A tire management method and system

Abstract

The application discloses a kind of tire management method and system, belong to tire management technical field, comprising: according to tire model, obtain tire configuration information, and distribute unique code for each equipment tire;Collect multi-source data, the multi-source data include tire pressure data, tire temperature data, equipment operation data, environmental road condition data and artificial inspection data;The multi-source data are cleaned, aligned and feature extraction, form time-aligned standardized feature vector;Construct multi-layer neural network model based on attention mechanism, the standardized feature vector is input into the model, dynamically fuses the feature information of each data source, outputs tire health score and abnormal early warning information;According to the actual use, maintenance and replacement of tire, complete tire full life cycle management;Through multi-source data fusion, the limitation of single data source is overcome, the evaluation of tire state is more comprehensive, diagnosis is more accurate, effectively reduces false alarm and leakage rate.

Description

Technical Field

[0001] This invention relates to the field of tire management technology, specifically a tire management method and system. Background Technology

[0002] In existing mines, data on tire condition (such as tire pressure and temperature), vehicle operation (such as load and speed), and mine environmental information (such as road slope) are typically managed by different systems or departments, lacking effective data integration and forming information silos. Simple tire pressure monitoring systems (TPMS) or IoT sensors achieve real-time monitoring of single or a few parameters such as tire pressure and temperature, but lack in-depth data analysis capabilities. They rarely integrate tire data with vehicle operation data and environmental data for in-depth analysis, making it impossible to comprehensively assess tire health and predict remaining life. Managers also find it difficult to analyze the true causes of abnormal tire wear or damage from a global perspective, making it impossible to achieve accurate life prediction and cost control.

[0003] Secondly, tire management for mining mobile equipment relies heavily on visual inspection and manual tapping by maintenance personnel to assess tire condition. This method is highly subjective and makes it difficult to detect early internal tire damage (such as delamination or slow leaks). Tires may suddenly fail due to external damage or abnormal tire pressure, leading to serious downtime or even safety accidents. The inability to provide timely and accurate warnings of these risks, coupled with a significant delay between detecting abnormalities and taking maintenance measures, makes preventative measures impossible. Summary of the Invention

[0004] On the one hand, in order to solve the problems of the prior art, the present invention provides a tire management method, including the following steps: Obtain tire configuration information based on tire model and assign a unique code to each device's tires; Collect multi-source data, including tire pressure data, tire temperature data, equipment operation data, environmental road condition data, and manual inspection data; The multi-source data is cleaned, aligned, and its features are extracted to form a time-aligned standardized feature vector. A multi-layer neural network model based on an attention mechanism is constructed. The standardized feature vector is input into the model, and feature information from various data sources is dynamically fused to output tire health scores and abnormal warning information. Based on the actual use, maintenance and replacement of tires, the entire life cycle management of tires is completed.

[0005] Furthermore, the tire pressure data and the tire temperature data are collected by tire pressure and temperature monitoring sensors installed inside the tire; The equipment's operating data is acquired through a CAN bus data acquisition unit, including speed, load, accelerator pedal percentage, brake pedal percentage, operating time, and cumulative mileage. The environmental road condition data is obtained by combining high-precision positioning terminal with mine map data, including road surface type and slope information; The manual inspection data is entered via a mobile app, including the wear and damage of the tire tread pattern.

[0006] Furthermore, the data cleaning includes removing outliers that exceed a set threshold, and the data alignment is achieved by allocating a uniform high-precision timestamp to ensure that the data time deviation within the same analysis time window does not exceed the set threshold.

[0007] Furthermore, physical constraints are introduced during the training of the attention-based multilayer neural network model to ensure that the prediction results conform to the basic physical laws of tire operation.

[0008] Furthermore, the anomaly warning information includes: multi-level warnings and anomaly type diagnosis; The abnormality types include: uneven wear, crown burst, lateral perforation, and delamination; The multi-level early warning system includes: attention warning, warning warning, and emergency warning.

[0009] Furthermore, the tire lifecycle management includes recording the tire's cumulative operating hours, mileage, maintenance status, replacement reasons, and wear level.

