Automobile brake hydraulic pressure leakage detection system based on space-time inversion twin network

Abstract

The invention discloses an automobile brake hydraulic pressure leakage detection system based on a space-time inversion twin network, and the system employs a space-time inversion twin network technology, and integrates a multi-physical field cooperative sensing network, a hardware layer, a data layer and an algorithm layer, so as to achieve the efficient and precise detection of automobile brake hydraulic pressure leakage. Key parameters such as pressure, flow velocity, vibration, temperature, dielectric constant and chemical components of the brake system are monitored in real time through various high-precision sensors, and comprehensiveness and accuracy of leakage detection are ensured. The CMOS-MEMS integrated sensor chip and the edge computing unit are responsible for data acquisition and primary processing, the multi-physics field joint calibration database and the space-time data compression encoder provide data support for the algorithm layer, a large number of parameter level basic models achieve intelligent analysis and decision, and finally the alarm module is used for timely reminding. The system is mainly used for improving the safety and reliability of an automobile braking system, and traffic accidents caused by braking hydraulic pressure leakage are effectively prevented.

Description

Technical Field

[0001] This invention relates to an automotive brake hydraulic leakage detection system, and more particularly to an automotive brake hydraulic leakage detection system based on a spatiotemporal inversion twin network. Background Technology

[0002] In the field of modern automobile manufacturing and application, with the continuous increase in car ownership and people's high regard for driving safety, the reliability of automotive braking systems has become a core element in automotive safety design. Traditional methods of monitoring automotive brake fluid mainly rely on periodic manual inspections, visually observing changes in brake fluid levels or performing simple pressure tests on the braking system. This makes it difficult to achieve real-time and accurate monitoring of brake fluid leaks. Manual inspections are not only subject to subjective errors but also fail to detect minute leaks in a timely manner, easily leading to the long-term accumulation of safety hazards, which may ultimately cause serious traffic accidents.

[0003] With the development of sensor technology, leak detection systems based on single sensors have emerged in the market. While these systems can achieve automatic monitoring to a certain extent, they only monitor one physical characteristic of the brake hydraulic fluid, such as pressure or flow rate. They lack comprehensive perception of the multi-physical field information of the braking system, and the detection results are easily affected by environmental interference, resulting in high false alarm and false negative rates. They are unable to meet the high precision and high reliability requirements of automotive braking systems for hydraulic leak detection under complex operating conditions. Furthermore, existing detection systems have shortcomings in data processing and analysis. They lack the ability to deeply integrate and intelligently analyze multi-source sensor data, making it difficult to extract effective leak characteristic information from massive amounts of data. They cannot accurately locate leaks or predict leak development trends. In terms of data transmission and storage, some systems have failed to effectively address the issues of transmission efficiency and storage optimization under large data volumes, resulting in data processing speeds lagging behind actual needs.

[0004] The automotive brake hydraulic fluid leakage detection system based on spatiotemporal inversion twin network proposed in this invention aims to solve the aforementioned problems in existing technologies. This system integrates a multi-physics collaborative sensing network, a hardware layer, a data layer, and an algorithm layer. It utilizes multiple high-precision sensors to collect key physical parameters of the braking system in real time, and combines this with an advanced spatiotemporal inversion twin network algorithm to achieve efficient and accurate detection and location of automotive brake hydraulic fluid leakage. This effectively prevents traffic accidents caused by brake hydraulic fluid leakage and ensures vehicle driving safety. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide an automotive brake hydraulic leakage detection system based on a spatiotemporal inversion twin network. By constructing an automotive brake hydraulic leakage detection system based on a spatiotemporal inversion twin network, the system can achieve real-time monitoring and accurate identification of automotive brake hydraulic leakage, effectively capture minute leakage characteristics, reduce false alarm rate and false alarm rate, and ensure the safety and stability of the automotive braking system.

[0006] Technical solution: The present invention provides an automotive brake hydraulic leakage detection system based on a spatiotemporal inversion twin network, comprising:

[0007] Data acquisition module: Collects key physical quantities data of the vehicle braking system, such as pressure, flow rate, vibration, temperature, dielectric constant, and chemical composition, through a sensor array;

[0008] Feature extraction and spatiotemporal inversion module: The feature extraction unit is used to extract temporal and spatial features; the feature fusion unit is used to fuse temporal and spatial features to construct a comprehensive feature vector; the spatiotemporal inversion unit is used to determine the leakage of automotive brake hydraulic fluid and locate the leakage point through reverse reasoning and analysis;

[0009] Hardware layer: Converts the acquired analog signals into digital signals, and performs preprocessing operations such as amplification, filtering, noise reduction, and compression encoding on the signals;

[0010] Data layer: Stores and manages data from the hardware layer. By establishing a multiphysics joint calibration database, it classifies, stores, organizes, and labels historical and real-time data; and optimizes the data using a spatiotemporal data compression encoder.

[0011] Algorithm layer: Combining a large number of parameter-level basic models, the Dropout algorithm is used to perform in-depth mining and intelligent analysis of multiphysics data stored in the data layer. By comparing real-time data with feature models under normal operating conditions, potential hydraulic leakage risks in the braking system are identified. When leakage signs are detected, corresponding alarm signals and decision suggestions are generated immediately.

[0012] Preferably, the sensor array comprises:

[0013] Nanopiezoelectric thin film array: Real-time monitoring of brake fluid pressure changes through the voltage output of the piezoelectric thin film array;

[0014] FBG fiber grating array: for monitoring vibration characteristics of braking systems;

[0015] MEMS ultrasonic Doppler sensor: Real-time monitoring of flow velocity changes;

[0016] Quantum dot infrared thermal imager: Tracking the thermodynamic characteristics of evaporation and sensing local temperature fluctuations;

[0017] Microwave resonant cavity sensor: monitoring the effect of cavitation on dielectric constant;

[0018] Terahertz time-domain spectrometer: detects minute changes in chemical composition by measuring molecular polarization relaxation time.

[0019] Preferably, the feature extraction includes:

[0020] (1) Temporal feature extraction: The temporal feature vector T includes the following elements:

[0021] T=[μ,σ,S k ,K u [AC,CC]

[0022]

[0023] Where μ is the mean of the signal; σ is the standard deviation of the signal, which measures the dispersion of the signal; x i y i S represents the number of sampling points; N represents the number of sampling points; S represents the number of sampling points. k Skewness, reflecting the asymmetry of signal distribution; K u τ is the kurtosis, which measures the sharpness of the signal peaks; AC is the autocorrelation coefficient, which measures the similarity of the signals at different times; τ is the time delay; CC is the cross-correlation coefficient, which calculates the correlation between the signal and the reference signal.

