A method, device, medium and equipment for diagnosing faults of ship steering gear

Abstract

The present invention provides a method, device, medium and equipment for diagnosing faults of a ship steering gear. The method comprises: determining an SDP image data set of a target signal and constructing a deep residual neural network; the first layer of the deep residual neural network is a convolutional neural network, and the convolutional neural network is used to perform convolution operations on the SDP image set to enhance the ability of extracting key features; an attention mechanism layer is introduced into the residual block of the deep residual neural network; a vibration signal reflects the vibration characteristics of the steering gear, and a current signal reflects the circuit connection of the steering gear. Multiple signals of the ship steering gear are fused in a two-dimensional polar coordinate axis to obtain an SDP image, which can increase the expression of fault features; and after determining the SDP image set by using a target conversion factor, the accuracy of the feature expression of the SDP image set can be improved, and then after the deep residual neural network model is trained by using the SDP image set, the recognition accuracy of the model is improved, thereby improving the accuracy of fault diagnosis.

Description

Technical Field

[0001] The present application relates to the technical field of ship fault diagnosis, and in particular to a method, device, medium and equipment for performing fault diagnosis on a ship steering gear. Background Art

[0002] The ship's steering gear is one of the key deck equipment of the ship. By manipulating the steering wheel or tiller, the ship's heading is adjusted to ensure that the ship maintains a stable route in the water environment. In complex water environments, the steering gear equipment is often affected by moisture, salt spray, vibration, etc., which eventually lead to steering gear failure. If the steering gear failure is not discovered early, it will cause major economic losses at the least, and serious accidents such as casualties on the ship, pollutant leakage, etc. Therefore, early detection of steering gear failure can effectively avoid the losses caused by steering gear failure, ensure the normal operation of the ship, and reduce unnecessary economic losses.

[0003] There are some technologies in the prior art that can be used to diagnose steering gear faults, but there are still some problems. For example, some technologies have high computational complexity and require a lot of computing resources and time to run and optimize the diagnostic model, which affects the efficiency of fault diagnosis. Other technologies are insufficient in effectively analyzing complex signals when the background noise signal is complex, which affects the accuracy of fault diagnosis. Summary of the invention

[0004] In view of the problems existing in the prior art, the embodiments of the present invention provide a method, device, medium and equipment for diagnosing faults of ship steering gear, so as to solve or partially solve the technical problem in the prior art that the accuracy of fault diagnosis cannot be ensured when diagnosing faults of ship steering gear.

[0005] A first aspect of the present invention provides a method for diagnosing a fault of a ship steering gear, the method comprising:

[0006] Collecting multiple groups of target signals of the ship steering gear in a normal state and collecting multiple groups of target signals of the ship steering gear in different fault states, and converting each group of the target signals into frequency domain signals; the target signals include: vibration signals and current signals;

[0007] Determine a target conversion factor of a symmetrical point pattern SDP image based on each group of the target signals and the corresponding frequency domain signals, and convert each group of target signals and the corresponding frequency domain signals in different states into a corresponding SDP image according to the target conversion factor to obtain an SDP image set;

[0008] The constructed deep residual neural network model is trained using the SDP image set, so that the deep residual neural network model learns the key features presented by the ship steering gear in a normal state and the key features presented in a fault state;

[0009] Acquire a target signal to be measured of the ship motor to be measured, convert the target signal to be measured into a frequency domain signal to be measured, and convert the target signal to be measured and the frequency domain signal to be measured into a SDP image to be measured;

[0010] The SDP image to be tested is diagnosed using the trained deep residual neural network model; wherein the first layer of the deep residual neural network is a convolutional neural network, and the convolutional neural network is used to perform convolution operations on the SDP image set to enhance the ability of key feature extraction; an attention mechanism layer is introduced into the residual block of the deep residual neural network, and the attention mechanism layer is used to adaptively learn the feature weights of each feature in the SDP image set to suppress several non-critical features in the SDP image set.

[0011] In the above scheme, the conversion factor includes: a delay factor and an amplification factor; the target conversion factor of the symmetrical point pattern SDP image is determined based on each group of the target signals and the corresponding frequency domain signals, including:

[0012] Determine all value combinations of the delay factor and the amplification factor according to the value range of the delay factor and the value range of the amplification factor;

[0013] The following processing is performed for each value combination: each group of target signals and corresponding frequency domain signals in each state are converted into reference SDP images according to the value combination; the number of reference SDP images in each value combination is consistent with the number of state types; for any two reference SDP images in the value combination, the similarity of the two reference SDP images is determined, thereby obtaining multiple similarities, and the average value of the multiple similarities is used as the benchmark similarity corresponding to the value combination;

[0014] Obtain the baseline similarities of all value combinations, and determine the value combination corresponding to the minimum baseline similarity as the target value combination;

[0015] The target conversion factor is determined according to the target value combination.

[0016] In the above solution, determining the similarity between the two reference SDP images includes:

[0017] According to the formula Determine the similarity R between the i′th reference SDP image and the j′th reference SDP image under the value combination i′,j′ (N,M); where

[0018] The N is a first pixel value matrix corresponding to the i′th reference SDP image, the M is a second pixel value matrix corresponding to the j′th reference SDP image, is the average value of all pixel values ​​in the first pixel value matrix, is the average value of all pixel values ​​in the second pixel value matrix, and the M mn is the pixel value corresponding to the mth row and nth column in the first pixel value matrix, wherein N mn is the pixel value corresponding to the mth row and nth column in the second pixel value matrix.

