Method for evaluating and monitoring state of health of battery of electric vehicle

Abstract

The invention provides an electric vehicle battery health state assessment monitoring method, which comprises the following steps: carrying out image mapping processing on a protection plate depression degree index and a bracket stress concentration point displacement estimation value by adopting a convolutional neural network, and converting a protection plate deformation parameter into a two-dimensional thermodynamic diagram representation form, obtaining a correlation characteristic graph of stress transmission between the protection plate and the support; analyzing a time sequence change mode in the correlation characteristic graph of stress transmission between the protection plate and the support, predicting the accumulated influence of the deepened depression of the protection plate on the displacement of the support, if the predicted displacement exceeds a preset threshold value, triggering an adjustment signal, and judging the updating demand of a dynamic detection strategy; and obtaining reference parameters in a similar road surface bulge scene from a historical impact data set according to the adjustment signal, fusing the current vehicle speed and impact strength data, and determining an optimized sensor sampling frequency and a stress monitoring focus area.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for assessing and monitoring the health status of electric vehicle batteries. Background Technology

[0002] As a core representative of new energy vehicles, the safety of the battery system in electric vehicles directly affects the reliability of vehicle operation and the safety of occupants. The battery pack, as the energy core of an electric vehicle, withstands complex external environmental impacts. Especially when the vehicle passes over obstacles such as speed bumps, manhole covers, or road damage, the battery pack may face bottom impact, leading to structural deformation or internal component failure. Monitoring the structural integrity of the battery pack under road bump impacts is not only crucial for extending battery life but also for ensuring overall vehicle safety and reducing accident risks. Currently, battery pack structural health monitoring mainly relies on static testing or limited real-time sensor data analysis, but these methods have significant limitations when dealing with dynamic road impact scenarios. Traditional monitoring systems often focus on stress distribution or deformation analysis under single operating conditions, making it difficult to adapt to impact changes at different speeds and on different road surfaces. For example, when a vehicle passes over a speed bump at different speeds, the force distribution at the bottom of the battery pack changes significantly, and existing methods struggle to capture these transient changes in real time, resulting in insufficient accuracy of monitoring results. When the battery pack is impacted by a road bump, the degree of local indentation of the bottom protective plate varies with the impact intensity and frequency. Such indentations not only affect the integrity of the battery pack casing but may also transmit stress to the fixing brackets between internal modules, causing displacement of stress concentration points on the brackets. Furthermore, deeper indentations in the protective plate lead to uneven stress distribution, which in turn affects the stability of the fixing brackets. This uneven stress distribution may cause fatigue damage to the brackets during long-term use, and even lead to displacement or connection failure of the battery modules. For example, when a vehicle frequently passes over speed bumps, the repeated deformation of the protective plate may cause the stress concentration points on the brackets to gradually shift, making it difficult for existing monitoring systems to dynamically adjust the detection focus, thus hindering accurate assessment of the battery pack's mechanical safety. Summary of the Invention

[0003] This invention provides a method for assessing and monitoring the health status of electric vehicle batteries, comprising:

[0004] S101. Real-time impact signal data is collected by a multi-axis accelerometer and strain sensor array deployed at the bottom of the battery pack. When the vehicle passes over a road bump, the deformation parameters of the protective plate area and the stress value changes of the internal fixed bracket are recorded simultaneously to obtain the original impact dataset.

[0005] S102. Based on the original impact dataset, Fourier transform is applied to extract frequency domain features and time domain peak features. Low-frequency components and high-frequency impact peaks related to the indentation of the protective plate are separated to determine the indentation degree index of the protective plate and the estimated displacement value of the stress concentration point of the support.

[0006] S103. A convolutional neural network is used to perform image mapping processing on the indentation degree index of the protective plate and the displacement estimate of the stress concentration point of the support. The deformation parameters of the protective plate are converted into a two-dimensional thermogram representation to obtain the correlation feature map of stress transmission between the protective plate and the support.

[0007] S104. Analyze the temporal change pattern in the correlation feature diagram of stress transmission between the protective plate and the support, predict the cumulative impact of the deepening of the protective plate indentation on the support displacement, and trigger an adjustment signal if the predicted displacement exceeds the preset threshold to determine the update requirements of the dynamic detection strategy.

[0008] S105. Based on the adjustment signal, obtain reference parameters for similar road surface protrusion scenarios from the historical impact dataset, fuse the current vehicle speed and impact intensity data, and determine the optimized sensor sampling frequency and stress monitoring focus area.

[0009] S106. The sensor array acquisition mode is reconfigured for the optimized sensor sampling frequency and monitoring focus area. In subsequent road impact events, an adaptive filtering algorithm is applied to capture transient deformation data and calculate the battery pack mechanical health assessment index that integrates multi-source information.

[0010] S107. A graded early warning mode is established based on the mechanical health assessment index of the battery pack. When the risk of bracket displacement is detected to reach the level two early warning standard, a structural safety reminder signal is generated and transmitted to the vehicle diagnostic system and instrument display module through the CAN bus to realize real-time monitoring and active protection of the electric vehicle battery health status.

[0011] Preferably, step S101 includes: acquiring the acceleration signal output by a multi-axis accelerometer array deployed on the bottom protective plate of the battery pack; processing the high-frequency noise components in the acceleration signal that are higher than a preset frequency threshold using a Butterworth low-pass filter; identifying the start time of the road surface protrusion impact event based on the change in the filtered acceleration amplitude; and recording the start time marker and triggering the data acquisition mode when the acceleration amplitude exceeds a preset impact threshold. After triggering the data acquisition mode, strain signals are read from the strain sensor array on the surface of the protective plate; the local deformation parameters of each measuring point of the protective plate are calculated based on the linear relationship between the resistance change rate of the strain gauge and the strain; the stress sensor output signals of the internal fixed support measuring points are acquired simultaneously; and the stress value change data of each measuring point of the support is obtained by converting the sensor output voltage and stress calibration curve. The start time marker is used to perform time-series alignment processing on the local deformation parameters and the stress value change data; the deformation parameters of each measuring point are combined into a protective plate deformation distribution matrix based on the spatial coordinates of the sensor array; and the stress values ​​of each measuring point are combined into a support stress distribution matrix. The original impact dataset containing spatiotemporal information is obtained through matrix splicing processing.

[0012] Preferably, step S102 includes: performing a Fast Fourier Transform on the deformation data of the protective plate in the original impact dataset to decompose the time-domain signal into a superposition of sinusoidal waves with different frequency components, obtaining the amplitude spectrum and phase spectrum in the frequency domain, identifying the dominant frequency component based on the energy magnitude of each frequency component in the amplitude spectrum, and classifying the frequency component below a preset low-frequency threshold as a low-frequency component of the slow deformation of the protective plate, thus obtaining a frequency domain feature vector related to the indentation of the protective plate. Bandpass filtering is performed on the low-frequency components in the frequency domain feature vector to retain the frequency bands related to the inherent frequency of the protective plate structure, and the root mean square value of the filtered signal is calculated as a deformation energy index. The degree of indentation of the protective plate is determined based on the relationship between the deformation energy index and the elastic modulus of the protective plate material. The stress peak value and its corresponding time are extracted from the support stress data in the original impact dataset. The stress propagation velocity is calculated by the ratio of the difference in the occurrence time of the stress peak value of adjacent measuring points to the distance between the measuring points, and the damage state of the support material is determined based on the stress propagation velocity. Based on the indentation degree index of the protective plate and the stress propagation speed, the correlation between the indentation degree and the stress distribution of the support is established by linear regression. When the stress value of a certain measuring point exceeds the preset stress concentration threshold, the spatial offset of the measuring point relative to the initial position under no-impact conditions is calculated to obtain the estimated displacement value of the stress concentration point of the support.

