Cabinet management method, device and equipment

Abstract

The invention relates to the technical field of computers, in particular to a cabinet management method, device and equipment. The cabinet management method comprises the following steps: according to multi-mode historical operation data of a cabinet, constructing a multi-mode operation state detection model of the cabinet; based on the multi-modal operation state detection model of the cabinet, determining a multi-modal operation state detection result corresponding to the multi-modal real-time operation data of the cabinet; based on an evidence fusion strategy, generating a Shannon support degree of a multi-mode operation state detection result; and adjusting the operation state of the cabinet according to the Shannon support degree of the multi-mode operation state detection result. According to the embodiment of the invention, the multi-modal operation state detection network can be constructed for the multi-modal data in the cabinet, and is more comprehensive and more accurate. Real-time monitoring of multiple cabinets in an equipment room is realized, and the labor cost is greatly reduced. The evidence fusion strategy is used for fusing the results of the multiple modes, the reasonability is higher, and countermeasures can be taken in time according to the fusion result to guarantee the safety of the cabinet.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of computer technology, and in particular to a cabinet management method, device and equipment. Background Art

[0002] With the continuous improvement of the degree of social informatization, the number of devices in the network cabinets between equipment rooms is increasing day by day, and cabinet monitoring is extremely important. Various abnormal events are prone to occur in network cabinets, such as excessive load on some equipment in the cabinet, shaking and tilting of the cabinet during operation, poor contact / short circuit of the power supply / line inside the cabinet, and excessive temperature in some areas of the cabinet. For the detection and prevention of these abnormal events, frequent inspections by on-duty personnel are required.

[0003] At present, the security management of network cabinets in equipment rooms mainly relies on relevant on-duty personnel to monitor network cabinets in real time through cameras. In addition, some equipment cabinet management methods use models based on deep learning and machine learning to monitor the operating status of network cabinets in equipment rooms in real time.

[0004] The existing methods that rely on manual inspections will generate large maintenance costs on the one hand, and on the other hand, they cannot handle abnormal events in a timely manner. For example, when the load on the cabinet equipment is too large, the corresponding load balancing strategy needs to be used to reduce the load of the equipment. When the line contact is poor / short circuit, the power supply of the corresponding cabinet needs to be turned off and repaired in time. When the temperature in some areas of the cabinet is too high, the ventilation equipment and other possible influencing factors need to be checked in time. Algorithms based on deep learning and machine learning often require data annotation of the historical operating data of the network cabinet between devices to train the model. The cost of data annotation is expensive and it is impossible to fully evaluate the operating status of the network cabinet between devices. Therefore, how to overcome the problems of incomplete evaluation factors, reliance on manual inspections, and expensive data acquisition in the existing methods, and propose a comprehensive, accurate, data-free and intelligent cabinet management method is a key problem that needs to be solved urgently. Summary of the invention

[0005] The purpose of the embodiments of this specification is to provide a cabinet management method, device and equipment to overcome the problems of incomplete evaluation factors, reliance on manual inspections, and expensive data acquisition in existing cabinet management methods, and to manage cabinets between equipment more comprehensively, accurately and intelligently.

[0006] This embodiment of the specification provides a cabinet management method, the cabinet management method comprising:

[0007] According to the multi-modal historical operation data of the cabinet, a multi-modal operation status detection model of the cabinet is constructed;

[0008] Based on the multimodal operation status detection model of the cabinet, determine the multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet;

[0009] Based on the evidence fusion strategy, generating the Shannon support of the multimodal operation status detection result;

[0010] The operating state of the cabinet is adjusted according to the Shannon support of the multi-modal operating state detection result.

[0011] In some embodiments, the cabinet management method further includes:

[0012] The multimodal historical operation data includes historical sound data, historical image data and historical electrical data; the multimodal operation status detection model includes a sound clustering network, an image memory network and an electrical detection network;

[0013] Constructing a sound clustering network based on the historical sound data;

[0014] constructing an image memory network according to the historical image data;

[0015] An electrical detection network is constructed based on the historical electrical data.

[0016] In some embodiments, the cabinet management method further includes:

[0017] constructing a sound reconstruction network according to the historical sound data;

[0018] Based on the sound reconstruction network, generating a low-dimensional sound vector corresponding to the historical sound data;

[0019] A sound clustering network is constructed according to the low-dimensional sound vector.

[0020] In some embodiments, the cabinet management method further includes:

[0021] constructing an image reconstruction network according to the historical image data;

[0022] Based on the image reconstruction network, generating a low-dimensional image vector corresponding to the historical image data;

[0023] An image memory network is constructed according to the low-dimensional image vector.

[0024] In some embodiments, the cabinet management method further includes:

[0025] constructing an electricity prompt text according to the historical electricity data;

[0026] According to the electrical prompt text, an electrical detection network is generated using a pre-trained large language model.

[0027] In some embodiments, the cabinet management method further includes:

[0028] The multimodal real-time operation data includes real-time sound data, real-time image data and real-time electrical data;

[0029] Determine a sound reconstruction loss and a sound clustering loss according to the real-time sound data;

[0030] Determine image reconstruction loss and image memory loss based on real-time image data;

[0031] Determine electrical detection results based on real-time electrical data;

[0032] According to the sound reconstruction loss, sound clustering loss, image reconstruction loss, image memory loss and electrical detection results, a multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet is determined.

[0033] In some embodiments, the cabinet management method further includes:

[0034] Calculate the Shannon entropy difference between the running state detection results of different modes;

[0035] Constructing a Shannon entropy difference matrix according to the Shannon entropy difference between the running state detection results of the different modes;

[0036] According to the Shannon entropy difference matrix, a Shannon support matrix of the multimodal operating state detection result is calculated.

[0037] In some embodiments, the cabinet management method further includes:

[0038] Determining the Shannon confidence interval of the multimodal operating state detection result according to the Shannon entropy difference matrix;

[0039] Calculating the Shannon uncertainty of the multi-modal operating state detection result according to the Shannon confidence interval of the multi-modal operating state detection result;

[0040] According to the Shannon support and the Shannon uncertainty of the multimodal running state detection results, the multimodal running state detection results are fused using the evidence fusion strategy;

[0041] If the fused multi-modal operation status detection result is greater than or equal to a preset detection threshold, determining the cause of the abnormal operation status of the cabinet according to the multi-modal operation status detection result;

[0042] Based on the cause of the abnormal operating status, the operating status of the cabinet is adjusted.

