CN-121722647-B - Security monitoring method, device, equipment and medium for control cabinet

CN121722647BCN 121722647 BCN121722647 BCN 121722647BCN-121722647-B

Abstract

The invention discloses a safety monitoring method, a device, equipment and a medium of a control cabinet, and relates to the field of safety monitoring of the control cabinet, wherein the method obtains an abnormal state prediction result of the control cabinet by acquiring multidimensional time sequence data of the current moment of the control cabinet and inputting a trained light-weight first abnormal state prediction model; and secondly, utilizing newer historical data, and transferring the multi-scale knowledge of the second abnormal state prediction model to the lightweight first abnormal state prediction model through comparison of knowledge distillation loss. According to the invention, the balance of the performance and the efficiency of the model is realized by optimizing the first comprehensive loss weight, the abnormal state label describing the overall running condition in the control cabinet is obtained, the transition from single-point threshold alarming to global state evaluation is realized, and the safety monitoring intelligent level of the control cabinet is improved to a certain extent.

Inventors

TIAN NINGNING
LIN FENG
SUN JIANWEI
FANG QINGXIANG

Assignees

埃斯凯(上海)电气科技股份有限公司

Dates

Publication Date: 20260508
Application Date: 20260227

Claims (9)

1. The safety monitoring method of the control cabinet is characterized by comprising the following steps of: Step 100, acquiring at least two groups of multidimensional time sequence monitoring data at the current moment in a control cabinet, wherein each group at least comprises three of temperature, current, vibration, humidity and smoke concentration; Step 200, at least two groups of multidimensional time sequence monitoring data at the current time in the control cabinet are input into a trained first abnormal state prediction model to obtain a first abnormal state result of the control cabinet; the training step of the trained first abnormal state prediction model comprises the following steps: Step 201, acquiring N1 groups of multidimensional time sequence monitoring data in a first historical period in the control cabinet, an abnormal state label corresponding to the N1 groups of multidimensional time sequence monitoring data, a first abnormal state prediction model initialized by parameters, and a trained second abnormal state prediction model, wherein N1 is a positive integer greater than 1; Step 202, using the second abnormal state prediction model as a teacher model and the first abnormal state prediction model initialized by the parameters as a student model, and inputting N1 groups of multi-dimensional time sequence monitoring data into the teacher model to obtain teacher multi-scale knowledge; step 203, the difference between the teacher multi-scale knowledge and the student multi-scale knowledge is minimized by comparing knowledge distillation loss, fine-grained knowledge of the teacher model is transferred to the student model to complete knowledge distillation, and the knowledge distillation loss updates parameters of the student model by using a back propagation algorithm to obtain a trained first abnormal state prediction model; The training step of the trained second abnormal state prediction model comprises the following steps: Step 2011, acquiring N2 groups of multidimensional time sequence monitoring data in a first historical period in the control cabinet, N2 groups of abnormal state labels corresponding to the multidimensional time sequence monitoring data, and a third abnormal state prediction model initialized by parameters, wherein N2 is a positive integer larger than 2; Step 2012, inputting the N2 groups of multi-dimensional time sequence monitoring data into the third abnormal state prediction model to obtain a third abnormal state result of the control cabinet; Step 2013, determining a first sub-loss, a second sub-loss and a third sub-loss of the third abnormal state prediction model according to the abnormal state label and the third abnormal state result; step 2014, determining a first weight of the first sub-loss, a second weight of the second sub-loss and a third weight of the third sub-loss in a preset set of weight combinations, wherein each weight combination comprises the weight of the first sub-loss, the weight of the second sub-loss and the weight of the third sub-loss; step 2015, determining a first comprehensive loss of the third abnormal state prediction model according to the first sub-loss, the second sub-loss, the third sub-loss, the first weight, the second weight and the third weight, and stopping model training when the first comprehensive loss meets a preset condition to obtain a trained second abnormal state prediction model.
