothing constant (preferably [0.2, 0.5]), the negative is taken as the exponent, and then the result of the exponentiation is obtained by subtracting the result of the exponentiation from the value 1. The dynamically updated step size parameter is used as the current weight and multiplied with the frequency domain phase angle offset feature of the current operation period to obtain the first product. The historical forgetting weight is obtained by subtracting the dynamically updated step size parameter from the value 1, and the historical forgetting weight is multiplied with the anti-interference warning sensitivity deviation of the previous period to obtain the second product. The first product and the second product are added to complete the iterative update of the anti-interference warning sensitivity deviation of the current period. Based on the convergence state of the updated anti-interference warning sensitivity deviation, the electrochemical state compensation coefficient is mapped to the output.

[0064] It should be further explained that the disturbance rejection warning sensitivity deviation is defined as a dynamic adaptive filtering parameter used to characterize the degree of microscopic electrochemical polarization hysteresis in the battery. Under strong grid disturbance conditions, after eliminating external high-frequency harmonics and random common-mode noise interference, it is the purity confidence assessment value of the frequency domain phase angle shift characteristics caused by actual micro-internal short circuits or dendrite growth precursors within the cell. This deviation is the convergence value of the actual degradation state obtained by the system iteratively filtering the nonlinear time difference gradient between the current anomaly characteristics and the historical stable benchmark through a dynamic gating mechanism. Its magnitude directly reflects the system's current sensitivity to capturing weak electrochemical anomalies and its noise rejection confidence.

[0065] The initial historical cache value of the anti-interference warning sensitivity deviation is forcibly reset to zero, or preferably, the frequency domain phase shift feature extracted for the first time is directly assigned to the anti-interference warning sensitivity deviation of the first cycle as the origin for calculating the subsequent time difference gradient. Regarding the output of the electrochemical state compensation coefficient based on this deviation, the specific numerical determination mechanism is as follows: the local main control logic continuously calculates the absolute value of the difference between the anti-interference warning sensitivity deviations of two adjacent cycles within a preset evaluation time window (e.g., 10 consecutive clock cycles). When this absolute value is continuously less than the system's preset steady-state convergence threshold (0.01), the system determines that it has entered the convergence state and directly extracts the anti-interference warning sensitivity deviation of the current cycle as the benchmark operator, further inputting it into the aforementioned nonlinear smoothing mapping logic (Sigmoid function) to calculate the penalty multiplier, thereby completing the final generation and output of the electrochemical state compensation coefficient.

[0066] When the polarization impedance changes abruptly, the dynamic update step size approaches 1, quickly discarding historical baggage and accepting the current anomaly; when in a stable charging state, the dynamic update step size approaches 0, maintaining strong anti-interference capability against small high-frequency noise, and resolving the contradiction between sensitivity and robustness.

[0067] Specific implementation instructions for the global background noise deviation of the distribution area: When extracting local high-frequency characteristics, if there are other high-power loads in the same distribution area, it is easy to induce severe harmonic surges in the AC bus. Without spatial environmental comparison, these background grid harmonics will directly pollute the local electrical characteristic flow, leading to erroneous physical power flow isolation actions. This action uses horizontal comparison of distribution area-level data to remove background noise not caused locally. Performing this cross-node acquisition action relies on the fieldbus or wireless mesh network environment within the charging pile group. This communication environment must meet the following conditions to ensure multi-source timing alignment: it must have multicast transparent transmission capability within the current distribution area subnet, and the network delay clock synchronization error between nodes must be within a time boundary of no more than 5 milliseconds. Then, the system receives bus electrical parameter sequences pushed by other adjacent charging nodes within the same distribution area that are in physical idle standby or absolute constant current operation state via the communication interface, and aggregates them into a background power grid fluctuation sequence for adjacent nodes. Specifically, the local microprocessor calls a broadband power line carrier (HPLC) communication module or twisted-pair transceiver coupled to the current distribution area as the physical receiving medium through the system's general peripheral bus. When a valid interrupt start signal is detected from the external carrier bus level, the local microprocessor activates the associated data receiving register, extracts the radio frequency data frame encapsulated according to the preset physical layer protocol, and verifies the checksum field at the end of the data frame using a check operator. If the verification passes, the local microprocessor maps the parameter bit stream in the data frame payload to static random access memory (SRAM) and assembles the bus electrical parameter sequence in the logical address space.

[0068] An extraction time window of 100 milliseconds is set, and the absolute value of the amplitude at each discrete time point in the background power grid fluctuation sequence of adjacent nodes is extracted. The 100-millisecond extraction time window is further divided into several continuous sub-intervals of two milliseconds each on the time axis. A local numerical comparison operation is performed in each sub-interval, and the maximum absolute value in the sub-interval is selected and retained as the local discrete peak point. The preset five-point moving average algorithm logic is invoked, and the discrete arithmetic mean operation is performed on all the extracted local discrete peak points in chronological order (specifically, the mean of the current peak point and the two adjacent peak points before and after it is calculated). The series of average values ​​obtained by calculation are sequentially connected to generate a time-domain energy envelope feature that filters out high-frequency spikes and is used to characterize the envelope trend of low-frequency macroscopic fluctuations.

[0069] The system retrieves the current timestamp from the clock register of the local master control unit and simultaneously parses the remote node clock tag encapsulated in the header of the background power grid fluctuation sequence data packet of the adjacent node; it performs a subtraction operation to obtain the time offset difference between the local timestamp and the remote clock tag, and performs a reverse data shift operation based on the time axis on the time-domain energy envelope feature in the memory sequence based on this time offset difference to complete the asynchronous alignment mechanism to offset communication delay; for the time-domain energy envelope features of all available adjacent nodes after time alignment, it performs an arithmetic mean operation on the envelope values ​​of all available nodes under the same time section index to generate a dynamic reference mean line characterizing the global environmental interference intensity of the current transformer area; The local node retrieves its own temporal energy envelope value at the same time and subtracts the dynamic reference mean line. Then, the absolute value of the difference is extracted and finally the absolute value is output as the global background noise deviation of the station area.