[0043] The embodiment of the present specification also provides a cabinet management device, the device comprising:

[0044] A construction module is used to construct a multi-modal operation status detection model of the cabinet according to the multi-modal historical operation data of the cabinet;

[0045] A determination module, based on the multimodal operation status detection model of the cabinet, determines the multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet;

[0046] A generation module, which generates the Shannon support of the multimodal operation status detection result based on the evidence fusion strategy;

[0047] An adjustment module adjusts the operating state of the cabinet according to the Shannon support of the multi-modal operating state detection result.

[0048] On the other hand, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the above cabinet management method.

[0049] It can be seen from the technical solution of a cabinet management method provided by the above embodiment of this specification that the embodiment of this specification can build a multimodal operation status detection model of the cabinet based on the multimodal historical operation data of the cabinet; based on the multimodal operation status detection model of the cabinet, determine the multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet; based on the evidence fusion strategy, generate the Shannon support of the multimodal operation status detection result; according to the Shannon support of the multimodal operation status detection result, adjust the operation status of the cabinet. Compared with the existing methods, the cabinet management method provided by the embodiment of the specification can build a multimodal operation status detection network for the multimodal data in the cabinet, which is more comprehensive and more accurate. Real-time monitoring of multiple cabinets between equipment is realized, which greatly reduces labor costs. The use of evidence fusion strategy to fuse the results of multiple modes is more reasonable and provides reliable data support for the background to take corresponding countermeasures. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the drawings required for use in the embodiments or the prior art description are briefly introduced below.

[0051] Figure 1 is a flow chart of a cabinet management method provided by an embodiment of this specification;

[0052] Figure 2 It is a logic flow chart of a cabinet management method provided by an embodiment of this specification;

[0053] Figure 3 This is a schematic diagram of the structure of a cabinet management device provided in an embodiment of this specification;

[0054] Figure 4 It is a schematic diagram of the structural composition of a computer device provided in an embodiment of this specification. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of this specification to clearly and completely describe the technical solutions in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0056] See also Figure 1 and Figure 2 . Figure 1 A flowchart of a cabinet management method provided in an embodiment of this specification. Figure 2 The present invention provides a logic flow chart of a cabinet management method according to an embodiment of the present invention. In specific implementation, the cabinet management method includes the following steps:

[0057] S101: Construct a multi-modal operation status detection model of the cabinet according to the multi-modal historical operation data of the cabinet.

[0058] In some embodiments, the multimodal historical operation data includes historical sound data, historical image data, and historical electrical data; the multimodal operation status detection model includes a sound clustering network, an image memory network, and an electrical detection network. Based on the historical sound data, a sound clustering network is constructed; based on the historical image data, an image memory network is constructed; and based on the historical electrical data, an electrical detection network is constructed.

[0059] By constructing a sound clustering network, an image memory network and an electrical detection network based on historical sound data, historical image data and historical electrical data respectively, the normal historical operating status of the cabinet can be more comprehensively characterized, laying the foundation for more comprehensive identification of the abnormal real-time operating status of the cabinet.

[0060] When the equipment in the network cabinet is operating abnormally, it may be manifested as abnormalities in one or more aspects such as sound, heat distribution, current, voltage, power, etc. When the equipment is operating normally, it will make a specific sound. When the equipment is operating abnormally, it may produce abnormal sounds that are inconsistent with the normal working sounds, such as ringing, hoarseness, friction, harshness, noise, etc. These abnormal sounds may indicate problems such as damage to the mechanical parts of the equipment, motor failure, or abnormal fan speed. The equipment in the network cabinet will generate heat during operation. Under normal circumstances, the temperature distribution in the cabinet should be uniform and will not exceed the normal operating temperature range of the equipment. When the equipment is operating abnormally, it may cause uneven heat distribution or local overheating. This may be caused by poor heat dissipation of the equipment, fan failure, or poor air flow inside the cabinet. These thermal distribution anomalies can be captured by infrared image sensors, so as to timely discover and deal with potential thermal problems. When the equipment in the network cabinet is running, excessive load or cable failure may cause abnormal current. If the load of the equipment exceeds the rated current, it will cause excessive current; and problems such as cable aging, damage, or loose wiring may cause current instability or abnormal fluctuations. Abnormal current may damage the equipment or affect the normal operation of the equipment. Therefore, by monitoring the current data of the equipment in the cabinet, the abnormal current can be discovered in time and corresponding measures can be taken to deal with it. Abnormal voltage is also a manifestation of abnormal operation of the equipment in the network cabinet. Too high or too low voltage may damage the equipment or affect its normal operation. For example, too high voltage may cause the internal components of the equipment to break down or burn; while too low voltage may cause the equipment to fail to work properly or performance degradation. By monitoring the voltage data of the equipment in the cabinet, voltage abnormalities can be discovered in time and corresponding protective measures can be taken. Power abnormalities are usually related to the energy consumption and performance of the equipment in the network cabinet. When the equipment is in an abnormal operating state, it may cause abnormal power fluctuations or exceed the normal range. This may be caused by damage to the internal components of the equipment, load changes, or power failures. By monitoring the power data of the equipment in the cabinet, the energy consumption of the equipment can be understood and potential energy consumption problems can be discovered. In summary, when the equipment in the network cabinet is in an abnormal operating state, it can be manifested as abnormalities in one or more aspects of sound, heat distribution, current, voltage, and power. In order to detect and handle these abnormal problems in time, it is necessary to regularly monitor and maintain the equipment in the cabinet, and take corresponding protective measures to ensure the stable operation of the equipment. By building multiple sound sensors and infrared image sensors into the network cabinet and recording the cabinet's historical electrical data, comprehensive monitoring of the cabinet's internal environment can be achieved.

[0061] Sound sensors can capture the sounds generated by the equipment inside the cabinet when it is running. These sound data can be used to analyze the operating status of the cabinet. For example, abnormal mechanical friction sound, motor hum, or fan speed changes may indicate that the equipment inside the cabinet is about to fail. By building a sound clustering network using the historical sound data of the cabinet under normal operating conditions, the sound feature distribution of the cabinet under normal operating conditions can be learned. Based on the learned sound features of the cabinet under normal operating conditions, potential equipment problems can be discovered and predicted in a timely manner, so that preventive measures can be taken to avoid service interruptions caused by equipment failures.

[0062] Infrared image sensors can capture the temperature distribution and hot spots inside the cabinet. Infrared image data is essential for detecting overheated equipment and potential fire risks. By building an image memory network using historical image data of the cabinet under normal operation, the thermal load characteristic distribution of the cabinet under normal operation can be learned. Based on the learned thermal load distribution characteristics of the cabinet under normal operation, the trend and law of temperature changes in the cabinet can be identified, thereby optimizing the cabinet's heat dissipation design and ensuring that the equipment operates within an appropriate temperature range.

