IDC machine room monitoring data-oriented deep learning model compression and deployment method

Abstract

The application discloses an IDC machine room monitoring data-oriented deep learning model compression and deployment method, comprising the following steps: S1: acquiring heterogeneous monitoring data collected by multi-modal sensors in the IDC machine room; S2: constructing an anomaly detection deep learning model comprising an image feature extraction branch and a time sequence feature extraction branch; S3: calculating a redundancy tolerance coefficient for hierarchical redundancy tolerance evaluation of the anomaly detection deep learning model and grading of each network layer; S4: performing differential compression on high-redundancy tolerance levels and medium-redundancy tolerance levels based on the hierarchical redundancy tolerance coefficient to generate a compressed model; and S5: deploying the compressed model to an edge node directly connected with an infrared thermal imaging array and an ultrasonic airflow velocity meter, and determining whether the IDC machine room monitoring data is in a stable state or a potential change state. The application automatically suppresses the parameters and computation amount of low information increment sensor channels through the compressed model, and solves the problem of waste of existing model parameter quantity.

Description

Technical Field

[0001] This invention relates to the field of data processing technology for data center monitoring, specifically to a method for compressing and deploying deep learning models for IDC data center monitoring data. Background Technology

[0002] With the deepening of digital transformation, Internet Data Centers (IDCs) are carrying an increasingly massive volume of critical business and data flows. The stability of the data center environment is directly related to the secure operation of social infrastructure such as financial transactions, communication networks, and online services. To ensure the security of the data center environment, modern IDC data centers generally deploy a large number of environmental monitoring and visual surveillance devices, including temperature and humidity sensors, smoke detectors, infrared thermal imagers, and visible light cameras, to monitor the data center microclimate, equipment operating status, and personnel activities around the clock.

[0003] Meanwhile, deep learning technologies, represented by Convolutional Neural Networks (CNNs) and Transformers, have demonstrated superior accuracy and generalization capabilities compared to traditional rule-matching methods in tasks such as environmental anomaly detection, equipment fault prediction, and video behavior analysis, providing technical support for the intelligent upgrade of IDC data center monitoring systems. However, most IDC data center monitoring data processing processes based on deep learning models suffer from the following problems:

[0004] 1. Existing deep learning models typically align all sensor data using the greatest common multiple and feed them into a unified neural network for feature extraction and anomaly detection using a full-channel input approach. However, data from different sensors within a data center differs fundamentally in temporal resolution, spatial coverage granularity, and physical meaning. For example, an infrared thermal imaging array outputs a 128×96 pixel temperature distribution map at a rate of 3-5 frames per second, reflecting the spatial temperature field behind the server rack; while an ultrasonic airflow meter outputs a single scalar wind speed value at a frequency of 20Hz, reflecting the airflow velocity in the cold aisle cross-section. While existing deep learning models simplify model design, they are prone to structural redundancy: the model must include a large number of parameters to process sensor channels that maintain a fixed sampling rate but have extremely low information increments. These channels contribute almost nothing to anomaly detection during most normal operating periods, leading to significant structural and computational redundancy in the model—that is, the model is forced to carry a large number of parameters that process useless or low-value information.

[0005] 2. A single compression strategy cannot be adapted to multi-level deep learning models: A single compression ratio can easily lead to insufficient compression of image branches, wasting edge computing power; at the same time, it can easily cause serious loss of temporal branch information, resulting in the failure to detect small abnormal signals. Summary of the Invention

[0006] To address this issue, the present invention provides a method for compressing and deploying deep learning models for monitoring data in IDC data centers, in order to solve the problems in the prior art where deep learning models tend to have a large amount of structural and computational redundancy, and where a single compression strategy cannot adapt to multi-level deep learning models.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for compressing and deploying deep learning models for IDC (Internet Data Center) monitoring data includes the following steps:

[0009] S1: Acquire heterogeneous monitoring data collected by multimodal sensors in the IDC data center, and perform data time synchronization and alignment to obtain the training dataset;

[0010] S2: Construct an anomaly detection deep learning model that includes image feature extraction branch and temporal feature extraction branch, and pre-train the anomaly detection deep learning model using the training dataset until it converges;

[0011] S3: Calculate the redundancy tolerance coefficient based on the structural complexity, effective factor, and information change rate of the network layers of the anomaly detection deep learning model. This coefficient is used to evaluate the hierarchical redundancy tolerance of the anomaly detection deep learning model and to divide each network layer into high redundancy tolerance level, medium redundancy tolerance level, and low redundancy tolerance level according to the value of the redundancy tolerance coefficient.

[0012] S4: Based on the hierarchical redundancy tolerance coefficient, perform differential compression on the high redundancy tolerance layer and the medium redundancy tolerance layer in the pre-trained anomaly detection deep learning model to generate a compressed model.

