IT equipment inspection report generation system and method based on deep neural network

Abstract

The invention relates to the technical field of information technology equipment maintenance, in particular to an IT equipment inspection report generation system and method based on a deep neural network. An acquisition model is adopted to generate a control instruction set and an acquisition instruction set to acquire key operation information of IT equipment, a plurality of data transmission chains form a data transmission group, a high-speed data I / O module is numbered in a group-number mode, a deep neural network model module is adopted to generate an acquisition model, a marshalling rule is generated, and a marshalling result is obtained. Performing pairing grouping on each multi-element intelligent acquisition control module and the high-speed data I / O module, determining a number in a group-number mode, establishing a grouping index, determining a group in the group-number mode, and providing prediction data for the multi-element intelligent acquisition control modules to perform interpolation processing on operation key information; according to the invention, the stability and accuracy of data acquisition are ensured, the IT operation and maintenance efficiency and quality are improved, and the burden of operation and maintenance personnel is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of information technology equipment maintenance, and specifically relates to a system and method for generating an IT equipment inspection report based on a deep neural network. Background Art

[0002] In the digital age, IT equipment has become a key support for enterprise operations and the operation of data centers. The traditional manual inspection mode of IT equipment has many drawbacks: firstly, the efficiency is extremely low. There are numerous devices in large data centers, and it is difficult for manual inspection to be frequent and detailed; secondly, the subjectivity is too strong. The inspection results are easily affected by the professional qualities and states of personnel, resulting in low quality, and potential faults are difficult to detect in a timely manner or misjudgments occur frequently; thirdly, the real-time performance is poor. There are blind spots in the fixed-cycle inspection. If a sudden device failure occurs during the interval period, the risk of business interruption increases significantly.

[0003] Although some existing inspection systems collect data and conduct preliminary analysis by means of simple sensing technologies, they are still insufficient in the face of complex faults, early micro-fault signs, and performance dynamic trend prediction, and cannot meet the increasingly strict refined requirements of IT operation and maintenance. The deep neural network technology has achieved remarkable results in many fields, providing an opportunity for the intelligent inspection of IT equipment. The GPT technology has outstanding advantages in natural language processing, can meet the requirements of generating inspection reports, and improve the inspection system. However, the existing inspection systems based on simple sensing technologies are difficult to handle complex faults and performance prediction. The main reason is that their data processing algorithms are relatively simple, it is difficult to accurately extract fault features from complex data, and the depth and breadth of model training are limited, and various fault modes and performance change situations cannot be effectively covered. Summary of the Invention

[0004] Aiming at the above deficiencies of the existing technologies, the present invention aims to provide a system and method for generating an IT equipment inspection report based on a deep neural network, so as to achieve automated, high-precision, and real-time monitoring and early warning, reduce equipment failure rates, reduce business downtime, comprehensively improve the efficiency and quality of IT operation and maintenance, and relieve the burden on operation and maintenance personnel.

[0005] To solve the above problems, the present invention adopts the following technical solutions:

[0006] On the one hand, the present invention provides a system for generating an IT equipment inspection report based on a deep neural network, including a multi-intelligent acquisition and control module, a deep neural network model module, and a high-speed data I / O module;

[0007] The multi-intelligence acquisition control module is set on the IT device, used to collect the basic parameter information of the IT device, and generate a control instruction set and an acquisition instruction set using an acquisition model to control the multi-intelligence acquisition control module to collect the key operation information of the IT device; the multi-intelligence acquisition control module is communicatively connected to the deep neural network model module through a high-speed data I / O module with a device number identifier;

[0008] One multi-intelligence acquisition control module and one high-speed data I / O module with a device number identifier form a data transmission chain, and multiple data transmission chains form a data transmission group; the high-speed data I / O modules within the data transmission group store all the data of the data transmission group, and the high-speed data I / O modules are numbered in a group-number manner;

[0009] The deep neural network model module is used to generate an acquisition model; generate a grouping rule, pair and group each multi-intelligence acquisition control module and high-speed data I / O module, determine the number in the group-number manner, establish a grouping index, and determine the group in the group-number manner; and when a timing interruption occurs in the collected key operation information, provide prediction data to the multi-intelligence acquisition control module to perform interpolation processing on the key operation information.

[0010] As an implementable manner, the generating the acquisition model includes:

[0011] Collect the basic parameter information of each IT device and the operation control information of the IT device. The basic parameter information includes brand, name, model, version number, and manufacturer; the operation control information includes acquisition instructions, control instructions, data types, acquisition frequencies, instruction correlations, and failure rates;

[0012] Construct the object features of the IT device according to the basic parameter information of the IT device. After inserting delimiters between the object features and the operation control information and between the operation control information, store them vectorially in a vector library to generate an acquisition model;

[0013] The acquisition model vectorizes the collected basic parameter information of the IT device, calculates the similarity between the vectorized basic parameter information and the vectorized object features in the vector library using cosine similarity or Euclidean distance, and uses the operation control information corresponding to the vectorized object features in the vector library with the highest similarity as the operation control information of the IT device.