[0013] Preferably, step S103 includes: establishing a two-dimensional coordinate system according to the physical arrangement of the sensor array based on the indentation degree index of the protective plate and the estimated displacement value of the stress concentration point of the support; interpolating the values ​​between discrete measurement points using bilinear interpolation; generating a continuously distributed numerical field on the regular grid points of the coordinate system to obtain the deformation distribution matrix of the protective plate and the displacement distribution matrix of the support. The deformation distribution matrix of the protective plate and the displacement distribution matrix of the support are normalized respectively, and mapped to a preset grayscale value range according to the normalized values. A heatmap of the deformation of the protective plate and a heatmap of the displacement of the support are generated through the correspondence between grayscale values ​​and pseudo-color. The pixel values ​​of the two heatmaps are fused using a weighted average method to obtain fused two-dimensional image data. A convolutional neural network is used to extract features from the two-dimensional image data. A pre-trained convolutional kernel is used to perform convolution operations on local areas of the image to extract spatial correlation features between the deformation mode of the protective plate and the displacement mode of the support. Max pooling is used in the pooling layer to reduce the dimension of the feature map while retaining significant features. Based on the spatially relevant features, the spatial resolution of the feature map is restored layer by layer through transposed convolution. After each layer of transposed convolution, a skip connection is made with the convolution features of the corresponding layer to fuse multi-scale spatial information and output a correlation feature map that reflects the stress transmission path between the protective plate and the support.

[0014] Preferably, step S104 includes: acquiring the associated feature map sequence of stress transfer between the protective plate and the support; arranging the feature map data generated by continuous impact events in chronological order; segmenting the time-series data using a fixed window size; calculating the difference between pixel values ​​at corresponding positions of the feature maps at adjacent times within each window; and obtaining a temporal difference matrix characterizing the rate of feature change. A long short-term memory network is used to process the temporal difference matrix, selectively retaining and forgetting historical information through gated recurrent units, encoding the cumulative effect of the protective plate indentation in the hidden layer state vector, and outputting a predicted value of the support displacement at the next time moment based on a nonlinear combination of the current input and historical state vectors. The predicted value is accumulated in a time sequence, with displacement increments added moment by moment starting from the initial state. When the accumulated displacement exceeds a preset threshold for allowable material deformation, a binary adjustment signal is generated. Based on the trigger time of the adjustment signal, the slope of displacement growth within a preset time period before triggering is calculated. If the slope exceeds a preset multiple of the historical average growth rate, the current detection sensitivity is determined to be insufficient, and a detection strategy update flag is output.

[0015] Preferably, step S105 includes: retrieving data records under the same time period and road conditions from the historical impact dataset based on the trigger time of the adjustment signal; measuring the similarity between the current protective plate deformation distribution and historical records using Euclidean distance; filtering out similar scenarios with distance values ​​less than a preset threshold; and extracting the sensor sampling rate and spatial resolution parameters in the similar scenarios as a reference configuration. Real-time readings from the vehicle speed sensor and the impact peak value measured by the accelerometer are obtained; the product of the square of the velocity and the impact peak value is calculated as the impact energy parameter; the reference sampling rate is linearly adjusted based on the ratio of the impact energy parameter to the standard energy value in the reference configuration to obtain a sensor sampling frequency adapted to the current impact intensity. By statistically analyzing the trigger counts of each sensor location in historical damage events, a Gaussian kernel function is used to smooth discrete damage points, generating a continuous damage probability density distribution map; and based on the boundary of the region where the probability density exceeds a preset threshold and the sensor sampling frequency, the optimized stress monitoring focus area and corresponding sampling configuration parameters are determined.

[0016] Preferably, step S106 includes: based on the optimized sensor sampling frequency and the set of coordinates of the monitoring focus area, sending a configuration command to the sensor array via a serial communication protocol to modify the sampling period of the sensors in the focus area to a high-frequency mode and the sampling period of the sensors in the non-focus area to a low-frequency mode, thus completing the partition configuration of the acquisition mode. When the road impact acceleration is detected to exceed a preset trigger threshold, a Kalman filter is used to process the sensor signals after the partition configuration. Through the recursive operation of the prediction equation and the observation equation, the noise component and the real signal are separated, and the transient characteristic data of the protective plate deformation are extracted. The deformation peak value, rise time, and recovery time parameters in the transient characteristic data are obtained. At the same time, the reference deformation value at the same location is read from the historical database, and the deviation ratio between the current deformation and the reference value is calculated as the local damage degree. Based on the distribution of the local damage degree in each area of ​​the protective plate, the overall damage degree value is calculated using the area-weighted average method. If the overall damage degree is less than the preset safety threshold, the output health level is excellent; if it is between the safety threshold and the warning threshold, the output health level is good; if it exceeds the warning threshold, the output health level is poor, thus obtaining a quantified mechanical health evaluation index for the battery pack.

[0017] Preferably, step S107 includes: establishing a graded early warning system based on the battery pack mechanical health assessment indicators, setting a normal state corresponding to a health level higher than a first threshold, a slight risk corresponding to a health level between the first and second thresholds, and a moderate risk corresponding to a health level lower than the second threshold; generating a secondary early warning indicator when the health level drops below the second threshold. According to the generation process of the structural safety reminder signal triggered by the secondary early warning indicator, the maximum displacement value and displacement rate are extracted from the bracket displacement monitoring data, and the displacement information is converted into a diagnostic fault code according to a preset encoding rule. A timestamp and sensor position identifier are added to form a complete reminder signal data packet. The reminder signal data packet is encapsulated into a standard frame format via the CAN bus protocol stack, the priority bit of the message identifier is set to high priority, and it is transmitted to the vehicle diagnostic system via the vehicle network physical layer for fault recording. Simultaneously, it is transmitted to the instrument display module to trigger the warning light to illuminate and text prompts, realizing real-time alarm and driver reminder for abnormal battery health status.

[0018] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0019] This invention discloses a method for assessing and monitoring the health status of electric vehicle batteries. Addressing the stress transmission and deformation issues of the battery pack's protective plate and internal supports when a vehicle passes over road bumps, the method collects real-time impact data using a multi-axis accelerometer and strain sensor array. Fourier transform is applied to separate low-frequency indentation and high-frequency impact features. A convolutional neural network is used to map deformation parameters into a two-dimensional heatmap, extracting stress transmission correlation features. A recurrent neural network further analyzes temporal changes, predicts support displacement risks, triggers adjustment signals to optimize sensor sampling frequency and monitoring focus areas, and applies an adaptive filtering algorithm to capture high-precision transient deformation data. The method calculates battery pack mechanical health assessment indicators, establishes a tiered early warning mode, and transmits the data to the vehicle diagnostic system via a CAN bus, achieving real-time monitoring and proactive protection. This invention significantly improves battery pack structural safety and vehicle operational reliability through multi-source data fusion and intelligent prediction. Attached Figure Description

[0020] Figure 1 This is a flowchart of a method for assessing and monitoring the health status of an electric vehicle battery according to the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0022] like Figure 1 This embodiment of a method for assessing and monitoring the health status of an electric vehicle battery may specifically include:

[0023] S101. Real-time impact signal data is collected by a multi-axis accelerometer and strain sensor array deployed at the bottom of the battery pack. When the vehicle passes over a road bump, the deformation parameters of the protective plate area and the stress value changes of the internal fixed bracket are recorded simultaneously to obtain the original impact dataset.

