A real-time monitoring method for electric shovel brake status

Abstract

This application relates to the field of data processing technology and discloses a method for real-time monitoring of the brake status of an electric shovel. The method comprises: synchronously collecting the hydraulic cylinder pressure pulsation, brake disc surface temperature gradient, brake pad thickness change rate, and motor feedback torque of the electric shovel brake through multi-source sensors to obtain a multi-physics field coupling feature vector; performing heavy-duty enhanced sampling of the multi-physics field coupling feature vector based on the brake power spectrum density to obtain a heavy-duty condition balance dataset; extracting features from the heavy-duty condition balance dataset using a heavy-duty condition adaptive multi-head attention mechanism to obtain global state features; performing multi-scale temporal convolution processing on the heavy-duty condition balance dataset based on the braking cycle characteristics to obtain temporal evolution features; and fusing and identifying the global state features and temporal evolution features based on brake power weighting to obtain the real-time status results of the electric shovel brake. This application improves the accuracy and real-time performance of electric shovel brake status monitoring.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method for real-time monitoring of the brake status of an electric shovel. Background Art

[0002] Existing electric shovel brake condition monitoring technologies primarily rely on a single sensor or a fixed acquisition frequency. These sensors monitor hydraulic system pressure changes, or temperature sensors monitor brake disc temperature changes to determine brake status. Traditional monitoring methods typically use simple threshold judgments or statistical feature-based analysis methods for condition identification. Some advanced solutions employ machine learning algorithms to classify a single physical quantity of the brake. Existing technologies primarily rely on traditional signal processing methods, such as frequency domain analysis and time domain statistical analysis, and employ conventional classification algorithms, such as neural networks or support vector machines, for condition identification.

[0003] However, the single monitoring parameter and fixed acquisition frequency fail to fully reflect the coupled characteristics of hydraulics, thermodynamics, mechanical wear, and electromagnetic fields in the electric shovel brake under heavy-load conditions, resulting in incomplete state information. Secondly, data imbalance is a serious issue. Abnormal state data of the electric shovel brake under intermittent heavy-load braking conditions is extremely scarce. Traditional data enhancement methods cannot maintain the consistency of physical features under heavy-load conditions, affecting model training effectiveness. Furthermore, existing feature extraction methods lack adaptability to changes in braking power and are unable to dynamically adjust feature extraction strategies based on actual operating conditions. Consequently, recognition accuracy varies significantly under different braking intensities.

[0004] How to achieve synchronous acquisition and effective preprocessing of multi-physical field coupling information to ensure the integrity of different physical quantity signals; on this basis, how to solve the data imbalance problem caused by the scarcity of abnormal samples under heavy load conditions and generate synthetic samples that maintain the consistency of physical features; further, it is necessary to solve the problem of lack of brake power adaptability in the feature extraction process, and dynamically adjust the attention weight and feature fusion strategy according to the actual working conditions; finally, it is necessary to solve the problem of effective fusion of multi-scale time series features and global state features to achieve accurate identification and real-time monitoring of the electric shovel brake status. Summary of the Invention

[0005] The present application provides a method for real-time monitoring of the electric shovel brake status, which is used to solve the technical problems of incomplete multi-physical field coupling information collection of the electric shovel brake under heavy load conditions, data imbalance caused by scarcity of abnormal state samples, lack of brake power adaptability in feature extraction, and poor multi-scale feature fusion effect, thereby improving the accuracy and real-time performance of electric shovel brake status monitoring.

[0006] In a first aspect, the present application provides a method for real-time monitoring of the brake status of an electric shovel, which includes: synchronously collecting the hydraulic cylinder pressure pulsation, brake disc surface temperature gradient, brake pad thickness change rate, and motor feedback torque of the electric shovel brake through multi-source sensors to obtain a brake multi-physics field coupling feature vector; performing heavy-load condition enhanced sampling processing on the brake multi-physics field coupling feature vector according to the braking power spectral density to obtain a heavy-load condition balance data set; extracting features from the heavy-load condition balance data set through a heavy-load condition adaptive multi-head attention mechanism to obtain a brake global state feature; performing multi-scale time series convolution processing on the heavy-load condition balance data set according to the braking cycle characteristics to obtain a brake time series evolution feature; and fusing and identifying the brake global state feature and the brake time series evolution feature based on braking power weighting to obtain a real-time status result of the electric shovel brake.

[0007] In a second aspect, the present application provides a real-time monitoring system for the brake status of an electric shovel, the real-time monitoring system for the brake status of an electric shovel comprising:

[0008] The acquisition module is used to synchronously collect the hydraulic cylinder pressure pulsation, brake disc surface temperature gradient, brake pad thickness change rate, and motor feedback torque of the electric shovel brake through multi-source sensors to obtain the brake multi-physics field coupling characteristic vector;

[0009] a sampling module, configured to perform heavy-load condition enhanced sampling processing on the brake multi-physics field coupling characteristic vector according to the braking power spectrum density to obtain a heavy-load condition balanced data set;

[0010] an extraction module, configured to extract features from the heavy-load condition balance dataset through a heavy-load condition adaptive multi-head attention mechanism to obtain a global brake state feature;

[0011] A convolution module is used to perform multi-scale time series convolution processing on the heavy load working condition balance data set according to the braking cycle characteristics to obtain the brake time series evolution characteristics;

[0012] The fusion module is used to fuse and identify the brake global state characteristics and the brake time sequence evolution characteristics based on the braking power weighting to obtain the real-time state result of the electric shovel brake.

[0013] In a third aspect, a device for real-time monitoring of the brake status of an electric shovel is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the device for real-time monitoring of the brake status of an electric shovel executes the above-mentioned method for real-time monitoring of the brake status of an electric shovel.

[0014] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium. When the computer-readable storage medium is run on a computer, the computer executes the above-mentioned method for real-time monitoring of the brake status of an electric shovel.

[0015] The technical solution provided in this application uses a multi-source sensor to synchronously collect the hydraulic cylinder pressure pulsation, brake disc surface temperature gradient, brake pad thickness change rate, and motor feedback torque of the electric shovel brake, thereby achieving comprehensive acquisition of brake multi-physics field coupling information. Compared with the traditional single-sensor monitoring method, it can capture the complete state information of the four physical fields of hydraulic, thermal, mechanical, and electromagnetic, avoiding the information loss problem caused by monitoring a single physical quantity. The technical feature of performing heavy-load condition enhanced sampling processing on the brake multi-physics field coupling feature vector based on the brake power spectrum density effectively solves the problem of extremely scarce abnormal state data of the electric shovel brake under intermittent heavy-load braking conditions. The synthetic samples generated by the SMOTE-HeavyLoad data enhancement method maintain the consistency of the physical characteristics of the heavy-load braking condition, so that the ratio of normal state and abnormal state samples in the training data set reaches a balanced state, significantly improving the recognition performance of the model under heavy-load conditions. The technical feature of extracting features from the heavy-load condition balance dataset through a heavy-load condition adaptive multi-head attention mechanism can dynamically adjust the attention weight allocation strategy according to the real-time braking power, automatically enhance the weights of temperature and pressure features under heavy-load conditions, and enhance the weights of displacement and torque features under light-load conditions. Compared with the fixed-weight attention mechanism, it significantly improves the feature extraction accuracy under different working conditions.