[0024] (2) Spatial domain feature extraction: The spatial domain feature vector S contains the following features:

[0025]

[0026] Among them, v max ,v min ,v avg These represent the maximum, minimum, and average values ​​of the spatial field quantity, respectively. The average value is calculated as: v avg =(1 / M)∑v j v j Here, M represents the field quantity value at the spatial sampling point, and M is the number of spatial sampling points. Let H be the gradient of the field quantity, reflecting the rate of change of the field quantity in space; H is the entropy of the spatial domain, measuring the complexity of the spatial distribution, and is calculated by the formula: H = -∑p(v j ) * logp(v j ), p(v j ) is the field quantity v j The probability of occurrence; E represents the energy characteristic of the spatial domain, calculated as: E = ∑v j ^2.

[0027] Preferably, the feature fusion combines the temporal feature vector T and the spatial feature vector S to construct a comprehensive feature vector F, and the fusion method adopts weighted fusion:

[0028] F = W T * T+W S * S

[0029] Among them, W T and W S The weight vectors for the temporal and spatial features are respectively, satisfying W T +W S =1.

[0030] Preferably, the spatiotemporal inversion unit uses the fused feature vector F to determine the leakage situation and locate the leakage point through reverse reasoning and analysis.

[0031] Leakage detection: Construct a leakage detection function D:

[0032] D = f(F; θ)

[0033] Where θ is the discrimination parameter, when D is greater than the threshold D th When the threshold D is reached, a leak is determined to have occurred. th The specific requirements for false positive and false negative rates are determined through training with historical data.

[0034] Leakage point location: Using a spatiotemporal inversion algorithm, based on the F feature vector and the spatiotemporal information of the sensor network, the location L of the leak point is calculated;

[0035] The positioning calculation formula is as follows:

[0036] L = argmin||H(L) - F||^2

[0037] In the formula, H(L) is the theoretical characteristic model corresponding to the location L of the leakage point.

[0038] Preferably, the step of classifying and storing the data is as follows:

[0039] (1) The data is classified and stored using time series analysis methods. The data is divided into time windows, and the formula for dividing the time windows is as follows:

[0040] T i =[t i ,t i +Δt)

[0041] In the formula, i = 1, 2, ..., N, where N is the number of time windows, and t i Let be the start time of the i-th time window;

[0042] (2) The formula for the hierarchical storage capacity of historical data is as follows:

[0043]

[0044] In the formula, k is the number of hierarchical storage levels, and C k Let k be the storage capacity of the k-th layer;

[0045] (3) Real-time data storage and updates:

[0046] The real-time data buffer size is S realtime When new data arrives, the buffer is updated according to the first-in, first-out principle. If the buffer is full, the oldest data is discarded and the latest data is stored to ensure that the buffer always contains the latest real-time data.

[0047] Preferably, the data optimization processing using a spatiotemporal data compression encoder includes:

[0048] (1) Data reception and preprocessing:

[0049] The data layer receives raw data from the hardware layer and performs preliminary cleaning on the braking system pressure, temperature, and flow data to remove obvious errors or outliers, ensuring the integrity and accuracy of the data.

[0050] (2) Spatiotemporal data compression coding:

[0051] (21) Data matrix construction:

[0052] The preprocessed data is used to construct a data matrix D based on time series and spatial location:

[0053] D∈R^{m×n}

[0054] In the formula, R represents the real number space; m represents the number of time series; and n represents the number of spatial locations.

[0055] (22) Singular value decomposition:

[0056] Perform singular value decomposition on the data matrix D:

[0057]

[0058] In the formula, U and V are orthogonal matrices; Σ is a diagonal matrix; and T is the matrix transpose.

[0059] (23) Singular value truncation:

[0060]

[0061] In the formula, U k V is a left singular vector; k D is a right singular vector; kΣ is an approximate matrix; k is the number of maximum singular values; Σ k It is a diagonal matrix containing only the first k singular values;

[0062] (24) Data compression and storage: U k , Σ k and V k Encode and store;

[0063] (3) Data query and decompression:

[0064] (31) Query request parsing: Receive query requests from the algorithm layer and parse out the time range and spatial location information of the required data;

[0065] (32) Encoded data retrieval: Based on the query request, locate and retrieve the corresponding U in the stored compressed data. k , Σ k and V k ;

[0066] (33) Data decompression and restoration: using the retrieved U k , Σ k and V k Through matrix multiplication D k =U k Σ k V k ^T restores an approximate original data matrix D k This provides the algorithm layer with high-quality data resources that can be directly used for intelligent analysis;

[0067] (4) Data quality assessment and feedback: Based on the analysis results and feedback information of the data by the algorithm layer, the compression and decompression process of the data layer is evaluated and optimized.

[0068] Preferably, the construction process of the large number of parameter-level basic models is as follows:

[0069] (1) Construct a multi-scale network of numerous parameter-level basic models:

[0070]

[0071] In the formula, h l This is the output of the l-th layer; LayerNorm is for normalization; h l-1 This is the output of the (l-1)th layer;

[0072] MultiHead is a multi-head attention mechanism that allows the model to learn different attention patterns in parallel across different representation subspaces; Q l ,K l V l These are the query vector, key vector, and value vector for the l-th layer, respectively; Wl Q W l K W l V These are the query weight matrix, key weight matrix, and value weight matrix, respectively; g is the global leak status code.

[0073] (2) Define the composite loss function L:

[0074]

[0075] In the formula, λ1, λ2, and λ3 are weighting coefficients used to balance the contributions of different loss terms to the total loss; L CE Cross-entropy loss; The weight matrix W l The square of the F-norm; E is the expected value; KL is the divergence, used to measure the difference between two probability distributions; p train p represents the probability distribution on the training set. test The probability distribution on the test set;

[0076] A large number of parameter-level basic models are combined with the Dropout algorithm, and synergistically optimized with a multi-scale network architecture and composite loss function to achieve overfitting suppression and improved generalization ability in hydraulic leakage detection of automotive braking systems. The specific steps of the Dropout algorithm are as follows:

[0077] (3) Dropout algorithm model construction:

[0078]

[0079] In the formula, h t h is the output vector at the current time t; t-1 The output vector at current time t-1; Q t K t V t These are the query vector, key vector, and value vector for the current time t, respectively; W t Q W t K W t V These are the linear transformation matrices for the query vector, key vector, and value vector, respectively. The regularized block vector after input to the Dropout algorithm;

[0080] (4) Enhanced regularization using composite loss function:

[0081]

[0082] In the formula, L zThis is the main part of the composite loss function, used to measure the difference between the model's predicted values ​​and the actual values; L ce Used to measure the difference between the predicted result and the true label; λ1 controls the weight of the cross-entropy loss in the total loss; W is the model weight matrix; λ is the Frobenius norm of the matrix to prevent overfitting; λ² is the control for L. z The strength of regularization; P train and P test The distributions of the training and test sets are given; KL(||) is used to measure the difference between the two distributions and reduce distribution bias; λ3 is the weight controlling the KL divergence term; E m For expectation operators; λ represents the transformed feature or hidden representation; λ4 controls the strength of sparsity regularization. λ is the rate of change of the feature over time; λ5 is the weight controlling the smoothness of the time series. For L z For the loss function when m j =1, the gradient includes the gradient of the cross-entropy loss and the sign function of the sparse regularization term; Represents the loss function L z right The gradient is used for model parameter updates; For loss function L z Hidden representation The partial derivatives are used for backpropagation; Here is a sign function used for sparsity regularization; when m j When m = 1, the gradient includes cross-entropy loss and sparse regularization term; when m = 1, the gradient includes cross-entropy loss and sparse regularization term. j When the gradient is 0, the gradient is 0, indicating that the feature is not included in the update.