2. The safety monitoring method of a control cabinet according to claim 1, wherein step 2014 comprises: in a preset set of M weight combinations, determining a third abnormal state result output by the third abnormal state prediction model after N2 groups of multi-dimensional time sequence monitoring data are input to the third abnormal state prediction model according to each weight combination, wherein M is a positive integer greater than 1; Counting the number of correctly predicted abnormal states, the number of misreported abnormal states and the number of missed abnormal states according to the third abnormal state result aiming at each weight combination; for each weight combination, determining the accuracy rate and recall rate of the third abnormal state result according to the correctly predicted abnormal state number, the misreported abnormal state number and the missed report abnormal state number; Determining a macroscopic F1 fraction of the third abnormal state result according to the accuracy rate and the recall rate for each weight combination; According to the high-low ordering of the M macroscopic F1 scores, selecting the weight combinations corresponding to the first M1 macroscopic F1 scores as weight combination candidate sets, and selecting a group of weight combinations in the weight combination candidate sets to obtain a first weight of a first sub-loss, a second weight of a second sub-loss and a third weight of a third sub-loss, wherein M1 is a positive integer larger than 1.
3. The method for monitoring safety of a control cabinet according to claim 2, wherein selecting the weight combinations corresponding to the first M1 macroscopic F1 scores as a weight combination candidate set according to the ranking of the M macroscopic F1 scores, and selecting a group of weight combinations in the weight combination candidate set to obtain a first weight of a first sub-loss, a second weight of a second sub-loss, and a third weight of a third sub-loss, comprises: For each weight combination, determining an index value of the third abnormal state result according to the correctly predicted abnormal state number, the misreported abnormal state number and the missed abnormal state number, wherein the index value comprises at least two of an accuracy rate, an accuracy rate and a recall rate; For each weight combination, determining an absolute error of the index value and the index threshold, a minimum absolute error of the index value and the index threshold and a maximum absolute error of the index value and the index threshold according to the index value and the preset index threshold of the third abnormal state result; Determining gray correlation degree according to the absolute error, the minimum absolute error, the maximum absolute error and a preset resolution coefficient aiming at each weight combination; And selecting a weight combination corresponding to the maximum gray correlation degree from M1 gray correlation degrees as an optimal weight combination to obtain a first weight of the first sub-loss, a second weight of the second sub-loss and a third weight of the third sub-loss.
4. The method according to claim 1, wherein determining the first and second sub-losses of the third abnormal state prediction model according to the abnormal state label and the third abnormal state result in step 2013 includes: Acquiring the total number of abnormal state labels output by the third abnormal state prediction model, the single-heat codes of the real abnormal state labels corresponding to the N2 groups of multi-dimensional time sequence monitoring data, the prediction probability of each abnormal state label output by the third abnormal state prediction model and the punishment weight of each abnormal state label; determining a first sub-loss of the third abnormal state prediction model according to the N2 groups of multi-dimensional time sequence monitoring data, the total number of abnormal state labels, the single thermal coding, the prediction probability and the penalty weight; And determining a second sub-loss of the third abnormal state prediction model according to the N2 groups of multi-dimensional time sequence monitoring data, the total number of abnormal state labels, the single thermal coding, the prediction probability and a preset focus parameter.
5. The method according to claim 1, wherein in step 2013, determining a third sub-loss of the third abnormal state prediction model according to the abnormal state label and the third abnormal state result includes: Acquiring a similar set of other data with the same abnormal state label as the N2 groups of the multi-dimensional time sequence monitoring data, a heterogeneous set of other data with different abnormal state labels as the N2 groups of the multi-dimensional time sequence monitoring data, and a feature vector corresponding to the N2 groups of the multi-dimensional time sequence monitoring data output by the third abnormal state prediction model; And determining a third sub-loss of the third abnormal state prediction model according to the N2 groups of multi-dimensional time sequence monitoring data, the same class set, the different class set, the characteristic vector and a preset temperature parameter.