[0070] Step S4: Based on the electrochemical state compensation coefficient, weight injection compensation is performed on the original multi-dimensional feature joint calculation logic. The fusion ratio of the charging waveform distortion and the hazard attenuation state indicator is dynamically adjusted according to the electrochemical state compensation coefficient to generate an adaptive safety judgment threshold. The specific implementation logic is as follows: It needs to be defined in advance that multidimensional feature joint calculation logic refers to the hardware-level algebraic operation control center that is embedded in the local microprocessor storage medium and directly scheduled by the arithmetic logic unit (ALU) or multiply-accumulate (MAC) physical operator.

[0071] This logic unit has an input-output mapping relationship in the system architecture: it uses the extracted charging waveform distortion degree, hidden danger attenuation status indicator, and electrochemical state compensation coefficient as heterogeneous data input sources; its core operation mechanism is to use the electrochemical state compensation coefficient as a penalty multiplier under the drive of the underlying clock pulse, perform a zero-sum transfer between the compensation margin weight of environmental parameters and the basic risk weight, and construct a normalized weight row matrix and a real-time feature column matrix to perform matrix multiplication; its final physical output result is determined as the adaptive safety judgment threshold for adjudicating the physical power flow isolation action of the terminal node on the underlying power supply circuit side.

[0072] The electrochemical state compensation coefficient is input into a preset nonlinear smoothing mapping logic. As fully disclosed in the aforementioned definition, this logic is specifically configured to combine the inflection point threshold of the electrochemical deterioration center with the steepness coefficient of the Sigmoid normalization mapping mechanism to forcibly constrain its dimensions and output a penalty multiplier between the lower limit value 0 and the upper limit value 1. The environmental parameter weighting equation is as follows: obtain the initial environmental parameter compensation margin weight in the multi-dimensional feature joint calculation logic, multiply the initial weight by the penalty multiplier, and calculate the stripping weight attenuation; perform a subtraction operation to deduct the stripping weight attenuation from the environmental parameter compensation margin weight to obtain the actual environmental compensation weight after shrinkage. The basic risk weights of the initially allocated hazard attenuation status indicators are obtained. The attenuation amount of the stripped weights is used as a zero-sum transfer compensation term and added to the basic risk weights to obtain the expanded actual risk decision weights. It is ensured that the sum of the actual environmental compensation weights and the actual risk decision weights is equivalent to the sum of the two initially allocated weights. A normalized parameter fusion matrix is ​​constructed using the reconstructed actual environmental compensation weights and the actual risk decision weights, and joint calculation logic is executed in conjunction with real-time features. The mathematical model for this matrix calculation is defined as follows:

[0073] in These are dynamically generated adaptive safety judgment thresholds; This is a preset initial empirical threshold constant for waveform distortion. To construct a one-dimensional normalized weight row matrix; For constructing a one-dimensional real-time feature column matrix; "" represents the inner product operation of the matrix. In the dynamic execution flow, the inverse of the actual risk decision weight with negative penalty attribute and the actual environmental compensation weight with positive compensation attribute are obtained. These two are arranged in sequence to construct a normalized weight row matrix with one row and two columns. The extracted hidden danger attenuation status identifier and the normalized global background noise deviation of the transformer area are obtained. (It should be noted that the normalization process specifically uses the aforementioned preset divergence boundary as the denominator benchmark to perform a division mapping on the currently obtained global background noise deviation of the transformer area to force a unified feature dimension for matrix operation.) These two are arranged in sequence to construct a real-time feature column matrix with two rows and one column. Matrix dot product operation is performed on the normalized weight row matrix and the real-time feature column matrix to output a scalar comprehensive bias coefficient. The value 1 and the comprehensive bias coefficient are added to obtain the correction multiplier. The preset initial empirical threshold constant of waveform distortion is multiplied with the correction multiplier, and the absolute value of the output is generated and determined as the final adaptive safety judgment threshold value.

[0074] Matrix multiplication is performed on the normalized weighted row matrix and the real-time feature column matrix. The local microprocessor acquires and decodes machine instructions via the system bus, controlling its internal arithmetic logic unit (ALU) or multiply-accumulate (MAC) physical operator to extract the first set of floating-point level signals representing the normalized weighted row matrix and the second set of floating-point level signals representing the real-time feature column matrix from the first and second general-purpose register arrays, respectively. The multiply-accumulate physical operator executes alternating floating-point multiplication and addition pulses on the underlying circuitry according to the physical clock cycle, continuously feeding intermediate results back to the accumulator. After the instruction cycle for the matrix dimension is completed, an accumulated voltage pulse result representing the final scalar value is generated and latched into the designated output result register, completing the physical generation of the comprehensive bias coefficient.

[0075] Step S5 involves performing a high-frequency joint comparison and verification, triggering either physical power flow isolation or a derating conservative power-off protection strategy. The specific implementation logic is as follows: A cyclic comparison is performed to determine whether the current feature offset (the charging waveform distortion degree extracted in real-time during the preceding step S1) exceeds the adaptive safety judgment threshold. If it does, a coordinated control command is immediately generated and issued: the first control branch responds by generating a pulse blocking signal, driving the underlying physical contactor to directly disconnect the power flow and perform hard isolation; the second logic branch synchronously captures the transient charging waveform distortion degree feature corresponding to the moment the physical contactor disconnects, and uses this transient feature to reset the long-term baseline parameters of the health prediction network. The static definition model for its reset iteration operation is as follows:

[0076] in To reset the long-term baseline parameters at the current time t after the update; The distortion characteristics of the transient charging waveform at the moment of disconnection are captured; To reset the historical long-term baseline parameters at time t-1 before the reset; The learning rate coefficient for disaster state is preferably set to a fixed constant within the range of [0.80, 0.95] in this embodiment. During the execution flow, the transient charging waveform distortion characteristics are acquired and multiplied with the disaster state learning rate coefficient to generate the current disaster bias component. The forgetting weight is obtained by subtracting the disaster state learning rate coefficient from the value 1, and multiplied with the cached historical long-term baseline parameters to generate the historical residual component. The current disaster bias component and the historical residual component are then added together to complete the overwrite reset of the long-term baseline parameters.