[0024] Acceleration signals from a multi-axis accelerometer array deployed on the bottom protective plate of the battery pack are acquired. High-frequency noise components exceeding a preset frequency threshold in the acceleration signals are processed using a Butterworth low-pass filter. The start time of the road surface protrusion impact event is identified based on the change in the filtered acceleration amplitude. When the acceleration amplitude exceeds a preset impact threshold, the start time is recorded and a data acquisition mode is triggered. After triggering the data acquisition mode, strain signals are read from the strain sensor array on the surface of the protective plate. Local deformation parameters at each measuring point of the protective plate are calculated based on the linear relationship between the resistance change rate of the strain gauge and the strain. Simultaneously, stress sensor output signals from the internal fixed support measuring points are acquired. Stress value change data at each measuring point of the support are obtained by converting the sensor output voltage and stress calibration curves. The start time marker is used to perform time-series alignment processing on the local deformation parameters and the stress value change data. Based on the spatial coordinates of the sensor array, the deformation parameters of each measuring point are used to form a deformation distribution matrix for the protective plate, and the stress values ​​of each measuring point are used to form a stress distribution matrix for the support. The original impact dataset containing spatiotemporal information is obtained through matrix stitching.

[0025] In one embodiment, a multi-axis accelerometer array is deployed in a 4×4 grid layout on the inner surface of the bottom protective plate of the battery pack, with each sensor node spaced 200 mm apart, covering the main stress-bearing areas of the protective plate. When the vehicle encounters a road bump during operation, the protective plate first bears the vertical impact load, and the accelerometers collect triaxial acceleration signals in real time.

[0026] Specifically, the Butterworth low-pass filter employs a fourth-order design, achieving a flat passband response through cascaded second-order sections. The filter's transfer function monotonically decreases at the cutoff frequency, effectively suppressing high-frequency vibration noise while preserving the main frequency components of road impact. When the filtered vertical acceleration amplitude exceeds a preset impact threshold, this moment is recorded as the start time marker of the impact event, which is used for subsequent time synchronization of multi-source sensor data.

[0027] It should be noted that the strain sensor uses a metal foil strain gauge, whose resistance change rate is linearly related to the strain. A Wheatstone bridge circuit converts the minute resistance change into a measurable voltage signal. Strain sensors are arranged on the surface of the protective plate according to the principle of dense arrangement in stress concentration areas and sparse arrangement in edge areas, focusing on monitoring the deformation in the central area of ​​the protective plate and near the reinforcing ribs.

[0028] The stress sensor uses a piezoresistive sensor. A calibration curve of the output voltage and stress value is established in advance through standard load tests. This curve is obtained through polynomial fitting to ensure good linearity within the working range.

[0029] For example, the deformation distribution matrix of the protective plate is constructed according to the physical location coordinates of the sensors. The row and column indices of the matrix correspond to the grid positions of the sensors on the protective plate, and the matrix element values ​​are the deformation parameters at the corresponding positions. The stress distribution matrix of the support is constructed using a similar method, reflecting the stress state at each measuring point of the fixed support. Through matrix concatenation, the two matrices are linked in the time dimension to form an original impact dataset containing complete information on the deformation of the protective plate and the stress of the support. This dataset preserves the spatiotemporal characteristics of the impact event.

[0030] S102. Based on the original impact dataset, Fourier transform is applied to extract frequency domain features and time domain peak features. Low-frequency components and high-frequency impact peaks related to the dent of the protective plate are separated to determine the dent degree index of the protective plate and the estimated displacement value of the stress concentration point of the support.

[0031] Fast Fourier Transform (FFT) is performed on the deformation data of the protective plate in the original impact dataset to decompose the time-domain signal into a superposition of sinusoidal waves with different frequency components. The amplitude and phase spectra in the frequency domain are obtained. The dominant frequency components are identified based on the energy magnitude of each frequency component in the amplitude spectrum. Frequency components below a preset low-frequency threshold are classified as low-frequency components of slow deformation of the protective plate, resulting in a frequency domain feature vector related to the plate's indentation. Bandpass filtering is applied to the low-frequency components in the frequency domain feature vector, retaining frequency bands related to the inherent frequencies of the protective plate structure. The root mean square (RMS) value of the filtered signal is calculated as a deformation energy index. The degree of indentation is determined based on the relationship between the deformation energy index and the elastic modulus of the protective plate material. The stress peak value and its corresponding time are extracted from the stress data of the support in the original impact dataset. The stress propagation velocity is calculated by the ratio of the difference in the time of occurrence of the stress peak value between adjacent measuring points to the distance between the measuring points. The damage state of the support material is determined based on the stress propagation velocity. Based on the indentation degree index of the protective plate and the stress propagation speed, the correlation between the indentation degree and the stress distribution of the support is established by linear regression. When the stress value of a certain measuring point exceeds the preset stress concentration threshold, the spatial offset of the measuring point relative to the initial position under no-impact conditions is calculated to obtain the estimated displacement value of the stress concentration point of the support.

[0032] In one embodiment, the Fast Fourier Transform (FFT) converts the continuous deformation signal of the protective panel into a finite-length digital sequence through discrete sampling. It employs a butterfly arithmetic unit to recursively decompose the calculation process, breaking down the N-point Discrete Fourier Transform into multiple 2-point basic operations, thus reducing the computational complexity from N... 2 The deformation data of the protective plate is reduced to NlogN. Before performing a Fourier transform, the data is first zero-mean normalized to eliminate the influence of the DC component on the spectrum analysis. The complex spectrum obtained after the transform contains real and imaginary parts. The amplitude spectrum is obtained by calculating the complex modulus, which reflects the magnitude of the amplitude of each frequency component. The phase spectrum is obtained by calculating the ratio of the real to the imaginary parts using the arctangent function.

[0033] Specifically, the amplitude spectrum in the frequency domain exhibits distinct energy distribution characteristics. The low-frequency band primarily reflects the overall bending deformation of the protective plate. This slow deformation is typically caused by static or quasi-static loads, with frequencies concentrated between 0.1 Hz and 10 Hz. When a vehicle passes over a speed bump, the protective plate first undergoes an overall indentation. This macroscopic deformation manifests as a concentration of energy in the low-frequency components of the spectrum. The preset low-frequency threshold is determined based on the natural frequency of the protective plate material. The natural frequency of aluminum alloy protective plates is typically in the range of 15 Hz to 30 Hz; therefore, frequency components below 15 Hz are classified as low-frequency components related to the indentation deformation.

[0034] The bandpass filter employs a Butterworth design, with the passband range set between 0.5 and 1.5 times the first-order natural frequency of the protective plate to ensure the preservation of frequency components closely related to the structural response. The filter order is fourth-order, achieving a balance between passband flatness and transition band steepness. Zero-phase filtering is used, first performing forward filtering, then reversing the signal for inverse filtering to eliminate the influence of phase distortion on the signal waveform. The filtered signal retains the vibrational characteristics of the protective plate near its natural frequency, which directly reflect the degree of elastic deformation of the protective plate.

[0035] It should be noted that the deformation energy index is obtained by calculating the root mean square (RMS) value of the filtered signal, which reflects the effective value of the signal. The elastic modulus of the protective plate material determines the amount of deformation under the same load; the larger the elastic modulus, the smaller the deformation produced by the same energy input. Through a pre-established calibration relationship, the deformation energy index is converted into the actual indentation depth value. The calibration process is carried out in a laboratory environment, applying concentrated loads of different degrees to the protective plate while simultaneously measuring the indentation depth and the corresponding deformation energy. A mathematical relationship between the two is established through polynomial fitting. The indentation degree index of the protective plate is defined as the ratio of the measured indentation depth to the thickness of the protective plate. This dimensionless index facilitates the assessment of the relative damage degree of protective plates with different thicknesses.

[0036] The processing of scaffold stress data focuses on peak feature extraction. The stress time history curve at each measuring point exhibits pulse-like variation characteristics, and valid peak values ​​are identified by setting a dynamic threshold. The dynamic threshold is determined based on the moving standard deviation of the signal, with three times the standard deviation used as the peak identification threshold. The amplitude and corresponding time of each peak are recorded to form a peak-time sequence. There is a time delay between stress peak values ​​at adjacent measuring points; this delay reflects the propagation process of stress waves in the scaffold material.