[0014] As an implementable manner, the when a timing interruption occurs in the collected key operation information, providing prediction data to the multi-intelligence acquisition control module to perform interpolation processing on the key operation information includes:

[0015] When a timing interruption occurs in the collected key operation information, starting from the collection moment of the first missing data, find the collection moments at three collection frequency intervals forward for each missing data's collection moment;

[0016] Find the data at the same moment 10 days forward for each of the three collection frequency intervals' collection moments respectively, obtaining 10 groups of datasets with 3 in each group;

[0017] Subtract two adjacent numerical values in each group to obtain two data differences, and divide them according to time periods to obtain 2 groups of datasets with 10 data differences in each group;

[0018] Take the median of the data differences in each group of the 2 groups of datasets with 10 data differences in each group, obtaining two medians;

[0019] If the deviation between the two medians is less than the first adjustment value, use the arithmetic mean of the two medians as the interpolation of the missing data and mark the interpolation with an interpolation mark; if the proportion of the data source of the selected data at the same moment in the recent 10 days that is interpolated exceeds the second adjustment value, then take the data at the same moment 1 day forward until the proportion of the data source that is interpolated does not exceed the second adjustment value.

[0020] As an implementable manner, the generation grouping rule pairs and groups each multi - intelligent acquisition control module and high - speed data I / O module, determines the number in the group - numbering method, establishes a grouping index, and determines the group in the group - numbering method, including:

[0021] Use a deep neural network to perform feature engineering processing on the basic parameter information of all IT devices to obtain object features. Let S represent the number of object features. Select X object features from the object features to screen all IT devices, obtaining IT devices containing X object features. Group the IT devices containing X object features into a data transmission group with K as the grouping width;

[0022] When the data transmission packet loss rate or data error rate of the deep neural network model module is higher than 5%, increase the value of K; when the number of IT devices containing X object features is greater than K, increase the value of X;

[0023] Use the pre - trained model RoBERTa to pre - train the deep neural network model with the grouping width K and X, as well as the data transmission packet loss rate and data error rate as inputs, update the deep neural network model, and adjust the grouping width K and X.

[0024] As an implementable manner, the deep neural network model module is used for dynamically electing high - speed data I / O modules; the dynamic election of high - speed data I / O modules includes:

[0025] The deep neural network model module receives the operation key information of the high-speed data I / O module in each data transfer group. The operation key information includes throughput parameters, time-sharing throughput parameters, I / O rate parameters, online duration parameters, fault status parameters, aging status, and quality inspection approaching expiration status. According to the decision-making parameter U provided by the decision maker, it adjusts the weight Q1 values corresponding to each throughput parameter, time-sharing throughput parameter, I / O rate parameter, and online duration parameter, and performs weighted summation to obtain the recommended value T. According to the decision-making parameter J provided by the decision maker, it adjusts the weight Q2 values corresponding to the fault status parameter, aging status, and quality inspection approaching expiration status, and performs weighted summation to obtain the detection value M. When it is detected that the value of M is greater than 1, it sorts according to the size of the T value and recommends the high-speed data I / O module with the largest T value in the data transfer group for data transfer, and notifies the multi-intelligent acquisition control module to adjust the I / O object.

[0026] As an implementable manner, the deep neural network of the deep neural network model module uses a multi-layer perceptron as the basic architecture, including an input layer, a hidden layer, and an output layer. It calculates its weights Q1 and Q2 through the input throughput parameters, time-sharing throughput parameters, I / O rate parameters, online duration parameters, fault status parameters, aging status, and quality inspection approaching expiration status, and uses the mean square error as the loss function to evaluate the error between the calculated weights Q1 and Q2 and the true weights Q1 and Q2. It updates the weights and biases of the network according to the gradient of the loss function through the backpropagation method to train the deep neural network model.

[0027] As an implementable manner, it further includes an inspection report generation module. The inspection report generation module is used to generate an inspection report based on the operation key information of the collected IT devices, a preset prompt word library, and the data of the expansion slot as inputs, based on the multi-modal GPT technology.

[0028] Manually annotate the operation key information of the collected IT devices, train the annotation model through the manually annotated data. The annotation model is based on the input operation key information of the collected IT devices to obtain the operation key information of the collected IT devices with annotations. Combine the annotations with the prompt words in the preset prompt word library, and input them into the multi-modal GPT model through the report generation rules provided by the expansion slot to generate an inspection report.

[0029] As an implementable manner, the generated inspection report presents a decision interface through a visualization interface, supports report preview, generation, push, and export and archiving, and is presented through a statistical large screen, desktop terminal, mobile terminal, and VR device terminal.

[0030] On the other hand, the present invention provides a method for generating an inspection report of IT devices based on a deep neural network, including:

[0031] The multi-intelligence acquisition control module is set on the IT device to collect the basic parameter information of the IT device, and a collection model is used to generate a control instruction set and a collection instruction set to control the multi-intelligence acquisition control module to collect the key operation information of the IT device; the multi-intelligence acquisition control module is communicatively connected to the deep neural network model module through a high-speed data I / O module with a device number identifier;

[0032] One multi-intelligence acquisition control module and one high-speed data I / O module with a device number identifier form a data transmission chain, and multiple data transmission chains form a data transmission group; the high-speed data I / O modules within the data transmission group store all the data of the data transmission group, and the high-speed data I / O modules are numbered in a group-number manner;

[0033] A deep neural network model module is used to generate a collection model; a grouping rule is generated to pair and group each multi-intelligence acquisition control module and high-speed data I / O module, determine the number in the group-number manner, establish a grouping index, and determine the group in the group-number manner; and when a timing interruption occurs in the collected key operation information, prediction data is provided to the multi-intelligence acquisition control module to perform interpolation processing on the key operation information.

