A heavy fork electrical system fault diagnosis and intelligent repair method

Abstract

The application discloses a kind of heavy cross electrical system fault diagnosis and intelligent repair method, specifically related to electrical system intelligent operation and maintenance technical field, the multi-source data of electrical parameter, mechanical parameter and environmental parameter are collected by sensor, the space-time attention feature extraction network is used to carry out fault space-time feature extraction, then based on real-time fault space-time feature vector and historical fault space-time feature vector similarity obtains fault intelligent repair strategy, corresponding intelligent repair instruction is issued through CAN bus, while monitoring execution result, the execution result data after repair operation is evaluated, and the evaluation abnormal result is fed back to management terminal to carry out man-machine interaction;The application is extracted by the network of spatial attention subnet and hollow convolution time series feature extractor, can comprehensively, accurately capture the characteristic information of system operating state, provide abundant feature basis for subsequent fault diagnosis and intelligent repair.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent operation and maintenance of electrical systems, in particular to a heavy fork electrical system fault diagnosis and intelligent repair method. BACKGROUND

[0002] In logistics transportation and warehousing operations, heavy forklifts are the core equipment in the industrial logistics field, and the reliability of their electrical systems directly determines production efficiency. However, as the complexity of equipment increases and the degree of intelligence improves, fault diagnosis and repair face many challenges.

[0003] Traditional fault diagnosis mainly relies on the experience and intuitive judgment of maintenance personnel, and preliminary judgments of fault location and cause are made by observing the running state of the forklift, listening to abnormal sounds, touching component temperatures, etc. Only a few key physical quantities of the electrical system, such as voltage and current, are usually detected, making it difficult to fully reflect the running state of the system. Traditional repair strategies are often based on pre-set rules and experience, lacking in pertinence and flexibility. After repair is completed, there is a lack of effective evaluation and feedback mechanism for repair effectiveness.

[0004] Currently, in the field of fault diagnosis and repair, existing research mostly focuses on one aspect of fault diagnosis or repair, lacking a systematic method that combines fault diagnosis, repair strategy generation, automatic repair execution, and repair effectiveness evaluation and iteration, making it difficult to achieve intelligent processing of heavy fork electrical system faults throughout the entire process. Moreover, heavy fork electrical systems involve multiple types of sensor data, such as voltage, current, temperature, vibration, etc. These data have different characteristics and dimensions, and how to effectively integrate these multi-source data to extract more valuable fault information is a problem that existing technology has not yet solved well. Therefore, it is of great practical significance to develop an efficient, accurate, and intelligent repair-capable heavy fork electrical system fault diagnosis method. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a heavy fork electrical system fault diagnosis and intelligent repair method to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a heavy fork electrical system fault diagnosis and intelligent repair method, comprising:

[0007] S1: Collect and store historical heavy fork electrical system operation and maintenance data, including running data under standard conditions, normal running data, fault running data, and fault repair strategies, while receiving and updating data generated during the operation of the heavy fork electrical system in real time, and constructing a heavy fork electrical system management database;

[0008] S2: Real-time synchronous acquisition of multi-source data of electrical parameters, mechanical parameters and environmental parameters through sensors deployed at the monitoring position of the heavy-duty electrical system, and fusion to construct a three-dimensional data matrix for diagnosing faults;

[0009] S3: Feature extraction of the three-dimensional data matrix after real-time acquisition and preprocessing through the spatio-temporal attention feature extraction network comprising the spatial attention subnet and the hollow convolution time sequence feature extractor, to obtain a fault spatio-temporal feature vector;

[0010] S4: Construction of a fault diagnosis and intelligent repair knowledge graph through historical fault spatio-temporal feature vectors and corresponding intelligent repair strategies, and acquisition of a fault intelligent repair strategy based on the similarity between real-time fault spatio-temporal feature vectors and historical fault spatio-temporal feature vectors;

[0011] S5: Based on the acquired fault intelligent repair strategy, corresponding intelligent repair instructions are issued through the CAN bus, and the execution results are monitored;

[0012] S6: Evaluation of the execution results after the repair operation, and feedback of the evaluation abnormal results to the management terminal for human-computer interaction.