[0063] In addition to sound and infrared image data, the historical electrical data of the cabinet under normal operating conditions can be recorded. The historical electrical data includes the voltage, current, and power data of multiple devices in the cabinet. By using the historical electrical data of the cabinet under normal operating conditions to build an electrical detection network, the electrical characteristic distribution of the cabinet under normal operating conditions can be learned, and then based on the learned electrical characteristic distribution of the cabinet under normal operating conditions, the energy consumption of the equipment can be understood, and equipment or time periods with abnormal energy consumption can be found, so as to take energy-saving measures, and it can also be used for equipment health monitoring and fault warning. For example, voltage fluctuations, abnormal current, or sudden power surges may indicate that the equipment is faulty or about to be damaged.

[0064] In summary, by acquiring the historical sound data, historical image data, and historical electrical data of the cabinet in normal operation, and building a sound clustering network, an image memory network, and an electrical detection network based on these data, a multimodal operation status detection model of the cabinet can be obtained. According to the multimodal operation status detection model of the cabinet, comprehensive monitoring and analysis of the internal environment of the cabinet can be achieved.

[0065] In some embodiments, a sound reconstruction network can be constructed based on the historical sound data; based on the sound reconstruction network, a low-dimensional sound vector corresponding to the historical sound data can be generated; and based on the low-dimensional sound vector, a sound clustering network can be constructed.

[0066] By constructing a sound reconstruction network and a sound clustering network, a foundation is laid for comprehensively evaluating the abnormality of the real-time sound data of the cabinet from the two aspects of sound reconstruction loss and sound clustering loss.

[0067] Multiple sound sensors built into the cabinet can be used to obtain historical sound data at multiple historical time points when the cabinet is operating normally. The collected historical sound data can be preprocessed, including denoising, standardization, segmentation into appropriate segments, etc., which will not be described in detail here. These historical sound data can be used to train a sound reconstruction network based on a deep learning algorithm. The sound reconstruction network can be based on a codec structure. The encoder can compress the input historical sound data into a low-dimensional sound vector that can be used to reconstruct the historical sound data. The low-dimensional sound vector contains the main feature information of the historical sound data. The decoder can decode the low-dimensional sound vector and reconstruct it back to the original sound data, which is as close to the real historical sound data as possible. Minimize the difference between the real historical sound data and the reconstructed historical sound data until the loss of the sound reconstruction network converges. Because the data used for training are all normal data, it can be considered that the sound reconstruction network has learned the sound feature distribution of the cabinet under normal operation.

[0068] The sound reconstruction network mainly focuses on the reconstruction of sound waveforms, so it usually has good recognition performance for abnormal sound data with large waveform changes. However, for abnormal sound data with small waveform changes, since these anomalies may be more subtle or difficult to detect, the sound reconstruction network may not be able to effectively capture these differences, resulting in reduced recognition performance. To compensate for this limitation, abnormal sound data can be further detected based on the low-dimensional sound vectors learned by the sound reconstruction network. These low-dimensional sound vectors are obtained after the original sound data is compressed by the encoder. They contain the main feature information of the sound data and reflect the deep semantic information of the sound data to a certain extent.

[0069] Therefore, abnormal sound data can be further detected based on the low-dimensional sound vector learned by the sound reconstruction network. The low-dimensional sound vector learned by the sound reconstruction network is compressed sound data, which contains deep semantic information about whether the sound data is abnormal or not. The low-dimensional sound vector can be clustered to obtain multiple sound cluster centers. For each low-dimensional sound vector, the distance between it and the nearest cluster center is calculated as the clustering loss. This distance reflects the degree of deviation of the low-dimensional sound vector from the distribution center of normal sound data.

[0070] Preferably, as the sound environment in the cabinet changes, the sound data of the cabinet can be continuously collected and used to update the sound reconstruction network. Similarly, as the sound environment of the cabinet changes, the sound clustering center may change, and the sound clustering network can be updated regularly. By regularly updating the sound reconstruction network and the sound clustering network, the accuracy and adaptability of the sound reconstruction network and the sound clustering network can be ensured.

[0071] In some embodiments, an image reconstruction network can be constructed based on the historical image data; based on the image reconstruction network, a low-dimensional image vector corresponding to the historical image data can be generated; and based on the low-dimensional image vector, an image memory network can be constructed.

[0072] Multiple infrared image sensors built into the cabinet can be used to obtain historical image data at multiple historical time points when the cabinet is operating normally. The historical image data reflects the heat distribution of multiple devices in the cabinet under normal conditions. The collected historical image data can be preprocessed, including denoising, standardization, segmentation into appropriate sizes, etc., which will not be repeated here. These historical image data can be used to train an image reconstruction network based on a deep learning algorithm. The image reconstruction network can be based on a codec structure. The encoder can compress the input historical image data into a low-dimensional image vector that can be used to reconstruct the historical image data. The low-dimensional image vector contains the main feature information of the historical image data. The decoder can decode the low-dimensional image vector and reconstruct it back to the original image data, which is as close to the real historical image data as possible. Minimize the difference between the real historical image data and the reconstructed historical image data until the loss of the image reconstruction network converges. Because the data used for training are all normal data, it can be considered that the image reconstruction network has learned the characteristic distribution of thermal load under normal operating conditions of the cabinet.

[0073] The image reconstruction network mainly focuses on the reconstruction of the heat load distribution, so it usually has good recognition performance for abnormal image data with large changes in the heat load distribution. However, for abnormal image data with small changes in the heat load distribution, since these anomalies may be more subtle or difficult to detect, the image reconstruction network may not be able to effectively capture these differences, resulting in a decrease in recognition performance. To compensate for this limitation, the abnormal image data can be further detected based on the low-dimensional image vectors learned by the image reconstruction network. These low-dimensional image vectors are obtained after the original image data is compressed by the encoder. They contain the main heat load distribution feature information of the image data and reflect the deep heat load distribution semantic information of the image data to a certain extent.