[0037] For example, the stress propagation velocity is calculated based on the cross-correlation analysis of the peak times of adjacent measuring points. A peak time sequence of a reference measuring point is selected, and a sliding correlation calculation is performed with the sequences of adjacent measuring points. The time offset corresponding to the maximum correlation coefficient is the time delay of stress wave propagation. The physical distance between the measuring points is known; the stress propagation velocity is obtained by dividing the distance by the time delay. In intact scaffold materials, the stress wave propagation velocity is close to the theoretical speed of sound of the material. When the scaffold develops microcracks or fatigue damage, the propagation velocity will decrease significantly. By comparing the deviation between the measured propagation velocity and the theoretical value, the degree of damage to the scaffold is determined.

[0038] Understandably, the linear regression method establishes a mathematical relationship between the degree of indentation of the protective plate and the stress distribution of the support, which can be expressed as follows:

[0039] σ i =β0+β1D+ε i

[0040] This is the basic form of a linear regression model, where σ i Let denot i represent the stress value at the i-th measuring point, D represent the degree of indentation of the protective plate, β0 represent the regression intercept, β1 represent the regression coefficient, and ε represent the stress value at the i-th measuring point. i This represents the random error term. The independent variable of the regression model is the degree of dent in the protective plate, and the dependent variable is the stress value at each measuring point. The regression coefficients are solved using the least squares method to minimize the sum of squared residuals between the predicted and measured stress values. The coefficient of determination R of the regression model is... 2 Used to evaluate goodness of fit, when Rfit 2 A value greater than 0.8 indicates the model has good predictive ability. Furthermore, stress concentration is determined based on the strength theory of materials mechanics. A preset stress concentration threshold is set at 0.7 times the material's yield strength, leaving a safety margin. When the stress value at a measuring point exceeds this threshold, it is marked as a stress concentration point. The displacement vector is calculated by comparing the spatial coordinates of the stress concentration points before and after the impact. The displacement estimation considers the geometric deformation effect caused by the dent in the protective plate. Coordinate transformation converts the displacement in the local coordinate system to the global coordinate system, obtaining the absolute displacement estimate of the stress concentration point on the support. This displacement estimation provides a quantitative indicator for assessing the integrity of the battery pack's internal structure; when the displacement exceeds the allowable range, it indicates a potential risk of mechanical failure.

[0041] S103. A convolutional neural network is used to perform image mapping processing on the indentation degree index of the protective plate and the displacement estimate of the stress concentration point of the support. The deformation parameters of the protective plate are converted into a two-dimensional heat map representation, and the correlation feature map of stress transmission between the protective plate and the support is obtained.

[0042] Based on the indentation degree index of the protective plate and the estimated displacement of the stress concentration points of the support, a two-dimensional coordinate system is established according to the physical arrangement of the sensor array. Bilinear interpolation is used to interpolate the values ​​between discrete measurement points, generating a continuously distributed numerical field on the regular grid points of the coordinate system, thus obtaining the deformation distribution matrix of the protective plate and the displacement distribution matrix of the support. The deformation distribution matrix of the protective plate and the displacement distribution matrix of the support are normalized respectively, and the normalized values ​​are mapped to a preset grayscale range. A heatmap of the protective plate deformation and a heatmap of the support displacement are generated based on the correspondence between grayscale values ​​and pseudo-color values. A weighted average method is used to fuse the pixel values ​​of the two heatmaps to obtain the fused two-dimensional image data. A convolutional neural network is used to extract features from the two-dimensional image data. A pre-trained convolutional kernel is used to perform convolution operations on local regions of the image to extract the spatial correlation features between the deformation mode of the protective plate and the displacement mode of the support. Max pooling is used in the pooling layer to reduce the dimensionality of the feature map while retaining salient features. Based on the spatially relevant features, the spatial resolution of the feature map is restored layer by layer through transposed convolution. After each layer of transposed convolution, a skip connection is made with the convolution features of the corresponding layer to fuse multi-scale spatial information and output a correlation feature map that reflects the stress transmission path between the protective plate and the support.

[0043] In one embodiment, the spatialization of the protective plate indentation index and the estimated displacement of the support stress concentration point is based on the actual arrangement of the sensor array. The sensors are deployed in a non-uniform distribution on the bottom of the protective plate, with a dense arrangement in the central region to capture detailed changes in the main stress areas and a sparse arrangement at the edges to cover a wider area. The established two-dimensional coordinate system has the geometric center of the protective plate as its origin, with the horizontal axis along the vehicle's driving direction and the vertical axis perpendicular to the driving direction. The position of each sensor is uniquely determined by its physical coordinates.

[0044] Specifically, the bilinear interpolation method is performed within a rectangular region formed by the four nearest neighbor measurement points. For each point to be interpolated, the rectangular cell containing it is first determined, and then weighting coefficients are calculated based on the point's relative position to the four vertices. The interpolation calculation is performed in two steps: first, two linear interpolations are performed along the horizontal axis to obtain the interpolation results at the upper and lower boundaries; then, a single linear interpolation is performed along the vertical axis to obtain the interpolated value. This method ensures the continuity of the interpolation function along the rectangular boundaries, avoiding abrupt changes in the numerical field. By performing interpolation operations at each node of the regular grid, discrete sensor measurement data are converted into a continuous spatial distribution field. The dimension of the protective plate deformation distribution matrix is ​​determined according to the required spatial resolution; a typical grid size is 50×50, corresponding to 2500 sampling points on the protective plate surface.

[0045] It should be noted that the normalization process uses a min-max normalization method, mapping the original values ​​to a standard range of 0 to 1. The maximum value of the protective plate deformation typically corresponds to the most severe dent, while the minimum value corresponds to the undeformed area. The normalized values ​​form a linear correspondence with gray levels, with 0 corresponding to pure black, 1 to pure white, and intermediate values ​​distributed proportionally across 256 gray levels. The pseudo-color mapping uses a color scheme commonly used in heatmaps, transitioning from dark blue to high values, through cyan, green, and yellow, to dark red. This color gradient visually reflects the spatial distribution characteristics of the values.

[0046] The two heatmaps were fused using a weighted average method, with weighting coefficients determined based on the contribution of the protective plate deformation and support displacement to overall safety. Protective plate deformation primarily reflects the direct impact of external shocks, with a weight of 0.6, while support displacement reflects the response of the internal structure, with a weight of 0.4. The fusion calculation was performed at the pixel level, with the fused value for each pixel equal to the weighted sum of the corresponding pixel values ​​from the two heatmaps. The fused two-dimensional image retains the original spatial resolution while comprehensively reflecting the state information of both the protective plate and the support.

[0047] Convolutional neural networks employ a classic feature extraction architecture, consisting of alternating combinations of multiple convolutional and pooling layers. Pre-trained convolutional kernels are derived from training on large-scale image datasets, having learned to detect basic visual features such as edges, textures, and shapes. The first convolutional layer uses 64 3×3 kernels with a stride of 1, padding to maintain the feature map size. Convolution operations are implemented using a sliding window mechanism; the output at each window position equals the inner product of the convolutional kernel and the corresponding image region plus a bias term. The activation function uses a modified linear unit (MRU), introducing non-linear transformation capabilities. Through the stacking of multiple convolutional layers, the network extracts feature representations from low to high levels layer by layer. Lower-level convolutions respond to local deformation patterns, while higher-level convolutions capture global stress propagation paths.

[0048] For example, max pooling selects the maximum value as the output within each 2×2 region, halving the spatial size of the feature map. This downsampling operation not only reduces computational complexity but also enhances the spatial invariance of the features. The pooled feature map retains the most salient activation responses while filtering out minor details. After three pooling operations, the spatial resolution of the feature map is reduced to 1 / 8 of the original, but the number of channels increases to 512, forming a high-dimensional feature representation.