[0013] S5: Deploy the compressed model to the edge node directly connected to the infrared thermal imaging array and ultrasonic airflow velocity meter, acquire the current frame temperature distribution image and single scalar wind speed value in real time, calculate the cosine similarity between the temperature distribution image and the previous frame and the absolute difference between the wind speed value and the previous moment. If both indicators are within the threshold, it is determined that the IDC data center monitoring data is in a stable state, and only the calculation corresponding to the non-sparsed channel is performed; otherwise, all network layers are fully executed.

[0014] Furthermore, the heterogeneous monitoring data includes first-type sensor data and second-type sensor data with high dimensionality and low temporal resolution; the first-type sensor data is a temperature distribution image sequence output by an infrared thermal imaging array; the second-type sensor data is a single scalar wind speed time series value output by an ultrasonic airflow velocity meter.

[0015] Furthermore, the specific steps of S2 are as follows:

[0016] S2.1: The image feature extraction branch uses a deep convolutional neural network algorithm as the backbone network, and performs convolution processing on the input temperature distribution image sequence layer by layer through multiple stacked convolutional blocks;

[0017] S2.2: The temporal feature extraction branch uses a two-layer bidirectional long short-term memory network as the backbone network to perform forward and backward convolution processing on the single scalar wind speed time series in the input second type of sensor data along the time dimension;

[0018] S2.3: The spatial temperature field feature map output by the image feature extraction branch and the temporal pattern feature vector output by the temporal feature extraction branch are received by the cross-modal attention fusion algorithm and concatenated to obtain the fused feature vector;

[0019] S2.4: Input the fused feature vector into the anomaly detection head network, and pre-train the anomaly detection deep learning model using the labeled training dataset until it converges.

[0020] Furthermore, the redundancy tolerance coefficient A larger value indicates a stronger tolerance for compression operations in the current layer; the calculation formula is as follows:

[0021]

[0022] in, The structural complexity is the ratio of the number of trainable parameters in the current layer to the total number of parameters in the model. The effective factor represents the ratio of the receptive field range of the input data at any position in the current layer's output feature map to the total size of the current layer's input feature map; The rate of information change is represented by the complement of the average cosine similarity between channels of the current layer's output feature map over N consecutive normal state time steps. , and All are preset weights.

[0023] Furthermore, the specific division of the high redundancy tolerance level, medium redundancy tolerance level, and low redundancy tolerance level is as follows:

[0024] If the redundancy tolerance coefficient ,but The layers are divided into high redundancy tolerance layers. In the typical structure of the IDC data center monitoring model, the high redundancy tolerance layers correspond to all deep convolutional layers in the third and fourth residual stages of the image feature extraction branch.

[0025] If low redundancy threshold ,but The layers are divided into medium redundancy tolerance layers, which correspond to the shallow convolutional layers in the first and second residual stages in the image feature extraction branch and the linear mapping layer in the cross-modal attention fusion algorithm.

[0026] like ,but The layers are divided into low redundancy tolerance layers, which correspond to all weight matrix layers of the two-layer bidirectional long short-term memory network in the temporal feature extraction branch and the first fully connected layer in the anomaly detection head network that directly receives the fused feature vector.

[0027] Furthermore, the specific working principle of S4 is as follows:

[0028] S4.1: For the high redundancy tolerance level divided in S3, use the redundancy tolerance coefficient calculated in S3 to screen the importance of the channel;

[0029] S4.2: For the moderate redundancy tolerance level partitioned in S3, the connections with absolute weight values ​​below a preset threshold are set to zero to form a sparse connection matrix.

[0030] S4.3: The low redundancy tolerance level of S3 is fully preserved, without any sparsification or zeroing operations.

[0031] S4.4: The high redundancy tolerance level after channel sparsification in S4.1, the medium redundancy tolerance level after weight sparsification in S4.2, and the low redundancy tolerance level that is retained in its original form in S4.3 are combined to obtain the compressed model.

[0032] Furthermore, the specific content of S4.1 is as follows:

[0033] For each convolutional layer within a high redundancy tolerance layer, the sum of the absolute values ​​of all elements in the matrix of activation values ​​for each output channel is calculated on the temperature distribution image sequence, serving as the channel importance score. The channel importance scores are then sorted in descending order, and the proportion of channels retained is determined by combining this with the redundancy tolerance coefficient. A higher redundancy tolerance coefficient results in a lower retention rate. For channels that do not enter the retention range after sorting, all corresponding convolutional kernel weights are set to zero. The calculation formula is as follows:

[0034]

[0035] in, , These represent the height and width of the output feature map of the convolutional layer in the spatial dimension, respectively. For the residual stage of the image feature extraction branch in S2.1, these values ​​decrease as the layer deepens. For the first When a frame is input, the layer No. The spatial location of each channel Activation value at that location, This represents the number of frames in the temperature distribution image sequence used for evaluation.

[0036] Furthermore, the specific content of S5 is as follows:

[0037] First, the compressed model generated by S4 is deployed to a node directly connected to the infrared thermal imaging array and the ultrasonic airflow velocity meter to receive data from the first type of sensor and the second type of sensor.