[0016] The technical feature performs multi-scale temporal convolution on a heavy-load condition balance dataset based on braking cycle characteristics. Using short-, medium-, and long-period convolution kernels specifically designed for the shovel brake duty cycle, this method simultaneously captures the transient response of a single braking action, the periodic changes of multiple consecutive braking cycles, and the degradation trend characteristics of long-term operation. This overcomes the limitation of traditional fixed-scale convolution kernels in adapting to the intermittent braking characteristics of shovels. The technical feature fuses and identifies the brake's global state features and brake's temporal evolution features based on brake power weighting. The feature fusion weights are dynamically adjusted based on the ratio of real-time braking power to historical maximum braking power, prioritizing global attention features under heavy-load conditions and temporal evolution features under light-load conditions, achieving condition-adaptive feature fusion. In the specific application of real-time shovel brake status monitoring, the application of core algorithms such as the heavy-load condition-adaptive multi-head attention mechanism and the multi-scale temporal convolutional network enables the system to automatically adjust feature extraction and fusion strategies based on the actual brake intensity. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 This is a schematic diagram of an embodiment of a method for real-time monitoring of the brake status of an electric shovel in an embodiment of the present application;

[0019] Figure 2 This is a schematic diagram of an embodiment of a real-time monitoring system for an electric shovel brake state in an embodiment of the present application;

[0020] Figure 3 The figure is a schematic block diagram of the structure of a device for real-time monitoring of the brake status of an electric shovel according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The embodiments of the present application provide a method for real-time monitoring of the brake status of an electric shovel. The terms "first," "second," "third," "fourth," and so on (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the numbers used in this way are interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatus.

[0022] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In one embodiment of the present application, a method for real-time monitoring of the brake status of an electric shovel includes:

[0023] Step S101: synchronously collect the hydraulic cylinder pressure pulsation, brake disc surface temperature gradient, brake pad thickness change rate, and motor feedback torque of the electric shovel brake through multi-source sensors to obtain a brake multi-physical field coupling characteristic vector;

[0024] Step S102: performing heavy-load condition enhanced sampling processing on the brake multi-physics field coupling characteristic vector according to the braking power spectrum density to obtain a heavy-load condition balanced data set;

[0025] Step S103: extract features from the heavy-load condition balance dataset through a heavy-load condition adaptive multi-head attention mechanism to obtain brake global state features;

[0026] Step S104: performing multi-scale time series convolution processing on the heavy-load working condition balance data set according to the braking cycle characteristics to obtain the brake time series evolution characteristics;

[0027] Step S105: Based on the brake power weighting, the brake global state characteristics and the brake time series evolution characteristics are fused and identified to obtain the real-time state result of the electric shovel brake.

[0028] It is understandable that the execution subject of the present application can be a real-time monitoring system for the brake status of an electric shovel, or a terminal or a server, which is not limited here. The embodiment of the present application is described by taking the server as the execution subject as an example.

[0029] Specifically, multiple sensors are deployed at key locations within the shovel's brake system to enable simultaneous acquisition of multi-physics data. Pressure sensors collect hydraulic cylinder pressure pulsation data at a frequency of 1000 Hz, infrared temperature sensors collect temperature gradient data at multiple points on the brake disc surface at a frequency of 100 Hz, laser displacement sensors measure the rate of change of brake pad thickness at a frequency of 500 Hz, and torque sensors collect motor feedback torque data at a frequency of 1000 Hz. This differentiated acquisition frequency strategy optimizes the characteristic frequency bands of different physical quantities. Pressure signals require high-frequency acquisition to capture the rapid response of the hydraulic system, while temperature signals change relatively slowly and therefore utilize lower frequencies. The collected raw data undergoes filtering preprocessing based on brake power spectral density analysis to remove strong electromagnetic interference and mechanical vibration noise in the mining environment, ultimately constructing the brake's multi-physics coupling characteristic vector.

[0030] To address the extremely scarce data on abnormal states of electric shovel brakes under intermittent heavy-load braking conditions, the SMOTE-HeavyLoad data augmentation method was employed. First, the instantaneous brake power was calculated based on the motor feedback torque and angular velocity. The time-domain power signal was then converted into the frequency-domain braking power spectral density (BPSD) using a fast Fourier transform (FFT). The FFT is a mathematical algorithm that converts time-domain signals into frequency-domain signals, analyzing the signal's frequency-domain characteristics by decomposing its frequency components. Based on the energy distribution characteristics of the BPSD, braking conditions were classified into three categories: light-load braking, medium-load braking, and heavy-load braking. Heavy-load braking was defined as conditions with a power density greater than 150 kW / Hz. The k-nearest neighbor algorithm was used to calculate the similarity between minority samples in the heavy-load braking condition. This algorithm determined the k most similar neighbor samples by calculating the Euclidean distance between samples. When the power spectral density difference between two samples was less than a preset threshold, a weighting factor was calculated based on the power spectral similarity and constrained synthesis was performed. The generated new sample maintained the consistency of the physical characteristics of the heavy-load braking condition.

[0031] The heavy-load adaptive multi-head attention mechanism dynamically adjusts the attention weight allocation strategy by comparing real-time braking power with a preset power threshold. When the real-time braking power exceeds the heavy-load power threshold of 150kW, attention weights are shifted toward temperature and pressure features. When the power is in the medium-load range, a balanced weight distribution is adopted. When the power is below the light-load threshold of 50kW, attention weights are shifted toward displacement and torque features. Eight parallel attention heads process different physical field information: heads 1-2 specifically process hydraulic pressure information, heads 3-4 process thermodynamic information, heads 5-6 process mechanical wear information, and heads 7-8 process motor torque information. The attention mechanism is a computational method that simulates human attention. It calculates the importance weights of different features through matrix multiplication operations on the query matrix, key matrix, and value matrix. The basic attention weights are weighted and modulated by the adaptive braking power weighting factor, and the global brake state characteristics are ultimately derived through weighted fusion.

[0032] Multi-scale temporal convolution processing uses convolution kernels of three different sizes to capture the state changes of the electric shovel brake at different time scales. The short-period convolution layer uses a 3×1 convolution kernel with a stride of 1, specifically capturing transient response characteristics during a single braking action, such as the rapid buildup and release of hydraulic pressure. Convolution is a mathematical operation that extracts features from input data using a sliding window. Local features are extracted by point-by-point multiplication and summation of the convolution kernel with the input data. The medium-period convolution layer uses a 15×1 convolution kernel with a stride of 3 to extract periodic changes during multiple consecutive braking operations. The long-period convolution layer uses a 45×1 convolution kernel with a stride of 9 to capture degradation trends during long-term operation. The brake state gating mechanism uses a gating unit to control the fusion weights of features at different time scales. The gating unit uses a sigmoid activation function to generate weights between 0 and 1 to control the flow of information.