[0083] (5) Dynamic probability adaptive mechanism p (l) :

[0084]

[0085] In the formula, exp is the exponential function term; α is the sensitivity parameter controlling the change in probability; Used to normalize the Frobenius norm, avoiding excessively large or small norm values ​​due to differences in dimensions.

[0086] Preferably, the Dropout algorithm is optimized for use in an automotive brake hydraulic leakage detection system. A gradient compensation mechanism is employed to optimize the Dropout algorithm, and the optimization steps are as follows:

[0087] (1) Dynamic stratified discard probability:

[0088]

[0089] In the formula, l is the current layer index; γ is the depth decay factor; β is the gradient sensitivity coefficient; Let be the gradient norm of the loss function at layer l;

[0090] (2) Gradient compensation mechanism:

[0091] Since Dropout randomly discards some neurons during training, it introduces variance in gradient estimation. Therefore, a gradient compensation term is introduced.

[0092]

[0093] In the formula, For L z gradient, For gradient operators; λ r This is the regularization coefficient, used to control the weight of the regularization term in gradient compensation; This is the regularization loss gradient, used to prevent the model from overfitting;

[0094]

[0095] In the formula, This represents taking the partial derivative with respect to the weight matrix W; This represents the summation of all terms from level 1 to level L; W l W represents the weight matrix of the Lth layer of the neural network. l (pre) This represents the weight matrix of the previous iteration of the neural network;

[0096] (3) Scaling compensation:

[0097]

[0098] In the formula, The optimized gradient; This is a scaling factor used to scale the compensated gradient when a neuron is activated; This represents the probability of a neuron being discarded, preventing overfitting.

[0099] Preferably, the feature model comparison step is as follows:

[0100] (1) Define the feature comparison function D(t):

[0101] D(t) = ||F real (t)-F model ||2

[0102] In the formula, F real (t) is the real-time feature vector; F modelThis is the feature model under normal operating conditions;

[0103] (2) Introduce time series analysis and calculate the moving average D. avg (t):

[0104]

[0105] In the formula, Δt is the length of the time interval [t-Δt,t]; τ is the integration variable, representing any point in time within the time interval [t-Δt,t]; D(τ) is the value of the distance function D at time τ; Indicates the time interval

[0106] Integrating the distance function D(τ) over the interval [t-Δt,t], i.e., accumulating the error;

[0107] (3) Determine the hydraulic leakage risk Leakage(t):

[0108]

[0109] In the formula, T is the leakage threshold determined by experimental data;

[0110] When a leak is detected, the algorithm layer immediately generates a corresponding alarm signal and decision suggestion. The steps for generating the alarm signal and making decision suggestions are as follows:

[0111] (4) Determine the alarm signal A(t):

[0112]

[0113] In the formula, α is the alarm intensity coefficient, which is used to adjust the amplitude of the alarm signal;

[0114] (5) The most suitable decision recommendation D s (t):

[0115]

[0116] In the formula, argmax is used to find the parameter value that makes the function reach its maximum value; Match(D(t),R i ) represents D(t) and R i The degree of matching, where R is a predefined set of decision rules.

[0117] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0118] (1) This invention constructs a vehicle brake hydraulic leakage detection system based on a spatiotemporal inversion twin network, which realizes real-time monitoring and accurate identification of vehicle brake hydraulic leakage, effectively captures minute leakage characteristics, reduces false alarm rate and false alarm rate, and ensures the safety and stability of vehicle braking system.

[0119] (2) This invention fully explores and utilizes multi-physics information to make the detection results more accurate and reliable. Even under complex working conditions, it can meet the high precision and high reliability requirements of automotive braking systems for hydraulic leakage detection. At the same time, by using a spatiotemporal inversion twin network algorithm to deeply fuse and intelligently analyze multi-source sensor data, it accurately identifies potential hydraulic leakage risks in the braking system, and achieves precise location of the leakage and prediction of the leakage development trend, further improving the performance and function of the detection system.

[0120] (3) The hardware layer of this invention uses a CMOS-MEMS integrated sensor chip and an edge computing unit to convert and preliminarily process the acquired analog signals, improving data quality and availability, reducing data transmission volume, and increasing data transmission efficiency. This not only enhances the real-time performance and stability of the system but also lays a solid foundation for subsequent data analysis and processing, ensuring that the entire detection system can operate stably and efficiently, thereby better leveraging its advantages in detecting hydraulic leakage in automotive brake systems and improving the overall performance and safety of automotive braking systems. Attached Figure Description

[0121] Figure 1 This is a system structure diagram of the present invention.

[0122] Figure 2 This is a flowchart of the Dropout algorithm.

[0123] Figure 3 A diagram illustrating the optimization process of the Dropout algorithm.

[0124] Figure 4 This is a comparison chart of the leak detection response of the present invention and existing systems.

[0125] Figure 5 This is a system feature analysis diagram as described in this invention.

[0126] Figure 6 This is a performance comparison chart between the present invention and existing systems.

[0127] Figure 7 This is the feature space evolution diagram of the system described in this invention. Detailed Implementation

[0128] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0129] A system for detecting hydraulic leakage in automotive brake fluid based on a spatiotemporal inversion twin network is disclosed. The automotive braking system comprises a master cylinder, brake lines, and wheel cylinders. The master cylinder acts as a power source, converting mechanical energy into hydraulic energy. The brake lines are responsible for transmitting brake fluid and connecting the master cylinder and wheel cylinders. The wheel cylinders, installed at the wheels, convert hydraulic energy into mechanical braking force. By rationally designing and arranging the various components of the braking system, the normal operation and reliability of the braking system are ensured.