[0074] Therefore, abnormal heat load distribution can be further detected based on the low-dimensional image vector learned by the image reconstruction network. The low-dimensional image vector learned by the image reconstruction network is compressed image data, which contains deep semantic information about whether the heat load distribution is abnormal or not. The low-dimensional image vector can be memorized using a deep memory strategy. According to the size of the low-dimensional image vector, the number and size of the memory cells can be determined. Each memory cell in the memory module can be initialized with a random value. These initialization values ​​will be gradually adjusted during the training process. In the process of training the image reconstruction network, for each historical image data, the corresponding low-dimensional image vector is first extracted through the image reconstruction network. Then, a real-valued vector is extracted along a specific dimension of the low-dimensional image vector (which can be all dimensions or a selected part). For each extracted real-valued vector, its similarity with all memory cells in the memory module is calculated. The similarity can be calculated using a metric such as cosine similarity and Euclidean distance. The similarity between each real-valued vector and all memory cells is normalized to obtain a weight update coefficient for each memory cell corresponding to each real-valued vector. According to the weight update coefficient, each memory cell in the memory module is updated. This can be achieved through weighted averaging, weighted summing, or other updating strategies. While training the reconstruction network, the memory module is also trained. By continuously and iteratively extracting real-valued vectors, calculating similarities, and updating memory units, the memory module gradually learns the deep semantic feature distribution of historical image data.

[0075] Preferably, as the thermal load environment in the cabinet changes, the image data of the cabinet can be continuously collected and used to update the image reconstruction network. Similarly, as the thermal load environment in the cabinet changes, the adaptability of the image memory network may decrease, and the image memory network can be updated regularly. By regularly updating the image reconstruction network and the image memory network, the accuracy and adaptability of the image reconstruction network and the image memory network can be ensured.

[0076] In some embodiments, an electrical prompt text may be constructed based on the historical electrical data; and an electrical detection network may be generated using a pre-trained large language model based on the electrical prompt text.

[0077] By using a pre-trained large language model to generate an electrical detection network, the electrical detection network can be accelerated based on the prior knowledge of the large language model to obtain historical electrical data of multiple devices in the cabinet (at least including historical voltage data, historical current data, and historical power data). First, an initial electrical classification network can be generated using a large language model. Construct electrical data prompt text based on historical electrical data. The constructed prompt text also includes expectations for the desired electrical classification network. The constructed electrical data prompt text can be input into the pre-trained large language model, and the large language model can adjust the initial electrical classification network based on the input prompt text. Using the fine-tuning ability of the large language model, adjust the parameters of the electrical detection model to make it more suitable for electrical data detection tasks. Through multiple iterative training, the prompt text and model parameters are continuously adjusted until satisfactory model performance is achieved.

[0078] Specifically, the historical electrical data of multiple historical time points when the cabinet is operating normally can be obtained, and the collected historical electrical data can be preprocessed, including denoising, standardization, segmentation into appropriate segments, etc., which will not be repeated here. Based on the preprocessed historical electrical data, an electrical prompt text can be constructed. The electrical prompt text includes the characteristics and sources of the data, and the expectations and constraints of the electrical detection network to be generated. The electrical prompt text may include information such as data format, feature range, classification target, model performance requirements, etc., which will not be repeated here. The constructed electrical prompt text can be input into a pre-trained large language model, and a request can be made to generate an initial electrical detection network. The electrical detection network can be a description, pseudocode, or actual code snippet of a neural network architecture. Some historical electrical data for verification can be reserved, and the cross-validation method can be used to evaluate the performance of the model to ensure the stability and generalization ability of the model on different data. The fine-tuned model is compared and analyzed with the unfine-tuned model to evaluate the improvement effect of fine-tuning on the model performance. In each iteration, the performance of the generated electrical detection network is evaluated. This can be done by testing the accuracy, robustness, and interpretability of the electrical detection network on a reserved historical electrical data validation dataset. Based on the evaluation results, the constraints or expectations in the electrical prompt text can be adjusted so that the pre-trained large language model can generate a more optimized electrical detection network.

[0079] The above training or optimization process can be completed by optimization algorithms such as gradient descent and stochastic gradient descent to minimize the loss function and improve the performance of the network.

[0080] Preferably, as the operation mode of the equipment in the cabinet changes, the electrical data of the cabinet can be continuously collected and used to update the electrical detection network. By regularly updating the electrical detection network, the accuracy and adaptability of the electrical detection network can be ensured.

[0081] When any of the deep learning / machine learning models in the above-mentioned sound reconstruction network, sound clustering network, image reconstruction network, image memory network, and electrical detection network is trained or optimized, the learning rate update process is as follows:

[0082] During the training process, gradient descent is used to update the deep learning / machine learning model parameters according to the gradient descent algorithm of the Newton-Leibniz formula. The calculation formula is:

[0083]

[0084] Where ωj represents the jth parameter of the deep learning / machine learning model, ωj' is the corresponding updated parameter; λ represents the learning rate of the deep learning / machine learning model, which is used to determine the convergence rate of the deep learning / machine learning model; It represents the partial derivative of the loss function F(ωj) with respect to ωj when the deviation between the predicted result and the actual result of the deep learning / machine learning model is known. The deviation between the predicted result and the actual result can be calculated by the loss function; the loss function is usually defined based on the sum of squared errors between the predicted result and the actual result or the absolute value of the error, logarithmic loss, and other methods. The deep learning / machine learning model adjusts the values ​​of various parameters in the model by the deviation between the predicted result and the actual result, so that the deviation becomes smaller and smaller. The deep learning / machine learning models used in the above-mentioned sound reconstruction network, sound clustering network, image reconstruction network, image memory network, and electrical detection network can specifically adopt CNN, RNN, ResNet, KNN, K-Means, GRU, LSTM and other models.

[0085] In some embodiments, the learning rate can be set to a fixed value, or strategies such as learning rate decay, adaptive learning rate, and polynomial learning rate can be used. The idea of ​​learning rate decay is to gradually reduce the learning rate as training progresses to improve the stability and generalization of model training. Adaptive learning rate is a type of strategy that can automatically adjust the learning rate. It can dynamically adjust the learning rate according to the gradient information of the parameters, thereby improving the training speed and performance of the model. The idea of ​​polynomial decay is to decay the learning rate through a polynomial function, thereby gradually reducing the learning rate during training. It can help the model converge better in the later stages of training and improve the generalization ability of the model.

[0086] S102: Based on the multimodal operation status detection model of the cabinet, determine a multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet.

[0087] In some embodiments, the multimodal real-time operation data includes real-time sound data, real-time image data and real-time electrical data. The multimodal real-time operation data of the cabinet includes real-time sound data, real-time image data and real-time electrical data, and the real-time sound data, real-time image data and real-time electrical data of the cabinet when it is running in real time can be obtained.

[0088] In some embodiments, based on the real-time sound data, a sound reconstruction loss and a sound clustering loss may be determined.

