Mining electric shovel load prediction and dynamic optimization method based on multi-modal data fusion driving

Abstract

The invention discloses a mining electric shovel load prediction and dynamic optimization method based on multi-modal data fusion driving, and belongs to the field of mining machinery intelligent control. The mining electric shovel load prediction and dynamic optimization method comprises the steps that according to multi-source heterogeneous data of at least three operation periods in an electric shovel historical operation database, full-dimension operation data covering four transmission mechanisms under typical working conditions is obtained; extracting electrical features, mechanical features and environmental features based on the obtained full-dimension operation data; constructing a multi-modal feature tensor; constructing a multi-modal LSTM prediction model, and carrying out training, verification and testing by adopting the obtained multi-modal feature tensor; the input layer of the multi-modal LSTM prediction model inputs a multi-modal feature tensor, and the output layer predicts four transmission mechanism load vectors in the future. Therefore, by integrating the multi-source heterogeneous data, the limitation of single data perception is broken through, and the feature expression ability of the complex working condition is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of mining electric shovel load prediction and dynamic optimization method based on multi-modal data fusion driving, belong to the intelligent control technical field of mine machinery. BACKGROUND

[0002] As the core equipment of open-pit mining, the operation performance of large mining electric shovel directly affects the efficiency and cost of mine production. However, the current electric shovel operation still highly depends on manual operation, facing multiple technical bottlenecks. The efficiency of manual operation is limited by the experience level and real-time judgment ability of the operator. In the scene of uneven distribution of ore and rock or complex geological conditions, the operator is difficult to accurately control the cutting depth and angle of the shovel, resulting in low full bucket rate and large fluctuation of single shovel loading cycle time, which not only reduces the effective loading capacity, but also increases the burden of subsequent sorting process due to the mixing of ore and waste rock; At the same time, the traditional operation mode lacks dynamic perception and prediction of load state, which is easy to cause shovel overload or underload problem, causing invalid consumption of energy, for example, frequent start and stop of electric shovel in overload condition will cause a large amount of invalid power consumption of hydraulic system, significantly increasing the operation energy consumption, and the randomness of manual operation makes the equipment in a long-term non-expected mechanical stress state, such as forcibly cutting into hard rock or improper turning operation, which will accelerate the wear of shovel tooth tip, hydraulic cylinder and transmission parts, and even cause early failure of key components. The delay of human operation response may also cause the equipment protection mechanism to fail to trigger in time, which poses a safety hazard.

[0003] Chinese patent CN116464119A discloses a kind of automatic operation control method of mining electric shovel considering digging mutation load, which mainly relies on visual detection or single current threshold to judge load mutation, when planning path by reconstructing three-dimensional topography of material pile with binocular camera, the visual detection accuracy significantly decreases (point cloud error > 15%) in complex environment such as dust and light change, and cannot reflect the dynamic changes of motor electrical parameters (such as harmonic distortion rate and instantaneous power fluctuation) in real time, resulting in load mutation misjudgment rate > 30%; It uses fixed trajectory planning or PID control, and the full bucket rate fluctuation range is 40%-65% in mutation conditions such as hard rock and frozen soil, while some research (such as BLISCO multidisciplinary collaborative optimization method) reduces the lifting force by 6% and the pushing pressure by 8.48% through structural design, but the optimization is concentrated in the hardware design stage, and cannot realize dynamic adjustment in the operation process; Traditional overload protection relies on physical sensor threshold triggering (such as shutdown after torque exceeds), with response delay > 300ms, which is easy to cause fatigue damage of gear box.

[0004] Chinese patent CN118707991A discloses a method for automatically controlling electric shovel digging trajectory based on TOF and ANN, which is a trajectory control method based on TOF camera, although it improves the long-distance detection accuracy, but does not fuse mechanical load spectrum data, resulting in lack of multi-dimensional state support for action decision.

[0005] Chinese patent CN118211498A discloses a dynamic digging force prediction method based on a convolutional recurrent neural network guided by dynamics, which realizes the application of a deep learning model that integrates physical prior in the load prediction of a mining shovel, but uses sensors to obtain data, and the data sources are mainly mechanical parameters, which are relatively single.

[0006] In addition, some mining machinery attempts to introduce an automatic control system, but the traditional method is mainly based on fixed threshold or simple feedback mechanism, which is difficult to adapt to complex working conditions such as sudden change of seam thickness and dynamic change of ore hardness, and the flexibility and adaptability of the control strategy are insufficient. In the prior art, there are problems of single data perception, rigid control strategy, lagging safety mechanism and insufficient scene adaptability. SUMMARY

[0007] The purpose of the present application is to provide a mining shovel load prediction and dynamic optimization method based on multi-modal data fusion driving, which integrates multi-source heterogeneous data to break through the limitations of single data perception and enhance the expression ability of complex working conditions. At the same time, a load joint prediction model of four transmission mechanisms (lifting, pushing, rotating and walking) is constructed to realize dynamic distribution and optimization of load, improve load prediction accuracy and working condition adaptability, reduce uneven load and power fluctuation of mechanisms, reduce equipment wear and energy consumption, and enhance operation safety. It has significant significance to improve the service life of equipment and reduce maintenance costs.