[0130] The energy conversion calculation formula for the brake master cylinder is as follows:

[0131]

[0132] In the formula, P power η is the power of the master cylinder; Q is the flow rate of the brake fluid; P is the hydraulic pressure output by the master cylinder; η is the efficiency of the master cylinder, reflecting the efficiency loss of the master cylinder in the energy conversion process.

[0133] The formula for calculating the flow rate of brake fluid transmitted through the brake line is:

[0134]

[0135] In the formula, v is the brake fluid flow rate; A is the cross-sectional area of ​​the brake line.

[0136] The formula for calculating the energy conversion of brake calipers is:

[0137]

[0138] In the formula, P mech The mechanical braking power generated by the brake caliper; η brake The energy conversion efficiency of the brake caliper; n brake This refers to the rotational speed of the brake disc.

[0139] The detection system includes:

[0140] (1) Data Acquisition Module: The multi-physics collaborative sensing network integrates multi-source sensor data and synchronously acquires key physical quantities of the braking system, such as pressure, flow rate, vibration, temperature, dielectric constant, and chemical composition, through a diverse sensor array; including:

[0141] Nanopiezoelectric thin film array: Real-time monitoring of minute changes in brake fluid pressure by using the voltage output from the piezoelectric thin film array;

[0142] FBG fiber grating array: Identifying structural modal changes and monitoring vibration characteristics of braking systems;

[0143] MEMS ultrasonic Doppler sensor: Analyzes fluid motion anomalies and enables real-time flow velocity monitoring;

[0144] Quantum dot infrared thermal imager: Tracking the thermodynamic characteristics of evaporation and sensing local temperature fluctuations;

[0145] Microwave resonant cavity sensor: monitoring the effect of cavitation on dielectric constant;

[0146] Terahertz time-domain spectrometer: detects minute changes in chemical composition by measuring molecular polarization relaxation time.

[0147] Among them, the piezoelectric film output voltage U out The calculation formula is:

[0148] U out =g·ΔP·A1

[0149] In the formula, g is the piezoelectric coefficient of the piezoelectric film; ΔP is the minute change in braking hydraulic force; and A1 is the effective force-bearing area of ​​the piezoelectric film.

[0150] The formula for calculating the change in reflected wavelength of a fiber optic grating is:

[0151] λ B =2n eff Λ

[0152] In the formula, λ B n is the reflection wavelength of the fiber optic grating; eff Λ represents the effective refractive index of the optical fiber; Λ represents the grating period.

[0153] When the structure vibrates, the optical fiber experiences strain ò, which affects the grating period Λ and the effective refractive index n. eff The change is calculated using the following formula:

[0154]

[0155] In the formula, ρ is the strain optical coefficient of the optical fiber; Δn eff This refers to the change in effective refractive index.

[0156] The formula for calculating fluid motion anomaly detection is:

[0157]

[0158] In the formula, σ is the mean; u is the standard deviation; n is the total number of samples; u i For the i-th observation; This represents the summation over the first to nth observations; when the relative change in flow velocity... The threshold α has been exceeded.

[0159] The formula for calculating the relationship between temperature change and infrared radiation intensity is:

[0160] I=η1·εσA2T4

[0161] In the formula, I is the infrared radiation intensity received by the quantum dot infrared thermal imager; η1 is the detection efficiency of the thermal imager; ε is the emissivity of the object; σ is the Stefan-Boltzmann constant; A2 is the surface area of ​​the object; and T is the absolute temperature of the object.

[0162] The formula for calculating the effect of cavitation on dielectric constant is:

[0163]

[0164] In the formula, ε eff To change the effective dielectric constant of the liquid; ε f ε is the dielectric constant of air; l is the dielectric constant of pure brake fluid; f is the volume fraction of air bubbles.

[0165] The formula for calculating the molecular polarization relaxation time is:

[0166]

[0167] In the formula, α1 is the terahertz absorption coefficient; e is the electron charge; m is the molecular mass; c is the speed of light; f1 is the terahertz wave frequency; and τ is the molecular polarization relaxation time.

[0168] The main code is as follows:

[0169]

[0170]

[0171] (2) Feature extraction and spatiotemporal inversion module: Multi-physics field collaborative sensing spatiotemporal inversion twin network. The feature extraction unit is used to extract temporal and spatial features; the feature fusion unit is used to fuse temporal and spatial features to construct a comprehensive feature vector; the spatiotemporal inversion unit is used to determine the leakage of automotive brake hydraulic fluid and locate the leakage point through reverse reasoning and analysis.

[0172] (21) The feature extraction steps are as follows:

[0173] (211) Temporal feature extraction:

[0174] The time-domain feature vector T contains the following elements:

[0175] T=[μ,σ,S k ,K u [AC,CC]

[0176]

[0177] Where μ is the mean of the signal; σ is the standard deviation of the signal, which measures the dispersion of the signal; x i y i S represents the number of sampling points; N represents the number of sampling points; S represents the number of sampling points. k Skewness, reflecting the asymmetry of signal distribution; K u τ is the kurtosis, which measures the sharpness of the signal peaks; AC is the autocorrelation coefficient, which measures the similarity of the signals at different times; τ is the time delay; CC is the cross-correlation coefficient, which calculates the correlation between the signal and the reference signal.

[0178] (212) Spatial feature extraction:

[0179] The spatial feature vector S contains the following features:

[0180]

[0181] In the formula, v max ,v min ,v avg These represent the maximum, minimum, and average values ​​of the spatial field quantity, respectively. The average value is calculated as: v avg =(1 / M)∑v j v j Here, M represents the field quantity value at the spatial sampling point, and M represents the number of spatial sampling points. The gradient of the field quantity reflects the rate of change of the field quantity in space. The discrete calculation formula is: Δd is the spatial sampling interval. H is the entropy of the spatial domain, which measures the complexity of the spatial distribution. The formula is: H=-∑p(v j ) * logp(v j ), p(v j ) is the field quantity v j The probability of occurrence. E represents the energy characteristic of the spatial domain, calculated using the formula: E = ∑v j ^2.

[0182] (22) The feature fusion steps are as follows:

[0183] The feature fusion unit fuses the temporal feature vector T and the spatial feature vector S to construct a comprehensive feature vector F.

[0184] The fusion method adopts weighted fusion:

[0185] F = W T * T+W S * S

[0186] In the formula, W T and W S The weight vectors for the temporal and spatial features are respectively, satisfying WT +W S =1. The weights can be determined based on prior knowledge or training data to ensure that the fused feature vector F best represents the leakage situation.

[0187] (23) The spatiotemporal inversion steps are as follows:

[0188] The spatiotemporal inversion unit uses the fused feature vector F to determine the leakage situation and locate the leakage point through reverse reasoning and analysis.