6. The method for monitoring safety of a control cabinet according to claim 1, wherein the second abnormal state prediction model is deployed at a cloud server outside the control cabinet, and the first abnormal state prediction model is deployed at an edge server inside the control cabinet, and after step 200, the method further comprises: Acquiring N3 groups of multidimensional time sequence monitoring data, N3 groups of abnormal state labels corresponding to the multidimensional time sequence monitoring data, the first abnormal state prediction model and the second abnormal state prediction model in a second historical period in the control cabinet, wherein N3 is a positive integer greater than 1 after the first historical period; the method comprises the steps that an edge server obtains a first abnormal state result output by a first abnormal state prediction model, and the edge server sends the first abnormal state result to a cloud server; and the cloud server retrains the second abnormal state prediction model according to the first abnormal state result to obtain a trained second abnormal state prediction model, and returns to step 202.
7. A safety monitoring device of a control cabinet, characterized in that the safety monitoring device comprises: The data acquisition module is used for acquiring at least two groups of multidimensional time sequence monitoring data at the current moment in the control cabinet, wherein each group at least comprises three of temperature, current, vibration, humidity and smoke concentration; The abnormal result output module is used for inputting at least two groups of multidimensional time sequence monitoring data at the current moment in the control cabinet into the trained first abnormal state prediction model to obtain a first abnormal state result of the control cabinet; The submodule of the abnormal result output module comprises: The model and data preparation sub-module is used for acquiring N1 groups of multidimensional time sequence monitoring data in a first historical period in the control cabinet, N1 groups of abnormal state labels corresponding to the multidimensional time sequence monitoring data, a first abnormal state prediction model initialized by parameters and a trained second abnormal state prediction model, wherein N1 is a positive integer larger than 1; The multi-scale knowledge extraction sub-module is used for taking the second abnormal state prediction model as a teacher model, taking the first abnormal state prediction model initialized by the parameters as a student model, inputting N1 groups of multi-dimensional time sequence monitoring data into the teacher model to obtain teacher multi-scale knowledge, and inputting N1 groups of multi-dimensional time sequence monitoring data into the student model to obtain student multi-scale knowledge; the knowledge alignment and model updating sub-module is used for minimizing the difference between the teacher multi-scale knowledge and the student multi-scale knowledge by comparing knowledge distillation loss, transmitting fine-grained knowledge of the teacher model to the student model to finish knowledge distillation, and updating parameters of the student model by using a back propagation algorithm by using the knowledge distillation loss to obtain a trained first abnormal state prediction model; The model and data preparation submodule comprises: the model and data preparation unit is used for acquiring N2 groups of multidimensional time sequence monitoring data in a first historical period in the control cabinet, N2 groups of abnormal state labels corresponding to the multidimensional time sequence monitoring data and a third abnormal state prediction model initialized by parameters, wherein N2 is a positive integer larger than 2; A third abnormal state result obtaining unit, configured to input N2 groups of the multidimensional time sequence monitoring data into the third abnormal state prediction model, to obtain a third abnormal state result of the control cabinet; The first comprehensive loss calculation unit is used for determining a first sub-loss, a second sub-loss and a third sub-loss of the third abnormal state prediction model according to the abnormal state label and the third abnormal state result; A first weight set calculation unit, configured to determine, in a preset set of weight combinations, a first weight of the first sub-loss, a second weight of the second sub-loss, and a third weight of the third sub-loss, where each weight combination includes the weight of the first sub-loss, the weight of the second sub-loss, and the weight of the third sub-loss; The second abnormal state prediction model updating unit is configured to determine a first comprehensive loss of the third abnormal state prediction model according to the first sub-loss, the second sub-loss, the third sub-loss, the first weight, the second weight and the third weight, and stop model training when the first comprehensive loss meets a preset condition, so as to obtain a trained second abnormal state prediction model.
8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the method of safety monitoring of a control cabinet according to any one of claims 1 to 6 when executing the computer program.
9. A computer readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the safety monitoring method of the control cabinet according to any one of claims 1 to 6.