[0077] Parallel monitoring is performed to determine whether the statistical variance of the global background noise deviation in the transformer area consistently exceeds the preset divergence boundary within five consecutive sliding statistical windows (or a business observation period with a cumulative observation time exceeding 100 milliseconds). If the statistical variance of the global background noise deviation in the transformer area continuously exceeds the preset divergence boundary, it is determined that the physical probe feedback chain has failed due to external strong noise pollution. At this time, the derating conservative power-off protection strategy is triggered: the environmental output weight of the electrochemical state compensation coefficient in the calculation logic (the actual environmental compensation weight obtained by zero-sum transfer reconstruction in the preceding step S4) is forcibly set to zero to block external uncertainty compensation. At the same time, the conservative defense line generation logic is executed, specifically extracting the initial empirical threshold constant of waveform distortion fixed in the original system and retrieving the preset conservative derating coefficient; the conservative derating coefficient is a dimensionless safety discount multiplier, which is preferably calibrated to [0.60, 0.75] in this embodiment. The initial empirical threshold constant of waveform distortion is multiplied by the conservative derating coefficient, and the absolute value of the product is output, directly covering all the original dynamic calculation processes. The adaptive security judgment threshold is reconstructed and solidified into the above-mentioned absolute value of the product, thereby forming a single-node conservative defense limit value that depends on the local domain's underlying characteristics and has high anti-disturbance physical redundancy.

[0078] It should be further explained that the generation and issuance of cooperative control instructions specifically involves the local microprocessor assembling a data packet with a byte-structured architecture as the cooperative control instruction in system memory. The main data structure of this data packet includes, in order: a header field representing the instruction priority, a two-byte control command field representing the target operation category, and a tail field representing cyclic redundancy check. The local microprocessor pushes the assembled cooperative control instruction into the transmit buffer queue of the general-purpose input / output (GPIO) controller to drive the corresponding external pins to generate a pulse blocking signal that causes a level transition. This level transition directly cuts off the drive current of the high-power relay coil connected in series in the power supply circuit, thereby forcibly disconnecting the underlying physical contactor. The two-byte control command field is a hexadecimal identifier used to trigger the hardware state machine; for executing physical power flow isolation actions, its preferred value in this embodiment is 0xFFFF to provide the highest fault tolerance against bit flips in the underlying driver parsing; as a non-limiting comparison, under normal power supply conditions, its preferred range is a heartbeat maintenance code (any discrete assignment within the range of 0x0000 to 0x00FF).

[0079] The specific implementation of the preset divergence boundary is as follows: When the same distribution substation encounters external physical-level noise pollution, the energy level of the grid noise floor will overwhelm the high-frequency probe signal, resulting in a "probe blind zone". If no intervention is taken at this time, the incorrect high-frequency characteristics will induce miscalculation of the penalty multiplier, leading to a large-scale erroneous power outage. This embodiment introduces a quantitative configuration parameter characterizing the "probe failure critical point". When the environment deteriorates to this point, the failed high-frequency feedback chain is decisively cut off, triggering a derating conservative power outage protection strategy. The logic for determining the preset divergence boundary is: the controlled physical experimental environment meets the following feature extraction conditions: an isolated distribution substation simulation sandbox is constructed, which is equipped with a high-power AC grid simulation source with programmable arbitrary waveform output function, and several charging node loads with real power conversion characteristics. In the system defense line judgment logic, the mathematical judgment for triggering the degradation fallback action is defined as follows: ; in This is the statistical variance operator for parameters within the sliding statistical window; The sequence of global background noise deviation for the transformer area is continuously output by the preceding steps; In this embodiment, the preferred value of the preset divergence boundary characterizing the tolerance limit of the power grid background noise is defined as a fixed constant in the interval [2.5, 3.8]. The offline calibration process with a preset divergence boundary is as follows: In a controlled physical experimental environment, a high-power AC power grid simulation source is used to continuously inject broadband interference electromagnetic waves conforming to the standard white noise distribution into the transformer area simulation bus, and the waveform data of each node under normal charging conditions are recorded; the energy gain level of the injected interference electromagnetic waves is gradually increased, and at each gain level, a stable injection residence time of at least ten minutes is maintained to simulate different distribution transformer area operating conditions from light pollution to severe pollution; at each injected energy gain level, the complete calculation action of extracting the global background noise deviation of the transformer area in the previous embodiment is executed synchronously to generate an offline deviation test sample library; the offline deviation test sample library is sliced ​​according to the time dimension, and the statistical variance value of the deviation data in each slice sequence is calculated; a probe signal-to-noise ratio evaluation logic is introduced, and when it is determined that the effective amplitude of the high-frequency ripple sequence of the local charging circuit superimposed on the bus is submerged by the background noise to the critical state where the signal-to-noise ratio is lower than one decibel (1dB), the injected energy gain level corresponding to this time is locked as the "physical probe failure condition"; Extract all statistical variance values ​​under the "physical probe failure condition" and construct the failure state probability density distribution curve; based on the normal distribution statistical algorithm, calculate the fifth percentile of the failure state probability density distribution curve (representing the tail confidence interval), and establish this percentile as the preset divergence boundary, and burn it into non-volatile memory.

[0080] Furthermore, based on the matrix calculation model analysis in step S4 above, it can be seen that when the adaptive safety judgment threshold approaches its theoretical minimum (i.e., the penalty multiplier approaches the upper limit, causing the complete stripping of environmental compensation margin, or the single-node conservative limit when triggering the derating conservative power-off protection strategy), it indicates that the current situation is facing an extremely severe internal electrochemical polarization degradation state (approaching thermal runaway precursor), or the underlying high-frequency physical probe feedback link is completely contaminated and failed by extreme common-mode noise. Under this trend, the control end deprives the tolerance to external disturbances and forcibly tightens the safety envelope, capturing weak charging waveform distortions with the highest priority sensitivity, ensuring that physical isolation actions are triggered as soon as a micro-short circuit occurs, blocking the thermal runaway heat accumulation path.