[0034] As an implementable manner, the generation of the collection model includes:

[0035] Collect the basic parameter information of each IT device and the operation control information of the IT device. The basic parameter information includes brand, name, model, version number, and manufacturer; the operation control information includes collection instructions, control instructions, data types, collection frequencies, instruction correlations, and failure rates;

[0036] Construct the object characteristics of the IT device according to the basic parameter information of the IT device. After inserting delimiters between the object characteristics, operation control information, and each operation control information, vectorize and store them in the vector library to generate a collection model;

[0037] The collection model vectorizes the collected basic parameter information of the IT device, calculates the similarity between the vectorized basic parameter information and the vectorized object characteristics in the vector library using cosine similarity or Euclidean distance, and uses the operation control information corresponding to the vectorized object characteristics in the vector library with the highest similarity as the operation control information of the IT device.

[0038] The beneficial effects of the present invention are as follows: The IT device inspection report generation system and method based on a deep neural network of the present invention realize the dynamic matching of the multi-intelligent acquisition control module and the high-speed data I / O module, ensure the stability and accuracy of data acquisition, achieve high-precision and real-time monitoring and early warning, reduce business downtime, comprehensively improve the efficiency and quality of IT operation and maintenance, and relieve the burden on operation and maintenance personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of an IT device inspection report generation system based on a deep neural network of the present invention.

[0040] Figure 2 It is a flowchart of an IT device inspection report generation method based on a deep neural network of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] The present invention will be further described in detail below in conjunction with specific embodiments.

[0042] It should be noted that these embodiments are only used to illustrate the present invention, rather than limiting the present invention. Under the premise of the concept of the present invention, simple improvements to this method all fall within the scope of protection required by the present invention.

[0043] Refer to Figure 1 , it is an IT device inspection report generation system based on a deep neural network, including a multi-intelligent acquisition control module 100, a deep neural network model module 200, and a high-speed data I / O module 300;

[0044] The multi-intelligent acquisition control module 100 is arranged on the IT device, and is used to collect the basic parameter information of the IT device, and generate a control instruction set and an acquisition instruction set by using an acquisition model to control the multi-intelligent acquisition control module to collect the key operation information of the IT device; the multi-intelligent acquisition control module is communicatively connected to the deep neural network model module through a high-speed data I / O module with a device number identifier;

[0045] A multi-intelligent acquisition control module 100 and a high-speed data I / O module 300 with a device number identifier form a data transmission chain, and multiple data transmission chains form a data transmission group; the high-speed data I / O modules in the data transmission group store all the data of the data transmission group, and the high-speed data I / O modules are numbered in a group-number manner, where the group represents the number of the data transmission group, and the number represents the multi-intelligent acquisition control module or the IT device number;

[0046] The deep neural network model module 300 is used to generate an acquisition model; generate grouping rules, pair and group each multi-intelligent acquisition control module and high-speed data I / O module, determine the number in the group-numbering method, establish a grouping index, and determine the group in the group-numbering method; and when a timing interruption occurs in the collected key operation information, provide prediction data to the multi-intelligent acquisition control module to perform interpolation processing on the key operation information.

[0047] The generation of the acquisition model includes:

[0048] Collect the basic parameter information of each IT device and the operation control information of the IT device. The basic parameter information includes brand, name, model, version number, and manufacturer; the operation control information includes acquisition instructions, control instructions, data types (status, switch quantity, instantaneous value, cumulative quantity), acquisition frequency, instruction correlation, and failure rate.

[0049] According to the data points / metrics of the collected IT devices, for example, there are a total of 20 including P1 - P20, P1 - P3 correspond to the acquisition instruction Ca, P4 - P7 correspond to the acquisition instruction Cb, P8 - P15 correspond to the acquisition instruction Cc, and P15 - P20 correspond to the acquisition instruction Cd. Different acquisition instructions and acquisition frequencies for different IT devices can be obtained according to the data points / metrics of the collected IT devices.

[0050] For example, the data volumes collected by Ca and Cc, or Ca and Cd, or Cb and Cc, or Cb and Cd need to be correlated in the same time period, and the instruction correlation can be obtained.

[0051] The control instructions include resetting the target IT device, memory optimization, file compression, performance mode change, security policy adjustment, traffic control, and instructions inherent to the device itself.

[0052] Construct the object characteristics of the IT device according to the basic parameter information of the IT device. After inserting delimiters between the object characteristics, the operation control information, and each operation control information, store them vectorially in the vector library to generate the acquisition model;

[0053] The acquisition model vectorizes the collected basic parameter information of the IT device, calculates the similarity between the vectorized basic parameter information and the vectorized object characteristics in the vector library using cosine similarity or Euclidean distance, and uses the operation control information corresponding to the vectorized object characteristics in the vector library with the highest similarity as the operation control information of the IT device.

[0054] This embodiment provides a feasible solution for the above-mentioned situation where when a timing interruption occurs in the collected key operation information, providing prediction data to the multi-intelligent acquisition control module to perform interpolation processing on the key operation information:

[0055] When a timing interruption occurs in the collected key operation information, starting from the collection moment of the first missing data, find the collection moments at three collection frequency intervals forward for the collection moment of each missing data;

[0056] Find the data at the same moment 10 days forward for each of the three collection frequency intervals to obtain 10 sets of data sets with 3 data in each set;

[0057] Subtract two adjacent values in each group to obtain two data differences, and divide them into 2 sets of data differences with 10 data in each set according to time periods;

[0058] Take the median of the data differences in each set of the 2 sets of data differences with 10 data in each set to obtain two medians;

[0059] If the deviation between the two medians is less than the first adjustment value, use the arithmetic mean of the two medians as the interpolation of the missing data and mark the interpolation with an interpolation mark; if the proportion of the data source of the selected data at the same moment in the recent 10 days that is interpolated exceeds the second adjustment value, then take the data at the same moment 1 day forward until the proportion of the data source that is interpolated does not exceed the second adjustment value.

