Rotor lamination risk position positioning and stress prediction method and system

Abstract

The application relates to the technical field of artificial intelligence, in particular to a rotor lamination risk position positioning and stress prediction method and system, which comprises the following steps: acquiring an RGB image of a rotor lamination, rotation speed information and material information; constructing a multi-channel feature map based on the RGB image and the rotation speed information, wherein the multi-channel feature map comprises a rotation speed feature mapping diagram; positioning the risk position of the rotor lamination and predicting the stress according to the multi-channel feature map to obtain a positioning result of the risk position and an initial predicted stress value; and correcting the stress value based on the material information and the initial predicted stress value to obtain a final corrected stress value. The method can improve the positioning and stress prediction efficiency of the risk position of the rotor lamination, shorten the prediction cycle, has high accuracy and low cost.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and system for locating the risk position and predicting the stress of rotor laminations. Background Technology

[0002] With the rapid development of the new energy vehicle industry, drive motors are constantly evolving towards higher speeds and higher power densities. In the actual operation of the motor, the rotor laminations, as the core electromagnetic and mechanical load-bearing components, directly determine the motor's safety performance through their structural reliability. Rotor laminations typically have multiple non-uniform slots to accommodate permanent magnets, forming extremely narrow metal regions—external and internal magnetic bridges—between the slots and their outer edges, and between slots themselves.

[0003] Under high-speed rotation conditions, these abrupt structural changes, such as magnetic bridges, are subjected to enormous centrifugal forces, making them high-risk locations prone to stress concentration. To ensure the safe and stable operation of the motor under complex conditions, it is necessary to accurately locate and predict the stress at these risk points of the rotor laminations. In related technologies, the location and stress prediction of risk points in rotor laminations mainly rely on manual finite element simulation analysis. However, this method suffers from problems such as cumbersome modeling and debugging, low prediction efficiency, and long development cycles, making it difficult to adapt to the needs of rapid iterative development of motors. Summary of the Invention

[0004] This application provides a method and system for locating the risk location and predicting the stress of rotor laminations, in order to solve the technical problems of low prediction efficiency and long cycle in the location and stress prediction of the risk location of rotor laminations in related technologies.

[0005] This application provides a method for locating the risk position and predicting the stress of rotor laminations. The method includes: Acquire RGB images, rotational speed information, and material information of the rotor laminations; Based on the RGB image and the rotation speed information, a multi-channel feature map is constructed, which includes a rotation speed feature map; Based on the multi-channel feature map, the risk location of the rotor lamination is located and the stress is predicted, so as to obtain the location result of the risk location and the initial predicted stress value. Based on the material information and the initial predicted stress value, the stress value is corrected to obtain the corrected final stress value.

[0006] In one embodiment of this application, the rotational speed information includes rotational speed values, and the step of constructing a multi-channel feature map based on the RGB image and the rotational speed information includes: The difference between the speed value and the preset minimum speed reference value is determined as a first value, the difference between the preset maximum speed reference value and the minimum speed reference value is a second value, and the ratio between the first value and the second value is a third value. The product of the third value and the preset channel intensity value is determined as the fourth value, and the sum of the fourth value and the preset normalization coefficient is the target pixel value. Based on the target pixel value, the rotational speed feature map is constructed. The rotational speed feature map is superimposed on the three-channel feature map of the RGB image to obtain the multi-channel feature map.

[0007] In one embodiment of this application, the step of locating the risk location and predicting the stress of the rotor lamination based on the multi-channel feature map, and obtaining the location result of the risk location and the initial predicted stress value, includes: The multi-channel feature map is input into a preset target model to locate the risk position of the rotor lamination and predict the stress, so as to obtain the location result and the initial predicted stress value output by the target model, wherein the target model includes: A feature extraction network is used to extract features based on the multi-channel feature map to obtain shallow features, mid-level features, and deep features. The target downsampling factor corresponding to the shallow features, the mid-level features, and the deep features shows an increasing trend. The target downsampling factor represents the downsampling factor relative to the size of the multi-channel feature map. A feature enhancement network is used to enhance the shallow features, the middle features, and the deep features to obtain shallow enhanced features, middle enhanced features, and deep enhanced features. The prediction network is used to locate the risk position and predict the stress of the rotor lamination based on the shallow enhancement features, the middle enhancement features, and the deep enhancement features, and outputs the location result of the risk position and the initial predicted stress value.

[0008] In one embodiment of this application, the feature extraction network includes multiple downsampling layers and multiple buffer layers, with the buffer layers located between adjacent downsampling layers; The multi-layer downsampling layer includes: a first downsampling layer, a second downsampling layer, a third downsampling layer, a fourth downsampling layer, and a fifth downsampling layer arranged in sequence. The input of the first downsampling layer is the multi-channel feature map, the output of the third downsampling layer is the shallow feature, the output of the fourth downsampling layer is the middle feature, and the output of the fifth downsampling layer is the deep feature. The multi-layer buffer layer includes a first buffer layer and a second buffer layer; the first buffer layer is located between the first downsampling layer and the second downsampling layer, and the second buffer layer is located between the third downsampling layer and the fourth downsampling layer.

[0009] In one embodiment of this application, the feature enhancement network includes: The first network branch is used to receive the shallow features, perform feature enhancement on the shallow features, obtain the shallow enhanced features, and output them. The second network branch is used to receive the mid-layer features, perform feature enhancement on the mid-layer features, obtain the mid-layer enhanced features, and output them. The shallow feature weight extraction branch is used to access the shallow features, and based on the shallow features and the preset channel attention mechanism, channel weighting is performed to obtain the target features; The third network branch is used to receive the deep features and the target features, perform feature enhancement based on the deep features and the target features, obtain the deep enhanced features, and output them.