[0038] Then, acquire a temperature distribution image from the first type of sensor and a single scalar wind speed value from the second type of sensor at the current moment, and calculate the cosine similarity between the current temperature distribution image and the previous temperature distribution image, as well as the absolute difference between the current wind speed value and the wind speed value at the previous moment.

[0039] If the cosine similarity is higher than the preset image stability threshold and the absolute difference is lower than the preset time-series stability threshold, the current IDC data center monitoring data is determined to be in a stable state; otherwise, it is determined to be in a potentially changing state.

[0040] This invention has the following advantages: Through the analysis of the structural complexity, effective factor and information change rate of each network layer in S3, and the channel sparsification of the high redundancy tolerance layer with low information change rate in the image feature extraction branch in S4, and the connection sparsification of the medium redundancy tolerance layer, this invention fully preserves the precision-sensitive low redundancy tolerance layer in the temporal feature extraction branch and the anomaly detection head network, so that the compressed model can automatically suppress the parameters and computational load of the low information increment sensor channel, and solve the problem of parameter waste caused by the existing model uniformly encoding the temperature distribution image sequence and the single scalar wind speed time series value;

[0041] Meanwhile, the compression model acquires the current frame's temperature distribution image and a single scalar wind speed value in real time, calculates the inter-frame cosine similarity and the absolute difference in wind speed to determine whether the IDC data center monitoring data is in a stable state or a potentially changing state. In the stable state of the IDC data center monitoring data, it skips the already sparsed channels and the convolutional blocks with the highest redundancy tolerance coefficient to avoid image branches consuming edge computing power during low information periods. In the potentially changing state, all network layers participate fully in inference, and low redundancy tolerance layers run with full-precision weights to ensure the ability of the temporal pattern feature vector to capture small-amplitude shifts in wind speed.

[0042] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description

[0043] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0044] Figure 1 This is a flowchart illustrating the implementation of the deep learning model compression and deployment method for IDC data center monitoring data according to the present invention. Detailed Implementation

[0045] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Technical engineers in the field can make some non-essential improvements and adjustments to the present invention based on the above-described content. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Please see Figure 1 A method for compressing and deploying deep learning models for IDC (Internet Data Center) monitoring data includes the following steps:

[0047] S1: Acquire heterogeneous monitoring data collected by multimodal sensors within the IDC data center, and perform time synchronization and alignment to obtain the training dataset. The heterogeneous monitoring data includes first-type sensor data with high dimensionality and low temporal resolution, and second-type sensor data with low dimensionality and high temporal resolution. The first-type sensor data consists of temperature distribution image sequences output by an infrared thermal imaging array with a frame rate of 3-5 frames / second and a resolution of 128×96 pixels. The second-type sensor data consists of single scalar wind speed time-series values ​​output by an ultrasonic airflow velocity meter at a frequency of 20Hz.

[0048] The multimodal sensor includes an infrared thermal imaging array and an ultrasonic airflow velocity meter. The infrared thermal imaging array uses an uncooled infrared focal plane array detector. It focuses the infrared radiation energy emitted from the back surface of the cabinet through an optical lens. Each pixel of the detector converts the received infrared radiation into an electrical signal. After being processed by the built-in readout circuit and analog-to-digital conversion, it forms a digital grayscale value. Then, using the radiometric calibration relationship, the grayscale value is mapped to an absolute temperature value. Finally, it continuously outputs a temperature distribution image sequence with a resolution of 128×96 pixels at a rate of 3-5 frames / second. Each pixel value in each frame represents the surface temperature at the corresponding spatial location.

[0049] Ultrasonic airflow velocity meters measure wind speed based on the time difference between the propagation of ultrasonic waves in a flowing gas along and against the current. The transducer alternately transmits and receives ultrasonic pulses along a fixed path in the cold aisle section, measuring the time of transit along the airflow direction and the time of transit against the current direction. Instantaneous airflow velocity is calculated using the geometric relationship between the difference in transit times and the known propagation path length, and a single scalar wind speed time series value is synchronously output at a frequency of 20Hz. This single scalar wind speed time series value, as low-dimensional, high-temporal-resolution type II sensor data, directly reflects the airflow velocity changes in the cold aisle cross-section. The calculation formula is as follows:

[0050]

[0051] in, This represents the propagation path length of the ultrasonic wave between the two transducers. The angle between the ultrasonic wave propagation path and the pipe axis. The time difference between the time spent crossing upstream and the time spent crossing downstream. , This is the reverse transit time of the ultrasonic wave propagating against the airflow direction. This is the downstream transit time of the ultrasonic wave as it propagates along the airflow direction.