[0089] By determining the sound reconstruction loss and sound clustering loss based on real-time sound data, a data foundation is laid for a more comprehensive assessment of the real-time operating status of the cabinet.

[0090] After the sound reconstruction network is trained, the sound reconstruction network learns the sound feature distribution of the cabinet under normal operation. Therefore, the difference loss between the real historical sound data and the reconstructed historical sound data is close to 0, and the mean and standard deviation of the reconstruction loss between all the real historical sound data and the reconstructed historical sound data can be obtained. In the real-time detection stage, the real-time sound data of the cabinet can be obtained for reconstruction, and then the reconstruction loss between the real real-time sound data and the reconstructed real-time sound data can be obtained. The reconstruction loss is subtracted from the mean and then divided by the standard deviation to obtain a sound reconstruction anomaly score in the interval [0,1]. Since the model is trained with normal data, it can be used to detect abnormal sounds in the cabinet. When the input sound data differs greatly from the reconstructed sound data output by the model, it can be considered that an abnormal sound has occurred. If the score is closer to 1, it means that the difference between the real-time sound data and the training data is greater, and therefore the degree of anomaly is higher. A threshold (for example, 0.8 or 0.9) can be set. When the score exceeds this threshold, it is considered that an abnormal sound has occurred in the cabinet. During the training process, the clustering loss of all low-dimensional sound vectors is calculated, and the mean and standard deviation of these losses are obtained. In the real-time detection stage, the real-time sound data of the cabinet can be input into the reconstruction network to obtain the corresponding low-dimensional sound vector while obtaining the reconstruction loss. The clustering loss from the low-dimensional sound vector to the cluster center is calculated. The clustering loss is subtracted from the mean and then divided by the standard deviation to obtain a sound clustering anomaly score in the interval [0,1].

[0091] In some embodiments, based on the real-time image data, image reconstruction loss and image memory loss may be determined.

[0092] By determining the image reconstruction loss and image memory loss based on real-time image data, a data foundation is laid for a more comprehensive assessment of the real-time operating status of the cabinet.

[0093] After the image reconstruction network is trained, the image reconstruction network learns the characteristic distribution of the heat load under the normal operation of the cabinet. Therefore, the difference loss between the real historical image data and the reconstructed historical image data is close to 0, and the mean and standard deviation of the reconstruction loss between all the real historical image data and the reconstructed historical image data can be obtained. In the real-time detection stage, the real-time image data of the cabinet can be obtained for reconstruction, and then the reconstruction loss between the real real-time image data and the reconstructed real-time image data can be obtained. The reconstruction loss is subtracted from the mean and then divided by the standard deviation to obtain an image reconstruction anomaly score in the interval [0,1]. Since the model is trained with normal data, it can be used to detect abnormal heat load distribution in the cabinet. When the input image data differs greatly from the reconstructed image data output by the model, it can be considered that an abnormal heat load distribution has occurred. If the score is closer to 1, it means that the difference between the real-time image data and the training data is greater, and therefore the degree of anomaly is higher. A threshold (such as 0.8 or 0.9) can be set. When the score exceeds this threshold, it is considered that an abnormal heat load distribution has occurred in the cabinet. During the training process, the memory loss of all low-dimensional image vectors is calculated, and the mean and standard deviation of these losses are obtained. In the real-time detection stage, the low-dimensional image vector corresponding to the real-time image data is reconstructed using the memory module, and the memory loss is calculated in the same way. The memory loss is subtracted from the mean and then divided by the standard deviation to obtain an image memory anomaly score in the interval [0,1].

[0094] In some embodiments, electrical detection results may be determined based on real-time electrical data.

[0095] By determining the electrical detection results based on real-time electrical data, a data foundation is laid for a more comprehensive assessment of the real-time operating status of the cabinet.

[0096] After the electrical detection network is trained, the electrical detection network learns the electrical feature distribution of the cabinet under normal operating conditions. In the real-time detection stage, the real-time electrical data of the cabinet can be obtained and directly input into the electrical detection network. The electrical detection network can output an electrical detection anomaly score in the interval [0,1]. Since the electrical detection network is generated with normal data, it can be used to detect abnormal electrical feature distributions in the cabinet. The closer the score is to 1, the greater the difference between the real-time electrical data and the training data, and therefore the higher the degree of anomaly. A threshold (for example, 0.8 or 0.9) can be set. When the score exceeds this threshold, it is considered that an abnormal electrical feature distribution has appeared in the cabinet.

[0097] In some embodiments, the multimodal operation state detection result corresponding to the multimodal real-time operation data of the cabinet can be determined according to the sound reconstruction loss, sound clustering loss, image reconstruction loss, image memory loss and electrical detection result. According to the sound reconstruction loss, sound clustering loss, image reconstruction loss, image memory loss and electrical detection result, the sound reconstruction abnormality score, the sound clustering abnormality score, the image reconstruction abnormality score, the image memory abnormality score and the electrical detection abnormality score can be determined respectively, and then the multimodal operation state detection result corresponding to the multimodal real-time operation data of the cabinet can be determined.

[0098] S103: Generate the Shannon support of the multimodal running status detection result based on the evidence fusion strategy.

[0099] In some embodiments, the Shannon entropy difference between the operating status detection results of different modes can be calculated; based on the Shannon entropy difference between the operating status detection results of different modes, a Shannon entropy difference matrix can be constructed; based on the Shannon entropy difference matrix, the Shannon support matrix of the multi-modal operating status detection results can be calculated.

[0100] By calculating the Shannon support matrix of the multi-modal operating status detection results, the foundation is laid for fusing the operating status detection results of different modes based on the Shannon support, so that the real-time operating status of the cabinet can be evaluated more comprehensively and reasonably.