[0010] In one embodiment of this application, the shallow feature weight extraction branch includes: a first two-dimensional convolutional layer, a second two-dimensional convolutional layer, and a channel attention mechanism layer connected in sequence, wherein the input of the first two-dimensional convolutional layer is the shallow feature; the channel attention mechanism layer includes: a global average pooling layer, a modified linear layer, and an activation function layer connected in sequence, wherein the output of the activation function layer is the target feature; The third network branch includes: The first spatial pyramid pooling network includes a first splicing layer and multiple parallel first soft pooling layers. The first soft pooling layers are used to perform soft pooling processing on the deep features to obtain first soft pooled features. The pooling kernels of the multiple first soft pooling layers are different. The first splicing layer is used to splice the deep features and the multiple first soft pooling features to obtain a first feature. The first target convolutional layer includes a first two-dimensional convolutional sub-layer, a first batch normalization sub-layer, and a first activation function sub-layer connected in sequence. The input of the first two-dimensional convolutional sub-layer is the first feature, and the output of the first activation function sub-layer is the second feature. The second feature is the data source for the second network branch to perform feature enhancement. A third two-dimensional convolutional layer is used to receive a third feature, which is the product of the target feature and the second feature; the third feature is convolved to obtain the deep enhancement feature.

[0011] In one embodiment of this application, the second network branch includes: The second spatial pyramid pooling network includes a second splicing layer and multiple parallel second soft pooling layers. The second soft pooling layers are used to perform soft pooling processing on the middle layer features to obtain second soft pooled features. The pooling kernels of the multiple second soft pooling layers are different. The second splicing layer is used to splice the middle layer features and the multiple second soft pooling features to obtain a fourth feature. The third splicing layer is used to splice the fourth and fifth features to obtain the sixth feature, wherein the fifth feature is the feature obtained by performing two-dimensional convolution and upsampling twice the space on the second feature; The second target convolutional layer includes a second two-dimensional convolutional sub-layer, a second batch normalization sub-layer, and a second activation function sub-layer connected in sequence. The input of the second two-dimensional convolutional sub-layer is the sixth feature, and the output of the second activation function sub-layer is the seventh feature. The seventh feature is the data source for feature enhancement of the first network branch. The fourth two-dimensional convolutional layer is used to perform two-dimensional convolution on the seventh feature to obtain the middle-layer enhanced feature.

[0012] In one embodiment of this application, the first network branch includes: The third spatial pyramid pooling network includes a third splicing layer and multiple parallel third soft pooling layers. The third soft pooling layers are used to perform soft pooling on the shallow features to obtain third soft pooled features. The pooling kernels of the multiple third soft pooling layers are different. The third splicing layer is used to splice the shallow features and the multiple third soft pooled features to obtain an eighth feature. The fourth splicing layer is used to splice the eighth and ninth features to obtain the tenth feature, wherein the ninth feature is obtained by performing two-dimensional convolution and double spatial upsampling on the seventh feature; The third target convolutional layer includes a third two-dimensional convolutional sub-layer, a third batch normalization sub-layer, and a third activation function sub-layer connected in sequence. The input of the third two-dimensional convolutional sub-layer is the tenth feature, and the output of the third activation function sub-layer is the eleventh feature. The fifth two-dimensional convolutional layer is used to perform two-dimensional convolution on the eleventh feature to obtain the shallow enhanced feature.

[0013] In one embodiment of this application, the material information includes the Poisson's ratio, elastic modulus, and density of the rotor lamination. The step of correcting the stress value based on the material information and the initial predicted stress value to obtain the corrected final stress value includes: The material information and the initial predicted stress value are input into a preset stress correction model to correct the stress value and obtain the final stress value. The stress correction model includes an input layer, a hidden layer, and an output layer connected in sequence. The hidden layer includes multiple parallel neurons for stress value correction.

[0014] This application also provides a rotor lamination risk location and stress prediction system, the system comprising: The data acquisition module is used to acquire RGB images, rotational speed information, and material information of the rotor laminations; A multi-channel feature map construction module is used to construct a multi-channel feature map based on the RGB image and the rotation speed information, wherein the multi-channel feature map includes a rotation speed feature map; The risk location and stress prediction module is used to locate the risk location and predict the stress of the rotor lamination based on the multi-channel feature map, and obtain the location result of the risk location and the initial predicted stress value. The stress value correction module is used to correct the stress value based on the material information and the initial predicted stress value to obtain the corrected final stress value.

[0015] The beneficial effects of this application are as follows: This application proposes a method and system for locating the risk location and predicting the stress of a rotor lamination. The method includes: acquiring RGB images, rotational speed information, and material information of the rotor lamination; constructing a multi-channel feature map based on the RGB images and rotational speed information, the multi-channel feature map including a rotational speed feature mapping map; locating the risk location and predicting the stress of the rotor lamination based on the multi-channel feature map, obtaining the location result of the risk location and the initial predicted stress value; and correcting the stress value based on the material information and the initial predicted stress value to obtain the corrected final stress value. This method can significantly improve the efficiency of locating the risk location and predicting the stress of the rotor lamination, shorten the prediction cycle, achieve high accuracy, and has low cost. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0017] In the attached diagram: Figure 1 This is a top view of a rotor lamination provided in one embodiment of this application; Figure 2 This is a flowchart illustrating the rotor lamination risk location and stress prediction method provided in one embodiment of this application; Figure 3This is a schematic diagram of a rotor lamination image with numerical simulation information provided in one embodiment of this application; Figure 4 This is a schematic diagram of a rotor lamination image with annotation information provided in one embodiment of this application; Figure 5 This is a schematic diagram of the target model in the rotor lamination risk location and stress prediction method provided in one embodiment of this application; Figure 6 This is a schematic diagram of the channel attention mechanism layer in the rotor lamination risk location and stress prediction method provided in one embodiment of this application; Figure 7 This is a schematic diagram of the structure of the first spatial pyramid pooling network in the rotor lamination risk location and stress prediction method provided in one embodiment of this application. Figure 8 This is a schematic diagram of the bottleneck network in the rotor lamination risk location and stress prediction method provided in one embodiment of this application. Figure 9 This is a schematic diagram of the stress correction model in the rotor lamination risk location and stress prediction method provided in one embodiment of this application; Figure 10 This is a schematic diagram of the rotor lamination risk location and stress prediction system provided in one embodiment of this application; Figure 11 This is a schematic diagram of the structure of an electronic device provided in one embodiment of this application. Detailed Implementation

[0018] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0019] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0020] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the present application. However, it will be apparent to those skilled in the art that embodiments of the present application may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present application.