[0052] Data time synchronization alignment uses the timestamp of each frame of the temperature distribution image sequence in the first type of sensor data as the alignment benchmark. The single scalar wind speed time series value of the second type of sensor data is matched to the time interval of two adjacent temperature distribution images according to the time nearest neighbor principle. The arithmetic mean of multiple wind speed values ​​in the interval is taken to obtain a single wind speed scalar that is time-aligned with the temperature distribution image of the current frame. After traversing all temperature distribution image sequences frame by frame, time-synchronized temperature distribution images and single scalar wind speed value sample pairs are formed. All sample pairs together constitute the training dataset.

[0053] S2: Construct an anomaly detection deep learning model that includes an image feature extraction branch and a temporal feature extraction branch. The image feature extraction branch receives data from the first type of sensor and extracts spatial feature maps, while the temporal feature extraction branch receives data from the second type of sensor and extracts temporal feature vectors. The spatial feature maps and temporal feature vectors are fused using a cross-modal attention fusion algorithm and then input into the anomaly detection head network. The anomaly detection deep learning model is pre-trained using the training dataset until it converges.

[0054] The specific steps of S2 are as follows:

[0055] S2.1: The image feature extraction branch uses a deep convolutional neural network (CNN) algorithm as the backbone network. It performs convolution processing on the input temperature distribution image sequence layer by layer through multiple stacked convolutional blocks. The shallow convolutional blocks extract low-level spatial features such as edges, textures and local temperature gradients in the temperature distribution image. The deep convolutional blocks gradually aggregate into a spatial temperature field feature map with a higher semantic level. The spatial temperature field feature map represents the temperature distribution pattern and spatial correlation of each spatial area on the back of the cabinet.

[0056] The aforementioned deep convolutional neural network algorithm generally refers to a deep residual convolutional neural network with 50 layers of trainable weights that performs classification or representation learning pre-training on a large general image dataset. In this invention, it is mainly used as the backbone network of the image feature extraction branch to receive the temperature distribution image sequence output by the infrared thermal imaging array and extract the spatial temperature field feature map on the back of the cabinet. Its working principle is as follows: The network consists of an input convolutional layer and four stacked residual stages. Each residual stage includes multiple residual blocks. Within each residual block, the input features are directly superimposed onto the output of the convolutional layer through skip connections. This allows the original pixel temperature information of the temperature distribution image sequence to be directly transmitted to the deep layer across the intermediate layers during forward propagation. The convolutional kernels of the shallow residual blocks capture low-level spatial temperature features such as temperature gradient edges and local hotspot contours. The convolutional kernels of the deep residual blocks learn the small deviation patterns of the temperature distribution relative to the desired steady-state temperature based on the identity mapping provided by the skip connections. Finally, the features of all spatial locations are aggregated by the global average pooling layer and output as a fixed-dimensional spatial temperature field feature map. This spatial temperature field feature map has extremely high cosine similarity between adjacent time steps under normal operating conditions, reflecting the physical characteristics of slow temperature field changes at the output frame rate of the infrared thermal imaging array.

[0057] S2.2: The temporal feature extraction branch uses a two-layer bidirectional long short-term memory network as its backbone to perform forward and backward convolution processing along the time dimension on the single scalar wind speed time series in the input second type of sensor data. Its hidden layer dimension ranges from 32 to 128, which is much smaller than the number of channels in the image feature extraction branch. At each time step, the current single scalar wind speed time series value and the hidden state of the previous time step are fed into this branch. The first layer of forward and backward long short-term memory units scans the sequence along the time axis in the forward and backward directions, respectively, capturing the rising and falling trends of airflow velocity and periodic pulsation patterns. The second layer further aggregates the bidirectional hidden states and outputs a compact temporal pattern feature vector. Since airflow velocity still fluctuates continuously under normal conditions, and abnormal precursors often manifest as small amplitude shifts, the temporal pattern feature vector of this branch is extremely sensitive to information loss and must retain sufficient numerical accuracy.

[0058] S2.3: Receive the spatial temperature field feature map output by the image feature extraction branch and the temporal pattern feature vector output by the temporal feature extraction branch through the cross-modal attention fusion algorithm.

[0059] The cross-modal attention fusion algorithm first flattens the spatial temperature field feature map into a spatial location sequence vector, and then extends the temporal pattern feature vector into a query vector of the same dimension as the spatial location vector through a linear mapping. Then, using the spatial location sequence vector as the key matrix and the temporal feature query vector as the query matrix, it calculates the query matrix and the... Original similarity score of the key matrix After normalization, the attention weight distribution is obtained. Then, the value matrix is ​​weighted and summed using these attention weights to generate a time-aware spatial context vector. The calculation formula is as follows:

[0060]

[0061] in, Time series feature query vector The first in A key vector at a spatial location, For feature dimension, i.e., vector and The length.

[0062] Temporally Aware Spatial Context Vectors The calculation formula is as follows:

[0063]

[0064] in, For the first The attention weight of the i-th spatial location represents the attention weight of the temporal mode to the i-th position at the current time step. The intensity of attention at each spatial location, satisfying , Key vector The corresponding number A vector of values ​​at spatial locations This represents the total number of spatial locations.