[0010] Furthermore, the steps for constructing a multi-layer neural network model based on the attention mechanism include: Four independent encoder networks are designed, each corresponding to one of four types of data sources: tire pressure and temperature data, CAN bus operation data, road surface type and slope data, and manual inspection data. Each encoder has the same structure and function. The preprocessed features of each data source with different dimensions and distributions are uniformly mapped to the same latent space. An attention fusion mechanism is constructed to achieve intelligent fusion of multi-source features through a multi-head attention mechanism. The steps are as follows: Step 1: Feature Matrix Concatenation. The output features from the four encoders are integrated to form a unified feature matrix. Step 2: Transform the feature matrix into a query matrix, key matrix, and value matrix respectively using the projection matrix, and compress the feature dimension to simplify the calculation; Step 3: Calculate the similarity between the query matrix and the key matrix to obtain the attention score. Introduce a scaling factor to avoid the gradient vanishing problem caused by excessively large calculation results. Then, convert the score into the initial weight distribution through a normalization function. Step 4: Use 8 attention heads to work simultaneously, each attention head learns different feature interaction patterns, and then concatenate the outputs of all attention heads; Step 5: Perform a linear transformation on the concatenated features to restore them to the target latent space dimension; extract the initial weights of each data source from the attention weight matrix, and normalize them to ensure that the sum of the weights is 1. Step 6: Based on the normalized weights, perform a weighted summation of the features from each data source to obtain the final multi-source fusion features; Design a multi-task output layer to output the model results; Furthermore, the output of tire health score and abnormal warning information includes: the output of tire health score and abnormal warning information is realized through a multi-task output layer, wherein the multi-task output layer includes a dual-branch structure; The dual-branch structure includes: a health regression branch and an anomaly classification branch; The health regression branch includes processing the multi-source fusion features through two hidden layers containing ReLU activation functions, and then compressing the output to the 0-1 range through the sigmoid function to obtain the tire health score (0 represents the tire is in critical condition, and 1 represents the tire is in healthy condition). The anomaly classification branch first processes the multi-source fusion features through a hidden layer containing a ReLU activation function, and then outputs the probability distribution of four types of anomalies—abrasion, crown burst, lateral perforation, and delamination—through a softmax function. The category with the highest probability is the anomaly type diagnosed by the model.

[0011] Furthermore, the introduction of physical constraints includes: in the data preprocessing stage, defining the reasonable range of data through physical laws, filtering outliers that violate physical common sense from the data source, and introducing basic physical constraints for the model; In the process of feature extraction and interaction, the physical laws related to tire operation are transformed into feature dimensions, so that the features learned by the model have physical constraint attributes. In the training process of a multi-layer neural network model based on the attention mechanism, physical laws are embedded as constraints into the training process to directly regulate the model's prediction logic.

[0012] On the other hand, this application provides a tire management system, including: The data acquisition module is used to collect data from multiple sources. A data preprocessing and storage module is used to process the multi-source data and store the processed data. The multi-source data fusion and analysis module has a built-in multi-layer neural network model based on the attention mechanism, which is used to output tire health scores and abnormal warning information. The early warning and decision support module is used to generate early warning notifications and maintenance suggestions; A visual human-computer interaction interface is used to display tire management-related data and information.

[0013] The beneficial effects of this invention are: By integrating multi-source data, the limitations of a single data source are overcome, resulting in a more comprehensive assessment and more accurate diagnosis of tire condition, effectively reducing false alarms and missed alarms.

[0014] The attention mechanism model adopted can dynamically adjust the weights of each data source, adapt to different devices and different working conditions (such as heavy load climbing and high speed driving), and has a high degree of intelligence.

[0015] It has enabled a shift from "reactive maintenance" to "predictive maintenance," allowing for the early detection of potential faults, the planning of maintenance windows, and the effective avoidance of serious accidents, thereby reducing unplanned downtime.

[0016] By scoring tire health and issuing early warnings for abnormalities, tire life can be effectively extended, maintenance costs optimized, and equipment utilization improved, bringing significant economic benefits to mining companies and realizing refined, full life-cycle management of tire assets. Attached Figure Description

[0017] Figure 1 A flowchart illustrating the multi-source data fusion method and system for managing tires of mining equipment provided by 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] It should be noted that the inventive points of this application are to solve the problem of information silos, realize predictive maintenance, and improve the intelligence and refinement of management.

[0020] Among these measures, the problem of information silos is solved by integrating heterogeneous data from multiple sources, including tires themselves, equipment operation, environmental road conditions, and manual inspections, breaking down data barriers and enabling comprehensive perception of tire status.

[0021] Predictive maintenance is achieved by using advanced fusion models to shift from passively responding to faults to proactively predicting risks, accurately assessing tire health and predicting remaining life, thereby avoiding sudden failures and reducing safety risks.

[0022] Enhance the intelligence and precision of management by using intelligent early warning and decision support to optimize tire maintenance plans, reduce total life cycle costs, and improve the uptime and operational efficiency of mining equipment.