[0081] When the adaptive safety judgment threshold approaches its theoretical maximum (i.e., the penalty multiplier approaches the lower limit, resulting in a low basic risk weight, and the environmental compensation margin is fully utilized), it indicates that the current battery cell is in a high health state, with no cumulative degradation from historical charge-discharge cycles, and no hysteresis abrupt change in the locally detected high-frequency ripple phase angle. Under this state, fluctuations exhibited by the underlying power grid are determined to originate purely from compliant global environmental disturbances. By releasing the environmental compensation weight to maximize the increase of the judgment threshold, false positives and erroneous power outages are eliminated while ensuring absolute safety, thus guaranteeing the continuity of the charging process.

[0082] The global background noise deviation in the distribution area is nonlinearly negatively correlated with the gain parameter of the high-purity differential mode ripple sequence; and nonlinearly positively correlated with the triggering state of the derating conservative power outage protection strategy (its threshold determination criterion is the preset divergence boundary in step S5 above). When the global interference in the distribution area surges, the locally captured high-frequency ripple will inevitably be mixed with a proportional amount of harmonic energy. Direct acceptance will lead to distortion of polarization impedance assessment. By using an exponential decay mapping with the natural constant as the base (specifically corresponding to the differential companding mathematical model implemented by introducing a suppression sensitivity constant in step S3 above), it is transformed into a dynamic decay coefficient, which can quickly and smoothly reduce the acceptance weight of local features before the noise overwhelms the probe. This ensures that in complex distribution networks, grid surges caused by the start-up and shutdown of external heavy loads will not be misjudged as a warning signal of internal battery fire, directly supporting the technical effect of "no oscillation of the defense line under strong grid disturbance environment" of this invention. For the frequency domain phase shift characteristics and their corresponding time difference gradient values, the time difference gradient values ​​are nonlinearly positively correlated with the dynamic update step size parameter (as defined in the exponential empirical mapping mechanism above); and further positively correlated with the contraction amplitude of the penalty multiplier and the adaptive safety judgment threshold. The damage to the solid electrolyte interface film caused by a micro-internal short circuit inside the battery manifests as a sharp increase in charge transfer impedance on the impedance spectrum, leading to phase hysteresis in the high-frequency current response on the time axis. Mapping its differential gradient to a step size parameter aligns with the mutation capture logic in adaptive filtering: when the data is stable, a large forgetting weight (a value of 1 minus a dynamic update step size parameter approaching 0) is used to maintain the baseline; when the data undergoes a mutation, the current weight is instantly amplified. Directly linking the mutation gradient to the penalty multiplier allows the control unit to instantly raise the defense sensitivity level at the cost of cutting off environmental tolerance within a few hundred milliseconds of capturing weak precursors to thermal runaway, achieving "seeing through the battery's micro-state and performing dynamic interception."

[0083] This embodiment is configured in the following digital twin and actual monitoring scenario: deploying an intelligent charging system including a local microprocessor, a broadband power line carrier communication module, and a high-frequency analog-to-digital conversion peripheral module. Multiple virtual and physical charging nodes are connected within the target area, and the tested batteries include new charging units with no degradation characteristics and service units with deep aging characteristics. The system's pre-static parameters for micro-node data processing are configured as follows: The initial empirical threshold constant for waveform distortion is obtained and fixed to a value of 100 in the local register; the inflection point threshold for deterioration center is obtained and fixed as a proportional characteristic quantity deviating 15% from the normal internal resistance baseline; the kurtosis coefficient is obtained and calibrated as a dimensionless constant of 10; the compensation margin weight of the initially allocated environmental parameters in the base core business logic is retrieved as 0.60, and the basic risk weight is retrieved as 0.40; the preset divergence boundary is set to 3; and the conservative derating coefficient is set to 0.65.

[0084] The data flow logic of this invention is executed as follows: the underlying analog-to-digital conversion peripheral continuously injects discrete voltage and current parameters into the local microprocessor; the main control logic periodically retrieves the aging probability values ​​of continuous regression from the health prediction network, and simultaneously extracts the high-purity differential mode ripple phase angle deviation value obtained from demodulation from the digital signal processing hardware acceleration kernel; the above heterogeneous variables are incorporated into the multi-dimensional feature joint calculation logic, which calculates the zero-sum transfer quantity according to the nonlinear mapping rule, and dynamically updates the actual environmental compensation weight and the actual risk decision weight; the internal arithmetic logic unit extracts the floating-point level signal and real-time feature level signal representing each weight, executes the multiply-accumulate physical instruction to perform matrix inner product operation, and finally generates and latches the adaptive safety judgment threshold value.

[0085] Table 1: Examples of system response calculations under different environmental disturbances

[0086] Table 1 shows the physical output of removing environmental noise and targeting and intercepting thermal runaway sentinel signals by performing "zero-sum reconstruction" and "multiply-addition calculation" under various macroscopic common-mode disturbances and microscopic polarization states.

[0087] Among them, the common mode fluctuation variance of the transformer area represents the statistical variance of the global background noise deviation of the corresponding transformer area; the normalized common mode deviation represents the global background noise deviation of the transformer area after normalization; the hazard attenuation probability represents the hazard attenuation status indicator; and the micro phase angle hysteresis deviation represents the frequency domain phase angle offset characteristics. Define the dynamic defense contraction rate as The system obtains the "adaptive safety judgment threshold" output in the current operation cycle and divides it by the "baseline safety judgment threshold" under the same common-mode disturbance level in the same transformer area environment but without abnormal micro-polarization. The percentage difference obtained by subtracting the value 1 from the aforementioned division is used to establish this parameter. This parameter quantitatively characterizes the system's aggressive response in shielding the false safety illusion of the external environment and forcibly compressing the safety envelope towards the physical absolute isolation boundary when encountering internal electrochemical mutations or probe physical failures. By comparing "Scenario 3" and "Scenario 4" in the table, the technical effect of performing weighted zero-sum transfer based on the electrochemical state compensation coefficient is directly verified. Under completely consistent external moderate disturbance intensity (the common-mode fluctuation variance of the transformer area is 1.80, and the normalization deviation is 0.600), existing comparison techniques without multi-dimensional decoupling will be unable to isolate background grid interference and will continue to give high tolerance thresholds.