[0019] In the above scheme, converting each group of target signals and corresponding frequency domain signals into corresponding SDP images according to the target conversion factor includes:

[0020] For the target signal and any signal point in the frequency domain signal of the target signal, the following processing is performed:

[0021] According to the formula The signal point is converted into a polar coordinate signal; the polar coordinate signal includes: the polar coordinate radius r(i), the angle θ(i) of β rotating clockwise, the angle φ(i) of β rotating counterclockwise, the β is the rotation angle of the mirror symmetry plane, the x i is the i-th signal point, the x min is the signal point with the smallest amplitude among all signal points, and x max is the signal point with the largest amplitude among all signal points, t is the delay factor in the target conversion factor, z is the amplification factor in the target conversion factor, and x i+t is the amplitude of the i+tth signal point;

[0022] Each signal point is mapped onto a polar coordinate axis according to the polar coordinates of each signal point to obtain the SDP image.

[0023] In the above scheme, the deep residual neural network model includes in sequence: a convolutional neural network, a convolutional layer, a batch normalization layer BN layer, an activation function layer, a first pooling layer, a plurality of residual blocks, a second pooling layer and a fully connected layer;

[0024] The convolutional neural network includes a plurality of convolutional layer combinations and pooling layers in sequence, and an activation function layer pooling layer is introduced after each convolutional layer; the convolutional neural network is used to perform a convolution operation on the input SDP image set to obtain a first feature image;

[0025] The convolution layer is used to perform a convolution operation on the first feature image to obtain a second feature image;

[0026] The BN layer is used to perform normalization processing on the second feature image to obtain a normalized feature image;

[0027] The activation function layer is used to perform nonlinear processing on the normalized feature image to obtain a nonlinear normalized feature image;

[0028] The first pooling layer is used to reduce the dimension of the nonlinear normalized feature image to obtain a low-dimensional feature image of the nonlinear normalized feature image;

[0029] The plurality of residual blocks are used to extract key feature images from the low-dimensional feature images;

[0030] The second pooling layer is used to continue the dimension reduction operation on the key feature image to obtain the key feature image after dimension reduction;

[0031] The fully connected layer classifies each key feature image based on a built-in classification function to obtain a key feature image in a normal state and a key feature image in a fault state.

[0032] In the above scheme, the residual block includes: convolution layer, activation function layer and attention mechanism layer, and a BN layer is introduced after each convolution layer;

[0033] The attention mechanism layer includes: an average pooling layer, a depth-separable convolution layer and an activation function layer in sequence; the input feature image of the attention mechanism layer is the output feature image of the last BN layer in the residual block;

[0034] The average pooling layer is used to average the pixel values ​​in each pooling window on the feature image output by the last BN layer in the residual block to generate a pooled feature image;

[0035] The depth-wise separable convolution layer is used to group the pooled feature images to form groups of feature maps, and generate convolution kernels corresponding to each group of feature maps, and respectively use the convolution kernels of each group of feature maps to convolve the corresponding group of feature maps to obtain convolution feature images; and then use a 1*1 convolution kernel to convolve each convolution feature product image to obtain a single-channel convolution image;

[0036] The activation function layer is used to perform nonlinear processing on the single-channel convolution image.

[0037] In the above scheme, the SDP image to be tested is input into the trained deep residual neural network model to obtain the fault diagnosis result, including:

[0038] Using the deep residual neural network model to perform feature analysis on the SDP image to be tested to obtain multiple fault feature values;

[0039] The fault feature classifier of the deep residual neural network model is used to determine the probability of each fault feature value, and the corresponding fault type is output according to the probability of the fault feature value.

[0040] A second aspect of the present invention provides a device for diagnosing a fault of a ship steering gear, the device comprising:

[0041] A collection unit is used to collect multiple groups of target signals of the ship steering gear in a normal state and multiple groups of target signals of the ship steering gear in different fault states, and convert each group of the target signals into frequency domain signals; the target signals include: vibration signals and current signals;

[0042] A first conversion unit is used to determine a target conversion factor of a symmetrical point pattern SDP image based on each group of the target signals and the corresponding frequency domain signals, and convert each group of target signals and the corresponding frequency domain signals in different states into a corresponding SDP image according to the target conversion factor to obtain an SDP image set;

[0043] A training unit, used to train the constructed deep residual neural network model using the SDP image set, so that the deep residual neural network model learns the key features presented by the ship steering gear in a normal state and the key features presented in a fault state;

[0044] A second conversion unit is used to obtain a target signal to be tested of the ship motor to be tested, convert the target signal to be tested into a frequency domain signal to be tested, and convert the target signal to be tested and the frequency domain signal to be tested into an SDP image to be tested;

[0045] The diagnostic unit uses the trained deep residual neural network model to diagnose the SDP image to be tested to obtain a fault diagnosis result; wherein the first layer of the deep residual neural network is a convolutional neural network, and the convolutional neural network is used to perform convolution operations on the input SDP image set to enhance the ability of key feature extraction; an attention mechanism layer is introduced into the residual block of the deep residual neural network, and the attention mechanism layer is used to adaptively learn the feature weight of each feature in the SDP image set, and use the feature weight to achieve weighting to suppress several non-critical features in the SDP image set.

[0046] According to a third aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the steps of any one of the methods described in the first aspect are implemented.

[0047] A fourth aspect of the present invention is a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the methods described in the first aspect when executing the program.