[0189] (231) Leakage assessment:

[0190] Construct the leakage detection function D:

[0191] D = f(F; θ)

[0192] In the formula, θ is the discrimination parameter. When D is greater than the threshold D th A leak is detected when the threshold D is reached. th It can be determined through training with historical data to meet specific requirements for false positive and false negative rates.

[0193] (232) Leakage location:

[0194] Using a spatiotemporal inversion algorithm, the location L of the leak point is calculated based on the F feature vector and the spatiotemporal information of the sensor network.

[0195] The positioning calculation formula is as follows:

[0196] L = argmin||H(L) - F||^2

[0197] In the formula, H(L) is the theoretical characteristic model corresponding to the leakage point location L, which is established through physical field simulation or empirical model.

[0198] The main code is as follows:

[0199]

[0200]

[0201] (3) Hardware layer: The CMOS-MEMS integrated sensor chip converts the acquired analog signals into digital signals and performs preliminary data processing operations such as signal amplification and filtering to improve data quality and availability. The edge computing unit further compresses and preprocesses the data to reduce the amount of data transmitted, improve data transmission efficiency, and ensure that the data can be transmitted to the data layer quickly and stably.

[0202] (31) The signal processing includes:

[0203] The raw pressure signals collected from the pressure sensors in a car's braking system are usually quite weak, making it difficult to perform accurate subsequent analysis and processing directly. Therefore, a high-precision amplifier circuit is first required to amplify the signal.

[0204] (311) Signal amplification steps:

[0205] G = V out / V in

[0206] In the formula, G is the gain setting of the amplifier circuit, and V out V is the amplified output voltage. in The input is the original pressure signal voltage.

[0207] (312) Filtering and noise reduction steps:

[0208] The pressure signal generated during braking is often accompanied by high-frequency noise interference. A Butterworth low-pass filter is used to filter the amplified signal. The amplitude-frequency response of the Butterworth filter is H(jω):

[0209]

[0210] In the formula, ω is the angular frequency. c ω is the cutoff angular frequency, and n is the filter order.

[0211] (313) Data compression and encoding steps:

[0212] Data compression is achieved using discrete cosine transform (DCT). This compression coding method reduces the amount of data transmitted. The compression formula is as follows:

[0213]

[0214] In the formula, Y(u,v) is the compressed data matrix; X(x,y) is the original data matrix;

[0215] The main code is as follows:

[0216]

[0217] (4) Data layer: Store and manage the data of the hardware layer. By establishing a multi-physics joint calibration database, classify, store, organize and label historical and real-time data; use a spatiotemporal data compression encoder to optimize the data; further improve data storage efficiency and query speed, and provide high-quality data resources for intelligent analysis of the algorithm layer.

[0218] (41) The data classification and storage steps are as follows:

[0219] (411) The data is classified and stored using time series analysis methods. The data is divided into time windows, and the time window division formula is as follows:

[0220] T i =[t i ,t i +Δt)

[0221] In the formula, i = 1, 2, ..., N, where N is the number of time windows, and t i Let be the start time of the i-th time window.

[0222] (412) The formula for the tiered storage capacity of historical data is as follows:

[0223]

[0224] In the formula, k is the number of hierarchical storage levels, and C k Let be the storage capacity of the k-th layer.

[0225] (413) Real-time data storage and updates:

[0226] The real-time data buffer size is S realtime When new data arrives, the buffer is updated according to the first-in, first-out principle. If the buffer is full, the oldest data is discarded and the latest data is stored to ensure that the buffer always contains the latest real-time data.

[0227] (42) The steps for optimizing the data using a spatiotemporal data compression encoder are as follows:

[0228] (421) Data reception and preprocessing: The data layer receives raw data from the hardware layer, performs preliminary cleaning of information such as braking system pressure, temperature, and flow, removes obvious errors or outliers, and ensures the integrity and accuracy of the data.

[0229] (422) Spatiotemporal data compression coding:

[0230] (4221) Data matrix construction:

[0231] The preprocessed data is used to construct a data matrix D based on time series and spatial location:

[0232] D∈R^{m×n}

[0233] In the formula, R represents the real number space; m represents the number of time series; and n represents the number of spatial locations.

[0234] (4222) Singular Value Decomposition (SVD):

[0235] Perform singular value decomposition on the data matrix D:

[0236]

[0237] In the formula, U and V are orthogonal matrices; Σ is a diagonal matrix; and T is the matrix transpose.

[0238] (4223) Singular value truncation:

[0239]

[0240] In the formula, U k V is a left singular vector; k D is a right singular vector; k Σ is an approximate matrix; k is the number of maximum singular values; Σ k It is a diagonal matrix containing only the first k singular values.

[0241] (4224) Data compression and storage:

[0242] Will U k , Σ k and V k Encoding and storing the data significantly reduces storage requirements and improves data storage efficiency compared to the original data matrix D.

[0243] (423) Data query and decompression:

[0244] (4231) Query request parsing: Receives query requests from the algorithm layer and parses out the time range and spatial location information of the required data.

[0245] (4232) Encoded Data Retrieval: Based on the query request, quickly locate and retrieve the corresponding U in the stored compressed data. k , Σ k and V k .

[0246] (4233) Data decompression and restoration: using the retrieved U k , Σ k and V k Through matrix multiplication D k =U k Σ k V k ^T restores an approximate original data matrix D k This provides the algorithm layer with high-quality data resources that can be directly used for intelligent analysis.

[0247] (424) Data Quality Assessment and Feedback: Based on the data analysis results and feedback information from the algorithm layer, the compression and decompression processes of the data layer are evaluated and optimized. This further improves data quality, ensuring more accurate and reliable data support for the intelligent analysis of the algorithm layer, thereby effectively improving the accuracy and timeliness of automotive brake hydraulic leakage detection.

[0248] (5) Algorithm layer: Combining a large number of parameter-level basic models, the Dropout algorithm is used to deeply mine and intelligently analyze the multiphysics data stored in the data layer. By comparing real-time data with the feature model under normal operation, potential hydraulic leakage risks in the braking system are identified. When leakage signs are detected, corresponding alarm signals and decision suggestions are generated immediately.

[0249] (51) The construction process of the large number of parameter-level basic models is as follows:

[0250] (511) Construct a multi-scale network of numerous parameter-level basic models:

[0251]

[0252] In the formula, h l This is the output of the l-th layer; LayerNorm is for normalization; h l-1 This is the output of the (l-1)th layer;

[0253] MultiHead is a multi-head attention mechanism that allows the model to learn different attention patterns in parallel across different representation subspaces, thereby capturing richer information; Q l ,K l V l These are the query vector, key vector, and value vector for the l-th layer, respectively; W l Q W l K W l V These are the query weight matrix, key weight matrix, and value weight matrix, respectively; g is the global leak status code.