[0088] This technical solution integrates high-frequency joint comparison and verification. In scenario four, when the physical probe obtains a micro-phase angle hysteresis deviation of 26.0% (exceeding the 15% deterioration inflection point and reaching a deviation difference of 11.0%), it calls the aforementioned steepness coefficient to execute an exponential nonlinear mapping instruction. After Sigmoid normalization, it generates a stripping weight decay of 0.45. The actual environmental compensation weight drops from the initial 0.60 to 0.15, while the basic risk weight of the hidden danger simultaneously rises sharply from 0.40 to 0.85. The above weights are extracted and matrix multiplication and accumulation are performed: the product of the negative of the actual risk decision weight (-0.85) and the hidden danger probability 0.20 is obtained, and this product is added to the product of the actual environmental compensation weight 0.15 and the normalized common mode deviation 0.600 to generate a comprehensive bias coefficient of -0.08. After adding this comprehensive bias coefficient to the value 1, multiplying by the empirical constant 100, the final adaptive safety judgment critical value of 92 is output.

[0089] Compared to the system's raised disturbance rejection threshold (130) in Scenario 3 when no polarization anomaly occurs, this method achieves a dynamic defense contraction rate of 29.2% at the moment of capturing the outpost of thermal runaway (using the quotient of 92 and 130 minus the value of 1). This set of data not only verifies the determinism at the matrix calculation level, but also quantitatively demonstrates that this technical solution has the ability to exclusively locate internal severe electrochemical degradation states under the cover of strong noise power grids. Thus, before irreversible micro-short circuit heat accumulation occurs, the defense line is precisely lowered and the physical contactor is driven to disconnect, achieving a substantial technical breakthrough in addressing the problem of missed detection by existing static defense lines.

[0090] For extreme operating conditions, scenarios five and six demonstrate the absolute physical interception effectiveness of the preset divergence boundary. When the common-mode fluctuation variance of the transformer area exceeds the preset divergence boundary of 3.0 (3.50), the probe signal-to-noise ratio is determined to be zero, and the parallel monitoring logic triggers a conservative power-off protection strategy with derating. The control branch forcibly performs a memory zeroing truncation operation on the environmental weight and deviation characteristic chain, only calling the initial empirical threshold of 100 for the solidified waveform distortion degree and the conservative derating coefficient of 0.65 to perform an absolute multiplication operation, reconstructing and locking the dynamic critical value to 65. In this state, compared with the highest environmental compensation value, the defense shrinkage rate reaches 50%, using a physical derating method that forcibly deprives half of the available power envelope limit to obtain fire-prevention redundancy when the system loses its high-frequency field of view.

[0091] Define the following application range: Interval 1: The high-frequency excitation self-proving steady-state interval, where the charging waveform distortion is ≤ 80% of the adaptive safety judgment threshold; its boundary is determined by the derivative characteristics of the penalty multiplier mapping equation in the linearly gradually varying region in the multi-dimensional feature joint calculation logic. Within this numerical interval, the polarization overshoot does not cross the central abrupt pole of the activation function, indicating that the electrochemical microphysical state has not undergone nonlinear deterioration. This method maintains normal full-power power output, only performing nonlinear differential sequence information entropy rolling calculation within the sliding time window in the background, and storing the waveform distortion in the historical feature stream, without triggering any intervention control commands.

[0092] Interval 2: The environmental stripping and defense convergence interval, where 80% of the adaptive safety judgment threshold < charging waveform distortion ≤ 100% of the adaptive safety judgment threshold; its boundary is determined by the penalty multiplier of the electrochemical state compensation coefficient entering the exponential surge segment (the region of maximum first derivative). Within this numerical range, it is determined that the polarization impedance degradation rate has exceeded the ability of macroscopic environmental interference to be masked.

[0093] This method performs a weighted zero-sum transfer operation, rapidly decaying the compensation margin weight of environmental parameters according to a non-linear ratio, while simultaneously and equivalently amplifying the basic risk weight of the hazard attenuation status indicator. At this point, the first control branch enters a pre-charge state, awaiting over-limit adjudication.

[0094] Interval 3: Physical cut-off and disaster marking interval, where the power waveform distortion is greater than the adaptive safety judgment threshold, or the common mode fluctuation variance of the transformer area is greater than the preset divergence boundary; its boundary is defined by the interrupt trigger threshold condition of the comparator hardware, and the statistical tail confidence interval in the controlled calibration experiment where the signal-to-noise ratio of the high-frequency probe is locked below 1 dB.

[0095] The first control branch generates a pulse blocking signal to directly drive the underlying physical contactor to disconnect the power flow and perform hard isolation; the second logic branch synchronously captures transient distortion characteristics and overwrites and resets the long-term baseline parameters with a preset disaster state learning rate coefficient (a fixed constant in the aforementioned step S5 between 0.80 and 0.95). If triggered by the environmental statistical variance exceeding the divergence boundary, an additional conservative derating coefficient is locked for physical derating protection.