[0048] The present invention provides a method, device, medium and equipment for diagnosing faults of a ship steering gear. The method comprises: collecting multiple groups of target signals of the ship steering gear in a normal state and collecting multiple groups of target signals of the ship steering gear in different fault states, and converting each group of the target signals into a frequency domain signal; the target signal comprises: a vibration signal and a current signal; determining a target conversion factor of a symmetrical point pattern SDP image based on each group of the target signals and the corresponding frequency domain signal, converting each group of target signals and the corresponding frequency domain signal in different states into a corresponding SDP image according to the target conversion factor, and obtaining an SDP image set; training a constructed deep residual neural network model using the SDP image set, so that the deep residual neural network model learns the key features presented by the ship steering gear in a normal state and the key features presented in a fault state; obtaining a target signal to be tested of a ship motor to be tested, converting the target signal to be tested into a frequency domain signal to be tested, and converting the target signal to be tested and the frequency domain signal to be tested into an SDP image to be tested; using The SDP image to be tested is diagnosed using the trained deep residual neural network model; wherein the first layer of the deep residual neural network is a convolutional neural network, and the convolutional neural network is used to perform convolution operations on the SDP image set to enhance the ability to extract key features; an attention mechanism layer is introduced into the residual block of the deep residual neural network, and the attention mechanism layer is used to adaptively learn the feature weights of each feature in the SDP image set to suppress several non-key features in the SDP image set; in this way, since the vibration signal can reflect the vibration characteristics of the steering gear, and the current signal can reflect the circuit connection of the steering gear, the various different signals of the ship steering gear are fused in the two-dimensional polar coordinate axis to obtain the SDP image, which can increase the expression of the fault characteristics; and the optimal target conversion factor is first determined. After the SDP image set is determined using the target conversion factor, the accuracy of the feature expression of the SDP image set can be improved. After the deep residual neural network model is trained using the SDP image set, the recognition accuracy of the model is improved, thereby improving the accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0050] Figure 1 A schematic flow chart of a method for diagnosing a fault of a ship steering gear according to an embodiment of the present invention is shown;

[0051] Figure 2A schematic diagram of a vibration signal collected in a normal state a according to an embodiment of the present invention is shown;

[0052] Figure 3 A schematic diagram of a vibration signal collected in a fault state b according to an embodiment of the present invention is shown;

[0053] Figure 4 A schematic diagram of a vibration signal collected under a fault state c according to an embodiment of the present invention is shown;

[0054] Figure 5 A schematic diagram of a vibration signal collected under a fault state d according to an embodiment of the present invention is shown;

[0055] Figure 6 A schematic diagram of a current signal collected in a normal state a according to an embodiment of the present invention is shown;

[0056] Figure 7 A schematic diagram of a current signal collected under a fault state b according to an embodiment of the present invention is shown;

[0057] Figure 8 A schematic diagram of a current signal collected under a fault state c according to an embodiment of the present invention is shown;

[0058] Fig. 9 A schematic diagram of a current signal collected under a fault state d according to an embodiment of the present invention is shown;

[0059] Fig.10 The figure shows an SDP image corresponding to a certain group of target signals collected in a normal state a according to an embodiment of the present invention;

[0060] Fig.11 The figure shows an SDP image corresponding to a certain group of target signals collected under fault state b according to an embodiment of the present invention;

[0061] Fig.12 The SDP image corresponding to a certain group of target signals collected under the fault state c according to one embodiment of the present invention is shown;

[0062] Fig.13 The SDP image corresponding to a certain group of target signals collected under the fault state d according to one embodiment of the present invention is shown;

[0063] Fig.14 A schematic diagram showing the accuracy of training and testing a deep residual neural network model according to an embodiment of the present invention is shown;

[0064] Fig.15 A schematic structural diagram of a device for diagnosing faults of a ship steering gear according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0065] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0066] The present invention provides a method for diagnosing a fault of a ship steering gear. Figure 1 As shown, the method mainly includes the following steps:

[0067] S110, collecting multiple groups of target signals of the ship steering gear in a normal state and collecting multiple groups of target signals of the ship steering gear in different fault states, and converting each group of the target signals into frequency domain signals; the target signals include: vibration signals and current signals.

[0068] In order to fully understand the operating status of the steering gear and reduce the probability of misjudgment and missed diagnosis, the present invention combines the vibration signal and the current signal for analysis to enhance the reliability and stability of the fault characteristics.

[0069] Then it is necessary to collect multiple groups of target signals of the ship's steering gear in normal state and multiple groups of target signals of the ship's steering gear in different fault states, and convert each group of target signals into frequency domain signals; the target signals include: vibration signals and current signals.

[0070] Normal status (a) and fault status can include the following:

[0071] (a) Test the cylinder, exhaust the air, and then steer left and then right;

[0072] (b) Test the cylinder, after exhausting the air, turn left and then right, the pump group is not working, and the oil circuit is closed;

[0073] (c) Cylinder test, after exhausting, turn left and then right, pump group not working, oil circuit closed;

[0074] (d) Measure the cylinder, exhaust the air, turn left and then right, the pump group is not working, the oil circuit is closed, the pump group is working, and the inlet and outlet resistances are increased.

[0075] refer to Figure 2 to Figure 5 , respectively, the vibration signal corresponding to the normal state (a), the vibration signal corresponding to the fault state (b), the vibration signal corresponding to the fault state (c), and the vibration signal corresponding to the fault state (d) are shown.