[0254] (512) Define the composite loss function L:

[0255]

[0256] In the formula, λ1, λ2, and λ3 are weighting coefficients used to balance the contributions of different loss terms to the total loss; L CE Cross-entropy loss; The weight matrix W l The square of the F-norm; E is the expected value; KL is the divergence, used to measure the difference between two probability distributions; p train p represents the probability distribution on the training set. test Let be the probability distribution on the test set.

[0257] A large number of parameter-level basic models are combined with the Dropout algorithm, along with a multi-scale network architecture and composite loss function for synergistic optimization, to achieve overfitting suppression and improved generalization ability in hydraulic leakage detection of automotive braking systems. The specific steps of the Dropout algorithm are as follows:

[0258] (513) Dropout algorithm model construction:

[0259]

[0260] In the formula, h t h is the output vector at the current time t; t-1 The output vector at current time t-1; Q t K t V t These are the query vector, key vector, and value vector for the current time t, respectively; W t Q W t K W t V These are the linear transformation matrices for the query vector, key vector, and value vector, respectively. The regularized module vector after input to the Dropout algorithm.

[0261] (514) Enhanced regularization using composite loss function:

[0262]

[0263] In the formula, L z This is the main part of the composite loss function, used to measure the difference between the model's predicted values ​​and the actual values; L ce Used to measure the difference between the predicted result and the true label; λ1 controls the weight of the cross-entropy loss in the total loss; W is the model weight matrix; λ is the Frobenius norm of the matrix to prevent overfitting; λ² is the control for L. z The strength of regularization; P train and P test The distributions of the training and test sets are given; KL(||) is used to measure the difference between the two distributions and reduce distribution bias; λ3 is the weight controlling the KL divergence term; E m For expectation operators; λ represents the transformed feature or hidden representation; λ4 controls the strength of sparsity regularization. λ is the rate of change of the feature over time; λ5 is the weight controlling the smoothness of the time series. For L z For the loss function when m j =1, the gradient includes the gradient of the cross-entropy loss and the sign function of the sparse regularization term. Represents the loss function L z right The gradient is used for updating model parameters. For loss function L z Hidden representation The partial derivatives are used for backpropagation; Here is a sign function used for sparsity regularization; when m j When m = 1, the gradient includes cross-entropy loss and sparse regularization term; when m = 1, the gradient includes cross-entropy loss and sparse regularization term. j When the gradient is 0, it means that the feature is not included in the update.

[0264] (515) Dynamic probability adaptive mechanism p (l) :

[0265]

[0266] In the formula, exp is the exponential function term; α is the sensitivity parameter controlling the change in probability; Used to normalize the Frobenius norm, avoiding excessively large or small norm values ​​due to differences in dimensions.

[0267] (516) Although the Dropout algorithm performs well in automotive brake hydraulic leakage detection systems, it still suffers from differences in the sensitivity of different network layers to dropping data and gradient estimation bias. Therefore, the Dropout algorithm is optimized using dynamic layered dropout probabilities and a gradient compensation mechanism. The optimization steps are as follows:

[0268] (5161) Dynamic stratified discard probability:

[0269]

[0270] In the formula, l is the index of the current layer (from 1 to the total number of layers L); γ is the depth decay factor; β is the gradient sensitivity coefficient; Let be the gradient norm of the loss function at layer l.

[0271] (5162) Gradient compensation mechanism:

[0272] Since Dropout randomly discards some neurons during training, it introduces variance in gradient estimation. Therefore, a gradient compensation term is introduced.

[0273]

[0274] In the formula, For L z gradient, For gradient operators; λ r This is the regularization coefficient, used to control the weight of the regularization term in gradient compensation; This is the regularization loss gradient, used to prevent the model from overfitting.

[0275]

[0276] In the formula, This represents taking the partial derivative with respect to the weight matrix W; This represents the summation of all terms from level 1 to level L; W l W represents the weight matrix of the Lth layer of the neural network. l (pre) This represents the weight matrix of the previous iteration of the neural network.

[0277] (5163) Scaling compensation:

[0278] In the formula, The optimized gradient; This is a scaling factor used to scale the compensated gradient when a neuron is activated; This represents the probability of a neuron being discarded, to prevent overfitting.

[0279] The improved Dropout algorithm significantly improves the handling of differences in the sensitivity of different network layers to dropout and gradient estimation bias, making it better suited for automotive brake hydraulic leakage detection systems and enabling the acquisition of the optimal loss function L. best .

[0280] (52) The feature model comparison steps are as follows:

[0281] (521) Define the characteristic contrast function D(t):

[0282] D(t) = ||F real (t)-F model ||2

[0283] In the formula, F real (t) is the real-time feature vector; F model This is the feature model under normal operating conditions.

[0284] (522) Introduce time series analysis and calculate the moving average D. avg (t):

[0285]

[0286] In the formula, Δt is the length of the time interval [t-Δt,t]; τ is the integration variable, representing any point in time within the time interval [t-Δt,t]; D(τ) is the value of the distance function D at time τ; This represents the integration of the distance function D(τ) over the time interval [t-Δt,t], i.e., the cumulative error.

[0287] (523) Determine the risk of hydraulic leakage (t):

[0288]

[0289] In the formula, T is the leakage threshold determined by experimental data.

[0290] Once a leak is detected, the algorithm layer immediately generates corresponding alarm signals and decision suggestions, providing a scientific basis for the alarm module. The alarm signal generation and decision suggestion steps are as follows:

[0291] (524) Determine the alarm signal A(t):

[0292]

[0293] In the formula, α is the alarm intensity coefficient, which is used to adjust the amplitude of the alarm signal.

[0294] (525) The most suitable decision recommendation D s (t):

[0295]

[0296] In the formula, argmax is used to find the parameter value that makes the function reach its maximum value; Match(D(t),R i ) represents D(t) and R i The degree of matching can be calculated through similarity metrics or threshold comparisons. R is a predefined set of decision rules.

[0297] The main code is as follows:

[0298]

[0299]

[0300] This invention has been verified through experiments, such as... Figure 4 The diagram compares the response speed of this invention with that of existing systems. Existing systems trigger an alarm based on a fixed threshold; a 25% pressure drop triggers an alarm after the pressure drop accumulates. This invention, through feature analysis, identifies minute leaks in advance, significantly shortening the response time by approximately 0.9 seconds. Figure 5 The algorithm layer of this invention illustrates the decision-making mechanism. Feature difference D(t) fuses temporal and spatial features, and the moving average Davg suppresses transient noise, ensuring robustness in leakage detection. Figure 6 The overfitting suppression mechanism of this invention is shown. After optimization, the generalization ability of the model is improved, and the accuracy of the test set reaches 95%. Figure 7The output of the visualization feature extraction module fuses temporal and spatial features. The system of this invention detects points that more closely approximate the leakage feature clusters, demonstrating the effectiveness of the feature fusion.

Claims

1. A system for detecting hydraulic leakage in automotive brakes based on spatiotemporal inversion twin networks, characterized in that, include: Data acquisition module: Collects key physical quantities data of the vehicle braking system, such as pressure, flow rate, vibration, temperature, dielectric constant, and chemical composition, through a sensor array; Feature extraction and spatiotemporal inversion module: The feature extraction unit is used to extract temporal and spatial features; feature The fusion unit is used to fuse temporal and spatial features to construct a comprehensive feature vector; the spatiotemporal inversion unit is used to determine the leakage of automotive brake hydraulic fluid and locate the leakage point through reverse reasoning and analysis. Hardware layer: Converts the acquired analog signals into digital signals, and performs preprocessing operations such as amplification, filtering, noise reduction, and compression encoding on the signals; Data layer: Stores and manages data from the hardware layer. By establishing a multiphysics joint calibration database, it classifies, stores, organizes, and labels historical and real-time data; and optimizes the data using a spatiotemporal data compression encoder. Algorithm layer: Combining a large number of parameter-level basic models, the Dropout algorithm is used to perform in-depth mining and intelligent analysis of multiphysics data stored in the data layer. By comparing real-time data with feature models under normal operating conditions, potential hydraulic leakage risks in the braking system are identified. When leakage signs are detected, corresponding alarm signals and decision suggestions are generated immediately.

2. The automotive brake hydraulic leakage detection system based on spatiotemporal inversion twin network according to claim 1, characterized in that, The sensor array includes: Nanopiezoelectric thin film array: Real-time monitoring of brake fluid pressure changes through the voltage output of the piezoelectric thin film array; FBG fiber grating array: for monitoring vibration characteristics of braking systems; MEMS ultrasonic Doppler sensor: Real-time monitoring of flow velocity changes; Quantum dot infrared thermal imager: Tracking the thermodynamic characteristics of evaporation and sensing local temperature fluctuations; Microwave resonant cavity sensor: monitoring the effect of cavitation on dielectric constant; Terahertz time-domain spectrometer: detects minute changes in chemical composition by measuring molecular polarization relaxation time.

3. The automotive brake hydraulic leakage detection system based on spatiotemporal inversion twin network according to claim 1, characterized in that, The feature extraction includes: (1) Temporal feature extraction: The temporal feature vector T includes the following elements: T=[μ,σ,S k ,K u ,AC,CC] Where μ is the mean of the signal; σ is the standard deviation of the signal, which measures the dispersion of the signal; x i y i S represents the number of sampling points; N represents the number of sampling points; S represents the number of sampling points. k Skewness, reflecting the asymmetry of signal distribution; K u τ is kurtosis, which measures the sharpness of signal peaks; AC is the autocorrelation coefficient, which measures the similarity of signals at different times; τ is the time delay; CC is the cross-correlation coefficient, which calculates the correlation between the signal and the reference signal. (2) Spatial domain feature extraction: The spatial domain feature vector S contains the following features: Among them, v max ,v min ,v avg These represent the maximum, minimum, and average values ​​of the spatial field quantity, respectively. The average value is calculated as: v avg =(1 / M)∑v j v j Here, M represents the field quantity value at the spatial sampling point, and M is the number of spatial sampling points. Let H be the gradient of the field quantity, reflecting the rate of change of the field quantity in space; H is the entropy of the spatial domain, measuring the complexity of the spatial distribution, and is calculated by the formula: H = -∑p(v j ) * logp(v j ), p(v j ) is the field quantity v j The probability of occurrence; E represents the energy characteristic of the spatial domain, calculated as: E = ∑v j ^2.

4. The automotive brake hydraulic leakage detection system based on spatiotemporal inversion twin network according to claim 1, characterized in that, The feature fusion method combines the temporal domain feature vector T and the spatial domain feature vector S to construct a comprehensive feature vector F. The fusion method uses weighted fusion. F=W T * T+W S * S Among them, W T and W S The weight vectors for the temporal and spatial features are respectively, satisfying W T +W S =1.

5. The automotive brake hydraulic leakage detection system based on spatiotemporal inversion twin network according to claim 1, characterized in that, The spatiotemporal inversion unit uses the fused feature vector F to determine the leakage situation and locate the leakage point through reverse reasoning and analysis. Leakage detection: Construct a leakage detection function D: D = f(F; θ) Where θ is the discrimination parameter, when D is greater than the threshold D th When the threshold D is reached, a leak is determined to have occurred. th The specific requirements for false positive and false negative rates are determined through training with historical data. Leakage point location: Using a spatiotemporal inversion algorithm, based on the F feature vector and the spatiotemporal information of the sensor network, the location L of the leak point is calculated; The positioning calculation formula is as follows: L=argmin||H(L)-F|| ^ 2 In the formula, H(L) is the theoretical characteristic model corresponding to the location L of the leakage point.

6. The automotive brake hydraulic leakage detection system based on spatiotemporal inversion twin network according to claim 1, characterized in that, The steps for classifying and storing the data are as follows: (1) The data is classified and stored using time series analysis methods. The data is divided into time windows, and the formula for dividing the time windows is as follows: T i =[t i ,t i +Δt) In the formula, i = 1, 2, ..., N, where N is the number of time windows, and t i Let be the start time of the i-th time window; (2) The formula for the hierarchical storage capacity of historical data is as follows: In the formula, k is the number of hierarchical storage levels, and C k Let k be the storage capacity of the k-th layer; (3) Real-time data storage and updates: The real-time data buffer size is S realtime When new data arrives, the buffer is updated according to the first-in, first-out principle. If the buffer is full, the oldest data is discarded and the latest data is stored to ensure that the buffer always contains the latest real-time data.

7. The automotive brake hydraulic leakage detection system based on spatiotemporal inversion twin network according to claim 1, characterized in that, The data optimization process using a spatiotemporal data compression encoder includes: (1) Data reception and preprocessing: The data layer receives raw data from the hardware layer and performs preliminary cleaning on the braking system pressure, temperature, and flow data to remove obvious errors or outliers, ensuring the integrity and accuracy of the data. (2) Spatiotemporal data compression coding: (21) Data matrix construction: The preprocessed data is used to construct a data matrix D based on time series and spatial location: D∈R ^ {m×n} In the formula, R represents the real number space; m represents the number of time series; and n represents the number of spatial locations. (22) Singular value decomposition: Perform singular value decomposition on the data matrix D: In the formula, U and V are orthogonal matrices; Σ is a diagonal matrix; and T is the matrix transpose. (23) Singular value truncation: In the formula, U k V is a left singular vector; k D is a right singular vector; k Σ is an approximate matrix; k is the number of maximum singular values; Σ k It is a diagonal matrix containing only the first k singular values; (24) Data compression and storage: U k , Σ k and V k Encode and store; (3) Data query and decompression: (31) Query request parsing: Receive query requests from the algorithm layer and parse out the time range and spatial location information of the required data; (32) Encoded data retrieval: Based on the query request, locate and retrieve the corresponding U in the stored compressed data. k , Σ k and V k ; (33) Data decompression and restoration: using the retrieved U k , Σ k and V k Through matrix multiplication D k =U k Σ k V k ^T restores an approximate original data matrix D k This provides the algorithm layer with high-quality data resources that can be directly used for intelligent analysis; (4) Data quality assessment and feedback: Based on the analysis results and feedback information of the data by the algorithm layer, the compression and decompression process of the data layer is evaluated and optimized.

8. The automotive brake hydraulic leakage detection system based on spatiotemporal inversion twin network according to claim 1, characterized in that, The construction process of the large number of parameter-level basic models is as follows: (1) Construct a multi-scale network of numerous parameter-level basic models: In the formula, h l This is the output of the l-th layer; LayerNorm is for normalization; h l-1 This represents the output of layer l-1; MultiHead is a multi-head attention mechanism that allows the model to learn different attention patterns in parallel across different representation subspaces; Q l ,K l V l These are the query vector, key vector, and value vector for the l-th layer, respectively; W l Q W l K W l V These are the query weight matrix, key weight matrix, and value weight matrix, respectively; g is the global leak status code. (2) Define the composite loss function L: In the formula, λ1, λ2, and λ3 are weighting coefficients used to balance the contributions of different loss terms to the total loss; L CE Cross-entropy loss; The weight matrix W l The square of the F-norm; E is the expected value; KL is the divergence, used to measure the difference between two probability distributions; p train p represents the probability distribution on the training set. test The probability distribution on the test set; A large number of parameter-level basic models are combined with the Dropout algorithm, and synergistically optimized with a multi-scale network architecture and composite loss function to achieve overfitting suppression and improved generalization ability in hydraulic leakage detection of automotive braking systems. The specific steps of the Dropout algorithm are as follows: (3) Dropout algorithm model construction: In the formula, h t h is the output vector at the current time t; t-1 The output vector at current time t-1; Q t K t V t These are the query vector, key vector, and value vector for the current time t, respectively; W t Q W t K W t V These are the linear transformation matrices for the query vector, key vector, and value vector, respectively. The regularized block vector after input to the Dropout algorithm; (4) Enhanced regularization using composite loss function: In the formula, L z This is the main part of the composite loss function, used to measure the difference between the model's predicted values ​​and the actual values; L ce Used to measure the difference between the predicted result and the true label; λ1 controls the weight of the cross-entropy loss in the total loss. W is the model weight matrix; λ is the Frobenius norm of the matrix to prevent overfitting; λ² is the control for L. z The strength of regularization; P train and P test The distributions of the training and test sets are given; KL(||) is used to measure the difference between the two distributions and reduce distribution bias; λ3 is the weight controlling the KL divergence term; E m For expectation operators; The transformed feature or hidden representation; λ4 controls the strength of sparsity regularization; λ is the rate of change of the feature over time; λ5 is the weight controlling the smoothness of the time series. For L z For the loss function when m j =1, the gradient includes the gradient of the cross-entropy loss and the sign function of the sparse regularization term; Represents the loss function L z right The gradient is used for model parameter updates; For loss function L z Hidden representation The partial derivatives are used for backpropagation; Here is a sign function used for sparsity regularization; when m j When m = 1, the gradient includes cross-entropy loss and sparse regularization term; when m = 1, the gradient includes cross-entropy loss and sparse regularization term. j When the gradient is 0, the gradient is 0, indicating that the feature is not included in the update. (5) Dynamic probability adaptive mechanism p (l) : In the formula, exp is the exponential function term; α is the sensitivity parameter controlling the change in probability; Used to normalize the Frobenius norm, avoiding excessively large or small norm values ​​due to differences in dimensions.

9. The automotive brake hydraulic leakage detection system based on spatiotemporal inversion twin network according to claim 1, characterized in that, The Dropout algorithm is optimized for use in an automotive brake hydraulic leakage detection system. A gradient compensation mechanism is employed to optimize the Dropout algorithm, and the optimization steps are as follows: (1) Dynamic stratified discard probability: In the formula, l is the current layer index; γ is the depth decay factor; β is the gradient sensitivity coefficient; Let be the gradient norm of the loss function at layer l; (2) Gradient compensation mechanism: Since Dropout randomly discards some neurons during training, it introduces variance in gradient estimation. Therefore, a gradient compensation term is introduced. In the formula, For L z gradient, For gradient operators; λ r This is the regularization coefficient, used to control the weight of the regularization term in gradient compensation; This is the regularization loss gradient, used to prevent the model from overfitting; In the formula, This represents taking the partial derivative with respect to the weight matrix W; This represents the summation of all terms from level 1 to level L; W l W represents the weight matrix of the Lth layer of the neural network. l (pre) This represents the weight matrix of the previous iteration of the neural network; (3) Scaling compensation: In the formula, The optimized gradient; This is a scaling factor used to scale the compensated gradient when a neuron is activated; This represents the probability of a neuron being discarded, preventing overfitting.

10. The automotive brake hydraulic leakage detection system based on spatiotemporal inversion twin network according to claim 1, characterized in that, The feature model comparison steps are as follows: (1) Define the feature comparison function D(t): D(t)=||F real (t)-F model ||2 In the formula, F real (t) is the real-time feature vector; F model This is the feature model under normal operating conditions; (2) Introduce time series analysis and calculate the moving average D. avg (t): In the formula, Δt is the length of the time interval [t-Δt,t]; τ is the integration variable, representing any point in time within the time interval [t-Δt,t]; D(τ) is the value of the distance function D at time τ; Indicates the time interval Integrating the distance function D(τ) over the interval [t-Δt,t], i.e., accumulating the error; (3) Determine the hydraulic leakage risk Leakage(t): In the formula, T is the leakage threshold determined by experimental data; When a leak is detected, the algorithm layer immediately generates a corresponding alarm signal and decision suggestion. The steps for generating the alarm signal and making decision suggestions are as follows: (4) Determine the alarm signal A(t): In the formula, α is the alarm intensity coefficient, which is used to adjust the amplitude of the alarm signal; (5) The most suitable decision recommendation D s (t): In the formula, argmax is used to find the parameter value that makes the function reach its maximum value; Match(D(t),R i ) represents D(t) and R i The degree of matching, where R is a predefined set of decision rules.

Images