[0096] Figure 1 The technical roadmap illustrates the core operational logic of the active fire prevention method of this invention. The core execution entity on the left side of the diagram is the electric bicycle charging pile. The first module triggered by this charging pile is multi-source electrical feature acquisition. This module corresponds to steps S1 and S2 of this invention. Its execution logic includes acquiring the underlying real-time electrical environment feature stream to extract the charging waveform distortion, inputting the historical charging cycle data stream into the health prediction network to deduce the hazard attenuation status indicator, and simultaneously acquiring the background power grid fluctuation sequence of adjacent nodes and the high-frequency ripple sequence of the local charging circuit of the local port. The process then proceeds to the second module, "differential companding and frequency domain mapping," which corresponds to step S3 of this invention. Its execution logic involves using the aforementioned low-frequency and high-frequency sequences to extract the global background noise deviation of the transformer area, and performing differential companding on the high-frequency ripple to eliminate environmental harmonic interference and generate an electrochemical state compensation coefficient. The process then proceeds to the third module, "Adaptive Safety Judgment Critical Value Reconstruction," which corresponds to step S4 of this invention. Its execution logic is based on the generated electrochemical state compensation coefficient, performing weight injection compensation, and dynamically adjusting the fusion ratio of waveform distortion and hidden danger attenuation indicators to reconstruct and generate an adaptive safety judgment critical value. Finally, the process flows to physical blocking and degradation fallback, which corresponds to step S5 of this invention. Its execution logic is to perform high-frequency joint comparison and verification. When the real-time feature offset exceeds the reconstruction critical value, a collaborative control command is generated to drive the underlying physical power flow isolation action. When grid pollution causes common-mode fluctuations to exceed the preset divergence boundary, a derating conservative power outage protection strategy is forcibly triggered to achieve an all-weather active safety closed loop.

[0097] exist Figure 2In the active protection architecture operation flow shown, step S1 is executed to obtain the underlying real-time electrical environment feature stream and perform a local sequence cleaning operation to extract the charging waveform distortion degree; simultaneously, the historical charging cycle data stream is input into a pre-set health prediction network to deduce the hazard attenuation status indicator. Then, step S2 is executed to obtain the adjacent node background grid fluctuation sequence representing adjacent non-local charging nodes, and the local charging circuit high-frequency ripple sequence representing the underlying power supply circuit of the local charging port. Next, the core step S3 is executed, using the adjacent node background grid fluctuation sequence and the local charging circuit high-frequency ripple sequence as input, executing spatial differential and frequency domain mapping logic, and performing differential companding operation on the local charging circuit high-frequency ripple sequence based on the global background noise deviation of the transformer area extracted from the low-frequency sequence to generate an electrochemical state compensation coefficient to represent the electrochemical microphysical state. Then, the process moves to step S4, where, based on the electrochemical state compensation coefficient, weight injection compensation is performed on the original multi-dimensional feature joint calculation logic, and the fusion ratio of the charging waveform distortion degree and the hazard attenuation status indicator is dynamically adjusted according to the electrochemical state compensation coefficient to generate an adaptive safety judgment threshold. Finally, in step S5, a high-frequency joint comparison and verification is performed. If it is determined that the current feature offset exceeds the adaptive safety judgment threshold, a collaborative control command is generated to instruct the terminal node on the power supply circuit side to perform physical power flow isolation action. When the global background noise deviation of the transformer area exceeds the preset divergence boundary, a derating conservative power outage protection strategy is triggered.

[0098] Figure 2 The central object architecture ECS represents the main body of the electric bicycle charging station. The REFS at the input end represents the underlying real-time electrical environment characteristic flow, and the CWD derived from this represents the charging waveform distortion. The bypass HDSF refers to the hazard attenuation status indicator derived by the prediction network. LFWS represents the captured background grid fluctuation sequence of adjacent nodes, while HFRS refers to the high-frequency ripple sequence of the underlying local charging circuit. These two converge at the logic center to generate the core ESCF, which is the electrochemical state compensation coefficient representing the micro-polarization state. Based on this compensation coefficient, the weights are dynamically adjusted, and the output ASDT represents the adaptive safety judgment threshold. When the characteristic offset exceeds the limit, the instruction is sent to the PCI shown in the local lead-out box, representing the physical power flow isolation action that drives the physical contactor to disconnect. Under transformer area pollution, the data flow will point to the SFP, representing the derating conservative power outage protection strategy that triggers cut-off compensation.

[0099] ECS stands for Electric Bicycle Charging Station, representing the core physical carrier and execution scenario of this invention. It is a low-level hardware hub equipped with multi-dimensional feature joint calculation logic and is responsible for maintaining the charging network status.

[0100] REFS stands for Real-Time Electrical Environment Feature Stream, which represents the physical data containing real-time charging voltage and current parameters, serving as the raw acquisition stream input source for local sequence cleaning operations.

[0101] CWD represents the distortion degree of the charging waveform. It is a characteristic parameter representing the disordered state of charging obtained by using a sliding time window to extract and calculate the information entropy of the nonlinear difference sequence of the underlying feature flow, and removing the macro trend term.

[0102] HDSF stands for Hidden Dangers and Degradation Status Identifier. It means that the historical charging cycle data stream is input into a preset health prediction network, and after the historical charging and discharging parameters are mapped into a time-series degradation vector, an identifier that characterizes the aging tendency of the battery is output.

[0103] LFWS represents the background power grid fluctuation sequence of adjacent nodes. It indicates that the bus parameters of other adjacent non-local charging nodes in the same distribution area that are in a physically unloaded or constant current working state are received, and the time-domain energy envelope features are extracted to characterize the intensity of macro-global environmental interference.

[0104] HFRS stands for High Frequency Ripple Sequence of Local Charging Circuit, which represents the intrinsic AC current ripple response sequence superimposed on the power supply network by the pulse width modulation switching frequency inherent in the AC / DC power conversion stage of the local power supply circuit.

[0105] ESCF stands for Electrochemical State Compensation Coefficient. It represents the convergence state output based on the anti-disturbance warning sensitivity deviation, which is used to characterize the electrochemical microphysical state by performing spatial difference and frequency domain mapping on low-frequency and high-frequency sequences as inputs, differential companding of high-frequency ripple to cancel background harmonics, and using the low-frequency and high-frequency sequences as inputs.

[0106] ASDT stands for Adaptive Safety Judgment Critical Value. It represents the threshold value of the defense line after the compensation coefficient is injected into the environmental parameter weighting equation, the fusion ratio of the power waveform distortion degree and the hazard attenuation status indicator is dynamically adjusted, and the boundary is narrowed after joint calculation and reconstruction.