[0033] The braking power weighted fusion recognition process first performs feature cascade on the global state features and time-series evolution features to form a comprehensive feature vector. Feature cascade is an operation that splices multiple feature vectors by dimension, merging feature information from different sources into a unified representation. The feature fusion weight is calculated based on the ratio of real-time braking power to the historical maximum braking power. This weight reflects the proportional relationship between the current braking intensity and the historical maximum intensity. The three-layer fully connected neural network performs nonlinear mapping on the power-weighted fusion features. The fully connected layer realizes the mapping from features to outputs through the linear transformation of the weight matrix and bias vector plus the nonlinear transformation of the activation function. The focal loss function is specifically used to deal with the problem of sample imbalance. It improves the classification performance by assigning higher weights to difficult-to-separate samples, and finally outputs the recognition results of four categories of states: normal state, slight wear, overheating failure, and severe wear.

[0034] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0035] The pressure sensor performs high-frequency acquisition and processing on the hydraulic cylinder pressure pulsation to obtain the hydraulic cylinder pressure pulsation data;

[0036] The infrared temperature sensor is arranged at multiple points to collect and process the temperature gradient of the brake disc surface to obtain the temperature gradient data of the brake disc surface;

[0037] The laser displacement sensor measures the brake pad thickness change rate with micron-level precision to obtain brake pad thickness change rate data;

[0038] The torque sensor synchronously collects and processes the motor feedback torque to obtain the motor feedback torque data;

[0039] Based on the brake power spectrum density analysis, the hydraulic cylinder pressure pulsation data, brake disc surface temperature gradient data, brake pad thickness change rate data, and motor feedback torque data are filtered and preprocessed to obtain the brake multi-physical field coupling characteristic vector.

[0040] Specifically, the pressure sensor uses a high-frequency acquisition method of 1000Hz to monitor the pressure pulsation signal inside the hydraulic cylinder in real time. High-frequency acquisition refers to the technical means of sampling a large amount of data per unit time, collecting 1000 pressure data points per second to capture the pressure fluctuation characteristics of the hydraulic system under heavy-load braking conditions. The hydraulic cylinder pressure pulsation data reflects the dynamic response process of the brake hydraulic system, including the pressure variation patterns in the three stages of pressure buildup, maintenance, and release. The infrared temperature sensor uses a multi-point arrangement to set multiple temperature measurement points at different radial and circumferential positions on the brake disc surface. Each temperature measurement point collects temperature data at a frequency of 100Hz. Multi-point arrangement acquisition refers to the method of simultaneously collecting data at multiple key locations of the measured object. The temperature gradient data on the brake disc surface reflects the spatial distribution characteristics of heat generation and transfer during braking. The temperature gradient refers to the rate of temperature change in space and is calculated by the difference between multiple temperature data points.

[0041] Laser displacement sensors utilize the principle of laser ranging to measure brake pad thickness with micron-level accuracy. Micron-level measurement refers to high-precision measurement technology with micron-level accuracy, with a measurement accuracy of 1 micron. Brake pad thickness change rate data is obtained by continuously measuring the brake pad thickness and calculating its rate of change over time, reflecting the rate and extent of brake pad wear. The torque sensor uses the same time base as other sensors for synchronous data acquisition. Synchronous data acquisition involves multiple sensors collecting data at the same time, ensuring temporal correspondence between different physical quantities. Motor feedback torque data records the changes in motor output torque during braking, reflecting the braking torque and effectiveness of the brake.

[0042] Braking power spectral density analysis is a method that performs frequency domain analysis of braking power signals based on power spectral density theory. Signal characteristics are identified by calculating the energy distribution of braking power at different frequency components. The filtering preprocessing process first calculates the instantaneous power of the brake and multiplies the motor feedback torque data with the corresponding angular velocity data to obtain the instantaneous power time series. A fast Fourier transform is then performed on the instantaneous power signal to convert the time domain power signal into a power spectral density function in the frequency domain, obtaining information on the energy distribution of the power signal at each frequency component. A bandpass filter is designed based on the frequency characteristics of the power spectral density. The filter's passband range is determined based on the operating frequency characteristics of the electric shovel brake, effectively removing interference signals such as 50Hz power frequency interference, high-frequency mechanical vibration noise, and low-frequency baseline drift in the mining environment.

[0043] The filtered hydraulic cylinder pressure pulsation data, brake disc surface temperature gradient data, brake pad thickness change rate data, and motor feedback torque data are aligned in time series to form a four-dimensional feature vector. The brake multi-physics field coupling feature vector is a comprehensive feature representation that combines the characteristic parameters of four physical fields: hydraulics, thermodynamics, mechanical wear, and electromagnetics. Each element in the vector corresponds to the measured value of a physical quantity at a specific moment. The hydraulic field characteristics reflect the operating state of the brake hydraulic system, the thermodynamic field characteristics reflect the heat generation and transfer patterns during braking, the mechanical wear field characteristics characterize the wear state and wear trend of the brake pad, and the electromagnetic field characteristics describe the torque output characteristics of the motor during braking.

[0044] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0045] The instantaneous power of the brake is calculated and processed according to the motor feedback torque and angular velocity to obtain the instantaneous power data of the brake;

[0046] The fast Fourier transform algorithm is used to perform frequency domain conversion on the brake instantaneous power data to obtain the brake power spectrum density characteristics;

[0047] The braking conditions are classified into three levels based on the energy distribution of the braking power spectrum density characteristics, namely light-load braking, medium-load braking and heavy-load braking.

[0048] The minority class samples of heavy-load braking conditions are processed by similarity calculation using the k-nearest neighbor algorithm to obtain the power spectrum similarity weight factor;

[0049] The heavy-load braking state samples are subjected to constrained synthesis generation processing according to the power spectrum similarity weight factor to obtain a heavy-load working condition balanced data set.

[0050] Specifically, the instantaneous power of the brake is calculated by multiplying the motor's feedback torque by the corresponding angular velocity. The formula for calculating instantaneous power is: power equals torque multiplied by angular velocity. The motor's feedback torque data is acquired in real time by a torque sensor, while the angular velocity data is obtained via an encoder or speed sensor. The instantaneous power data of the brake reflects the dynamic characteristics of energy conversion during braking, with the power value directly corresponding to changes in braking intensity and load. The instantaneous power data is arranged in a time series according to the time of acquisition, forming a time-domain signal showing the power variation over time. This signal contains information about the power characteristics of the brake under different operating conditions.