[0008] To achieve the above technical purpose, the present application will adopt the following technical solution:

[0009] A mining shovel load prediction and dynamic optimization method based on multi-modal data fusion driving, comprising the following steps:

[0010] According to the multi-source heterogeneous data of at least 3 operation cycles in the shovel historical operation database, full-dimensional operation data of the four transmission mechanisms under typical working conditions are obtained, including but not limited to: real-time monitoring data of electrical system, dynamic acquisition data of mechanical load spectrum and multi-dimensional perception data of working environment; the four transmission mechanisms are lifting mechanism, pushing mechanism, rotating mechanism and walking mechanism;

[0011] Extracting electrical features based on the obtained real-time monitoring data of electrical system;

[0012] Extracting mechanical features based on the obtained dynamic acquisition data of mechanical load spectrum;

[0013] Extracting environmental features based on the obtained multi-dimensional perception data of working environment;

[0014] The obtained electrical features, mechanical features and environmental features are time-scale aligned and then subjected to cross-modal correlation analysis to screen high-correlation features whose mutual information meets a preset value, and the screened high-correlation features are assembled to form a multi-modal feature tensor;

[0015] A multi-modal LSTM prediction model is constructed, and the obtained multi-modal feature tensor is trained, verified and tested; the multi-modal LSTM prediction model comprises an input layer, a feature extraction layer, an attention mechanism layer and an output layer, wherein:

[0016] The input layer is subjected to sliding processing within a time window W to obtain corresponding multi-modal time series features;

[0017] The feature extraction layer is constructed based on a bidirectional LSTM structure, receives the multi-modal time series features output by the input layer, and captures the long and short term time series dependency of the input multi-modal time series features and then outputs to the attention mechanism layer;

[0018] The attention mechanism layer comprises a time attention layer, a modal attention layer and an institution attention layer; the time attention layer is used to calculate the time step weight t , focusing on the key time series features of the input data; the modal attention layer is used to calculate the modal weight m , to adaptively fuse electrical features, mechanical features and environmental features; the institution attention layer is used to calculate the weight j of the four transmission mechanisms, highlighting the load change sensitive mechanism;

[0019] The output layer predicts the future four transmission mechanism load vectors, corresponding to the lifting, pushing, rotating and walking mechanism load prediction values.

[0020] Preferably, the typical working conditions include hard rock excavation working condition, slope walking working condition and full load rotating working condition.

[0021] Preferably, the real-time monitoring data of the electrical system includes: lifting current, lifting voltage and lifting power monitored in real time for the lifting motor, pushing current, pushing voltage and pushing power monitored in real time for the pushing motor, rotating current, rotating voltage and rotating power monitored in real time for the rotating motor, and walking current, walking voltage and walking power monitored in real time for the walking motor;

[0022] The mechanical load spectrum dynamic acquisition data includes: lifting torque T l , lifting speed v l and lifting acceleration a l monitored in real time for the lifting mechanism, pushing torque T p , pushing displacement x p and pushing pressure Fp , the slewing torque T of the slewing mechanism r , the slewing angle θ r , and the slewing angular velocity ω r , the walking torque T of the walking mechanism w , the walking speed v w , and the walking ground pressure p w ;

[0023] The multi-dimensional perception data of the working environment includes material characteristic data and environmental condition data; the material characteristic data includes hardness, density, and humidity; and the environmental condition data includes temperature, wind speed, and slope.

[0024] Preferably, the obtained full-dimensional operation data is preprocessed, and then electrical characteristics are extracted based on real-time monitoring data of the electrical system, mechanical characteristics are extracted based on dynamic acquisition data of the mechanical load spectrum, and environmental characteristics are extracted based on multi-dimensional perception data of the working environment.

[0025] Preferably, the obtained full-dimensional operation data is preprocessed, including missing value processing, abnormal value correction, and normalization processing.

[0026] When the obtained full-dimensional operation data is subjected to missing value processing, a time series interpolation method is used.

[0027] When the obtained full-dimensional operation data is subjected to abnormal value correction, an isolation forest algorithm is used for correction: an abnormal value is detected by a pre-set abnormal score threshold, and a sliding window median filter is used for correction when an abnormal value is detected.

[0028] The obtained full-dimensional operation data is subjected to normalization processing to retain the original distribution characteristics of each data.

[0029] Preferably, when the obtained full-dimensional operation data is subjected to missing value processing by using the time series interpolation method, specifically: linear interpolation is used for a segment with a missing value of <5%, bidirectional LSTM prediction is used to fill in a segment with a missing value of >5%, and the filling error is controlled within 3%.

[0030] Preferably, the electrical characteristics extracted based on real-time monitoring data of the electrical system include current effective value, current peak value factor, power spectrum density, harmonic content, and sample entropy, permutation entropy, and multi-scale entropy of the current signal.

[0031] The mechanical characteristics extracted based on dynamic acquisition data of the mechanical load spectrum include dynamic load characteristics and vibration characteristics; the dynamic load characteristics include load fluctuation coefficient, load mutation rate, and load distribution imbalance coefficient; and the vibration characteristics include wavelet packet energy feature vector, vibration intensity, and kurtosis index.

[0032] The environmental features extracted based on the multi-dimensional perception data of the working environment include material characteristic features and environmental condition features.

[0033] The material characteristic features are material hardness prediction values output by a material hardness prediction model, and the material hardness prediction model is based on a random forest regression algorithm and is expressed as:

[0034] H m = f (I RMS , P avg , T max , THD) + ∈

[0035] In the formula, H m represents the material hardness prediction value; f represents a mathematical model constructed based on a random forest regression algorithm; I RMS represents the current effective value; P avg represents the power average; T max represents the maximum temperature; THD represents the total harmonic distortion rate; and ∈ represents an error.

[0036] The environmental condition features include a temperature correction coefficient, a humidity correction coefficient, and a slope influence factor.

[0037] Preferably, a load balance distribution model is constructed to perform load balance distribution on the predicted load vectors of the four transmission mechanisms output by the multi-modal LSTM prediction model, and the load balance distribution model is expressed as:

[0038]

[0039] In the formula, Y t+k∣t is the predicted load vector output by the multi-modal LSTM prediction model, Y ref is a reference load vector, u t is a control input vector, is the current / voltage of any mechanism in the four transmission mechanisms, LDIC t is a load distribution imbalance coefficient, λ b1 is a control smoothness weight, and λ b2 is a control load balance weight.