[0049] Understandably, the transposed convolution operation gradually restores the spatial resolution of the feature map through a learned upsampling kernel. The kernel parameters of the transposed convolution are learned through backpropagation, ensuring that the reconstructed feature map retains key spatial information. Each transposed convolution doubles the size of the feature map, and after three transposed convolutions, the original resolution is restored. The skip connection mechanism concatenates feature maps from the encoding path with feature maps of corresponding scales from the decoding path. This connection method allows high-resolution detail information to be directly transmitted to the output layer, avoiding information loss in deep networks. Furthermore, the fusion of multi-scale spatial information is achieved through feature map concatenation. Feature maps from different levels contain spatial patterns of different scales. Shallow features retain details of local deformation of the protective plate, while deep features encode the global distribution of stress in the support structure. The concatenation operation is performed in the channel dimension, adding the number of channels of the two feature maps to form a richer feature representation. The output associated feature map is a two-dimensional matrix of the same size as the input image, where the value at each position reflects the correlation strength between the deformation of the protective plate and the stress in the support structure at that position. Regions with high correlation strength indicate the main paths of stress transmission, and the identification of these paths is of great significance for assessing the structural integrity of the battery pack.

[0050] S104. Analyze the temporal change pattern in the correlation feature diagram of stress transmission between the protective plate and the support, predict the cumulative impact of the deepening of the protective plate indentation on the support displacement, and trigger an adjustment signal if the predicted displacement exceeds the preset threshold to determine the update requirements of the dynamic detection strategy.

[0051] The sequence of associated feature maps of stress transfer between the protective plate and the support is obtained. Feature map data generated by consecutive impact events are arranged chronologically. The time-series data is segmented using a fixed window size. The difference between pixel values ​​at corresponding positions in the feature maps of adjacent moments within each window is calculated to obtain a temporal difference matrix representing the rate of feature change. A Long Short-Term Memory (LSTM) network is used to process the temporal difference matrix. Gated recurrent units selectively retain and forget historical information, encoding the cumulative effect of protective plate indentation in the hidden layer state vector. Based on a nonlinear combination of the current input and historical state vectors, a predicted value of the support displacement at the next moment is output. The predicted value is accumulated over time, with displacement increments added moment by moment starting from the initial state. When the accumulated displacement exceeds a preset threshold for allowable material deformation, a binary adjustment signal is generated. Based on the trigger time of the adjustment signal, the slope of displacement growth within a preset time period before triggering is calculated. If the slope exceeds a preset multiple of the historical average growth rate, the current detection sensitivity is deemed insufficient, and a detection strategy update flag is output.

[0052] In one embodiment, the acquisition of the associated feature map sequence is based on continuous records of road impact events. Each time a vehicle passes over a speed bump, manhole cover, or pothole, a feature map reflecting the stress transfer relationship between the guard plate and the support is generated. These feature maps are stored in a buffer in timestamp order, forming time-series data. The sequence length is dynamically adjusted according to the historical impact frequency, with longer historical sequences retained for high-frequency impact sections to capture the cumulative effect.

[0053] Specifically, the choice of a fixed window size needs to balance the integrity of temporal features with computational efficiency. The window width is set to include data from 5 to 10 consecutive time points, a range that covers typical shock response cycles. For a system with a sampling frequency of 100 Hz, the window corresponds to a time span of 50 to 100 milliseconds. The window slides along the time axis in single-time-step increments to ensure sufficient overlap between adjacent windows and avoid missing transient change information. The temporal difference matrix is ​​obtained through point-by-point subtraction, where each element represents the rate of change of the eigenvalue at the corresponding spatial location over time. Positive values ​​indicate stress enhancement, negative values ​​indicate stress release, and zero values ​​indicate a steady state. The difference operation eliminates the static background and highlights dynamic change features, which is crucial for identifying the development trend of protective plate dents.

[0054] It's important to note that the gated recurrent unit (ROU) of a Long Short-Term Memory (LSTM) network comprises three key components: the input gate, the forget gate, and the output gate. The input gate determines which information from the current input is worth preserving in the cell state. A sigmoid activation function limits the gate value to between 0 and 1, where 0 represents complete blocking and 1 represents complete passage. The forget gate controls which information from the previous cell state should be discarded; this selective forgetting mechanism prevents the accumulation of irrelevant historical information. The output gate determines the output based on the current cell state, ensuring that only relevant information is passed to the next layer. These three gates are connected to the input data and the previous hidden state via a learnable weight matrix, with the weights optimized using backpropagation.

[0055] The hidden layer state vector has a dimension of 128 or 256, which is sufficient to encode the complex spatiotemporal patterns of the protective plate indentation. Each element in the state vector corresponds to a specific cumulative effect pattern, such as persistent indentation in the central region, stress diffusion in the edge region, and fatigue accumulation at the support connection points. Through linear and nonlinear transformations of the vector space, the network learns the mapping relationship from historical states to future displacements. The nonlinear combination is achieved through the hyperbolic tangent activation function, introducing the model's nonlinear modeling capability, enabling it to capture the complex nonlinear relationship between the protective plate deformation and the support displacement.

[0056] The accumulation process of displacement predictions takes into account the non-uniformity of time intervals. In actual road driving, the intervals between impact events are not fixed; multiple impacts may occur consecutively in a short period, or there may be long periods without impacts. The accumulation calculation uses a time-weighted method, assigning higher weight to recent displacement increments and exponentially decreasing weight to longer-term increments. This weighting mechanism reflects the stress relaxation characteristics of materials, meaning that early accumulated stress is partially released over time. The comparison between accumulated displacement and allowable deformation of the material is based on the principles of materials mechanics. The allowable deformation of aluminum alloy protective plates is typically taken as 50% to 70% of the yield strain to ensure operation within the elastic range.

[0057] For example, the triggering of the binary adjustment signal follows a hysteresis comparison principle. When the cumulative displacement first exceeds a threshold, the signal jumps from 0 to 1, but only returns to 0 when the displacement decreases to below 80% of the threshold. This hysteresis characteristic avoids frequent signal switching when the displacement fluctuates around the threshold. The duration and frequency of the adjustment signal are recorded as an auxiliary indicator for assessing the health of the battery pack.

[0058] Understandably, the selection of the preset time period before triggering needs to cover typical damage development cycles. For fatigue damage, the time period is set to the most recent 10 to 20 impact cycles; for sudden damage, the time period is shortened to 3 to 5 cycles. The displacement growth slope is calculated through least-squares linear fitting, and the slope of the fitted line is the average growth rate. The historical average growth rate is based on long-term statistical data under normal operating conditions and is continuously updated using the moving average method. When the real-time slope exceeds 1.5 to 2 times the historical average, it indicates accelerated damage development, requiring increased detection sensitivity. Furthermore, the detection strategy update identifier contains information in multiple dimensions: the update type indicates whether to increase the sampling frequency or expand the monitoring range; the update urgency is divided into three levels: regular, urgent, and critical; and the update parameters provide specific adjustment suggestions. This information is encoded in a structured data packet, facilitating subsequent parsing and application by the strategy execution module. The output of the update identifier triggers a closed-loop control process. After the new detection strategy is applied, the system continuously monitors its effect. If the displacement growth rate returns to normal, the new strategy is maintained; if the problem persists or worsens, the strategy parameters are further adjusted to achieve adaptive dynamic optimization.

[0059] S105. Based on the adjustment signal, obtain reference parameters for similar road surface protrusion scenarios from the historical impact dataset, fuse the current vehicle speed and impact intensity data, and determine the optimized sensor sampling frequency and stress monitoring focus area.