[0051] The Fast Fourier Transform (FFT) algorithm is an efficient discrete Fourier transform (DFT) calculation method used to convert time-domain signals into frequency-domain signals for spectrum analysis. The Fast Fourier Transform (FFT) decomposes a discrete Fourier transform of length N into multiple shorter transforms using a divide-and-conquer strategy, significantly reducing computational complexity. After the brake's instantaneous power data is processed by the Fast Fourier Transform (FFT), the time-domain power signal is decomposed into complex representations of different frequency components, each of which contains amplitude and phase information. The braking power spectral density feature is obtained by calculating the power density of the power signal at each frequency point. The power density value represents the power within a unit frequency bandwidth and reflects the energy distribution pattern of the power signal in the frequency domain.

[0052] The classification of braking conditions is based on the energy distribution characteristics of the braking power spectrum density. Energy distribution refers to the degree of energy concentration and distribution pattern of the power signal within different frequency ranges. Light-load braking conditions correspond to situations where the power spectrum density peak is less than 50kW / Hz. At this time, the load on the brake is small, and the spectral energy of the power signal is mainly concentrated in the low-frequency band. The power spectrum density peak of the medium-load braking condition is between 50kW / Hz and 150kW / Hz, and the spectral energy distribution of the power signal is relatively uniform, covering medium-low frequency and some high-frequency components. The power spectrum density peak of the heavy-load braking condition exceeds 150kW / Hz, and the spectral energy of the power signal extends to the high-frequency band, reflecting the complex dynamic characteristics of the brake under high-intensity working conditions. The classification criteria for the three operating conditions are determined based on the statistical analysis results of a large amount of actual operating data of electric shovel brakes.

[0053] The k-nearest neighbor algorithm determines the k nearest neighbors of each sample by calculating the distance metric between samples. The distance metric uses the Euclidean distance calculation method. Minority samples of heavy-load braking conditions refer to the small number of heavy-load braking state samples in the training data set. These samples are scarce due to the low frequency of heavy-load conditions. The similarity calculation process first extracts the power spectrum density feature vector of each heavy-load braking sample, and then calculates the power spectrum density difference between the target sample and other samples. The power spectrum similarity weight factor is calculated based on the comparison result of the power spectrum density difference value and the preset threshold. When the power spectrum density difference between the two samples is less than the threshold, the weight factor value is close to 1, indicating that the samples are highly similar. When the difference value approaches or exceeds the threshold, the weight factor value tends to 0.

[0054] Constrained synthetic generation is an improved version of the SMOTE algorithm, specifically designed to generate samples based on the physical characteristic constraints of heavy-duty braking conditions. SMOTE, short for synthetic minority oversampling technology, generates new synthetic samples by linearly interpolating between minority class samples and their nearest neighbors. Constrained synthetic generation adds a power spectrum similarity constraint to the traditional SMOTE algorithm, ensuring that the generated synthetic samples maintain the physical characteristics of heavy-duty braking conditions. The synthetic sample generation process first selects a minority class sample from a heavy-duty braking condition as a baseline sample, then randomly selects a reference sample from its k nearest neighbors. The eigenvalue of the new sample is calculated by adding the baseline sample's eigenvalue to a weighting factor multiplied by the difference between the baseline and reference samples. The weighting factor ranges from 0 to the power spectrum similarity weighting factor. A balanced dataset for heavy-duty braking conditions is generated by merging the original heavy-duty braking samples with the generated synthetic samples, ensuring a relative balance between the number of heavy-duty braking condition samples and the number of samples from other conditions.

[0055] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0056] The attention weight allocation strategy is dynamically adjusted based on the comparison result between the real-time braking power and the preset power threshold to obtain the braking power adaptive weight factor;

[0057] The heavy-load working condition balance data set is processed through eight parallel attention heads for multi-physics field information processing to obtain hydraulic pressure attention features, thermodynamic attention features, mechanical wear attention features, and motor torque attention features.

[0058] Based on the query matrix, key matrix, and value matrix, the attention weight of each attention head is calculated to obtain the basic attention weight distribution;

[0059] The adaptive attention weight distribution is obtained by weighted modulation of the basic attention weight distribution by the adaptive weight factor of the braking power;

[0060] According to the adaptive attention weight distribution, the hydraulic pressure attention features, thermodynamic attention features, mechanical wear attention features and motor torque attention features are weightedly fused to obtain the global state features of the brake.

[0061] Specifically, the real-time braking power is compared with preset power thresholds by numerically comparing the current braking power value with a heavy-load power threshold of 150kW and a light-load power threshold of 50kW. The comparison results determine the adjustment direction of the attention weight allocation strategy. Dynamic adjustment processing uses a piecewise function to calculate the weight allocation ratio based on the power comparison results. When the real-time braking power exceeds the heavy-load power threshold, the temperature feature weight is set to 0.4, the pressure feature weight is set to 0.35, the displacement feature weight is set to 0.15, and the torque feature weight is set to 0.1. When the real-time braking power is between the light-load and heavy-load thresholds, the weights of the four features are evenly distributed, with a value of 0.25. When the real-time braking power is below the light-load power threshold, the displacement feature weight is set to 0.4, the torque feature weight is set to 0.35, the temperature feature weight is set to 0.15, and the pressure feature weight is set to 0.1. The adaptive braking power weight factor is normalized to ensure that the sum of all weights is equal to 1, reflecting the importance of each physical field feature under different braking conditions.

[0062] The eight parallel attention heads utilize a multi-head attention mechanism, a parallel computing architecture. Each attention head independently processes a specific dimension of physical field information. The multi-head attention mechanism is an attention computing method in deep learning that captures different aspects of data through parallel computation across multiple attention heads. Each sample in the heavy-load balancing dataset contains time series data for four physical quantities: hydraulic cylinder pressure, brake disc temperature, brake pad displacement, and motor torque. The eight attention heads process these data according to their functional divisions. Attention heads 1 and 2 specifically process hydraulic cylinder pressure pulsation data, extracting hydraulic system characteristics such as pressure buildup rate, peak pressure, and pressure hold time. Attention heads 3 and 4 specifically process brake disc surface temperature gradient data, extracting thermodynamic characteristics such as temperature rise rate, maximum temperature, and temperature distribution uniformity. Attention heads 5 and 6 specifically process brake pad thickness change data, extracting mechanical wear characteristics such as wear rate, cumulative wear, and wear pattern. Attention heads 7 and 8 specifically process motor feedback torque data, extracting motor torque characteristics such as torque drop rate, braking torque, and torque ripple.