[0060] This embodiment provides the above-mentioned generation grouping rules, pairs and groups each multi-intelligent acquisition control module and high-speed data I / O module, determines the number in the group-numbering method, establishes a grouping index, and determines a feasible solution for the group in the group-numbering method:

[0061] Use a deep neural network to perform feature engineering processing on the basic parameter information of all IT devices to obtain object features. Let S represent the number of object features (such as model number, version number, manufacturer). Determine X (such as 2, that is, any two of the model number, version number, and manufacturer) object features in the object features to screen all IT devices, and obtain IT devices containing X object features (obtain IT devices containing any two data among the model number, version number, and manufacturer). Use K as the grouping width to form a data transmission group with K IT devices containing X object features;

[0062] When the data transmission packet loss rate or data error rate of the deep neural network model module is higher than 5%, increase the value of K; when the number of IT devices containing X object features is greater than K, increase the value of X;

[0063] Use the grouping width K and X, as well as the data transmission packet loss rate and data error rate as inputs, and use the pre-trained model RoBERTa to pre-train the deep neural network model, update the deep neural network model, and adjust the grouping width K and X.

[0064] After determining the new group number, the deep neural network model module will accurately notify the multi-intelligent acquisition control module of the new grouping information to ensure that each multi-intelligent acquisition control module can record the high-speed data I / O module paired with it in a timely manner. At the same time, the deep neural network model module maintains information interaction with the high-speed data I / O module to ensure that the grouped high-speed data I / O modules can correctly serve the corresponding multi-intelligent acquisition control modules.

[0065] The deep neural network model module is used for dynamically electing high-speed data I / O modules; the dynamic election of high-speed data I / O modules includes:

[0066] The deep neural network model module receives the operation key information of the high-speed data I / O modules within each data transmission group. The operation key information includes throughput parameters, time-sharing throughput parameters, I / O rate parameters, online duration parameters, fault status parameters, aging status, and quality inspection approaching expiration status. According to the decision-making parameter U provided by the decision maker, it adjusts the weight Q1 values corresponding to each of the throughput parameter, time-sharing throughput parameter, I / O rate parameter, and online duration parameter (the specific values of Q1 corresponding to each parameter are not necessarily the same), and performs weighted summation to obtain the recommended value T; according to the decision-making parameter J provided by the decision maker, it adjusts the weight Q2 values corresponding to the fault status parameter, aging status, and quality inspection approaching expiration status respectively, and performs weighted summation to obtain the detection value M; when it is detected that the value of M is greater than 1, it sorts according to the value of T, recommends the high-speed data I / O module with the largest T value within the data transmission group for data transmission, and notifies the multi-intelligent acquisition control module to adjust the I / O object.

[0067] Among them, the deep neural network of the deep neural network model module uses a multi-layer perceptron as the basic architecture, including an input layer, a hidden layer, and an output layer; it calculates the weights Q1 and Q2 respectively through the input throughput parameters, time-sharing throughput parameters, I / O rate parameters, online duration parameters, fault status parameters, aging status, and quality inspection approaching expiration status, and uses the mean square error as the loss function to evaluate the error between the calculated weights Q1 and Q2 and the true weights Q1 and Q2; it updates the weights and biases of the network according to the gradient of the loss function through the backpropagation method to train the deep neural network model.

[0068] For the input, a vector is constructed, and the number of neurons in the input layer is determined according to its dimension. For example, the number of neurons in the input layer is the sum of the numbers of the throughput parameter, time-sharing throughput parameter, I / O rate parameter, online duration parameter, fault status parameter, aging status, and quality inspection approaching expiration status parameters.

[0069] The number of hidden layers and the number of neurons can be determined according to specific problems and data volume. In the present invention, 3 hidden layers are adopted, and each hidden layer contains 64 to 256 neurons to capture the complex relationships in the data.

[0070] The number of neurons in the output layer depends on the number of weights to be adjusted. For the parameters corresponding to Q1, the output layer will have the number of neurons corresponding to the Q1 weights. The same applies to the parameters corresponding to Q2.

[0071] For adjusting the weights Q1 of the first type of parameters according to the decision parameter U provided by the decision maker, the mean squared error (MSE) can be used as the loss function. That is, the average of the sum of the squares of the errors between the predicted Q1 value and the actual Q1 value. The formula is MSE = , where yi is the actual Q1 value, is the predicted Q1 value. The adjustment of Q2 is the same.

[0072] During the training process, the decision parameter U or J and the corresponding operating parameter feature vector are used as inputs together. The network outputs the predicted weights Q1 or Q2, and the weights and biases of the network are updated according to the gradient of the loss function through the backpropagation algorithm.

[0073] U is the preference parameter of the user for the first type of parameters (such as throughput, online duration, etc.). For example, the user may hope that the current system pays more attention to "throughput" rather than "online duration". At this time, U will increase this preference, and the deep neural network will adjust the weight distribution of Q1 according to U.

[0074] Input: The first type of parameters + decision parameter U.

[0075] Label: The expected weight Q1 (generated from historical data or manual annotation). For example, when U represents "high throughput priority", the weight of the throughput parameter is large.

[0076] Loss function: Mean squared error (MSE), calculating the gap between the predicted Q1 value and the true Q1 value.

[0077] Backpropagation: Update the network parameters through gradient descent to make the predicted Q1 approach the true Q1.

[0078] Encoding of U: The decision parameter U can be a scalar (such as U = 0.8 representing "high throughput preference"), or a vector (such as U = [0.3, 0.7] representing multi-objective weights).