[0107] PCI stands for Physical Power Flow Isolation Action, which refers to a hard isolation hardware action structure in which the pulse blocking signal generated by the first control branch in response to the coordinated control command directly drives the physical contactor at the terminal of the power supply circuit to disconnect the power flow.

[0108] SFP stands for Degradation Conservative Power-Off Protection Strategy. It refers to the mechanism triggered when the statistical variance of the global background noise deviation of the distribution area continuously exceeds the preset divergence boundary. The compensation coefficient weight is forcibly set to zero, thus constructing a single-node conservative defense line that relies entirely on local feature cleaning data.

[0109] Figure 4Numerical verification aims to test the theoretical response characteristics of this invention under preset standardized boundary conditions (different combinations of disturbances and polarization abrupt changes, ranging from a clean power grid to extreme common-mode pollution). In the figure, the horizontal axis represents discrete standardized verification test scenarios; the left vertical axis represents the characteristic proportion of micro-phase angle hysteresis deviation calculated based on the input-defined common-mode fluctuation variance of the power grid and high-frequency probe demodulation; and the right vertical axis represents the adaptive safety judgment threshold calculated based on the nonlinear mapping rule and matrix calculations of this invention. The parallel bars represent the quantized input states of the power grid environmental disturbance and battery micro-polarization, respectively, and the superimposed solid line segments represent the corresponding calculated dynamic defense threshold. Figure 4 Numerical calculations clearly reveal the exponential convergence trend of the critical value in response to micro-polarization deterioration. The core approach of "nonlinear penalty multiplier and weighted zero-sum transfer" proposed in this method can effectively decouple environmental noise and target internal faults. In the standardized test scenario (corresponding to moderate grid disturbance and battery health state), the calculated critical value piecewise linearity remains within a high compensation range to shield against external common-mode interference. However, in the adjacent polarization abrupt change test scenario, when the column representing the phase angle hysteresis deviation shows a significant jump, the calculated adaptive safety judgment critical value piecewise linearity exhibits a steep downward slope shape through the determined matrix inner product operation logic. This quantitative abrupt change characteristic directly confirms that after introducing the electrochemical state compensation coefficient and performing zero-sum mapping, it is possible to forcibly deprive the tolerance of the external environment at the moment of encountering a deteriorated electrochemical abrupt change masked by external noise.

[0110] The computational logic involved in this application can be constructed using algorithms such as regression analysis in machine learning, establishing a mathematical model by analyzing the inherent trends and interrelationships of the collected parameters. This process can be implemented using specialized computational tools (such as Python's Scikit-learn library or the R language environment). Throughout all calculations, to eliminate the influence of different physical dimensions and ensure that data is compared and analyzed on the same scale, the input parameters in each formula are dimensionless. The dimensionless techniques used include, but are not limited to, max-min normalization or Z-score standardization.

[0111] To decouple the core algorithm from specific application strategies and ensure the configurability and ease of debugging of the technical solution, all configurable operating parameters in the specific implementation path of this invention are read through a standardized "configuration interface". The data source of this configuration interface is a "data storage module" (e.g., a non-transitory computer-readable storage medium, such as a configuration file, database entry, or cloud configuration service), which is configured to store configuration data in key-value pair format.

[0112] It should be emphasized that the foregoing embodiments are merely illustrative of preferred implementations of the present invention and are not intended to limit the scope of protection of the present invention. This application also provides a computer-readable storage medium having computer program instructions stored thereon.

Claims

1. A method for active fire prevention in electric bicycle charging stations, characterized in that, The specific steps include: S1: Obtain the underlying real-time electrical environment feature stream and perform local sequence cleaning operations to extract the charging waveform distortion degree; input the historical charging cycle data stream into the preset health prediction network to deduce the hidden danger decay status indicator; S2: Obtain the background grid fluctuation sequence of adjacent nodes representing adjacent non-local charging nodes, and obtain the high-frequency ripple sequence of local charging circuit representing the underlying power supply circuit of local charging port. S3: Using the background power grid fluctuation sequence of the adjacent nodes and the high-frequency ripple sequence of the local charging circuit as input, execute spatial differential and frequency domain mapping logic, and perform differential companding operation on the high-frequency ripple sequence of the local charging circuit based on the global background noise deviation of the transformer area extracted from the background power grid fluctuation sequence of the adjacent nodes, so as to generate electrochemical state compensation coefficients for characterizing the electrochemical microphysical state. S4: Based on the electrochemical state compensation coefficient, the original multi-dimensional feature joint calculation logic is weighted and compensated. The fusion ratio of the charging waveform distortion degree and the hidden danger attenuation state indicator is dynamically adjusted according to the electrochemical state compensation coefficient to generate an adaptive safety judgment threshold. S5: Perform high-frequency joint comparison and verification. If it is determined that the distortion of the charging waveform extracted in real time exceeds the adaptive safety judgment threshold, generate a collaborative control command to instruct the terminal node on the power supply circuit side to perform physical power flow isolation action, and trigger the derating conservative power-off protection strategy when the global background noise deviation of the transformer area exceeds the preset divergence boundary.

2. The active fire prevention method for electric bicycle charging stations according to claim 1, characterized in that: Acquire the underlying real-time electrical environment feature stream and perform local sequence cleaning operations to extract the power waveform distortion, including: Extract the real-time charging voltage and current parameters from the underlying real-time electrical environment feature stream; The real-time charging voltage and current parameters are extracted sequentially using a sliding time window, and the information entropy of the nonlinear differential sequence within the sliding time window is calculated. The information entropy of the nonlinear difference sequence after removing the macro trend term is determined as the distortion degree of the charging waveform; Historical charging cycle data streams are input into a pre-set health prediction network to deduce the status indicators of potential hazard decay, including: The historical charge-discharge cycle parameters of the associated business identifier are mapped to a time-series decay vector, and the potential decay status identifier, which characterizes the internal aging tendency of the battery, is output based on the health prediction network.