[0101] After determining the running status detection results of different modes, the results of each mode can be regarded as one or more evidences. Specifically, the sound reconstruction anomaly score, the sound clustering anomaly score, the image reconstruction anomaly score, the image memory anomaly score and the electrical detection anomaly score can be used as evidence to judge the real-time running status of the cabinet. The Shannon entropy of each piece of evidence can be calculated, that is, the Shannon entropy corresponding to the sound reconstruction anomaly score, the sound clustering anomaly score, the image reconstruction anomaly score, the image memory anomaly score and the electrical detection anomaly score can be calculated respectively. According to the Shannon entropy of each anomaly score (evidence), the Shannon entropy difference between the two anomaly scores (evidence) can be calculated. According to the Shannon entropy difference between the two anomaly scores (evidence), the Shannon entropy difference matrix corresponding to the real-time running status of the cabinet can be constructed. According to the Shannon entropy difference matrix, the information difference and correlation between different evidences can be analyzed. If the Shannon entropy difference between two pieces of evidence is small, it means that they are close in uncertainty and they are similar in information content and structure; if the difference is large, it means that they are different or complementary in information content and structure. This similarity, difference and complementarity can be used to more comprehensively judge the real-time operating status of the cabinet. According to the calculated Shannon entropy difference matrix corresponding to the real-time operating status of the cabinet, the Shannon support matrix corresponding to the real-time operating status of the cabinet can be calculated. The e index value corresponding to each entropy difference of the Shannon entropy difference matrix can be calculated, and the Shannon support matrix corresponding to the real-time operating status of the cabinet can be constructed after the e index value corresponding to each entropy difference is inverted. According to the calculated Shannon support matrix corresponding to the real-time operating status of the cabinet, the Shannon support of each anomaly score (evidence) can be obtained. Specifically, all Shannon support values ​​related to the current anomaly score in the Shannon support matrix can be summed, and all support values ​​in the Shannon support matrix can be summed, and the former divided by the latter value is used as the support of the current anomaly score (evidence). According to the calculated Shannon support corresponding to each anomaly score (evidence).

[0102] S104: adjusting the operating state of the cabinet according to the Shannon support of the multi-modal operating state detection result.

[0103] In some embodiments, based on the Shannon support of the multimodal operating status detection result, the Shannon confidence interval of the multimodal operating status detection result can be determined; based on the Shannon confidence interval of the multimodal operating status detection result, the Shannon uncertainty of the multimodal operating status detection result can be calculated; based on the Shannon support and Shannon uncertainty of the multimodal operating status detection result, the multimodal operating status detection result can be fused using the evidence fusion strategy.

[0104] Comprehensively considering the Shannon support and Shannon uncertainty between the recognition results of different modes can effectively describe the degree of trust between the recognition results of different modes, and the degree of aggregation of the recognition results of different modes can also be reflected, making the most of the recognition results of different modes. By fusing the multi-modal operation status detection results according to the Shannon support and Shannon uncertainty of the multi-modal operation status detection results, the real-time operation status of the cabinet is evaluated more comprehensively and reasonably, and then corresponding measures can be taken based on the fusion results.

[0105] According to the calculated Shannon entropy difference matrix corresponding to the real-time operating status of the cabinet, the Shannon confidence interval of each anomaly score (evidence) can be calculated. The Shannon entropy difference related to each anomaly score in the Shannon entropy difference matrix can be counted, and then the Shannon confidence interval of each anomaly score (evidence) can be obtained. The distance between the Shannon confidence interval of each anomaly score (evidence) and the interval [0,1] can be obtained by using a definite integral, and this distance represents the Shannon certainty of the current evidence. The Shannon uncertainty of each anomaly score (evidence) can be obtained by subtracting the distance between the Shannon confidence interval of each anomaly score (evidence) obtained by using a definite integral and the interval [0,1]. The support is based on the mutuality between evidences and describes whether the change trends between evidences are related, while the uncertainty of evidence is based on the characteristics of the evidence itself and is used to reflect the degree of aggregation of evidence. The higher the uncertainty, the lower the degree of aggregation of evidence. Therefore, the support between evidences and the uncertainty of the evidence itself do not affect each other, and both belong to the information possessed by the evidence itself. Considering only one of them will destroy the integrity of the evidence information, which is not conducive to the effective correction of conflicting evidence. According to the Shannon support and Shannon uncertainty of each anomaly score (evidence), the weight coefficient of each anomaly score (evidence) can be calculated more comprehensively. The weight coefficient of each anomaly score (evidence) can be defined as: weight coefficient w = Shannon support v*(1-Shannon uncertainty t). The weight coefficient of each anomaly score (evidence) is normalized, and all anomaly scores are fused according to the normalized weight coefficient to obtain a fused anomaly score that characterizes the real-time operating status of the cabinet.

[0106] In some embodiments, if the fused multimodal operating status detection result is greater than or equal to a preset detection threshold, the cause of the abnormal operating status of the cabinet can be determined based on the multimodal operating status detection result; based on the cause of the abnormal operating status, the operating status of the cabinet can be adjusted.

[0107] By using the multi-modal operation status detection results to determine the cause of the abnormal operation status of the cabinet, corresponding measures can be taken to adjust the operation status of the cabinet, thereby avoiding the occurrence of some accidents and avoiding the loss of cabinet property resources.

[0108] A detection threshold (such as 0.75 or 0.80) can be set in advance. If the fusion anomaly score is greater than or equal to the preset detection threshold, it means that the current operating state of the cabinet is abnormal. The cause of the abnormal operating state of the cabinet can be further determined based on the sound reconstruction anomaly score, the sound clustering anomaly score, the image reconstruction anomaly score, the image memory anomaly score, and the electrical detection anomaly score, and the operating state of the cabinet can be adjusted in time according to the abnormal cause. For example, the set detection threshold is 0.50, and the fusion anomaly score is 0.81, indicating that the current operating state of the cabinet is abnormal. Further checking the sound reconstruction anomaly score, the sound clustering anomaly score, the image reconstruction anomaly score, the image memory anomaly score, and the electrical detection anomaly score, which are 0.91, 0.73, 0.92, 0.83, and 0.12, respectively, it can be determined that the thermal load and sound of the equipment in the cabinet are abnormal. Real-time image data of cabinet reconstruction can be obtained to locate the area where abnormal events occur. It can be found that abnormal sounds are caused by fan failure and local equipment temperature rises due to untimely equipment heat dissipation. A shutdown order can be issued to stop the cabinet from working to avoid damage to the equipment in the cabinet and repair the fan failure in time.

[0109] In some embodiments, the maintenance priority of the cabinet can be updated according to the fusion anomaly score. If the fusion anomaly score is less than the preset detection threshold, the maintenance priority of the cabinet can be updated according to the fusion anomaly score, and a mapping relationship between the fusion anomaly score and the maintenance priority can be established to update the maintenance priority of the cabinet in real time. For example, a weak detection threshold (such as 0.4) can be set, the number of times the fusion anomaly score of each cabinet is greater than or equal to the weak detection threshold is recorded, and a number threshold is set. If the number of times the fusion anomaly score of the cabinet is greater than or equal to the weak detection threshold is greater than or equal to the number threshold, the cabinet can be overhauled. The specific content of the overhaul may include checking the hardware status, software configuration, network connection, etc. of the cabinet, and performing corresponding repairs or replacements. The purpose of the overhaul is to ensure that the cabinet can operate normally, meet business needs, and reduce potential security risks.