[0063] The query matrix, key matrix, and value matrix are the three core components of the attention mechanism, used to calculate attention weights and weighted feature representations, respectively. The query matrix maps input features to query vectors through a linear transformation, the key matrix maps input features to key vectors, and the value matrix maps input features to value vectors. The attention weight calculation process first calculates the matrix product of the query matrix and the transposed key matrix, and then divides it by a scaling factor for numerical stabilization. The scaling factor is usually set to the square root of the feature dimension. The calculated result is normalized using the softmax function, which converts any real vector into a probability distribution to ensure that the sum of the attention weights is equal to 1. The basic attention weight distribution represents the degree of attention each attention head pays to different time steps and feature dimensions when processing the input data. Larger weights indicate more important information at the corresponding position.

[0064] The weighted modulation of the basic attention weight distribution by the adaptive brake power weight factor is achieved through element-by-element multiplication. The weighted modulation process multiplies the basic attention weight by the adaptive weight factor element by element. The purpose of weighted modulation is to dynamically adjust the importance weights of each physical field feature based on the power characteristics of the current braking condition, strengthening the weights of temperature and pressure features under heavy load conditions and displacement and torque features under light load conditions. The adaptive attention weight distribution is renormalized to ensure that the modulated weight distribution still meets the constraints of the probability distribution. The modulated weight distribution can automatically adjust the focus of feature extraction based on the actual operating state of the brake, focusing on the most relevant physical field information under different operating conditions.

[0065] The weighted fusion process linearly combines the four types of attention features through an adaptive attention weight distribution. The weighted fusion calculation process is to multiply each type of attention feature by the corresponding weight value and then sum them. The hydraulic pressure attention feature contains the dynamic response characteristics of the brake hydraulic system, the thermodynamic attention feature contains the heat generation and transfer characteristics during the braking process, the mechanical wear attention feature contains the wear status and trend characteristics of the brake pad, and the motor torque attention feature contains the torque output characteristics of the motor during the braking process. The brake global state feature is the weighted combination result of the four types of attention features, which comprehensively reflects the overall state information of the brake in the four physical fields of hydraulics, thermodynamics, mechanical wear, and electromagnetics. The dimensionality of the global state feature is equal to the dimensionality of a single attention feature, but the information content includes the coupling characteristics of multiple physical fields.

[0066] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0067] According to the single braking cycle characteristics of the electric shovel, a short-cycle convolution kernel is applied to the heavy-load working condition balance data set to obtain the brake transient response characteristics.

[0068] According to the characteristics of multiple consecutive braking cycles of the electric shovel, the heavy-load working condition balance data set is processed with a medium-cycle convolution kernel to obtain the periodic variation characteristics of the brake.

[0069] Based on the long-term working cycle characteristics of the electric shovel, the heavy-load working condition balance data set is processed with a long-period convolution kernel to obtain the brake degradation trend characteristics;

[0070] The brake state gating mechanism is used to perform gate control processing on the brake transient response characteristics, brake periodic change characteristics, and brake degradation trend characteristics to obtain multi-scale feature fusion weights;

[0071] Based on the multi-scale feature fusion weights, the brake transient response characteristics, brake periodic change characteristics, and brake degradation trend characteristics are weightedly fused to obtain the brake time series evolution characteristics.

[0072] Specifically, the characteristics of a single braking cycle on an electric shovel refer to the variations in various physical quantities during a complete braking maneuver. A single braking cycle typically lasts 1 to 3 seconds and includes four phases: brake activation, pressure buildup, brake hold, and pressure release. A short-period convolution kernel uses a 3×1 kernel size with a step size of 1, specifically designed to capture the rapidly changing characteristics of a single braking process. Convolution kernel processing extracts features from a balanced dataset under heavy load conditions by sliding a convolution window across the time dimension. The convolution operation multiplies the kernel weights element-by-element with the input data at the corresponding position and then sums the results. Brake transient response characteristics primarily include the rapid buildup of hydraulic pressure, the transient rise in brake disc temperature, the displacement jump at the moment of brake pad contact, and the sharp drop in motor torque. The small size of the short-period convolution kernel accurately captures the rapid millisecond-level changes during braking. The temporal resolution of the convolution output feature map remains the same as that of the input data, ensuring the temporal accuracy of transient features.

[0073] The characteristics of continuous multiple braking cycles for electric shovels reflect the correlation and periodicity between multiple braking actions during continuous operation. A continuous braking cycle typically spans 10 to 30 seconds and encompasses multiple single braking actions and their intervals. A medium-periodic convolution kernel with a kernel size of 15×1 and a step size of 3 is used to extract medium-timescale features spanning multiple braking actions. A kernel size of 15 corresponds to a time window covering five single braking cycles, and a step size of 3 means that the convolution window moves three time steps each time the next convolution operation is performed. Periodic variations in brake performance include the cumulative increase and intermittent heat dissipation of brake disc temperature, the periodic fluctuation pattern of hydraulic system pressure, the gradual accumulation of brake pad wear, and the changing trend of motor torque during continuous braking. The medium-periodic convolution kernel, with its large receptive field, is able to capture the periodic patterns and cumulative effects of continuous brake operation. The temporal resolution of the convolution output is reduced by a factor of 3 compared to the input data, highlighting medium-timescale variations.

[0074] The long-duration duty cycle characteristics of electric shovels describe the long-term evolution of brake performance over hours of continuous operation. A long-duration duty cycle typically lasts 5 to 15 minutes and includes dozens of complete braking cycles. A long-period convolution kernel with a kernel size of 45×1 and a step size of 9 is specifically designed to capture the long-term degradation and performance decay characteristics of brakes. A kernel size of 45 corresponds to a time window covering 15 consecutive braking cycles, enabling the capture of slowly changing trends over extended periods of time. Signatures of brake degradation trends include long-term wear trends in brake pad thickness, baseline drift in brake disc surface temperature, slow decay in hydraulic system pressure, and a gradual decrease in motor torque output capacity. The large receptive field of the long-period convolution kernel enables the identification of performance degradation patterns over long periods of use. The temporal resolution of the convolution output is further reduced, emphasizing trends over long timescales.

[0075] The brake state gating mechanism adopts the design concept of a gated recurrent unit, which controls the information flow and fusion weights of features at different time scales. The gating unit consists of two core components: an update gate and a reset gate. The update gate determines how much historical information is retained, while the reset gate determines how much is forgotten. The gating unit control process first takes the three time scale features as input and calculates the activation values ​​of the update and reset gates through a linear transformation and a sigmoid activation function. The sigmoid activation function maps the input values ​​to a range between 0 and 1. An output value close to 1 indicates that the information is fully passed, while an output value close to 0 indicates that the information is blocked. The multi-scale feature fusion weights are determined by the output value of the gating unit, and the weights reflect the importance of each time scale feature in the current braking state. The gating mechanism adaptively adjusts the contribution ratio of different time scale features based on the actual operating state of the brake, prioritizing transient features when the brake is operating normally and degradation trends when the brake performance is degraded.