[0013] Technical effects and advantages of the present application:

[0014] 1. The present application uses a multi-physical quantity synchronous acquisition module to arrange sensors at the monitoring position of the heavy-duty electrical system, and synchronously acquires multi-source data such as electrical parameters, mechanical parameters and environmental parameters in real time, which can more comprehensively reflect the running condition of the system, effectively avoid misjudgment or missed judgment of faults due to a single parameter abnormality, and significantly improve the accuracy and reliability of fault diagnosis;

[0015] 2. The present application uses a network comprising a spatial attention subnet and a hollow convolution time sequence feature extractor for feature extraction, which can comprehensively and accurately capture feature information of the system running state, adaptively adjust the attention degree to different spatial positions and time sequence features, adapt to various complex operating conditions, and provide rich feature basis for subsequent identification of fault state labels and intelligent repair;

[0016] 3. The present application uses historical fault spatio-temporal feature vectors and corresponding intelligent repair strategies to construct a knowledge graph, quickly acquires a fault intelligent repair strategy based on the similarity between real-time fault spatio-temporal feature vectors and historical fault spatio-temporal feature vectors, reduces the time for manual analysis and decision-making, issues corresponding intelligent repair instructions through the CAN bus, monitors the execution results, and continuously optimizes the system according to the monitoring results to improve the adaptive ability. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The present application is a whole process schematic diagram.

[0018] Figure 2 The method flowchart of the present application is shown in the figure. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0020] Please refer to Figure 1 The present application provides an AI-based offshore wind turbine cabin structure production control system, which comprises a heavy fork electrical system management database, a multi-physical quantity synchronous acquisition module, a space-time attention feature extraction module, a fault-repair dynamic mapping module, an intelligent repair strategy execution module, and a repair effect evaluation and feedback module.

[0021] The heavy fork electrical system management database is connected with the rest of the modules, the multi-physical quantity synchronous acquisition module is connected with the space-time attention feature extraction module, the fault-repair dynamic mapping module is connected with the space-time attention feature extraction module and the intelligent repair strategy execution module respectively, and the repair effect evaluation and feedback module is connected with the fault-repair dynamic mapping module and the intelligent repair strategy execution module respectively.

[0022] Heavy fork electrical system management database: used for collecting and storing historical heavy fork electrical system operation and maintenance data, including running data under standard conditions, normal running data, fault running data and fault repair strategy, while receiving and updating data generated during the operation of the heavy fork electrical system in real time, and constructing the heavy fork electrical system management database;

[0023] Multi-physical quantity synchronous acquisition module: through the sensors arranged at the monitoring position of the heavy fork electrical system, real-time synchronous acquisition of multi-source data of electrical parameters, mechanical parameters and environmental parameters is carried out, and fusion is carried out to construct a three-dimensional data matrix for diagnosing faults, and transmitted to the space-time attention feature extraction module;

[0024] Space-time attention feature extraction module: through the space-time attention feature extraction network comprising the space attention subnetwork and the hollow convolution time sequence feature extractor which has been constructed, the three-dimensional data matrix after real-time acquisition and preprocessing operation is subjected to feature extraction to obtain a real-time fault space-time feature vector, and transmitted to the fault-repair dynamic mapping module;

[0025] Fault-repair dynamic mapping module: a fault diagnosis and intelligent repair knowledge graph is constructed through historical fault spatiotemporal feature vectors and corresponding intelligent repair strategies, a fault intelligent repair strategy is obtained based on the similarity between real-time fault spatiotemporal feature vectors and historical fault spatiotemporal feature vectors, and the intelligent repair strategy is transmitted to the intelligent repair strategy execution module;

[0026] Intelligent repair strategy execution module: based on the obtained fault intelligent repair strategy, corresponding intelligent repair instructions are issued through a CAN bus, the execution results are monitored, and the execution results are transmitted to the repair effect evaluation and feedback module;

[0027] Repair effect evaluation and feedback module: the execution result data after the repair operation is evaluated, and the evaluation abnormal result is fed back to the management terminal for human-computer interaction.

[0028] Please refer to Figure 2 As shown in FIG. 1, a heavy fork electrical system fault diagnosis and intelligent repair method comprises the following steps: S1: collecting and storing historical heavy fork electrical system operation and maintenance data, including running data under standard conditions, normal running data, fault running data and fault repair strategy, and simultaneously receiving and updating data generated during the operation of the heavy fork electrical system, to construct a heavy fork electrical system management database; S2: through sensors deployed at monitoring positions of the heavy fork electrical system, real-time synchronous collection of multi-source data of electrical parameters, mechanical parameters and environmental parameters is performed, and fusion is performed to construct a three-dimensional data matrix for diagnosing faults; S3: through a spatiotemporal attention feature extraction network comprising a spatial attention subnet and a hollow convolution time sequence feature extractor, feature extraction is performed on the three-dimensional data matrix after real-time collection and preprocessing, to obtain a fault spatiotemporal feature vector; S4: a fault diagnosis and intelligent repair knowledge graph is constructed through historical fault spatiotemporal feature vectors and corresponding intelligent repair strategies, and a fault intelligent repair strategy is obtained based on the similarity between real-time fault spatiotemporal feature vectors and historical fault spatiotemporal feature vectors; S5: based on the obtained fault intelligent repair strategy, corresponding intelligent repair instructions are issued through a CAN bus, and the execution results are monitored; S6: the execution result data after the repair operation is evaluated, and the evaluation abnormal result is fed back to the management terminal for human-computer interaction.

