Machine room sensor reading identification method and system based on improved YOLOv3

Abstract

The invention discloses a machine room sensor reading identification method and system based on improved YOLOv3, and relates to the technical field of data machine room environment sensor reading identification. The method comprises the following steps: acquiring a sensor reading image in a data room; reading area labeling is carried out on the sensor reading image, and a labeling data set file containing reading area position information is generated; training a pre-constructed improved YOLOv3 model by using the annotation data set file to obtain a reading area identification model; identifying a to-be-identified sensor reading image based on the reading area identification model to obtain a sensor reading POI area subgraph; and performing reading recognition on the sensor reading POI region sub-graph by using an OCR recognition model to obtain a sensor reading value. According to the invention, high-precision and rapid positioning and identification can be realized, and the application requirement of large-scale and automatic inspection by using an intelligent terminal is completely met.

Description

Technical Field

[0001] This invention relates to the field of data center environmental sensor reading recognition technology, specifically to a data center sensor reading recognition method and system based on an improved YOLOv3. Background Technology

[0002] With the rapid advancement of new infrastructure construction, the demand for high-value equipment, centered on information networks and represented by GPU servers, is growing rapidly. This has led to a continuous expansion in the scale and number of data centers, resulting in an increasing number of objects requiring operation and maintenance management, and a constant increase in complexity. As critical infrastructure for enterprises and institutions, data centers require regular inspections. Traditional inspections rely on manual on-site reading and recording of the readings on devices such as air conditioners, thermometers, hygrometers, and electricity meters. This process is time-consuming, labor-intensive, inefficient, and has high labor costs, making it difficult to meet the needs of modern and intelligent operation and maintenance.

[0003] Currently, data center environment sensor reading recognition methods based on image processing technology mainly include image morphology-based methods, convolutional neural network-based methods, and template matching-based methods.

[0004] (1) Image morphology-based method: This method is based on the edge features of the sensor reading image. Through feature engineering, specific structural elements are selected to perform morphological processing on the image, such as dilation, erosion, opening and closing operations, to extract candidate feature points. Then, projection analysis is performed in the horizontal and vertical directions, and the text region is located by combining the number of feature points with a set threshold. Then, the characters are recognized by OCR technology, and the results are verified by using the pattern of the number sequence (such as 5 to 6 consecutive digits) combined with regular expressions to obtain the final reading. However, this method is sensitive to image quality. In natural scenes, when the image is distorted due to factors such as device tilt, glass reflection or reflection, the recognition accuracy drops significantly.

[0005] (2) Convolutional Neural Network-Based Method: This method relies on a large number of labeled sensor reading image samples. After preprocessing such as normalization and data augmentation (e.g., rotation, brightness adjustment), the reading regions and values ​​are manually labeled, and a target detection or end-to-end recognition model is trained based on a deep learning framework. Although the recognition accuracy is high, it requires a large amount of training data, has high labeling costs, and the computational overhead of model training and inference is high, requiring high hardware resources and making it difficult to achieve efficient deployment on mobile terminals such as smartphones.

[0006] (3) Template matching method: This method achieves numerical recognition by matching the similarity between the input image and the standard numerical images in the pre-stored template library. Its main limitations are high computational complexity, unsuitability for large-size image processing, and the ability to only recognize models and styles already in the template library. It cannot complete matching and recognition when faced with new or unrecorded targets, and has poor generalization ability.

[0007] In view of the above, this application is hereby submitted. Summary of the Invention

[0008] The technical problem to be solved by this invention is that traditional methods (such as image morphology and template matching) are sensitive to image quality and have poor generalization ability, while general deep learning models (such as the original YOLOv3) cannot be deployed on mobile devices due to excessive computation, resulting in the inability to achieve high-precision and fast positioning and recognition. The purpose of this invention is to provide a method and system for identifying sensor readings in data centers based on an improved YOLOv3. This method utilizes a small sample (e.g., 378 images) of sensor readings (electricity meters, temperature and humidity sensors, etc.) collected by a mobile terminal device in a natural data center environment. The reading regions of these images are labeled. Then, based on the original YOLOv3 framework, the backbone network is replaced with the computationally complex Darknet-53 and the lightweight and efficient MobileNetV3, resulting in the reading region detection model MobileNetV3-YOLOv3. This model is used to identify newly acquired data center sensor reading images, extracting the reading regions. Finally, OCR technology is used to identify the extracted reading regions and output the final reading, meeting the requirements for high-precision and rapid sensor reading acquisition in large-scale applications.

[0009] This invention is achieved through the following technical solution:

[0010] In a first aspect, the present invention provides a method for identifying sensor readings in a computer room based on an improved YOLOv3, the method comprising:

[0011] Acquire sensor readings within the data center. These sensor readings are RGB format images of the sensors within the data center, captured using a mobile terminal in a natural environment.

[0012] The sensor reading images are labeled with the reading regions to generate a labeled dataset file containing the location information of the reading regions.

[0013] The pre-built improved YOLOv3 model was trained using labeled dataset files to obtain a display region recognition model. The improved YOLOv3 model replaces the backbone network Darknet-53 of the original YOLOv3 model with the lightweight network MobileNetV3, and utilizes its depthwise separable convolution and attention mechanism to reduce model complexity while ensuring feature extraction capability.

[0014] The sensor reading image to be identified is based on the reading region recognition model to obtain the sensor reading POI region sub-image;

[0015] The sensor reading values ​​are obtained by using an OCR recognition model to identify the readings of the POI region sub-image.

[0016] Furthermore, the method also includes:

[0017] Image preprocessing is performed on the sensor reading images; the preprocessing involves normalizing the spatial resolution of the sensor reading images to a preset value.

[0018] Furthermore, the sensor reading images are annotated to represent the reading regions, generating an annotated dataset file containing the location information of the reading regions, including:

[0019] The sensor reading image is bounded by a rectangular label to select a sub-image of the reading region, and the coordinates of the four corners of the sub-image of the reading region are recorded.

[0020] The sub-image data of the display region is stored line by line in txt file format, generating a labeled dataset file containing the coordinates of the location points of the sub-images of the display region;

[0021] Each line is formatted as image_name.jpg ["points": [coordinates of the top left corner of the rectangle], [coordinates of the top right corner of the rectangle], [coordinates of the bottom right corner of the rectangle], [coordinates of the bottom left corner of the rectangle]].

[0022] Further improvements to the YOLOv3 model include:

[0023] The original YOLOv3 multi-scale detection head structure is retained, that is, target prediction is performed at three different scales (19×19, 38×38, 76×76);

[0024] The original YOLOv3 Feature Pyramid Network (FPN) structure is retained to fuse high-level semantic information with low-level spatial details;

[0025] The output format of the multi-scale detector head structure is defined as: bounding box coordinates (x, y, w, h), confidence score, and class probability; where x, y, w, and h are the coordinates of the four points of the bounding box, respectively.

[0026] The original YOLOv3 backbone network Darknet-53 was removed, and the lightweight network MobileNetV3 was introduced as a new feature extraction network.

[0027] Feature maps from multiple stages of the lightweight network MobileNetV3, including the outputs of the third, fourth, and fifth stages, are extracted and used as inputs to the Feature Pyramid Network (FPN) structure to achieve multi-scale feature fusion.

[0028] Define the input image size as uniformly normalized. , where m is the number of stage outputs in the lightweight network MobileNetV3.

[0029] Furthermore, the pre-built improved YOLOv3 model is trained using an annotated dataset file to obtain a display region recognition model, including:

[0030] Read the labeled dataset file and store the labeled dataset into the first list-type data structure object all_samples, with the size of the labeled dataset being all_samples_size;

[0031] The labeled dataset is divided into training, validation, and test sets, with the ratio of training, validation, and test set data volume set to 8:1:1.

[0032] Generate a random number idx with a value ranging from 0 to all_samples_size-1;

[0033] Pick As the training set, it is stored in the second list-type data structure object train_dataset;

[0034] Get all_samples[ + As a validation set, it is stored in the third list-type data structure object valid_dataset;

[0035] Pick As a test set, it is stored in the fourth list-type data structure object test_dataset;

[0036] The training iteration count (epochs) is set to 40, the batch size to 8, the base learning rate (base_lr) to 0.0001, and the loss of the improved YOLOv3 model is set to 0.7.

[0037] Based on the above parameters, the pre-built improved YOLOv3 model is trained to obtain the display region recognition model, and the optimal model parameters obtained from the training are saved to the local file system.

[0038] Furthermore, an OCR recognition model is used to identify the sensor readings in the POI region sub-image to obtain the sensor reading values, including:

[0039] The sensor reading POI region sub-image is converted to grayscale, and an adaptive thresholding method is used to generate a binary image from the grayscale sensor reading POI region sub-image.

[0040] Character-level segmentation of binary images is performed based on connected component analysis to obtain segmented character images;

[0041] Each segmented character image is uniformly scaled to a standard-sized pixel matrix and center-aligned and padded. The HOG gradient histogram of the pixel matrix is ​​then calculated as a feature value.

[0042] The features of each character are input into a pre-trained convolutional neural network single-character classifier, which outputs a single character of the meter reading.

[0043] The single characters identified in sequence are concatenated into an original string, and then post-processed according to the fixed syntax rules of the sensor readings to obtain the sensor reading values.

[0044] Secondly, this invention also provides a data center sensor reading recognition system based on an improved YOLOv3, the system comprising:

[0045] The acquisition unit is used to acquire sensor reading images in the data center. The sensor reading images are RGB format images of the sensors in the data center collected by a mobile terminal in a natural environment.

[0046] The annotation dataset generation unit is used to annotate the display area of ​​the sensor display image and generate an annotation dataset file containing the location information of the display area.

[0047] The model training unit is used to train the pre-built improved YOLOv3 model using labeled dataset files to obtain the display region recognition model. The improved YOLOv3 model replaces the backbone network Darknet-53 of the original YOLOv3 model with the lightweight network MobileNetV3, and utilizes its depthwise separable convolution and attention mechanism to reduce model complexity while ensuring feature extraction capability.

[0048] The display area recognition unit is used to recognize the sensor display image to be recognized based on the display area recognition model, and obtain a sensor display POI region sub-image;

[0049] The reading recognition unit is used to recognize the readings of the sensor reading POI area sub-image using an OCR recognition model to obtain the sensor reading values.

[0050] Furthermore, the system also includes:

[0051] The preprocessing unit is used to perform image preprocessing on the sensor reading image; the preprocessing involves normalizing the spatial resolution of the sensor reading image to a preset value.

[0052] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the above-mentioned method for identifying data center sensor readings based on the improved YOLOv3.

[0053] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for identifying data center sensor readings based on the improved YOLOv3.

[0054] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0055] This invention relates to a method and system for identifying sensor readings in data centers based on an improved YOLOv3 architecture. The method utilizes a small sample (e.g., 378 images) of sensor readings (electricity meters, temperature and humidity sensors, etc.) collected by a mobile terminal device in a natural data center environment. The reading regions of these images are labeled. Then, based on the original YOLOv3 framework, the backbone network is replaced with the computationally complex Darknet-53 with the lightweight and efficient MobileNetV3, resulting in the reading region detection model MobileNetV3-YOLOv3. This model is used to identify newly acquired sensor reading images in the data center, extracting the reading regions. Finally, OCR technology is used to identify the extracted reading regions and output the final reading values. This method meets the requirements for high-precision and rapid sensor reading acquisition in large-scale applications. Attached Figure Description

[0056] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a flowchart of the data center sensor reading recognition method based on the improved YOLOv3 of this invention;

[0058] Figure 2 This is a detailed flowchart of the data center sensor reading recognition method based on the improved YOLOv3 of this invention;

[0059] Figure 3 This is a diagram of the improved YOLOv3 model structure for this invention;

[0060] Figure 4 Images used to verify this invention;

[0061] Figure 5 To verify the image recognition results of this invention;

[0062] Figure 6 This is a graph showing the loss function results during the training process of this invention;

[0063] Figure 7 This is a block diagram of the computer room sensor reading recognition system based on the improved YOLOv3 of this invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0065] This invention addresses the urgent needs of smart terminal applications by proposing a truly lightweight, low-computational-complexity, and small-sample-training-suitable method for rapid identification of sensor readings in data center environments. During the training phase, by replacing the original Darknet-53 backbone network of YOLOv3 with the efficient MobileNetV3 network, the number of model parameters and computational cost (FLOPs) are significantly reduced, substantially decreasing the computational resource overhead and time cost required for training. MobileNetV3 employs depthwise separable convolution and Neural Architecture Search (NAS) techniques, achieving extreme model lightweighting while maintaining feature extraction capabilities. This allows the model to converge quickly even with only 378 small sample data points, effectively solving the problem of deep learning models' dependence on large-scale labeled data. During the inference phase, the improved YOLOv3 model inherits the efficient detection head of YOLOv3 and combines it with the lightweight characteristics of MobileNetV3, enabling real-time, low-power inference operations on mobile terminals such as smartphones. This method overcomes the shortcomings of traditional methods (such as image morphology and template matching) in terms of sensitivity to image quality and poor generalization ability. It also avoids the problem of being unable to deploy on mobile devices due to the excessive computational cost of general-purpose deep learning models (such as the original YOLOv3). From verification images (such as...) Figure 5 The recognition results show that, even in natural scenes, the present invention can still achieve high-precision and rapid positioning and recognition of sensor reading images in computer rooms with complex conditions such as tilt, reflection, and shadows, fully meeting the application requirements of large-scale and automated inspection using smart terminals.

[0066] Example 1

[0067] like Figure 1 and Figure 2As shown, this invention relates to a data center sensor reading recognition method based on an improved YOLOv3, which includes:

[0068] Step 1: Acquire sensor reading images in the data center. The sensor reading images are RGB format images of the sensors in the data center, collected using a mobile terminal in a natural environment.

[0069] In this embodiment, the sensor reading image is in .jpg format, and the pixel attribute values ​​of the original image are read. Where M represents the number of rows and N represents the number of columns.

[0070] Step 2: Perform image preprocessing on the sensor reading image; the preprocessing involves normalizing the spatial resolution of the sensor reading image to a preset value.

[0071] In this embodiment, the spatial resolution of the sensor reading images is uniformly converted to... This generates the corresponding pixel matrix.

[0072] Step 3: Label the reading regions of the preprocessed sensor reading images to generate a labeled dataset file containing the location information of the reading regions;

[0073] In this embodiment, step 3 specifically includes:

[0074] Step 31: Select the sub-image of the reading region of the sensor reading image based on the rectangular label, and record the coordinates of the four corners of the sub-image of the reading region;

[0075] Step 32: Using a txt file format, store the sub-image data of the display area line by line to generate a labeled dataset file containing the coordinates of the location points of the sub-images of the display area;

[0076] Each line is formatted as image_name.jpg ["points": [coordinates of the top left corner of the rectangle], [coordinates of the top right corner of the rectangle], [coordinates of the bottom right corner of the rectangle], [coordinates of the bottom left corner of the rectangle]].

[0077] Step 4: Train the pre-built improved YOLOv3 model using the labeled dataset file to obtain the display region recognition model; The improved YOLOv3 model is to replace the backbone network Darknet-53 of the original YOLOv3 model with the lightweight network MobileNetV3, and use its depthwise separable convolution and attention mechanism to reduce the model complexity while ensuring the feature extraction capability.

[0078] In this embodiment, the improved YOLOv3 model includes:

[0079] a) Retain the original multi-scale detection head structure of YOLOv3, that is, perform target prediction at three different scales (19×19, 38×38, 76×76);

[0080] b) Retain the original YOLOv3 Feature Pyramid Network (FPN) structure to fuse high-level semantic information with low-level spatial details;

[0081] c) Define the output format of the multi-scale detector head structure as: bounding box coordinates (x, y, w, h), confidence score, and class probability; where x, y, w, and h are the coordinates of the four points of the bounding box, respectively.

[0082] d) Remove the original YOLOv3 backbone network Darknet-53 and introduce the lightweight network MobileNetV3 as a new feature extraction network;

[0083] e) Extract feature maps from multiple stages in the lightweight network MobileNetV3, including the output of the third stage (Stage3), the output of the fourth stage (Stage4), and the output of the fifth stage (Stage5), and use them as input to the Feature Pyramid Network (FPN) structure to achieve multi-scale feature fusion;

[0084] f) Define the input image size as uniformly normalized. It adapts to the input requirements of YOLOv3.

[0085] like Figure 3 As shown, Figure 3 To improve the YOLOv3 model structure diagram, the specific explanation is as follows:

[0086] (1) Input image : Used to provide standardized high-resolution image input to the detection network, ensuring that the visual features of fine-grained targets are preserved.

[0087] (2) The backbone network is MobileNetV3.

[0088] Conv 3×3, s=2: This is the initial downsampling layer of the backbone network. It quickly reduces the spatial dimension of the feature map through a convolution operation with a stride of 2, while extracting basic texture and edge information.

[0089] MBConv Block Stage 1–2: Composed of multiple MobileNetV3 inverse residual modules, used to complete early feature abstraction with low computational overhead, gradually compressing redundancy and enhancing semantic expression;

[0090] MBConv Block Stage 3: Generates high-resolution shallow feature maps, preserving accurate spatial relationships and providing geometric support for the subsequent localization of large-scale targets;

[0091] MBConv Block Stage 4: Outputs medium-scale feature maps, achieving a balance between semantic richness and spatial detail, suitable for the recognition of medium-granularity objects;

[0092] MBConv Block Stage 5: Constructs deep semantic feature representations to capture global contextual information, thereby enhancing the ability to distinguish small targets or densely laid-out regions.

[0093] (2) Feature Pyramid FPN

[0094] P1_OUT: 76×76×C: As the bottom output of the feature pyramid, it directly inherits the shallow high-resolution features and is used to maintain the boundary accuracy and structural integrity of large target detection.

[0095] P2_OUT: 38×38×C: By fusing Stage 4 features with upsampled high-level semantics through lateral connections, the classification confidence and localization robustness of medium-scale elements are improved.

[0096] P3_OUT: 19×19×C: Integrates Stage 5 deep semantics and top-down contextual enhancement information to effectively alleviate the problem of missed detection of small targets due to feature sparsity.

[0097] (3) YOLOv3 detection head

[0098] Detection Head 1: Deployed on a high-resolution feature map, it is specifically designed to predict the bounding boxes, categories, and confidence scores of large targets, ensuring complete coverage of the sensor reading area.

[0099] Detection Head 2: It operates on medium-scale feature maps to achieve accurate detection of diverse intermediate-granularity targets, balancing efficiency and accuracy;

[0100] Detection Head 3: Based on low-resolution but high-semantic feature maps, it focuses on the detection of small-sized targets and uses strong semantic priors to compensate for the loss of spatial information.

[0101] (4) Output

[0102] Multi-scale detection results: The outputs of the three detection heads are fused to form a final detection result covering the entire scale range, providing structured input for subsequent display recognition.

[0103] In this embodiment, a pre-built improved YOLOv3 model is trained using a labeled dataset file to obtain a display region recognition model, including:

[0104] (a) Read the labeled dataset file and store the labeled dataset into the first list-type data structure object all_samples, with the size of the labeled dataset being all_samples_size;

[0105] (b) Divide the labeled dataset into training set, validation set and test set, and set the data volume ratio of training set, validation set and test set to 8:1:1;

[0106] (c) Generate a random number idx with a value ranging from 0 to all_samples_size-1;

[0107] (d) Take As the training set, it is stored in the second list-type data structure object train_dataset;

[0108] (e) Get all_samples[ + As a validation set, it is stored in the third list-type data structure object valid_dataset;

[0109] (f) Take As a test set, it is stored in the fourth list-type data structure object test_dataset;

[0110] (g) Set the number of training iterations (epochs) to 40, the batch size to 8, the base learning rate (base_lr) to 0.0001, and the loss of the improved YOLOv3 model to 0.7;

[0111] (h) According to the above parameters, perform the training process on the pre-built improved YOLOv3 model to obtain the display region recognition model, and save the optimal model parameters obtained by training to the local file system.

[0112] In addition, the output of the loss function during training is as follows: Figure 6 As shown, Figure 6 The horizontal axis represents the number of training iterations, and the vertical axis represents the loss value. loss_cls (classification loss): measures the accuracy of the model's judgment of object categories; loss_obj (confidence loss): calculates the difference in confidence between the predicted bounding box and the ground truth bounding box; loss_xy (center coordinate loss): evaluates the accuracy of the center coordinates of the predicted bounding box; loss_wh (width and height loss): measures the error in the width and height of the predicted bounding box.

[0113] Step 5: Based on the reading region recognition model, identify the sensor reading image to be identified to obtain a sensor reading POI region sub-image;

[0114] In this embodiment, step 5 specifically includes:

[0115] Step 51: Scale the resolution of the sensor reading image in the data center to be identified to [size missing]. And input the trained reading region recognition model; the data center sensor reading image to be recognized is as follows: Figure 4 As shown.

[0116] Step 52: Save the identified sensor reading POI region sub-map to the local file system; the target detection results of the reading region are as follows: Figure 5 As shown.

[0117] Step 6: Use the OCR recognition model to recognize the sensor readings in the POI region sub-image to obtain the sensor reading values.

[0118] In this embodiment, step 6 specifically includes:

[0119] Step 61: The sensor reading POI region sub-image is converted to grayscale, and an adaptive thresholding method (e.g., Gaussian weighted local thresholding) is used to generate a binary image from the grayscale sensor reading POI region sub-image, effectively dealing with uneven illumination and background texture interference; then morphological opening operation (kernel=3×3) is applied to remove isolated noise points and burrs;

[0120] Step 62: Perform character-level segmentation on the binary image based on connected component analysis to obtain the segmented character image; including:

[0121] 1) Traverse all foreground connected regions and filter out candidate blocks that meet the graphic characteristics of the sensor display (aspect ratio between 2.5 and 6.0, area between 3000 and 25,000 pixels).

[0122] 2) Sort the candidate blocks according to their horizontal coordinates to form a character sequence from left to right;

[0123] Step 63: Scale each segmented character image to a standard size (28×28 pixels) pixel matrix, perform center alignment and padding, and calculate the HOG gradient histogram of the pixel matrix as feature values.

[0124] Step 64: Input the features of each character into a pre-trained convolutional neural network (CNN) single-character classifier and output the single character of the telephone number; for example, the output category set is set to telephone number characters: {'0','1','2',…,'9','.'};

[0125] Step 65 involves concatenating the sequentially identified single characters into an original string, and then performing post-processing according to the fixed syntax rules for sensor readings to obtain the sensor reading value. The regular expression used in the post-processing is as follows: .

[0126] After step 6 above, the sensor reading POI area is obtained as follows: Figure 5 As shown.

[0127] Example 2

[0128] like Figure 7 As shown, the difference between this embodiment and Embodiment 1 is that this embodiment provides a data center sensor reading recognition system based on improved YOLOv3. This system corresponds one-to-one with the data center sensor reading recognition method based on improved YOLOv3 in Embodiment 1. The system includes:

[0129] The acquisition unit is used to acquire sensor reading images in the data center. The sensor reading images are RGB format images of the sensors in the data center collected by a mobile terminal in a natural environment.

[0130] The annotation dataset generation unit is used to annotate the display area of ​​the sensor display image and generate an annotation dataset file containing the location information of the display area.

[0131] The model training unit is used to train the pre-built improved YOLOv3 model using labeled dataset files to obtain the display region recognition model. The improved YOLOv3 model replaces the backbone network Darknet-53 of the original YOLOv3 model with the lightweight network MobileNetV3, and utilizes its depthwise separable convolution and attention mechanism to reduce model complexity while ensuring feature extraction capability.

[0132] The display area recognition unit is used to recognize the sensor display image to be recognized based on the display area recognition model, and obtain a sensor display POI region sub-image;

[0133] The reading recognition unit is used to recognize the readings of the sensor reading POI area sub-image using an OCR recognition model to obtain the sensor reading values.

[0134] As a further implementation, the system also includes:

[0135] The preprocessing unit is used to perform image preprocessing on the sensor reading image; the preprocessing involves normalizing the spatial resolution of the sensor reading image to a preset value.

[0136] The execution process of each unit can be carried out according to the steps of the data center sensor reading recognition method based on improved YOLOv3 in Example 1, and will not be described in detail in this example.

[0137] Meanwhile, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the above-mentioned method for identifying data center sensor readings based on the improved YOLOv3.

[0138] Meanwhile, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for identifying data center sensor readings based on the improved YOLOv3.

[0139] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0140] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0141] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0142] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0143] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying sensor readings in a computer room based on an improved YOLOv3, characterized in that, The method includes: Acquire sensor reading images, which are RGB format images of sensors collected in a data center under natural conditions; The sensor reading images are labeled with reading regions to generate a labeled dataset file containing the location information of the reading regions; The pre-built improved YOLOv3 model is trained using the labeled dataset file to obtain a display region recognition model; the improved YOLOv3 model is obtained by replacing the skeleton network Darknet-53 of the original YOLOv3 model with the lightweight network MobileNetV3. Based on the aforementioned reading region recognition model, the sensor reading image to be recognized is identified to obtain a sensor reading POI region sub-image; The sensor reading values ​​are obtained by using an OCR recognition model to identify the readings of the POI region sub-image.

2. The method for identifying sensor readings in a computer room based on improved YOLOv3 according to claim 1, characterized in that, The method also includes: The sensor reading image is preprocessed; the preprocessing involves normalizing the spatial resolution of the sensor reading image to a preset value.

3. The method for identifying sensor readings in a computer room based on improved YOLOv3 according to claim 1, characterized in that, The sensor reading image is annotated to represent the reading region, generating an annotated dataset file containing the location information of the reading region, including: The sensor reading image is bounded by a rectangular label to select a sub-image of the reading region, and the coordinates of the four corners of the sub-image of the reading region are recorded. The sub-image data of the display area is stored line by line in txt file format to generate a labeled dataset file containing the coordinates of the position points of the sub-image data of the display area; Each line is formatted as image_name.jpg ["points": [coordinates of the top left corner of the rectangle], [coordinates of the top right corner of the rectangle], [coordinates of the bottom right corner of the rectangle], [coordinates of the bottom left corner of the rectangle]].

4. The method for identifying sensor readings in a computer room based on improved YOLOv3 according to claim 1, characterized in that, The improved YOLOv3 model includes: The original YOLOv3 multi-scale detection head structure is retained, that is, target prediction is performed at three different scales (19×19, 38×38, 76×76); The original YOLOv3 Feature Pyramid Network (FPN) structure is retained to fuse high-level semantic information with low-level spatial details; The output format of the multi-scale detector head structure is defined as: bounding box coordinates (x, y, w, h), confidence score, and class probability; where x, y, w, and h are the coordinates of the four points of the bounding box, respectively. The original YOLOv3 backbone network Darknet-53 was removed, and the lightweight network MobileNetV3 was introduced as a new feature extraction network. Feature maps from multiple stages of the lightweight network MobileNetV3, including the outputs of the third, fourth, and fifth stages, are extracted and used as inputs to the Feature Pyramid Network (FPN) structure to achieve multi-scale feature fusion. Define the input image size as uniformly normalized. , where m is the number of stage outputs in the lightweight network MobileNetV3.

5. The method for identifying sensor readings in a computer room based on improved YOLOv3 according to claim 1, characterized in that, The pre-built improved YOLOv3 model is trained using the labeled dataset file to obtain a display region recognition model, including: Read the labeled dataset file and store the labeled dataset into a first list-type data structure object all_samples, with the size of the labeled dataset being all_samples_size; The labeled dataset is divided into a training set, a validation set, and a test set, and the ratio of the data volume of the training set, validation set, and test set is set to 8:1:

1. Generate a random number idx with a value ranging from 0 to all_samples_size-1; Pick As the training set, it is stored in the second list-type data structure object train_dataset; Get all_samples[ + As a validation set, it is stored in the third list-type data structure object valid_dataset; Pick As a test set, it is stored in the fourth list-type data structure object test_dataset; The training iteration count (epochs) is set to 40, the batch size to 8, the base learning rate (base_lr) to 0.0001, and the loss of the improved YOLOv3 model is set to 0.

7. Based on the above parameters, the pre-built improved YOLOv3 model is trained to obtain the display region recognition model, and the optimal model parameters obtained from the training are saved to the local file system.

6. The method for identifying sensor readings in a computer room based on improved YOLOv3 according to claim 1, characterized in that, The sensor readings are identified using an OCR recognition model in the POI region sub-image, yielding sensor reading values, including: The sensor reading POI region sub-image is converted to grayscale, and an adaptive thresholding method is used to generate a binary image from the grayscale sensor reading POI region sub-image. The binary image is segmented into characters based on connected component analysis to obtain segmented character images; Each segmented character image is uniformly scaled to a standard-sized pixel matrix and center-aligned and padded. The HOG oriented gradient histogram of the pixel matrix is ​​then calculated as a feature value. The features of each character are input into a pre-trained convolutional neural network single-character classifier, which outputs a single character of the meter reading. The single characters identified in sequence are concatenated into an original string, and then post-processed according to the fixed syntax rules of the sensor readings to obtain the sensor reading values.

7. A data center sensor reading recognition system based on improved YOLOv3, characterized in that, The system includes: The acquisition unit is used to acquire sensor reading images, which are RGB format images of sensors collected in a data center under natural conditions. The annotation dataset generation unit is used to annotate the display area of ​​the sensor display image and generate an annotation dataset file containing the location information of the display area. The model training unit is used to train the pre-built improved YOLOv3 model using the labeled dataset file to obtain the display region recognition model; the improved YOLOv3 model is to replace the skeleton network Darknet-53 of the original YOLOv3 model with the lightweight network MobileNetV3. The reading region recognition unit is used to recognize the sensor reading image to be recognized based on the reading region recognition model, and obtain a sensor reading POI region sub-image; The reading recognition unit is used to perform reading recognition on the sensor reading POI area sub-image using an OCR recognition model to obtain the sensor reading value.

8. The data center sensor reading recognition system based on improved YOLOv3 according to claim 7, characterized in that, The system also includes: The preprocessing unit is used to perform image preprocessing on the sensor reading image; the preprocessing is to normalize the spatial resolution of the sensor reading image to a preset value.

9. An electronic 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 computer program, it implements the data center sensor reading recognition method based on the improved YOLOv3 as described in any one of claims 1 to 6.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the data center sensor reading recognition method based on the improved YOLOv3 as described in any one of claims 1 to 6.

Images