[0021] With the booming development of the new energy vehicle industry, drive motors, as the core power source of vehicles, are undergoing profound technological evolution towards higher speeds, higher power densities, and higher efficiency. In permanent magnet synchronous motors, the rotor laminations are crucial electromagnetic and mechanical load-bearing composite components. To accommodate permanent magnets and guide magnetic lines of force, the rotor laminations typically have multiple weight-reducing holes or non-uniform slots inside. Figure 1 This is a top view of a rotor lamination provided in one embodiment of this application. The structure of the rotor lamination can be referenced. Figure 1 Between these slots and the outer edge of the rotor, as well as between adjacent slots, extremely narrow metal connection areas are inevitably formed, which are referred to in engineering as external magnetic bridges and internal magnetic bridges, respectively.

[0022] From an electromagnetic design perspective, the narrower the width of the outer and inner magnetic bridges, the weaker the leakage flux generated by the permanent magnets, resulting in higher torque density and overall operating efficiency of the motor. However, from the perspective of solid mechanics and structural strength, when the motor is operating at high speed, the rotor assembly generates enormous centrifugal forces. These centrifugal forces are primarily borne by the outer and inner magnetic bridges. Due to the drastic abrupt change in the geometric cross-section of the magnetic bridge region, severe stress concentration is easily triggered at this point. Once the local maximum stress exceeds the yield strength or even tensile strength of the rotor lamination material, the rotor lamination will undergo irreversible plastic deformation or even fracture failure, leading to consequences such as motor rubbing and seizure. Therefore, accurate positioning and high-precision stress prediction of the risk locations of the rotor laminations (including the outer and inner magnetic bridges) are indispensable core aspects of motor development. However, methods such as manual finite element simulation analysis in related technologies suffer from problems such as cumbersome modeling and debugging, low prediction efficiency, and long cycles. Based on this, this application proposes a method and system for locating the risk locations and predicting the stress of rotor laminations to solve the aforementioned technical problems.

[0023] The method and system for locating risk positions and predicting stress in rotor laminations provided in this application will be explained and illustrated below by way of examples.

[0024] Please see Figure 2 , Figure 2 This is a flowchart illustrating the rotor lamination risk location and stress prediction method provided in one embodiment of this application, as shown below. Figure 2 As shown, the method includes: S210: Acquire RGB images, rotational speed information, and material information of the rotor laminations.

[0025] In some examples of this embodiment, the RGB image is a 3-channel color image that contains two-dimensional topological information of the rotor laminations, such as the outer magnetic bridge, inner magnetic bridge, and the shape, size, and relative positional relationship of the non-uniform slots. Rotational speed information reflects the dynamic centrifugal force load level experienced by the rotor during actual operation. Material information reflects the mechanical properties of the rotor lamination material.

[0026] S220: Based on RGB images and rotational speed information, a multi-channel feature map is constructed, which includes a rotational speed feature map.

[0027] In step S220, a multi-channel feature map containing a speed feature map is constructed based on the RGB image and speed information, which effectively breaks through the limitation of traditional image processing relying solely on visual features. The one-dimensional physical load scalar (speed information) is transformed into a two-dimensional spatial feature map and stitched with the RGB image in the channel dimension, thereby achieving deep coupling between structural features and operating condition features on the data input side. This helps improve the accuracy of subsequent rotor lamination risk location positioning and stress prediction.

[0028] S230: Based on the multi-channel feature map, locate the risk position of the rotor lamination and predict the stress, and obtain the location result of the risk position and the initial predicted stress value.

[0029] In some examples of this embodiment, the risky locations of the rotor laminations include locations where stress concentration is likely to occur, such as the outer magnetic bridge and the inner magnetic bridge.

[0030] S240: Based on material information and initial predicted stress values, stress values ​​are corrected to obtain the corrected final stress values.

[0031] In step S240, material information is introduced into the prediction process as a physical constraint, which can better correct the stress prediction deviation caused by the lack of material information in step S230, improve the accuracy of stress prediction, and help improve the accuracy and physical interpretability of stress prediction.

[0032] The rotor lamination risk location and stress prediction method in this embodiment can significantly improve the efficiency of rotor lamination risk location and stress prediction, shorten the prediction cycle, achieve high accuracy, low cost, and strong feasibility. This method can achieve millisecond-level stress assessment of risk locations, significantly shortening the rotor lamination structure design iteration cycle and improving R&D efficiency.

[0033] In some embodiments, the rotational speed information includes rotational speed values. Based on the RGB image and the rotational speed information, a multi-channel feature map is constructed, including: S2201: The difference between the speed value and the preset minimum speed reference value is determined as the first value, the difference between the preset maximum speed reference value and the minimum speed reference value is the second value, and the ratio between the first value and the second value is the third value.

[0034] S2202: The product of the third value and the preset channel intensity value is determined as the fourth value, and the sum of the fourth value and the preset normalization coefficient is the target pixel value.

[0035] In some examples of this embodiment, the mathematical expression of the target pixel value can be: (Equation 1) Where k represents the preset channel intensity value, which can be the maximum quantization value of a single channel (255), etc. The representation is mapped to [0, The target pixel value within the range, Characterizing the rotational speed value, Characterizing the minimum speed reference value, Characterizing the reference value of maximum speed, The normalization coefficient is used to represent the normalization factor, which can be set according to the actual situation, such as 0.5, etc. " represents the dot product symbol.

[0036] Understandably, by executing the above steps S2201 to S2202, the normalization of the speed operating condition is achieved quite well.

[0037] S2203: Based on the target pixel value, complete the construction of the rotational speed feature map.

[0038] In some examples of this embodiment, a two-dimensional matrix with the same size as the RGB image (e.g., 416×416) can be generated based on the calculated target pixel value. The value of all elements in the two-dimensional matrix is ​​the target pixel value, thereby completing the construction of the rotational speed feature map.

[0039] S2204: Superimpose the rotational speed feature map onto the three-channel feature map of the RGB image to obtain a multi-channel feature map.

[0040] In some examples of this embodiment, the multi-channel feature map is a four-channel input feature map.

