Hydraulic pump condition monitoring system based on multimodal information fusion and digital twin

Abstract

This invention discloses a hydraulic pump condition monitoring system based on multimodal information fusion and digital twin, specifically relating to the field of hydraulic equipment condition monitoring technology. The system includes a hydraulic pump multimodal data acquisition module, a hydraulic pump multimodal feature engineering module, a hydraulic pump condition hierarchical fusion diagnosis module, a hydraulic pump health status and risk assessment module, and a digital twin and decision support module. The hydraulic pump multimodal data acquisition module synchronously collects key operating parameters of the hydraulic pump through a distributed sensor array and obtains a hydraulic pump multimodal parameter set after preprocessing. The invention utilizes the hydraulic pump condition hierarchical fusion diagnosis module, combining the advantages of algorithms such as RF, CNN, and LSTM+Attention, to construct cavitation diagnosis sub-models, bearing wear diagnosis sub-models, and internal leakage diagnosis sub-models, enabling diagnosis for different fault types and achieving automation and intelligence from feature extraction to fusion diagnosis.

Description

Technical Field

[0001] This invention relates to the field of hydraulic equipment condition monitoring technology, specifically a hydraulic pump condition monitoring system based on multimodal information fusion and digital twin. Background Technology

[0002] In modern industry, hydraulic pumps, as key power transmission and control components, are widely used in many industries such as engineering machinery, metallurgy, mining, and aerospace. The stability and reliability of their operation are directly related to the efficiency and safety of the entire production system.

[0003] Traditional hydraulic pump condition monitoring methods include single-parameter monitoring, periodic inspection and maintenance, and simple rule-based diagnostic systems. Single-parameter monitoring involves setting fixed thresholds and triggering an alarm when the monitored parameter exceeds the normal range. Periodic inspection and maintenance involve manual inspection of the hydraulic pump, using observation, listening, and touch to determine if there are any abnormalities. Simple rule-based diagnostic methods diagnose faults based on preset fault modes and corresponding parameter variation ranges. These methods achieve hydraulic pump condition monitoring.

[0004] While traditional hydraulic pump condition monitoring methods can monitor the condition of hydraulic pumps, they also have limitations. For example, single-parameter monitoring methods only cover a single fault mode, making it difficult to comprehensively and accurately diagnose the fault state of the hydraulic pump. Regular inspections and maintenance rely on post-event analysis or regular inspections, making it difficult to detect problems in the early stages of a fault and carrying the risk of missed detections and misdiagnoses. Diagnostic methods based on simple rules often lack quantitative indicators, only providing qualitative judgments of normal or abnormal, which is difficult to meet the refined management requirements of modern industrial production for equipment maintenance. Moreover, they lack the ability to deeply mine and intelligently analyze data, making it difficult to predict the future operating state and remaining service life of hydraulic pumps. Therefore, developing a system that can monitor the condition of hydraulic pumps in real time, accurately, and comprehensively and predict their remaining service life is of great significance. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a hydraulic pump condition monitoring system based on multimodal information fusion and digital twin, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a hydraulic pump condition monitoring system based on multimodal information fusion and digital twin, comprising:

[0007] Hydraulic pump multimodal data acquisition module: Synchronously acquires key operating parameters of the hydraulic pump through a distributed sensor array, and obtains the hydraulic pump multimodal parameter set after preprocessing;

[0008] Hydraulic pump multimodal feature engineering module: Extracts features from the multimodal parameter set of hydraulic pump to obtain a multimodal feature dataset of hydraulic pump, including basic parameter features, vibration signal features, and motor and flow characteristics;

[0009] Hydraulic pump status hierarchical fusion diagnosis module: includes a parallel expert model diagnosis unit and a fusion diagnosis unit. Based on the hydraulic pump multimodal feature dataset, it first performs parallel expert model diagnosis, then performs fusion diagnosis, obtains hydraulic pump status diagnosis results, and transmits them to the hydraulic pump health status and risk assessment module.

[0010] Hydraulic pump health status and risk assessment module: Based on the hydraulic pump status diagnosis results, the hydraulic pump health index is obtained through mapping rules, the hydraulic pump risk index is obtained based on the hydraulic pump health index, and the remaining useful life is predicted to obtain the hydraulic pump health status and risk assessment results.

[0011] Digital Twin and Decision Support Module: Based on the hydraulic pump status diagnosis results, different maintenance strategy parameters are input, and the built-in digital twin model of the hydraulic pump is used to virtually debug and optimize the maintenance strategy parameters to obtain the optimal maintenance suggestions for human-computer interaction at the management end.