[0065] Attention weight The calculation formula is as follows:

[0066] in, For query matrix and the first The original similarity score of the key matrix.

[0067] Simultaneously, following the same attention mechanism framework, the symmetry direction is calculated: the temporal mode feature vector is flattened and linearly mapped into a key matrix and a value matrix; the spatial temperature field feature map is compressed into a query vector through global average pooling or linear mapping. The scale dot product similarity between the query vector and the key vector at each temporal location is calculated, and after normalization, the attention weight distribution of the spatial mode on each time step (or each hidden state position) is obtained. Then, the temporal value matrix is ​​weighted and summed using this attention weight to generate a spatially aware temporal context vector. This vector indicates which time points or patterns in the airflow velocity sequence the model should focus on under the current spatial temperature field state.

[0068] The temporally sensed spatial context vector is concatenated with the spatially sensed temporally sensed context vector along the feature dimension to form the final fused feature vector. This fused feature vector includes cross-modal correlation information in two directions: the temperature distribution region most relevant to airflow velocity changes and the temporally sensed mode most relevant to temperature field state.

[0069] S2.4: Input the fused feature vector into the anomaly determination head network, and pre-train the anomaly detection deep learning model using the labeled training dataset until convergence; the anomaly determination head network is the lightweight classification sub-network at the end of the anomaly detection deep learning model, which is responsible for receiving the fused feature vector and mapping it to the final anomaly determination result.

[0070] The anomaly detection head network receives a fused feature vector as input, formed by concatenating a time-aware spatial context vector and a spatially-aware temporal context vector. The anomaly detection head network consists of three sequentially connected fully connected layers. The first fully connected layer linearly maps the fused feature vector from 2D to 512D, processes it using the ReLU non-linear activation function, and then inputs it to the second fully connected layer. The second fully connected layer further compresses the feature dimension to 256D and also performs ReLU activation. The third fully connected layer linearly maps the 256D feature to a 1D scalar value, and then uses the Sigmoid activation function to map this scalar value to the [0,1] interval, outputting the anomaly probability at the current time step.

[0071] Samples are extracted from the historical records of temperature distribution image sequences and single scalar wind speed time series values ​​in the training dataset, using a sliding window method based on preset time windows. Maintenance personnel then label each time window sample with normal status and abnormal status types such as local hotspots, cold aisle airflow blockage, power distribution anomalies, and cooling failures, based on historical alarm records and maintenance logs. The labeled training set samples are then fed batch by batch into the anomaly detection deep learning model for forward propagation, and the cross-entropy loss function value between the model's output anomaly probability and the true label is calculated. The error backpropagation algorithm updates all trainable parameters of the anomaly detection deep learning model batch by batch with a preset learning rate. Forward and backpropagation are iteratively performed until the model's anomaly detection accuracy on the independent validation set reaches the preset convergence condition. The calculation formula is as follows:

[0072]

[0073] in, For the true labels of the training samples, , This indicates that the current time step is in any abnormal state (local hotspot, cold aisle airflow blockage, power distribution abnormality, or cooling failure). 0 indicates the normal state; The anomalous probability is obtained by forward propagation of the fused feature vector through the anomaly detection head network and activation by Sigmoid.

[0074] The aforementioned backpropagation algorithm generally refers to a deep neural network parameter optimization algorithm based on the chain rule of differentiation. It is mainly used in the pre-training stage of anomaly detection deep learning models to calculate the gradient of the cross-entropy loss function value with respect to all trainable parameters in the anomaly judgment head network, cross-modal attention fusion algorithm, temporal feature extraction branch, and image feature extraction branch. The gradient information is then propagated backward along the network layers, and the weights of each layer are updated with a preset learning rate. This iteratively minimizes the difference between the anomaly probability output by the model and the true label until the model converges.

[0075] S3: Evaluate the layered redundancy tolerance of the anomaly detection deep learning model. Calculate the redundancy tolerance coefficient based on the structural complexity, efficiency factor, and information change rate of the network layers in the model. A higher redundancy tolerance coefficient indicates a stronger tolerance for compression operations. Based on this coefficient, network layers are categorized into high redundancy tolerance, medium redundancy tolerance, and low redundancy tolerance levels, providing a basis for subsequent differentiated compression. The calculation formula is as follows:

[0076]

[0077] in, Structural complexity is the ratio of the number of trainable parameters in this layer to the total number of parameters in the model, with a value range of [0,1]. A higher structural complexity indicates that the layer carries a larger proportion of the parameter capacity in the model, and indicates that there are more redundant channels or connections. The effective factor represents the ratio of the receptive field of the input data corresponding to any position in the output feature map of this layer to the total size of the input feature map of this layer, and its value ranges from [0,1]. The larger the effective factor, the more likely that a single neuron in this layer has the ability to aggregate large-scale or even global spatial information, and the contribution of local details to the final judgment is relatively reduced. The information change rate is defined as 1 minus the average cosine similarity between channels of the current layer's output feature map over N consecutive normal state time steps. The smaller the information change rate, the more unchanged the output of the layer is in steady state, and the higher its tolerance to compression. , and All are preset weights. .