[0041] In some examples of this embodiment, the rotational speed feature map can be used as the fourth channel and superimposed on the three-channel feature map of an RGB image with a size of 416×416×3 to obtain a four-channel feature map with a size of 416×416×4.

[0042] Understandably, by overlaying the rotational speed feature map onto the three-channel feature map of the RGB image, channel fusion of the RGB image and the rotational speed condition is effectively achieved. In subsequent prediction, the convolutional kernel can simultaneously capture the local correlation between structural abrupt changes and rotational speed intensity within the receptive field, which helps the target model to adaptively perceive different rotational speed conditions, significantly improving the robustness and accuracy of prediction, and also enhancing the model's multi-condition generalization ability.

[0043] In some embodiments, based on the multi-channel feature map, the risk location of the rotor lamination is located and stress is predicted to obtain the location result of the risk location and the initial predicted stress value, including: S2301: Input the multi-channel feature map into the preset target model to locate the risk position of the rotor lamination and predict the stress, so as to obtain the positioning result and the initial predicted stress value output by the target model.

[0044] In some examples of this embodiment, the training process of the target model includes: First, a first training set and a first validation set are constructed. Specifically, a rotor lamination geometric model is established, where the width of the outer magnetic bridge can range from 0.4 to 1.2 mm, and the width of the inner magnetic bridge can range from 0.8 to 2.5 mm. The speed range is set from 10,000 rpm to 22,000 rpm, increasing in 200 rpm increments, and includes a 1.2x overspeed condition. Using the rotor lamination geometric model as the base model, a finite element model is established. Using the finite element model, multiple stress distribution structure images are generated. Based on these multiple stress distribution structure images, different speed conditions and materials are set to complete the batch generation of rotor lamination images with numerical simulation information. The rotor lamination images with numerical simulation information can be referenced... Figure 3 . Figure 3 In this context, Value represents numerical simulation information, specifically the simulated stress value. Figure 3 The red area in the image represents the inner magnetic bridge, and the marked area around the rotor lamination represents the outer magnetic bridge. After batch generation of rotor lamination images with numerical simulation information, these images can be scaled proportionally, such as to 416x416 pixels. Information annotation is then performed on these rotor lamination images, including marking the rectangles containing the outer and inner magnetic bridges, as well as the corresponding actual stress values, resulting in rotor lamination images with annotation information. Figure 4 An illustrative example is shown of rotor lamination images with annotation information. Figure 4The image contains the true stress value corresponding to the position within the rectangle, as well as the true confidence level (0 or 1). Both the first training set and the first validation set contain multiple rotor lamination images with labeled information. The true stress value is the value obtained after normalization. For example, the difference between the original stress value obtained from numerical simulation and the preset minimum stress value (the minimum value among the original stress values ​​obtained from numerical simulation) is determined as the first intermediate value, the difference between the preset maximum stress value (the maximum value among the original stress values ​​obtained from numerical simulation) and the minimum stress value is determined as the second intermediate value, and the ratio between the first intermediate value and the second intermediate value is determined as the normalized true stress value.

[0045] Secondly, based on the rotor lamination image with annotation information and the speed range, a multi-channel feature map sample containing a speed feature map is constructed. The method for constructing the multi-channel feature map sample can refer to steps S2201 to S2204 in the above embodiment, and will not be repeated here.

[0046] Then, the multi-channel feature map samples are input into a preset initial model (the target model that has not yet been trained), and the predicted location information, prediction confidence, and predicted stress value of the rectangular boxes at the risk locations (outer magnetic bridge, inner magnetic bridge) output by the initial model are obtained. The predicted location information of the rectangular boxes includes the predicted width and height values ​​of the rectangular boxes and the predicted coordinates of the center point.

[0047] Subsequently, based on the predicted location information, prediction confidence, predicted stress value, and corresponding annotation information (true location information, true confidence, and true stress value of the rectangular frame at the risk location (outer magnetic bridge, inner magnetic bridge), and a preset loss function, a composite loss is obtained. The true location information of the rectangular frame includes the true width and height values ​​of the rectangular frame and the true coordinates of the center point.

[0048] The mathematical expression for this composite loss can be: (Equation 2) in, Characterizing composite loss, Characterizing the width and height loss of the rectangular frame, Characterizing the loss at the center position of the rectangular box, Characterizing confidence loss, Characterizes stress loss.

[0049] (Equation 3) in, This represents the actual width and height values ​​of the i-th rectangle. This represents the predicted width and height values ​​of the i-th rectangle. This indicates the width and height of the rectangle.

[0050] (Equation 4) in, Indicates the number of rectangles. The true coordinates of the center point of the i-th rectangle are represented. The predicted coordinates of the center point of the i-th rectangle are represented. Representing the coordinates of the center point, This indicates that the logarithm is being calculated.

[0051] (Equation 5) in, Characterizes the true confidence level of the i-th rectangle. The prediction confidence level is represented by the i-th rectangle.

[0052] (Equation 6) in, The number of multi-channel feature map samples. This indicates the number of risk locations in a multi-channel feature map sample. This represents the product of the number of risk locations in a multi-channel feature map sample and the number of multi-channel feature map samples. This represents the true stress value at the ith risk location in the p-th multi-channel feature map sample. This represents the predicted stress value at the i-th risk location in the p-th multi-channel feature map sample.

[0053] Finally, based on this composite loss, the initial model is trained to obtain the target model. Additionally, the target model can be validated using the first validation set.

[0054] In the above training process, the initial learning rate can be 0.01, the weight decay can be set to 0.0001, the momentum can be 0.9, the batch size can be set to 10, the epoch can be set to 40, and an early stopping mechanism is introduced to calculate the composite loss after each training round to prevent overfitting.