[0012] The technical effects and advantages of this invention are as follows:

[0013] 1. This invention comprehensively collects the mechanical, hydraulic, thermodynamic, and electrical signals of a hydraulic pump through a multi-modal data acquisition module; at the same time, it extracts features from the multi-modal parameter set of the hydraulic pump through a multi-modal feature engineering module, providing a data foundation for accurate diagnosis, improving the accuracy of fault diagnosis, and reducing the possibility of misdiagnosis and missed diagnosis.

[0014] 2. This invention utilizes a hydraulic pump state-layered fusion diagnostic module, combining the advantages of algorithms such as RF, CNN, and LSTM+Attention, to construct cavitation diagnostic sub-models, bearing wear diagnostic sub-models, and internal leakage diagnostic sub-models, enabling diagnosis of different fault types and achieving automation and intelligence from feature extraction to fusion diagnosis.

[0015] 3. This invention introduces the health index HI and risk index RI through the hydraulic pump health status and risk assessment module, and realizes the prediction of remaining useful life RUL, so that maintenance is changed from "after the fact" to "before the fact", potential faults can be identified in advance and corresponding maintenance measures can be taken. The system can effectively avoid unplanned downtime and improve the reliability and efficiency of the production line. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0017] Figure 2This is a schematic diagram of the hydraulic pump status layered fusion diagnostic module of the present invention. Detailed Implementation

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

[0019] Please see Figure 1 As shown, the present invention provides a hydraulic pump condition monitoring system based on multimodal information fusion and digital twin, including a hydraulic pump multimodal data acquisition module, a hydraulic pump multimodal feature engineering module, a hydraulic pump condition hierarchical fusion diagnosis module, a hydraulic pump health status and risk assessment module, and a digital twin and decision support module.

[0020] The hydraulic pump multimodal data acquisition module is connected to the hydraulic pump multimodal feature engineering module, the hydraulic pump state hierarchical fusion diagnosis module is connected to the hydraulic pump multimodal feature engineering module and the hydraulic pump health status and risk assessment module respectively, and the digital twin and decision support module is connected to the hydraulic pump health status and risk assessment module.

[0021] Hydraulic pump multimodal data acquisition module: Synchronously acquires key operating parameters of the hydraulic pump through a distributed sensor array, and obtains the hydraulic pump multimodal parameter set after preprocessing;

[0022] The key operating parameters of the hydraulic pump include inlet vacuum, outlet pressure, three-dimensional vibration signal, flow rate, temperature, motor power, and motor current; the X / Y / Z axis vibration acceleration is collected based on the vibration signal.

[0023] This embodiment requires specific explanation of the following: Inlet vacuum: A vacuum pressure sensor is used with a sampling frequency of 1kHz, covering an absolute pressure range of 0-1 atm; Outlet pressure: A high-frequency dynamic pressure sensor is used with a sampling frequency of 10kHz, supporting FFT spectrum analysis; Vibration signal: A triaxial accelerometer is used with a sampling frequency of 50kHz to collect three-dimensional vibration data; Flow rate: A turbine or ultrasonic flow meter is used with a sampling frequency of 100Hz, with a measurement accuracy of ≤±1%; Temperature: A PT100 temperature sensor is used with a sampling frequency of 1Hz, measuring a range of -50℃ to 200℃; Motor power: A power sensor is used with a sampling frequency of 100Hz to collect active and reactive power; Motor current: A current sensor is used with a sampling frequency of 1kHz to collect stator current signals for spectrum analysis.

[0024] The preprocessing includes denoising the operating parameters using denoising techniques (such as wavelet denoising) and aligning the operating parameters using timestamp interpolation methods (such as linear interpolation, spline interpolation, etc.) to obtain a multimodal parameter set for the hydraulic pump.

[0025] This embodiment specifically explains the timestamp-based interpolation data alignment method, which ensures that multimodal data from different sensors remain consistent in the time dimension. This guarantees that the features extracted by subsequent feature engineering and the data used by the hierarchical fusion diagnostic model are all based on the same time reference, thereby improving the accuracy and reliability of fault diagnosis.

[0026] Hydraulic pump multimodal feature engineering module: Extracts features from the multimodal parameter set of hydraulic pump to obtain a multimodal feature dataset of hydraulic pump, including basic parameter features, vibration signal features, and motor and flow characteristics;

[0027] The basic parameter characteristics are based on the inlet vacuum degree P. in and export pressure P out The ratio of the inlet vacuum pressure difference R is calculated. p Within the set sliding window, the maximum value P of the outlet pressure is collected respectively. out，max Minimum value P out，min and mean P out，avg Calculate the outlet pressure deviation rate DR p DR p =(P out，max -P out，min ) / P out，avg ;