[0076] The weighted fusion process linearly combines the three time-scale features through the multi-scale feature fusion weights. The mathematical expression of weighted fusion is the sum of the products of each feature and its corresponding weight. The brake transient response characteristics, brake periodic change characteristics, and brake degradation trend characteristics represent the short-term, medium-term, and long-term time evolution information respectively. The fusion result of the three can fully describe the time series dynamic characteristics of the brake. The brake time series evolution characteristics are a comprehensive representation of multi-scale time characteristics, which contain complete time scale information from millisecond-level transient responses to minute-level long-term trends. The dimensionality of the time series evolution characteristics is equal to the dimensionality of a single time scale feature, but the information content covers the state change laws of the brake at different time scales.

[0077] In a specific embodiment, the process of performing the short-period convolution kernel processing on the heavy-load working condition balance data set according to the single braking cycle characteristics of the electric shovel may specifically include the following steps:

[0078] A 3×1 short-period convolution kernel is used to perform convolution operations on the heavy-load working condition balance data set with a step size of 1 to obtain the hydraulic pressure buildup process characteristics;

[0079] A short-period convolution kernel of size 3×1 is used to perform convolution operation on the heavy-load working condition balance data set with a step size of 1 to obtain the hydraulic pressure release process characteristics;

[0080] The ReLU activation function is used to perform nonlinear transformation on the hydraulic pressure building process characteristics and the hydraulic pressure release process characteristics to obtain the transient characteristics after activation;

[0081] The batch normalization algorithm is used to perform data normalization on the activated transient features to obtain normalized transient features;

[0082] The normalized transient features are processed by time series slicing based on a single braking time window to obtain the brake transient response characteristics.

[0083] Specifically, a 3×1 short-period convolution kernel has a receptive field of three time steps in the temporal dimension, enabling simultaneous processing of data from three consecutive time points. The kernel weight parameters are trained and optimized using a backpropagation algorithm. A convolution operation with a stride of 1 means that the kernel moves forward one time step each time it performs the next convolution calculation. This dense sliding approach ensures sufficient feature extraction for each time point in the time series. The hydraulic pressure buildup process features are extracted for the period of time when the hydraulic cylinder pressure rises from a low to a high pressure state in a heavy-duty balancing dataset. The convolution kernel learns characteristic patterns such as the slope, acceleration, and amplitude of the pressure rise to identify the pressure buildup process. During the convolution operation, the kernel weights are element-wise multiplied by the input data window and then summed to generate a scalar output value representing the characteristic strength of the hydraulic pressure buildup process within the current time window. The hydraulic pressure buildup process features reflect the response speed and pressure-building capacity of the brake hydraulic system, including key information such as the starting time of the pressure rise, the rate of rise, and the time it takes to reach the target pressure.

[0084] The hydraulic pressure release process features are extracted using the same 3×1 convolutional kernel structure, but the training target is the period when the hydraulic cylinder pressure drops from a high-pressure state to a low-pressure state. The convolutional kernel learns to identify characteristic patterns of pressure drops, including characteristic parameters such as the trigger moment of pressure release, the slope of the drop, the pressure hold time, and the final residual pressure. A convolution operation with a step size of 1 ensures accurate capture of changes at every time point during the pressure release process, and the temporal resolution of the convolution output feature map remains consistent with the input data. The hydraulic pressure release process features and the pressure buildup process features form complementary information representations in the feature space, jointly describing the complete dynamic response characteristics of the hydraulic system. Although the two convolutional kernels have the same structure but different weight parameters, they specifically learn and extract features for the different physical processes of pressure rise and fall, respectively.

[0085] The full name of the ReLU activation function is the rectified linear unit. Its mathematical expression is that the output is equal to the maximum value of the input value and zero, that is, the negative input is set to zero, and the positive input remains unchanged. The nonlinear transformation processing takes the characteristics of the hydraulic pressure buildup process and the hydraulic pressure release process as input, introduces nonlinear characteristics through the ReLU function, and enhances the expressive ability of the neural network. The role of the activation function is to eliminate the negative components in the features, retain the feature information of the positive activation, and provide nonlinear input for subsequent network layers. After activation, the transient characteristics retain the key feature information of the hydraulic pressure buildup and release process, but remove possible negative interference signals. The ReLU activation function has low computational complexity and simple gradient calculation, which helps to accelerate the training and inference processes of the neural network.

[0086] Batch normalization is a commonly used data normalization technique in deep learning. It normalizes features by calculating the mean and variance of batches of data. Data normalization first calculates the sample mean and sample variance of the activated transient features along the batch dimension. Each feature value is then normalized by subtracting the mean and dividing by the standard deviation. The mathematical formula for batch normalization is: the normalized value equals the original value minus the batch mean, divided by the batch standard deviation plus the square root of a small constant, which is used to prevent division by zero errors. Normalized transient features have a distribution characteristic of zero mean and unit variance, which helps stabilize the training process of neural networks and prevent gradient explosion or vanishing problems. Batch normalization also includes learnable scaling and bias parameters, allowing the network to adaptively adjust the distribution range of normalized features.

[0087] The single braking time window is a time division unit determined based on the actual operating characteristics of the electric shovel brake. It is usually set to a time length of 1 to 3 seconds, corresponding to a complete braking action cycle. Time series slicing processing segments the continuous normalized transient features according to the single braking time window. Each slice contains the feature information of a complete braking process. The slicing process first identifies the start and end times of the braking action, and then divides the continuous feature sequence into independent braking events based on these time markers. The brake transient response feature is a statistical representation of the features within each time window slice, including statistics such as the mean, variance, maximum, and minimum values ​​of the features within the time period. Time series slicing processing ensures the temporal alignment of feature extraction, making the features of different braking events comparable and consistent.

[0088] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0089] Perform feature cascade processing on the brake global state features and the brake temporal evolution features to obtain a comprehensive feature vector;

[0090] The feature fusion weight is calculated based on the ratio of the real-time braking power to the historical maximum braking power to obtain the braking power weighting coefficient;

[0091] Based on the braking power weighting coefficient, the brake global state characteristics and brake time series evolution characteristics are weighted and fused to obtain the power weighted fusion characteristics;

[0092] The three-layer fully connected neural network performs nonlinear mapping on the power-weighted fusion features to obtain the state classification probability distribution;

[0093] The focal loss function is used to calculate the state classification probability distribution and obtain the recognition results of four types of states: normal state, slight wear, overheating failure, and severe wear.

[0094] The operating status of the electric shovel brake is determined and processed in real time according to the four types of status recognition results to obtain the real-time status results of the electric shovel brake.

[0095] Specifically, feature concatenation combines the brake global state features and brake temporal evolution features into a single feature representation through dimensional splicing. The cascade operation performs vector concatenation on the feature dimensions. The brake global state features contain multi-physics field coupling information extracted through a heavy-load condition adaptive multi-head attention mechanism, with a dimension of 128. The brake temporal evolution features contain time evolution information obtained through multi-scale temporal convolution processing, with a dimension of 256. Feature concatenation concatenates the two feature vectors in dimensional order, resulting in a composite feature vector with a dimension of 128+256=384. The first 128 elements correspond to the global state features, and the last 256 elements correspond to the temporal evolution features. The composite feature vector contains both the spatial state information and the temporal evolution information of the brake, providing a complete feature foundation for subsequent state recognition. Cascade processing ensures the complete preservation of feature information from different sources and avoids information loss that may occur during feature fusion.