[0055] Figure 5 This is a schematic diagram of the target model in the rotor lamination risk location and stress prediction method provided in one embodiment of this application. Please refer to... Figure 5 In some embodiments, the target model includes: The feature extraction network is used to extract features based on multi-channel feature maps, obtaining shallow features, mid-level features, and deep features. The target downsampling factor corresponding to the shallow features, mid-level features, and deep features shows an increasing trend. The target downsampling factor represents the downsampling factor relative to the size of the multi-channel feature map. Feature enhancement networks are used to enhance shallow, medium, and deep features, resulting in shallow enhanced features, medium enhanced features, and deep enhanced features. The prediction network is used to locate the risk location and predict the stress of rotor laminations based on shallow, medium and deep reinforcement features, and outputs the location result of the risk location and the initial predicted stress value.

[0056] In some examples of this embodiment, the prediction network may employ three convolutional layers (e.g., Figure 5 The system uses the first, second, and third convolutional layers (in the database) to obtain the location results of the risk location and the initial predicted stress value, and outputs them. Figure 5 The feature map size output by each network layer is illustrated in the example, and will not be repeated here.

[0057] Understandably, setting up the above model structure can help improve the model's prediction accuracy and reduce the amount of model data.

[0058] In some embodiments, the feature extraction network includes multiple downsampling layers and multiple buffer layers, with the buffer layers located between adjacent downsampling layers; The multi-layer downsampling layer includes: a first downsampling layer, a second downsampling layer, a third downsampling layer, a fourth downsampling layer, and a fifth downsampling layer arranged in sequence. The input of the first downsampling layer is a multi-channel feature map, the output of the third downsampling layer is a shallow feature, the output of the fourth downsampling layer is a medium feature, and the output of the fifth downsampling layer is a deep feature. The multi-layer buffer layer includes a first buffer layer and a second buffer layer; the first buffer layer is located between the first downsampling layer and the second downsampling layer, and the second buffer layer is located between the third downsampling layer and the fourth downsampling layer.

[0059] In some examples of this embodiment, the feature extraction network employs a lightweight network architecture. The feature extraction network progressively reduces the spatial resolution of the multi-channel feature map through five consecutive downsampling layers. For example, the first downsampling layer receives a 416×416×4 multi-channel feature map; after the first three downsampling steps, the third downsampling layer outputs shallow features with a spatial resolution of 52×52 (downsampling by 8 times); the fourth downsampling layer outputs mid-level features with a spatial resolution of 26×26 (downsampling by 16 times); and the fifth downsampling layer outputs deep features with a spatial resolution of 13×13 (downsampling by 32 times).

[0060] In some examples of this embodiment, the buffer layer may contain depthwise separable convolutions and linear bottleneck structures, which greatly reduces the number of multiply-accumulate operations and the number of model parameters.

[0061] Furthermore, the feature extraction network employs the ReLU6 (rectified linear unit with an upper limit of 6) activation function in the first, second, and third downsampling layers, as well as the first buffer layer, to ensure computational speed. The h-Swish activation function (hard Swish activation function) is used in the second buffer layer, fourth downsampling layer, and fifth downsampling layer to improve the stability of high-dimensional feature extraction.

[0062] This lightweight design effectively reduces the number of model parameters and computational load, thereby improving inference speed.

[0063] In some embodiments, the feature enhancement network includes: The first network branch is used to receive shallow features, perform feature enhancement on the shallow features, obtain shallow enhanced features, and output them. The second network branch is used to receive the mid-layer features, perform feature enhancement on the mid-layer features, obtain the mid-layer enhanced features, and output them. The shallow feature weight extraction branch is used to access shallow features. Based on the shallow features and the preset channel attention mechanism, channel weighting is performed to obtain the target features. The third network branch is used to receive deep features and target features, perform feature enhancement based on the deep features and target features, obtain deep enhanced features, and output them.

[0064] Understandably, the outer and inner magnetic bridges of the rotor lamination occupy a very small proportion of pixels in the entire image, making it a typical small target detection task. While shallow features retain good spatial resolution, the semantic information between channels is rather cluttered. This embodiment introduces a shallow feature weight extraction branch and utilizes a channel attention mechanism to redistribute global information from shallow features. This allows the network to focus on the feature channels most valuable for risk location and stress prediction, significantly improving the detection recall and localization accuracy of small targets.

[0065] In some embodiments, the shallow feature weight extraction branch includes: a first two-dimensional convolutional layer, a second two-dimensional convolutional layer, and a channel attention mechanism layer connected in sequence. The input of the first two-dimensional convolutional layer is the shallow feature. The channel attention mechanism layer includes: a global average pooling layer, a modified linear layer, and an activation function layer connected in sequence. The output of the activation function layer is the target feature.

[0066] Figure 6 This is a schematic diagram of the channel attention mechanism layer in the rotor lamination risk location and stress prediction method provided in one embodiment of this application. Please refer to [the diagram]. Figure 6The global average pooling layer compresses the two-dimensional spatial features of each channel into a single real number. That is, it compresses the H×W (height × width) two-dimensional spatial features of each channel into a single numerical value, representing the average activation intensity of that channel globally. After compression, the feature shape changes from H×W×C (height × width × number of channels) to 1×1×C. The corrected linear layer (with a compression factor of up to 4) learns the non-linear dependencies between channels; in this layer, the number of channels is reduced to a fraction of the original value. r is a preset reduction rate, such as 16. The shape of the feature changes from 1×1×C to 1×1×C / r. The activation function layer (such as the Sigmoid function) outputs a normalized weight vector, which is multiplied with the shallow features to obtain the target feature.

[0067] In some embodiments, the third network branch includes: The first spatial pyramid pooling network SPP1 includes a first concatenation layer and multiple parallel first soft pooling layers. The first soft pooling layers are used to perform soft pooling on deep features to obtain first soft pooled features. The pooling kernels of the multiple first soft pooling layers are different. The first concatenation layer is used to concatenate the deep features and the multiple first soft pooling features to obtain the first feature. The first target convolutional layer includes a first two-dimensional convolutional sub-layer, a first batch normalization sub-layer, and a first activation function sub-layer connected in sequence. The input of the first two-dimensional convolutional sub-layer is a first feature, and the output of the first activation function sub-layer is a second feature. The second feature is the data source for the second network branch to perform feature enhancement. The third two-dimensional convolutional layer is used to receive the third feature, which is the product of the target feature and the second feature; convolution is performed on the third feature to obtain the deep enhancement feature.

