CN114993710B - Remote control type electric-drive folding forklift hub rapid maintenance device and method thereof - Google Patents

Remote control type electric-drive folding forklift hub rapid maintenance device and method thereof Download PDF

Info

Publication number
CN114993710B
CN114993710B CN202210624763.8A CN202210624763A CN114993710B CN 114993710 B CN114993710 B CN 114993710B CN 202210624763 A CN202210624763 A CN 202210624763A CN 114993710 B CN114993710 B CN 114993710B
Authority
CN
China
Prior art keywords
feature
vector
neural network
feature vector
convolutional neural
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202210624763.8A
Other languages
Chinese (zh)
Other versions
CN114993710A (en
Inventor
戴肖肖
蒋连杰
张汉章
陈春喜
张建东
李博文
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang Jialift Warehouse Equipment Co ltd
Original Assignee
Zhejiang Jialift Warehouse Equipment Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang Jialift Warehouse Equipment Co ltd filed Critical Zhejiang Jialift Warehouse Equipment Co ltd
Priority to CN202210624763.8A priority Critical patent/CN114993710B/en
Publication of CN114993710A publication Critical patent/CN114993710A/en
Application granted granted Critical
Publication of CN114993710B publication Critical patent/CN114993710B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M17/00—Testing of vehicles
    • G01M17/007—Wheeled or endless-tracked vehicles
    • G01M17/013—Wheels
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/20—Administration of product repair or maintenance

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Human Resources & Organizations (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • Operations Research (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Economics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Quality & Reliability (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Forklifts And Lifting Vehicles (AREA)

Abstract

The application relates to the field of intelligent maintenance of forklift hubs, and particularly discloses a remote control type electric-drive folding forklift hub quick maintenance device and a method thereof.