[0110] In some embodiments, the maintenance priority can be determined according to the purpose of each cabinet door opening, that is, the possible impact on the cabinet and the duration of the cabinet door opening. There are many purposes for opening the cabinet, and different weights can be assigned to each purpose. The purpose of opening is divided into maintenance and non-maintenance. The maintenance priority of the cabinet can be lowered after each maintenance. When the cabinet door is opened for maintenance operations, it usually means that the cabinet has a certain degree of failure or requires preventive maintenance. Different weights can be assigned according to the specific content of the maintenance (such as replacing hardware, adjusting configuration, cleaning, etc.). For example, a maintenance operation of replacing key hardware may have a higher weight because it has a greater impact on the stability and reliability of the cabinet. Sometimes the cabinet door may not be opened for maintenance, but for inspection, installation of new equipment, debugging, etc. These operations may have a smaller impact on the cabinet, so a lower weight can be assigned. However, if non-maintenance operations are frequent or last for a long time, it may also have an adverse effect on the stability of the cabinet, so it still needs attention. The duration of opening the cabinet door is also an important consideration. Opening the cabinet door for a long time may cause pollutants such as dust and moisture to enter the cabinet, increasing the risk of line failure. Therefore, the maintenance priority can be adjusted according to the duration of opening the cabinet door. For example, if the duration of a cabinet door being opened exceeds a preset threshold (such as 1 hour), the cabinet maintenance priority can be increased accordingly. By continuously monitoring the operating status and multimodal fusion score of the cabinet, and regularly evaluating the purpose and duration of the cabinet door being opened, the cabinet maintenance priority can be dynamically adjusted. This ensures that the cabinet maintenance needs are responded to in a timely manner while avoiding unnecessary maintenance operations.

[0111] As can be seen from the above, the cabinet management method provided in the embodiment of this specification can construct a multimodal operation status detection model of the cabinet according to the multimodal historical operation data of the cabinet; based on the multimodal operation status detection model of the cabinet, determine the multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet; based on the evidence fusion strategy, generate the Shannon support of the multimodal operation status detection result; according to the Shannon support of the multimodal operation status detection result, adjust the operation status of the cabinet. In this way, a multimodal operation status detection model of the cabinet is constructed according to the multimodal historical operation data of the cabinet to implement the acquisition of the multimodal operation status detection result of the cabinet, without relying on manual inspection and without data annotation for model training, and the real-time operation status of the cabinet can be obtained more comprehensively, accurately and intelligently.

[0112] Based on the above cabinet management method, this specification also proposes an embodiment of a cabinet management device. Figure 3 As shown, the cabinet management device may specifically include the following modules:

[0113] The construction module 301 can be used to construct a multi-modal operation status detection model of the cabinet according to the multi-modal historical operation data of the cabinet;

[0114] The determination module 302 may be used to determine a multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet based on the multimodal operation status detection model of the cabinet;

[0115] A generating module 303 may be used to generate the Shannon support of the multimodal running state detection result based on an evidence fusion strategy;

[0116] The adjustment module 304 may be configured to adjust the operating state of the cabinet according to the Shannon support of the multi-modal operating state detection result.

[0117] In some embodiments, the above-mentioned construction module 301 can be specifically used to construct a sound clustering network based on the historical sound data; construct an image memory network based on the historical image data; and construct an electrical detection network based on the historical electrical data.

[0118] In some embodiments, the above-mentioned construction module 301 can also be specifically used to construct a sound reconstruction network based on the historical sound data; generate a low-dimensional sound vector corresponding to the historical sound data based on the sound reconstruction network; and construct a sound clustering network based on the low-dimensional sound vector.

[0119] In some embodiments, the above-mentioned construction module 301 can also be specifically used to construct an image reconstruction network based on the historical image data; based on the image reconstruction network, generate a low-dimensional image vector corresponding to the historical image data; and construct an image memory network based on the low-dimensional image vector.

[0120] In some embodiments, the above-mentioned construction module 301 can also be specifically used to construct an electricity prompt text according to the historical electricity data; and generate an electricity detection network according to the electricity prompt text using a pre-trained large language model.

[0121] In some embodiments, the above-mentioned determination module 302 can be specifically used to determine the sound reconstruction loss and the sound clustering loss based on the real-time sound data; determine the image reconstruction loss and the image memory loss based on the real-time image data; determine the electrical detection result based on the real-time electrical data; and determine the multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet based on the sound reconstruction loss, the sound clustering loss, the image reconstruction loss, the image memory loss and the electrical detection result.

[0122] In some embodiments, the above-mentioned generation module 303 can be specifically used to calculate the Shannon entropy difference between the operating status detection results of different modes; construct a Shannon entropy difference matrix based on the Shannon entropy difference between the operating status detection results of different modes; and calculate the Shannon support matrix of the multi-modal operating status detection results based on the Shannon entropy difference matrix.

[0123] In some embodiments, the above-mentioned adjustment module 304 can be specifically used to determine the Shannon confidence interval of the multimodal operating status detection result based on the Shannon entropy difference matrix; calculate the Shannon uncertainty of the multimodal operating status detection result based on the Shannon confidence interval of the multimodal operating status detection result; fuse the multimodal operating status detection result using the evidence fusion strategy based on the Shannon support and Shannon uncertainty of the multimodal operating status detection result; if the fused multimodal operating status detection result is greater than or equal to a preset detection threshold, determine the cause of the abnormal operating status of the cabinet based on the multimodal operating status detection result; and adjust the operating status of the cabinet based on the cause of the abnormal operating status.

[0124] It should be noted that the units, devices or modules described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. For the convenience of description, the above devices are described separately by functions divided into various modules. Of course, when implementing this specification, the functions of each module can be implemented in the same or more software and / or hardware, or the modules that implement the same function can be implemented by a combination of multiple sub-modules or sub-units. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0125] As can be seen from the above, based on the cabinet management device provided in the embodiment of this specification, a multimodal operation status detection model of the cabinet can be constructed according to the multimodal historical operation data of the cabinet; based on the multimodal operation status detection model of the cabinet, the multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet is determined; based on the evidence fusion strategy, the Shannon support of the multimodal operation status detection result is generated; according to the Shannon support of the multimodal operation status detection result, the operation status of the cabinet is adjusted. In this way, a multimodal operation status detection model of the cabinet is constructed according to the multimodal historical operation data of the cabinet to implement the acquisition of the multimodal operation status detection result of the cabinet, without relying on manual inspections and without data annotation for model training, and the real-time operation status of the cabinet can be obtained more comprehensively, accurately and intelligently.