[0068] The aforementioned first spatial pyramid pooling network SPP1, as a multi-scale pooling module, introduces soft pooling to replace the traditional MaxPool (maximum pooling). Max pooling directly discards all activation values ​​within the pooling kernel except for the maximum value, resulting in a significant loss of sub-pixel-level stress gradient information in the micro-magnetic bridge region. SoftPool retains all activation values ​​within the local region through a weighted operation based on Softmax (soft maxima), and its calculation formula is as follows: (Equation 7) in, Characterize the pooling kernel region, Representing the d-th activation value in R, The weight assigned to the d-th activation value is represented by the weight assigned to the d-th activation value. It is a natural constant. The j-th activation value in R is represented.

[0069] (Equation 8) in, This represents the first soft pooling feature.

[0070] Multiple first soft pooling layers with different pooling kernel sizes (such as 5×5, 9×9, 13×13, etc.) extract multi-scale first soft pooling features, which are then concatenated with deep features through a first splicing layer. This design endows the network with rich multi-scale information with very few additional parameters, greatly improving the prediction accuracy of small magnetic bridge stress gradients.

[0071] Figure 7 This is a schematic diagram of the structure of the first spatial pyramid pooling network SPP1 in the rotor lamination risk location and stress prediction method provided in one embodiment of this application. Figure 7 The network structure is illustrated using three parallel first soft pooling layers as an example.

[0072] In some embodiments, the second network branch includes: The second spatial pyramid pooling network SPP2 includes a second concatenation layer and multiple parallel second soft pooling layers. The second soft pooling layers are used to perform soft pooling on the mid-layer features to obtain second soft pooled features. The pooling kernels of the multiple second soft pooling layers are different. The second concatenation layer is used to concatenate the mid-layer features and multiple second soft pooling features to obtain a fourth feature. The third concatenation layer is used to concatenate the fourth and fifth features to obtain the sixth feature. The fifth feature is obtained by performing two-dimensional convolution and upsampling of the second feature twice. The second target convolutional layer includes a second two-dimensional convolutional sub-layer, a second batch normalization sub-layer, and a second activation function sub-layer connected in sequence. The input of the second two-dimensional convolutional sub-layer is the sixth feature, and the output of the second activation function sub-layer is the seventh feature. The seventh feature is the data source for feature enhancement in the first network branch. The fourth two-dimensional convolutional layer is used to perform two-dimensional convolution on the seventh feature to obtain the mid-layer enhanced feature.

[0073] Understandably, adopting the above network architecture can help improve prediction accuracy.

[0074] In some embodiments, the first network branch includes: The third spatial pyramid pooling network SPP3 includes a third splicing layer and multiple parallel third soft pooling layers. The third soft pooling layers are used to perform soft pooling on shallow features to obtain third soft pooled features. The pooling kernels of the multiple third soft pooling layers are different. The third splicing layer is used to splice the shallow features and multiple third soft pooling features to obtain the eighth feature. The fourth concatenation layer is used to concatenate the eighth and ninth features to obtain the tenth feature. The ninth feature is obtained by performing a two-dimensional convolution and a double spatial upsampling on the seventh feature. The third target convolutional layer includes a third two-dimensional convolutional sub-layer, a third batch normalization sub-layer, and a third activation function sub-layer connected in sequence. The input of the third two-dimensional convolutional sub-layer is the tenth feature, and the output of the third activation function sub-layer is the eleventh feature. The fifth two-dimensional convolutional layer is used to perform two-dimensional convolution on the eleventh feature to obtain shallow enhanced features.

[0075] Through the design of the second and first network branches described above, top-down feature pyramid fusion was achieved. Shallow features, while having their receptive field expanded by the SoftPool, retained high-resolution edge information and were integrated with powerful semantic guidance from deeper layers. This directly addresses the issues of missed detections and bounding box drift in complex backgrounds involving small magnetic bridges, achieving high-precision risk location locking.

[0076] In some embodiments, the first downsampling layer may employ a two-dimensional convolutional network. The second downsampling layer, the third downsampling layer, the fourth downsampling layer, the fifth downsampling layer, the first buffer layer, and the second buffer layer may all employ bottleneck networks. Figure 8 This is a schematic diagram of the bottleneck network in the rotor lamination risk location and stress prediction method provided in one embodiment of this application. Please refer to... Figure 8 The bottleneck network consists of a dilation layer, a deep convolutional layer, a channel attention mechanism layer, and a linear bottleneck layer, connected in sequence. The dilation layer uses a regular 1×1 convolution to expand the number of channels in the input feature map (e.g., to 4 to 6 times the original number), allowing the network to extract richer feature information in a higher-dimensional space. The deep convolutional layer uses 3×3 or 5×5 depthwise separable convolutions for feature extraction, thus processing high-dimensional feature maps while significantly reducing the number of parameters. The linear bottleneck layer uses a regular 1×1 convolution to compress the number of channels back to a lower dimension.

[0077] In some embodiments, the material information includes the Poisson's ratio, elastic modulus, and density of the rotor laminations. Based on the material information and the initial predicted stress value, a stress value correction is performed to obtain the corrected final stress value, including: S2401: Input the material information and the initial predicted stress value into the preset stress correction model, correct the stress value, and obtain the final stress value.

[0078] In some embodiments, training the stress correction model includes: First, a second training set and a second validation set are constructed. Specifically, various materials are selected, and finite element analysis is performed at 1000 rpm intervals using Poisson's ratio, elastic modulus, and density as parameters to generate multiple sets of sample data, thus completing the second training set and the second validation set. The sample data includes simulated material information, simulated predicted stress values, and stress label values.

[0079] Secondly, the simulated material information and the simulated predicted stress values ​​are input into the initial propagation model (the stress correction model that has not been trained) to obtain the stress correction values ​​output by the initial propagation model.

[0080] Finally, based on the stress correction value, the corresponding stress label value, and the preset mean absolute error loss function, the initial propagation model is trained to obtain the stress correction model.