Description

Remote control type electric-drive folding forklift hub rapid maintenance device and method thereof
Technical Field
The application relates to the field of intelligent detection of forklift hubs, in particular to a remote control type electric-drive folding forklift hub quick maintenance device and a method thereof.
Background
The main operation mode of the logistics warehouse is the transportation operation of loading and unloading goods by means of heavy mechanical equipment such as a forklift and a traveling crane. Frequent transport operations not only place a great demand on personnel, but also place great pressure on the maintenance of the transport equipment.
When the forklift loads and unloads goods, the forklift body jolts seriously in the transportation process due to the fact that the ground surface is uneven and the solid tires are arranged on the steering wheels of the forklift. Therefore, the plane supporting bearing above the steering knuckle is very easy to damage, and the damaged bearing cannot be found and replaced in time, and the reason is that the existing forklift generally uses a hydraulic power-assisted steering system, when the plane supporting bearing is worn or even collapsed, a forklift driver cannot feel the change of the steering wheel strength in the driving process, and when the steering knuckle and the steering main pin plane of the steering axle are worn to be locked, the forklift has a steering fault, and the damaged part can be found.
At present, the fault detection of the factory leaving the forklift can only depend on manual inspection, and the mode can not only be inaccurate, but also be influenced by human factors. Therefore, in order to accurately judge the fault type of the forklift hub intelligently and maintain the forklift hub more quickly, the remote control type electric driving folding forklift hub quick maintenance device is expected.
Disclosure of Invention
The present application is proposed to solve the above-mentioned technical problems. The embodiment of the application provides a remote control formula drives folding fork truck wheel hub quick maintenance device of electricity and method thereof, and it uses the intelligent detection technique based on artificial intelligence to come to the pulse signal who gathers when detection device is close to fork truck wheel hub and carries out the analysis, with accurately detecting fork truck wheel hub's fault type to maintain sooner, and then guarantee fork truck's normal operation avoids the emergence of accident.
According to an aspect of the application, a remote control formula electricity drives folding fork truck wheel hub quick maintenance device is provided, it includes: a training module comprising: the device comprises a pulse signal acquisition unit, a detection device and a control unit, wherein the pulse signal acquisition unit is used for acquiring a pulse signal acquired when the detection device approaches a forklift hub, and the detection device comprises a Hall magnetic proximity switch and a permanent magnet; a sampling unit for intercepting a plurality of sampling windows from the pulse signal along a timing dimension with a preset sampling window; a frequency domain feature extraction unit, configured to extract frequency domain features of each of the sampling windows from the plurality of sampling windows based on fourier transform; the time sequence coding unit is used for enabling the frequency domain characteristics of each sampling window to pass through a time sequence coder comprising a one-dimensional convolutional layer and a full-link layer so as to obtain a first characteristic vector; a first convolution unit, configured to pass the waveform map of the pulse signal through a depth separable convolutional neural network to obtain a first feature map, where the depth separable convolutional neural network is in a convolution operation, and filters of different layers are used to convolve input data in two spatial dimensions W and H and a channel dimension C, respectively; a second convolution unit to pass the first feature map through a filter-based convolutional neural network to obtain a second feature vector; a scale migration certainty loss function value calculation unit configured to calculate a scale migration certainty loss function value between the first feature vector and the second feature vector, the scale migration certainty loss function value being a natural exponent function value raised by a power of a quotient of a feature value obtained by multiplying the first feature vector by a transpose of the second feature vector, and a Frobenius norm of a feature matrix obtained by dividing a feature value obtained by multiplying the transpose of the first feature vector by the second feature vector; a joint encoding unit for calculating a transposed vector of the first eigenvector multiplied by the second eigenvector to obtain a classification eigenvector matrix; the classification loss function value calculation unit is used for enabling the classification characteristic matrix to pass through a classifier so as to obtain a classification loss function value; and a training unit to train the temporal encoder, the deep separable convolutional neural network, and the filter-based convolutional neural network with a weighted sum between the classification loss function value and the scale migration certainty loss function value as a loss function value; and an inference module comprising: the detection signal acquisition unit is used for acquiring a pulse signal acquired when the detection device approaches the forklift hub; an intercepting unit for intercepting a plurality of sampling windows from the pulse signal along a timing dimension with a preset sampling window; a statistical feature extraction unit, configured to extract frequency domain features of each of the sampling windows from the plurality of sampling windows based on fourier transform; the first coding unit is used for enabling the frequency domain characteristics of each sampling window to pass through the time sequence coder which is trained by the training module and comprises the one-dimensional convolutional layer and the full-connection layer so as to obtain a first characteristic vector; a second coding unit, configured to pass the waveform map of the pulse signal through the deep separable convolutional neural network trained by the training module to obtain a first feature map; a third encoding unit, configured to pass the first feature map through a filter-based convolutional neural network trained by the training module to obtain a second feature vector; a fourth encoding unit configured to calculate a transposed vector of the first eigenvector by multiplying the second eigenvector to obtain a classification eigenvector matrix; and the classification result generating unit is used for enabling the classification characteristic matrix to pass through a classifier to obtain a classification result, and the classification result is used for indicating whether the forklift hub has a fault or not.
According to another aspect of the application, a method for rapidly repairing a device of a remote control type electric drive folding forklift hub is provided, which comprises the following steps: a training phase comprising: acquiring a pulse signal acquired when a detection device approaches a forklift hub, wherein the detection device comprises a Hall magnetic proximity switch and a permanent magnet; intercepting a plurality of sampling windows from the pulse signal along a timing dimension with a preset sampling window; extracting frequency domain features of each of the sampling windows from the plurality of sampling windows based on a Fourier transform; passing the frequency domain features of each sampling window through a time sequence encoder comprising a one-dimensional convolutional layer and a full-link layer to obtain a first feature vector; passing the waveform map of the pulse signal through a depth separable convolutional neural network to obtain a first feature map, wherein the depth separable convolutional neural network is in a convolution operation, and filters of different layers are used for convolving input data in two spatial dimensions W and H and a channel dimension C respectively; passing the first feature map through a filter-based convolutional neural network to obtain a second feature vector; calculating a scale migration certainty loss function value between the first feature vector and the second feature vector, the scale migration certainty loss function value being a natural exponent function value raised to a power of a quotient of an eigenvalue obtained by multiplying the first feature vector by a transpose of the second feature vector, and a Frobenius norm of an eigenmatrix obtained by multiplying the transpose of the first feature vector by the second feature vector; computing a transposed vector of the first eigenvector multiplied by the second eigenvector to obtain a classification eigenvector matrix; passing the classification feature matrix through a classifier to obtain a classification loss function value; and training the time-series encoder, the deep separable convolutional neural network, and the filter-based convolutional neural network with a weighted sum between the classification loss function value and the scale migration deterministic loss function value as a loss function value; and an inference phase comprising: acquiring a pulse signal acquired when the detection device approaches the forklift hub; intercepting a plurality of sampling windows from the pulse signal along a timing dimension with a preset sampling window; extracting frequency domain features of each of the sampling windows from the plurality of sampling windows based on a fourier transform; passing the frequency domain features of each sampling window through the time sequence encoder which is trained by a training module and comprises the one-dimensional convolutional layer and the full-link layer to obtain a first feature vector; passing the waveform map of the pulse signal through the deep separable convolutional neural network completed by training through the training module to obtain a first feature map; passing the first feature map through a filter-based convolutional neural network trained by the training module to obtain a second feature vector; computing a transposed vector of the first eigenvector multiplied by the second eigenvector to obtain a classification eigenvector matrix; and enabling the classification characteristic matrix to pass through a classifier to obtain a classification result, wherein the classification result is used for indicating whether the forklift hub has a fault or not.
Compared with the prior art, the remote control type electrically-driven folding forklift hub rapid maintenance device and the method thereof can start from the fusion relevance characteristic of the time domain characteristic and the frequency domain characteristic through the pulse signal acquired when the detection device Hall magnetic proximity switch and the permanent magnet are close to the forklift hub, and timely and rapidly detect the fault type of the forklift hub so as to maintain more rapidly, thereby ensuring the normal operation of the forklift and avoiding the occurrence of accidents.
Drawings
The above and other objects, features and advantages of the present application will become more apparent by describing in more detail embodiments of the present application with reference to the attached drawings. The accompanying drawings are included to provide a further understanding of the embodiments of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings, like reference numbers generally represent like parts or steps.
Fig. 1 is a scene schematic diagram of a remote control type electrically-driven folding forklift hub quick maintenance device according to an embodiment of the application.
Fig. 2 is a block diagram of a remote control type electric drive folding forklift hub quick maintenance device according to an embodiment of the application.
Fig. 3A is a flowchart of a training phase in a method for quickly repairing a device on a hub of a remotely controlled electrically driven folding forklift according to an embodiment of the present application.