[0023] Example 1 Please see Figure 1This invention provides a tire management method, comprising the following steps: Obtain tire configuration information based on tire model and assign a unique code to each device's tires; The system automatically associates the tire knowledge base with the equipment model, including information such as tire distribution and reasonable thresholds for tire data. Each tire is assigned a unique code upon entering the database, which is used for subsequent management records such as tire warnings, replacements, and scrapping.

[0024] Collect multi-source data, including tire pressure data, tire temperature data, equipment operation data, environmental road condition data, and manual inspection data; The tire pressure data and tire temperature data are collected by a tire pressure and temperature monitoring sensor installed inside the tire; a high-temperature resistant and impact-resistant sensor is used, which is directly installed inside the tire valve or on the rim, and collects tire pressure (kPa) and tire temperature (°C) data at a frequency of at least 1Hz.

[0025] The equipment's operating data is acquired through a CAN bus data acquisition unit, including speed, load, accelerator pedal percentage, brake pedal percentage, operating time, and cumulative mileage. The CAN bus data acquisition unit connects through the device's OBD interface and parses and acquires engine data via the J1939 protocol, including but not limited to: vehicle speed (km / h), calculated load (kg), accelerator pedal percentage (%), brake pedal percentage (%), engine operating time (h), and cumulative mileage (km).

[0026] The environmental road condition data is obtained by combining high-precision positioning terminal with mine map data, including road surface type and slope information; The device's latitude, longitude, altitude, speed, and heading angle are obtained in real time through a single Beidou high-precision positioning terminal, with positioning accuracy reaching the centimeter level. An industrial-grade 4G / 5G router is used to package, encrypt, and transmit the aforementioned sensor data to the cloud platform. The manual inspection data is entered via a mobile terminal APP, including the wear and damage of the tire tread pattern; Maintenance and management personnel can access the system via a smartphone, tablet, or PC web interface with a dedicated app installed. Maintenance and management personnel can access the system via a smartphone, tablet, or PC web interface with a dedicated app installed. The multi-source data is cleaned, aligned, and its features are extracted to form a time-aligned standardized feature vector. The data cleaning includes removing outliers that exceed a set threshold, and the data alignment is achieved by allocating a uniform high-precision timestamp to ensure that the data time deviation within the same analysis time window does not exceed a set threshold. Data preprocessing includes: The first layer of cleaning employs hard filtering based on physical rules, directly removing outliers that significantly exceed set thresholds. For example, the reasonable range for tire pressure is 50-900 kPa, determined based on the technical specifications and actual operating conditions of mining equipment tires. The second layer of cleaning uses an improved Z-score method based on a statistical model. The core advantage of this method is its insensitivity to outliers, making it particularly suitable for processing industrial data that may contain multiple outliers. The specific calculation process is as follows: first, the median of the data sequence is calculated as a location estimate; then, the absolute deviation of the median is calculated as a measure of dispersion. For each data point, its improved Z-score value is calculated; when the absolute value exceeds 3.5, it is identified as a statistical outlier. For identified outliers and missing values, we use a linear interpolation method based on time context for imputation. This method considers the temporal characteristics of the data, estimating the true location of outliers through the linear relationship between adjacent valid data points, thus maintaining the continuity of the data sequence while minimizing the impact of outliers on subsequent analysis.

[0027] Data time alignment includes: The primary prerequisite for multi-source data fusion is establishing a unified time reference system. Since the clocks of different data acquisition systems may have slight deviations and their sampling frequencies differ, we must establish a precise time alignment scheme. We choose the 1Hz sampling frequency of the tire pressure monitoring system as the benchmark because it is the highest and most stable sampling frequency among all data sources. Timestamps from other data sources are aligned to this unified time axis using nearest-neighbor interpolation. Specifically, for each target time point, we select the data point with the smallest time difference in the original time series as the value for that moment. For discrete data such as manual inspections, we employ a persistence method. Inspection results remain valid for subsequent time periods until the next inspection data update. While this method sacrifices some time precision, it aligns with the actual characteristics of inspection data, namely, that inspection results are considered valid until the next inspection.

[0028] Feature extraction includes: The first step is time-domain feature extraction, which reflects the short-term variation patterns and statistical characteristics of the data. We use a sliding window method to calculate a series of statistical features, with the window size typically set to 60 seconds. This duration captures meaningful changes in operating conditions without introducing excessive computational delay.

[0029] Multi-source interaction feature construction: The temperature-pressure coupling feature, based on the ideal gas law, calculates the deviation between actual pressure and theoretically expected pressure. This feature can effectively identify slow tire leaks or sensor malfunctions, as pressure changes in such situations deviate from the normal temperature-pressure relationship.