[0078] The calculation formula is as follows:

[0079]

[0080] in, For layer Time step The output feature map is vectorized. This is the preset number of consecutive normal state time steps. This indicates that the average value is taken over each channel dimension. This is the cosine similarity function.

[0081] After calculating the redundancy tolerance coefficient of each trained network layer in the anomaly detection deep learning model, a preset high redundancy threshold is applied. and low redundancy threshold (satisfy All network layers are divided into three redundancy levels according to the following correspondence:

[0082] like ,but The layers are divided into high redundancy tolerance layers. In the typical structure of the IDC data center monitoring model, this layer corresponds to all deep convolutional layers in the third and fourth residual stages of the image feature extraction branch.

[0083] like ,but The layers are divided into medium redundancy-tolerant layers, which correspond to the shallow convolutional layers in the first and second residual stages of the image feature extraction branch and the linear mapping layer in the cross-modal attention fusion algorithm.

[0084] like ,but The layers are divided into low redundancy-tolerant layers, which correspond to all weight matrix layers of the two-layer bidirectional long short-term memory network in the temporal feature extraction branch and the first fully connected layer in the anomaly detection head network that directly receives the fused feature vector.

[0085] S4: Based on the hierarchical redundancy tolerance coefficient, perform differential compression on the high redundancy tolerance layer and the medium redundancy tolerance layer in the pre-trained anomaly detection deep learning model to generate a compressed model. This facilitates the use of the evaluation results of S3 to perform hierarchical differential compression and eliminate the structural redundancy proposed in S1.

[0086] The specific working principle of S4 is as follows:

[0087] S4.1: For the high redundancy tolerance layer defined in S3, i.e., all deep convolutional layers within the third and fourth residual stages of the image feature extraction branch, the redundancy tolerance coefficients calculated in S3 are used for channel importance screening. For each convolutional layer within this layer, the sum of the absolute values ​​of all elements in the matrix of activation values ​​for each output channel is calculated on the temperature distribution image sequence, serving as a measure of channel importance. The calculation formula is as follows:

[0088]

[0089] in, , These represent the height and width of the output feature map of the convolutional layer in the spatial dimension, respectively. For the residual stage of the image feature extraction branch in S2.1, these values ​​decrease as the layer deepens. For the first When a frame is input, the layer No. The spatial location of each channel The activation value at the location is determined layer by layer by the image feature extraction branch through the temperature distribution image of the first type of sensor data in S1. This represents the number of frames in the temperature distribution image sequence used for evaluation.

[0090] The channel importance metrics are sorted in descending order, and the proportion of channels to be retained is determined by combining this with the redundancy tolerance coefficient of the layer. The larger the value, the lower the retention rate. For channels that do not enter the retention interval after sorting, all corresponding convolution kernel weights are set to zero, and a sparse mask is applied to each channel at the output of that layer, so that channels with zero weights are directly skipped during the inference stage and do not participate in subsequent calculations. Because S3 has evaluated that this level has a high effective factor and a low rate of information change, the spatial temperature field feature map remains almost unchanged in steady state, and the spatial temperature distribution pattern on the back of the cabinet can still be characterized after channel sparsification.

[0091] S4.2: For the moderate redundancy-tolerant layers defined in S3, namely the shallow convolutional layers in the first and second residual stages of the image feature extraction branch and the linear mapping layers in the cross-modal attention fusion algorithm, connections with absolute weight values ​​below a preset threshold are set to zero, forming a sparse connection matrix. This directly reduces the number of effective connections and lowers the computational cost during edge inference. This step compresses the number of effective connections without changing the layer structure, thus reducing the computational cost during edge inference.

[0092] S4.3: The low redundancy tolerance layer of S3, namely all weight matrix layers of the two-layer bidirectional long short-term memory network of the sequential feature extraction branch and the first fully connected layer in the anomaly detection head network that directly receives the fused feature vector, retains the original structure completely without any sparsification or zeroing operations, so as to ensure that the anomaly precursor signals such as the small amplitude shift mentioned in S2.2 are not lost due to compression.

[0093] S4.4: The high redundancy tolerance level after channel sparsification in S4.1, the medium redundancy tolerance level after weight sparsification in S4.2, and the low redundancy tolerance level retained in its original form in S4.3 are combined to obtain the compressed model. Under the full-channel input condition of S1, this model automatically suppresses the parameters and computational load corresponding to the low information increment sensor channels.