Fig. 3B is a flow chart of an inference phase in a method for quickly servicing a device for remotely controlling an electrically driven folding forklift hub according to an embodiment of the application.
Fig. 4 is a schematic configuration diagram of a training phase in a method for rapidly repairing a device for remotely controlling a hub of an electrically-driven folding forklift according to an embodiment of the application.
Fig. 5 is a schematic configuration diagram of an inference stage in a method for rapidly repairing a device for remotely controlling a hub of an electrically-driven folding forklift according to an embodiment of the application.
Detailed Description
Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. It should be understood that the described embodiments are only some embodiments of the present application and not all embodiments of the present application, and that the present application is not limited by the example embodiments described herein.
Overview of a scene
As described above, in order to improve the service life of the forklift and facilitate maintenance, it is desirable to perform analysis and diagnosis of the failure type of the hub of the forklift so as to perform early warning. It should be understood that there may be differences in "representation" layers of a forklift with a fault or a hidden fault, for example, if the detection device disclosed in patent No. CN 211468564 is used to approach the hub of the forklift, there may be differences in the pulse signals acquired by the detection device, that is, the fault type of the forklift can be classified and judged based on the acquired pulse signals and early-warning is performed to facilitate maintenance, which is essentially a classification problem, that is, the pulse signals acquired by the detection device are subjected to feature extraction and analysis based on a deep neural network model, and then the classifier is used to accurately classify and diagnose the fault of the hub, so as to ensure normal operation of the forklift and avoid accidents.
Specifically, in the technical scheme of this application, at first, through detection device, pulse signal is gathered when hall magnetism proximity switch and permanent magnet are close to fork truck wheel hub for example. In one specific example, the Hall magnetic proximity switch is horizontally arranged on one side of a rear axle of the forklift truck close to the steering knuckle, and the permanent magnet is arranged on one side of the steering knuckle close to the Hall magnetic proximity switch. The action principle is as follows: when the permanent magnet enters or exits the induction range of the Hall magnetic proximity switch, the switch correspondingly starts to trigger or stops triggering a series of pulse signals. Consequently, this application can be by hall magnetism proximity switch for the response change characteristic when the permanent magnet takes place vertical displacement come the wheel hub fault type of short-term test fork truck to in time maintain fork truck's wheel hub fast.
Because the pulse signal is continuous and the time domain is expanded infinitely, but an infinite number of signals cannot be processed in practice, the pulse signal needs to be converted into a signal which can be processed by a computer, and in order to better extract a local implicit characteristic in the frequency domain, a characteristic which is not obvious in the time domain is more obvious in the frequency domain, for example, because a noise signal and the pulse signal are mixed together, recognition of noise in the time domain is difficult, and if a high frequency appears in the frequency domain, the noise signal is a noise signal, so that the characteristic can be distinguished more easily in the frequency domain by the noise signal obviously. Therefore, in order to improve the accuracy of the fault diagnosis of the forklift hub, a plurality of sampling windows are further intercepted from the pulse signal along a time sequence dimension by a preset sampling window, and the frequency domain characteristics of each sampling window are extracted from the plurality of sampling windows based on Fourier transform. In this way, the frequency domain features of each sampling window can be subjected to feature extraction in a time sequence encoder comprising a one-dimensional convolutional layer and a fully-connected layer to obtain a first feature vector. Therefore, the implicit characteristics of the frequency domain characteristics of each sampling window can be extracted, and the relevance implicit characteristics among the frequency domain characteristics of each sampling window can also be extracted, so that the frequency domain characteristics are not regarded as independent frequency domain numerical characteristics, and the essential properties of frequency domain data can be reflected better.
After the frequency domain data is subjected to feature extraction, the time domain data needs to be further processed to integrate feature information of the frequency domain data and the time domain data for accurate classification. Specifically, first, a waveform map of the pulse signal is passed through a depth separable convolutional neural network to obtain a first feature map. Here, the deep separable convolutional neural network is in a convolution operation, with different layers of filters used to convolve the input data in two spatial dimensions W and H and a channel dimension C, respectively. It should be understood that the pulse signal may have noise during the acquisition process, and therefore, in order to eliminate the noise, the deep separable convolutional neural network is used to denoise the noise, which can be beneficial to improving the subsequent classification accuracy. That is, in the convolution operation, filters of different layers are used to perform convolution on two spatial dimensions W and H and a channel dimension C, respectively, that is, the convolution operation is performed not only on the space of the image dimension but also on the channel dimension, so that a three-dimensional block structure in a high-dimensional feature can be mined, thereby performing denoising of an original image based on the principle of three-dimensional block-matching and filtering (3D block-matching and filtering).
And then, processing the obtained first feature map in a convolutional neural network based on a filter to filter out useless features, and further excavating more needed hidden feature information to obtain a second feature vector. Here, the formula of the filter-based convolutional neural network is expressed as:
f i ＝Sigmoid(N i ×f i-1 +B i )
wherein, f i-1 Is the input of the i-th convolutional neural network, f i Is the output of the ith convolutional neural network, N i A filter which is the ith convolutional neural network, and B i Sigmoid represents a nonlinear activation function, which is a bias matrix of the i-th layer convolutional neural network.
Further, consider the first feature vector V 1 Representing the time-domain correlation of frequency-domain features, and a second feature vector V 2 Is a local correlation characterization of pixel-level feature representation in time domain, in order to make the probability distribution representation of the two in a high-dimensional feature space consistent, a scale migration certainty loss function is introduced,expressed as:
Figure BDA0003676531740000061
||·|| F the Frobenius norm of the matrix is represented. Wherein the first feature vector V 1 And a second eigenvector V 2 Are all in the form of row vectors.
It should be appreciated that the scale migration deterministic loss function preserves the long range (long range) relationship under scale migration of high-dimensional features through the embedding of relative positions between feature vectors, by training the model with it, it is possible to make the first feature vector V 1 And a second eigenvector V 2 To a certain extent, ensures the consistency of the probability distribution in the high-dimensional feature space.
Therefore, after the time sequence encoder, the depth separable convolutional neural network and the filter-based convolutional neural network are trained, the convolutional neural network can be used in an actual inference stage to more accurately judge the fault type of the forklift hub, and then the forklift hub can be timely and rapidly maintained to eliminate potential safety hazards.
Based on this, this application has provided a remote control formula electricity drives folding fork truck wheel hub quick maintenance device, and it includes training module and inference module. Wherein, the training module includes: the device comprises a pulse signal acquisition unit, a detection device and a control unit, wherein the pulse signal acquisition unit is used for acquiring a pulse signal acquired when the detection device approaches a forklift hub, and the detection device comprises a Hall magnetic proximity switch and a permanent magnet; a sampling unit for intercepting a plurality of sampling windows from the pulse signal along a timing dimension with a preset sampling window; a frequency domain feature extraction unit, configured to extract frequency domain features of each of the sampling windows from the plurality of sampling windows based on fourier transform; the time sequence coding unit is used for enabling the frequency domain characteristics of each sampling window to pass through a time sequence coder comprising a one-dimensional convolution layer and a full-link layer so as to obtain a first characteristic vector; a first convolution unit, configured to pass the waveform map of the pulse signal through a depth separable convolutional neural network to obtain a first feature map, where the depth separable convolutional neural network is in a convolution operation, and filters of different layers are used to convolve input data in two spatial dimensions W and H and a channel dimension C, respectively; a second convolution unit for passing the first feature map through a filter-based convolutional neural network to obtain a second feature vector; a scale migration certainty loss function value calculation unit configured to calculate a scale migration certainty loss function value between the first feature vector and the second feature vector, the scale migration certainty loss function value being a natural index function value raised by a power of a quotient of an eigenvalue obtained by dividing an eigenvalue obtained by multiplying the first feature vector by a transpose of the second feature vector by a Frobenius norm of an eigen matrix obtained by multiplying the transpose of the first feature vector by the second feature vector; a joint encoding unit for calculating a transposed vector of the first eigenvector multiplied by the second eigenvector to obtain a classification eigenvector matrix; the classification loss function value calculation unit is used for enabling the classification characteristic matrix to pass through a classifier so as to obtain a classification loss function value; and a training unit to train the temporal encoder, the deep separable convolutional neural network, and the filter-based convolutional neural network with a weighted sum between the classification loss function value and the scale migration deterministic loss function value as a loss function value. Wherein, the inference module comprises: the detection signal acquisition unit is used for acquiring pulse signals acquired when the detection device approaches the forklift hub; an intercepting unit for intercepting a plurality of sampling windows from the pulse signal along a timing dimension with a preset sampling window; a statistical feature extraction unit, configured to extract frequency domain features of each of the sampling windows from the plurality of sampling windows based on fourier transform; the first coding unit is used for enabling the frequency domain characteristics of each sampling window to pass through the time sequence coder which is trained by the training module and comprises the one-dimensional convolutional layer and the full-connection layer so as to obtain a first characteristic vector; a second coding unit, configured to pass the waveform map of the pulse signal through the deep separable convolutional neural network trained by the training module to obtain a first feature map; a third encoding unit, configured to pass the first feature map through a filter-based convolutional neural network trained by the training module to obtain a second feature vector; a fourth encoding unit configured to calculate a transposed vector of the first eigenvector by multiplying the second eigenvector to obtain a classification eigenvector matrix; and the classification result generating unit is used for enabling the classification characteristic matrix to pass through a classifier to obtain a classification result, and the classification result is used for indicating whether the forklift hub has a fault or not.