[0040] Preferably, a power optimization model is constructed to perform power optimization on the predicted load vectors of the four transmission mechanisms output by the multi-modal LSTM prediction model, and the power optimization model is expressed as:

[0041]

[0042] Constraint condition:

[0043] In the formula, η i is the efficiency coefficient of the i-th mechanism, and P iP represents the power of the i th mechanism; λ is the power change rate penalty coefficient; F l P represents the force of the i th mechanism; F i,max P is the upper limit of the load of the i th mechanism, P grid P is the upper limit of the power supply of the power grid; P total P represents the total power threshold.

[0044] i=l, p, r, w, respectively represent the lifting, pushing, rotating and walking mechanisms.

[0045] Preferably, based on the dynamic load distribution algorithm, the load distribution ratio of the four driving mechanisms is dynamically adjusted:

[0046]

[0047] In the formula: r i P represents the load distribution ratio of the i th mechanism; i and j are indexes of the four driving mechanisms, i and j=l, p, r, w, respectively represent the lifting, pushing, rotating and walking mechanisms; H i P is the health index of the i th mechanism; k i P is the adjustment coefficient of the i th mechanism, which controls the coefficient of the load deviation term H j P is the health index of the j th mechanism; F j,ref P is the force value of the j th mechanism in the last sampling period at the current sampling time; F j,max P is the maximum allowable force value of the j th mechanism in the last sampling period at the current sampling time; F j P is the real-time force value of the j th mechanism at the current sampling time.

[0048] Based on the above technical purposes, compared with the prior art, the present application has the following advantages:

[0049] (1) The present application breaks through the limitations of traditional single data perception by fusing electrical, mechanical and environmental data; adopts bidirectional LSTM to fill in missing data, improves the algorithm to correct abnormal values and normalizes the processing, improves the data integrity and availability, and provides high-quality input for model training.

[0050] (2) The present application deeply mines the connotation of each modal data through time-frequency domain analysis, nonlinear features and vibration features, combines mutual information to screen high-correlation features, and constructs a multi-modal feature tensor; realizes cross-modal information interaction, enhances the feature expression ability for complex working conditions (such as hard rock and overload), and lays a foundation for accurate load prediction.

[0051] ​(3) The application is based on a three-layer Bi-LSTM and attention mechanism model, which captures long-term and short-term time sequence dependencies and dynamically focuses on key features, thereby improving the accuracy and working condition adaptability of load prediction; through load balancing distribution, power optimization and dynamic distribution algorithm, the uneven load of the mechanism and power fluctuation are reduced, the equipment wear and energy consumption are reduced, and the operation safety is enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 It is a process diagram for multi-modal information acquisition, feature extraction and modal interaction;

[0053] Figure 2 It is a load prediction and optimization model flowchart;

[0054] Figure 3 It is a comparison chart of the output power of the lifting motor for verifying the load optimization model of the application;

[0055] Figure 4 It is a comparison chart of the output power of the push motor for verifying the load optimization model of the application;

[0056] Figure 5 It is a comparison chart of the lifting torque for verifying the load optimization model of the application;

[0057] Figure 6 It is a comparison chart of the push torque for verifying the load optimization model of the application;

[0058] Figure 7 It is a comparison chart of the lifting force for verifying the load optimization model of the application;

[0059] Figure 8 It is a comparison chart of the push force for verifying the load optimization model of the application. DETAILED DESCRIPTION

[0060] 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. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Unless otherwise specifically stated, the relative arrangement, expressions, and values ​​of components and steps set forth in these embodiments do not limit the scope of the present invention. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0061] like Figure 1 , Figure 2 As shown, the method for load prediction and dynamic optimization of mining electric shovels based on multimodal data fusion driven by the present invention specifically includes the following steps:

[0062] I. Historical Data Acquisition and Preprocessing

[0063] Based on multi-source heterogeneous data from at least three operating cycles in the historical operation database of electric shovels, it is necessary to obtain full-dimensional operation data covering the four major transmission mechanisms (lifting, pushing, slewing, and traveling) under typical working conditions (hard rock excavation, slope traveling, and full-load slewing), including but not limited to: real-time monitoring data of the electrical system, dynamic acquisition data of the mechanical load spectrum, and multi-dimensional perception data of the working environment.

[0064] 1. Historical data types

[0065] (1) Real-time monitoring data of electrical system:

[0066] Hoisting motor: Real-time monitoring of the electrical data of the hoisting motor, including three-phase current (I lA ,I lB ,I lC (A) DC bus voltage (l) (V), power (P) l (kW), denoted as boost current, boost voltage, and boost power.

[0067] Push-press motor: Real-time monitoring of the electrical data of the push-press motor, including three-phase current (I pA ,I pB ,I pC (A) DC bus voltage (U) p(V), power (P p ), recorded as thrust current, thrust voltage and thrust power.

[0068] Rotary motor: real-time monitoring of electrical data of rotary motor, including three-phase current (I rA , I rB , I rC (A), DC bus voltage (U r (V), power (P r (kW), recorded as rotary current, rotary voltage and rotary power.

[0069] Walking motor: real-time monitoring of electrical data of walking motor, including three-phase current (I wA , I wB , I wC (A), DC bus voltage (U w (V), power (P w (kW), recorded as walking current, walking voltage and walking power.

[0070] (2) Dynamic data acquisition of mechanical load spectrum:

[0071] Lifting mechanism: real-time monitoring of dynamic data of mechanical load spectrum of lifting motor, including torque (T l (N·m), speed (v l (m / s), acceleration (a l (m / s 2 ), recorded as lifting torque T l , lifting speed v l and lifting acceleration a l .

[0072] Thrust mechanism: real-time monitoring of dynamic data of mechanical load spectrum of lifting motor, including torque (T p (N·m), displacement (x p (m), pressure (F p (N), recorded as thrust torque T p , thrust displacement x p and thrust pressure F p .

[0073] Rotary mechanism: real-time monitoring of dynamic data of mechanical load spectrum of lifting motor, including torque (T r (N·m), angle (θ r (rad), angular velocity (ω r (rad / s), recorded as rotary torque T r , rotary angle θ r and rotary angular velocity ω r .