[0094] S5: Deploy the compressed model to edge nodes directly connected to the infrared thermal imaging array and ultrasonic airflow velocity meter. It acquires the current frame's temperature distribution image and a single scalar wind speed value in real time. It calculates the cosine similarity between the temperature distribution image and the previous frame, as well as the absolute difference between the wind speed value and the previous moment. If both indicators are within the threshold, the IDC data center monitoring data is considered to be in a stable state, and only the calculations corresponding to the non-sparsed channels are performed; otherwise, all network layers are executed. This allows for dynamic adjustment of the inference computation load based on the real-time changes in the input data, reducing edge computing power consumption during stable periods while maintaining the ability to detect minor anomalies.

[0095] First, the compressed model generated by S4 is deployed to a node directly connected to the infrared thermal imaging array and the ultrasonic airflow velocity meter to receive data from both types of sensors.

[0096] Next, acquire a temperature distribution image from the first type of sensor and a single scalar wind speed value from the second type of sensor at the current moment. Calculate the cosine similarity between the current temperature distribution image and the previous temperature distribution image, as well as the absolute difference between the current wind speed value and the wind speed value at the previous moment.

[0097] If the cosine similarity is higher than the preset image stability threshold and the absolute difference is lower than the preset time-series stability threshold, the current IDC data center monitoring data is determined to be in a stable state; otherwise, it is determined to be in a potentially changing state.

[0098] When the monitoring data in the IDC data center is stable, the compression model utilizes the low rate of change of information in the high redundancy tolerance level of S3 to skip all calculations corresponding to the channels that have been set to zero in S4.1, and additionally bypasses the calculations in that level. The highest-order convolutional blocks further reduce edge computing power consumption and avoid the problem of insufficient image branch compression caused by single compression.

[0099] When in a state of potential change, all network layers of the compressed model participate in forward inference in full according to the structure after S4 compression. The low redundancy tolerance layer runs with full-precision weights to ensure the ability of the time series model feature vector to capture small-amplitude shifts in wind speed.

[0100] Meanwhile, the anomaly detection head network of the compressed model outputs the anomaly probability at each time step. When the anomaly probability exceeds the preset alarm threshold for several consecutive inference cycles, a local alarm is triggered at the edge node, and the fused feature vector is uploaded to the central management platform.

[0101] This invention analyzes the structural complexity, effective factor, and information change rate of each network layer in S3, and implements channel sparsification for high redundancy tolerance layers with low information change rate in the image feature extraction branch in S4, and implements connection sparsification for medium redundancy tolerance layers. At the same time, it fully preserves the precision-sensitive low redundancy tolerance layers in the temporal feature extraction branch and the anomaly detection head network, so that the compression model can automatically suppress the parameters and computational load of the low information increment sensor channels. This solves the problem of parameter waste caused by the existing model uniformly encoding the temperature distribution image sequence and the single scalar wind speed time series value.

[0102] Meanwhile, the compression model acquires the current frame's temperature distribution image and a single scalar wind speed value in real time, calculates the inter-frame cosine similarity and the absolute difference in wind speed to determine whether the IDC data center monitoring data is in a stable state or a potentially changing state. In the stable state of the IDC data center monitoring data, it skips the already sparsed channels and the convolutional blocks with the highest redundancy tolerance coefficient to avoid image branches consuming edge computing power during low information periods. In the potentially changing state, all network layers participate fully in inference, and low redundancy tolerance layers run with full-precision weights to ensure the ability of the temporal pattern feature vector to capture small-amplitude shifts in wind speed.

[0103] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for compressing and deploying deep learning models for IDC (Internet Data Center) monitoring data, characterized in that, Includes the following steps: S1: Acquire heterogeneous monitoring data collected by multimodal sensors in the IDC data center, and perform data time synchronization and alignment to obtain the training dataset; S2: Construct an anomaly detection deep learning model that includes image feature extraction branch and temporal feature extraction branch, and pre-train the anomaly detection deep learning model using the training dataset until it converges; S3: Calculate the redundancy tolerance coefficient based on the structural complexity, effective factor, and information change rate of the network layers of the anomaly detection deep learning model. This coefficient is used to evaluate the hierarchical redundancy tolerance of the anomaly detection deep learning model and to divide each network layer into high redundancy tolerance level, medium redundancy tolerance level, and low redundancy tolerance level according to the value of the redundancy tolerance coefficient. S4: Based on the hierarchical redundancy tolerance coefficient, perform differential compression on the high redundancy tolerance layer and the medium redundancy tolerance layer in the pre-trained anomaly detection deep learning model to generate a compressed model. S5: Deploy the compressed model to an edge node directly connected to the infrared thermal imaging array and ultrasonic airflow velocity meter. Acquire the current frame temperature distribution image and a single scalar wind speed value in real time. Calculate the cosine similarity between the current frame temperature distribution image and the previous frame, as well as the absolute difference between the wind speed value and the previous moment. If both indicators are within the threshold, the IDC data center monitoring data is determined to be in a stable state; otherwise, it is determined to be in a potentially changing state. In the stable state of the IDC data center monitoring data, skip the sparsed channels and the convolutional block with the highest redundancy tolerance coefficient in the compressed model. In the potentially changing state, all network layers of the compressed model participate fully in inference, and the low redundancy tolerance layer runs with full precision weights. The redundancy tolerance coefficient A larger value indicates a stronger tolerance for compression operations in the current layer; the calculation formula is as follows: ， in, The structural complexity is the ratio of the number of trainable parameters in the current layer to the total number of parameters in the model. The effective factor represents the ratio of the receptive field range of the input data at any position in the current layer's output feature map to the total size of the current layer's input feature map; The rate of information change is represented by 1 minus the average cosine similarity between channels of the current layer's output feature map over N consecutive normal state time steps. , and All are preset weights.