Fig. 1 illustrates a scene schematic diagram of a remote-control electric-driven folding forklift hub quick maintenance device according to an embodiment of the application. As shown in fig. 1, in the training phase of the application scenario, first, pulse signals of a hall magnetic proximity switch (e.g., T1 as illustrated in fig. 1) and a permanent magnet (e.g., T2 as illustrated in fig. 1) of a detection device (e.g., D as illustrated in fig. 1) are acquired when approaching a forklift hub (e.g., F as illustrated in fig. 1). Then, the obtained pulse signal is inputted into a server (for example, S as illustrated in fig. 1) deployed with a remote electric-driven folding forklift hub quick repair algorithm, wherein the server is capable of training the time-sequential encoder, the deep separable convolutional neural network, and the filter-based convolutional neural network of the remote electric-driven folding forklift hub quick repair device with the pulse signal based on the remote electric-driven folding forklift hub quick repair algorithm.
After training is completed, in the inference phase, first, pulse signals of a hall magnetic proximity switch (e.g., T1 as illustrated in fig. 1) and a permanent magnet (e.g., T2 as illustrated in fig. 1) of a detection device (e.g., D as illustrated in fig. 1) are acquired when approaching a forklift hub (e.g., F as illustrated in fig. 1). The pulse signal is then input into a server (e.g., S as illustrated in fig. 1) deployed with a remote electric-driven folding forklift hub quick repair algorithm, wherein the server is capable of processing the pulse signal with the remote electric-driven folding forklift hub quick repair algorithm to generate a classification result indicating whether the forklift hub is faulty or not. And then, based on the classification result of trouble right fork truck wheel hub carries out the fault classification diagnosis to in time maintain fast, guaranteed fork truck normal operating's security.
Having described the general principles of the present application, various non-limiting embodiments of the present application will now be described with reference to the accompanying drawings.
Exemplary System
Fig. 2 illustrates a block diagram of a remote controlled electric drive folding forklift hub quick service device according to an embodiment of the application. As shown in fig. 2, the device 200 for quickly repairing a hub of a remotely-controlled electrically-driven folding forklift according to an embodiment of the present application includes: a training module 210 and an inference module 220. Wherein, the training module 210 includes: the pulse signal acquisition unit 2101 is used for acquiring pulse signals acquired when a detection device is close to a forklift hub, and the detection device comprises a Hall magnetic proximity switch and a permanent magnet; a sampling unit 2102 configured to intercept a plurality of sampling windows from the pulse signal along a timing dimension with a preset sampling window; a frequency domain feature extraction unit 2103, configured to extract frequency domain features of each of the sampling windows from the plurality of sampling windows based on fourier transform; a time sequence coding unit 2104 for passing the frequency domain features of each of the sampling windows through a time sequence coder including a one-dimensional convolutional layer and a fully-connected layer to obtain a first feature vector; a first convolution unit 2105 configured to pass the waveform map of the pulse signal through a depth separable convolutional neural network to obtain a first feature map, wherein the depth separable convolutional neural network is configured to perform convolution operations on input data in two spatial dimensions W and H and a channel dimension C by using filters of different layers respectively; a second convolution unit 2106 for passing the first feature map through a filter-based convolutional neural network to obtain a second feature vector; a scale migration certainty loss function value calculation unit 2107 configured to calculate a scale migration certainty loss function value between the first feature vector and the second feature vector, the scale migration certainty loss function value being a natural index function value having a power of a quotient of a feature value obtained by multiplying the first feature vector by a transpose of the second feature vector, and a Frobenius norm of a feature matrix obtained by dividing a feature value obtained by multiplying the transpose of the first feature vector by the second feature vector; a joint encoding unit 2108 for calculating a transposed vector of the first eigenvector multiplied by the second eigenvector to obtain a classification eigenvector matrix; a classification loss function value calculation unit 2109, configured to pass the classification feature matrix through a classifier to obtain a classification loss function value; and a training unit 2110 for training the time-series encoder, the deep separable convolutional neural network, and the filter-based convolutional neural network with a weighted sum between the classification loss function value and the scale migration deterministic loss function value as a loss function value. The inference module 220 includes: a detection signal acquisition unit 221 configured to acquire a pulse signal acquired when the detection device approaches the forklift hub; a clipping unit 222 for clipping a plurality of sampling windows from the pulse signal along a timing dimension with a preset sampling window; a statistical feature extraction unit 223, configured to extract frequency domain features of each of the sampling windows from the plurality of sampling windows based on fourier transform; a first encoding unit 224, configured to pass the frequency domain features of each sampling window through the time-series encoder that is trained by the training module and includes the one-dimensional convolutional layer and the fully-connected layer to obtain a first feature vector; a second encoding unit 225, configured to pass the waveform map of the pulse signal through the deep separable convolutional neural network trained by the training module to obtain a first feature map; a third encoding unit 226, configured to pass the first feature map through a filter-based convolutional neural network trained by the training module to obtain a second feature vector; a fourth encoding unit 227, configured to calculate a transposed vector of the first feature vector by multiplying the second feature vector to obtain a classification feature matrix; and a classification result generating unit 228, configured to pass the classification feature matrix through a classifier to obtain a classification result, where the classification result is used to indicate whether the forklift hub has a fault.
Specifically, in the embodiment of the present application, in the training module 210, the pulse signal acquisition unit 2101 and the sampling unit 2102 are configured to acquire a pulse signal acquired when the detection device is close to a hub of a forklift, where the detection device includes a hall magnetic proximity switch and a permanent magnet. As mentioned above, it should be understood that when the existing forklift is used for loading and unloading goods, due to the uneven ground surface, and the solid tires mounted on the steering wheels of the forklift, the vehicle body jolts seriously during transportation, so that the plane support bearing above the steering knuckle is very easy to damage, and the damaged bearing cannot be found and replaced in time. Therefore, the analysis and diagnosis of the fault type of the hub of the forklift are needed, and this is essentially a classification problem, that is, the pulse signal acquired by the detection device is subjected to feature extraction and analysis based on a deep neural network model, and then the classifier is used for accurately classifying and diagnosing the fault of the hub, so as to ensure the normal operation of the forklift and avoid the occurrence of accidents.
That is, specifically, in the technical solution of the present application, first, a pulse signal thereof is acquired while approaching the forklift hub by a detection device, such as a hall magnetic proximity switch and a permanent magnet. In a specific example, the Hall magnetic proximity switch is horizontally arranged on one side of a rear axle of the forklift near a steering knuckle, and the permanent magnet is arranged on one side of the steering knuckle near the Hall magnetic proximity switch. The action principle is as follows: when the permanent magnet enters or exits the induction range of the Hall magnetic proximity switch, the switch correspondingly starts to trigger or stops triggering a series of pulse signals. Consequently, this application can by hall magnetism proximity switch for response change characteristic when the permanent magnet takes place vertical displacement comes short-term test fork truck's wheel hub fault type to in time maintain fork truck's wheel hub fast.
Specifically, in the embodiment of the present application, in the training module 210, the sampling unit 2102, the frequency domain feature extraction unit 2103, and the time sequence encoding unit 2104 are configured to intercept a plurality of sampling windows from the pulse signal along a time sequence dimension with a preset sampling window, extract frequency domain features of each of the sampling windows from the plurality of sampling windows based on fourier transform, and pass the frequency domain features of each of the sampling windows through a time sequence encoder including a one-dimensional convolutional layer and a fully-connected layer to obtain a first feature vector. It should be understood that, since the pulse signal is continuous and the time domain is infinitely expanded, but in practice, an infinite number of signals cannot be processed, the pulse signal needs to be converted into a signal which can be processed by a computer, and in order to better extract a local implicit feature in the frequency domain, a feature which is not obvious in the time domain is more obvious in the frequency domain, for example, because a noise signal and the pulse signal are mixed together, identification of noise is difficult in the time domain, and if a high frequency occurs in the frequency domain, the noise signal is a noise signal, so that the noise signal is obviously easier to distinguish the feature in the frequency domain.
Therefore, in the technical scheme of the application, in order to improve the accuracy of the fault diagnosis of the forklift hub, a plurality of sampling windows are further intercepted from a pulse signal along a time sequence dimension through a preset sampling window, and frequency domain features of the sampling windows are extracted from the sampling windows on the basis of Fourier transform, so that time domain features in the sampling windows are converted into a feature space of a frequency domain. In this way, the frequency domain features of each sampling window can be subjected to time-dimension feature extraction in a time-sequence encoder comprising a one-dimensional convolutional layer and a full-link layer to obtain a first feature vector. It should be understood that, in this way, not only the implicit features of the frequency domain features of each sampling window can be extracted, but also the relevance implicit features between the frequency domain features of each sampling window can be extracted, so that the implicit features are not regarded as independent frequency domain numerical features, and the essential properties of the frequency domain data can be reflected better.