[0096] The ratio of real-time braking power to historical maximum braking power is calculated by dividing the current braking intensity relative to the historical maximum intensity. This ratio reflects the relative strength of the brake's current workload. The historical maximum braking power is calculated by summing the maximum braking power values ​​recorded during the brake's past operation and represents the maximum workload the brake has experienced. The feature fusion weight calculation uses the normalized ratio as the weight coefficient. The formula is: the weight is equal to the real-time braking power divided by the historical maximum braking power plus a small constant. The small constant is set to 0.001 to prevent division by zero errors. The braking power weight coefficient ranges from 0 to 1. When the real-time braking power is close to the historical maximum, the weight is close to 1, and when the real-time braking power is very low, the weight is close to 0. This weight coefficient controls the contribution ratio of global state features and time-series evolution features in the fusion process, giving more weight to global state features under high-power conditions and to time-series evolution features under low-power conditions.

[0097] The weighted fusion process uses a linear weighted combination to fuse the two features according to the braking power weighting coefficient. The fusion formula is: the weighted feature is equal to the weight coefficient multiplied by the global state feature plus one minus the weight coefficient multiplied by the time series evolution feature. The braking power weighting coefficient determines the weight of the global state feature, and one minus the weight coefficient determines the weight of the time series evolution feature. The sum of the two weights is always equal to 1. The dimension of the power-weighted fusion feature is the same as the original feature vector, which is 384 dimensions, but the feature content dynamically adjusts the proportion of global and time series information according to the current braking power. The weighted fusion mechanism makes the system pay more attention to the coupled state information of multiple physical fields under heavy-load braking conditions, and pays more attention to the evolution trend information of the time series under light-load braking conditions. The fused features not only maintain the complete information of the original features, but also highlight the most relevant feature components according to the characteristics of the working conditions.

[0098] The three-layer fully connected neural network uses a layer-by-layer linear transformation and nonlinear activation structure to perform nonlinear mapping on the power-weighted fusion features. The network consists of an input layer, two hidden layers, and an output layer. The first hidden layer contains 512 neurons, which uses a linear transformation to map the 384-dimensional input features into a 512-dimensional intermediate representation. The mathematical expression for the linear transformation is that the output equals the weight matrix multiplied by the input vector plus the bias vector. The second hidden layer contains 256 neurons, further compressing the 512-dimensional features into a 256-dimensional representation. The third hidden layer contains 128 neurons, ultimately compressing the features to 128 dimensions. Each hidden layer is followed by a ReLU activation function for nonlinear transformation, enhancing the network's nonlinear representation capabilities. The output layer contains four neurons, one for each of the four state categories: normal, minor wear, overheating failure, and severe wear. The softmax activation function is used to convert the neuron outputs into a probability distribution. The state classification probability distribution represents the probability of the brake belonging to each state category, with the sum of the four probabilities equal to 1.

[0099] The focal loss function is a loss function specifically designed to address class imbalance issues. It improves classification performance by assigning higher weights to difficult-to-classify samples. The loss calculation process first compares the state classification probability distribution with the true label and calculates the difference between the predicted probability and the true probability for each sample. The mathematical formula for focal loss is that the loss is equal to the negative modulation factor multiplied by the logarithm of the true probability. The modulation factor is equal to alpha multiplied by one minus the predicted probability to the power of gamma, where alpha is set to 0.75 and gamma is set to 2. The role of the modulation factor is to reduce the loss weight of easy-to-classify samples and increase the loss weight of difficult-to-classify samples, so that the network focuses more on learning the characteristics of difficult samples. The results of the loss calculation are used in the backpropagation algorithm to update the network weights and optimize the network parameters through gradient descent. The four-category state recognition results are obtained by selecting the category label corresponding to the maximum probability value in the probability distribution. When the probability of the normal state is the highest, the normal state label is output, and when the probability of other states is the highest, the corresponding abnormal state label is output.

[0100] The real-time judgment processing makes the final state judgment based on the four types of state recognition results combined with the confidence threshold. The confidence threshold is used to control the sensitivity and accuracy of state recognition. When the maximum probability value exceeds the preset confidence threshold of 0.8, the system outputs the corresponding state recognition result. When the maximum probability value is lower than the threshold, the system outputs an uncertain state and recommends further testing. The real-time state result of the electric shovel brake contains information such as state category, confidence probability, and judgment timestamp. The state category indicates the current health status of the brake, and the confidence probability reflects the reliability of the judgment result. The real-time state result also includes corresponding maintenance recommendations. It is recommended to continue monitoring in normal status, strengthen inspections in case of minor wear, check the cooling system in case of overheating failure, and replace the brake pads immediately in case of severe wear.

[0101] The above describes the method for real-time monitoring of the brake status of an electric shovel in the embodiment of the present application. The following describes the system for real-time monitoring of the brake status of an electric shovel in the embodiment of the present application. Figure 2 In one embodiment of the present application, a system for real-time monitoring of the brake status of an electric shovel includes:

[0102] The acquisition module is used to synchronously collect the hydraulic cylinder pressure pulsation, brake disc surface temperature gradient, brake pad thickness change rate, and motor feedback torque of the electric shovel brake through multi-source sensors to obtain the brake multi-physics field coupling characteristic vector;

[0103] a sampling module, configured to perform heavy-load condition enhanced sampling processing on the brake multi-physics field coupling characteristic vector according to the braking power spectrum density to obtain a heavy-load condition balanced data set;

[0104] an extraction module, configured to extract features from the heavy-load condition balance dataset through a heavy-load condition adaptive multi-head attention mechanism to obtain a global brake state feature;

[0105] A convolution module is used to perform multi-scale time series convolution processing on the heavy load working condition balance data set according to the braking cycle characteristics to obtain the brake time series evolution characteristics;

[0106] The fusion module is used to fuse and identify the brake global state characteristics and the brake time sequence evolution characteristics based on the braking power weighting to obtain the real-time state result of the electric shovel brake.

[0107] above Figure 2 The system for real-time monitoring of the brake status of an electric shovel in an embodiment of the present invention is described in detail from the perspective of modular functional entities. The device for real-time monitoring of the brake status of an electric shovel in an embodiment of the present invention is described in detail from the perspective of hardware processing.