[0030] The load-speed product characteristic reflects the instantaneous power load of the tire and is the main driving factor of tire wear. By monitoring the changing trend of this characteristic, the remaining service life of the tire can be predicted. The dynamic load factor describes the load intensity per unit speed and is of great value in identifying overload conditions.

[0031] The road-load interaction characteristics combine environmental information and operating status, identifying heavy-load conditions under harsh road surface conditions through an indicator function. These conditions cause the greatest damage to tires and require special attention. The cumulative stress characteristics calculate the cumulative load borne by the tire throughout its lifespan through integration, providing important data for predictive maintenance.

[0032] Among them, multi-source interactive features (temperature-pressure coupling, load-velocity product, etc.) serve as a bridge connecting "raw data acquisition" and "model prediction (health score + anomaly warning)." Their role extends throughout the entire subsequent model processing, from feature input, encoder mapping, attention fusion to the final prediction output. However, they are not individually labeled but rather integrated into each stage as "preprocessed features," becoming a key support for accurate model prediction. The specific related logic can be broken down into the following four levels: I. Multi-source interaction features are the core component of "standardized feature vectors" and provide "meaningful input" to the model. The document clearly states that after "cleaning, aligning and feature extraction", multi-source data forms "time-aligned standardized feature vectors", and multi-source interaction features are the core output of the "feature extraction" stage (different from basic statistical features from a single data source, such as average tire pressure and maximum speed).

[0033] Raw data (such as isolated tire pressure, tire temperature, and load values) can only reflect a single state and cannot show the correlation between data (such as "whether the increase in tire temperature is accompanied by a normal increase in tire pressure" or "whether the heavy load occurs on a rough road surface"). Multi-source interactive features transform scattered raw data into "effective information that directly reflects tire health" through physical laws (such as the ideal gas law) and operating condition logic (such as "heavy load + harsh road surface aggravates wear"). For example, the temperature-pressure coupling feature calculates deviations based on the ideal gas law (pressure is proportional to temperature when volume is constant), transforming "isolated tire pressure / temperature data" into a direct basis for "determining whether there is a slow leak or whether the sensor is malfunctioning." Road surface-load interaction characteristics: By using indicator functions (such as "road surface type = gravel road and load > rated value, then marked as 1"), "environmental data + operational data" are transformed into key signals for "identifying high-damage operating conditions".

[0034] These interactive features, together with basic statistical features (such as tire pressure standard deviation and cumulative mileage), constitute a "standardized feature vector," which is the direct input to the model (a multi-layer neural network based on an attention mechanism). Without these interactive features, the model can only process single-dimensional data and cannot capture the "complex influencing factors" of tire condition, resulting in a significant decrease in prediction accuracy.

[0035] Second, multi-source interactive features are integrated into the latent space through "encoder mapping" and become "fusionable feature units". The core role of the "4 independent encoders" is to "map preprocessed features of different dimensions and distributions to the same latent space", and multi-source interactive features are the core part of the "preprocessed features" and directly determine the "feature vector quality" of the encoder output: For the "tire pressure and temperature data source", the encoder processes not only isolated tire pressure / temperature values, but also "temperature-pressure coupling features" (such as the deviation of tire pressure from the theoretical value). After the feature is transformed into a latent space vector, the model can quickly identify the abnormal pattern of "tire temperature rises but tire pressure does not rise proportionally" (corresponding to slow air leakage). For the "CAN bus operation data source", the encoder processes "load-speed product characteristics" and "cumulative stress characteristics". After these characteristics are mapped into latent space vectors, the "instantaneous load" (affecting short-term wear) and "total life cycle load" (affecting remaining life) borne by the tire can be quantified, providing a key basis for the "load dimension" for health scoring; For the "environmental road condition data source", the encoder processes "road surface-load interaction features". After the feature is mapped, the model can associate high-risk working conditions of "severe road surface + heavy load", providing the basis for judging the "working condition cause" for abnormal warnings such as "uneven wear and crown burst".

[0036] Without these interactive features, the encoder can only map single data (such as only mapping "load value" instead of "load + speed product"), and the latent space vector will lose key information such as "load intensity and physical correlation", causing subsequent fusion and prediction to deviate from the actual working conditions.