More specifically, in an embodiment of the present application, the time-series encoding unit includes: the input vector arrangement subunit is used for arranging the frequency domain characteristics of each sampling window into a one-dimensional input vector; a full-connection subunit, configured to perform full-connection coding on the input vector obtained by the input vector arrangement subunit by using a full-connection layer of the time sequence encoder according to a following formula to extract a high-dimensional implicit feature of a feature value at each position in the input vector, where the formula is:
Figure BDA0003676531740000111
wherein X is the input vector, Y is the output vector, W is the weight matrix, B is the offset vector,
Figure BDA0003676531740000112
represents a matrix multiplication; a one-dimensional convolution subunit, configured to perform one-dimensional convolution encoding on the input vector obtained by the input vector arrangement subunit by using a one-dimensional convolution layer of the time-series encoder according to the following formula to extract a high-dimensional implicit correlation feature between feature values of each position in the input vector, where the formula is:
Figure BDA0003676531740000113
wherein, a is the width of the convolution kernel in the x direction, F is the parameter vector of the convolution kernel, G is the local vector matrix operated with the convolution kernel function, and w is the size of the convolution kernel.
Specifically, in the embodiment of the present application, in the training module 210, the first convolution unit 2105 and the second convolution unit 2106 are configured to pass the waveform map of the pulse signal through a deep separable convolutional neural network to obtain a first feature map, where the deep separable convolutional neural network is configured to convolve input data in two spatial dimensions W and H and a channel dimension C, respectively, with filters of different layers, and pass the first feature map through a filter-based convolutional neural network to obtain a second feature vector. That is, after the feature extraction is performed on the frequency domain data, the time domain data needs to be further processed to integrate feature information of the frequency domain data and the time domain data for accurate classification. Specifically, in the technical solution of the present application, first, a waveform diagram of the pulse signal is passed through a deep separable convolutional neural network to obtain a first feature diagram. Here, the deep separable convolutional neural network is in a convolution operation, filters of different layers are used to convolve input data in two spatial dimensions W and H and a channel dimension C, respectively. It should be understood that the pulse signal may have noise during the acquisition process, and therefore, in order to eliminate the noise, the deep separable convolutional neural network is used to denoise the pulse signal, which can be beneficial to improving the subsequent classification accuracy. That is, in the convolution operation, filters of different layers are used to perform convolution on two spatial dimensions W and H and a channel dimension C, respectively, that is, the convolution operation is performed not only on the space of the image dimension but also on the channel dimension, so that a three-dimensional block structure in a high-dimensional feature can be mined, thereby performing denoising of an original image based on the principle of three-dimensional block matching and filtering (3D block-matching and filtering).
And then, processing the obtained first feature map in a convolutional neural network based on a filter to filter out useless features, and further excavating more needed hidden feature information for classification to obtain a second feature vector.
More specifically, in this embodiment of the present application, the second convolution unit is further configured to: processing the first feature map using the convolutional filter-based neural network with the following formula to generate the second feature vector;
wherein the formula is:
f i ＝Sigmoid(N i ×f i-1 +B i )
wherein f is i-1 As input to the ith convolutional neural network, f i Is the output of the ith convolutional neural network, N i A filter which is the ith convolutional neural network, and B i Sigmoid represents the nonlinear activation function for the bias matrix of the ith convolutional neural network.
Specifically, in this embodiment of the application, in the training module 210, the scale migration certainty loss function value calculation unit 2107 is configured to calculate a scale migration certainty loss function value between the first feature vector and the second feature vector, where the scale migration certainty loss function value is obtained by dividing a feature value obtained by multiplying the first feature vector by a transpose of the second feature vector by a transpose of the first feature vector and multiplying the second feature vectorThe quotient of the Frobenius norm of the feature matrix is the natural exponential function value of the power. It should be understood that the first eigenvector V is considered 1 Representing a time-domain correlation of frequency-domain features, and said second feature vector V 2 The method is a local correlation characterization of a pixel-level feature representation of a time domain, and in order to make probability distribution representations of the two in a high-dimensional feature space tend to be consistent, in the technical scheme of the application, a scale migration deterministic loss function is further introduced to train a deep neural network. It should be appreciated that the scale migration deterministic loss function preserves the long range (long range) relationship under scale migration of high-dimensional features by embedding relative positions between the feature vectors, with which the model can be trained to make the first feature vector V 1 And said second feature vector V 2 The feature expression of (2) ensures the consistency of probability distribution in a high-dimensional feature space to a certain extent, and further improves the accuracy of subsequent fault classification.
More specifically, in an embodiment of the present application, the scale migration deterministic loss function value calculating unit is further configured to: calculating the scale migration certainty loss function value between the first feature vector and the second feature vector in the following formula;
wherein the formula is:
Figure BDA0003676531740000131
wherein V 1 Representing said first feature vector, V 2 Representing said second feature vector, V 1 And V 2 Are all in the form of row vectors, and | · | | non-calculation F The Frobenius norm of the matrix is represented.
Specifically, in the embodiment of the present application, in the training module 210, the joint encoding unit 2108, the classification loss function value calculating unit 2109 and the training unit 2110 are configured to calculate a transposed vector of the first feature vector to be multiplied by the second feature vector to obtain a classification feature matrix, and pass the classification feature matrix through a classifier to obtain a classification loss functionValues, and training the time series encoder, the deep separable convolutional neural network, and the filter-based convolutional neural network with a weighted sum between the classification loss function values and the scale migration deterministic loss function values as loss function values. That is, in the technical solution of the present application, in order to train the deep neural network model to make the fault diagnosis accuracy for the forklift hub higher, a classification feature matrix having associated feature information distribution between the first feature vector and the second feature vector is further calculated to be processed by a classifier, so as to obtain a classification loss function value. Accordingly, in one specific example, the classifier processes the classification feature matrix to generate a classification result according to the following formula: softmax { (W) n ,B n ):…:(W 1 ,B 1 ) L Project (F) }, where Project (F) denotes the projection of the classification feature matrix as a vector, W 1 To W n As a weight matrix for each fully connected layer, B 1 To B n A bias matrix representing the layers of the fully-connected layer; and calculating a cross entropy value between the classification result and a real value as the classification loss function value. The time series encoder, the deep separable convolutional neural network, and the filter-based convolutional neural network may then be further trained with a weighted sum between the classification loss function value and the scale migration deterministic loss function value as a loss function value.
More specifically, in an embodiment of the present application, the joint encoding unit is further configured to: calculating the classification feature matrix based on the first feature vector and the second feature vector by the following formula;
wherein the formula is:
Figure BDA0003676531740000132
wherein
Figure BDA0003676531740000133
Representing multiplication of vectors, M representing the classification feature matrix, V 1 Representing said first feature vector, V 2 Representing the second feature vector in the second set of feature vectors,
Figure BDA0003676531740000134
a transposed vector representing the first feature vector.
After training is complete, the inference module is engaged, i.e., after training the timing encoder, the deep separable convolutional neural network, and the filter-based convolutional neural network, the trained timing encoder, the deep separable convolutional neural network, and the filter-based convolutional neural network are used in actual inference. And then, a classification result used for representing whether the forklift hub has a fault can be obtained, so that the fault type of the forklift hub can be more accurately judged, and the forklift hub can be timely and quickly maintained to eliminate potential safety hazards.
Specifically, in the embodiment of the present application, first, a pulse signal acquired when the detection device approaches the forklift hub is acquired. Next, a plurality of sampling windows are truncated from the pulse signal along a timing dimension with a preset sampling window. And then extracting the frequency domain features of each sampling window from the plurality of sampling windows based on Fourier transform. Then, the frequency domain features of each sampling window are passed through the time sequence encoder which is trained by a training module and comprises a one-dimensional convolutional layer and a full-link layer to obtain a first feature vector. Then, the waveform diagram of the pulse signal is passed through the deep separable convolutional neural network which is trained by the training module to obtain a first feature diagram. Then, the first feature map is passed through a filter-based convolutional neural network trained by the training module to obtain a second feature vector. Then, a transposed vector of the first feature vector is multiplied by the second feature vector to obtain a classification feature matrix. And finally, enabling the classification characteristic matrix to pass through a classifier to obtain a classification result, wherein the classification result is used for representing whether the forklift hub has a fault or not.
In conclusion, based on this application embodiment the quick maintenance device 200 of remote control formula electricity drive folding fork truck wheel hub is elucidated, and it can come from the fusion relevance characteristic of time domain characteristic and frequency domain characteristic through the pulse signal that detection device hall magnetism proximity switch and permanent magnet gathered when being close to fork truck wheel hub, in time detects fast fork truck wheel hub's fault type to maintain more fast, and then ensure fork truck's normal operation, avoid the emergence of accident.