Description

Security monitoring method, device, equipment and medium for control cabinet Technical Field The invention relates to the field of safety monitoring of control cabinets, in particular to a safety monitoring method, device, medium and equipment of a control cabinet. Background The safety pre-warning method of the existing control cabinet generally adopts a single-parameter independent monitoring mode, namely, independent monitoring units are arranged aiming at key operation parameters (such as temperature, current, voltage, vibration, humidity, smoke concentration and the like) in the control cabinet, each monitoring unit detects corresponding parameters in real time or periodically based on a preset fixed threshold value, and when a certain parameter value exceeds the preset threshold value, a corresponding abnormal label is output and an alarm is triggered. Although the method realizes the monitoring of the operation state of the control cabinet to a certain extent, the multi-parameter coupling system is simplified into the monitoring of a plurality of independent variables, and a certain missing report problem exists. For example, when the preset temperature threshold is 60 ℃ and the voltage threshold is 240V, and the temperature reaches 55 ℃ and the voltage synchronously fluctuates to 235V due to poor heat dissipation or abnormal load of the control cabinet, the two parameters do not exceed the preset threshold, but the cooperative abnormal change of the two parameters clearly indicates that the control cabinet is in a transition state of critical failure. However, the conventional method can judge that the control cabinet operates normally because the parameter criteria are mutually independent, so that the optimal early warning time is missed, and thermal damage or electrical faults can be caused. Disclosure of Invention The embodiment of the invention provides a safety monitoring method, device, equipment and medium of a control cabinet, which are used for solving the problem that in the prior art, a single parameter independent monitoring mode is adopted to output abnormal states of single monitoring data, a multi-parameter coupling system is simplified to monitor a plurality of independent variables, and a certain report missing exists. In a first aspect, the present invention provides a safety monitoring method for a control cabinet, where the safety monitoring method for a control cabinet includes: Step 100, acquiring at least two groups of multidimensional time sequence monitoring data at the current moment in a control cabinet, wherein each group at least comprises three of temperature, current, vibration, humidity and smoke concentration; Step 200, at least two groups of multidimensional time sequence monitoring data at the current time in the control cabinet are input into a trained first abnormal state prediction model to obtain a first abnormal state result of the control cabinet; the training step of the trained first abnormal state prediction model comprises the following steps: Step 201, acquiring N1 groups of multidimensional time sequence monitoring data in a first historical period in the control cabinet, an abnormal state label corresponding to the N1 groups of multidimensional time sequence monitoring data, a first abnormal state prediction model initialized by parameters, and a trained second abnormal state prediction model, wherein N1 is a positive integer greater than 1; Step 202, using the second abnormal state prediction model as a teacher model and the first abnormal state prediction model initialized by the parameters as a student model, and inputting N1 groups of multi-dimensional time sequence monitoring data into the teacher model to obtain teacher multi-scale knowledge; And 203, minimizing the difference between the teacher multi-scale knowledge and the student multi-scale knowledge by comparing the knowledge distillation loss, transmitting the fine-grained knowledge of the teacher model to the student model to finish knowledge distillation, and updating the parameters of the student model by using a back propagation algorithm by using the knowledge distillation loss to obtain a trained first abnormal state prediction model. In a second aspect, the present invention provides a safety monitoring device for a control cabinet, the safety monitoring device for a control cabinet comprising: The data acquisition module is used for acquiring at least two groups of multidimensional time sequence monitoring data at the current moment in the control cabinet, wherein each group at least comprises three of temperature, current, vibration, humidity and smoke concentration; The abnormal result output module is used for inputting at least two groups of multidimensional time sequence monitoring data at the current moment in the control cabinet into the trained first abnormal state prediction model to obtain a first abnormal state result of the control cabinet; The submodule of the abnorma