[0074] Walking mechanism: Real-time monitoring of the mechanical load spectrum dynamic data of the lifting motor, including torque (T w ) (N·m), speed (v w ) (m / s), ground specific pressure (p w ) (Pa), denoted as walking torque T w , walking speed v w and walking ground specific pressure p w .

[0075] (3) Multidimensional perception data of working environment:

[0076] Material properties: hardness (Mohs) (Mohs), density (p) (kg / m 3 ), humidity (H) (%).

[0077] Environmental conditions: temperature (T) (℃), wind speed (v wind ) (m / s), slope (a) (°).

[0078] 2. Data preprocessing process

[0079] (1) Missing value processing:

[0080] Using time series interpolation method, linear interpolation is used for the segment with missing value <5%, and bidirectional LSTM prediction is used for the segment with missing value >5%:

[0081]

[0082] The filling error is controlled within 3%.

[0083] (2) Abnormal value correction:

[0084] Based on the improved isolated forest algorithm, set the abnormal score threshold to 0.85, and use sliding window median filtering correction for the detected abnormal values:

[0085] x′ t = median(x t-k:t+k )

[0086] Where the window size k = 5, and the correction rate is >99%.

[0087] (3) Standardization:

[0088] Z-Score normalization is used for parameters with different dimensions to retain the original distribution characteristics: Where μ and σ are the mean and standard deviation of historical data, respectively, and x t represents the original data of a parameter at time t; x' tZ-score normalized standardized data of the original data at the t-th moment, unitless.

[0089] II. Multi-modal feature extraction and fusion

[0090] 1. Electrical feature extraction

[0091] (1) Time-frequency domain features:

[0092] Current effective value (RMS):

[0093] I i represents the instantaneous current value (unit: A) at the i-th sampling moment, covering the three-phase currents of the lifting, pushing, rotating and walking motors (such as I lAi , I lBi , I lCi ), N represents the total number of current samples in a single calculation period, which is determined according to the sampling frequency. In this application, the sampling frequency is 10 Hz, and if the current effective value in 1 second is calculated, N = 10.

[0094] Current peak factor (CF):

[0095] I peak represents the maximum value of the three-phase current (such as I lA , I lB , I lC instantaneous value) of a specific motor (lifting motor, pushing motor, rotating motor or walking motor) of the mining shovel in a set calculation period.

[0096] Power spectral density (PSD): calculated by Welch method, resolution 0.1 Hz.

[0097] Harmonic content: extract 2-13 harmonic amplitudes and phase angles.

[0098] (2) Nonlinear features:

[0099] Sample entropy of current signal (Samp-En): reflects the complexity of the system.

[0100] Permutation entropy (Perm-En): quantifies signal randomness, Perm-En < 0.8 under hard rock working conditions.

[0101] Multiscale entropy (MSE): analyzes complexity at different time scales.

[0102] 2. Mechanical feature extraction

[0103] (1) Dynamic load features:

[0104] Load fluctuation coefficient (LFC): Reflects load stability.

[0105] σi(t) = F(t) - μi F σi(t) = F(t) - μi F μi = 1 / T ∑ F(t)dt

[0106] Load change rate (LCR): Change threshold 20 kN / s.

[0107] F(t) = F(t - Δt) + (F(t) - F(t - Δt)) / Δt t F(t) = F(t - Δt) + (F(t) - F(t - Δt)) / Δt t-1 F(t) = F(t - Δt) + (F(t) - F(t - Δt)) / Δt

[0108] Load distribution imbalance coefficient (LDIC):

[0109] where Fi is the load of the ith transmission mechanism, is the average load, i = 1, 2, 3, 4, representing the lifting, pushing, rotating, and walking mechanisms, respectively.

[0110] (2) Vibration characteristics:

[0111] Wavelet packet energy feature vector: db6 wavelet packet decomposition is performed on the vibration signal, and 32 sub-band energies are extracted.

[0112] Vibration intensity (VI): Evaluate the vibration level of the equipment.

[0113] a(t) = 1 / T ∑ a(t)dt

[0114] Kurtosis index Detect impact failure;

[0115] E[x4] = μ4

[0116] 3. Environmental feature extraction

[0117] (1) Material characteristic features:

[0118] Hardness prediction model of material based on historical data: H m = f(I RMS , P avg , T max , THD) + ε

[0119] H m represents the predicted value of material hardness; f represents a mathematical model constructed based on the random forest regression algorithm; I RMS represents the effective value of current; P avg represents the average power; T max represents the maximum temperature; THD is the total harmonic distortion rate, which represents the total harmonic distortion rate of the current signal of the motor, quantifies the deviation of the harmonic component in the current signal relative to the fundamental component, and reflects the nonlinear characteristics of the motor load; ε represents.

[0120] Using the random forest regression algorithm, the feature importance ranking is: the effective value of current I RMS (the weight is set to 0.35) > the average power P avg (the weight is set to 0.25) > the temperature T max (the weight is set to 0.2) > the total harmonic distortion rate THD (the weight is set to 0.2), and the model RMSE = 0.35 Mohs.

[0121] H m directly determines the load optimization strategy of the electric shovel: if H m > 6 (hard rock), the digging speed needs to be reduced, the loads of the four mechanisms are balanced, and overloading is avoided; if H m < 3 (soft rock), the operating efficiency can be improved, and the lifting / pushing force can be appropriately increased.

[0122] (2) Environmental condition features:

[0123] Temperature correction coefficient:

[0124] In the formula, T represents the real-time temperature of the operating environment of the mining electric shovel, specifically the air temperature around the operating area of the electric shovel, not the heating temperature of the device itself (such as the motor and the reduction gearbox), and the unit is Celsius (℃).