As described above, the remote control type electric-driven folding forklift hub quick repair device 200 according to the embodiment of the present application can be implemented in various terminal devices, such as a server of a remote control type electric-driven folding forklift hub quick repair algorithm, and the like. In one example, the remote control type electric drive folding forklift hub rapid maintenance device 200 according to the embodiment of the application can be integrated into a terminal device as a software module and/or a hardware module. For example, the remote control type electric drive folding forklift hub rapid maintenance device 200 can be a software module in an operating system of the terminal device, or can be an application program developed for the terminal device; of course, the remote-controlled electric-driven folding forklift hub quick-repair device 200 can also be one of the hardware modules of the terminal equipment.
Alternatively, in another example, the remote controlled electric powered folding forklift hub quick repair device 200 and the terminal device may be separate devices, and the remote controlled electric powered folding forklift hub quick repair device 200 may be connected to the terminal device through a wired and/or wireless network, and transmit interactive information according to an agreed data format.
Exemplary method
Fig. 3A illustrates a flow chart of a training phase in a method of remotely controlling an electrically driven folding forklift hub quick service device according to an embodiment of the application. As shown in fig. 3A, a method for rapidly repairing a device of a remote-controlled electrically-driven foldable forklift hub according to an embodiment of the present application includes: a training phase comprising the steps of: s110, acquiring a pulse signal acquired when a detection device approaches a forklift hub, wherein the detection device comprises a Hall magnetic proximity switch and a permanent magnet; s120, intercepting a plurality of sampling windows from the pulse signal along a time sequence dimension by using a preset sampling window; s130, extracting frequency domain characteristics of each sampling window from the plurality of sampling windows based on Fourier transform; s140, passing the frequency domain characteristics of each sampling window through a time sequence encoder comprising a one-dimensional convolution layer and a full-connection layer to obtain a first characteristic vector; s150, passing the waveform diagram of the pulse signal through a depth separable convolutional neural network to obtain a first characteristic diagram, wherein in the convolution operation of the depth separable convolutional neural network, filters of different layers are used for performing convolution on input data in two spatial dimensions W and H and a channel dimension C respectively; s160, passing the first feature map through a filter-based convolutional neural network to obtain a second feature vector; s170, calculating a scale migration certainty loss function value between the first feature vector and the second feature vector, the scale migration certainty loss function value being a natural exponent function value raised by a power of a quotient of a feature value obtained by multiplying the first feature vector by a transpose of the second feature vector, and a Frobenius norm of a feature matrix obtained by multiplying the transpose of the first feature vector by the second feature vector; s180, calculating a transposed vector of the first feature vector and multiplying the second feature vector to obtain a classification feature matrix; s190, enabling the classification characteristic matrix to pass through a classifier to obtain a classification loss function value; and S200, training the time series encoder, the deep separable convolutional neural network, and the filter-based convolutional neural network with a weighted sum between the classification loss function value and the scale migration certainty loss function value as a loss function value.
Fig. 3B illustrates a flow chart of an inference phase in a method for quickly servicing a device for remotely controlled electrically powered folding forklift hubs in accordance with an embodiment of the present application. Fig. 3B illustrates a method for quickly repairing a device on a hub of a remotely controlled electrically powered folding forklift according to an embodiment of the present application, including: an inference phase comprising the steps of: s210, acquiring a pulse signal acquired when the detection device approaches the forklift hub; s220, intercepting a plurality of sampling windows from the pulse signal along a time sequence dimension by using a preset sampling window; s230, extracting frequency domain characteristics of each sampling window from the plurality of sampling windows based on Fourier transform; s240, the frequency domain characteristics of each sampling window pass through the time sequence encoder which is trained by a training module and comprises the one-dimensional convolutional layer and the full connection layer to obtain a first characteristic vector; s250, passing the oscillogram of the pulse signal through the deep separable convolutional neural network trained by the training module to obtain a first feature map; s260, passing the first feature map through a filter-based convolutional neural network trained by the training module to obtain a second feature vector; s270, calculating a transposed vector of the first feature vector and multiplying the second feature vector to obtain a classification feature matrix; and S280, enabling the classification characteristic matrix to pass through a classifier to obtain a classification result, wherein the classification result is used for indicating whether the forklift hub has a fault or not.
Fig. 4 illustrates an architecture diagram of a training phase in a method for rapidly repairing a device for remotely controlling a hub of an electrically driven folding forklift according to an embodiment of the present application. As shown in fig. 4, in the training phase, in the network architecture, first, a plurality of sampling windows (e.g., P1 as illustrated in fig. 4) are intercepted from the pulse signal (e.g., P as illustrated in fig. 4) along a timing dimension by a preset sampling window; then, extracting frequency domain features of each of the sampling windows from the plurality of sampling windows based on fourier transform (e.g., P2 as illustrated in fig. 4); then, passing the frequency domain features of each of the sampling windows through a time-sequence encoder (e.g., E as illustrated in fig. 4) containing one-dimensional convolutional layers and fully-connected layers to obtain a first feature vector (e.g., VF1 as illustrated in fig. 4); next, passing a waveform map (e.g., Q as illustrated in fig. 4) of the pulse signal through a deep separable convolutional neural network (e.g., CNN1 as illustrated in fig. 4) to obtain a first feature map (e.g., F1 as illustrated in fig. 4); then, passing the first feature map through a filter-based convolutional neural network (e.g., CNN2 as illustrated in fig. 4) to obtain a second feature vector (e.g., VF2 as illustrated in fig. 4); then, calculating a scale-migration deterministic loss function value (e.g., SLV as illustrated in fig. 4) between the first and second feature vectors; then, a transposed vector of the first feature vector is multiplied by the second feature vector to obtain a classification feature matrix (e.g., MF as illustrated in fig. 4); then, passing the classification feature matrix through a classifier (e.g., a classifier as illustrated in fig. 4) to obtain a classification loss function value (e.g., a CLV as illustrated in fig. 4); and, finally, training the time-series encoder, the deep separable convolutional neural network, and the filter-based convolutional neural network with a weighted sum between the classification loss function values and the scale migration deterministic loss function values as loss function values (e.g., LV as illustrated in fig. 4).
Fig. 5 illustrates an architecture diagram of an inference phase in a method for rapidly repairing a device for remotely controlling a hub of an electrically driven folding forklift according to an embodiment of the application. As shown in fig. 5, in the inference phase, in the network architecture, first, a plurality of sampling windows (e.g., P1 as illustrated in fig. 5) are intercepted from the pulse signal (e.g., P as illustrated in fig. 5) along a timing dimension with a preset sampling window; then, extracting frequency domain features (e.g., P2 as illustrated in fig. 5) of each of the sampling windows from the plurality of sampling windows based on fourier transform; then, passing the frequency domain features of each of the sampling windows through the time-sequence encoder (e.g., E0 as illustrated in fig. 5) containing one-dimensional convolutional layers and fully-connected layers trained by a training module to obtain a first feature vector (e.g., VF1 as illustrated in fig. 5); then, passing a waveform map (e.g., Q as illustrated in fig. 5) of the pulse signal through the deep separable convolutional neural network (e.g., CN1 as illustrated in fig. 5) completed by training by the training module to obtain a first feature map (e.g., F1 as illustrated in fig. 5); then, passing the first feature map through a filter-based convolutional neural network (e.g., CN2 as illustrated in fig. 5) trained by the training module to obtain a second feature vector (e.g., VF2 as illustrated in fig. 5); then, a transposed vector of the first eigenvector is calculated multiplied by the second eigenvector to obtain a classification feature matrix (e.g., MF as illustrated in fig. 5); and, finally, passing the classification feature matrix through a classifier (e.g., a classifier as illustrated in fig. 5) to obtain a classification result, which is used to indicate whether the forklift hub has a fault.
In summary, the method for rapidly maintaining the remote control type electrically-driven folding forklift hub is clarified, and the fault type of the forklift hub can be rapidly detected in time from the fusion relevance characteristic of the time domain characteristic and the frequency domain characteristic through the pulse signal acquired when the detection device hall magnetic proximity switch and the permanent magnet are close to the forklift hub, so that the maintenance can be performed more rapidly, the normal operation of the forklift is further ensured, and the occurrence of accidents is avoided.
The basic principles of the present application have been described above with reference to specific embodiments, but it should be noted that advantages, effects, etc. mentioned in the present application are only examples and are not limiting, and the advantages, effects, etc. must not be considered to be possessed by various embodiments of the present application. Furthermore, the foregoing disclosure of specific details is for the purpose of illustration and description and is not intended to be limiting, since the foregoing disclosure is not intended to be exhaustive or to limit the disclosure to the precise details disclosed.
The block diagrams of devices, apparatuses, systems referred to in this application are only given as illustrative examples and are not intended to require or imply that the connections, arrangements, configurations, etc. must be made in the manner shown in the block diagrams. These devices, apparatuses, devices, systems may be connected, arranged, configured in any manner, as will be appreciated by those skilled in the art. Words such as "including," "comprising," "having," and the like are open-ended words that mean "including, but not limited to," and are used interchangeably therewith. As used herein, the words "or" and "refer to, and are used interchangeably with, the word" and/or, "unless the context clearly dictates otherwise. The word "such as" is used herein to mean, and is used interchangeably with, the phrase "such as but not limited to".
It should also be noted that in the devices, apparatuses, and methods of the present application, each component or step can be decomposed and/or re-combined. These decompositions and/or recombinations should be considered as equivalents of the present application.
The previous description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other aspects without departing from the scope of the application. Thus, the present application is not intended to be limited to the aspects shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (10)