2. The method for compressing and deploying deep learning models for IDC data center monitoring data according to claim 1, characterized in that, The heterogeneous monitoring data includes first-type sensor data and second-type sensor data with high dimensionality and low temporal resolution; the first-type sensor data is a temperature distribution image sequence output by an infrared thermal imaging array; the second-type sensor data is a single scalar wind speed time series value output by an ultrasonic airflow velocity meter.

3. The method for compressing and deploying deep learning models for IDC data center monitoring data according to claim 1, characterized in that, The specific steps of S2 are as follows: S2.1: The image feature extraction branch uses a deep convolutional neural network algorithm as the backbone network, and performs convolution processing on the input temperature distribution image sequence layer by layer through multiple stacked convolutional blocks; S2.2: The temporal feature extraction branch uses a two-layer bidirectional long short-term memory network as the backbone network to perform forward and backward convolution processing on the single scalar wind speed time series in the input second type of sensor data along the time dimension; S2.3: The spatial temperature field feature map output by the image feature extraction branch and the temporal pattern feature vector output by the temporal feature extraction branch are received by the cross-modal attention fusion algorithm and concatenated to obtain the fused feature vector; S2.4: Input the fused feature vector into the anomaly detection head network, and pre-train the anomaly detection deep learning model using the labeled training dataset until it converges.

4. The method for compressing and deploying deep learning models for IDC data center monitoring data according to claim 1, characterized in that, The specific divisions of the high redundancy tolerance level, medium redundancy tolerance level, and low redundancy tolerance level are as follows: If the redundancy tolerance coefficient ,but The layers are divided into high redundancy tolerance levels; If low redundancy threshold ,but The layers are divided into medium redundancy tolerance levels; like ,but The layers are divided into low redundancy tolerance levels.

5. The method for compressing and deploying deep learning models for IDC data center monitoring data according to claim 1, characterized in that, The specific working principle of S4 is as follows: S4.1: For the high redundancy tolerance level divided in S3, use the redundancy tolerance coefficient calculated in S3 to screen the importance of the channel; S4.2: For the moderate redundancy tolerance level partitioned in S3, the connections with absolute weight values ​​below a preset threshold are set to zero to form a sparse connection matrix. S4.3: The low redundancy tolerance level of S3 is fully preserved, without any sparsification or zeroing operations. S4.4: The high redundancy tolerance level after channel sparsification in S4.1, the medium redundancy tolerance level after weight sparsification in S4.2, and the low redundancy tolerance level that is retained in its original form in S4.3 are combined to obtain the compressed model.

6. The method for compressing and deploying deep learning models for IDC data center monitoring data according to claim 5, characterized in that, The specific content of S4.1 is as follows: For each convolutional layer within a high redundancy tolerance layer, the sum of the absolute values ​​of all elements in the matrix of activation values ​​for each output channel is calculated on the temperature distribution image sequence, serving as the channel importance score. The channel importance scores are then sorted in descending order, and the proportion of channels retained is determined by combining this with the redundancy tolerance coefficient. A higher redundancy tolerance coefficient results in a lower retention rate. For channels that do not enter the retention range after sorting, all corresponding convolutional kernel weights are set to zero. The calculation formula is as follows: ， in, , These represent the height and width of the output feature map of the convolutional layer in the spatial dimension, respectively. For the residual stage of the image feature extraction branch in S2.1, these values ​​decrease as the layer deepens. For the first When a frame is input, the layer No. The spatial location of each channel Activation value at that location, This represents the number of frames in the temperature distribution image sequence used for evaluation.

7. The method for compressing and deploying deep learning models for IDC data center monitoring data according to claim 1, characterized in that, The specific content of S5 is as follows: First, the compressed model generated by S4 is deployed to a node directly connected to the infrared thermal imaging array and the ultrasonic airflow velocity meter to receive data from the first type of sensor and the second type of sensor. Then, acquire a temperature distribution image from the first type of sensor and a single scalar wind speed value from the second type of sensor at the current moment, and calculate the cosine similarity between the current temperature distribution image and the previous temperature distribution image, as well as the absolute difference between the current wind speed value and the wind speed value at the previous moment. If the cosine similarity is higher than the preset image stability threshold and the absolute difference is lower than the preset time-series stability threshold, the current IDC data center monitoring data is determined to be in a stable state; otherwise, it is determined to be in a potentially changing state.

Images