[0108] Reference Figure 3 In an embodiment of the present invention, a device for real-time monitoring of the brake status of an electric shovel is further provided. The device for real-time monitoring of the brake status of an electric shovel can be a server, and its internal structure can be as follows: Figure 3 As shown. The real-time monitoring device for the electric shovel brake status includes a processor, a memory, a display screen, an input device, a network interface, and a database connected via a system bus. The computer-designed processor is used to provide computing and control capabilities. The memory of the real-time monitoring device for the electric shovel brake status includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the real-time monitoring device for the electric shovel brake status is used to store the corresponding data in this embodiment. The network interface of the real-time monitoring device for the electric shovel brake status is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the above method is implemented.

[0109] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the device for real-time monitoring of the brake status of an electric shovel to which the solution of the present invention is applied.

[0110] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions. When the instructions are executed on a computer, the computer executes the steps of the method for real-time monitoring of the electric shovel brake status.

[0111] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0112] If the integrated unit is implemented as a software functional unit and sold or used as a standalone product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product, stored in a storage medium, includes instructions for enabling a device for real-time monitoring of the brake status of an electric shovel (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0113] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for real-time monitoring of the brake status of an electric shovel, characterized in that: The method comprises: The hydraulic cylinder pressure pulsation, brake disc surface temperature gradient, brake pad thickness change rate, and motor feedback torque of the electric shovel brake are synchronously collected by multi-source sensors to obtain the brake multi-physical field coupling characteristic vector, including: using a pressure sensor to perform high-frequency collection and processing on the hydraulic cylinder pressure pulsation to obtain hydraulic cylinder pressure pulsation data; using an infrared temperature sensor to perform multi-point arrangement collection and processing on the brake disc surface temperature gradient to obtain brake disc surface temperature gradient data; using a laser displacement sensor to perform micron-level precision measurement and processing on the brake pad thickness change rate to obtain brake pad thickness change rate data; using a torque sensor to synchronously collect and process the motor feedback torque to obtain motor feedback torque data; and performing filtering preprocessing on the hydraulic cylinder pressure pulsation data, the brake disc surface temperature gradient data, the brake pad thickness change rate data, and the motor feedback torque data based on brake power spectrum density analysis to obtain the brake multi-physical field coupling characteristic vector; The heavy-load working condition enhanced sampling processing is performed on the multi-physical field coupling characteristic vector of the brake according to the braking power spectrum density to obtain a heavy-load working condition balanced data set, including: calculating and processing the instantaneous power of the brake according to the feedback torque and angular velocity of the motor to obtain the instantaneous power data of the brake; performing frequency domain conversion processing on the instantaneous power data of the brake using a fast Fourier transform algorithm to obtain a braking power spectrum density feature; performing hierarchical classification processing on the braking working condition based on the energy distribution of the braking power spectrum density feature to obtain three working condition levels: light-load braking, medium-load braking, and heavy-load braking; performing similarity calculation processing on the minority class samples of the heavy-load braking working condition through a k-nearest neighbor algorithm to obtain a power spectrum similarity weight factor; performing constrained synthesis generation processing on the heavy-load braking state samples according to the power spectrum similarity weight factor to obtain the heavy-load working condition balanced data set; The heavy-load condition balance dataset is subjected to feature extraction through a heavy-load condition adaptive multi-head attention mechanism to obtain a global brake state feature, including: dynamically adjusting the attention weight allocation strategy based on the comparison result between the real-time braking power and the preset power threshold to obtain a braking power adaptive weight factor; the heavy-load condition balance dataset is subjected to multi-physics field information processing through eight parallel attention heads to obtain hydraulic pressure attention features, thermodynamic attention features, mechanical wear attention features, and motor torque attention features; Performing attention weight calculation processing on each attention head based on the query matrix, key matrix, and value matrix to obtain a basic attention weight distribution; performing weighted modulation processing on the basic attention weight distribution by applying the brake power adaptive weight factor to obtain an adaptive attention weight distribution; performing weighted fusion processing on the hydraulic pressure attention feature, the thermodynamic attention feature, the mechanical wear attention feature, and the motor torque attention feature according to the adaptive attention weight distribution to obtain the brake global state feature; Performing multi-scale time series convolution processing on the heavy-load working condition balance data set according to the braking cycle characteristics to obtain brake time series evolution characteristics, including: performing short-cycle convolution kernel processing on the heavy-load working condition balance data set according to the single braking cycle characteristics of the electric shovel to obtain brake transient response characteristics; performing medium-cycle convolution kernel processing on the heavy-load working condition balance data set according to the continuous multiple braking cycle characteristics of the electric shovel to obtain brake periodic change characteristics; performing long-cycle convolution kernel processing on the heavy-load working condition balance data set according to the long-term working cycle characteristics of the electric shovel to obtain brake degradation trend characteristics; performing gate control unit control processing on the brake transient response characteristics, the brake periodic change characteristics, and the brake degradation trend characteristics by a braking state gating mechanism to obtain multi-scale feature fusion weights; performing weighted fusion processing on the brake transient response characteristics, the brake periodic change characteristics, and the brake degradation trend characteristics based on the multi-scale feature fusion weights to obtain the brake time series evolution characteristics; Based on the braking power weighting, the brake global state characteristics and the brake time-series evolution characteristics are fused and identified to obtain the real-time state result of the electric shovel brake, including: performing feature cascade processing on the brake global state characteristics and the brake time-series evolution characteristics to obtain a comprehensive feature vector; calculating and processing the feature fusion weight according to the ratio of the real-time braking power to the historical maximum braking power to obtain a braking power weighting coefficient; performing weighted fusion processing on the brake global state characteristics and the brake time-series evolution characteristics based on the braking power weighting coefficient to obtain a power-weighted fusion feature; performing nonlinear mapping processing on the power-weighted fusion feature using a three-layer fully connected neural network to obtain a state classification probability distribution; performing loss calculation processing on the state classification probability distribution using a focalloss loss function to obtain four state recognition results: normal state, slight wear, overheating failure, and severe wear; and performing real-time determination processing on the electric shovel brake operation state based on the four state recognition results to obtain the real-time state result of the electric shovel brake.

2. The method for real-time monitoring of the brake status of an electric shovel according to claim 1, characterized in that: The short-cycle convolution kernel processing is performed on the heavy-load working condition balance data set according to the single braking cycle characteristics of the electric shovel to obtain the brake transient response characteristics, including: A short-period convolution kernel of size 3×1 is used to perform a convolution operation with a step size of 1 on the heavy-load working condition balance data set to obtain the hydraulic pressure building process characteristics; Performing a convolution operation with a step size of 1 on the heavy load working condition balance data set using the 3×1 short-period convolution kernel to obtain a hydraulic pressure release process feature; Performing nonlinear transformation processing on the hydraulic pressure establishment process characteristics and the hydraulic pressure release process characteristics using a ReLU activation function to obtain post-activation transient characteristics; Performing data normalization processing on the activated transient features using a batch normalization algorithm to obtain normalized transient features; The normalized transient characteristics are subjected to time series slicing processing based on a single braking time window to obtain the brake transient response characteristics.

Images