1. The utility model provides a remote control formula electricity drives folding fork truck wheel hub quick maintenance device which characterized in that includes:
a training module comprising:
the device comprises a pulse signal acquisition unit, a detection device and a control unit, wherein the pulse signal acquisition unit is used for acquiring a pulse signal acquired when the detection device approaches a forklift hub, and the detection device comprises a Hall magnetic proximity switch and a permanent magnet;
a sampling unit for intercepting a plurality of sampling windows from the pulse signal along a timing dimension with a preset sampling window;
a frequency domain feature extraction unit, configured to extract frequency domain features of each of the sampling windows from the plurality of sampling windows based on fourier transform;
the time sequence coding unit is used for enabling the frequency domain characteristics of each sampling window to pass through a time sequence coder comprising a one-dimensional convolution layer and a full-link layer so as to obtain a first characteristic vector;
a first convolution unit, configured to pass a waveform diagram of the pulse signal through a deep separable convolutional neural network to obtain a first feature diagram, where the deep separable convolutional neural network is in a convolution operation, and filters of different layers are used to convolve input data in two spatial dimensions W and H and a channel dimension C, respectively;
a second convolution unit to pass the first feature map through a filter-based convolutional neural network to obtain a second feature vector;
a scale migration certainty loss function value calculation unit configured to calculate a scale migration certainty loss function value between the first feature vector and the second feature vector, the scale migration certainty loss function value being a natural exponent function value raised by a power of a quotient of a feature value obtained by multiplying the first feature vector by a transpose of the second feature vector, and a Frobenius norm of a feature matrix obtained by dividing a feature value obtained by multiplying the transpose of the first feature vector by the second feature vector;
a joint encoding unit for calculating a transposed vector of the first eigenvector multiplied by the second eigenvector to obtain a classification eigenvector matrix;
the classification loss function value calculation unit is used for enabling the classification characteristic matrix to pass through a classifier so as to obtain a classification loss function value; and
a training unit to train the temporal encoder, the deep separable convolutional neural network, and the filter-based convolutional neural network with a weighted sum between the classification loss function value and the scale migration certainty loss function value as a loss function value; and
an inference module comprising:
the detection signal acquisition unit is used for acquiring pulse signals acquired when the detection device approaches the forklift hub;
an intercepting unit for intercepting a plurality of sampling windows from the pulse signal along a timing dimension with a preset sampling window;
a statistical feature extraction unit, configured to extract frequency domain features of each of the sampling windows from the plurality of sampling windows based on fourier transform;
the first coding unit is used for enabling the frequency domain characteristics of each sampling window to pass through the time sequence coder which is trained by the training module and comprises the one-dimensional convolutional layer and the full-connection layer so as to obtain a first characteristic vector;
a second coding unit, configured to pass the waveform map of the pulse signal through the deep separable convolutional neural network trained by the training module to obtain a first feature map;
a third encoding unit, configured to pass the first feature map through a filter-based convolutional neural network trained by the training module to obtain a second feature vector;
a fourth encoding unit configured to calculate a transposed vector of the first eigenvector by multiplying the second eigenvector to obtain a classification eigenvector matrix; and
and the classification result generating unit is used for enabling the classification characteristic matrix to pass through a classifier to obtain a classification result, and the classification result is used for indicating whether the forklift hub has a fault or not.
2. The remote control type electric drive folding forklift hub rapid maintenance device according to claim 1, wherein the time sequence coding unit comprises:
the input vector arrangement subunit is used for arranging the frequency domain characteristics of each sampling window into a one-dimensional input vector;
a full-connection subunit, configured to perform full-connection coding on the input vector obtained by the input vector arrangement subunit by using a full-connection layer of the time sequence encoder according to a following formula to extract a high-dimensional implicit feature of a feature value at each position in the input vector, where the formula is:
Figure FDA0003967039480000021
wherein X is the input vector, Y is the output vector, W is the weight matrix, B is the offset vector,
Figure FDA0003967039480000022
represents a matrix multiplication;
a one-dimensional convolution subunit, configured to perform one-dimensional convolution encoding on the input vector obtained by the input vector arrangement subunit by using a one-dimensional convolution layer of the time-series encoder according to the following formula to extract a high-dimensional implicit correlation feature between feature values of each position in the input vector, where the formula is:
Figure FDA0003967039480000031
wherein, a is the width of the convolution kernel in the x direction, F is the parameter vector of the convolution kernel, G is the local vector matrix operated with the convolution kernel function, and w is the size of the convolution kernel.
3. The remote controlled electric drive folding forklift hub quick service device according to claim 2, wherein said second convolution unit is further configured to: processing the first feature map using the convolutional filter-based neural network with the following formula to generate the second feature vector;
wherein the formula is:
f i ＝Sigmoid(N i ×f i-1 +B i )
wherein f is i-1 Is the input of the i-th convolutional neural network, f i Is the output of the ith convolutional neural network, N i A filter which is the ith convolutional neural network, and B i Sigmoid represents the nonlinear activation function for the bias matrix of the ith convolutional neural network.
4. The remote controlled electric powered folding forklift hub quick service device according to claim 3, wherein said scale migration deterministic loss function value calculation unit is further configured to: calculating the scale migration certainty loss function value between the first feature vector and the second feature vector in the following formula;
wherein the formula is:
Figure FDA0003967039480000032
wherein V 1 Representing said first feature vector, V 2 Representing said second feature vector, V 1 And V 2 Are all in the form of row vectors, and | · | | luminance F The Frobenius norm of the matrix is represented.
5. The remote controlled electric drive folding forklift hub quick service device according to claim 4, wherein said joint coding unit is further configured to: calculating the classification feature matrix based on the first feature vector and the second feature vector in the following formula;
wherein the formula is:
Figure FDA0003967039480000033
wherein
Figure FDA0003967039480000034
Representing vector multiplication, M representing the classification feature matrix, V 1 Representing said first feature vector, V 2 Representing the second feature vector in the second set of feature vectors,
Figure FDA0003967039480000035
a transposed vector representing the first feature vector.
6. The remote controlled electric powered folding forklift hub quick service device according to claim 5, wherein said classification loss function value calculation unit is further configured to: the classifier processes the classification feature matrix to generate a classification result according to the following formula: softmax { (W) n ,B n ):…:(W 1 ,B 1 ) L Proiect (F) }, where Project (F) denotes the projection of the classification feature matrix as a vector, W 1 To W n As a weight matrix for each fully connected layer, B 1 To B n A bias matrix representing the layers of the fully-connected layer; and calculating a cross entropy value between the classification result and a real value as the classification loss function value.
7. The utility model provides a remote control formula electricity drives folding fork truck wheel hub quick maintenance device's application method which characterized in that includes:
a training phase comprising:
acquiring a pulse signal acquired when a detection device approaches a forklift hub, wherein the detection device comprises a Hall magnetic proximity switch and a permanent magnet;
intercepting a plurality of sampling windows from the pulse signal along a timing dimension with a preset sampling window;
extracting frequency domain features of each of the sampling windows from the plurality of sampling windows based on a Fourier transform;
passing the frequency domain features of each sampling window through a time sequence encoder comprising a one-dimensional convolutional layer and a full-link layer to obtain a first feature vector;
passing the waveform map of the pulse signal through a depth separable convolutional neural network to obtain a first feature map, wherein the depth separable convolutional neural network is in a convolution operation, and filters of different layers are used for convolving input data in two spatial dimensions W and H and a channel dimension C respectively;
passing the first feature map through a filter-based convolutional neural network to obtain a second feature vector;
calculating a scale migration certainty loss function value between the first feature vector and the second feature vector, the scale migration certainty loss function value being a natural exponent function value raised to a power of a quotient of an eigenvalue obtained by multiplying the first feature vector by a transpose of the second feature vector, and a Frobenius norm of an eigenmatrix obtained by multiplying the transpose of the first feature vector by the second feature vector;
computing a transposed vector of the first eigenvector multiplied by the second eigenvector to obtain a classification eigenvector matrix;
passing the classification feature matrix through a classifier to obtain a classification loss function value; and
training the time-series encoder, the deep separable convolutional neural network, and the filter-based convolutional neural network with a weighted sum between the classification loss function value and the scale migration deterministic loss function value as a loss function value; and
an inference phase comprising:
acquiring a pulse signal acquired when the detection device approaches the forklift hub;
intercepting a plurality of sampling windows from the pulse signal along a timing dimension with a preset sampling window;
extracting frequency domain features of each of the sampling windows from the plurality of sampling windows based on a fourier transform;
passing the frequency domain features of each sampling window through the time sequence encoder which is trained by a training module and comprises the one-dimensional convolutional layer and the full-link layer to obtain a first feature vector;
passing the waveform map of the pulse signal through the deep separable convolutional neural network trained by the training module to obtain a first feature map;
passing the first feature map through a filter-based convolutional neural network trained by the training module to obtain a second feature vector;
computing a transposed vector of the first eigenvector multiplied by the second eigenvector to obtain a classification eigenvector matrix; and
and passing the classification characteristic matrix through a classifier to obtain a classification result, wherein the classification result is used for representing whether the forklift hub has a fault or not.
8. The method of using the apparatus for rapidly repairing a hub of a remotely controlled electrically powered folding forklift as claimed in claim 7, wherein the step of passing the frequency domain feature of each of the sampling windows through a time sequence encoder comprising a one-dimensional convolution layer and a full link layer to obtain a first feature vector comprises:
arranging the frequency domain characteristics of each sampling window into a one-dimensional input vector;
performing full-connection coding on the input vector obtained by the input vector arrangement subunit by using a full-connection layer of the time sequence encoder according to the following formula to extract high-dimensional implicit features of feature values of each position in the input vector, wherein the formula is as follows:
Figure FDA0003967039480000051
wherein X is the input vector and Y is the outputA vector is given, W is the weight matrix, B is the offset vector,
Figure FDA0003967039480000052
represents a matrix multiplication;
performing one-dimensional convolution encoding on the input vector obtained by the input vector arrangement subunit by using a one-dimensional convolution layer of the time sequence encoder according to the following formula to extract high-dimensional implicit correlation features among feature values of all positions in the input vector, wherein the formula is as follows:
Figure FDA0003967039480000061
wherein, a is the width of the convolution kernel in the x direction, F is the parameter vector of the convolution kernel, G is the local vector matrix operated with the convolution kernel function, and w is the size of the convolution kernel.
9. The method of using a remote controlled electric powered folding forklift hub quick service device as recited in claim 8, wherein the step of passing the first feature map through a filter-based convolutional neural network to obtain a second feature vector comprises:
processing the first feature map using the convolutional filter-based neural network to generate the second feature vector;
wherein the formula is:
f i ＝Sigmoid(N i ×f i-1 +B i )
wherein f is i-1 Is the input of the i-th convolutional neural network, f i Is the output of the ith convolutional neural network, N i A filter which is an i-th layer convolutional neural network, and B i Sigmoid represents the nonlinear activation function for the bias matrix of the ith convolutional neural network.
10. The method of using a remote controlled electric powered folding forklift hub quick service device as claimed in claim 9, wherein calculating a scale migration deterministic loss function value between said first eigenvector and said second eigenvector comprises:
calculating the scale migration certainty loss function value between the first feature vector and the second feature vector in the following formula;
wherein the formula is:
Figure FDA0003967039480000062
wherein V 1 Representing said first feature vector, V 2 Representing said second feature vector, V 1 And V 2 Are all in the form of row vectors, and | · | | luminance F The Frobenius norm of the matrix is represented.
CN202210624763.8A 2022-06-02 2022-06-02 Remote control type electric-drive folding forklift hub rapid maintenance device and method thereof Active CN114993710B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210624763.8A CN114993710B (en) 2022-06-02 2022-06-02 Remote control type electric-drive folding forklift hub rapid maintenance device and method thereof