[0081] The mathematical expression for the mean absolute error loss function is: (Equation 9) in, The mean absolute error loss is represented by F, the number of sample data points is represented by G, and the number of risk locations is represented by G. This represents the product of the number of sample data points and the number of risk locations. This represents the stress label value at the g-th risk location of the f-th sample data. This represents the stress correction value at the g-th risk location of the f-th sample data.

[0082] Understandably, the above method can yield a stress correction model with high accuracy. This stress correction model can be a backpropagation model, etc.

[0083] In some embodiments, the stress correction model includes an input layer, a hidden layer, and an output layer connected in sequence, wherein the hidden layer includes a plurality of parallel neurons for stress value correction.

[0084] Figure 9 This is a schematic diagram of the stress correction model in the rotor lamination risk location and stress prediction method provided in one embodiment of this application. Please refer to [the diagram]. Figure 9 Each parameter (x1, x2, x3, x4) in the input layer is input to each neuron in the hidden layer. The neuron adjusts its stress value based on its preset weights and the input parameters. Figure 9 The example demonstrates 6 neurons. The corrected formula can be Y=WX+B, Z= In this system, Y represents the data obtained after weighting by neurons, which includes y1, y2, y3, y4, y5, and y6. W represents the weight of each neuron, X represents the parameters input to the neuron, and Z represents the output of the neuron, which includes z1, z2, z3, z4, z5, and z6. Characterizes the preset correction function, for example: The output layer obtains the final stress value based on the output of each neuron.

[0085] Figure 10 This is a schematic diagram of the rotor lamination risk location and stress prediction system provided in one embodiment of this application. Please refer to [the diagram]. Figure 10 This embodiment also provides a rotor lamination risk location and stress prediction system, which includes: The data acquisition module 1010 is used to acquire RGB images, rotational speed information, and material information of the rotor laminations; The multi-channel feature map construction module 1020 is used to construct a multi-channel feature map based on the RGB image and rotation speed information. The multi-channel feature map includes a rotation speed feature map. The risk location and stress prediction module 1030 is used to locate the risk location and predict the stress of the rotor lamination based on the multi-channel feature map, and obtain the location result of the risk location and the initial predicted stress value. The stress value correction module 1040 is used to correct the stress value based on the material information and the initial predicted stress value, so as to obtain the corrected final stress value.

[0086] It should be noted that the rotor lamination risk location and stress prediction method and system provided in the above embodiments belong to the same concept. The specific operation methods of each module have been described in detail in the method embodiments and will not be repeated here. In practical applications, the rotor lamination risk location and stress prediction system provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0087] Figure 11 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 11As shown, the electronic device may include a processor 1110, a communication interface 1120, a memory 1130, and a communication bus 1140. The processor 1110, communication interface 1120, and memory 1130 communicate with each other via the communication bus 1140. The processor 1110 can call logical instructions in the memory 1130 to execute a rotor lamination risk location and stress prediction method. This method includes: acquiring RGB images, rotational speed information, and material information of the rotor lamination; constructing a multi-channel feature map based on the RGB image and rotational speed information, the multi-channel feature map including a rotational speed feature mapping map; performing rotor lamination risk location and stress prediction based on the multi-channel feature map to obtain the risk location result and initial predicted stress value; and correcting the stress value based on the material information and the initial predicted stress value to obtain the corrected final stress value.

[0088] Furthermore, the logical instructions in the aforementioned memory 1130 can be implemented in the form of software functional units. When sold or used as independent products, these software functional units can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0089] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the rotor lamination risk location and stress prediction method provided by the above methods. The method includes: acquiring RGB images, rotational speed information, and material information of the rotor lamination; constructing a multi-channel feature map based on the RGB images and rotational speed information, the multi-channel feature map including a rotational speed feature map; performing rotor lamination risk location and stress prediction based on the multi-channel feature map to obtain the risk location result and initial predicted stress value; and correcting the stress value based on the material information and the initial predicted stress value to obtain the corrected final stress value.

[0090] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0091] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0092] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for locating risk positions and predicting stress in rotor laminations, characterized in that, The method includes: Acquire RGB images, rotational speed information, and material information of the rotor laminations; Based on the RGB image and the rotation speed information, a multi-channel feature map is constructed, which includes a rotation speed feature map; Based on the multi-channel feature map, the risk location of the rotor lamination is located and the stress is predicted, so as to obtain the location result of the risk location and the initial predicted stress value. Based on the material information and the initial predicted stress value, the stress value is corrected to obtain the corrected final stress value.

2. The method for locating the risk position and predicting the stress of rotor laminations according to claim 1, characterized in that, The rotational speed information includes rotational speed values. The construction of a multi-channel feature map based on the RGB image and the rotational speed information includes: The difference between the speed value and the preset minimum speed reference value is determined as a first value, the difference between the preset maximum speed reference value and the minimum speed reference value is a second value, and the ratio between the first value and the second value is a third value. The product of the third value and the preset channel intensity value is determined as the fourth value, and the sum of the fourth value and the preset normalization coefficient is the target pixel value. Based on the target pixel value, the rotational speed feature map is constructed. The rotational speed feature map is superimposed on the three-channel feature map of the RGB image to obtain the multi-channel feature map.

3. The method for locating the risk position and predicting the stress of rotor laminations according to claim 1, characterized in that, The step of locating the risk location and predicting the stress of the rotor lamination based on the multi-channel feature map, and obtaining the location result of the risk location and the initial predicted stress value, includes: The multi-channel feature map is input into a preset target model to locate the risk position of the rotor lamination and predict the stress, so as to obtain the location result and the initial predicted stress value output by the target model, wherein the target model includes: A feature extraction network is used to extract features based on the multi-channel feature map to obtain shallow features, mid-level features, and deep features. The target downsampling factor corresponding to the shallow features, the mid-level features, and the deep features shows an increasing trend. The target downsampling factor represents the downsampling factor relative to the size of the multi-channel feature map. A feature enhancement network is used to enhance the shallow features, the middle features, and the deep features to obtain shallow enhanced features, middle enhanced features, and deep enhanced features. The prediction network is used to locate the risk position and predict the stress of the rotor lamination based on the shallow enhancement features, the middle enhancement features, and the deep enhancement features, and outputs the location result of the risk position and the initial predicted stress value.