3. The active fire prevention method for electric bicycle charging stations according to claim 2, characterized in that: Obtain the background grid fluctuation sequence of adjacent non-local charging nodes, including: receiving the bus electrical parameter sequence of other charging nodes in the same distribution area that are in a physically unloaded or constant current operating state; Obtain the high-frequency ripple sequence of the local charging circuit, which characterizes the underlying power supply circuit of the local charging port, including: Obtain the intrinsic AC current ripple response sequence excited by the inherent pulse width modulation switching frequency of the AC / DC power conversion stage and superimposed on the output power supply network.

4. The active fire prevention method for electric bicycle charging stations according to claim 3, characterized in that: Using the background grid fluctuation sequence of the adjacent nodes and the high-frequency ripple sequence of the local charging circuit as inputs, spatial difference and frequency domain mapping logic is executed to generate electrochemical state compensation coefficients, including: The temporal energy envelope features of the background power grid fluctuation sequence of the adjacent nodes are extracted, and the global background noise deviation of the transformer area, which characterizes the intensity of global environmental interference, is generated through an asynchronous alignment mechanism. Using the global background noise deviation of the transformer area as a dynamic constraint, differential companding is performed on the gain parameter of the high-frequency ripple sequence of the local charging circuit to cancel the background grid harmonic interference mixed in the high-frequency ripple sequence of the local charging circuit, thereby obtaining a high-purity differential mode ripple sequence. Perform discrete frequency domain phase angle demodulation operation to extract the frequency domain phase angle offset features between the high-purity differential mode ripple sequence and the original excitation pulse width modulation signal; The anti-interference warning sensitivity deviation is calculated in real time, and the frequency domain phase angle offset feature is used to update the anti-interference warning sensitivity deviation. The electrochemical state compensation coefficient is output based on the convergence state of the anti-interference warning sensitivity deviation. The anti-interference warning sensitivity deviation is configured as a detection resolution gain index characterizing the fused purified waveform for weak electrochemical polarization impedance.

5. The active fire prevention method for electric bicycle charging stations according to claim 4, characterized in that: Based on the electrochemical state compensation coefficient, the fusion ratio of the charging waveform distortion degree and the hidden danger attenuation state indicator is dynamically adjusted to generate an adaptive safety judgment threshold, including: The electrochemical state compensation coefficient is used as a multiplication penalty factor and injected into the environmental parameter weighting equation in the multidimensional feature joint calculation logic. In response to the impedance degradation trend indicated by the electrochemical state compensation coefficient, the compensation margin weight of the environmental parameter is reduced proportionally, and the basic risk weight of the hazard attenuation state indicator is simultaneously amplified. The adaptive safety judgment threshold value after narrowing the boundary is calculated and output.

6. The active fire prevention method for electric bicycle charging stations according to claim 5, characterized in that: Generate coordinated control commands to instruct the terminal nodes on the power supply circuit side to perform physical power flow isolation actions, and trigger a derating conservative power-off protection strategy when the global background noise deviation of the transformer area exceeds a preset divergence boundary, including: When the coordinated control command is issued, the first control branch generates a pulse blocking signal to drive the physical contactor to disconnect the power flow, and the second logic branch synchronously captures the distortion characteristics of the charging waveform at the moment of breaking the critical value and resets the long-term baseline parameters of the health prediction network. When the statistical variance of the global background noise deviation of the substation area continuously exceeds the preset divergence boundary, the output weight of the electrochemical state compensation coefficient is forcibly set to zero, and the adaptive safety judgment threshold is reconstructed into a single-node conservative defense limit value that is completely dependent on the local feature cleaning data.

7. The active fire prevention method for electric bicycle charging stations according to claim 6, characterized in that: Updating the anti-interference warning sensitivity deviation using the frequency domain phase angle offset feature includes: Extract the frequency domain phase angle offset features within the current operation cycle, and obtain the smoothed reference phase angle feature quantity within the historical time window; Calculate the time difference gradient value between the frequency domain phase offset feature and the reference phase feature; The time difference gradient value is mapped to a dynamic update step size parameter between zero and one, wherein the dynamic update step size parameter has a non-linear positive correlation with the time difference gradient value. The current weight of the frequency domain phase offset feature is calculated using the dynamic update step size parameter, and the historical forgetting weight is calculated using the difference between the numerical value and the dynamic update step size parameter. The first product is obtained by multiplying the current weight by the frequency domain phase offset feature, and the second product is obtained by multiplying the historical forgetting weight by the anti-interference warning sensitivity deviation of the previous cycle. The first product is added to the second product to iteratively update the anti-interference warning sensitivity deviation in the current operation cycle.

8. The active fire prevention method for electric bicycle charging stations according to claim 7, characterized in that: The electrochemical state compensation coefficient is used as a multiplication penalty factor and injected into the environmental parameter weighting equation in the multidimensional feature joint calculation logic. The compensation margin weight of the environmental parameters is proportionally reduced, while the basic risk weight of the hazard attenuation state indicator is simultaneously amplified, including: The electrochemical state compensation coefficient is input into a preset nonlinear smoothing mapping logic, and the output is a penalty multiplier between the lower limit value of zero and the upper limit value of one. Obtain the initial compensation margin weight in the environmental parameter weighting equation, and multiply the compensation margin weight by the penalty multiplier to obtain the stripping weight attenuation.

9. The active fire prevention method for electric bicycle charging stations according to claim 8, characterized in that: Subtract the stripping weight attenuation from the compensation margin weight to obtain the actual environmental compensation weight after shrinkage; The initial allocation of the basic risk weight is obtained, and the stripping weight decay is added to the basic risk weight as a zero-sum transfer compensation term to obtain the expanded actual risk decision weight; wherein, the sum of the values ​​of the actual environmental compensation weight and the actual risk decision weight is equal to the sum of the compensation margin weight and the initial allocation value of the basic risk weight. A normalized parameter fusion matrix is ​​constructed using the actual environmental compensation weight and the actual risk decision weight to generate the adaptive safety judgment threshold.

Images