[0029] S1: collecting and storing historical heavy fork electrical system operation and maintenance data, including running data under standard conditions, normal running data, fault running data and fault repair strategy, and simultaneously receiving and updating data generated during the operation of the heavy fork electrical system, to construct a heavy fork electrical system management database;

[0030] S2: through sensors deployed at monitoring positions of the heavy fork electrical system, real-time synchronous collection of multi-source data of electrical parameters, mechanical parameters and environmental parameters is performed, and fusion is performed to construct a three-dimensional data matrix for diagnosing faults, comprising the following steps:

[0031] S2.1: Collecting electrical parameters with a sampling frequency f1 (e.g. 1kHZ), mechanical parameters with a sampling frequency f2 (e.g. 100Hz) and environmental parameters with a sampling frequency f2 (e.g. 10Hz) by sensor technology; the electrical parameters include three-phase voltage U a,b,c (t), three-phase current I a,b,c (t) and three-phase current harmonic distortion rate THD a,b,c (t), the mechanical parameters include x-direction vibration acceleration a x (t), y-direction vibration acceleration a y (t) and z-direction vibration acceleration a z (t), the environmental parameters include temperature T(t), humidity H(t) and dust concentration D(t);

[0032] It needs to be specifically explained in this embodiment that the heavy-duty fork electrical system monitoring positions include motor end, controller end, hydraulic pump end, cab, etc.; the sensors include but are not limited to voltage sensor, current sensor, vibration sensor, temperature and humidity composite sensor, dust concentration sensor, etc.

[0033] It needs to be specifically explained in this embodiment that the current in the heavy-duty fork electrical system is collected in real time by using the current sensor, and the discrete sequence I(n) of the current changing with time is obtained, n is the sampling number of time point t, the collected current discrete sequence is processed by FFT (Fast Fourier Transform), the current harmonic distortion rate THD(t) (for the existing formula) is calculated, the three-phase current is traversed, and the three-phase current harmonic distortion rate THD a,b,c (t) is obtained; in order to accurately describe the mechanical vibration of the heavy-duty fork system, three mutually perpendicular directions are established, for example, in the heavy-duty fork system, the vehicle body of the heavy-duty forklift may produce vibration in the front-back (assuming x-direction), left-right (assuming y-direction) and vertical (assuming z-direction) directions during operation, a x reflects the vibration intensity of the object along the x-axis direction, a y reflects the vibration intensity of the object along the y-axis direction, a z reflects the vibration intensity of the object along the z-axis direction.

[0034] S2.2: Adopting PTP precise clock protocol (IEEE 1588) to synchronize the time stamp of each sensor, the error is less than or equal to threshold value Δt (e.g. 50ns), calculating the synchronization time t sync of each sensor, t sync =t mas +Δt pro -Δt deL , t mas is the master clock reference time, and Δt proFor the propagation delay, Δt deL For the clock offset compensation value;

[0035] S2.3: Fusion of the synchronized electrical parameters, mechanical parameters and environmental parameters to construct a three-dimensional data matrix M(t) ∈ R 3×3×n , n is the sampling number of time point t, ;

[0036] The embodiment needs to be specifically explained that the three-phase voltage U a,b,c (t) as a whole, the three-phase current I a,b,c (t) as a whole and the three-phase current harmonic distortion rate THD a,b,c (t) as a whole.

[0037] S3: Through the constructed spatio-temporal attention feature extraction network containing the spatial attention subnetwork and the hollow convolution time sequence feature extractor, the three-dimensional data matrix after the real-time collection and preprocessing operation is subjected to feature extraction to obtain a fault spatio-temporal feature vector, including the following steps:

[0038] S3.1: The construction of the spatio-temporal attention feature extraction network includes:

[0039] S3.1.1: First, obtain N groups (for example, 5000 groups) of historical three-dimensional data matrices M(t) of different fault states of the cross electrical system from the cross electrical system management database, and perform preprocessing operations on the three-dimensional data matrices, including denoising processing, outlier processing and data normalization processing, to obtain the preprocessed three-dimensional data matrix M y (t) ∈ R 3×3×n ; Then, the N groups of preprocessed three-dimensional data matrices M y (t) ∈ R 3×3×n and the corresponding fault state one-hot encoding vectors are taken as a training data set D, , [M N (t), yN] is the fault state label of the Nth three-dimensional data matrix M N (t), and the fault state label is a one-hot encoding vector, which is composed of no fault, three fault types and corresponding three severity levels, the three fault types include electrical fault, mechanical fault and environmental fault, and the severity levels include mild, moderate and severe, a total of 3x3+1=10 fault state label categories, for example, the one-hot encoding vector of no fault is [1, 0, 0, 0, 0, 0, 0, 0, 0, 0], the one-hot encoding vector of mild electrical fault is [0, 1, 0, 0, 0, 0, 0], 1 represents electrical fault, and 0.5 represents the fault degree coefficient of electrical fault;

[0040] The embodiment needs to be specifically explained that the preprocessing operation is the prior art, and the denoising can adopt filtering algorithms such as median filtering, mean filtering and the like; the abnormal value processing can use statistical methods such as the 3σ principle; the normalization can map the data to the [0, 1] interval, and the commonly used methods include minimum-maximum normalization.

[0041] The embodiment needs to be specifically explained that the fault degree is the sum of the absolute values of the differences between all fault elements in each fault type in the three-dimensional data matrix and the corresponding elements under the standard condition and the elements under the standard condition, and then the fault degree values of the fault elements are compared with the corresponding degree value ranges of the preset light, medium and heavy, to obtain the severity of the fault category corresponding to the fault element, for example, the mechanical fault degree = [(|a x -a x,0 | / a x,0 )+(|a y -a y,0 | / a y,0 )+(|a z -a z,0 | / a z,0 )], a x,0 , a y,0 and a z,0 are the accelerations in the x direction, the y direction and the z direction under the standard condition.

[0042] S3.1.2: First, add an adaptive gating function ΘM y (t) to the spatial attention weight Ms(M y (t)) in the CBAM attention module structure, calculate the spatial attention weight Ms t (M y (t)) after adding the adaptive gating function, Ms t (M y (t))=Ms(M y (t))×ΘM y (t), Ms(M y (t))=σ(Conv 7×7 ([AvgPool(M y (t));MaxPool(M y (t))])),ΘM y (t)=σ(k×M y (t)+b),k is a learnable weight parameter, b is a learnable bias parameter, Ms(M y (t)) is obtained by first performing average pooling AvgPool(M y (t) and maximum pooling MaxPool(M y (t) on the input three-dimensional data matrix M y(t)) operation, and after concatenating the two pooling results in the channel dimension, a 7x7 convolution Conv 7×7 and a sigmoid activation function σ, and then the spatial attention weight Ms t (M y (t)) is multiplied element-wise with the input three-dimensional data matrix M y (t) to obtain the output spatial feature Fs of the spatial attention subnetwork, Fs=Ms t (M y (t) is multiplied element-wise with the input three-dimensional data matrix M y (t).

[0043] It needs to be specifically explained in this embodiment that CBAM (Convolutional Block Attention Module) is a convolutional block attention module, which is composed of a channel attention (ChannelAtention) and a spatial attention (SpatiaAtenion) two submodules; the design goal of CBAM is to improve the feature expression ability of the convolutional neural network by explicitly modeling the attention of the channel and spatial two dimensions, and in this application, the fault spatiotemporal feature expression ability of the convolutional neural network is improved.

[0044] S3.1.3: First, the output Fs of the spatial attention subnetwork is taken as the input of the empty convolution time series feature extractor, and the empty convolution structure is a 3-layer empty convolution stack with a dilation rate of 2, 4 and 8, respectively, where K represents the dilation rate. The receptive field calculation formula of the empty convolution is RF, , RF pre is the receptive field size of the previous layer, and the size of the initial layer receptive field is the size of the convolution kernel, k is the size of the convolution kernel (for example, k=3), stride is the step size (for example, stride=1), n is the number of empty convolution layers, and dilation i is the dilation rate of the i-th layer; then the input Fs is subjected to time series feature extraction by empty convolution to obtain the output time series feature Ft of the empty convolution time series feature extractor, Ft=Conv d-8 (Conv d-4 (Conv d-2 (Fs))), Conv d-K is a convolution operation with a dilation rate of K;

[0045] S3.1.4: First, the spatial feature Fs obtained by the spatial attention subnetwork and the time series feature Ft obtained by the empty convolution time series extractor are subjected to feature fusion to obtain a fault spatiotemporal feature vector F with a dimension of 512, F=Flatten([Fs;Ft]), and Flatten() is a flattening operation; then the fault spatiotemporal feature vector F is input into a fully connected layer and a Softmax function to obtain a fault state label prediction probability yy y y =Softmax(W F T F+b F ), W F W is an m1×n1 dimensional vector, where n1 is the dimension of the fault spatiotemporal feature vector F, m1 is the number of classification categories, i.e., 10 fault state label categories. F T For W F transpose, b F The bias vector is a learnable vector with m1 dimensions; then, the prediction result y is constructed using the cross-entropy loss function. y The loss function L(θ) between the true label y and the actual label y is... y c Let y be the c-th element in the one-hot encoding of label y. c y To predict the probability y y The c-th element in the algorithm represents the learnable parameters of the network. Then, a backpropagation algorithm (e.g., stochastic gradient descent, Adam optimization algorithm) is used to calculate the gradient ∇ of the network parameters. θ L adjusts the value of θ and finds the minimum value of the loss function L(θ). When the value of the loss function changes less than the threshold for d1 consecutive times (e.g., 5 times), it indicates that the training is over, d1 < d; finally, the spatiotemporal attention feature extraction network is output.

[0046] In this embodiment, it should be specifically noted that in the fault diagnosis scenario of heavy forklift electrical system, the larger the sensing field, the more long-distance dependencies can be captured in the time series data, which helps to discover features such as the development of intermittent faults over time and the accumulation process of gradual faults.

[0047] S3.2: Using the constructed spatiotemporal attention feature extraction network, features are extracted from the preprocessed 3D data matrix acquired in real time to obtain the real-time fault spatiotemporal feature vector F. now Then, the real-time fault spatiotemporal feature vector F now The input is fed into a fully connected layer and a softmax function to obtain the fault state label; if the real-time fault spatiotemporal feature vector F now If the corresponding fault status label is "no fault", no further action is required; otherwise, further action is required.

[0048] S4: Construct a fault diagnosis and intelligent repair knowledge graph using historical fault spatiotemporal feature vectors and corresponding intelligent repair strategies. Obtain intelligent fault repair strategies based on the similarity between real-time fault spatiotemporal feature vectors and historical fault spatiotemporal feature vectors, including the following steps:

[0049] S4.1: First, define the fault feature node set V F and the intelligent repair strategy node set V R , , F N is the Nth set of historical fault spatiotemporal feature vectors, N is the number of historical fault cases; then the BERT model is used to hierarchically encode the intelligent repair strategy text to obtain the text feature vector R, R = BERT([w1, w2,...wm1]), wm1 represents the m1th word unit after tokenization and part-of-speech tagging, and the intelligent repair strategy node set V R , , R N is the intelligent repair strategy text vector corresponding to F N ; finally, a directed edge E N from F N to R N is established for each set of historical fault cases (F N , R N ), and a preliminary fault diagnosis and intelligent repair knowledge graph READ1 is constructed;

[0050] It needs to be specifically explained that BERT (Bidirectional Encoder Representations from Transformers) is a pre-training language model based on the Transformer architecture. Through unsupervised learning on a large-scale intelligent repair strategy text corpus, it has learned rich language representations and semantic information. Through a bidirectional Transformer encoder, it can capture the context information in the text and better understand the meaning of words in different contexts.

[0051] S4.2: For the weight W N of each directed edge E N of the initial fault diagnosis and intelligent repair knowledge graph READ1, the edge weight contribution degree Δw ij of the remaining N-1 sets of historical fault cases is calculated and accumulated to obtain , , Δw ij is the edge weight contribution degree of the ith node and the jth node, i≠j, i∈N, j∈N, α is a balance coefficient (for example, α = 0.75), Sim(F i , F j ) is the cosine similarity of the ith and jth historical fault spatiotemporal feature vectors F i and F j , and ||R i -R j|| is the distance norm (measuring the distance between vectors, the smaller the distance, the closer the exponential result to 1) of the corresponding ith and jth intelligent repair strategy text vectors, SF i th1 is the corresponding threshold value (for example, 0.8) of Sim(F i ,F j ); then the weight W N of all directed edges E N is traversed to build a complete fault diagnosis and intelligent repair knowledge graph READ2;

[0052] S4.3: based on the similarity function Sim, traverse C N 2 groups (C N 2 is the number of combinations, indicating the selection of 2 groups from N groups of historical fault cases) of historical fault spatiotemporal feature vectors, obtain the similarity SF i of each 2 groups of historical fault spatiotemporal feature vectors, if SF i ≥ threshold value SF i th1 , then merge into one node, otherwise exist as an independent node, traverse the similarity of C N 2 groups of historical fault spatiotemporal feature vectors, update the fault feature node set V F to V F g ; based on the distance function (for example, Euclidean distance), traverse C N 2 groups of intelligent repair strategy nodes, obtain the distance function value SR i of each 2 groups of intelligent repair strategy text vectors, if SR i < threshold value SR i th (for example, 0.3), then merge into one node, otherwise exist as an independent node, traverse the distance function value of C N 2 groups of intelligent repair strategy text vectors, update the repair strategy node set V R to V R g , and finally obtain the updated fault diagnosis and intelligent repair knowledge graph READ3;

[0053] S4.4: calculate the cosine similarity Sim(F now ,F F ) between the real-time fault spatiotemporal feature vector F g and each node F N in the fault node set V now in the fault diagnosis and intelligent repair knowledge graph READ3; N), filter out those with similarity ≥ threshold SF i th1 The faulty nodes are used as the candidate fault feature node set VF now Then, for candidate fault feature nodes F... N ∈VF now To the intelligent repair strategy node R N Given the path, obtain the set of path edge weights W. N (VF now The optimal repair strategy R is selected by filtering the repair strategy text corresponding to the path with the highest path edge weight. opt This refers to intelligent fault repair strategies (e.g., motor winding temperature > 150℃ AND abnormal vibration acceleration THEN generation strategy: shutdown inspection + winding insulation test + stator replacement if necessary); if the cosine similarity Sim(F now ,F N ) < threshold SF i th1 This triggers the management terminal decision interface;

[0054] S4.5: If the real-time fault spatiotemporal feature vector F now With the set of fault characteristic nodes V F g node F in N cosine similarity Sim(F) now ,F N )≥threshold SF i th2 (e.g., 0.95), threshold SF i th2 > Threshold SF i th1 Update the fault spatiotemporal feature vector of the node; otherwise, add a fault feature node and a corresponding intelligent repair strategy node. If the freshness fh of historical fault cases is less than the corresponding threshold fh... th (For example, if fhth=0.2), then the faulty case is eliminated; otherwise, it is retained, fh=f(t) case )×g(up),f(t) case The basic time decay function is... , t case T represents the number of days since a historical failure. max This represents the maximum number of days since a historical failure (e.g., 730 days), where g(up) represents the upgrade and correction items for the heavy forklift system equipment. up=1 represents historical fault cases after the equipment upgrade, up=0 represents historical fault cases before the equipment upgrade, and β is a correction factor (e.g., β=0.5); for example, case A: t case=365 days, up=1 (upgraded case), fh=(1-365 / 730)×1=0.5, 0.5≥0.2, case retained;

[0055] S5: Based on the acquired intelligent fault repair strategy, send the corresponding intelligent repair command through the CAN bus, and monitor the execution result. If the execution fails, trigger S4 to reacquire the intelligent fault repair strategy. The intelligent repair command is sent through the CAN bus message, which includes the fault status label category and the intelligent repair strategy text.

[0056] In this embodiment, it should be specifically explained that, for example, if the fault is an electrical fault and the power module control mechanism fails to successfully adjust the output voltage waveform, the monitoring system will send a command or switch to a backup control strategy, such as activating a backup filter circuit.

[0057] S6: Evaluate the execution result data after the repair operation, and feed back any abnormal evaluation results to the management terminal for human-computer interaction, including the following steps:

[0058] S6.1: Monitor the execution results of the repair strategy and the real-time status data of the system via the CAN bus. The data on the execution results of the repair strategy includes the number of successful fault repairs (sc), the total number of repairs (ta), and the time (t) for each repair. I The spatiotemporal feature vector F of the I-th fault after repair I h And the spatiotemporal feature vector F of the I-th fault under standard conditions I b Calculate the performance evaluation index η. , t I,h Let ||·||2 be the average repair time of the three most recent repair operations that had the same repair operation as the I-th repair operation, and let ||·||2 be the L2 norm, where I∈ta;

[0059] S6.2: The performance evaluation metric η is less than the corresponding threshold η th This triggers incremental learning of the fault diagnosis and intelligent repair knowledge graph READ3. new =READ3∪[(F now ,F h )], F now and F h These are the real-time fault spatiotemporal feature vector and the repaired fault spatiotemporal feature vector, respectively, and are fed back to the management terminal. Otherwise, it indicates that the execution result is good.

[0060] 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.

[0061] Finally: the above only is the preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement etc. made within the spirit and principle of the present application, should be included in the protection scope of the present application

[0062] included in the protection scope of the present application.

Claims

1. A method for fault diagnosis and intelligent repair of a heavy fork electrical system, characterized in that: Comprise: S1: Collect and store historical cross-over electrical system operation data, including standard operating data, normal operating data, fault operating data and fault repair strategy, while receiving and updating data generated during the operation of the cross-over electrical system in real time, and build a cross-over electrical system management database; S2: Through the sensors deployed in the monitoring position of the cross-over electrical system, real-time synchronous acquisition of multi-source data of electrical parameters, mechanical parameters and environmental parameters is carried out, and a three-dimensional data matrix for diagnosing faults is constructed by fusion; S3: Through the spatio-temporal attention feature extraction network including the spatial attention subnetwork and the hollow convolution time sequence feature extractor which has been constructed, feature extraction is carried out on the three-dimensional data matrix collected in real time after preprocessing, and a fault spatio-temporal feature vector is obtained; S3.1.1: First, obtain N sets of historical three-dimensional data matrices M(t) of different fault states of the heavy fork electrical system from the heavy fork electrical system management database, and perform preprocessing operations on the three-dimensional data matrices to obtain preprocessed three-dimensional data matrices M y (t)∈R 3×3×n ; Then take the N sets of preprocessed three-dimensional data matrices M y (t)∈R 3×3×n and the corresponding fault state one-hot encoding vectors as the training data set D, , [M N (t), yN] is the fault state label of the Nth set of three-dimensional data matrices M N (t), and the fault state label is a one-hot encoding vector. There are 10 fault state label categories, including: no fault, electrical fault mild, electrical fault moderate, electrical fault severe, mechanical fault mild, mechanical fault moderate, mechanical fault severe, environmental fault mild, environmental fault moderate, and environmental fault severe. S3.1.2: First, the spatial attention weight Ms(M y (t)) is added with an adaptive gating function ΘM y (t), and the spatial attention weight Ms(M t (t) after adding the adaptive gating function is calculated. y (t), Ms(M t (t)) = Ms(M y (t)) x ΘM y (t), Ms(M y (t)) = σ(Conv y ([AvgPool(M 7×7 (t)); MaxPool(M y (t))]), ΘM y (t) = σ(k x M y (t) + b), k is a learnable weight parameter, b is a learnable bias parameter, and Ms(M y (t) is obtained by first performing average pooling AvgPool(M y (t) and maximum pooling MaxPool(M y (t) operations on the input three-dimensional data matrix M y (t), then concatenating the two pooling results in the channel dimension, and then passing through a 7 x 7 convolution Conv y and a sigmoid activation function σ; then the spatial attention weight Ms(M 7×7 (t) is multiplied element by element with the input three-dimensional data matrix M t (t) to obtain the output spatial feature Fs of the spatial attention subnetwork, Fs = Ms(M y (t)) o M y (t). t (t). y (t). y (t). S3.1.3: First, the output Fs of the spatial attention subnetwork is taken as the input of the empty convolution time sequence feature extractor, and the empty convolution structure is a 3-layer empty convolution stack with empty rates of 2, 4 and 8, denoted as K. The receptive field calculation formula of the empty convolution is RF, , RF pre is the receptive field size of the previous layer, k is the kernel size, stride is the step, n is the number of empty convolution layers, and dilation i is the empty rate of the i-th layer; then the input Fs is extracted by the empty convolution to obtain the output time sequence feature Ft of the empty convolution time sequence feature extractor, Ft=Conv d-8 (Conv d-4 (Conv d-2 (Fs))), Conv d-K is the convolution operation with the empty rate K; S4: A fault diagnosis and intelligent repair knowledge graph is constructed through the historical fault spatio-temporal feature vector and the corresponding intelligent repair strategy, and the fault intelligent repair strategy is obtained based on the similarity between the real-time fault spatio-temporal feature vector and the historical fault spatio-temporal feature vector; S4.1: First, define the fault feature node set V F and the intelligent repair strategy node set V R , , F N is the Nth group of historical fault spatiotemporal feature vectors, N is the number of historical fault cases; Then the BERT model is used to hierarchically encode the intelligent repair strategy text to obtain the text feature vector R, R=BERT([w1,w2,...wm1]), wm1 represents the m1th word unit after tokenization and part-of-speech tagging, and the intelligent repair strategy node set V R , , R N is the intelligent repair strategy text vector corresponding to F N ; Finally, a directed edge E N from F N to R N is established for each group of historical fault cases (F N , R N ), and a preliminary fault diagnosis and intelligent repair knowledge graph READ1 is constructed; S4.2: For each directed edge E of the initial fault diagnosis and intelligent repair knowledge graph READ1 N , the weight W N is calculated by accumulating the edge weight contribution Δw ij of the remaining N-1 groups of historical fault cases, that is, ; then traverse the weight W N of all directed edges E N to build a complete fault diagnosis and intelligent repair knowledge graph READ2; S4.3: traversing C based on the similarity function Sim N 2 group historical fault spatiotemporal feature vectors, obtaining the similarity SF of every two groups of historical fault spatiotemporal feature vectors i , if SF i ≥ threshold SF i th1 , then merging into one node, otherwise existing as an independent node, traversing C N 2 the similarity of the group historical fault spatiotemporal feature vectors, updating the fault feature node set V F to V F g ; traversing C based on the distance function N 2 group intelligent repair strategy nodes, obtaining the distance function value SR of every two groups of intelligent repair strategy text vectors i , if SR i < threshold SR i th , then merging into one node, otherwise existing as an independent node, traversing C N 2 the distance function value of the group intelligent repair strategy text vectors, updating the repair strategy node set V R to V R g , finally obtaining the updated fault diagnosis and intelligent repair knowledge graph READ3; S4.4: Calculate the real-time fault spatiotemporal feature vector F now The set of fault nodes V in the Fault Diagnosis and Intelligent Repair Knowledge Graph READ3 F g Each node F in N cosine similarity Sim(F) now ,F N ), filter out those with similarity ≥ threshold SF i th1 The faulty nodes are used as the candidate fault feature node set VF now Then, for candidate fault feature nodes F... N ∈VF now To the intelligent repair strategy node R N Given the path, obtain the set of path edge weights W. N (VF now The optimal repair strategy R is selected by filtering the repair strategy text corresponding to the path with the highest path edge weight. opt That is, the intelligent fault repair strategy; if the cosine similarity Sim(F now ,F N ) < threshold SF i th1 If so, the management terminal decision interface will be triggered; S4.5: If the real-time fault spatiotemporal feature vector F now With the set of fault characteristic nodes V F g node F in N cosine similarity Sim(F) now ,F N )≥threshold SF i th2 Threshold SF i th2 > Threshold SF i th1 Update the fault spatiotemporal feature vector of the node; otherwise, add a fault feature node and a corresponding intelligent repair strategy node. If the freshness fh of historical fault cases is less than the corresponding threshold fh... th If the faulty case is positive, it will be eliminated; otherwise, it will be retained. fh = f(t) case )×g(up),f(t) case ) is the base time decay function, and g(up) is the equipment upgrade correction term for the heavy forklift system; S5: Based on the obtained fault intelligent repair strategy, the corresponding intelligent repair instruction is issued through the CAN bus, and the execution result is monitored at the same time; S6: The execution result data after the repair operation is evaluated, and the evaluation abnormal result is fed back to the management terminal for human-computer interaction.

