Method for constructing a prediction model for automotive brake pad wear based on digital twins

Abstract

This invention relates to the field of automotive braking system technology, specifically disclosing a method for constructing a brake pad wear prediction model based on digital twins. The method involves collecting source domain data including braking condition time-series parameters, wear thickness sequences, and material batch identifiers; obtaining batch-independent wear features through feature encoding mapping and batch adversarial constraints; constructing a wear prediction mapping to output a basic wear depth sequence and adjusting it in reverse until convergence; when introducing a new batch of material, collecting a small amount of measured data, freezing the two mappings, and adjusting the batch discrimination mapping to calculate the wear coefficient correction; embedding the correction into the digital twin model, acquiring the current batch and operating condition parameters in real time, obtaining the basic wear depth through feature encoding mapping and wear prediction mapping, and then multiplying it by the dynamic correction obtained based on the operating condition distribution comparison to output the final wear prediction result; this invention achieves rapid calibration and accurate prediction of brake pad wear across batches and operating conditions.

Description

Technical Field

[0001] This invention relates to the field of automotive braking system technology, and more specifically to a method for constructing a digital twin-based automotive brake pad wear prediction model. Background Technology

[0002] As a critical safety component of the vehicle braking system, the accurate prediction of brake pad wear is crucial for maintenance cycle planning and ensuring driving safety. Digital twin technology offers a new technical approach for brake pad wear prediction by constructing a virtual model corresponding to the physical entity. Existing technologies commonly employ wear prediction methods including physical models based on finite element simulation and long short-term memory network models based on historical wear data.

[0003] Existing brake pad wear prediction models for digital twins struggle to adapt to new batches of materials (e.g., formula changes) or new driving conditions (e.g., gentle driving) when the training data distribution is fixed, leading to increased prediction bias. Furthermore, acquiring new batches of material wear characteristics requires extensive bench testing or real-vehicle road testing, resulting in long cycles and high costs. Therefore, providing a predictive model construction method that requires only a small amount of new batch measured data for rapid model calibration and can dynamically adjust the prediction output based on the differences between real-time operating conditions and the training data distribution is a pressing technical problem in this field. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing a digital twin-based automotive brake pad wear prediction model to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: A method for constructing a digital twin-based automotive brake pad wear prediction model includes the following steps: S1, the source domain data collected includes: multiple sets of braking condition timing parameters, wear thickness sequence corresponding to each set of braking condition timing parameters, and material batch identifier of the brake pad to which each set of data belongs; S2, perform feature encoding mapping on the timing parameters of braking conditions to obtain wear feature representation, and introduce batch adversarial constraints to obtain batch-independent wear features by minimizing the mutual information between the wear feature representation and the material batch identifier; S3, input batch-independent wear features into the wear prediction mapping, output the basic wear depth sequence, and adjust the feature encoding mapping and wear prediction mapping in reverse according to the deviation between the basic wear depth sequence and the wear thickness sequence until the deviation converges; S4. When a new batch of material is introduced, a small amount of measured wear data of the new batch of material is collected, the feature coding mapping and wear prediction mapping are frozen, only the batch discrimination mapping is adjusted to distinguish between the new and old batches, and the wear coefficient correction amount of the new batch of material relative to the source domain batch is calculated from the adjusted batch discrimination mapping. S5 embeds the wear coefficient correction amount into the digital twin model. During the real-time prediction process, it obtains the material batch identifier of the current brake pad and the real-time braking condition time sequence parameters. The basic wear depth is obtained by feature encoding mapping and wear prediction mapping in sequence, and then multiplied by the corresponding wear coefficient correction amount to output the final wear prediction result.

[0006] As a further aspect of the present invention: obtaining batch-independent wear characteristics specifically includes: The braking condition timing parameters are input into a fully connected encoding mapping, and the initial wear characteristics are output. The initial wear characteristics are simultaneously input into the wear prediction branch and the batch discrimination branch. A gradient inversion layer is set before the batch discrimination branch to invert the gradient sign during backpropagation. Simultaneously, a mutual information estimation network is constructed, using the divergence of the product of the joint distribution of initial wear features and material batch identifiers and the marginal distribution as the loss term. Through adversarial training, the distribution of initial wear features under different batch identifiers is aligned, and finally batch-independent wear features are output.

[0007] As a further aspect of the present invention: a gradient inversion layer is set before the batch discrimination branch to invert the gradient sign during backpropagation, specifically including: The initial wear characteristics are received and forwarded to the batch discrimination branch to calculate the batch discrimination loss; During backpropagation, the gradient of the batch discrimination loss with respect to the initial wear characteristics is multiplied by a negative preset coefficient to obtain the inverted gradient; The inverted gradient is backpropagated to the fully connected encoding map to update its parameters, while the parameters of the batch discrimination branch are updated in the normal gradient direction. The preset coefficient increases linearly from zero to one with the number of training rounds.

[0008] As a further aspect of the present invention: the wear prediction mapping specifically includes: Batch-independent wear features are split into multiple time segments according to time sequence, and the features within each segment are weighted and aggregated to obtain the segment wear contribution value; The wear contribution value of each segment is sequentially input into the loop mapping with memory unit. The memory unit maintains the long-term wear accumulation state and outputs the initial wear depth step by step. The residuals of the products of the initial wear depth and the braking force and temperature parameters at the corresponding time step are superimposed to generate the basic wear depth sequence. The absolute deviation between the basic wear depth sequence and the wear thickness sequence is used as the loss. Backpropagation updates the cyclic mapping, the weights of the weighted aggregation, and the coefficients of the residual superposition in sequence.

[0009] As a further aspect of the present invention: the output of the initial wear depth step-by-step specifically includes: Get the memory state output from the previous time step; The attenuation coefficient is calculated based on the braking force and temperature at the current time step, so that the attenuation coefficient is negatively correlated with the braking force and positively correlated with the temperature. The fragment wear contribution value of the current time step is weighted and fused with the memory state of the previous time step according to the decay coefficient to obtain the current memory state; Apply a nonlinear compression mapping to the current memory state and output the initial wear depth at the current time step.

[0010] As a further aspect of the present invention: the calculation of the wear coefficient correction amount of the new batch of material relative to the source domain batch from the adjusted batch discrimination mapping specifically includes: The braking condition time series parameters in a small amount of measured wear data of the new batch of materials are sequentially processed by frozen feature encoding mapping and wear prediction mapping to output the basic wear depth sequence of the new batch. By comparing the basic wear depth sequence of the new batch with the measured wear thickness sequence corresponding to the new batch of materials step by step, multiple wear deviation ratios are obtained. Weighted median filtering is applied to each wear deviation ratio, where the weights are positively correlated with the braking intensity of the corresponding time step. After removing ratios that deviate from the median by more than a preset threshold, the median of the remaining ratios is taken and used as the wear coefficient correction amount.

[0011] As a further aspect of the present invention: the output of the new batch basic wear depth sequence specifically includes: The braking condition timing parameters in the new batch of measured wear data are resampled along the time axis to align with the sampling interval of the source domain data. The aligned time series parameters are segmented, and the weighted moving average of each parameter in each segment is calculated to obtain a smoothed operating condition sequence. The smooth working condition sequence is input into the frozen feature encoding mapping step by step, and the frozen wear features are output. The frozen wear features are input into the frozen wear prediction mapping, and the intermediate depth value at each time step is output. The intermediate depth values ​​at all times are spliced ​​together in chronological order to form a new batch of basic wear depth sequence.

[0012] As a further aspect of the present invention: the output of the final wear prediction result specifically includes: The current real-time braking condition time series parameters are compared with the source domain condition parameter set used during feature encoding mapping training, and the local offset of each time step is calculated. A time-series accumulation is performed on the local offset to obtain a confidence weight between 0 and 1, which decreases as the offset increases; Multiply the wear coefficient correction by the confidence weight to obtain the dynamic correction. Multiply the base wear depth by the dynamic correction amount to output the final wear prediction result.

[0013] As a further aspect of the present invention: the calculation of the local offset at each time step specifically includes: Extract multiple parameter values ​​from the current time step in the current real-time braking condition timing parameters to form a real-time parameter vector; Extract multiple reference parameter vectors corresponding to the current time step position from the source domain operating condition parameter set, calculate the sum of the absolute differences between the components of each reference parameter vector and the real-time parameter vector, and obtain the original deviation sequence. Sort the deviation values ​​in the original deviation sequence from smallest to largest, and take the deviation value that is in the preset percentile after sorting as the candidate offset of the time step. The candidate offset is fused with the local offset of the previous time step by an exponential weighting to obtain the local offset of the current time step.

[0014] The beneficial effects of this invention are: (1) For brake pad materials with new formulations and new processes, only a small amount of measured wear data needs to be collected to quickly obtain the wear coefficient correction amount. There is no need to repeat the training process of the complete feature encoding mapping and wear prediction mapping, which reduces the workload of bench calibration and vehicle verification before the new batch of materials are put into application.

[0015] (2) In the real-time prediction process, the confidence weight of the correction amount is dynamically adjusted by comparing the current operating condition parameters with the source domain operating condition distribution step by step. When the vehicle operating conditions (such as extreme temperature and emergency braking frequency) exceed the coverage of the source domain data, the prediction results automatically reduce their dependence on the fixed correction amount, avoid prediction deviation caused by operating condition extrapolation, and improve the robustness of the digital twin model in different driving environments. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the process for obtaining batch-independent wear characteristics in this invention; Figure 3This is a flowchart of the wear prediction mapping in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention provides a method for constructing a vehicle brake pad wear prediction model based on digital twins, comprising the following steps: S1, the source domain data collected includes: multiple sets of braking condition timing parameters, wear thickness sequence corresponding to each set of braking condition timing parameters, and material batch identifier of the brake pad to which each set of data belongs; S2, perform feature encoding mapping on the timing parameters of braking conditions to obtain wear feature representation, and introduce batch adversarial constraints to obtain batch-independent wear features by minimizing the mutual information between the wear feature representation and the material batch identifier; S3, input batch-independent wear features into the wear prediction mapping, output the basic wear depth sequence, and adjust the feature encoding mapping and wear prediction mapping in reverse according to the deviation between the basic wear depth sequence and the wear thickness sequence until the deviation converges; S4. When a new batch of material is introduced, a small amount of measured wear data of the new batch of material is collected, the feature coding mapping and wear prediction mapping are frozen, only the batch discrimination mapping is adjusted to distinguish between the new and old batches, and the wear coefficient correction amount of the new batch of material relative to the source domain batch is calculated from the adjusted batch discrimination mapping. S5 embeds the wear coefficient correction amount into the digital twin model. During the real-time prediction process, it obtains the material batch identifier of the current brake pad and the real-time braking condition time sequence parameters. The basic wear depth is obtained by feature encoding mapping and wear prediction mapping in sequence, and then multiplied by the corresponding wear coefficient correction amount to output the final wear prediction result.