[0037] Third, multi-source interaction features influence "attention weight allocation," causing the model to prioritize "highly correlated features." The core of the "attention fusion mechanism" is "dynamically adjusting the weights of each data source to adapt to different operating conditions." The "physical correlation and operating condition risk" information carried by multi-source interaction features is the key basis for the model to judge the "importance of data sources." When the tire is in a "high temperature environment", the "tire pressure and tire temperature data source" where the "temperature-pressure coupling feature" is located will be given higher attention weight - because at this time the "deviation between tire pressure and tire temperature" is the core of determining whether there is a slow leak, and the model will prioritize the information output by this interaction feature. When the tire is in a "gravel road + heavy load" condition, the weight of the "environmental road condition data source" and "CAN bus data source" where the "road surface-load interaction feature" is located will be increased - because this interaction feature identifies a "high damage condition", and the model needs to prioritize the integration of the interaction information of these two data sources to determine whether there is a risk of uneven wear or side penetration. When assessing "remaining tire life", the weight of the "CAN bus data source" where the "cumulative stress feature" is located will be increased - because this interactive feature directly reflects the load accumulation throughout the tire's life cycle and is the core of predicting remaining life. The model will rely heavily on its information to calculate the health score.

[0038] Overall, multi-source interactive features guide the attention mechanism to achieve "intelligent weight allocation" by "carrying working condition priority information," avoiding the model from "treating" all data sources "equally," thereby improving the specificity of predictions (such as not wasting computing power on irrelevant inspection data during heavy loads).

[0039] Fourth, multi-source interaction features provide "physical and logical support" for "health score + anomaly warning," ensuring that the predictions match the actual model output of "health score (0-1)" and "anomaly warning (type + level)." Essentially, this quantifies and classifies the "risk level reflected by multi-source interaction features"—these interaction features directly determine the rationality of the prediction results. Basis for calculating health score: The larger the "load-velocity product characteristic" (higher instantaneous load) and the larger the "cumulative stress characteristic" (higher life cycle load), the faster the health score drops (e.g., during heavy-load uphill climbing, the health score may drop from 0.8 to 0.5). The greater the deviation of the "temperature-pressure coupling characteristic" (e.g., the tire temperature rises by 10°C but the tire pressure only rises by 5 kPa, which is far below the theoretical value), the lower the health score (e.g., when the deviation exceeds the threshold, the health score drops below 0.3, triggering an emergency warning).

[0040] Criteria for judging abnormal warnings: If the "temperature-pressure coupling characteristic" continues to deviate (and sensor failure is ruled out), the model will diagnose it as a "slow leak" anomaly, and trigger a "warning alert" based on the health score (e.g., 0.4). If the "Road-Load Interaction Features" frequently marks "Gravel Road + Heavy Load" and the "Load-Velocity Product Features" exceed the threshold for a long time, the model will diagnose it as "Uneven Wear" (due to heavy load + poor road surface aggravating uneven tread wear) and trigger the corresponding warning. If the "cumulative stress characteristic" reaches 80% of the tire's rated life, even if the current single data (such as tire pressure) is normal, the health score will drop below 0.6, triggering a "caution warning" (indicating insufficient remaining life).

[0041] Multi-source data fusion model and early warning: Construct a multi-layer neural network model based on an attention mechanism to dynamically weight and fuse features from different data sources.

[0042] The model can automatically learn the importance of each data source to the tire health status under different working conditions and output a comprehensive tire health score: a quantitative health status index between 0 and 1.

[0043] Physical constraints are introduced during the training of the attention-based multilayer neural network model to ensure that the prediction results conform to the basic physical laws of tire operation. The abnormal warning information includes: multi-level warnings and abnormality type diagnosis; The abnormality types include: uneven wear, crown burst, lateral perforation, and delamination; The multi-level early warning system includes: attention warning, warning warning, and emergency warning; The specific implementation process of the multi-source data fusion model and early warning includes: 1. The multi-source data fusion model is designed based on a combination of deep neural networks and attention mechanisms, aiming to overcome the limitations of traditional single-source data analysis. The core idea of ​​the model is to first extract features independently from each data source through a hierarchical processing mechanism, and then dynamically weight and fuse them through an attention mechanism to ultimately achieve accurate health status assessment. This architecture can adapt to the changing importance of each data source under different operating conditions. For example, it focuses more on tire pressure and temperature data in high-temperature environments, and more on road surface and load data in harsh road conditions.

[0044] 2. The feature encoding layer employs four independent encoder networks, each processing feature data from different sources. Each encoder consists of two fully connected layers, using the ReLU activation function to introduce nonlinear transformation capabilities. The core function of the encoder is to map features of different dimensions and distributions to a unified latent space, laying the foundation for subsequent fusion processing. (1) Tire pressure and temperature encoder: ; (2) CAN data encoder: ; (3) Road surface type data encoder: ; (4) Inspection data encoder: ; In the formula, This represents tire pressure and temperature data. This indicates CAN bus operation data. This indicates the road surface type and slope data. This represents data from manual inspection. (⋅) represents the activation function, which is used to introduce nonlinear transformations to enhance the model's ability to fit complex features; This represents the weight matrix of the corresponding data source encoder, used to perform linear transformation on the input features; This represents the preprocessed original feature vector of the corresponding data source (standardized data after cleaning, alignment, and feature extraction). This represents the bias term of the corresponding data source encoder, used to adjust the offset of the linear transformation.