4. The method for locating the risk position and predicting the stress of rotor laminations according to claim 3, characterized in that, The feature extraction network includes multiple downsampling layers and multiple buffer layers, with the buffer layers located between adjacent downsampling layers. The multi-layer downsampling layer includes: a first downsampling layer, a second downsampling layer, a third downsampling layer, a fourth downsampling layer, and a fifth downsampling layer arranged in sequence. The input of the first downsampling layer is the multi-channel feature map, the output of the third downsampling layer is the shallow feature, the output of the fourth downsampling layer is the middle feature, and the output of the fifth downsampling layer is the deep feature. The multi-layer buffer layer includes a first buffer layer and a second buffer layer; the first buffer layer is located between the first downsampling layer and the second downsampling layer, and the second buffer layer is located between the third downsampling layer and the fourth downsampling layer.

5. The method for locating the risk position and predicting the stress of rotor laminations according to claim 3, characterized in that, The feature enhancement network includes: The first network branch is used to receive the shallow features, perform feature enhancement on the shallow features, obtain the shallow enhanced features, and output them. The second network branch is used to receive the mid-layer features, perform feature enhancement on the mid-layer features, obtain the mid-layer enhanced features, and output them. The shallow feature weight extraction branch is used to access the shallow features, and based on the shallow features and the preset channel attention mechanism, channel weighting is performed to obtain the target features; The third network branch is used to receive the deep features and the target features, perform feature enhancement based on the deep features and the target features, obtain the deep enhanced features, and output them.

6. The method for locating the risk position and predicting the stress of rotor laminations according to claim 5, characterized in that, The shallow feature weight extraction branch includes: a first two-dimensional convolutional layer, a second two-dimensional convolutional layer, and a channel attention mechanism layer connected in sequence. The input of the first two-dimensional convolutional layer is the shallow feature. The channel attention mechanism layer includes: a global average pooling layer, a modified linear layer, and an activation function layer connected in sequence. The output of the activation function layer is the target feature. The third network branch includes: The first spatial pyramid pooling network includes a first splicing layer and multiple parallel first soft pooling layers. The first soft pooling layers are used to perform soft pooling processing on the deep features to obtain first soft pooled features. The pooling kernels of the multiple first soft pooling layers are different. The first splicing layer is used to splice the deep features and the multiple first soft pooling features to obtain a first feature. The first target convolutional layer includes a first two-dimensional convolutional sub-layer, a first batch normalization sub-layer, and a first activation function sub-layer connected in sequence. The input of the first two-dimensional convolutional sub-layer is the first feature, and the output of the first activation function sub-layer is the second feature. The second feature is the data source for the second network branch to perform feature enhancement. A third two-dimensional convolutional layer is used to receive a third feature, which is the product of the target feature and the second feature; the third feature is convolved to obtain the deep enhancement feature.

7. The method for locating the risk position and predicting the stress of rotor laminations according to claim 6, characterized in that, The second network branch includes: The second spatial pyramid pooling network includes a second splicing layer and multiple parallel second soft pooling layers. The second soft pooling layers are used to perform soft pooling processing on the middle layer features to obtain second soft pooled features. The pooling kernels of the multiple second soft pooling layers are different. The second splicing layer is used to splice the middle layer features and the multiple second soft pooling features to obtain a fourth feature. The third splicing layer is used to splice the fourth and fifth features to obtain the sixth feature, wherein the fifth feature is the feature obtained by performing two-dimensional convolution and upsampling twice the space on the second feature; The second target convolutional layer includes a second two-dimensional convolutional sub-layer, a second batch normalization sub-layer, and a second activation function sub-layer connected in sequence. The input of the second two-dimensional convolutional sub-layer is the sixth feature, and the output of the second activation function sub-layer is the seventh feature. The seventh feature is the data source for feature enhancement of the first network branch. The fourth two-dimensional convolutional layer is used to perform two-dimensional convolution on the seventh feature to obtain the middle-layer enhanced feature.

8. The method for locating the risk position and predicting the stress of rotor laminations according to claim 7, characterized in that, The first network branch includes: The third spatial pyramid pooling network includes a third splicing layer and multiple parallel third soft pooling layers. The third soft pooling layers are used to perform soft pooling on the shallow features to obtain third soft pooled features. The pooling kernels of the multiple third soft pooling layers are different. The third splicing layer is used to splice the shallow features and the multiple third soft pooled features to obtain an eighth feature. The fourth splicing layer is used to splice the eighth and ninth features to obtain the tenth feature, wherein the ninth feature is obtained by performing two-dimensional convolution and double spatial upsampling on the seventh feature; The third target convolutional layer includes a third two-dimensional convolutional sub-layer, a third batch normalization sub-layer, and a third activation function sub-layer connected in sequence. The input of the third two-dimensional convolutional sub-layer is the tenth feature, and the output of the third activation function sub-layer is the eleventh feature. The fifth two-dimensional convolutional layer is used to perform two-dimensional convolution on the eleventh feature to obtain the shallow enhanced feature.

9. The method for locating the risk position and predicting the stress of rotor laminations according to claim 1, characterized in that, The material information includes the Poisson's ratio, elastic modulus, and density of the rotor laminations. The stress value correction based on the material information and the initial predicted stress value to obtain the corrected final stress value includes: The material information and the initial predicted stress value are input into a preset stress correction model to correct the stress value and obtain the final stress value. The stress correction model includes an input layer, a hidden layer, and an output layer connected in sequence. The hidden layer includes multiple parallel neurons for stress value correction.

10. A rotor lamination risk location positioning and stress prediction system, characterized in that, The system includes: The data acquisition module is used to acquire RGB images, rotational speed information, and material information of the rotor laminations; A multi-channel feature map construction module is used to construct a multi-channel feature map based on the RGB image and the rotation speed information, wherein the multi-channel feature map includes a rotation speed feature map; The risk location and stress prediction module is used to locate the risk location and predict the stress of the rotor lamination based on the multi-channel feature map, and obtain the location result of the risk location and the initial predicted stress value. The stress value correction module is used to correct the stress value based on the material information and the initial predicted stress value to obtain the corrected final stress value.

Images