[0076] refer to Figure 6 to Figure 9, respectively, the current signal corresponding to the normal state (a), the current signal corresponding to the fault state (b), the current signal corresponding to the fault state (c), and the current signal corresponding to the fault state (d) are shown.

[0077] In order to improve the subsequent training accuracy of the neural network, the target signals collected under different states can also be set with corresponding state labels. For example, the state label of multiple groups of target signals and corresponding frequency domain signals under normal state (a) can be 0, the state label of multiple groups of target signals and corresponding frequency domain signals under fault state (b) can be 1, the state label of multiple groups of target signals and corresponding frequency domain signals under fault state (c) can be 2, and the state label of multiple groups of target signals and corresponding frequency domain signals under fault state (d) can be 3.

[0078] Then the target signal is converted into a frequency domain signal according to formula (1):

[0079]

[0080] In formula (1), when the target signal is a vibration signal, x(n) is the nth vibration signal, N′ is the total number of vibration signals in a group of target signals, and X(k) is the frequency domain signal corresponding to the nth vibration signal. is a complex term, k is the serial number of the frequency domain signal, and i is the imaginary unit.

[0081] When the target signal is a current signal, x(n) is the nth current signal, N′ is the total number of current signals in a group of target signals, and X(k) is the frequency domain signal corresponding to the nth current signal. is a complex term, k is the serial number of the frequency domain signal, and i is the imaginary unit.

[0082] In this way, each current signal and the frequency domain signal corresponding to each current signal, each vibration signal and the frequency domain signal corresponding to each vibration signal can be obtained.

[0083] In practical applications, overlapping sampling method can be used to collect target signals. The number of sampling points for each group of target signals can be 256. The number of groups of target signals can also be determined based on actual conditions (for example, 1000), which is not limited here.

[0084] S111, determining a target conversion factor of a symmetrical point pattern SDP image based on each group of the target signals and the corresponding frequency domain signals, converting each group of target signals and the corresponding frequency domain signals in different states into a corresponding SDP image according to the target conversion factor, to obtain an SDP image set.

[0085] In order to enhance the expression of fault characteristics, the present invention needs to fuse the signal analysis results of multi-source signals at different scales on the two-dimensional polar coordinate axis. Therefore, the present invention uses the symmetrical point pattern SDP image conversion method to convert the above-mentioned current signal, the frequency domain signal of the current signal, the vibration signal and the frequency domain signal of the vibration signal into an SDP image for fault analysis.

[0086] It should be noted that when using the SDP image conversion method, the method involves conversion factors, including: delay factor and amplification factor; the delay factor controls the sparseness of the SDP image petals, and the amplification factor controls the opening and closing degree of the SDP image petals. However, if the delay factor and the amplification factor are too large or too small, the SDP image will not be able to express the fault characteristics to the maximum extent, so it is necessary to determine the target conversion factor first.

[0087] In one embodiment, determining a target conversion factor of a symmetrical point pattern SDP image based on each group of target signals and corresponding frequency domain signals includes:

[0088] Determine all value combinations of the delay factor and the amplification factor according to the value range of the delay factor and the value range of the amplification factor;

[0089] The following processing is performed for each value combination: each group of target signals and corresponding frequency domain signals in each state are converted into reference SDP images according to the value combination; the number of reference SDP images in each value combination is consistent with the number of state types; for any two reference SDP images in the value combination, the similarity of the two reference SDP images is determined, thereby obtaining multiple similarities, and the average value of the multiple similarities is used as the benchmark similarity corresponding to the value combination;

[0090] Obtain the baseline similarities of all value combinations, and determine the value combination corresponding to the minimum baseline similarity as the target value combination;

[0091] The target conversion factor is determined based on the target value combination.

[0092] In one implementation, determining the similarity between the two reference SDP images includes:

[0093] According to formula (2), the similarity R between the i′th reference SDP image and the j′th reference SDP image under the value combination is determined: i′,j′ (N,M);

[0094]

[0095] In formula (2), N is the first pixel value matrix corresponding to the i′th reference SDP image, M is the second pixel value matrix corresponding to the j′th reference SDP image, is the average value of all pixel values ​​in the first pixel value matrix, is the average value of all pixel values ​​in the second pixel value matrix, and the M mn is the pixel value corresponding to the mth row and nth column in the first pixel value matrix, wherein N mn is the pixel value corresponding to the mth row and nth column in the second pixel value matrix.

[0096] Specifically, the delay factor and the amplification factor both have a range of values. For example, the delay factor may have a range of values ​​of [1, 10], and the amplification factor may have a range of values ​​of [20o, 60o]. Then, the value combinations of the delay factor and the amplification factor may include [1, 20o], [1, 21o], ... [10, 60o]. Under each value combination, the target signals of the above four states and the corresponding frequency domain signals may be converted into corresponding reference SDP images. That is, under each value combination, four reference SDP images are included, and each reference SDP image corresponds to a state.

[0097] Then, for any two reference SDP images in each value combination, the similarity between the two reference SDP images is determined according to the above formula (1). Since one value combination corresponds to 4 reference SDP images, 6 similarities can be obtained, and then the average of the 6 similarities is used as the benchmark similarity under the value combination.

[0098] In this way, there is a reference similarity for each value combination, and then the minimum reference similarity is determined among all the reference similarities, and the value combination corresponding to the minimum reference similarity is determined as the target value combination.

[0099] For example, assuming that the baseline similarity corresponding to the value combination [8, 35o] is the smallest, then the target value combination is [8, 35o], the delay factor t in the target conversion factor is 8, and the amplification factor z is 35.