[0060] Based on the trigger time of the adjustment signal, data records under the same time period and road conditions are retrieved from the historical impact dataset. The similarity between the current protective plate deformation distribution and historical records is measured using Euclidean distance. Similar scenarios with distance values ​​less than a preset threshold are selected, and the sensor sampling rate and spatial resolution parameters in these similar scenarios are extracted as reference configurations. Real-time readings from the vehicle speed sensor and the impact peak value measured by the accelerometer are acquired. The product of the squared velocity and the impact peak value is calculated as the impact energy parameter. Based on the ratio of the impact energy parameter to the standard energy value in the reference configuration, the reference sampling rate is linearly adjusted to obtain a sensor sampling frequency suitable for the current impact intensity. By statistically analyzing the trigger counts at each sensor location in historical damage events, a Gaussian kernel function is used to smooth discrete damage points, generating a continuous damage probability density distribution map. Based on the boundaries of regions with probability densities exceeding a preset threshold and the sensor sampling frequency, the optimized stress monitoring focus area and corresponding sampling configuration parameters are determined.

[0061] In one embodiment, the historical impact dataset is stored and categorized according to road condition type, weather conditions, and time period. Each historical record contains complete matrix data of the guardrail deformation distribution, corresponding sensor configuration parameters, and the final damage assessment result. After the adjustment signal is triggered, environmental parameters are extracted for 5 seconds before and after the current moment, including road surface roughness estimates, ambient temperature, and humidity, as auxiliary conditions for scene matching.

[0062] Specifically, the Euclidean distance is calculated in the feature space of the protective panel deformation distribution. The current deformation distribution matrix is ​​expanded into a one-dimensional vector, and the distance value is calculated one by one with the deformation vectors in the historical records. The distance calculation is normalized to eliminate the influence of different dimensions. The similarity threshold is determined based on the statistical distribution of historical data, and the lower quartile of the distance distribution is taken as the screening criterion to ensure that the selected similar scenes are sufficiently representative.

[0063] It should be noted that the impact energy parameter reflects the amount of kinetic energy transferred to the battery pack during the collision. The velocity square term reflects the quadratic relationship between kinetic energy and velocity, while the impact peak represents the maximum instantaneous acceleration. The product of the two provides a comprehensive measure of impact intensity. The standard energy value is obtained through laboratory calibration and corresponds to the average impact level under typical urban road conditions. When the actual impact energy is higher than the standard value, the sampling frequency is increased proportionally to ensure that more transient details are captured; conversely, it is appropriately reduced to save computational resources.

[0064] The bandwidth parameter of the Gaussian kernel function is adaptively adjusted according to the spatial resolution of the sensor. The kernel function transforms discrete damage points into a continuous probability field, where the probability density value at each spatial location represents the likelihood of damage occurring at that location. The smaller the bandwidth, the more concentrated the generated probability distribution is near the actual damage point; the larger the bandwidth, the more pronounced the smoothing effect, and the more effectively potential high-risk areas can be identified.

[0065] Areas with damage probability density exceeding a threshold are identified using a contour line extraction algorithm. The threshold is set at 60% of the maximum probability density, a proportion that ensures coverage of major risk areas while avoiding excessive expansion of the monitoring range. The extracted area boundaries are simplified using polygon processing to reduce the number of vertices for easier real-time processing. The monitoring focus area configuration includes the center coordinates, coverage radius, and corresponding sampling frequency parameters for each area, forming a zoned and hierarchical monitoring scheme.

[0066] S106. The sensor array acquisition mode is reconfigured based on the optimized sensor sampling frequency and monitoring focus area. In subsequent road impact events, an adaptive filtering algorithm is applied to capture transient deformation data, and the mechanical health assessment index of the battery pack, which integrates multi-source information, is calculated.

[0067] Based on the optimized sensor sampling frequency and the coordinate set of the monitoring focus area, a configuration command is sent to the sensor array via a serial communication protocol to modify the sampling period of sensors within the focus area to a high-frequency mode and the sampling period of sensors outside the focus area to a low-frequency mode, thus completing the partitioned configuration of the acquisition mode. When the road impact acceleration exceeds a preset trigger threshold, a Kalman filter is used to process the sensor signals after the partitioned configuration. Through recursive calculations of the prediction equation and the observation equation, noise components and the true signal are separated, and transient characteristic data of the protective plate deformation are extracted. The deformation peak value, rise time, and recovery time parameters in the transient characteristic data are obtained. At the same time, the reference deformation value at the same location is read from the historical database, and the deviation ratio between the current deformation and the reference value is calculated as the local damage degree. Based on the distribution of the local damage degree in each area of ​​the protective plate, the overall damage degree value is calculated using an area-weighted average method. If the overall damage degree is less than the preset safety threshold, the output health level is excellent; if it is between the safety threshold and the warning threshold, the output health level is good; if it exceeds the warning threshold, the output health level is poor, thus obtaining a quantified mechanical health assessment index for the battery pack.

[0068] In one embodiment, the serial communication protocol adopts the RS-485 bus standard, supporting multi-point communication and long-distance transmission. Configuration commands are encapsulated in data frame format, including the device address, function code, data field, and checksum. Each sensor node is assigned a unique address identifier, ranging from 01H to FFH, supporting up to 255 sensor nodes. The function code defines the configuration operation type: 03H indicates reading the current configuration, and 06H indicates writing a new configuration. The data field contains a sampling period parameter in milliseconds, set to 10ms for high-frequency mode and 100ms for low-frequency mode. The checksum employs a cyclic redundancy check algorithm to ensure the integrity of data transmission.

[0069] Specifically, the zonal configuration is implemented based on the spatial distribution of the monitoring focus area. Sensor addresses within the focus area are obtained through a lookup table, and these sensors receive high-frequency sampling configuration commands. Sensors in non-focus areas maintain low-frequency sampling to reduce data transmission load and power consumption. Configuration commands are sent via broadcast, and all sensors, upon receiving the command, determine whether to perform a configuration update based on their own addresses. After the configuration takes effect, the sensors return confirmation frames, and the main controller counts the number of confirmations to verify the success rate of the configuration.

[0070] It should be noted that the prediction equation of the Kalman filter is based on the system's state transition model. The state vector contains three components: displacement, velocity, and acceleration of the protective plate. The state transition matrix is ​​determined according to the sampling time interval and the laws of physical motion. The prediction covariance matrix reflects the uncertainty of state prediction and is updated through the process noise covariance matrix. The observation equation associates the sensor measurements with the state vector. The observation matrix is ​​usually a subset of the identity matrix, representing the direct measurement of certain state components. The measurement noise covariance matrix is ​​set according to the sensor's accuracy specifications, reflecting the statistical characteristics of the measurement error. The Kalman gain is calculated based on the ratio of the prediction covariance to the measurement noise covariance, determining the weighting of the predicted and measured values ​​in the state update. The recursive operation is performed at each sampling time: first, state prediction is performed, then state correction is performed based on the new measured values, and the filtered state estimate is output.

[0071] Transient characteristic parameters were extracted based on the time-domain waveform of the impact response. The deformation peak was obtained by searching for the maximum value within a time window after the trigger moment, with the time window length set to 200 ms, covering the main part of a typical impact response. The rise time was defined as the time required for the signal to rise from 10% peak value to 90% peak value, reflecting the rate of impact energy transfer. The recovery time was measured as the time interval from peak value to 10% peak value, characterizing the elastic recovery properties of the shielding plate. These time-domain parameters collectively describe the dynamic characteristics of the impact response.

[0072] The baseline deformation value is established through long-term statistics. During the initial operation phase, the system collects deformation data under normal operating conditions and performs statistical analysis on historical data for each sensor location. The baseline value is taken as the median of historical data to avoid the influence of extreme values. The baseline value is updated periodically using a sliding window method, retaining data from the most recent 30 days for statistical analysis. The formula for calculating the deviation ratio between the current deformation and the baseline value is the difference between the current value and the baseline value, divided by the baseline value, yielding a dimensionless relative deviation index.