[0020] In S1, the source domain data collected includes: multiple sets of braking condition timing parameters, wear thickness sequences corresponding to each set of braking condition timing parameters, and the material batch identifier of the brake pad to which each set of data belongs, specifically including: Multiple sets of braking condition timing parameters were collected. Thin-film thermocouple temperature sensors were installed on a set of test vehicles, embedded at the center of the back of the brake pad friction material, to measure the friction interface temperature during braking. Simultaneously, piezoelectric force sensors were installed at the brake caliper piston to measure braking pressure. Wheel speed sensor signals were read via the vehicle's CAN bus and converted to obtain the wheel linear velocity. All sensors collected data synchronously at a sampling frequency of 100 Hz. For each braking event, from the start of brake pedal travel until the vehicle speed dropped to zero, three continuous curves were recorded showing the changes in braking pressure, friction interface temperature, and wheel linear velocity over time, forming a set of braking condition timing parameters.

[0021] Collect wear thickness sequences corresponding to the timing parameters of each braking condition. After each braking event, park the vehicle on a level surface and use an ultrasonic pulse-echo thickness gauge. Hold the probe vertically against a preset measurement point on the outer surface of the brake pad backing plate. This measurement point is laser-etched on the backing plate to ensure consistent measurement position each time. The thickness gauge displays the current remaining thickness value, records this value, and associates it with the corresponding braking condition timing parameters. Repeat the above process, measuring the thickness again after a certain number of braking events (e.g., 50 or 100 times), thus obtaining a wear thickness sequence that varies with time or number of braking events.

[0022] Record the material batch identifier of the brake pad to which each set of data belongs. During the brake pad manufacturing process, each batch of friction material formulation is assigned a unique batch code, which is stamped or laser-engraved on the edge of the brake pad backing plate. Before installing the brake pads onto the test vehicle, the batch code is entered into the data acquisition system by visual inspection or barcode scanning, and then linked to the subsequently acquired braking condition timing parameters and wear thickness sequence, thereby completing the labeling of the material batch identifier in the source domain data.

[0023] Please see Figure 2 As shown, in S2, feature encoding mapping is performed on the timing parameters of the braking condition to obtain the wear feature representation. Simultaneously, batch adversarial constraints are introduced. By minimizing the mutual information between the wear feature representation and the material batch identifier, batch-independent wear features are obtained, specifically including: A fully connected encoding map is constructed, consisting of three sequentially connected fully connected layers: the first layer has an input node count equal to the number of parameters at each time step in the braking condition time-series parameters (e.g., each time step includes three parameters: braking pressure, friction interface temperature, and wheel linear velocity; if the time series length is 100 time steps, the parameters of the 100 time steps are expanded into 300 input nodes), and the first layer has 128 output nodes; the second layer has 128 input nodes and 64 output nodes; the third layer has 64 input nodes and 32 output nodes. A linear rectified function is used as the activation function after each fully connected layer. A set of braking condition time-series parameters is arranged step-by-step and input into this encoding map, outputting a 32-dimensional vector, called the initial wear feature.

[0024] The wear prediction branch and batch discrimination branch are constructed as follows: The wear prediction branch consists of two fully connected layers. The first layer takes 32 input nodes and outputs 16 nodes, while the second layer takes 16 input nodes and outputs 1 node. This branch ultimately outputs a scalar representing the wear depth at the current time step. The batch discrimination branch also consists of two fully connected layers. The first layer takes 32 input nodes and outputs 16 nodes, while the second layer takes 16 input nodes and outputs 2 nodes. The two output values ​​correspond to the unnormalized scores of the old and new batches, respectively. A gradient inversion layer is set before the input of the batch discrimination branch. During forward propagation, this gradient inversion layer passes the initial wear features of the input to the batch discrimination branch intact. During backpropagation, perform the following three steps: The first step is to receive the gradient values ​​of the batch discrimination branch loss function with respect to the batch discrimination branch input; The second step is to multiply the gradient value by a preset coefficient, which ranges from negative one to zero, specifically a negative preset coefficient (for example, if the preset coefficient is 0.5, multiply by negative 0.5). The third step involves using the gradient value multiplied by a negative coefficient as the gradient for backpropagation to the fully connected encoding map, thereby updating the parameters of the fully connected encoding map. Simultaneously, the parameters of the batch discrimination branch are updated in the normal gradient direction (i.e., without reversal).

[0025] The aforementioned preset coefficient increases linearly from zero to one as the number of training rounds increases. Specifically, the total number of training rounds is set to 200. The preset coefficient is 0 in the first round, and increases by 0.005 for each round until it reaches 1 in the 200th round. During the linear increase, the preset coefficient always takes a positive value, and is actually multiplied by the opposite of this positive value during backpropagation.

[0026] A mutual information estimation network is constructed to minimize the mutual information between initial wear characteristics and material batch identifiers. This mutual information estimation network is a separate neural network consisting of three fully connected layers: The number of input nodes in the first layer is the dimension of the initial wear feature (32 dimensions) plus the dimension of the material batch identifier (the batch identifier is encoded as a 2-dimensional unique heat vector, for a total of 34 input nodes), and the number of output nodes is 64. The second layer has 64 input nodes and 32 output nodes; the third layer has 32 input nodes and 1 output node, outputting a single real value.

[0027] The training objective of this network is to estimate the lower bound of the mutual information between the initial wear characteristics and the material batch identifier.

[0028] The specific implementation is as follows: A batch of samples is randomly sampled from the training data. Each sample contains an initial wear feature vector and its corresponding one-hot encoded material batch identifier. This pair of data is input into the mutual information estimation network to obtain the first output value. The initial wear feature vectors from the same batch of samples are paired with the randomly shuffled material batch identifiers and input into the same network to obtain the second output value. The arithmetic mean of the first output value is calculated, and the arithmetic mean of the second output value is subtracted from the arithmetic mean of the second output value. The difference is the mutual information estimate. This mutual information estimate is added as a loss term to the total loss, and adversarial training is used to make the mutual information estimate approach zero.

[0029] During training, the fully connected encoding mapping attempts to minimize the mutual information estimate, while the mutual information estimation network itself attempts to maximize the estimate, and the two optimize alternately.

[0030] Through the combined effect of the adversarial training and gradient inversion layer, the initial wear features output by the fully connected encoding map gradually become independent of the material batch identifier, making the feature distribution under different batch identifiers tend to be consistent. The 32-dimensional feature vector output by the fully connected encoding map at this point is then used as the batch-independent wear features for subsequent wear prediction.

[0031] Please see Figure 3 As shown, in S3, batch-independent wear features are input into the wear prediction mapping, and a basic wear depth sequence is output. The feature encoding mapping and wear prediction mapping are then adjusted inversely based on the deviation between the basic wear depth sequence and the wear thickness sequence until the deviation converges. Specifically, this includes: The process involves constructing time-segment splits and weighted aggregation: Batch-independent wear features are divided into multiple consecutive time segments according to the time axis. Each segment contains a fixed number of time steps, for example, 10 time steps per segment. Let the total number of time steps be T, then the number of segments is T divided by 10 and rounded up. For each time segment, a weighted sum is performed on the batch-independent wear feature vectors across all time steps within that segment. The weight of each time step is determined by a learnable weight coefficient, initially set to 1 divided by the segment length and updated during subsequent backpropagation. The weighted sum yields the aggregated vector for that segment, called the segment wear contribution value. Each time segment corresponds to one segment wear contribution value, thus compressing a time series of length T into a contribution value sequence of length equal to the number of segments.

[0032] Construct a circular mapping with a memory unit: This circular mapping maintains a memory state, which is a scalar value, initially set to 0. Perform the following operations sequentially for each time segment.

[0033] The first step is to obtain the memory state of the previous time segment (the previous memory state of the first segment is 0).

[0034] The second step is to calculate a braking force attenuation coefficient based on the average braking force and average temperature within the current time segment. The average braking force is calculated as the arithmetic mean of the braking pressure values ​​across all time steps within the current segment; the average temperature is calculated in the same way. The formula for calculating the braking force attenuation coefficient is as follows: ;in, Indicates the first The decay coefficient for each time segment, with a value ranging from 0 to 1; This represents the average braking pressure within the current segment, expressed in kilonewtons. This represents the average friction interface temperature within the current segment, expressed in degrees Celsius. This formula ensures that when both the average braking force and average temperature increase simultaneously, the exponential term... As the denominator increases, the attenuation coefficient decreases; conversely, when both decrease, the attenuation coefficient approaches 1.

[0035] The third step is to calculate the fragment wear contribution value of the current time segment (denoted as...). The memory state of the previous time segment (denoted as) The current memory state is obtained by performing weighted fusion. The fusion method is as follows: ,in The above refers to the attenuation coefficient. This formula indicates that a larger attenuation coefficient means more historical memory is retained; a smaller attenuation coefficient means more contribution from the current segment is accepted.

[0036] The fourth step is to analyze the current memory state. A nonlinear compression mapping is applied, specifically using the hyperbolic tangent function, meaning the output value equals the hyperbolic tangent function applied to... The result is that the output value is the initial wear depth of the current time segment. This nonlinear compression mapping restricts the value range of the memory state to between -1 and +1. Since the wear depth is a non-negative value, it will be corrected later by residual superposition.

[0037] Constructing residual superposition to generate a basic wear depth sequence: For each time step within each time segment, add the initial wear depth corresponding to that segment output by the above cyclic mapping to the product of braking force and temperature at that time step. Specifically, let the... The time step index corresponding to each time segment is: Then, the base wear depth at that time step is equal to the initial wear depth of that segment plus a residual term. The residual term is calculated as follows: [The remaining text appears to be incomplete and requires further context.] The braking force value (in kilonewtons) at each time step is multiplied by the temperature value (in degrees Celsius), and then the product is multiplied by a learnable coefficient, initially set to 0.01. The calculation results of all time steps are arranged in chronological order to obtain the basic wear depth sequence.

[0038] Construction of loss function and backpropagation process: The absolute deviation between the base wear depth sequence and the wear thickness sequence collected from the source domain data is used as the loss value.

[0039] Specifically, for each time step, the absolute value of the difference between the baseline wear depth and the measured wear thickness corresponding to that time step is calculated. The absolute values ​​of all time steps are summed, and then divided by the total number of time steps to obtain the average absolute deviation. Using this average absolute deviation as the optimization objective, the parameters of the following three parts are updated sequentially using the gradient descent method: First, the parameters in the cyclic mapping include the learnable parameters in the calculation of the decay coefficient used when fusing memory states (the exponential part in the formula actually implies the product of braking force and temperature, which does not require additional parameters, but the coefficients in the residual superposition need to be updated). Second, the learnable weight coefficients at each time step in the weighted aggregation process; Thirdly, there are the learnable coefficients in the residual stacking. During backpropagation, the gradient starts from the loss value and passes through the residual stacking layer, the cyclic mapping layer, and the weighted aggregation layer in sequence, finally reaching the feature encoding mapping (but the parameters of the feature encoding mapping are also updated in this step, because its output is used as input).

[0040] Repeat the forward calculation and backpropagation process as described above until the mean absolute deviation converges to below a preset threshold, for example, if the loss value decreases by less than 0.001 mm over 10 consecutive training epochs, then stop training. At this point, the trained feature encoding map and wear prediction map are obtained, which can be used for rapid calibration and real-time prediction of subsequent batches of materials.

[0041] In S4, when a new batch of material is introduced, a small amount of measured wear data of the new batch is collected. The feature encoding mapping and wear prediction mapping are frozen. Only the batch discrimination mapping is adjusted to distinguish between new and old batches. The wear coefficient correction amount of the new batch of material relative to the source domain batch is calculated from the adjusted batch discrimination mapping. Specifically, this includes: The braking condition timing parameters in the new batch of measured wear data are resampled along the time axis to align with the sampling interval of the source domain data. Specifically, the sampling frequency during source domain data acquisition is 100 Hz, meaning that braking pressure, friction interface temperature, and wheel linear velocity are recorded at a time point every 0.01 seconds.

[0042] If the new batch of measured data uses a different sampling frequency (e.g., 50 Hz), it is resampled to 100 Hz using linear interpolation. The specific operation of linear interpolation is as follows: for a missing time point, the two nearest neighbor measured values ​​before and after that time point are taken, and their arithmetic mean is calculated as the estimated value for that time point. After resampling, the new batch of data and the source domain data have the same physical time interval at each time step.

[0043] The aligned timing parameters are then segmented and smoothed. The entire resampled braking event timing data is divided into multiple consecutive time windows, each containing 20 time steps. For each time window, the arithmetic mean of braking pressure, friction interface temperature, and wheel linear velocity for all time steps within the window are calculated to obtain the smoothed braking pressure, smoothed temperature, and smoothed speed for that window.

[0044] A weighted moving average is used to further smooth the data between adjacent windows, with the weights set as follows: the current window has a weight of 0.6, the previous window has a weight of 0.2, and the next window has a weight of 0.2. All windows are traversed in chronological order to obtain a smoothed condition sequence, the length of which is equal to the number of windows.

[0045] The smoothed operating condition sequence is input into the frozen feature encoding map step by step. Here, "frozen" means that the parameters of the feature encoding map remain fixed in the fourth step, i.e., using the parameter values ​​saved after training in the third step, without further gradient updates. For each time step in the smoothed operating condition sequence, the three values ​​of smoothed braking pressure, smoothed temperature, and smoothed speed are arranged sequentially into a three-dimensional vector and input into the feature encoding map. This feature encoding map contains three fully connected layers, each with 3 nodes for the input, 128 nodes for the first hidden layer, 64 nodes for the second hidden layer, and 32 nodes for the output. Each layer is followed by a linear rectified activation function. After mapping, each time step outputs a 32-dimensional vector, called the frozen wear feature.

[0046] The frozen wear features are input into the frozen wear prediction map, which outputs the intermediate depth value at each time step. The wear prediction map is also in a frozen state, and it contains time segment splitting, weighted aggregation, cyclic mapping with memory units, and residual superposition.

[0047] Specifically, the 32-dimensional frozen wear features of each time step are arranged chronologically, and every 10 consecutive time steps are divided into a segment. The features within each segment are weighted and aggregated (the weight coefficients are the values ​​obtained from training in step three) to obtain the segment wear contribution value. This contribution value is sequentially input into a cyclic mapping, which maintains a memory state with an initial value of 0. The attenuation coefficient is calculated based on the average braking force and average temperature of the current segment, and the memory state is updated according to the fusion formula. After compression using the hyperbolic tangent function, the initial wear depth is output. This initial wear depth is then added to the product of braking force and temperature for each time step within the segment (multiplied by the training-obtained residual coefficient of 0.01) to obtain the intermediate depth value for each time step. The intermediate depth values ​​of all time steps are concatenated chronologically to form a new batch of basic wear depth sequences.

[0048] This completes the output of the basic wear depth sequence for the new batch and the calculation of the wear coefficient correction. The first step involves comparing the baseline wear depth sequence of the new batch with the corresponding measured wear thickness sequence for each time step. The measured wear thickness sequence is the remaining thickness value measured at the same measurement point after each braking event using an ultrasonic pulse-echo thickness gauge when acquiring the new batch data. For each time step, the baseline wear depth is divided by the measured wear thickness for that time step to obtain the wear deviation ratio. If the measured thickness is zero, that time step is skipped.

[0049] The second step involves performing weighted median filtering on each wear deviation ratio: The braking intensity at each time step is calculated, defined as the product of the braking pressure and the wheel linear velocity at that time step. The braking intensity value at each time step is divided by the maximum braking intensity in the entire sequence to obtain a normalized weight, with the normalized weight ranging from 0 to 1. Each wear deviation ratio is then assigned a weight according to its corresponding normalized weight at that time step; that is, ratios with larger weights have higher priority in subsequent median filtering.

[0050] Sort all wear deviation ratios in ascending order of their numerical value, and record the weight corresponding to each ratio. Start accumulating the weights from the smallest ratio, and when the accumulated weights reach half of the total weights, that ratio is the weighted median.

[0051] The third step is to remove ratios that deviate from the weighted median by more than a preset threshold. The preset threshold is set to 20% of the weighted median.

[0052] The specific operation is as follows: For each wear deviation ratio, calculate the absolute value of the difference between it and the weighted median. If the absolute value is greater than 20% of the weighted median, then the ratio is removed. The remaining ratios that are not removed constitute the set of valid ratios.

[0053] The fourth step is to take the median of the remaining effective ratio set (if the number of values ​​in the set is odd, take the middle value; if it is even, take the arithmetic mean of the two middle values), and use this median as the wear coefficient correction amount. This correction amount reflects the overall scaling ratio of the new batch of material in terms of wear rate relative to the source batch.

[0054] For example, if the correction value is 0.7, it means that under the same working conditions, the wear depth of the new batch of material is only 70% of that of the source batch of material. This correction value is used in the real-time prediction in the subsequent fifth step.

[0055] In S5, the wear coefficient correction is embedded in the digital twin model. During real-time prediction, the material batch identifier of the current brake pad and the real-time braking condition time series parameters are obtained. The basic wear depth is obtained by feature encoding mapping and wear prediction mapping, and then multiplied by the corresponding wear coefficient correction to output the final wear prediction result, which specifically includes: During real-time prediction, the material batch identifier of the current brake pad and real-time braking condition timing parameters are acquired. These real-time braking condition timing parameters include braking pressure, friction interface temperature, and wheel linear velocity, all of which are collected in real-time at a sampling frequency of 100 Hz. Simultaneously, the wear coefficient correction amount corresponding to this material batch is read from the storage unit.

[0056] The current real-time braking condition time-series parameters are compared with the source domain condition parameter set used during feature encoding mapping training, and the local offset for each time step is calculated. The source domain condition parameter set refers to the braking condition time-series parameters in all source domain data collected in the first step, arranged by time step index. Each time step position (e.g., the 1st time step, the 2nd time step, up to the maximum time step T) contains multiple samples of braking pressure, temperature, and speed values. For each time step in the current real-time sequence, the following sub-steps are performed to calculate the local offset for that time step.

[0057] The first step is to extract the three parameter values ​​of the current time step, namely the braking pressure value, temperature value and speed value, and arrange them in order to form a three-dimensional vector, which is called the real-time parameter vector.

[0058] The second step is to extract all reference parameter vectors that are at the same time position as the current time step from the source domain operating condition parameter set.

[0059] Specifically, if the index of the current time step is i, then there is a three-dimensional reference parameter vector at the i-th time step for each braking event sample in the source domain. Assuming there are M braking event samples in the source domain, M reference parameter vectors are extracted. For each reference parameter vector, the absolute value of the difference between it and the corresponding component of the real-time parameter vector is calculated. The absolute differences of the three components are summed to obtain an initial deviation value. The above calculation is repeated for all M reference parameter vectors to obtain M initial deviation values, which are then arranged in their original order to form an initial deviation sequence.

[0060] The third step is to sort the deviation values ​​in the original deviation sequence in ascending order. The deviation value at the 90th percentile after sorting is taken as the candidate offset for that time step. The specific calculation method for the 90th percentile is as follows: number the index of the sorted sequence starting from 1, calculate the position index equal to M multiplied by 0.9 and rounded up, and the deviation value at that position is the candidate offset.

[0061] The fourth step involves exponentially weighting the candidate offset with the local offset from the previous time step to obtain the local offset for the current time step. Let the local offset from the previous time step be a known value (the previous offset in the first time step is set to 0), and the current candidate offset be a new value. The fusion method is: the current local offset equals the previous local offset multiplied by 0.7 plus the current candidate offset multiplied by 0.3. These coefficients 0.7 and 0.3 are preset values, where 0.7 is called the attenuation factor to smooth the offset change.

[0062] Perform temporal accumulation on the local offset of each time step calculated above to obtain a confidence weight between 0 and 1, and the confidence weight decreases as the local offset increases.

[0063] The specific method for time-series accumulation is as follows: Let the accumulated value of the first time step be 1 minus the normalized local offset of the first time step. The normalization method is to divide the local offset by the maximum deviation of all samples across all time steps in the source domain (this maximum deviation is statistically obtained during the training phase and set to 100). That is, the accumulated value of the first time step equals 1 minus (the first local offset divided by 100), and this result is then constrained to between 0 and 1. For subsequent time steps, the current confidence weight equals the confidence weight of the previous time step multiplied by (1 minus the normalized local offset of the current time step). When the local offsets of multiple consecutive time steps are large, the confidence weight will rapidly decay to near 0; when the local offset is small, the confidence weight remains close to 1.

[0064] The wear coefficient correction is multiplied by the confidence weight to obtain the dynamic correction. The dynamic correction is a scalar value that dynamically adjusts according to the difference between the current real-time operating conditions and the source domain operating conditions: when the real-time operating conditions are highly consistent with the source domain operating conditions, the confidence weight is close to 1, and the dynamic correction is close to the fixed wear coefficient correction; when the real-time operating conditions differ greatly from the source domain operating conditions (e.g., extreme high temperature or high braking force occurs), the confidence weight decreases, and the dynamic correction decreases accordingly to avoid over-reliance on the fixed correction obtained from source domain training.

[0065] The base wear depth (i.e., the depth value output by feature encoding mapping and wear prediction mapping at the current time step) is multiplied by the dynamic correction amount, and the product is used as the final wear prediction result for the current time step. This final wear prediction result is output sequentially by time step to obtain the wear prediction sequence for the entire braking event or the entire service life, which can be used for real-time status updates and maintenance early warnings in the digital twin model.

[0066] It should be noted that the construction process of the digital twin model is as follows: Real-time collected braking condition time-series parameters and the current brake pad material batch identifier are used as inputs. These are sequentially processed through a pre-trained feature encoding mapping and wear prediction mapping to output a basic wear depth sequence. This sequence is then multiplied by a wear coefficient correction value retrieved based on the material batch identifier to obtain the final wear prediction result. Specifically, the feature encoding mapping is responsible for converting the three time-series parameters—brake pressure, friction interface temperature, and wheel linear velocity—into batch-independent wear features. The wear prediction mapping is responsible for outputting the basic wear depth after time segmentation, weighted aggregation, cyclic mapping with memory units, and residual superposition of these features. The wear coefficient correction value is determined by S4 and stored in a correction table corresponding to the material batch identifier. This digital twin model is deployed on an onboard computing unit or a cloud-based prediction platform, outputting wear prediction values ​​step-by-step at an update frequency of 100 Hz.

[0067] The working principle of this invention is as follows: Source domain data is collected, including braking condition time-series parameters, wear thickness sequences, and material batch identifiers. Feature encoding mapping is performed on the braking condition time-series parameters, and batch adversarial constraints are introduced. By minimizing the mutual information between wear features and batch identifiers, batch-independent wear features are obtained. These batch-independent wear features are input into the wear prediction mapping, outputting a basic wear depth sequence. The two mappings are then adjusted inversely based on their deviation from the wear thickness sequence until convergence. When a new batch of material is introduced, a small amount of measured data is collected, the feature encoding mapping and wear prediction mapping are frozen, and only the batch discrimination mapping is adjusted to distinguish between new and old batches. A wear coefficient correction is calculated. This correction is embedded into a digital twin model, and the current batch identifier and operating condition parameters are obtained in real time. The basic wear depth is obtained through feature encoding mapping and wear prediction mapping, and then multiplied by the dynamic correction obtained by comparing the operating conditions with the source domain distribution to output the final wear prediction result. This method achieves accurate wear prediction across batches and operating conditions.

[0068] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for constructing a prediction model for automotive brake pad wear based on digital twins, characterized in that, Includes the following steps: S1, the source domain data collected includes: multiple sets of braking condition timing parameters, wear thickness sequence corresponding to each set of braking condition timing parameters, and material batch identifier of the brake pad to which each set of data belongs; S2, perform feature encoding mapping on the timing parameters of braking conditions to obtain wear feature representation, and introduce batch adversarial constraints to obtain batch-independent wear features by minimizing the mutual information between the wear feature representation and the material batch identifier; S3, input batch-independent wear features into the wear prediction mapping, output the basic wear depth sequence, and adjust the feature encoding mapping and wear prediction mapping in reverse according to the deviation between the basic wear depth sequence and the wear thickness sequence until the deviation converges; S4. When a new batch of material is introduced, a small amount of measured wear data of the new batch of material is collected, the feature coding mapping and wear prediction mapping are frozen, only the batch discrimination mapping is adjusted to distinguish between the new and old batches, and the wear coefficient correction amount of the new batch of material relative to the source domain batch is calculated from the adjusted batch discrimination mapping. S5 embeds the wear coefficient correction amount into the digital twin model. During the real-time prediction process, it obtains the material batch identifier of the current brake pad and the real-time braking condition time sequence parameters. The basic wear depth is obtained by feature encoding mapping and wear prediction mapping in sequence, and then multiplied by the corresponding wear coefficient correction amount to output the final wear prediction result.