[0100] After the target conversion factor is determined, in one embodiment, each group of target signals and corresponding frequency domain signals are converted into a corresponding SDP image according to the target conversion factor, including:

[0101] For the target signal and any signal point in the frequency domain signal of the target signal, the following processing is performed:

[0102] According to formula (3), the signal point is converted into polar coordinate signal:

[0103]

[0104] The polar coordinate signal includes: the polar coordinate radius r(i), the angle θ(i) of β rotating clockwise, the angle φ(i) of β rotating counterclockwise, β is the rotation angle of the mirror symmetry plane, x i is the amplitude of the ith signal point, x min is the signal point with the smallest amplitude among all signal points, x max is the signal point with the largest amplitude among all signal points, t is the delay factor in the target conversion factor, z is the amplification factor in the target conversion factor, x i+t is the amplitude of the i+tth signal point.

[0105] Each signal point is mapped onto the polar coordinate axis according to the polar coordinates of each signal point to obtain an SDP image.

[0106] Under the target conversion factor, the SDP image corresponding to a set of target signals in the normal state (a) is as follows: Fig.10 As shown; the SDP image corresponding to a group of target signals in the fault state (b) is as follows Fig.11 As shown; the SDP image corresponding to a group of target signals under fault state (c) is as follows Fig.12 As shown in the figure, the SDP image corresponding to a group of target signals under the fault state (d) is as follows Fig.13 shown.

[0107] In this way, each SDP image contains both the current signal and vibration signal in the time domain and the current signal and vibration signal in the frequency domain, which can be used to analyze the possible faults of the servo from multiple angles, thereby enhancing the expression of fault characteristics.

[0108] After SDP conversion is performed on all groups of target signals, an SDP image set can be obtained.

[0109] It should be noted that the specific implementation method for converting the target signals in the four states and the corresponding frequency domain signals into the corresponding reference SDP images under each value combination can also be implemented according to formula (3).

[0110] S112, using the SDP image set to train the constructed deep residual neural network model, so that the deep residual neural network model learns the key features of the ship steering gear in a normal state and the key features of the ship steering gear in a fault state.

[0111] After the SDP image set is determined, the SDP image set is used as a sample set to train the constructed deep residual neural network model, so that the deep residual neural network model can learn the key features of the ship's steering gear in a normal state and the key features in a fault state.

[0112] In the present invention, the deep residual neural network model includes: a convolutional neural network, a convolutional layer, a batch normalization layer BN layer, an activation function layer, a first pooling layer, a plurality of residual blocks, a second pooling layer and a fully connected layer in sequence;

[0113] The convolutional neural network includes several convolutional layer combinations and pooling layers in sequence, and an activation function layer pooling layer is introduced after each convolutional layer; the convolutional neural network is used to perform convolution operations on the input SDP image set to obtain the first feature image; the first layer of the deep residual neural network is set to a convolutional neural network, which can improve the accuracy of extracting key features compared to directly using a separate convolutional layer to extract features from the input SDP image set;

[0114] The convolution layer is used to perform a convolution operation on the first feature image to obtain a second feature image;

[0115] The BN layer is used to normalize the second feature image to obtain a normalized feature image;

[0116] The activation function layer is used to perform nonlinear processing on the normalized feature image to obtain a nonlinear normalized feature image;

[0117] The first pooling layer is used to reduce the dimension of the nonlinear normalized feature image to obtain a low-dimensional feature image of the nonlinear normalized feature image;

[0118] Several residual blocks are used to extract key feature images from low-dimensional feature images;

[0119] The second pooling layer is used to continue the dimension reduction operation on the key feature image to obtain the key feature image after dimension reduction;

[0120] The fully connected layer classifies each key feature image based on the built-in classification function to obtain the key feature image of the normal state and the key feature image of the fault state.

[0121] Among them, an attention mechanism layer is introduced in the residual block, and the attention mechanism layer is used to adaptively learn the feature weight of each feature in the SDP image set, and weighting is achieved using the feature weight to suppress several non-critical features in the SDP image set.

[0122] The residual block includes: convolution layer, activation function layer, convolution layer and attention mechanism layer in sequence, and a BN layer is introduced after each convolution layer; the attention mechanism layer includes: average pooling layer, depth-separable convolution layer and activation function layer in sequence; the input feature image of the attention mechanism layer is the output feature image of the last BN layer in the residual block.

[0123] The average pooling layer is used to average the pixel values ​​in each pooling window on the feature image output by the last BN layer in the residual block to generate a pooled feature image;

[0124] In order to reduce the complexity of the calculation process and improve the efficiency of image processing, the present invention uses a deep separable convolution layer to separate and convolve the pooled feature image, which is achieved as follows:

[0125] The depth-separable convolution layer is used to group the pooled feature images to form groups of feature maps, and generate convolution kernels corresponding to each group of feature maps; convolve the corresponding group of feature maps using the convolution kernel of each group of feature maps to obtain the corresponding convolution feature images; and then convolve each of the convolution feature product images using a 1*1 convolution kernel to obtain a single-channel convolution image;

[0126] For example, assuming that the pooled feature image is a 6-channel feature image, the 6-channel image can be divided into 2 groups of feature images, and a corresponding convolution kernel is generated for each group of feature images. The convolution kernel of each group of feature images is used to convolve the corresponding group of feature images to obtain two groups of 3-channel convolution feature images, and the two groups of 3-channel images are recombined to obtain a combined image. The combined image is then convolved with a 1*1 convolution kernel to output a feature image with 1 channel number.