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210624763.8A CN114993710B (en) 2022-06-02 2022-06-02 Remote control type electric-drive folding forklift hub rapid maintenance device and method thereof

Publications (2)

Publication Number Publication Date
CN114993710A CN114993710A (en) 2022-09-02
CN114993710B true CN114993710B (en) 2023-01-31

Family

ID=83030633

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210624763.8A Active CN114993710B (en) 2022-06-02 2022-06-02 Remote control type electric-drive folding forklift hub rapid maintenance device and method thereof

Country Status (1)

Country Link
CN (1) CN114993710B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116794553B (en) * 2023-04-07 2024-03-08 浙江万能弹簧机械有限公司 Intelligent fault diagnosis method and system for high-frequency power supply

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106874957A (en) * 2017-02-27 2017-06-20 苏州大学 A kind of Fault Diagnosis of Roller Bearings
CN211468564U (en) * 2019-12-06 2020-09-11 上港集团物流有限公司兴宝仓储分公司 Detection device for detecting steering fault of forklift
CN111783331A (en) * 2020-06-04 2020-10-16 中国人民解放军军事科学院国防科技创新研究院 Satellite component temperature field determination method based on transfer learning
CN112001903A (en) * 2020-08-21 2020-11-27 深圳市华汉伟业科技有限公司 Defect detection network construction method, abnormality detection method and system, and storage medium
CN113670610A (en) * 2021-06-09 2021-11-19 广州大学 Fault detection method, system and medium based on wavelet transformation and neural network
WO2022011754A1 (en) * 2020-07-16 2022-01-20 苏州大学 Fault diagnosis method based on adaptive manifold embedded dynamic distribution alignment

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106874957A (en) * 2017-02-27 2017-06-20 苏州大学 A kind of Fault Diagnosis of Roller Bearings
CN211468564U (en) * 2019-12-06 2020-09-11 上港集团物流有限公司兴宝仓储分公司 Detection device for detecting steering fault of forklift
CN111783331A (en) * 2020-06-04 2020-10-16 中国人民解放军军事科学院国防科技创新研究院 Satellite component temperature field determination method based on transfer learning
WO2022011754A1 (en) * 2020-07-16 2022-01-20 苏州大学 Fault diagnosis method based on adaptive manifold embedded dynamic distribution alignment
CN112001903A (en) * 2020-08-21 2020-11-27 深圳市华汉伟业科技有限公司 Defect detection network construction method, abnormality detection method and system, and storage medium
CN113670610A (en) * 2021-06-09 2021-11-19 广州大学 Fault detection method, system and medium based on wavelet transformation and neural network

Also Published As

Publication number Publication date
CN114993710A (en) 2022-09-02

Similar Documents

Publication Publication Date Title
CN116639010B (en) Intelligent control system and method for charging pile
CN107703920B (en) Fault detection method of train braking system based on multivariate time sequence
CN114993710B (en) Remote control type electric-drive folding forklift hub rapid maintenance device and method thereof
CN103838229B (en) Diagnosis method and device of electric car
CN107219457A (en) Frame-type circuit breaker fault diagnosis and degree assessment method based on operation annex electric current
CN115248355A (en) Mechanical equipment fault detection system and detection method thereof
CN104809722A (en) Electrical device fault diagnosis method based on infrared thermography
Chang et al. Wire mismatch detection using a convolutional neural network and fault localization based on time–frequency-domain reflectometry
CN116552306B (en) Monitoring system and method for direct current pile
CN110516380B (en) Deep reinforcement test method and system based on vehicle driving simulation data
CN113188807B (en) Automatic abs result judging algorithm
CN112507915A (en) Method for identifying loosening state of bolt connection structure based on vibration response information
EP4009534A1 (en) Automotive physical layer (phy) cable fault diagnosis
Zhang et al. Feature engineering and artificial intelligence-supported approaches used for electric powertrain fault diagnosis: A review
CN116992226A (en) Water pump motor fault detection method and system
CN106301610A (en) The adaptive failure detection of a kind of superhet and diagnostic method and device
CN115266159A (en) Fault diagnosis method and system for train traction system
CN116842379A (en) Mechanical bearing residual service life prediction method based on DRSN-CS and BiGRU+MLP models
CN116298880A (en) Micro-motor reliability comprehensive test system and method thereof
CN113156913B (en) ABS fault diagnosis system and method
CN109858573B (en) Method for preventing lifting of collecting card based on neural network
CN106093626B (en) A method of the small fault for diagnosing CRH5 type high ferro inverters
Yi et al. Safety belt wearing detection algorithm based on human joint points
CN116001579A (en) Emergency power-off method and system for new energy vehicle
CN116204821B (en) Vibration evaluation method and system for rail transit vehicle

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant