An equipment monitoring method and system based on image recognition

Abstract

The present invention discloses a device monitoring method and system based on image recognition, which relates to the technical field of image recognition. It includes obtaining real-time image data and performing preprocessing, integrating multi-level feature vectors into a comprehensive feature vector through feature fusion; inputting the comprehensive feature vector into a classification model constructed based on deep features and environmental dynamic adaptation, with enhanced environmental adaptability and generalization ability, applying the classification model to monitor the device status in real time, calculating the abnormal probability of the device, evaluating the current device status based on the probability of the device being in an abnormal state, and generating a device status monitoring report. By collecting and processing real-time images, the present invention improves the monitoring efficiency. The deep learning technology extracts features, enhancing the recognition ability. After feature fusion, the model sensitivity is improved, and the abnormal detection is more accurate. The adaptive classifier ensures the monitoring accuracy, analyzes the device status in real time, evaluates the health status, and provides data support for maintenance, reducing faults.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and particularly to a device monitoring method and system based on image recognition. Background Art

[0002] Traditional device monitoring methods rely on sensors to directly measure physical parameters of devices, such as vibration, temperature, and pressure. The rise of image recognition technology provides a new perspective for device monitoring. Image recognition technology combined with deep learning algorithms, especially deep convolutional neural networks, can extract features from complex image data to achieve non-contact monitoring of the operating state of devices. This technological innovation not only broadens the scope of monitoring but also improves the accuracy and efficiency of monitoring.

[0003] Most existing device monitoring methods based on image recognition ignore the influence of environmental factors, which results in poor generalization ability and adaptability of the monitoring system. In different working environments, the same device may exhibit different visual features. If not taken into account, it will directly affect the accuracy of the monitoring results. Secondly, traditional monitoring systems often focus on single-level feature extraction and fail to fully utilize multi-level information in images, which limits the system's ability to recognize complex device states. Moreover, existing technologies have limitations in real-time monitoring and anomaly detection. Especially when dealing with dynamically changing environments and device parameters, it is difficult to quickly respond and accurately evaluate the device state, resulting in high monitoring latency and false alarm rates. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a device monitoring method and system based on image recognition to solve the problems of poor environmental adaptability, single feature extraction, and limitations in real-time monitoring.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, an embodiment of the present invention provides a device monitoring method based on image recognition, which includes obtaining real-time image data and performing preprocessing; using a pre-trained deep convolutional neural network to perform deep feature extraction on the preprocessed image; integrating multi-level feature vectors into a comprehensive feature vector through feature fusion; inputting the comprehensive feature vector into a classification model with enhanced environmental adaptability and generalization ability constructed based on deep features and environmental dynamic adjustment; applying the classification model to perform real-time monitoring of the device state, calculating the anomaly probability of the device; and evaluating the current state of the device based on the probability of the device being in an abnormal state, and generating a device state monitoring report.

[0008] As a preferred solution of the device monitoring method based on image recognition according to the present invention, wherein: the acquisition of real-time image data and preprocessing are specifically carried out as follows:

[0009] Use an industrial-grade high-frame-rate camera, connected with a high-performance image acquisition card, to collect real-time image data;

[0010] Adopt the bilinear interpolation algorithm to perform size normalization on the acquired original image data. For the image after size normalization, use RGB to grayscale conversion to convert the color image into a grayscale image, and adopt a median filter to filter the noise of the grayscale image.

[0011] As a preferred solution of the device monitoring method based on image recognition according to the present invention, wherein: the use of a pre-trained deep convolutional neural network to perform deep feature extraction on the preprocessed image is specifically carried out as follows:

[0012] Set the input image I to perform deep feature extraction through the convolutional layer of the pre-trained VGG16 model;

[0013] Through the VGG16 model, perform deep feature extraction in different convolutional layers to obtain a multi-layer feature map set {F1, F2,..., F i}, where F i represents the feature map generated by the i-th convolutional layer.

[0014] The integration of multi-level feature vectors into a comprehensive feature vector through feature fusion is specifically carried out as follows:

[0015] Generate a query vector Q(i, j) and a key vector K(k, l) from the feature map through linear transformation;

[0016] Before generating the query vector Q(i, j) and the key vector K(k, l) from the feature map through linear transformation, it is necessary to perform Glorot uniform initialization on W Q and W K ;

[0017] For each layer of feature map, calculate the attention weight matrix W, and use the attention mechanism to combine the feature information at different positions into a comprehensive feature vector;

[0018] Use the attention mechanism to combine the feature information at different positions into a comprehensive feature vector.

[0019] The input of the comprehensive feature vector into a classification model with enhanced environmental adaptability and generalization ability constructed based on deep features and environmental dynamic adaptation is specifically carried out as follows:

[0020] Define an adaptive dynamic weight adjustment function, G(F *i (i,j) , E, θ G ), using the radial basis function RBF kernel function as the basic building block, input the comprehensive feature vector into the adaptive dynamic weight adjustment function G(F * i (i,j) , E, θ G ), combining the information of the feature vector itself and the influence of environmental changes, and enhancing the adaptability and generalization ability of the model by dynamically adjusting the weights inside the model;

[0021] Use the adaptive dynamic weight adjustment function G to adjust the weights in the classification model.

[0022] The application of the classification model to monitor the device status in real time and calculate the abnormal probability of the current device, the specific steps are as follows:

[0023] By considering the absolute differences in the changes of internal parameters of the device, environmental factor changes, and the dynamic evolution of features, accurately calculate the probability that the device is in an abnormal state;

[0024] Through the fixed percentile method, define an abnormal behavior critical value T. When the device is running, through P ay value, compare the P ay value calculated in real time with the set abnormal behavior critical value T. If P ay >T, it is determined that the device is in an abnormal state. If P ay <T, it is determined that the device is in a normal state;

[0025] As a preferred scheme of the device monitoring method based on image recognition described in the present invention, wherein: based on the probability that the device is in an abnormal state, evaluate the current device status and generate a device status monitoring report, the specific steps are as follows:

[0026] Through the fixed percentile method, define an abnormal behavior critical value T. When the device is running, through P ay value, compare the P ay value calculated in real time with the set abnormal behavior critical value T. If P ay >T, it is determined that the device is in an abnormal state. If P ay <T, it is determined that the device is in a normal state;

[0027] Integrate the device status evaluation results, the abnormal state probability, and the environmental parameter change trend, conduct a comprehensive analysis, and formulate countermeasures based on the analysis results.

[0028] Second aspect, the present invention provides a device monitoring system based on image recognition, including a data acquisition and preprocessing module, a depth feature extraction and fusion module, an environmental dynamic debugging module, a device status anomaly detection module, and a device status evaluation module. The data acquisition and preprocessing module is used to obtain real-time image data and preprocess the images. The depth feature extraction and fusion module is used to extract depth features from the preprocessed images, fuse feature vectors of different levels to form a comprehensive feature vector. The environmental dynamic debugging module is used to automatically adjust model parameters according to different environmental conditions. The device status anomaly detection module is used to use a classification model to monitor the device status in real time and calculate the anomaly probability of the device in the current state. The device status evaluation module is used to evaluate the overall operation status of the current device according to the probability of the device anomaly status and automatically generate a detailed device status monitoring report.

[0029] Third aspect, an embodiment of the present invention provides a computer device, including a memory and a processor. The memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the device monitoring method based on image recognition as described in the first aspect of the present invention is implemented.

[0030] Fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, wherein: when the computer program is executed by the processor, any step of the device monitoring method based on image recognition as described in the first aspect of the present invention is implemented.

[0031] The beneficial effects of the present invention are as follows: By acquiring and preprocessing real-time image data, this method can quickly adapt to the device operating environment and improve the monitoring efficiency. Using a deep convolutional neural network for accurate feature extraction, thereby improving the recognition accuracy and reliability of the device status. Through feature fusion, integrating multi-level information, enhancing the model sensitivity, and improving the accuracy and timeliness of anomaly detection. Inputting the fused feature vector into a classification model with self-adjusting ability to ensure high-accuracy monitoring in different environments. Monitoring the device status in real time and calculating the anomaly probability to achieve continuous monitoring of the device health status and timely discovery of potential problems. Finally, evaluating the device status based on the anomaly status probability and generating a report, providing a scientific basis for device maintenance and reducing the failure risk. Description of the Drawings

[0032] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0033] Figure 1 It is a flowchart of the device monitoring method based on image recognition in Embodiment 1.

[0034] Figure 2 It is a flowchart of the device monitoring system based on image recognition in Embodiment 1. Detailed implementation manners

[0035] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the detailed implementation manners of the present invention with reference to the accompanying drawings of the specification.

[0036] Embodiment 1, referring to Figure 1 , which is the first embodiment of the present invention. This embodiment provides a device monitoring method based on image recognition, including the following steps:

[0037] S1. Obtain real-time image data and perform preprocessing.

[0038] Furthermore, use an industrial-grade high-frame-rate camera (which has excellent light adaptability and high-speed data transmission capabilities) in combination with a high-performance image acquisition card to ensure that the data transmission rate is not less than 60 frames, and collect real-time image data;

[0039] For the obtained original image data, use the bilinear interpolation algorithm to standardize its size to 224×224 pixels to keep the geometric shape of the image unchanged. For the image after size standardization, use RGB to grayscale conversion to convert the color image into a grayscale image to reduce the amount of calculation. For the grayscale image, use a 3×3 median filter for noise filtering;

[0040] It should also be noted that for image size standardization: the target size is set to 640×480 640×480 pixels, and the bilinear interpolation algorithm is used for scaling to keep the geometric shape of the image unchanged. The image after size standardization will undergo color space conversion to further optimize the image quality to meet the input requirements of the deep learning model;

[0041] Grayscale conversion: Use the standard formula of RGB to grayscale to convert the color image into a grayscale image to reduce the amount of calculation. The expression is:

[0042] Y = 0.299R + 0.587G + 0.114B Y = 0.299R + 0.587G + 0.114B;

[0043] Among them, Y is the grayscale value, R is the red in the original image, G is the green in the original image, and B is the blue in the original image;

[0044] The grayscale image will then undergo noise filtering to enhance the clarity and contrast of the image and create better conditions for deep feature extraction;

[0045] The working principle of the median filter is to take a fixed-size neighborhood for each pixel in the image, and then replace the value of the central pixel with the median of all pixel values in that neighborhood. Applying this operation to each pixel point according to the window size can effectively remove salt-and-pepper noise.

[0046] S2. Use a pre-trained deep convolutional neural network to perform deep feature extraction on the preprocessed image.

[0047] Furthermore, it is set that the input image I undergoes deep feature extraction through the convolutional layers of the pre-trained VGG16 model, and the expression is:

[0048] F = Cv(I; W, b);

[0049] Among them, Cv represents the convolution operation, W represents the weight of the convolution kernel, b represents the bias of the convolution kernel, and F represents the feature map;

[0050] Through the VGG16 model for deep feature extraction in different convolutional layers, a multi-layer feature map set {F1, F2,... F i} is obtained, where F i represents the feature map generated by the i-th convolutional layer.

[0051] It should also be noted that the process of deep feature extraction is essentially a process of converting an image into a feature map. In deep learning, the feature map F is usually a three-dimensional array with a shape of H′×W′×D, where H′ and W′ are the height and width of the feature map, and D is the depth or number of channels;

[0052] The VGG16 model extracts deep features layer by layer from the input image through multiple convolutional layers in its structure. The initial layers focus on simple features such as edges, and subsequent layers capture more complex concepts. This process starts from the input of the preprocessed image, goes through convolution, activation, pooling until multi-layer feature maps are formed, and finally converges into an information-rich feature vector, which is suitable for tasks such as classification and detection. The feature fusion strategy enhances the model performance and adapts to specific application requirements.

[0053] S3. Integrate multi-level feature vectors into a comprehensive feature vector through feature fusion.

[0054] Furthermore, through linear transformation, query vectors Q(i, j) and key vectors K(k, l) are generated from the feature map, and the expressions are:

[0055] Q(i,j) = F i (i,j)·W Q ;

[0056] K(k,l) = F i (k,l)·W K ;

[0057] Among them, W Q is the weight matrix for generating the query vector Q(i, j), and W K is the weight matrix for generating the key vector K(k, l), and W Q and W K have dimensions that match the number of channels of the feature map and can be obtained through random initialization or pre-training;

[0058] Before generating the query vector Q(i, j) and the key vector K(k, l) from the feature map through linear transformation, it is necessary to perform Glorot uniform initialization on W Q and W K ;

[0059] For each layer of the feature map, calculate the attention weight matrix W, and use the attention mechanism to combine the feature information at different positions into a comprehensive feature vector. The expression is:

[0060]

[0061] where d k is the dimension of the key vector K, usually determined by the number of channels of the feature map F. exp(Q(i, j)K(k', l')T) is the sum of the unnormalized attention scores calculated for all positions (k', l'), and W(i, j, k, l) represents the attention weight from position (i, j) to position (k, l);

[0062] Use the attention mechanism to combine the feature information at different positions into a comprehensive feature vector. The expression is:

[0063]

[0064] where F * i (i,j) represents the attention-weighted feature value at position (i, j) after the i-th layer of the feature map is processed by the attention mechanism.

[0065] It should also be noted that during the feature fusion process, the model first generates query and key vectors through a specific linear transformation to capture the correlations between positions on the feature map. Then, it calculates the attention weight matrix, which quantifies the important relationships between different positions and realizes fine control of the global features. Finally, based on these weights, the model aggregates the local features at each level to generate a comprehensive feature vector that combines multi-level information, enhancing the model's ability to understand and represent complex patterns. This mechanism is particularly helpful for strengthening key features, suppressing irrelevant background information, and improving the overall recognition accuracy.

[0066] S4. Input the comprehensive feature vector into a classification model that is constructed based on depth features and environmental dynamic adaptation and has enhanced environmental adaptability and generalization ability.

[0067] Furthermore, define an adaptive dynamic weight adjustment function, G(F * i (i,j) , E, θ G ). Using the radial basis function (RBF) kernel function as the basic building block, input the comprehensive feature vector into the adaptive dynamic weight adjustment function G(F * i (i,j) , E, θ G ). By combining the information of the feature vector itself and the influence of environmental changes, enhance the adaptability and generalization ability of the model by dynamically adjusting the weights inside the model. The expression is:

[0068]

[0069] where F * i (i,j) represents the comprehensive feature vector obtained through the attention mechanism, E represents the environmental factor matrix, θ G is the parameter set of G(F * i (i,j) , E, θ G ), w m is the weight coefficient of the comprehensive feature vector F * i (i,j) , β m is the weight coefficient of the environmental factor matrix E, μ m is the reference point of the comprehensive feature vector, V m is the reference point of the environmental factor matrix, γ is used to control the sensitivity of the RBF kernel to the distance between input features, η controls the distance sensitivity between each element in the environmental factor matrix E and the reference point V m , and M is the number of RBF units;

[0070] The parameters in the G function, such as M, μ m , v m , w m , β m , etc., can be set by experience or obtained through training and learning.

[0071] Use the adaptive dynamic weight adjustment function G to adjust the weights in the classification model. The expression is:

[0072]

[0073] where W(i,j) is the original weight of the classification model at (i, j), W n (i,j) is the updated weight.

[0074] It should also be noted that in the G(F i (i,j) , E, θ G ) function, the range of the output weight adjustment factor is between 0 and infinity. When the comprehensive feature vector and the environmental factor matrix highly match the reference point, the weight adjustment factor will be larger, indicating that these features and environmental conditions significantly contribute to the classification performance; conversely, the weight adjustment factor is smaller, meaning that the features or the environment have less impact on the classification result.

[0075] S5. Apply the classification model to monitor the device status in real time and calculate the anomaly probability of the current device.

[0076] Furthermore, by considering the absolute differences in the changes of internal device parameters, environmental factor changes, and the dynamic evolution of features, accurately calculate the probability that the device is in an abnormal state. The expression is:

[0077]

[0078] where, P ay (Y = 1|F * i (i,j) , E) represents the probability that the device is in an abnormal state given the comprehensive feature vector F * i (i,j) and the environmental factor matrix E. σ is the mapping function, and b is the bias term;

[0079] By using the fixed percentile method, define an abnormal behavior threshold T. When the device is running, compare the value of P ay with the set abnormal behavior threshold T. If P ay > T, it is determined that the device is in an abnormal state. If P ay < T, it is determined that the device is in a normal state; ay < T, it is determined that the device is in a normal state;

[0080] It should also be noted that this formula combines three key aspects: weight change, environmental factor change, and feature change. By using the log function, it transforms them into the probability that the device is in an abnormal state, ensuring that the value range is between 0 and 1 for intuitive understanding. By introducing the log function, even when there are large changes in weights or features, it will not cause drastic fluctuations in the probability, enhancing the stability of the model. In addition, this formula considers the multi-dimensional characteristics of device status evaluation, improving the accuracy and reliability of prediction.

[0081] S6. Evaluate the current device status based on the probability of the device being in an abnormal state, and generate a device status monitoring report.

[0082] Furthermore, by using the fixed percentile method, define an abnormal behavior threshold T. When the device is running, through the value of P ay The value of P ay calculated in real time is compared with the set abnormal behavior threshold T. If P ay > T, it is determined that the device is in an abnormal state. If P ay < T, it is determined that the device is in a normal state;

[0083] Integrate the device status evaluation results and the change trend of the environmental parameters of the abnormal state probability, conduct a comprehensive analysis, and formulate countermeasures based on the analysis results.

[0084] It should also be noted that the fixed percentile method is a statistical method used to determine the outlier threshold in a dataset. Specifically, select a specific percentile as the critical point, such as the 95th percentile (which means 95% of the data is below this value), and this value is T. In the device monitoring scenario, if the calculated Pay index exceeds this pre-set T value, it is considered that the current observed value deviates from the normal range, which may indicate an abnormal device state. This method effectively utilizes the data distribution characteristics to identify abnormalities without making assumptions about the specific distribution form of the data, so it has high flexibility and practicality.

[0085] This embodiment also provides a device monitoring system based on image recognition, including: a data acquisition and preprocessing module, a deep feature extraction and fusion module, an environmental dynamic adjustment module, a device status abnormal detection module, and a device status evaluation module. The data acquisition and preprocessing module is used to obtain real-time image data and preprocess the images. The deep feature extraction and fusion module is used to extract deep features from the preprocessed images and fuse the feature vectors of different levels to form a comprehensive feature vector. The environmental dynamic adjustment module is used to automatically adjust the model parameters according to different environmental conditions. The device status abnormal detection module is used to use a classification model to monitor the device status in real time and calculate the abnormal probability of the device in the current state. The device status evaluation module is used to evaluate the overall operating condition of the current device according to the probability of the device abnormal state and automatically generate a detailed device status monitoring report.

[0086] This embodiment also provides a computer device applicable to the situation of the device monitoring method based on image recognition, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the device monitoring method based on image recognition as proposed in the above embodiment.

[0087] The computer device may be a terminal, which includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner, and the wireless manner can be implemented through WIFI, carrier networks, NFC (Near Field Communication), or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, trackball, or touchpad provided on the housing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0088] This embodiment also provides a storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the device monitoring method based on image recognition proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, abbreviated as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, abbreviated as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, abbreviated as EPROM), programmable read-only memory (Programmable Red-Only Memory, abbreviated as PROM), read-only memory (Read-Only Memory, abbreviated as ROM), magnetic memory, flash memory, a magnetic disk, or an optical disc.

[0089] In summary, the present invention ensures the high-quality and high-speed acquisition of real-time image data by using high-performance image acquisition hardware, optimizes the image quality and reduces the computational burden through image preprocessing techniques, extracts deep features of the image with the help of the deep learning model VGG16, enhances the richness and expressiveness of the features, adopts feature fusion and attention mechanism to accurately locate the key information in the image, significantly improves the environmental adaptability and classification accuracy of the model, enables the model to respond to environmental changes in real time through adaptive dynamic weight adjustment, maintains the stability of the monitoring effect, calculates the probability of abnormal states of the computing device based on the changes of features, environment and weights, realizes fault warning, and the system comprehensively analyzes the device state, automatically generates a detailed monitoring report, and guides maintenance decisions. This series of innovative designs greatly improve the accuracy, efficiency and adaptability of device monitoring, provide a strong technical guarantee for the health management and fault prevention of industrial devices, and ensure the double improvement of production safety and efficiency.

[0090] Example 2, referring to Table 1, is the second example of the present invention. To further verify the advancement of the present invention, experimental simulation data of the device monitoring method based on image recognition are given.

[0091] Five different time periods are selected, and the accuracy rate of the traditional method and the accuracy rate of the method of the present invention, as well as the response time of the traditional method and the response time of the method of the present invention are set respectively for each time period.

[0092] First, the obtained original image data are subjected to size normalization processing by the bilinear interpolation algorithm, and then the computational burden is reduced through RGB to grayscale conversion, and the median filter is applied to remove noise.

[0093] Secondly, the pre-trained VGG16 model is used to extract deep features from the preprocessed image, and the feature vectors extracted by different convolutional layers are input into the attention mechanism to generate a comprehensive feature vector.

[0094] Then, the comprehensive feature vector is input into a classification model with an adaptive dynamic weight adjustment function, and the model can adjust its internal weights according to the dynamic changes of the environment and the device state.

[0095] Finally, the classification model is applied to monitor the device state in real time, calculate the abnormal probability of the device, judge the device state according to the critical value T, and generate a monitoring report.

[0096] Specifically, as shown in Table 1:

[0097] Table 1 Experimental record table

[0098]

[0099] Through the data analysis of the above table, it can be clearly seen that the present invention has high accuracy and fast response time. For example, in the morning period under low light conditions, the accuracy of the traditional method is 75%, while the method of the present invention reaches 92%, significantly improving the accuracy of anomaly detection. At the same time, the response time of the present invention is reduced from 180 milliseconds to 120 milliseconds. A faster response time means that potential anomalies can be detected and processed more quickly. In addition, the improvement in the adaptive adjustment efficiency, from 5 times per hour to 7 times per hour, shows that the system can better adapt and optimize its performance under different environmental conditions, ensuring the efficient operation of the monitoring system.

[0100] The device monitoring based on image recognition of the present invention can not only significantly improve the accuracy of anomaly detection, but also greatly shorten the response time of the system, especially the ability to quickly adapt to environmental changes in a short period of time, which reflects the high efficiency and reliability of the present invention in the field of automated production line monitoring.

[0101] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A device monitoring method based on image recognition, characterized in that: include, Acquire real-time image data and perform preprocessing; Use pre-trained deep convolutional neural network to extract deep features from pre-processed images; Integrate multi-level feature vectors into a comprehensive feature vector through feature fusion; The comprehensive feature vector is input into a classification model built based on deep features and dynamic adaptation of the environment, with enhanced environmental adaptability and generalization ability; Apply classification models to monitor equipment status in real time and calculate the probability of equipment abnormality; Based on the probability that the equipment is in an abnormal state, evaluate the current equipment status and generate an equipment status monitoring report; The multi-level feature vectors are integrated into a comprehensive feature vector by feature fusion, and the specific steps are as follows: Generate query vector Q(i, j) and key vector K(k, l) from the feature map through linear transformation; Before starting, we need to transform W to generate the query vector Q(i, j) and key vector K(k, l) from the feature graph through linear transformation. Q and W K Use Glorot uniform initialization; For each layer of feature maps, the attention weight matrix W is calculated, and the feature information at different positions is combined into a comprehensive feature vector using the attention mechanism; The attention mechanism is used to combine feature information at different positions into a comprehensive feature vector; The step of inputting the comprehensive feature vector into a classification model with enhanced environmental adaptability and generalization capability, which is constructed based on deep features and dynamic environmental adaptation, is as follows: Define an adaptive dynamic weight adjustment function, G(F * i （i,j） , E, θ G ), using the radial basis function RBF kernel function as the basic building block, the comprehensive feature vector is input into the adaptive dynamic weight adjustment function G (F * i （i,j） , E, θ G ), combining the information of the feature vector itself and the impact of environmental changes, and dynamically adjusting the weights within the model to enhance the adaptability and generalization ability of the model; Using adaptive dynamic weight adjustment function G to adjust the weights in the classification model; By considering the absolute differences in the changes in internal parameters of the device, changes in environmental factors, and the dynamic evolution of characteristics, the probability of the device being in an abnormal state can be accurately calculated. The expression is: Among them, P ay (Y=1|F * i （i,j） , E) represents a given comprehensive feature vector F * i （i,j） The probability that the device is in an abnormal state under the condition of the environmental factor matrix E, σ is the mapping function, and b is the bias term; By using the fixed percentile method, an abnormal behavior threshold value T is defined. When the device is running, through the value of P ay , the real-time calculated value of P ay is compared with the set abnormal behavior threshold value T. If P ay >T, it is determined that the device is in an abnormal state. If P ay <T, it is determined that the device is in a normal state.

2. The device monitoring method based on image recognition according to claim 1, characterized in that: The specific steps of acquiring real-time image data and preprocessing are as follows: Use industrial-grade high-frame-rate cameras and high-performance image acquisition cards to collect real-time image data; The acquired original image data is size-standardized using a bilinear interpolation algorithm. After the size-standardized image, the color image is converted into a grayscale image using RGB to grayscale conversion, and the grayscale image is noise-filtered using a median filter.

3. The device monitoring method based on image recognition according to claim 2, characterized in that: The pre-trained deep convolutional neural network is used to extract deep features from the pre-processed image. The specific steps are as follows: Set the input image I to pass through the convolution layer of the pre-trained VGG16 model for deep feature extraction; The VGG16 model is used to extract deep features in different convolutional layers to obtain a multi-layer feature map set {F1, F2, ... F i }, where F i Represents the feature map produced by the i-th convolutional layer.

4. A device monitoring system based on image recognition, based on the device monitoring method based on image recognition according to any one of claims 1 to 3, characterized in that: Including data acquisition and preprocessing module, deep feature extraction and fusion module, environment dynamic debugging module, equipment status anomaly detection module, equipment status assessment module; The data acquisition and preprocessing module is used to acquire real-time image data and preprocess the image; The deep feature extraction and fusion module is used to extract deep features from the preprocessed image and fuse feature vectors at different levels to form a comprehensive feature vector; The environment dynamic debugging module is used to automatically adjust model parameters according to different environmental conditions; The device state anomaly detection module is used to use a classification model to monitor the device state in real time and calculate the abnormal probability of the device in the current state; The equipment status evaluation module is used to evaluate the overall operating status of the current equipment according to the probability of abnormal equipment status, and automatically generate a detailed equipment status monitoring report; The multi-level feature vectors are integrated into a comprehensive feature vector by feature fusion, and the specific steps are as follows: Generate query vector Q(i, j) and key vector K(k, l) from the feature map through linear transformation; Before starting, we need to transform W to generate the query vector Q(i, j) and key vector K(k, l) from the feature graph through linear transformation. Q and W K Use Glorot uniform initialization; For each layer of feature maps, the attention weight matrix W is calculated, and the feature information at different positions is combined into a comprehensive feature vector using the attention mechanism; The attention mechanism is used to combine feature information at different positions into a comprehensive feature vector; The step of inputting the comprehensive feature vector into a classification model with enhanced environmental adaptability and generalization capability, which is constructed based on deep features and dynamic environmental adaptation, is as follows: Define an adaptive dynamic weight adjustment function, G(F * i （i,j） , E, θ G ), using the radial basis function RBF kernel function as the basic building block, the comprehensive feature vector is input into the adaptive dynamic weight adjustment function G (F * i （i,j） , E, θ G ), combining the information of the feature vector itself and the impact of environmental changes, and dynamically adjusting the weights within the model to enhance the adaptability and generalization ability of the model; Using adaptive dynamic weight adjustment function G to adjust the weights in the classification model; By considering the absolute differences in the changes in internal parameters of the device, changes in environmental factors, and the dynamic evolution of characteristics, the probability of the device being in an abnormal state can be accurately calculated. The expression is: Among them, P ay (Y=1|F * i （i,j） , E) represents a given comprehensive feature vector F * i （i,j） The probability that the device is in an abnormal state under the condition of the environmental factor matrix E, σ is the mapping function, and b is the bias term; By using the fixed percentile method, an abnormal behavior threshold value T is defined. When the device is running, through the value of P ay , the value of P ay calculated in real time is compared with the set abnormal behavior threshold value T. If P ay >T, it is determined that the device is in an abnormal state. If P ay <T, it is determined that the device is in a normal state.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the device monitoring method based on image recognition described in any one of claims 1 to 3 are implemented.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the device monitoring method based on image recognition described in any one of claims 1 to 3 are implemented.

Images