[0127] The activation function layer is used to perform nonlinear processing on single-channel convolution images.

[0128] It is understandable that when the pooled feature image is grouped and convolved using the depthwise separable convolutional layer, the multi-channel feature maps are grouped and then convolved separately. Compared with the process of convolving the multi-channel feature maps at the same time, the complexity of the calculation process can be greatly reduced and the image processing efficiency can be improved.

[0129] In actual training, the SDP images in each state of the SDP image dataset are divided into training set and test set according to a 4:1 ratio. The training set is used to train the deep residual neural network model, and then the test set is used to test the accuracy of the trained deep residual neural network model. When the test accuracy meets the requirements, it indicates that the deep residual neural network model training is completed.

[0130] Among them, reference Fig.14 , the loss value and model prediction accuracy of the training set and the test set after 50 rounds of iterations. From the loss value change and model prediction accuracy change graph, it can be seen that the model's loss value and prediction accuracy fluctuated greatly in the early stage, but after about 3 rounds of iterations, the model fluctuation gradually stabilized. This phenomenon shows that the deep residual network model of the present invention has a faster convergence ability and a higher model stability.

[0131] S113, obtaining a target signal to be tested of the ship motor to be tested, converting the target signal to be tested into a frequency domain signal to be tested, and converting the target signal to be tested and the frequency domain signal to be tested into an SDP image to be tested.

[0132] Before using the trained deep residual network model to diagnose the SDP image to be tested, it is necessary to obtain the target signal to be tested of the ship motor to be tested, convert the target signal to be tested into a frequency domain signal to be tested, and convert the target signal to be tested and the frequency domain signal to be tested into the SDP image to be tested.

[0133] Specifically, the target signal to be measured can be converted into the corresponding frequency domain signal to be measured according to formula (1), and the target signal to be measured and the frequency domain signal to be measured can be converted into the SDP image to be measured using formula (3) under the target conversion factor; the specific conversion method can refer to the corresponding description above, so it will not be repeated here.

[0134] S114, diagnosing the SDP image to be tested by using the trained deep residual neural network model.

[0135] Then, the SDP image to be tested is input into the trained deep residual neural network model, and the trained deep residual neural network model is used to diagnose the SDP image to be tested, which specifically includes:

[0136] A deep residual neural network model is used to perform feature analysis on the SDP image to be tested, and multiple fault feature values ​​are obtained;

[0137] The fault feature classifier of the deep residual neural network model is used to determine the probability of each fault feature value, and the corresponding fault type is output according to the probability of the fault feature value.

[0138] Specifically, after passing through the fully connected layer, multiple fault feature values ​​y can be output s (x), in order to obtain the characteristic value y of each fault s (x), the fault feature classifier needs to first use the following formula to convert y s (x) is mapped to (0, +∞) and then normalized to (0, 1) using the following formula:

[0139]

[0140]

[0141] Where s is the fault category and K is the total number of fault categories.

[0142] It can be seen that after the fault feature value is converted by softmax, it meets the characteristics of the probability expression, that is, for any type of fault category, a corresponding fault probability is finally output. Therefore, it is equivalent to expressing the probabilities of different fault feature values, and the specific fault type can be determined according to the fault probability. Generally speaking, the fault category corresponding to the maximum fault probability is determined as the final fault type.

[0143] The present invention takes into account that the vibration signal can reflect the vibration characteristics of the steering gear, and the current signal can reflect the circuit connection status of the steering gear. Therefore, the various signals of the ship steering gear are fused in the two-dimensional polar coordinate axis to obtain the SDP image, which can increase the expression of the fault characteristics; and the optimal target conversion factor is first determined. After the SDP image set is determined by using the target conversion factor, the accuracy of the feature expression of the SDP image set can be improved. After the deep residual neural network model is trained by using the SDP image set, the recognition accuracy of the model is improved, thereby improving the accuracy of fault diagnosis.

[0144] Based on the same inventive concept as in the above-mentioned embodiment, this embodiment also provides a device for diagnosing faults of a ship steering gear, such as Fig.15 As shown, the device comprises:

[0145] The acquisition unit 151 is used to acquire multiple groups of target signals of the ship steering gear in a normal state and multiple groups of target signals of the ship steering gear in different fault states, and convert each group of the target signals into frequency domain signals; the target signals include: vibration signals and current signals;

[0146] The first conversion unit 152 is used to determine a target conversion factor of a symmetrical point pattern SDP image based on each group of the target signals and the corresponding frequency domain signals, and convert each group of target signals and the corresponding frequency domain signals in different states into a corresponding SDP image according to the target conversion factor to obtain an SDP image set;

[0147] A training unit 153 is used to train the constructed deep residual neural network model using the SDP image set, so that the deep residual neural network model learns the key features presented by the ship steering gear in a normal state and the key features presented in a fault state;

[0148] The second conversion unit 154 is used to obtain a target signal to be tested of the ship motor to be tested, convert the target signal to be tested into a frequency domain signal to be tested, and convert the target signal to be tested and the frequency domain signal to be tested into a SDP image to be tested;

[0149] The diagnosis unit 155 uses the trained deep residual neural network model to diagnose the SDP image to be tested to obtain a fault diagnosis result; wherein the first layer of the deep residual neural network is a convolutional neural network, and the convolutional neural network is used to perform convolution operations on the input SDP image set to enhance the ability of key feature extraction; an attention mechanism layer is introduced into the residual block of the deep residual neural network, and the attention mechanism layer is used to adaptively learn the feature weight of each feature in the SDP image set, and use the feature weight to achieve weighting to suppress several non-critical features in the SDP image set.