[0073] For example, the local damage degree is calculated separately in the gridded regions of the protective plate. Each grid region contains several sensor measurement points, and the local damage degree within the region is the average of the deviation ratios of each measurement point. The area-weighted averaging method considers the differences in importance between different regions, assigning higher weights to the central region and lower weights to the peripheral regions. The weighting coefficients are determined according to a Gaussian distribution function based on the distance from the region to the center of the protective plate, ensuring a smooth weight transition. The overall damage degree value is equal to the sum of the products of the local damage degree of each region and its corresponding weight.

[0074] Understandably, the health level thresholds are set based on the material's fatigue characteristics and safety factor. The safety threshold corresponds to 30% of the material's fatigue limit; within this range, the protective plate can withstand an unlimited number of cyclic loads. The warning threshold is set at 60% of the fatigue limit; exceeding this value requires enhanced monitoring and preventative maintenance. The specific threshold values ​​are determined through fatigue testing, taking into account the influence of environmental factors such as temperature and humidity. Furthermore, in addition to the three qualitative levels of excellent, good, and poor, the health level also outputs a quantitative health score from 0 to 100. The health score is calculated using a piecewise linear mapping function: 100 when the overall damage level is 0, 70 when the safety threshold is reached, and 40 when the warning threshold is reached; beyond the warning threshold, the health score decreases linearly. This quantitative representation facilitates trend analysis and remaining life prediction, providing a quantitative basis for maintenance decisions.

[0075] For example, when a vehicle passes over multiple speed bumps consecutively, the system updates its health assessment in real time. The health changes after each impact are recorded and trend curves are plotted. Slope analysis identifies the rate of health degradation. When the degradation rate exceeds a preset limit, it is recommended to shorten the inspection cycle or schedule preventative replacement, achieving condition-based maintenance management.

[0076] S107. A graded early warning mode is established based on the mechanical health assessment index of the battery pack. When the risk of bracket displacement is detected to reach the level two early warning standard, a structural safety reminder signal is generated and transmitted to the vehicle diagnostic system and instrument display module through the CAN bus to realize real-time monitoring and active protection of the electric vehicle battery health status.

[0077] A tiered early warning system is established based on battery pack mechanical health assessment indicators. Normal condition corresponds to a health level above a first threshold, mild risk corresponds to a health level between the first and second thresholds, and moderate risk corresponds to a health level below the second threshold. A secondary warning indicator is generated when the health level drops below the second threshold. Following the generation process of the structural safety reminder signal triggered by the secondary warning indicator, the maximum displacement value and displacement rate are extracted from the bracket displacement monitoring data. The displacement information is converted into a diagnostic fault code according to a preset encoding rule, and a complete reminder signal data packet is formed by adding a timestamp and sensor location identifier. The reminder signal data packet is encapsulated into a standard frame format via the CAN bus protocol stack, with the message identifier priority bit set to high priority. It is then transmitted via the vehicle network physical layer to the vehicle diagnostic system for fault recording, and simultaneously transmitted to the instrument display module to trigger the warning light and text prompt, realizing real-time alarm and driver reminder for abnormal battery health conditions.

[0078] In one embodiment, the tiered early warning system employs a three-tiered threshold structure. The first threshold is set at 70 points for the health index, the second at 40 points, and the third at 20 points. These thresholds are determined based on statistical analysis of extensive real-vehicle test data, taking into account the battery pack stress characteristics and material fatigue patterns under different road conditions. As the health index gradually decreases from 100 points, the system compares the current value with the thresholds at each level in real time, dynamically updating the early warning status.

[0079] Specifically, the signal generation process after a Level 2 warning is triggered follows a standardized procedure. First, key parameters are extracted from the support displacement monitoring data, including the maximum displacement amplitude, the coordinate index of the displacement location, and the displacement rate. These parameters are converted according to the diagnostic fault code encoding standard. The fault code uses a four-digit hexadecimal format, with the first two digits indicating the fault type and the last two digits indicating the fault level and location. The timestamp uses a Unix time format, accurate to the millisecond, ensuring accurate recording of the fault occurrence time.

[0080] It should be noted that the CAN bus message encapsulation conforms to the ISO 11898 standard. The standard frame format contains an 11-bit identifier, with the first 3 bits indicating message priority. The structural safety warning signal is assigned a priority code of 010, second only to emergency braking and steering control signals. The data field is 8 bytes long; the first 4 bytes store compressed information of the fault code and timestamp, and the last 4 bytes contain displacement data and reserved extension bits. After receiving the CAN message, the vehicle diagnostic system reconstructs the complete fault information through the fault code parsing module. The fault information is written to non-volatile memory, forming a fault history record for easy tracing and analysis by maintenance personnel. Simultaneously, the diagnostic system generates diagnostic codes conforming to the OBD-II standard, supporting reading by general diagnostic equipment.

[0081] The instrument cluster's alarm system utilizes both visual and audible alerts. Visual alerts are provided by illuminating the battery fault indicator light on the instrument panel, which uses an amber LED and flashes at 2Hz. A text message is displayed on the central control screen stating, "Battery pack structure abnormal, please check as soon as possible." Audible alerts consist of an intermittent buzzer, lasting 3 seconds followed by a 2-second pause, repeated 3 times. This multimodal alarm system ensures that the driver can promptly detect abnormal conditions and take appropriate safety measures.

[0082] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for evaluating and monitoring the state of health of an electric vehicle battery, characterized in that, The method comprises the following steps: S101, collecting real-time impact signal data through a multi-axis acceleration sensor and a strain sensor array deployed at the bottom of the battery pack, synchronously recording the deformation parameters of the protection plate area and the stress value changes of the internal fixed support when the vehicle passes through the road bumps, and obtaining the original impact data set; S102, applying Fourier transform to the original impact data set to extract frequency domain features and time domain peak value features, separating the low frequency components related to the protection plate depression and the high frequency impact peak value, determining the protection plate depression degree index and the support stress concentration point displacement estimation value; S103, using a convolutional neural network to image map the protection plate depression degree index and the support stress concentration point displacement estimation value, converting the protection plate deformation parameters into a two-dimensional heat map representation, and obtaining the correlation feature map of stress transmission between the protection plate and the support; S104, analyzing the time sequence change pattern in the correlation feature map of stress transmission between the protection plate and the support, predicting the cumulative effect of protection plate depression deepening on support displacement, and if the predicted displacement exceeds the preset threshold, triggering an adjustment signal and determining the update requirement of the dynamic detection strategy; S105, obtaining reference parameters under similar road bump scenarios from historical impact data sets according to the adjustment signal, fusing current vehicle speed and impact intensity data, and determining the optimized sensor sampling frequency and stress monitoring focus area; S106, reconfiguring the acquisition mode of the sensor array according to the optimized sensor sampling frequency and monitoring focus area, applying an adaptive filtering algorithm to capture transient deformation data in subsequent road impact events, and calculating a battery pack mechanical health degree evaluation index fused with multi-source information; S107, establishing a hierarchical warning mode based on the battery pack mechanical health degree evaluation index, generating a structure safety reminder signal when the support displacement risk reaches the secondary warning standard, and transmitting the signal to the vehicle diagnostic system and instrument display module through the CAN bus, realizing real-time monitoring and active protection of the battery health status of the electric vehicle.

2. The method of claim 1, wherein, The step S101 comprises: acquiring acceleration signals output by a multi-axis acceleration sensor array arranged on a bottom protection plate of a battery pack, processing high-frequency noise components higher than a preset frequency threshold in the acceleration signals through a Butterworth low-pass filter, identifying a starting moment of a road bump impact event according to a change in the filtered acceleration amplitude, recording a starting moment marker when the acceleration amplitude exceeds a preset impact threshold, and triggering a data acquisition mode. After triggering the data acquisition mode, strain signals are read from a strain sensor array on the surface of the protection plate, local deformation parameters of each measuring point of the protection plate are calculated according to a linear relationship between a resistance change rate of the strain gauge and a strain value, stress sensor output signals of measuring points of an internal fixed support are synchronously acquired, and stress value change data of each measuring point of the support are converted according to a calibration curve of the sensor output voltage and the stress. The starting moment marker is used for time sequence alignment processing of the local deformation parameters and the stress value change data, deformation distribution matrixes of the protection plate are formed by combining deformation parameters of each measuring point according to spatial position coordinates of the sensor array, stress distribution matrixes of the support are formed by combining stress values of each measuring point, and original impact data sets containing time and space information are obtained through matrix splicing processing.