2. The method according to claim 1, wherein: The fault spatiotemporal feature vector in S3 is obtained: through the constructed spatiotemporal attention feature extraction network, the three-dimensional data matrix after the real-time collection and preprocessing operation is subjected to feature extraction to obtain a real-time fault spatiotemporal feature vector F now ; Then the real-time fault space-time feature vector F now is input into the full connection layer and the Softmax function to obtain a fault state label; if the real-time fault space-time feature vector F now corresponds to a fault-free state label, no subsequent operation is needed, otherwise, subsequent operation is performed.

3. The method of claim 1, wherein: The S6 implementation comprises the following steps: S6.1: Monitor the repair strategy execution results and system real-time state data through CAN bus, the repair strategy execution results data including the number of successful fault repair sc, the total repair number ta, the repair time t each time I , the first fault spatiotemporal feature vector F after repair I h , and the first fault spatiotemporal feature vector F under standard conditions I b , and calculate the execution result evaluation index η; S6.2: the execution result evaluation index η is less than the corresponding threshold value η th , then trigger fault diagnosis and intelligent repair knowledge graph READ3 incremental learning, READ3 new = READ3∪[(F now , F h )], F now and F h are real-time fault space-time feature vectors and repaired fault space-time feature vectors respectively, and feedback to the management terminal, otherwise it means that the execution result is good.

Images