[0150] Since the device introduced in the embodiment of the present invention is a device used to implement the method for diagnosing a fault of a ship steering gear in the embodiment of the present invention, based on the method introduced in the embodiment of the present invention, a person skilled in the art can understand the specific structure and deformation of the device, so it is not described here in detail. All devices used in the method of the embodiment of the present invention belong to the scope of protection of the present invention.

[0151] Based on the same inventive concept, this embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any step of the method described above when executing the computer program.

[0152] Based on the same inventive concept, this embodiment provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the methods described above are implemented.

[0153] Through one or more embodiments of the present invention, the present invention has the following beneficial effects or advantages:

[0154] The present invention provides a method, device, medium and equipment for diagnosing faults of a ship steering gear. The method comprises: collecting multiple groups of target signals of the ship steering gear in a normal state and collecting multiple groups of target signals of the ship steering gear in different fault states, and converting each group of the target signals into a frequency domain signal; the target signal comprises: a vibration signal and a current signal; determining a target conversion factor of a symmetrical point pattern SDP image based on each group of the target signals and the corresponding frequency domain signal, converting each group of target signals and the corresponding frequency domain signal in different states into a corresponding SDP image according to the target conversion factor, and obtaining an SDP image set; training a constructed deep residual neural network model using the SDP image set, so that the deep residual neural network model learns the key features presented by the ship steering gear in a normal state and the key features presented in a fault state; obtaining a target signal to be tested of a ship motor to be tested, converting the target signal to be tested into a frequency domain signal to be tested, and converting the target signal to be tested and the frequency domain signal to be tested into an SDP image to be tested; using The SDP image to be tested is diagnosed using the trained deep residual neural network model; wherein the first layer of the deep residual neural network is a convolutional neural network, and the convolutional neural network is used to perform convolution operations on the SDP image set to enhance the ability to extract key features; an attention mechanism layer is introduced into the residual block of the deep residual neural network, and the attention mechanism layer is used to adaptively learn the feature weights of each feature in the SDP image set to suppress several non-key features in the SDP image set; in this way, since the vibration signal can reflect the vibration characteristics of the steering gear, and the current signal can reflect the circuit connection of the steering gear, the various different signals of the ship steering gear are fused in the two-dimensional polar coordinate axis to obtain the SDP image, which can increase the expression of the fault characteristics; and the optimal target conversion factor is first determined. After the SDP image set is determined using the target conversion factor, the accuracy of the feature expression of the SDP image set can be improved. After the deep residual neural network model is trained using the SDP image set, the recognition accuracy of the model is improved, thereby improving the accuracy of fault diagnosis.

[0155] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0156] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for diagnosing a fault of a ship steering gear, characterized in that: The method comprises: Collecting multiple groups of target signals of the ship steering gear in a normal state and collecting multiple groups of target signals of the ship steering gear in different fault states, and converting each group of the target signals into frequency domain signals; the target signals include: vibration signals and current signals; Determine a target conversion factor of a symmetrical point pattern SDP image based on each group of the target signals and the corresponding frequency domain signals, and convert each group of target signals and the corresponding frequency domain signals in different states into a corresponding SDP image according to the target conversion factor to obtain an SDP image set; The constructed deep residual neural network model is trained using the SDP image set, so that the deep residual neural network model learns the key features presented by the ship steering gear in a normal state and the key features presented in a fault state; Acquire a target signal to be measured of the ship motor to be measured, convert the target signal to be measured into a frequency domain signal to be measured, and convert the target signal to be measured and the frequency domain signal to be measured into a SDP image to be measured; The SDP image to be tested is diagnosed using the trained deep residual neural network model; wherein the first layer of the deep residual neural network is a convolutional neural network, and the convolutional neural network is used to perform convolution operations on the SDP image set to enhance the ability of key feature extraction; an attention mechanism layer is introduced into the residual block of the deep residual neural network, and the attention mechanism layer is used to adaptively learn the feature weights of each feature in the SDP image set to suppress several non-critical features in the SDP image set.

2. The method according to claim 1, characterized in that The conversion factor includes: a delay factor and an amplification factor; the target conversion factor of the symmetrical point mode SDP image is determined based on each group of the target signals and the corresponding frequency domain signals, including: Determine all value combinations of the delay factor and the amplification factor according to the value range of the delay factor and the value range of the amplification factor; The following processing is performed for each value combination: each group of target signals and corresponding frequency domain signals in each state are converted into reference SDP images according to the value combination; the number of reference SDP images in each value combination is consistent with the number of state types; for any two reference SDP images in the value combination, the similarity of the two reference SDP images is determined, thereby obtaining multiple similarities, and the average value of the multiple similarities is used as the benchmark similarity corresponding to the value combination; Obtain the baseline similarities of all value combinations, and determine the value combination corresponding to the minimum baseline similarity as the target value combination; The target conversion factor is determined according to the target value combination.