[0045] 3. Attention Fusion Mechanism: The importance weights of each data source are dynamically learned through a multi-head attention mechanism. First, the output features of the four encoders are concatenated to form a feature matrix. Each row represents a feature vector of a data source.

[0046] (1) Attention calculation process: ; Where the projection matrix This compresses the original feature dimensions to 8 dimensions, making it easier to calculate the attention distribution.

[0047] Represents the characteristic matrix; This represents the query matrix, used to calculate the similarity with the key matrix; The key matrix is ​​used to calculate attention weights by matching it with the query matrix. The value matrix is ​​used to obtain the fused features by weighted summation based on attention weights. (2) Calculation of attention for scaling dot product: ; in It is the dimension of the key vector, the scaling factor. This is used to prevent the gradient of the softmax function from vanishing due to an excessively large dot product.

[0048] This represents the transpose of the query matrix and the key matrix, used to calculate the similarity (attention score) between features from different data sources. This represents the normalization function, which converts the attention score into a weight distribution (the sum of the weights is 1). (3) Multi-head attention mechanism: The model employs eight attention heads for parallel computation, with each head learning a different feature interaction pattern: ; Indicates the first The output of each of the eight attention heads (a total of eight parallel attention heads), each head learns different feature interaction patterns; Indicates the first A dedicated projection matrix for each attention head; The outputs of each head are concatenated and then linearly transformed to obtain the final output: ; (⋅) represents the concatenation function, which concatenates the outputs of the 8 attention heads into an R4×64 matrix; This represents the linear transformation weight matrix, used to map the concatenated 64-dimensional features to the final fused features, with dimension R. 64×32 (Restored to 32-dimensional hidden space); (4) Dynamic weight allocation strategy: From the attention weight matrix The importance weights of each data source are extracted. The initial weights of each data source are obtained by averaging along the column directions of the attention matrix. ; Indicates the first Initial weights for each data source; This represents the attention score between the i-th data source and the j-th data source; Normalization is performed using the softmax function to ensure that the sum of the weights is 1. ; This represents the final fused feature vector of multi-source data; Indicates the first The normalized weights of each data source (processed by softmax) are used to dynamically allocate the importance of each data source. This represents the feature vector obtained after the i-th type of data source is processed by an independent encoder network. The final fused features are obtained by weighted summation: ; This dynamic weighting mechanism allows the model to automatically adjust the contribution of each data source based on the current operating conditions. For example, when detecting abnormal tire pressure, the weight of the TPMS data source will automatically increase; when analyzing wear patterns, the weights of inspection data and CAN data will increase accordingly.

[0049] (5) Multi-task output layer design: The output layer employs a dual-branch structure, handling health regression and anomaly classification tasks separately. This design allows the two tasks to share the feature extraction layer while simultaneously learning task-specific features.

[0050] Health regression branch: ; ; ; The sigmoid function The output is compressed to the [0,1] range to represent the tire's health score.

[0051] The outputs of the two hidden layers in the health regression branch are both 32-dimensional. , This represents the weight matrix of the hidden layer in the health branch; , This represents the bias term of the hidden layer in the health branch; This indicates the tire health rating, ranging from [0,1] (0 = critically ill, 1 = healthy). This represents the sigmoid activation function, used to compress the output to the [0,1] interval; This represents the weight vector of the output layer of the health branch; This represents the bias term in the output layer of the health branch; Anomaly classification branches: ; ; The softmax function ensures that the output is a probability distribution of four types of abnormalities (uneven wear, crown burst, lateral perforation, and delamination), and the sum of the probabilities of each type is 1.

[0052] The hidden layer output represents the anomaly classification branch, with a dimension of 32. , The weight matrix representing the anomaly classification branch ( Hidden layer (output layer) , Bias terms representing the anomaly classification branches; This represents the probability distribution of abnormality types (model output), corresponding to 4 types of abnormalities (unilateral wear, crown burst, lateral perforation, and delamination), with the sum of the probabilities of each type being 1; softmax This represents the normalization function, used to convert the hidden layer output into a probability distribution; 4. Early Warning Mechanism: The health score is a comprehensive quantitative indicator, ranging from 0 to 1, calculated using a multi-source data fusion model. We have set three key thresholds to implement tiered early warning systems: (1) Level 1: Be alert (health level 0.6-0.8) When the tire health score drops below 0.8 but above 0.6, the system triggers a "Caution" level warning. This indicates that the tire is beginning to show signs of early degradation, but has not yet reached the point where immediate repair is required. The system will generate a message: "Tire health has dropped to [specific value], it is recommended to increase daily monitoring and pay attention to tire wear." At this stage, the system will increase the data collection frequency from once per minute to once every 30 seconds to more closely track changes in tire condition.