3. The method of claim 1, wherein, The step S102 comprises: performing fast Fourier transform on the protection plate deformation data in the original impact data sets, decomposing time domain signals into a form of superposition of sine waves of different frequency components, acquiring amplitude spectrum and phase spectrum in the frequency domain, identifying a dominant frequency component according to energy sizes of frequency components in the amplitude spectrum, classifying low-frequency components of the protection plate slow deformation when the frequency components are lower than a preset low-frequency threshold, and obtaining a frequency domain feature vector related to the protection plate depression. Band-pass filter processing is performed on the low-frequency components in the frequency domain feature vector, a frequency band related to a structure inherent frequency of the protection plate is retained, a root mean square value of the filtered signal is calculated as a deformation energy index, a protection plate depression degree index is determined according to a relationship between the deformation energy index and an elastic modulus of the protection plate material. Stress peak values of each measuring point and corresponding moments are extracted from the support stress data of the original impact data sets, a stress propagation speed is calculated through a ratio of a difference between stress peak value occurrence moments of adjacent measuring points to a distance between the measuring points, and a damage state of the support material is judged according to the stress propagation speed. Based on the protection plate depression degree index and the stress propagation speed, an association relationship between the depression degree and the support stress distribution is established through a linear regression method, a spatial displacement of a measuring point relative to an initial position in a non-impact state is calculated when a stress value of the measuring point exceeds a preset stress concentration threshold, and a support stress concentration point displacement estimation value is obtained.

4. The method of claim 1, wherein, Step S103 comprises: according to the protective plate recess degree index and the support stress concentration point displacement estimation value, a two-dimensional coordinate system is established according to the physical arrangement position of the sensor array, the numerical values between the discrete measuring points are calculated by the bilinear interpolation method, the continuous distribution numerical field is generated on the regular grid points of the coordinate system, and the protective plate deformation distribution matrix and the support displacement distribution matrix are obtained. The protective plate deformation distribution matrix and the support displacement distribution matrix are normalized respectively, the normalized numerical values are mapped to the preset gray value interval, the protective plate deformation thermal map and the support displacement thermal map are generated through the corresponding relationship between the gray value and the pseudo color, the pixel values of the two thermal maps are fused by using the weighted average method, and the fused two-dimensional image data is obtained. The convolutional neural network is used for feature extraction of the two-dimensional image data, the local area of the image is convolved by using the pre-trained convolution kernel, the spatial correlation features of the protective plate deformation mode and the support displacement mode are extracted, and the feature map dimension is reduced and the significant features are retained through the maximum pooling operation in the pooling layer. According to the spatial correlation features, the spatial resolution of the feature map is restored layer by layer through the transpose convolution operation, the convolution features of the corresponding layer are connected through the jump connection after each layer of transpose convolution, the multi-scale spatial information is fused, and the correlation feature map reflecting the stress transmission path between the protective plate and the support is output.

5. The method of claim 1, wherein, Step S104 comprises: obtaining the correlation feature map sequence of the stress transmission between the protective plate and the support, arranging the feature map data generated by the continuous impact events in time sequence, segmenting the time sequence data by using a fixed window size, calculating the difference value of the pixel values at the corresponding positions of the adjacent time features in each window, and obtaining the time sequence difference matrix representing the feature change rate. The long short-term memory network is used for processing the time sequence difference matrix, the historical information is selectively retained and forgotten through the gating cycle unit, the cumulative effect of the protective plate recess is encoded in the hidden layer state vector, the prediction value of the support displacement at the next moment is output according to the nonlinear combination of the current input and the historical state vector. The prediction value is accumulated in time sequence, the displacement increment is superimposed moment by moment from the initial state, and when the cumulative displacement exceeds the preset threshold of the material allowable deformation, a binary adjustment signal is generated. According to the triggering moment of the adjustment signal, the slope of the displacement growth in the preset time period before triggering is calculated, and if the slope exceeds the preset multiple of the historical average growth rate, it is determined that the current detection sensitivity is insufficient, and a detection strategy updating identifier is output.

6. The method of claim 1, wherein, Step S105 includes: according to the trigger time of the adjustment signal, retrieving the data record under the same time period and road condition from the historical impact data set, measuring the similarity degree between the current protective plate deformation distribution and the historical record by the Euclidean distance, screening out the similar scenes with the distance value less than the preset threshold, and extracting the sampling rate and spatial resolution parameters of the sensor in the similar scene as the reference configuration. The real-time reading of the vehicle-mounted speed sensor and the impact peak value measured by the acceleration sensor are obtained, the product of the speed square and the impact peak value is calculated as the impact energy parameter, the reference sampling rate is linearly adjusted according to the ratio of the impact energy parameter to the standard energy value in the reference configuration, and the sensor sampling frequency suitable for the current impact intensity is obtained. By counting the trigger times of each sensor position in the historical damage event, the discrete damage points are smoothed by using the Gaussian kernel function, and a continuous damage probability density distribution map is generated. According to the region boundary of the probability density exceeding the preset threshold and the sensor sampling frequency, the optimized stress monitoring focus area and the corresponding sampling configuration parameters are determined.

7. The method of claim 1, wherein, Step S106 includes: according to the optimized sensor sampling frequency and the monitoring focus area coordinate set, sending a configuration command to the sensor array through a serial communication protocol, modifying the sampling period of the sensors in the focus area to a high-frequency mode, and the sampling period of the sensors in the non-focus area to a low-frequency mode, and completing the partition configuration of the acquisition mode. When detecting that the road impact acceleration exceeds the preset trigger threshold, a Kalman filter is used to process the sensor signal after the partition configuration, and through the recursive operation of the prediction equation and the observation equation, the noise component and the real signal are separated, and the transient characteristic data of the protective plate deformation are extracted. The deformation peak value, rise time and recovery time parameters in the transient characteristic data are obtained, and the reference deformation value of the same position is read from the historical database, and the deviation proportion of the current deformation and the reference value is calculated as the local damage degree. According to the distribution of the local damage degree in each region of the protective plate, the area weighted average method is used to calculate the overall damage degree value, and if the overall damage degree is less than the preset safety threshold, the health level is output as excellent, if it is between the safety threshold and the warning threshold, the health level is output as good, and if it exceeds the warning threshold, the health level is output as poor, and the quantitative battery pack mechanical health degree evaluation index is obtained.

8. The method of claim 1, wherein, Step S107 includes: establishing a hierarchical early warning system according to the battery pack mechanical health degree evaluation index, The normal state corresponds to a high health degree higher than a first threshold value, the mild risk corresponds to a health degree between the first threshold value and a second threshold value, the moderate risk corresponds to a health degree lower than the second threshold value, and a secondary warning mark is generated when the health degree falls below the second threshold value. According to the secondary warning mark, a structure safety reminder signal generation process is triggered, a maximum displacement value and a displacement rate are extracted from the support displacement monitoring data, the displacement information is converted into a diagnostic fault code according to a preset coding rule, and a complete reminder signal data packet is formed after adding a timestamp and a sensor position mark. The reminder signal data packet is encapsulated into a standard frame format through a CAN bus protocol stack, the priority bit of the message identifier is set to high priority, transmitted to the vehicle diagnostic system through the vehicle network physical layer for fault recording, and transmitted to the instrument display module to trigger the warning light to light up and the text prompt, realizing real-time alarm and driver reminder of the abnormal state of the battery health.

Images