[0125] Humidity correction coefficient: k H = 1 + 0.015 (H-30);

[0126] In the formula, H represents the relative humidity of the operating area of the mining electric shovel, that is, the percentage of water vapor content in the air at the operating site to the saturated water vapor content of the air at the same temperature (unit: %), which reflects the degree of air humidity, not the humidity inside the electric shovel device or the humidity of the material itself.

[0127] Slope impact factor: k α = cos(a) + 0.2 sin(a).

[0128] In the formula: a represents the ground slope angle of the current working area of the electric shovel, that is, the included angle between the surface of the working site and the horizontal plane, which is used to quantify the terrain inclination degree when the electric shovel is working, rather than the inclination angle of the electric shovel body (the body inclination needs to be corrected by the leveling system, and a only reflects the external terrain slope).

[0129] 4. Feature fusion strategy

[0130] (1) Time scale alignment:

[0131] The data of different sampling rates are uniformly resampled to 10 Hz by using cubic spline interpolation, and the time synchronization error is <10 ms.

[0132] (2) Cross-modal correlation analysis:

[0133] The mutual information (Mutual Information, MI) of the electrical parameters and the mechanical load is calculated, and the high correlation features (MI>0.6) are screened.

[0134] Mutual information is a quantitative index of cross-modal correlation strength, and mutual information comes from information theory, which is used to measure the "information sharing degree" between two random variables (such as "electrical feature A" and "mechanical feature B"). The higher the mutual information value is, the stronger the correlation between the two variables is, and the change of one side can effectively reflect the change of the other side.

[0135] For two random variables X (such as a certain electrical feature, such as the RMS of the lifting motor current) and Y (such as a certain mechanical feature, such as the lifting torque), the calculation formula of the mutual information MI(X, Y) is:

[0136]

[0137] Parameter definition: P(x): the marginal probability distribution of variable X (such as the probability of the current RMS being 1200A); P(y): the marginal probability distribution of variable Y (such as the probability of the lifting torque being 15000N·m); P(x, y): the joint probability distribution of variables X and Y (such as the probability of the current RMS being 1200A and the lifting torque being 15000N·m appearing at the same time); the mutual information value range: MI(X, Y)≥0, when MI=0, X and Y are completely irrelevant; the larger the MI value is, the stronger the correlation is.

[0138] In the present application, for the lifting mechanism, the mutual information MI of the lifting motor current and the lifting torque is calculated, and the mutual information threshold for screening high correlation features is set to 0.82.

[0139] For the thrust mechanism, the mutual information MI between the thrust motor power and the thrust force was calculated, and the mutual information threshold for screening high correlation features was set to 0.78.

[0140] For the rotation mechanism, the mutual information MI between the rotation motor harmonic content and the rotation mechanism vibration was calculated, and the mutual information threshold for screening high correlation features was set to 0.65.

[0141] (3) Construct a multi-modal feature tensor:

[0142] Where X e (electricity), X m (mechanical), X e (environment) are 128-dimensional, 96-dimensional, and 64-dimensional feature matrices, respectively, T is the time step, and D=288 is the total feature dimension.

[0143] Three, four transmission mechanism load prediction and optimization model

[0144] 1. Multi-modal LSTM prediction model

[0145] (1) Model architecture:

[0146] Input layer: 288-dimensional multi-modal feature vector Time window W=600 (corresponding to 60 seconds of data).

[0147] Feature extraction layer: 3-layer bidirectional LSTM (Bi-LSTM) structure with hidden layer dimensions of 256, 192, and 128 to capture long and short-term temporal dependencies.

[0148] Attention mechanism layer:

[0149] Temporal attention: Calculate the weight α t of each time step to focus on key temporal features.

[0150] Modality attention: Calculate the weight β m of each modality to adaptively fuse electrical, mechanical, and environmental information.

[0151] Mechanism attention: Calculate the weight γ i of the four transmission mechanisms to highlight the load change sensitive mechanisms.

[0152] Output layer: Predict the four transmission mechanism load vector for the next 10 seconds: Where are the predicted values of the lifting, thrust, rotation, and walking mechanisms, respectively.

[0153] (2) Training strategy:

[0154] Loss function: Weighted hybrid loss function

[0155] where: w i is the weight of the mechanism, w i = 0.3 for the lifting / pushing mechanism, w i = 0.2 for the rotating / walking mechanism.

[0156] λ a1 = 0.6, λ a2 = 0.3, λ a3 = 0.1 to balance the sensitivity of different errors.

[0157] Optimizer: Adam optimizer with an initial learning rate of 0.0005 and cosine annealing learning rate scheduling.

[0158] 2. Load optimization strategy

[0159] (1) Load balancing distribution model:

[0160]

[0161] where: Y t+k∣t is the predicted load vector, Y ref is the reference load vector. k represents the prediction step index, which identifies the prediction node at the future time unit from the current time, is the key parameter connecting the current control input u t and the load optimization target Y at future time.

[0162] u t is the control input vector (motor current / voltage).

[0163] LDIC t is the load distribution imbalance coefficient.

[0164] λ b1 = 0.2, λ b2 = 0.5, which controls the smoothness and load balance weight.

[0165] (2) Power optimization model:

[0166] Power optimization model:

[0167] Constraint condition:

[0168] where: η i is the efficiency coefficient of the i-th mechanism.

[0169] λ = 0.1 is the power change rate penalty coefficient.

[0170] F i,maxPmax is the upper limit of the mechanism load grid Pgrid is the upper limit of the grid power supply.

[0171] (3) Dynamic load distribution algorithm:

[0172] Based on the predicted load and the health status of the equipment, the load distribution ratio of the four driving mechanisms is dynamically adjusted:

[0173]

[0174] Where: H i is the health index of the ith mechanism (0-1), evaluated by vibration, temperature, etc.

[0175] k i is the adjustment coefficient, k i = 0.3 for the lifting / pushing mechanism, k i = 0.2 for the rotating / walking mechanism.

[0176] Application example

[0177] I. Data preparation

[0178] Select the operating data of the electric shovel within one minute from the historical data as the analysis sample, covering the electrical parameters related to the lifting motor and the pushing motor (such as lifting current, pushing current, etc.), mechanical parameters (such as lifting torque, pushing force, etc.). The data involved in the analysis include lifting current, pushing current, lifting torque, pushing force, lifting motor output power, pushing motor output power, lifting force, pushing torque, etc.

[0179] II. Feature extraction

[0180] 1. Electrical feature extraction

[0181] Current effective value (RMS): For current signal I(t), its effective value calculation formula is:

[0182]

[0183] In the case of discrete data, taking the lifting current as an example, the lifting current effective value is calculated by:

[0184]

[0185] Where n is the sample number, I lift_i is the lifting current value at the ith time. After calculation, the lifting current effective value rms_lift_current = 1138.6077943817118A, and the pushing current effective value rms_push_current = 355.91162940201826A.

[0186] Current peak factor (CF): The peak factor is defined as the ratio of the current peak value I peak to the current effective value RMS, i.e.:

[0187]

[0188] For the lifting current, its current peak value peak_lift_current is expressed as:

[0189] peak_lift_current = max(I lift_i );

[0190] The calculated lifting current peak factor cf_lift_current = 1.8155856741895526;

[0191] For the pushing current, its current peak value peak_push_current is expressed as:

[0192] peak_push_current = max(I push_i ),

[0193] The pushing current peak factor cf_push_current = 1.6617355296697878.

[0194] Motor output power average: The average of the lifting motor output power and the pushing motor output power is directly obtained from the data.

[0195] Lifting motor output power average

[0196] Pushing motor output power average

[0197] where P lift_i and P push_i are the lifting motor and pushing motor output power at the i-th moment, respectively.

[0198] Based on the collected historical data, the following can be calculated according to the above formula:

[0199] Lifting motor output power average P lift_avg = 955.34 kW.

[0200] Pushing motor output power average P push_avg = 178.68 kW.

[0201] 2. Mechanical feature extraction

[0202] Load fluctuation coefficient (LFC): for torque or force signals, first calculate the mean μ and standard deviation σ, then calculate the ratio of the two, that is, the corresponding load fluctuation coefficient LFC can be obtained, which is specifically expressed as follows:

[0203]

[0204] Take the lifting torque as an example, its mean μ lift_torque is expressed as:

[0205]

[0206] Its standard deviation σ lift_torque is expressed as:

[0207]

[0208] In the above formula, T lift_i represents the lifting torque at the i-th moment;

[0209] Based on the collected historical data, the following formula can be calculated according to the above formula:

[0210] The mean μ of the lifting torque lift_torque = 14469.73215302186 N·m;

[0211] The standard deviation σ lift_torque = 10441.390498525048 N·m;

[0212] The load fluctuation coefficient lfc_lift_torque = 0.7216021960948646;

[0213] The mean μ of the push force push_force = -94.01814231093627 kN;

[0214] The standard deviation σ push_force = 437.8227174316655 kN;

[0215] The load fluctuation coefficient lfc_push_force = -4.656789707498162.

[0216] Three, load prediction

[0217] The above extracted electrical features (lifting current effective value, lifting current peak value factor, push pressure current effective value, push pressure current peak value factor) and mechanical features (lifting torque mean, lifting torque standard deviation, push force mean, push force standard deviation) are combined into a feature vector Input into the multi-modal LSTM prediction model to predict the load of the four transmission mechanisms in the future 10 seconds.

[0218] Assuming that the model has been trained and can be loaded through the function load_model, the prediction function is predict, and the predicted load of the four mechanisms in the next 10 seconds is [0.56965478, 0.66568179, 0.86584565, 0.1134091] (the values here are only examples and actually represent the load of each mechanism).

[0219] Four, load optimization decision

[0220] 1. Load balancing distribution

[0221] The preset mechanism health index, in this invention, the preset lifting mechanism health index H l = 0.8, the preset pushing mechanism health index H p = 0.85, the preset rotating mechanism health index H r = 0.9, and the preset walking mechanism health index H w = 0.88.

[0222] The total health index total_health = H l + H p + H r + H w = 0.8 + 0.85 + 0.9 + 0.88 = 3.43.

[0223] The load distribution ratio r i of each mechanism is calculated as follows:

[0224] Where: i = l, p, r, w respectively represent the lifting, pushing, rotating, and walking mechanisms.

[0225] The lifting mechanism load distribution ratio is calculated as follows:

[0226] The pushing mechanism load distribution ratio is calculated as follows:

[0227] The rotating mechanism load distribution ratio is calculated as follows:

[0228] The walking mechanism load distribution ratio is calculated as follows:

[0229] 2. Optimization distribution of power, force, and torque

[0230] Power distribution:

[0231] The preset total power threshold P total = 10000kW, and the calculation formula of the distribution power of each mechanism is as follows:

[0232] P​​​​​​​​​​​​​​​​​​​​​​​​​​​i = P total x r i .

[0233] Wherein: r i represents the load distribution ratio of each mechanism, i=l, p, r, w respectively represent the lifting, pushing, rotating and walking mechanisms.

[0234] The power of each mechanism before optimization is respectively: the lifting mechanism P l0 = 2000kW, the pushing mechanism P p0 = 2600kW, the rotating mechanism P r0 = 2700kW, and the walking mechanism P w0 = 2700kW.

[0235] The distributed power of the lifting mechanism P l = P total x r l = 10000 x 0.2332 = 2332kW.

[0236] The distributed power of the pushing mechanism P p = P total x r p = 10000 x 0.2478 = 2478kW.

[0237] The distributed power of the rotating mechanism P r = P total x r r = 10000 x 0.2624 = 2624kW.

[0238] The distributed power of the walking mechanism P w = P total x r w = 10000 x 0.2566 = 2566kW

[0239] The distribution of force:

[0240] First, calculate the total force before optimization total_force = F l0 + F p0 + F r0 + F w0 ;

[0241] Wherein: F l0 represents the lifting force before optimization analysis_data, and F l0 = 2126.60kN is obtained by collection; F p0 represents the pushing force before optimization analysis_data, and F p0 = -118.00kN is obtained by collection; F r0 represents the rotating force before optimization, and the rotating force F r0= 0 kN; F w0 represents walking force, and F w0 = 0 kN (actual should be obtained according to actual situation). Then the total force total_force before optimization is:

[0242] total_force = 2126.60 + (-118.00) + 0 + 0 = 2008.60 kN.

[0243] The distribution formula of the optimized force F i of each mechanism is:

[0244] F i = total_force x r i

[0245] In the formula, r i represents the load distribution ratio of each mechanism. Then:

[0246] The optimized force F l of the lifting mechanism is: l = 2008.60 x 0.2332 ≈ 468.40 kN.

[0247] The optimized force F p of the pushing mechanism is: p = 2008.60 x 0.2478 ≈ 497.73 kN.

[0248] The optimized force F r of the rotating mechanism is: r = 2008.60 x 0.2624 ≈ 527.06 kN.

[0249] The optimized force F w of the walking mechanism is: w = 2008.60 x 0.2566 ≈ 515.41 kN.

[0250] Distribution of torque:

[0251] First, calculate the total torque total_torque before optimization = T l0 + T p0 + T r0 + T w0 ;

[0252] In the formula, T l0 represents the lifting torque analysis_data before optimization, and T l0 = 15657.70 kN.m; T p0The analysis_data representing the optimized pushing torque is obtained by collecting T p0 = -1163.64 kN.m; assuming the slewing torque T r0 = 0 kN.m, the walking torque T w0 = 0 kN.m (which should be actually obtained according to the actual situation).

[0253] Then total_torque = 15657.70 + (-1163.64) + 0 + 0 = 14494.06 kN.m.

[0254] The distribution formula of the optimized torque T i of each mechanism is T i = total_torque x r i .

[0255] The optimized torque T l of the lifting mechanism is T l = 14494.06 x 0.2332 = 3380.01 kN.m;

[0256] The optimized torque T p of the pushing mechanism is T p = 14494.06 x 0.2478 = 3591.63 kN.m;

[0257] The optimized torque T r of the slewing mechanism is T r = 14494.06 x 0.2624 = 3803.24 kN.m;

[0258] The optimized torque T w of the walking mechanism is T w = 14494.06 x 0.2566 = 3719.18 kN.m.

[0259] In summary, through the above load optimization decision process, the optimized distribution of the power, force and torque of each mechanism of the electric shovel is completed, so that the load distribution of each mechanism is more reasonable, and the overall operation performance and equipment service life of the electric shovel are expected to be improved. The following is a comparison of the lifting motor output power, pushing motor output power, lifting torque, pushing torque, lifting force and pushing force of the lifting mechanism and the pushing mechanism of the mining electric shovel before and after optimization.

[0260] Table 1 Comparison of parameters before and after optimization

[0261]

[0262]

[0263] The above descriptions are only the preferred embodiment of the present application, and are not used to limit the present application. Any modification, equivalent replacement and improvement made in the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for load prediction and dynamic optimization of mining electric shovels based on multimodal data fusion, characterized in that, Includes the following steps: Based on multi-source heterogeneous data from at least three operating cycles in the electric shovel's historical operation database, we obtained comprehensive operational data covering the four major transmission mechanisms under typical working conditions, including but not limited to: real-time monitoring data of the electrical system, dynamic acquisition data of the mechanical load spectrum, and multi-dimensional perception data of the working environment; the four major transmission mechanisms are the lifting mechanism, the pushing mechanism, the slewing mechanism, and the traveling mechanism. Electrical features are extracted based on the obtained real-time monitoring data of the electrical system; Mechanical features are extracted based on the dynamically acquired mechanical load spectrum data. Environmental features are extracted based on the obtained multi-dimensional perception data of the working environment; After aligning the obtained electrical, mechanical, and environmental features by time scale, cross-modal correlation analysis is performed to screen out highly correlated features whose mutual information meets the preset value, and the screened highly correlated features are used to form a multimodal feature tensor. A multimodal LSTM prediction model is constructed, and the obtained multimodal feature tensors are used for training, validation, and testing. The multimodal LSTM prediction model includes an input layer, a feature extraction layer, an attention mechanism layer, and an output layer, wherein: The input layer performs a sliding process on the input multimodal feature tensor within a time window W to obtain the corresponding multimodal temporal features; The feature extraction layer, built on a bidirectional LSTM structure, receives the multimodal temporal features output by the input layer, captures the long and short-term temporal dependencies of the input multimodal temporal features, and outputs them to the attention mechanism layer. The attention mechanism layer includes a temporal attention layer, a modal attention layer, and a mechanism attention layer; the temporal attention layer is used to calculate the weight α for each time step. t It focuses on the key temporal features of the input data; the modal attention layer is used to calculate the modal weights β. m The system adaptively integrates electrical, mechanical, and environmental features; the mechanism attention layer is used to calculate the weights γ of the four major transmission mechanisms. j Emphasizing load-sensitive mechanisms; The output layer predicts the load vectors of the four major transmission mechanisms in the future, corresponding to the predicted load values ​​of the lifting, pushing, slewing, and traveling mechanisms.

2. The method for load prediction and dynamic optimization of mining electric shovels based on multimodal data fusion as described in claim 1, characterized in that, Typical operating conditions include hard rock excavation, slope walking, and full-load turning.

3. The method for load prediction and dynamic optimization of mining electric shovels based on multimodal data fusion as described in claim 1, characterized in that, Real-time monitoring data of the electrical system includes: real-time monitoring of lifting current, lifting voltage and lifting power for lifting motors; real-time monitoring of pushing current, pushing voltage and pushing power for pushing motors; real-time monitoring of slewing current, slewing voltage and slewing power for rotary motors; and real-time monitoring of traveling current, traveling voltage and traveling power for traveling motors. Dynamic acquisition data of mechanical load spectrum, including: lifting torque T monitored in real time for the lifting mechanism. l Speed ​​increase v l And increase acceleration α l For the real-time monitoring of the pushing torque T of the pushing mechanism p , Thrust displacement x p and the pushing pressure F p For the real-time monitoring of the rotation torque T of the slewing mechanism r Rotation angle θ r and rotational angular velocity ω r For the real-time monitoring of the traveling torque T of the traveling mechanism w Walking speed v w and walking grounding specific voltage p w ; Multidimensional sensing data of the working environment includes material characteristic data and environmental condition data; material characteristic data includes hardness, density and humidity; environmental condition data includes temperature, wind speed and slope.

4. The method for load prediction and dynamic optimization of mining electric shovels based on multimodal data fusion as described in claim 3, characterized in that, After preprocessing the obtained full-dimensional operational data, electrical features are extracted based on real-time monitoring data of the electrical system, mechanical features are extracted based on dynamic acquisition data of mechanical load spectrum, and environmental features are extracted based on multi-dimensional perception data of the working environment.

5. The method for load prediction and dynamic optimization of mining electric shovels based on multimodal data fusion as described in claim 4, characterized in that, The obtained full-dimensional operational data is preprocessed, including missing value handling, outlier correction, and normalization. When handling missing values ​​in the obtained full-dimensional operational data, time series interpolation is used; When correcting outliers in the obtained full-dimensional operational data, the Isolation Forest algorithm is used: an outlier detection is performed by setting a preset outlier score threshold, and when an outlier is detected, a sliding window midpoint filter is used for correction. The obtained full-dimensional operational data is normalized to preserve the original distribution characteristics of each data point.

6. The method for load prediction and dynamic optimization of mining electric shovels based on multimodal data fusion as described in claim 5, characterized in that, When using time series interpolation to handle missing values ​​in the obtained full-dimensional running data, the specific steps are as follows: linear interpolation is used for segments with missing values ​​<5%, and bidirectional LSTM prediction is used to fill segments with missing values ​​>5%, with the filling error controlled within 3%.

7. The method for load prediction and dynamic optimization of mining electric shovels based on multimodal data fusion as described in claim 5, characterized in that, Electrical features extracted from real-time monitoring data of electrical systems include RMS current, current peak factor, power spectral density, harmonic content, and sample entropy, permutation entropy, and multi-scale entropy of current signals. Mechanical features extracted from dynamic acquisition data of mechanical load spectrum, including dynamic load features and vibration features; Dynamic load characteristics include load fluctuation coefficient, load abrupt change rate, and load distribution imbalance coefficient; Vibration characteristics include wavelet packet energy feature vectors, vibration intensity, and kurtosis index; Environmental features extracted from multi-dimensional sensing data of the work environment, including material characteristics and environmental condition characteristics; in: The material characteristic feature is the predicted material hardness value output by the material hardness prediction model. The material hardness prediction model is based on the random forest regression algorithm and is expressed as follows: H m =f(I RMS ,P avg ,T max ,THD)+∈ In the formula: H m represents the predicted value of material hardness; f represents the mathematical model constructed based on the random forest regression algorithm; I RMS Indicates the effective value of the current; P avg T represents the average power; max Indicates the maximum temperature; THD represents the total harmonic distortion; ∈ represents... Environmental conditions include temperature correction factor, humidity correction factor, and slope influence factor.

8. The method for load prediction and dynamic optimization of mining electric shovels based on multimodal data fusion as described in claim 1, characterized in that, A load balancing distribution model is constructed to perform load balancing distribution on the predicted load vectors of the four transmission mechanisms output by the multimodal LSTM prediction model; the load balancing distribution model is expressed as: Where: Y t+k∣t The predicted load vector output by the multimodal LSTM prediction model, Y ref u is the reference load vector. t The control input vector is the current / voltage of any one of the four major transmission mechanisms; LDIC t λ is the load distribution imbalance coefficient. b1 To control the smoothness weight; λ b2 To control the load balance weight.

9. The method for load prediction and dynamic optimization of mining electric shovels based on multimodal data fusion as described in claim 1, characterized in that, A power optimization model is constructed to optimize the power of the predicted load vectors of the four transmission mechanisms output by the multimodal LSTM prediction model; the power optimization model is expressed as: Constraints: Where: η i P represents the efficiency coefficient of the i-th mechanism; i Let represent the power of the i-th mechanism; λ is the power change rate penalty coefficient. F l F represents the force of the i-th mechanism. i,max P is the upper limit of the load of the i-th mechanism. grid The upper limit of power supply to the grid; P total This represents the total power threshold. i = l, p, r, w, representing the lifting, pushing, rotating, and traveling mechanisms, respectively.

10. The method for load prediction and dynamic optimization of mining electric shovels based on multimodal data fusion as described in claim 1, characterized in that, Based on a dynamic load distribution algorithm, the load distribution ratio of the four major transmission mechanisms is dynamically adjusted: In the formula: r i This indicates the load distribution ratio of the i-th mechanism; i and j are the indices of the four major transmission mechanisms, i, j = l, p, r, w, representing the lifting, pushing, slewing, and traveling mechanisms respectively; H i Let k be the health index of the i-th institution; i The adjustment coefficient for the i-th mechanism controls the load deviation term. coefficient; H j Let F be the health index of the j-th institution; j,ref F represents the force value of the j-th mechanism in the previous sampling period at the current sampling time. j,max F represents the maximum permissible force value of the j-th mechanism in the previous sampling period at the current sampling time. j Let be the real-time force value of the j-th mechanism at the current sampling time.

Images