[0053] (2) Level 2: Warning (health level 0.3-0.6) When the tire health score drops further to below 0.6 but above 0.3, the system upgrades to a "Warning" level alert. This level indicates that the tire has significant abnormal wear or potential failure risk, requiring planned maintenance. The alert message will clearly state: "Tire health has dropped to [specific value], posing a risk of [specific abnormality type]. Professional inspection is recommended within [recommended time frame]." The system will automatically generate a work order, recommend a suitable maintenance time window, and begin tracking the execution of the maintenance schedule.

[0054] (3) Level 3: Emergency Warning (Health level below 0.3) This is the highest level of warning, triggered when the tire health score drops below 0.3. This indicates a serious safety hazard with the tire and a potential for malfunction at any time. The system will immediately issue an emergency alert: Tire health is critical, with [specific hazard type] risk; please stop immediately and check. Simultaneously, the system will send the warning information through multiple channels (in-vehicle display, mobile app push notifications, SMS notifications, etc.) to ensure relevant personnel are promptly informed and can take appropriate action.

[0055] Complete the full life cycle management of tires based on their actual use, maintenance, and replacement.

[0056] The tire lifecycle management includes recording the tire's cumulative operating hours, mileage, maintenance status, reasons for replacement, and degree of wear.

[0057] After each warning response, maintenance personnel are required to record detailed processing information via a mobile app, including: the actual tire condition detected, the specific measures taken, the replaced parts, and the repair time. This data, together with the warning information, forms a complete "warning-response" closed-loop record, providing a data foundation for subsequent analysis and optimization. The system automatically calculates the average response time from the occurrence of the warning to the completion of maintenance, and uses this as a key indicator for evaluating maintenance efficiency.

[0058] Take a Lingong RTH136 mining truck as an example: 1. Data upload: When the vehicle is in operation, all sensor data is uploaded to the cloud in real time.

[0059] 2. Real-time analysis and early warning: The cloud model performs inference once per second. When the system detects that the temperature of a certain tire is rising continuously and the tire pressure is fluctuating abnormally when the vehicle is going uphill under heavy load, it judges that the vehicle is overloaded based on CAN data. The attention mechanism gives higher weight to TPMS and CAN data. The health score continues to drop to 0.55, triggering a "warning" level warning and indicating "risk of uneven wear and overheating".

[0060] 3. Decision-making and execution: The maintenance personnel receive a push notification on their app, and the system automatically generates a repair work order. Based on the recommendations, the maintenance team immediately conducts a tire inspection after the shift ends, discovering and correcting the alignment parameter issue that was causing uneven tire wear, thus preventing further tire damage.

[0061] 4. Closed-loop and optimization: The results of this early warning and maintenance are recorded and fed back to the system for future model iteration and optimization.

[0062] Example 2 This application provides a tire management system, including: The data acquisition module collects data from multiple sources, including tire sensors, CAN bus data collectors, positioning terminals, and mobile apps. The data preprocessing and storage module cleans, aligns, and extracts features from the multi-source data collected by the data acquisition module, and then stores it in the database. The multi-source data fusion and analysis module has a built-in multi-layer neural network model based on the attention mechanism, which is used to calculate the health of the tire, diagnose abnormalities and predict its lifespan, and output the tire health score and abnormal warning information. The early warning and decision support module generates early warning information, maintenance work orders, and optimization suggestions based on the analysis results of the multi-source data fusion analysis module. A visual human-computer interaction interface displays tire status, warning information, and system analysis results to administrators in the form of charts, dashboards, etc.

[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A tire management method, characterized in that, Includes the following steps: Obtain tire configuration information based on tire model and assign a unique code to each device's tires; Collect multi-source data, including tire pressure data, tire temperature data, equipment operation data, environmental road condition data, and manual inspection data; The multi-source data is cleaned, aligned, and its features are extracted to form a time-aligned standardized feature vector. A multi-layer neural network model based on an attention mechanism is constructed. The standardized feature vector is input into the model, and feature information from various data sources is dynamically fused to output tire health scores and abnormal warning information. Based on the actual use, maintenance and replacement of tires, the entire life cycle management of tires is completed.