[0079] Dynamic adjustment: When the user updates U (such as changing from "throughput priority" to "online duration priority"), the deep neural network will re-predict Q1 according to the new U value.

[0080] For the operating parameters of the new high-speed data I / O module, input them into the trained deep neural network to obtain the predicted Q1 and Q2 values. According to the predicted Q1 and Q2 values, calculate the recommended value T and the detected value M. When M > 1, the election module sorts according to the T value and recommends the high-speed data I / O module with the highest T value within the group. Notify the multi-intelligent acquisition control module to adjust the I / O object to ensure the stability and performance optimization of the system.

[0081] The system also includes an inspection report generation module 400, which is used to generate an inspection report based on the key operating information of the collected IT devices, the preset prompt word library, and the data in the expansion slot as inputs, using multi-modal GPT technology;

[0082] Manually annotate the key operating information of the collected IT devices, train the annotation model with the manually annotated data. The annotation model, based on the input key operating information of the collected IT devices, obtains the key operating information of the collected IT devices with annotations. Combine the annotations with the prompt words in the preset prompt word library, and input them into the multi-modal GPT model through the report generation rules / templates provided by the expansion slot to generate an inspection report. The expansion slot supports the input and correction of conditional parameters.

[0083] The generated inspection report presents a decision-making interface through a visualization interface, supporting report preview, generation, push, export and archiving, and is presented through a statistical large screen, desktop terminal, mobile terminal, and VR device terminal.

[0084] After receiving the report generated by the inspection report generation module, the report can be previewed, and in the report output unit, select to export, push, or generate control instructions to the scheduler to finally quickly avoid potential risks of the device or repair faults.

[0085] For different forms of IT devices, reasonably install the multi-intelligent acquisition control module and the high-speed data I / O module outside or inside them. When installing, consider the signal transmission stability and anti-interference ability to ensure the smoothness of the data transmission channel. For large server clusters, due to the complex device structure and strong electromagnetic interference, when installing the multi-intelligent acquisition control module, a location far from strong interference sources should be selected, and shielded cables should be used to connect the sensors and the module; for devices such as small switches, the module can be installed externally to ensure normal operation of the device. The deep neural network model module is installed on a server with GPU computing resources, and the inspection report generation module driven by multi-modal GPT technology is installed on an application server.

[0086] The underlying layer of the multi - intelligent acquisition control module and the high - speed data I / O module is written in embedded C language, which is deeply adapted to the hardware drivers of multi - sensors to achieve low - power and highly reliable operation. The deep neural network model relies on the mainstream open - source framework of TensorFlow to build a training script system, and regularly introduces fresh data for incremental training and optimization iteration. Specifically, new device operation data is collected once a week. After screening and pre - processing the data, it is added to the training set to retrain the model. Each time during training, parameters such as the learning rate and optimizer are adjusted according to the performance of the model on the validation set to improve the accuracy of the model.

[0087] See Figure 2 , a method for generating an IT device inspection report based on a deep neural network, including:

[0088] S100. Set the multi - intelligent acquisition control module on the IT device, collect the basic parameter information of the IT device, and use the acquisition model to generate a control instruction set and an acquisition instruction set to control the multi - intelligent acquisition control module to collect the key operation information of the IT device; the multi - intelligent acquisition control module is communicatively connected to the deep neural network model module through a high - speed data I / O module with a device number identifier.

[0089] S200. A multi - intelligent acquisition control module and a high - speed data I / O module with a device number identifier form a data transmission chain, and multiple data transmission chains form a data transmission group; each high - speed data I / O module in the data transmission group stores all the data of the data transmission group, and the high - speed data I / O module is numbered in the group - number way; the group represents the number of the data transmission group, and the number represents the multi - intelligent acquisition control module or the IT device number.

[0090] S300. Use the deep neural network model module to generate an acquisition model; generate a grouping rule, pair and group each multi - intelligent acquisition control module and high - speed data I / O module to determine the number in the group - number way, establish a grouping index, and determine the group in the group - number way; and when a timing interruption occurs in the collected key operation information, provide prediction data for the multi - intelligent acquisition control module to perform interpolation processing on the key operation information.

[0091] The above - mentioned generation of the acquisition model includes:

[0092] Collect the basic parameter information of each IT device and the operation control information of the IT device. The basic parameter information includes brand, name, model, version number, and manufacturer; the operation control information includes acquisition instructions, control instructions, data types, acquisition frequencies, instruction correlations, and failure rates.

[0093] Construct the object features of the IT device based on the basic parameter information of the IT device. After inserting delimiters between the object features, the operation control information, and each piece of operation control information, vectorize and store them in the vector library to generate an acquisition model.

[0094] The acquisition model vectorizes the basic parameter information of the IT device collected through the basic parameter information of the IT device collected. It calculates the similarity between the vectorized basic parameter information and the vectorized object features in the vector library using cosine similarity or Euclidean distance, and uses the operation control information corresponding to the vectorized object features in the vector library with the highest similarity as the operation control information of the IT device.

[0095] When a timing interruption occurs in the collected operation critical information, the interpolation process of the operation critical information provided by the multi-intelligent acquisition control module with prediction data includes:

[0096] When a timing interruption occurs in the collected operation critical information, starting from the acquisition moment of the first missing data, find the acquisition moments at three acquisition frequency intervals forward for each missing data's acquisition moment.

[0097] Find the data at the same moment 10 days forward for each of the three acquisition frequency intervals' acquisition moments to obtain 10 groups of datasets, with 3 data in each group.

[0098] Subtract two adjacent values in each group to obtain two data differences, and divide them into 2 groups of datasets with 10 data differences in each group according to time periods.

[0099] Take the median of the data differences in each group of the 2 groups of datasets with 10 data differences in each group to obtain two medians.

[0100] If the deviation between the two medians is less than the first adjustment value, use the arithmetic mean of the two medians as the interpolation of the missing data and mark the interpolation with an interpolation mark. If the proportion of the source of the data at the same moment in the selected recent 10 days that is interpolated exceeds the second adjustment value, then take the data at the same moment 1 day forward until the proportion of the source of the interpolated data does not exceed the second adjustment value.

[0101] For the above generation grouping rules, pair and group each multi-intelligent acquisition control module and high-speed data I / O module, determine the number in the group-numbering method, establish a grouping index, and determine the group in the group-numbering method, including:

[0102] The basic parameter information of all IT devices is processed by a deep neural network for feature engineering to obtain object features. Let S represent the number of object features (such as model number, version number, manufacturer). Among the object features, X object features are determined (such as 2, that is, any two of the model number, version number, and manufacturer are selected as the object features for screening) to screen all IT devices, and IT devices containing X object features are obtained (IT devices containing any two data among the model number, version number, and manufacturer). With K as the grouping width, the IT devices containing X object features are grouped into a data transmission group with K devices in each group;

[0103] When the data transmission packet loss rate or data error rate of the deep neural network model module is higher than 5%, the value of K is increased; when the number of IT devices containing X object features is greater than K, the value of X is increased;

[0104] Using the grouping width K and X, as well as the data transmission packet loss rate and data error rate as inputs, the pre-trained model RoBERTa is used to pre-train the deep neural network model, update the deep neural network model, and adjust the grouping width K and X.

[0105] After determining the new group number, the deep neural network model module will accurately notify the multi-intelligent acquisition control module of the new grouping information to ensure that each multi-intelligent acquisition control module can timely record the high-speed data I / O module paired with it. At the same time, the deep neural network model module maintains information interaction with the high-speed data I / O module to ensure that the grouped high-speed data I / O module can correctly serve the corresponding multi-intelligent acquisition control module.

[0106] Use the deep neural network model module to dynamically elect the high-speed data I / O module; the dynamic election of the high-speed data I / O module includes:

[0107] The deep neural network model module receives the key operation information of the high-speed data I / O module in each data transmission group. The key operation information includes throughput parameters, time-sharing throughput parameters, I / O rate parameters, online duration parameters, fault status parameters, aging status, and quality inspection approaching expiration status. According to the decision-making parameter U provided by the decision maker, the weight Q1 values corresponding to each throughput parameter, time-sharing throughput parameter, I / O rate parameter, and online duration parameter are adjusted (the specific values of Q1 corresponding to each parameter are not necessarily the same), and the weighted sum is used to obtain the recommended value T; according to the decision-making parameter J provided by the decision maker, the weight Q2 values corresponding to the fault status parameter, aging status, and quality inspection approaching expiration status are adjusted, and the weighted sum is used to obtain the detection value M; when it is detected that the value of M is greater than 1, according to the size of the T value, the high-speed data I / O module with the largest T value in the data transmission group is recommended for data transmission, and the multi-intelligent acquisition control module is notified to adjust the I / O object.

[0108] Among them, the deep neural network in the deep neural network model module uses a multi-layer perceptron as the basic architecture, including an input layer, a hidden layer, and an output layer; the weights Q1 and Q2 are calculated respectively through the input throughput parameter, time-sharing throughput parameter, I / O rate parameter, online duration parameter, fault status parameter, aging status, and quality inspection approaching expiration status, and the mean squared error is used as the loss function to evaluate the error between the calculated weights Q1 and Q2 and the true weights Q1 and Q2; the weights and biases of the network are updated according to the gradient of the loss function through the backpropagation method to train the deep neural network model.

[0109] For the input construction vector, the number of neurons in the input layer is determined according to its dimension. For example, the number of neurons in the input layer is the sum of the number of throughput parameter, time-sharing throughput parameter, I / O rate parameter, online duration parameter, fault status parameter, aging status, and quality inspection approaching expiration status parameters.

[0110] The number of hidden layers and the number of neurons can be determined according to the specific problem and the amount of data. There are 3 hidden layers, and each hidden layer contains 64 to 256 neurons to capture the complex relationships in the data.

[0111] The number of neurons in the output layer depends on the number of weights to be adjusted. For the parameters corresponding to Q1, the output layer will have the number of neurons corresponding to the Q1 weight. The same applies to the parameters corresponding to Q2.

[0112] For adjusting the weight Q1 of the first type of parameters according to the decision parameter U provided by the decision maker, the mean squared error (MSE) can be used as the loss function. That is, the average of the sum of the squares of the errors between the predicted Q1 value and the actual Q1 value, and the formula is MSE = , where yi is the actual Q1 value, is the predicted Q1 value. The adjustment of Q2 is the same.

[0113] During the training process, the decision parameter U or J and the corresponding operating parameter feature vector are used as inputs together, and the network outputs the predicted weights Q1 or Q2, and the weights and biases of the network are updated according to the gradient of the loss function through the backpropagation algorithm.

[0114] U is the preference parameter of the user for the first type of parameters (such as throughput, online duration, etc.). For example, the user may hope that the current system pays more attention to "throughput" rather than "online duration". At this time, U reflects this preference, and the deep neural network will adjust the weight allocation of Q1 according to U.

[0115] Input: the first type of parameters + decision parameter U.

[0116] Label: Desired weight Q1 (generated from historical data or manual annotation). For example, when U represents "high throughput priority", the weight of the throughput parameter is large.

[0117] Loss function: Mean Squared Error (MSE), which calculates the gap between the predicted Q1 value and the true Q1 value:

[0118] Backpropagation: Update the network parameters through gradient descent to make the predicted Q1 approach the true Q1.

[0119] Encoding of U: The decision parameter U can be a scalar (e.g., U = 0.8 represents "high throughput preference") or a vector (e.g., U = [0.3, 0.7] represents multi-objective weights).

[0120] Dynamic adjustment: When the user updates U (e.g., changes from "throughput priority" to "online duration priority"), the deep neural network will re-predict Q1 based on the new U value.

[0121] For the operating parameters of the new high-speed data I / O module, input them into the trained deep neural network to obtain the predicted Q1 and Q2 values. According to the predicted Q1 and Q2 values, calculate the recommended value T and the detection value M. When M > 1, the election module sorts according to the T value and recommends the high-speed data I / O module with the highest T value within the group. Notify the multi-intelligent acquisition control module to adjust the I / O object to ensure the stability and performance optimization of the system.

[0122] This method also uses the inspection report generation module to generate an inspection report based on the key operating information of the collected IT devices, the preset prompt word library, and the data in the expansion slot as inputs, using multi-modal GPT technology;

[0123] Manually annotate the key operating information of the collected IT devices, train the annotation model with the manually annotated data. The annotation model, based on the input key operating information of the collected IT devices, obtains the key operating information of the collected IT devices with annotations. Combine the annotations with the prompt words in the preset prompt word library, and input them into the multi-modal GPT model through the report generation rules provided by the expansion slot to generate an inspection report. The expansion slot supports the input and correction of conditional parameters.

[0124] The generated inspection report presents the decision-making interface through a visualization interface, supports report preview, generation, push, export and archiving, and is displayed through a statistical large screen, desktop terminal, mobile terminal, and VR device terminal.

[0125] After receiving the report generated by the inspection tour report generation module, the report can be previewed, and export, push, or generate control instructions to the scheduler can be selected in the report output unit to finally quickly avoid potential risks of the device or repair faults.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described by referring to the preferred embodiments of the present invention, those of ordinary skill in the art should understand that various changes can be made in form and details without departing from the spirit and scope of the present invention defined by the appended claims.

Claims

1. An IT device inspection report generation system based on a deep neural network, characterized in that, It includes a multiple-intelligence acquisition control module, a deep neural network model module, and a high-speed data I / O module; The multiple-intelligence acquisition control module is set on the IT device and is used to collect the basic parameter information of the IT device, and generate a control instruction set and an acquisition instruction set using an acquisition model to control the multiple-intelligence acquisition control module to collect the key operation information of the IT device; the multiple-intelligence acquisition control module is communicatively connected to the deep neural network model module through a high-speed data I / O module with a device number identifier; One multiple-intelligence acquisition control module and one high-speed data I / O module with a device number identifier form a data transmission chain, and multiple data transmission chains form a data transmission group; the high-speed data I / O modules within the data transmission group store all the data of the data transmission group, and the high-speed data I / O modules are numbered in a group-number manner; The deep neural network model module is used to generate an acquisition model; generate a grouping rule, pair and group each multiple-intelligence acquisition control module and high-speed data I / O module, determine the number in the group-number manner, establish a grouping index, and determine the group in the group-number manner; and when a timing interruption occurs in the collected key operation information, provide prediction data to the multiple-intelligence acquisition control module to perform interpolation processing on the key operation information.

2. The IT device inspection report generation system based on a deep neural network according to claim 1, wherein The generating the acquisition model includes: Collecting the basic parameter information of each IT device and the operation control information of the IT device, where the basic parameter information includes brand, name, model, version number, and manufacturer; the operation control information includes acquisition instructions, control instructions, data types, acquisition frequencies, instruction correlations, and failure rates; Constructing the object characteristics of the IT device according to the basic parameter information of the IT device, inserting delimiters between the object characteristics and the operation control information and between each operation control information, and storing them vectorially in a vector library to generate an acquisition model; The acquisition model vectorizes the collected basic parameter information of the IT device, calculates the similarity between the vectorized basic parameter information and the vectorized object characteristics in the vector library using cosine similarity or Euclidean distance, and uses the operation control information corresponding to the vectorized object characteristics in the vector library with the highest similarity as the operation control information of the IT device.

3. The IT device inspection report generation system based on a deep neural network according to claim 1, characterized in that, The when a timing interruption occurs in the collected key operation information, providing prediction data to the multiple-intelligence acquisition control module to perform interpolation processing on the key operation information includes: When a timing interruption occurs in the collected key operation information, starting from the acquisition moment of the first missing data, find the acquisition moments three acquisition frequency intervals forward for each missing data acquisition moment; Find the data at the same moment 10 days forward for each of the three acquisition frequency intervals of the acquisition moments to obtain 10 sets of datasets, with 3 data in each set; Subtract each adjacent pair of values in each set to obtain two data differences, and divide them into 2 sets of datasets with 10 data differences in each set according to time periods; Take the median of the data differences in each set of the 2 sets of datasets with 10 data differences in each set to obtain two medians; If the deviation between the two medians is less than the first adjustment value, the arithmetic mean of the two medians is used as the interpolation of the missing data, and the interpolation is marked with an interpolation mark; if the proportion of the data sources of the selected data at the same time in the recent 10 days that are interpolated exceeds the second adjustment value, then the data at the same time one day earlier is taken until the proportion of the data sources that are interpolated does not exceed the second adjustment value.

4. The IT device inspection report generation system based on a deep neural network according to claim 1, wherein, The generation grouping rule pairs and groups each multi-intelligence acquisition control module and high-speed data I / O module, determines the number in the group-numbering method, establishes a grouping index, and determines the groups in the group-numbering method, including: Use a deep neural network to perform feature engineering processing on the basic parameter information of all IT devices to obtain object features. Let S represent the number of object features. Determine X object features from the object features to screen all IT devices, and obtain IT devices containing X object features. Use K as the grouping width to form a data transmission group with K IT devices containing X object features; When the data transmission packet loss rate or data error rate of the deep neural network model module is higher than 5%, increase the value of K; when the number of IT devices containing X object features is greater than K, increase the value of X; Use the grouping width K and X, as well as the data transmission packet loss rate and data error rate as inputs, and use the pre-trained model RoBERTa to pre-train the deep neural network model, update the deep neural network model, and adjust the grouping width K and X.

5. The IT device inspection report generation system based on a deep neural network according to claim 1, characterized in that The deep neural network model module is used to dynamically elect a high-speed data I / O module; The dynamic election of the high-speed data I / O module includes: The deep neural network model module receives the operation key information of the high-speed data I / O module in each data transmission group. The operation key information includes throughput parameters, time-sharing throughput parameters, I / O rate parameters, online duration parameters, fault status parameters, aging status, and quality inspection approaching expiration status. According to the decision parameter U provided by the decision maker, adjust the weight Q1 values corresponding to each throughput parameter, time-sharing throughput parameter, I / O rate parameter, and online duration parameter respectively, and perform weighted summation to obtain the recommended value T; according to the decision parameter J provided by the decision maker, adjust the weight Q2 values corresponding to the fault status parameter, aging status, and quality inspection approaching expiration status respectively, and perform weighted summation to obtain the detection value M; when it is detected that the value of M is greater than 1, sort according to the size of the T value, recommend the high-speed data I / O module with the largest T value in the data transmission group for data transmission, and notify the multi-intelligence acquisition control module to adjust the I / O object.

6. The IT device inspection report generation system based on a deep neural network according to claim 5, wherein, The deep neural network of the deep neural network model module uses a multi-layer perceptron as the basic architecture, including an input layer, a hidden layer, and an output layer; calculate its weights Q1 and Q2 respectively through the input throughput parameters, time-sharing throughput parameters, I / O rate parameters, online duration parameters, fault status parameters, aging status, and quality inspection approaching expiration status, and use the mean square error as the loss function to evaluate the error between the calculated weights Q1 and Q2 and the true weights Q1 and Q2; update the weights and biases of the network according to the gradient of the loss function through the backpropagation method, and train the deep neural network model.

7. The IT device inspection report generation system based on a deep neural network according to claim 6, wherein, It also includes an inspection report generation module, which is used to generate an inspection report based on the collected key operation information of IT devices, a preset prompt word library, and the data in the expansion slot as input, based on multi-modal GPT technology; Manually annotate the collected key operation information of IT devices, train a labeling model with the data after manual annotation. The labeling model is based on the input collected key operation information of IT devices to obtain the collected key operation information of IT devices with labels. Combine the labels with the prompt words in the preset prompt word library, and input them into the multi-modal GPT model through the report generation rules provided by the expansion slot to generate an inspection report.

8. The IT device inspection report generation system based on a deep neural network according to claim 7, wherein, The generated inspection report presents a decision-making interface through a visualization interface, supports report preview, generation, push, and export and archiving, and is displayed through a statistical dashboard, desktop, mobile phone, and VR device terminal.

9. A method for generating an IT device inspection report based on a deep neural network, characterized in that, It includes: Set the multi-intelligent acquisition control module on the IT device to collect the basic parameter information of the IT device, and use the acquisition model to generate a control instruction set and an acquisition instruction set to control the multi-intelligent acquisition control module to collect the key operation information of the IT device; the multi-intelligent acquisition control module is communicatively connected to the deep neural network model module through a high-speed data I / O module with a device number identifier; One multi-intelligent acquisition control module and one high-speed data I / O module with a device number identifier form a data transmission chain, and multiple data transmission chains form a data transmission group; the high-speed data I / O modules in the data transmission group store all the data of the data transmission group, and the high-speed data I / O modules are numbered in the group-number mode; Use the deep neural network model module to generate an acquisition model; generate a grouping rule, pair and group each multi-intelligent acquisition control module and high-speed data I / O module, determine the number in the group-number mode, establish a grouping index, and determine the group in the group-number mode; and when there is a timing interruption in the collected key operation information, provide prediction data to the multi-intelligent acquisition control module to interpolate the key operation information.

10. The method for generating an IT device inspection report based on a deep neural network according to claim 9, wherein, The generation of the acquisition model includes: Collect the basic parameter information of each IT device and the operation control information of the IT device. The basic parameter information includes brand, name, model, version number, and manufacturer; the operation control information includes acquisition instructions, control instructions, data types, acquisition frequencies, instruction correlations, and failure rates; Construct the object features of the IT device according to the basic parameter information of the IT device. After inserting delimiters between the object features and the operation control information and each operation control information, vectorize and store them in the vector library to generate an acquisition model; The acquisition model vectorizes the collected basic parameter information of the IT device, calculates the similarity between the vectorized basic parameter information and the vectorized object features in the vector library using cosine similarity or Euclidean distance, and uses the operation control information corresponding to the vectorized object features in the vector library with the highest similarity as the operation control information of the IT device.

Images