2. The method for constructing a digital twin-based automotive brake pad wear prediction model according to claim 1, characterized in that, The acquisition of batch-independent wear characteristics specifically includes: The braking condition timing parameters are input into a fully connected encoding mapping, and the initial wear characteristics are output. The initial wear characteristics are simultaneously input into the wear prediction branch and the batch discrimination branch. A gradient inversion layer is set before the batch discrimination branch to invert the gradient sign during backpropagation. Simultaneously, a mutual information estimation network is constructed, using the divergence of the product of the joint distribution of initial wear features and material batch identifiers and the marginal distribution as the loss term. Through adversarial training, the distribution of initial wear features under different batch identifiers is aligned, and finally batch-independent wear features are output.

3. The method for constructing a digital twin-based automotive brake pad wear prediction model according to claim 2, characterized in that, A gradient inversion layer is set before the batch discrimination branch to invert the gradient sign during backpropagation, specifically including: The initial wear characteristics are received and forwarded to the batch discrimination branch to calculate the batch discrimination loss; During backpropagation, the gradient of the batch discrimination loss with respect to the initial wear characteristics is multiplied by a negative preset coefficient to obtain the inverted gradient; The inverted gradient is backpropagated to the fully connected encoding map to update its parameters, while the parameters of the batch discrimination branch are updated in the normal gradient direction. The preset coefficient increases linearly from zero to one with the number of training rounds.

4. The method for constructing a vehicle brake pad wear prediction model based on digital twins according to claim 1, characterized in that, The wear prediction mapping specifically includes: Batch-independent wear features are split into multiple time segments according to time sequence, and the features within each segment are weighted and aggregated to obtain the segment wear contribution value; The wear contribution value of each segment is sequentially input into the loop mapping with memory unit. The memory unit maintains the long-term wear accumulation state and outputs the initial wear depth step by step. The residuals of the products of the initial wear depth and the braking force and temperature parameters at the corresponding time step are superimposed to generate the basic wear depth sequence. The absolute deviation between the basic wear depth sequence and the wear thickness sequence is used as the loss. Backpropagation updates the cyclic mapping, the weights of the weighted aggregation, and the coefficients of the residual superposition in sequence.

5. The method for constructing a vehicle brake pad wear prediction model based on digital twins according to claim 4, characterized in that, The output of the initial wear depth step-by-step specifically includes: Get the memory state output from the previous time step; The attenuation coefficient is calculated based on the braking force and temperature at the current time step, so that the attenuation coefficient is negatively correlated with the braking force and positively correlated with the temperature. The fragment wear contribution value of the current time step is weighted and fused with the memory state of the previous time step according to the decay coefficient to obtain the current memory state; Apply a nonlinear compression mapping to the current memory state and output the initial wear depth at the current time step.

6. The method for constructing a vehicle brake pad wear prediction model based on digital twins according to claim 1, characterized in that, The calculation of the wear coefficient correction amount of the new batch of material relative to the source domain batch by the adjusted batch discrimination mapping specifically includes: The braking condition time series parameters in a small amount of measured wear data of the new batch of materials are sequentially processed by frozen feature encoding mapping and wear prediction mapping to output the basic wear depth sequence of the new batch. By comparing the basic wear depth sequence of the new batch with the measured wear thickness sequence corresponding to the new batch of materials step by step, multiple wear deviation ratios are obtained. Weighted median filtering is applied to each wear deviation ratio, where the weights are positively correlated with the braking intensity of the corresponding time step. After removing ratios that deviate from the median by more than a preset threshold, the median of the remaining ratios is taken and used as the wear coefficient correction amount.

7. The method for constructing a vehicle brake pad wear prediction model based on digital twins according to claim 6, characterized in that, The output of the new batch of basic wear depth sequence specifically includes: The braking condition timing parameters in the new batch of measured wear data are resampled along the time axis to align with the sampling interval of the source domain data. The aligned time series parameters are segmented, and the weighted moving average of each parameter in each segment is calculated to obtain a smoothed operating condition sequence. The smooth working condition sequence is input into the frozen feature encoding mapping step by step, and the frozen wear features are output. The frozen wear features are input into the frozen wear prediction mapping, and the intermediate depth value at each time step is output. The intermediate depth values ​​at all times are spliced ​​together in chronological order to form a new batch of basic wear depth sequence.

8. The method for constructing a digital twin-based automotive brake pad wear prediction model according to claim 1, characterized in that, The final wear prediction result output specifically includes: The current real-time braking condition time series parameters are compared with the source domain condition parameter set used during feature encoding mapping training, and the local offset of each time step is calculated. A time-series accumulation is performed on the local offset to obtain a confidence weight between 0 and 1, which decreases as the offset increases; Multiply the wear coefficient correction by the confidence weight to obtain the dynamic correction. Multiply the base wear depth by the dynamic correction amount to output the final wear prediction result.

9. The method for constructing a vehicle brake pad wear prediction model based on digital twins according to claim 8, characterized in that, The calculation of the local offset at each time step specifically includes: Extract multiple parameter values ​​from the current time step in the current real-time braking condition timing parameters to form a real-time parameter vector; Extract multiple reference parameter vectors corresponding to the current time step position from the source domain operating condition parameter set, calculate the sum of the absolute differences between the components of each reference parameter vector and the real-time parameter vector, and obtain the original deviation sequence. Sort the deviation values ​​in the original deviation sequence from smallest to largest, and take the deviation value that is in the preset percentile after sorting as the candidate offset of the time step. The candidate offset is fused with the local offset of the previous time step by an exponential weighting to obtain the local offset of the current time step.

Images