2. The tire management method according to claim 1, characterized in that, The tire pressure data and the tire temperature data are collected by tire pressure and temperature monitoring sensors installed inside the tire. The equipment's operating data is acquired through a CAN bus data acquisition unit, including speed, load, accelerator pedal percentage, brake pedal percentage, operating time, and cumulative mileage. The environmental road condition data is obtained by combining high-precision positioning terminal with mine map data, including road surface type and slope information; The manual inspection data is entered via a mobile app, including the wear and damage of the tire tread pattern.

3. The tire management method according to claim 1, characterized in that, The data cleaning includes removing outliers that exceed a set threshold, and the data alignment is achieved by allocating a uniform high-precision timestamp to ensure that the time deviation of the data within the same analysis time window does not exceed the set threshold.

4. The tire management method according to claim 1, characterized in that, Physical constraints are introduced during the training of the attention-based multilayer neural network model to ensure that the prediction results conform to the basic physical laws of tire operation.

5. The tire management method according to claim 1, characterized in that, The abnormal warning information includes: multi-level warnings and abnormality type diagnosis; The abnormality types include: uneven wear, crown burst, lateral perforation, and delamination; The multi-level early warning system includes: attention warning, warning warning, and emergency warning.

6. The tire management method according to claim 1, characterized in that, The tire lifecycle management includes recording the tire's cumulative operating hours, mileage, maintenance status, reasons for replacement, and degree of wear.

7. The tire management method according to claim 1, characterized in that, The steps to construct a multi-layer neural network model based on the attention mechanism include: Four independent encoder networks are designed, each corresponding to one of four types of data sources: tire pressure and temperature data, CAN bus operation data, road surface type and slope data, and manual inspection data. Each encoder has the same structure and function. The preprocessed features of each data source with different dimensions and distributions are uniformly mapped to the same latent space. An attention fusion mechanism is constructed to achieve intelligent fusion of multi-source features through a multi-head attention mechanism. The steps are as follows: Step 1: Feature Matrix Concatenation. The output features from the four encoders are integrated to form a unified feature matrix. Step 2: Transform the feature matrix into a query matrix, key matrix, and value matrix respectively using the projection matrix, and compress the feature dimension to simplify the calculation; Step 3: Calculate the similarity between the query matrix and the key matrix to obtain the attention score. Introduce a scaling factor to avoid the gradient vanishing problem caused by excessively large calculation results. Then, convert the score into the initial weight distribution through a normalization function. Step 4: Use 8 attention heads to work simultaneously, each attention head learns different feature interaction patterns, and then concatenate the outputs of all attention heads; Step 5: Perform a linear transformation on the concatenated features to restore them to the target latent space dimension; extract the initial weights of each data source from the attention weight matrix, and normalize them to ensure that the sum of the weights is 1. Step 6: Based on the normalized weights, perform a weighted summation of the features from each data source to obtain the final multi-source fusion features; Design a multi-task output layer to output the model results.

8. The tire management method according to claim 7, characterized in that, The output of tire health score and abnormality warning information includes: the output of tire health score and abnormality warning information is realized through a multi-task output layer, wherein the multi-task output layer includes a dual-branch structure; The dual-branch structure includes: a health regression branch and an anomaly classification branch; The health regression branch includes processing the multi-source fusion features through two hidden layers containing ReLU activation functions, and then compressing the output to the 0-1 range through the sigmoid function to obtain the tire health score (0 represents the tire is in critical condition, and 1 represents the tire is in healthy condition). The anomaly classification branch first processes the multi-source fusion features through a hidden layer containing a ReLU activation function, and then outputs the probability distribution of four types of anomalies—abrasion, crown burst, lateral perforation, and delamination—through a softmax function. The category with the highest probability is the anomaly type diagnosed by the model.

9. The tire management method according to claim 4, characterized in that, Introducing physical constraints includes: in the data preprocessing stage, defining the reasonable range of data through physical laws, filtering outliers that violate physical common sense from the data source, and introducing basic physical constraints into the model; In the process of feature extraction and interaction, the physical laws related to tire operation are transformed into feature dimensions, so that the features learned by the model have physical constraint attributes. In the training process of a multi-layer neural network model based on the attention mechanism, physical laws are embedded as constraints into the training process to directly regulate the model's prediction logic.

10. A tire management method system, applied in the method according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect data from multiple sources. A data preprocessing and storage module is used to process the multi-source data and store the processed data. The multi-source data fusion and analysis module has a built-in multi-layer neural network model based on the attention mechanism, which is used to output tire health scores and abnormal warning information. The early warning and decision support module is used to generate early warning notifications and maintenance suggestions; A visual human-computer interaction interface is used to display tire management-related data and information.

Images