[0126] The embodiment of the present specification also provides a computer device for a cabinet management method, including a processor and a memory for storing processor executable instructions. When the processor is implemented, it can perform the following steps according to the instructions: construct a multimodal operation status detection model of the cabinet according to the multimodal historical operation data of the cabinet; based on the multimodal operation status detection model of the cabinet, determine the multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet; based on the evidence fusion strategy, generate the Shannon support of the multimodal operation status detection result; according to the Shannon support of the multimodal operation status detection result, adjust the operation status of the cabinet.

[0127] In order to complete the above instructions more accurately, refer to Figure 4 As shown, the embodiment of this specification also provides another specific computer device, wherein the electronic device includes a network communication port 401, a processor 402 and a memory 403, and the above structures are connected through internal cables so that each structure can perform specific data interaction.

[0128] The processor 402 can be specifically used to build a multimodal operation status detection model of the cabinet based on the multimodal historical operation data of the cabinet; determine the multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet based on the multimodal operation status detection model of the cabinet; generate the Shannon support of the multimodal operation status detection result based on the evidence fusion strategy; and adjust the operation status of the cabinet according to the Shannon support of the multimodal operation status detection result.

[0129] The memory 403 may be specifically used to store corresponding instruction programs.

[0130] In this embodiment, the network communication port 401 can be a virtual port that is bound to different communication protocols so that different data can be sent or received. For example, the network communication port can be a port responsible for web data communication, a port responsible for FTP data communication, or a port responsible for email data communication. In addition, the network communication port can also be a physical communication interface or communication chip. For example, it can be a wireless mobile network communication chip, such as GSM, CDMA, etc.; it can also be a Wifi chip; it can also be a Bluetooth chip.

[0131] In this embodiment, the processor 402 may be implemented in any appropriate manner. For example, the processor may take the form of a microprocessor or processor and a computer-readable medium storing a computer-readable program code (such as software or firmware) executable by the (micro)processor, a logic gate, a switch, an application-specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller, etc. This specification does not limit this.

[0132] In this embodiment, the memory 403 includes a volatile memory and a non-volatile memory. The memory 403 may include multiple levels. In a digital system, anything that can store binary data can be a memory; in an integrated circuit, a circuit with a storage function without a physical form is also called a memory, such as RAM, FIFO, etc.; in a system, a storage device with a physical form is also called a memory, such as a memory stick, TF card, etc.

[0133] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented 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.

[0134] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0135] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0136] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0137] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. 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 in the scope of protection of the present invention.

Claims

1. A cabinet management method, characterized in that: include: According to the multi-modal historical operation data of the cabinet, a multi-modal operation status detection model of the cabinet is constructed; Based on the multimodal operation status detection model of the cabinet, determine the multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet; Based on the evidence fusion strategy, generating the Shannon support of the multimodal operation status detection result; The operating state of the cabinet is adjusted according to the Shannon support of the multi-modal operating state detection result.

2. The method according to claim 1, characterized in that: The multimodal historical operation data includes historical sound data, historical image data and historical electrical data; the multimodal operation status detection model includes a sound clustering network, an image memory network and an electrical detection network; The method of constructing a multi-modal operation status detection model of the cabinet according to the multi-modal historical operation data of the cabinet includes: Constructing a sound clustering network based on the historical sound data; constructing an image memory network according to the historical image data; An electrical detection network is constructed based on the historical electrical data.

3. The method according to claim 2, characterized in that: The step of constructing a sound clustering network according to the historical sound data includes: constructing a sound reconstruction network according to the historical sound data; Based on the sound reconstruction network, generating a low-dimensional sound vector corresponding to the historical sound data; A sound clustering network is constructed according to the low-dimensional sound vector.

4. The method according to claim 2, characterized in that: The step of constructing an image memory network according to the historical image data includes: constructing an image reconstruction network according to the historical image data; Based on the image reconstruction network, generating a low-dimensional image vector corresponding to the historical image data; An image memory network is constructed according to the low-dimensional image vector.

5. The method according to claim 2, characterized in that: The step of constructing an electrical detection network according to the historical electrical data includes: constructing an electricity prompt text according to the historical electricity data; According to the electrical prompt text, an electrical detection network is generated using a pre-trained large language model.

6. The method according to claim 1, characterized in that: The multimodal real-time operation data includes real-time sound data, real-time image data and real-time electrical data; The method of determining a multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet based on the multimodal operation status detection model of the cabinet includes: Determine a sound reconstruction loss and a sound clustering loss according to the real-time sound data; Determine image reconstruction loss and image memory loss based on real-time image data; Determine electrical detection results based on real-time electrical data; According to the sound reconstruction loss, sound clustering loss, image reconstruction loss, image memory loss and electrical detection results, a multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet is determined.

7. The method according to claim 1, characterized in that: The method of generating the Shannon support of the multimodal running state detection result based on the evidence fusion strategy includes: Calculate the Shannon entropy difference between the running state detection results of different modes; Constructing a Shannon entropy difference matrix according to the Shannon entropy difference between the running state detection results of the different modes; According to the Shannon entropy difference matrix, a Shannon support matrix of the multimodal operating state detection result is calculated.

8. The method according to claim 7, characterized in that: The adjusting the operating state of the cabinet according to the Shannon support of the multi-modal operating state detection result includes: Determining the Shannon confidence interval of the multimodal operating state detection result according to the Shannon entropy difference matrix; Calculating the Shannon uncertainty of the multi-modal operating state detection result according to the Shannon confidence interval of the multi-modal operating state detection result; According to the Shannon support and the Shannon uncertainty of the multimodal running state detection results, the multimodal running state detection results are fused using the evidence fusion strategy; If the fused multi-modal operation status detection result is greater than or equal to a preset detection threshold, determining the cause of the abnormal operation status of the cabinet according to the multi-modal operation status detection result; Based on the cause of the abnormal operating status, the operating status of the cabinet is adjusted.

9. A cabinet management device, characterized in that: The device comprises: A construction module, used to construct a multi-modal operation status detection model of the cabinet according to the multi-modal historical operation data of the cabinet; A determination module, used to determine a multimodal operation status detection result corresponding to the multimodal real-time operation data of the cabinet based on the multimodal operation status detection model of the cabinet; A generation module, used for generating the Shannon support of the multimodal running state detection result based on the evidence fusion strategy; An adjustment module is used to adjust the operating state of the cabinet according to the Shannon support of the multi-modal operating state detection result.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 8 is implemented.

Images