3. The method according to claim 2, characterized in that The determining the similarity of the two reference SDP images comprises: According to the formula Determine the similarity R between the i′th reference SDP image and the j′th reference SDP image under the value combination i′,j′ (N,M); where The N is a first pixel value matrix corresponding to the i′th reference SDP image, the M is a second pixel value matrix corresponding to the j′th reference SDP image, is the average value of all pixel values ​​in the first pixel value matrix, is the average value of all pixel values ​​in the second pixel value matrix, and the M mn is the pixel value corresponding to the mth row and nth column in the first pixel value matrix, wherein N mn is the pixel value corresponding to the mth row and nth column in the second pixel value matrix.

4. The method according to claim 1, characterized in that The converting each group of target signals and the corresponding frequency domain signals into a corresponding SDP image according to the target conversion factor includes: For the target signal and any signal point in the frequency domain signal of the target signal, the following processing is performed: According to the formula The signal point is converted into a polar coordinate signal; the polar coordinate signal includes: the polar coordinate radius r(i), the angle θ(i) of β rotating clockwise, the angle φ(i) of β rotating counterclockwise, the β is the rotation angle of the mirror symmetry plane, the x i is the i-th signal point, the x min is the signal point with the smallest amplitude among all signal points, and x max is the signal point with the largest amplitude among all signal points, t is the delay factor in the target conversion factor, z is the amplification factor in the target conversion factor, and x i+t is the amplitude of the i+tth signal point; Each signal point is mapped onto a polar coordinate axis according to the polar coordinates of each signal point to obtain the SDP image.

5. The method according to claim 1, characterized in that The deep residual neural network model includes: a convolutional neural network, a first convolutional layer, a batch normalization layer (BN layer), an activation function layer, a first pooling layer, a plurality of residual blocks, a second pooling layer and a fully connected layer in sequence; The convolutional neural network includes a plurality of second convolutional layer combinations and pooling layers in sequence, and an activation function layer is introduced after each second convolutional layer; the convolutional neural network is used to perform a convolution operation on the input SDP image set to obtain a first feature image; The first convolution layer is used to perform a convolution operation on the first feature image to obtain a second feature image; The BN layer is used to perform normalization processing on the second feature image to obtain a normalized feature image; The activation function layer is used to perform nonlinear processing on the normalized feature image to obtain a nonlinear normalized feature image; The first pooling layer is used to reduce the dimension of the nonlinear normalized feature image to obtain a low-dimensional feature image of the nonlinear normalized feature image; The plurality of residual blocks are used to extract key feature images from the low-dimensional feature images; The second pooling layer is used to continue the dimension reduction operation on the key feature image to obtain the key feature image after dimension reduction; The fully connected layer classifies each key feature image based on a built-in classification function to obtain a key feature image in a normal state and a key feature image in a fault state.

6. The method according to claim 5, characterized in that The residual block consists of: Convolutional layer, activation function layer and attention mechanism layer, a BN layer is introduced after each convolutional layer; among them, The attention mechanism layer includes: an average pooling layer, a depth-separable convolution layer and an activation function layer in sequence; the input feature image of the attention mechanism layer is the output feature image of the last BN layer in the residual block; The average pooling layer is used to average the pixel values ​​in each pooling window on the feature image output by the last BN layer in the residual block to generate a pooled feature image; The depth-separable convolution layer is used to group the pooled feature images to form groups of feature maps, and generate convolution kernels corresponding to each group of feature maps; convolve the corresponding group of feature maps using the convolution kernel of each group of feature maps to obtain the corresponding convolution feature images; and then convolve each of the convolution feature product images using a 1*1 convolution kernel to obtain a single-channel convolution image; The activation function layer is used to perform nonlinear processing on the single-channel convolution image.

7. The method according to claim 1, characterized in that The step of inputting the SDP image to be tested into the trained deep residual neural network model to obtain a fault diagnosis result includes: Using the deep residual neural network model to perform feature analysis on the SDP image to be tested to obtain multiple fault feature values; The fault feature classifier of the deep residual neural network model is used to determine the probability of each fault feature value, and the corresponding fault type is output according to the probability of the fault feature value.

8. A device for diagnosing faults of a ship steering gear, characterized in that: The device comprises: A collection unit is used to collect multiple groups of target signals of the ship steering gear in a normal state and multiple groups of target signals of the ship steering gear in different fault states, and convert each group of the target signals into frequency domain signals; the target signals include: vibration signals and current signals; A first conversion unit is used to determine a target conversion factor of a symmetrical point pattern SDP image based on each group of the target signals and the corresponding frequency domain signals, and convert each group of target signals and the corresponding frequency domain signals in different states into a corresponding SDP image according to the target conversion factor to obtain an SDP image set; A training unit, used to train the constructed deep residual neural network model using the SDP image set, so that the deep residual neural network model learns the key features presented by the ship steering gear in a normal state and the key features presented in a fault state; A second conversion unit is used to obtain a target signal to be tested of the ship motor to be tested, convert the target signal to be tested into a frequency domain signal to be tested, and convert the target signal to be tested and the frequency domain signal to be tested into an SDP image to be tested; The diagnostic unit uses the trained deep residual neural network model to diagnose the SDP image to be tested to obtain a fault diagnosis result; wherein the first layer of the deep residual neural network is a convolutional neural network, and the convolutional neural network is used to perform convolution operations on the input SDP image set to enhance the ability of key feature extraction; an attention mechanism layer is introduced into the residual block of the deep residual neural network, and the attention mechanism layer is used to adaptively learn the feature weight of each feature in the SDP image set, and use the feature weight to achieve weighting to suppress several non-critical features in the SDP image set.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method according to any one of claims 1 to 7 are implemented.

Images