[0028] The vibration signal characteristics are as follows: Based on the vibration signal, three-dimensional vibration signals are collected, including three-dimensional vibration signals of the hydraulic pump and bearings; Fourier transform is performed on the vibration signals to obtain the power spectral density S(f) of each dimension of the hydraulic pump, and the 500-1000Hz broadband power P of each dimension is obtained. 500-1000 , The vibration-averaged broadband power P is obtained by averaging the three-dimensional broadband power. 500-1000 avg For each dimension of the bearing vibration signal, calculate the fault characteristic frequency f of the outer ring, inner ring, and ball bearing respectively. o f i and f b , Where n is the motor speed, z is the number of balls, d is the ball diameter, D is the pitch circle diameter, and θ is the contact angle. , In most cases, when referring to the contact angle, it usually means the contact angle between the inner ring and the ball, because the inner ring usually rotates with the shaft, while the outer ring is fixed in the bearing housing; however, in specially designed bearings, the contact angle may also refer to the angle between the outer ring and the ball.

[0029] In this embodiment, it should be specifically explained that the power spectral density S(f) of the vibration signal can be obtained using the existing formula of Welch's power spectral density. For each dimension of the vibration acceleration signal a(t), the DC offset of the signal light is removed, and the long time-domain signal a(t) is divided into n1 shorter, overlapping segments. Each segment is multiplied by a window function w(n) (e.g., a Hanning window) to obtain the windowed signal x of each segment. k (n), and perform a Fast Fourier Transform to obtain its Discrete Fourier Transform X. k (f), calculate the signal power spectrum of this segment, and average the signal power spectra of K segments to obtain the vibration signal power spectral density S(f) for each dimension.

[0030] The motor and flow characteristics are based on the measured power P. me Measured flow rate Q (in L / min), outlet pressure P out (Unit: bar) to obtain theoretical power P th (Unit is kW, bar × L / min = kW), P th =Q×P out / 600η, where η is the theoretical volumetric efficiency of the hydraulic pump (e.g., 0.95), based on the measured power P. me With theoretical power P th The difference is used to obtain the power efficiency deviation ΔP; based on the rated flow rate Q of the hydraulic pump. ra Calculate the flow attenuation rate DR Q DR Q =(QQ ra ) / Q ra ;

[0031] Please see Figure 2 As shown, the hydraulic pump status hierarchical fusion diagnosis module includes a parallel expert model diagnosis unit and a fusion diagnosis unit. Based on the hydraulic pump multimodal feature dataset, it first performs parallel expert model diagnosis, then performs fusion diagnosis, obtains the hydraulic pump status diagnosis results, and transmits them to the hydraulic pump health status and risk assessment module.

[0032] The parallel expert model diagnostic unit includes a cavitation diagnostic sub-model, a bearing wear diagnostic sub-model, and an internal leakage diagnostic sub-model.

[0033] The cavitation diagnostic sub-model is based on the multimodal feature dataset of a hydraulic pump, and the vibration average broadband power P is... 500-1000 avgImported vacuum degree P in and export pressure deviation rate DR p The multi-source feature vectors formed are used as the input feature vector X of the pre-defined random forest (RF) model. cav X cav =[P 500-1000 avg P in DR p Suppose that the random forest consists of K decision trees {h} k (x)} k=1 K For the input feature vector X cav Each decision tree outputs the probability p of cavitation occurring. k，cav p k，cav =h k (X cav Next, the cavitation confidence level C is obtained by averaging the outputs of the K decision trees. cav , Finally, based on the cavitation confidence level, the cavitation severity level L is classified. cav , Cavitation refers to the vaporization of a liquid into bubbles during the operation of a hydraulic pump, caused by a local pressure drop below the liquid's vaporization pressure. It is typically identified by monitoring vibration and pressure change parameters. The output of the cavitation diagnostic sub-model includes the cavitation confidence level C. cav Cavitation severity level L cav ;

[0034] The bearing wear diagnostic sub-model includes:

[0035] A1: Constructing Input Features: Based on the multimodal feature dataset of hydraulic pumps, extract the three-dimensional vibration time-series data X of bearings. vib ∈R 3×T T represents the time step, 3 represents the three-dimensional vibration direction, and the characteristic frequencies f of the outer ring, inner ring, and ball bearing faults are extracted. o f i f b Based on the power spectral density S(f) of the vibration signal, the average power values ​​Sf in the three-dimensional vibration directions of the outer ring, inner ring, and balls at the characteristic frequency of bearing failure are obtained respectively. o，avg (f0), S i，avg (f i ), S b，avg (f b ), construct the input feature vector X of the pre-defined 1D convolutional neural network (1D-CNN) in X in =[X vib ;S o，avg (f0), S i，avg (fi ), S b，avg (f b )];

[0036] A2: Local temporal feature extraction: First, perform convolutional layer 1, y1=ReLU(X in *w1+b1), where * represents the convolution operation, w1 is the convolution kernel, and w1∈R. k1×c1 k1=32 is the kernel size, c1=16 is the number of output channels, b1 is the bias term, and ReLU activation function is used to extract local features such as impact and periodic pulses from vibration signals, such as periodic vibrations caused by bearing wear; then pooling layer 1 is performed, y p1 =MaxPooling(y1, s1=2), where the pooling window stride s1 is 2, and the output is y1. p1 ∈R c1×T / 2 MaxPooling is a pooling operation;

[0037] A3: Feature frequency fusion: First, perform convolutional layer 2, y2=ReLU(y p1 *w2+b2), where w2 is the convolution kernel, and w2∈R k2×c2 k2=16, c2=32, b2 is a bias term, which further integrates local temporal features with characteristic frequency power to enhance fault frequency correlation features; then pooling layer 2 is performed, y p2 =MaxPooling(y2, s2=2), outputs y p2 ∈R c1×T / 2 Flatten the pooling output into a vector y. flat ∈R N N is the dimension after flattening, resulting in the fully connected layer y. fc1 =ReLu(W3×y flat +b3), W3 is a fully connected layer y fc1 The weight matrix is ​​given by b3, where b3 is the corresponding bias term; the probability distribution P of the three types of faults is output through the Softmax function, P = Softmax(W4 × y fc1 +b4), W3 is the weight matrix of the output layer, and b4 is the corresponding bias term;

[0038] A4: Output results: including the inner ring failure probability P in outer ring failure probability P out and the probability of ball bearing failure P ball and bearing wear confidence level C bear , β is the weight (e.g., 0.6), C bear ∈[0, 1], the higher the value, the more reliable the diagnosis. If C bearIf the value is greater than or equal to the corresponding threshold (e.g., 0.6), it is determined to be the corresponding fault type, and the item with the highest probability is selected; otherwise, it is determined to be a bearingless fault.

[0039] The internal leakage diagnostic sub-model: First, a multivariate time series X is constructed. le The time series length is t, X le =[Q(t), T(t), P out (t), P me (t), ΔP(t), DR Q [(t)], Q(t), T(t), P out (t), P me (t), ΔP(t) and DR Q (t) represents the measured flow rate, temperature, outlet pressure, power, power efficiency deviation, and flow rate decay rate, respectively; then, the multivariate time series X is processed through an LSTM layer (Long Short-Term Memory network). le Processing yields the hidden state sequence {h} t} t=1 tt h t Let be the hidden state at time t; then, the attention weight 'a' is calculated using an attention mechanism. t , e t =W h ×h t +b h W h and b h These are the weight matrix and bias term of the hidden state sequence, respectively. t Let h be the attention weights of the hidden state at time t; then, the hidden states are weighted and fused to obtain the fused feature h. att , Next, feature h is fused. att The input is fed into a fully connected layer and passes through the Sigmoid function (Sigmoid(W)). fc ×h att +b fc The internal leakage rate R is obtained. le W fc and b fc These are the weight matrix and bias term of the fully connected layer, respectively; finally, the prediction error (MAE) and feature consistency (CV) of the LSTM+Attention model are combined to obtain the inner leakage confidence C. le C le =(1-MAE)×γ+(1-CV)×(1-γ), where MAE is the model-predicted leakage rate R. le Compared with the true value R le，0 The mean absolute error, CV, is the power efficiency deviation ΔP and the flow rate attenuation rate DR.Q The coefficient of variation (CV) is the ratio of the standard deviation (σ(ΔP), σ(DRQ)) of the combination of two variables to the mean (μ(ΔP), μ(DRQ)) of the combination. In internal leakage diagnosis, ΔP (power efficiency deviation) and DRQ (flow attenuation rate) are physically strongly correlated features. By measuring the synchronicity of changes in the two variables, the physical interpretability of the confidence level of internal leakage diagnosis is enhanced. The output results of the internal leakage diagnosis sub-model are obtained, including the internal leakage rate R. le and internal leakage confidence level C le ;

[0040] In this embodiment, it should be specifically noted that LSTM is an existing technology, and its hidden state is updated through a forget gate, an input gate, a cell state gate, and an output gate to obtain the hidden state sequence {h}. t} t=1 tt .

[0041] The fusion diagnostic unit includes:

[0042] B1: Define the identification framework: Based on the DS evidence theory fusion, define the identification framework Θ, Θ={normal, cavitation, bearing wear, internal leakage, composite fault}, composite fault is the situation where two or more of the above single faults exist at the same time, bearing wear includes inner ring fault probability, outer ring fault probability and ball fault.

[0043] B2: Basic probability allocation: Based on the output of the cavitation diagnosis sub-model, including the cavitation confidence level C. cav Cavitation severity level L cav The basic probability allocation function m1(A) of the cavitation diagnostic sub-model is constructed through function f1, satisfying m1(A1)+m1(Θ)=1, where m1(A1) and m1(Θ) are the basic probability allocation results, respectively. The basic probability allocation function m2(A) for constructing the bearing wear diagnosis sub-model using function g1 and the basic probability allocation function m3(A) for constructing the internal leakage diagnosis sub-model using function h1 are logically the same as the basic probability allocation function m1(A) for the cavitation diagnosis sub-model. P max Let P be the probability of failure in the inner ring. in outer ring failure probability P out and the probability of ball bearing failure P ball The maximum value in the range, bearing wear confidence level C bear , Internal leakage rate R le and internal leakage confidence level C le In m2(A), A = {bearing wear}, and in m3(A), A = {internal leakage}.

[0044] B3: Evidence Combination: The basic probability assignment m(A) after combining the three sub-models is: The numerator represents the sum of the products of the basic probabilities of the corresponding subsets of each submodel when the intersection of the subsets A1, A2 and A3 from the basic probability assignments of each submodel is A. The denominator represents the fact that the intersection of the subsets is empty, and the conflict measure is subtracted from 1 for normalization.

[0045] B4: Hydraulic pump condition diagnosis result: The comprehensive health status HS of the hydraulic pump is obtained based on the combined basic probability allocation m(A), HS=argmax A∈Θ m(A), for example, if m{cavitation} is the largest among all m(A), then the overall health status is judged as cavitation failure;

[0046] Hydraulic pump health status and risk assessment module: Based on the hydraulic pump status diagnosis results, the hydraulic pump health index is obtained through mapping rules, the hydraulic pump risk index is obtained based on the hydraulic pump health index, and the remaining useful life is predicted to obtain the hydraulic pump health status and risk assessment results.

[0047] The hydraulic pump health index is based on the hydraulic pump condition diagnosis result HS, multiplied by 100 and mapped to a score of 1-100. This includes the normal hydraulic pump health index (HI) score range, the HI score range for cavitation, the HI score range for bearing wear, the HI score range for internal leakage, and the HI score range for combined faults. A volumetric efficiency attenuation coefficient kη is introduced to correct the HI, resulting in the corrected HI. th HI th =HI×kη, where kη is the current efficiency η of the hydraulic pump. me The ratio of the theoretical efficiency η to the current efficiency of the hydraulic pump, which is the ratio of the output power to the input power.

[0048] The specific HI scores in this embodiment need to be explained in detail. For example, the normal HI score range is 95-100, the HI score range for cavitation is 80-95 for level 1 and 60-80 for level 2, the HI score range for internal leakage is 50-70, and the HI score range for complex faults is <50.

[0049] The hydraulic pump risk index is calculated as follows: a predefined fault severity weight S (e.g., for minor leakage, S=1.0, obtained by analyzing historical fault cases and assigning fixed weight values ​​to different fault types) is used to calculate the risk index RI, RI=(100-HI)×S / 10; then, risk levels are classified: if RI belongs to the preset high risk level (e.g., 7.5), an immediate shutdown maintenance recommendation is triggered; if RI belongs to the medium risk level (e.g., 5<RI≤7.5), a planned maintenance recommendation is triggered; if RI belongs to the low risk level (e.g., RI≤5), a continuous monitoring recommendation is triggered.

[0050] The remaining useful life prediction: A pre-trained LSTM model is used to predict the next n2 time steps, with the HI sequence [HI] of the most recent m1 time steps as input. TT-m1+1 , ..., HI TT ], TT represents the current time, and the next time HI is predicted. TT+1 Then [HI] TT-m1+2 , ..., HI TT HI TT As a new input, predict HI TT+2 By analogy, the predicted HI decay curve result HI(TT1) is obtained, where TT1 > TT; a time threshold HI is set. fail (For example, 20), solve for HI(TT1)≤HI fail The minimum time t fail Then the remaining useful lifetime is RUL=t fail -t me , t me The current moment;

[0051] Digital Twin and Decision Support Module: Based on the hydraulic pump status diagnosis results, different maintenance strategy parameters are input, and the maintenance strategy parameters are virtually debugged and optimized using the built-in digital twin model of the hydraulic pump to obtain the optimal maintenance suggestions for human-computer interaction at the management end, including the following steps;

[0052] C1: Based on the hydraulic pump condition diagnosis results, input different maintenance strategy parameters θ into the built-in digital twin model of the hydraulic pump, θ=[θ1, θ2, ..., θ...]. M ], θ M The parameters for the Mth maintenance strategy are (e.g., hydraulic pump speed, system set pressure, etc.); then, based on the remaining useful life (RUL), the RUL after implementing the maintenance strategy θ is obtained. θ RUL θ =RUL0+ΔRUL(θ), where ΔRUL(θ) is the lifetime extension time;

[0053] C2: Strategy Benefit Calculation: Benefit value B(θ) is the output value RUL brought about by the lifespan extension. θ ×C pr Total maintenance cost C ma The ratio of (θ) is obtained, C pr The output value per unit time is calculated based on the economic value produced by the hydraulic pump, with the unit being yuan / s. The total maintenance cost is also in yuan, referring to the cost of all maintenance strategies converted into yuan according to preset rules. Based on the benefit value B(θ) of virtual debugging, the maintenance strategy with the greatest benefit is selected.

[0054] This embodiment needs to specifically explain the maintenance strategy costs, such as the energy cost of adjusting pressure and the time cost of downtime maintenance. The preset rules are: for example, the energy cost of adjusting pressure multiplied by the electricity price is the cost in yuan, and the downtime maintenance time multiplied by the output value per unit time is the cost in yuan.

[0055] C3: Optimal Maintenance Recommendation: This includes determining the latest maintenance time t based on the Remaining Useful Life (RUL) predicted by the LSTM model and the production schedule. ma Maintenance recommendations are determined based on the hydraulic pump risk index. Based on the benefit value B(θ) of virtual commissioning, the maintenance strategy with the greatest benefit is selected. At the same time, the hydraulic pump health index HI, hydraulic pump risk index RI, and remaining useful life RUL are displayed. Finally, the optimal maintenance recommendation is sent to the management terminal for human-computer interaction.

[0056] This embodiment specifically explains that the built-in digital twin model of the hydraulic pump integrates physical models such as the hydraulic pump flow continuity equation and Bernoulli equation with the data-driven model of the hydraulic pump state diagnosis output to realize real-time mapping of the hydraulic pump operating state, thereby constructing a digital twin model of the hydraulic pump.

[0057] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0058] In conclusion, the above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be considered as such.

[0059] It is included within the scope of protection of this invention.

Claims

1. A hydraulic pump condition monitoring system based on multimodal information fusion and digital twin, characterized in that: include: Hydraulic pump multimodal data acquisition module: Synchronously acquires key operating parameters of the hydraulic pump through a distributed sensor array, and obtains the hydraulic pump multimodal parameter set after preprocessing; Hydraulic pump multimodal feature engineering module: Extracts features from the multimodal parameter set of hydraulic pump to obtain a multimodal feature dataset of hydraulic pump, including basic parameter features, vibration signal features, and motor and flow characteristics; Hydraulic pump status hierarchical fusion diagnosis module: includes a parallel expert model diagnosis unit and a fusion diagnosis unit. Based on the hydraulic pump multimodal feature dataset, it first performs parallel expert model diagnosis, then performs fusion diagnosis, obtains hydraulic pump status diagnosis results, and transmits them to the hydraulic pump health status and risk assessment module. The parallel expert model diagnostic unit includes a cavitation diagnostic sub-model, a bearing wear diagnostic sub-model, and an internal leakage diagnostic sub-model. The fusion diagnostic unit includes: B1: Define the identification framework: Based on the DS evidence theory fusion, define the identification framework Θ, Θ={normal, cavitation, bearing wear, internal leakage, composite fault}, composite fault is the situation where two or more of the above single faults exist at the same time, bearing wear includes inner ring fault, outer ring fault and ball fault; B2: Basic probability allocation: Based on the output of the cavitation diagnosis sub-model, including the cavitation confidence level C. cav Cavitation severity level L cav The basic probability allocation function m1(A) of the cavitation diagnostic sub-model is constructed using function f1, satisfying m1(A) i )+m1(Θ)=1, m1(A1) and m1(Θ) are the basic probability allocation result and the frame Θ probability allocation result, respectively; then, the basic probability allocation function m2(A) of the bearing wear diagnosis sub-model is constructed through function g1 and the basic probability allocation function m3(A) of the internal leakage diagnosis sub-model is constructed through function h1, with the same logic as the basic probability allocation function m1(A) of the cavitation diagnosis sub-model; B3: Evidence Combination: The basic probability assignment m(A) after combining the three sub-models is: The numerator represents the sum of the products of the basic probabilities of the corresponding subsets of each submodel when the intersection of the subsets A1, A2 and A3 from the basic probability assignments of each submodel is A. The denominator represents the fact that the intersection of the subsets is empty, and the conflict measure is subtracted from 1 for normalization. B4: Hydraulic pump condition diagnosis result: Based on the combined basic probability allocation m(A), the comprehensive health status HS of the hydraulic pump is obtained, HS=argmax A∈Θ m(A); Hydraulic pump health status and risk assessment module: Based on the hydraulic pump status diagnosis results, the hydraulic pump health index is obtained through mapping rules, the hydraulic pump risk index is obtained based on the hydraulic pump health index, and the remaining useful life is predicted to obtain the hydraulic pump health status and risk assessment results. Digital Twin and Decision Support Module: Based on the hydraulic pump status diagnosis results, different maintenance strategy parameters are input, and the built-in digital twin model of the hydraulic pump is used to virtually debug and optimize the maintenance strategy parameters to obtain the optimal maintenance suggestions for human-computer interaction at the management end.

2. The hydraulic pump condition monitoring system based on multimodal information fusion and digital twin as described in claim 1, characterized in that: The cavitation diagnosis sub-model in the parallel expert model diagnosis unit: based on the multimodal feature dataset of the hydraulic pump, the vibration average broadband power P... 500-1000 avg Imported vacuum degree P in and export pressure deviation rate DR p The multi-source feature vectors formed are used as the input feature vector X of the pre-defined random forest model. cav X cav =[P 500-1000 avg P in DR p Suppose that the random forest consists of K decision trees {h} k (x)} k=1 K For the input feature vector X cav Each decision tree outputs the probability p of cavitation occurring. k，cav p k，cav =h k (X cav Next, the cavitation confidence level C is obtained by averaging the outputs of the K decision trees. cav Finally, based on the cavitation confidence level, the cavitation severity level L is classified. cav ; The output of the cavitation diagnosis sub-model is obtained, including the cavitation confidence level C. cav Cavitation severity level L cav .

3. The hydraulic pump condition monitoring system based on multimodal information fusion and digital twin as described in claim 1, characterized in that: The bearing wear diagnosis sub-model in the parallel expert model diagnosis unit includes: A1: Constructing Input Features: Based on the multimodal feature dataset of hydraulic pumps, extract the three-dimensional vibration time-series data X of bearings. vib ∈R 3 ×T T represents the time step, 3 represents the three-dimensional vibration direction, and the characteristic frequencies f of the outer ring, inner ring, and ball bearing faults are extracted. o f i f b Based on the power spectral density S(f) of the vibration signal, the average power values ​​Sf in the three-dimensional vibration directions of the outer ring, inner ring, and balls at the characteristic frequency of bearing failure are obtained respectively. o，avg (f0), S i，avg (f i ), S b，avg (f b ), construct the input feature vector X of the pre-defined 1D convolutional neural network. in X in =[X vib ;S o，avg (f0), S i，avg (f i ), S b，avg (f b )]; A2: Local temporal feature extraction: First, perform convolutional layer 1, y1=ReLU(X in *w1+b1), where * represents the convolution operation, w1 is the convolution kernel, and w1∈R. k1×c1 k1=32 is the kernel size, c1=16 is the number of output channels, b1 is the bias term, and ReLU activation function is used; then pooling layer 1 is performed, y p1 =MaxPooling(y1, s1=2), the pooling window stride s1 is 2, and the output is y1. p1 ∈R c1×T / 2 MaxPooling is a pooling operation.

4. The hydraulic pump condition monitoring system based on multimodal information fusion and digital twin as described in claim 1, characterized in that: The bearing wear diagnosis sub-model in the parallel expert model diagnosis unit also includes: A3: fusing feature frequencies: first, performing convolutional layer 2, y2=ReLU(y p1 *w2+b2), where w2 is the convolution kernel, and w2∈R k2×c2 k2=16, c2=32, b2 is a bias term, which further integrates local temporal features with characteristic frequency power to enhance fault frequency correlation features; then pooling layer 2 is performed, y p2 =MaxPooling(y2, s2=2), outputs y p2 ∈R c1×T / 2 Flatten the pooling output into a vector y. flat ∈R N N is the dimension after flattening, resulting in the fully connected layer y. fc1 =ReLu(W3×y flat +b3), W3 is a fully connected layer y fc1 The weight matrix is ​​given by b3, where b3 is the corresponding bias term; the probability distribution P of the three types of faults is output through the Softmax function, P = Softmax(W4 × y fc1 +b4), W3 is the weight matrix of the output layer, and b4 is the corresponding bias term; A4: Output results: including the inner ring failure probability P in outer ring failure probability P out and the probability of ball bearing failure P ball and bearing wear confidence level C bear C bear ∈[0,1], if C bear If the value is greater than or equal to the corresponding threshold, it is determined to be the corresponding fault type, and the item with the highest probability is selected; otherwise, it is determined to be a bearingless fault.

5. The hydraulic pump condition monitoring system based on multimodal information fusion and digital twin as described in claim 1, characterized in that: The internal leakage diagnosis sub-model in the parallel expert model diagnosis unit: First, a multivariate time series X is constructed. le The time series length is t, X le =[Q(t), T(t), P out (t), P me (t), ΔP(t), DR Q [(t)], Q(t), T(t), P out (t), P me (t), ΔP(t) and DR Q (t) represents the measured flow rate, temperature, outlet pressure, power, power efficiency deviation, and flow rate decay rate, respectively; then, the multivariate time series X is processed through an LSTM layer. le Processing yields the hidden state sequence {h} t } t=1 tt h t Let be the hidden state at time t; then, the attention weight 'a' is calculated using an attention mechanism. t a t Let h be the attention weights of the hidden state at time t; then, the hidden states are weighted and fused to obtain the fused feature h. att , Next, feature h is fused. att The input is fed into a fully connected layer and passes through the Sigmoid function (Sigmoid(W)). fc ×h att +b fc The internal leakage rate R is obtained. le W fc and b fc These are the weight matrix and bias term of the fully connected layer, respectively; finally, the prediction error (MAE) and feature consistency (CV) of the LSTM+Attention model are combined to obtain the inner leakage confidence C. le C le =(1-MAE)×γ+(1-CV)×(1-γ), where MAE is the model-predicted leakage rate R. le Compared with the true value R le，0 The mean absolute error, CV, is the power efficiency deviation ΔP and the flow rate attenuation rate DR. Q coefficient of variation; The output of the internal leakage diagnostic sub-model is obtained, including the internal leakage rate R. le and internal leakage confidence level C le .

6. The hydraulic pump condition monitoring system based on multimodal information fusion and digital twin according to claim 1, characterized in that: The hydraulic pump health index is based on the hydraulic pump condition diagnosis result HS, which is multiplied by 100 and mapped to a score of 1-100. It includes the normal hydraulic pump health index HI score range, the HI score range for cavitation, the HI score range for bearing wear, the HI score range for internal leakage, and the HI score range for compound faults. By introducing a volumetric efficiency attenuation coefficient kη to correct HI, the corrected HI is finally obtained. th HI th =HI×kη, where kη is the current efficiency η of the hydraulic pump. me The ratio of the theoretical efficiency η to the current efficiency of the hydraulic pump, which is the ratio of the output power to the input power. The risk index of the hydraulic pump is calculated by pre-defining the failure severity weight S and calculating the risk index RI, where RI = (100 - HI) × S / 10. Then, risk levels are classified: if RI is at the preset high-risk level, an immediate shutdown maintenance recommendation is triggered; if RI is at the medium-risk level, a planned maintenance recommendation is triggered; if RI is at the low-risk level, a continuous monitoring recommendation is triggered. The remaining useful life prediction: A pre-trained LSTM model is used to predict the next n2 time steps, with the HI sequence [HI] of the most recent m1 time steps as input. TT-m1+1 , ..., HI TT ], TT represents the current time, and the next time HI is predicted. TT+1 Then [HI] TT-m1+2 , ..., HI TT HI TT+1 As a new input, predict HI TT+2 By analogy, the predicted HI decay curve result HI(TT1) is obtained, where TT1 > TT; a time threshold HI is set. fail Solve for HI(TT1)≤HI fail The minimum time t fail Then the remaining useful lifetime is RUL=t fail -t me , t me This refers to the current moment.

7. The hydraulic pump condition monitoring system based on multimodal information fusion and digital twin as described in claim 1, characterized in that: The digital twin and decision support module includes: C1: Based on the hydraulic pump condition diagnosis results, inputting different maintenance strategy parameters θ into the built-in digital twin model of the hydraulic pump, θ=[θ1, θ2, ..., θ...]. M ], θ M Let θ be the parameter of the Mth maintenance strategy; then, based on the remaining useful life (RUL), the RUL after implementing the maintenance strategy θ is obtained. θ RUL θ =RUL0+ΔRUL(θ), where ΔRUL(θ) is the lifetime extension time; C2: Strategy Benefit Calculation: Benefit value B(θ) is the output value RUL brought about by the lifespan extension. θ ×C pr Total maintenance cost C ma The ratio of (θ) is obtained, C pr The output value per unit time is calculated based on the economic value produced by the hydraulic pump, with the unit being yuan / s. The total maintenance cost is also in yuan, referring to the cost of all maintenance strategies converted into yuan according to preset rules. Based on the benefit value B(θ) of virtual debugging, the maintenance strategy with the greatest benefit is selected. C3: Optimal Maintenance Recommendation: This includes determining the latest maintenance time t based on the Remaining Useful Life (RUL) predicted by the LSTM model and the production schedule. ma Maintenance recommendations are determined based on the hydraulic pump risk index. Based on the benefit value B(θ) of virtual commissioning, the maintenance strategy with the greatest benefit is selected. At the same time, the hydraulic pump health index HI, hydraulic pump risk index RI, and remaining useful life RUL are displayed. Finally, the optimal maintenance recommendation is sent to the management terminal for human-computer interaction.

Images