An automatic detection and monitoring system for the operating state of an electrically controlled switch

Abstract

The application relates to an automatic detection and monitoring system for the running state of an electric control switch, in particular to the field of automatically detecting and monitoring the running state of an electric control switch, and realizes full-process automatic monitoring and intelligent operation and maintenance of the running state of the electric control switch through the cooperative work of multiple modules, can deeply integrate multi-source monitoring data, accurately evaluates the health state of the switch, and prospectively predicts the remaining service life and quantitatively predicts the uncertainty, on the basis of which, the system combines the prediction result with the criticality of the equipment, dynamically calculates the maintenance priority, and finally generates an adaptive dispatching scheme which takes into account the risk reduction and resource efficiency, so as to significantly improve the reliability, safety and intelligent operation and maintenance level of the power supply system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automatic detection and monitoring of the running state of electrically controlled switches, and more particularly, to an automatic detection and monitoring system for the running state of electrically controlled switches. BACKGROUND

[0002] With the continuous improvement of smart grid and industrial automation level, the running reliability of electrically controlled switches, as the key execution unit in power supply and distribution systems and power storage systems, is directly related to the stability and safety of the entire power grid. Such switch devices are widely used in power distribution reconstruction of old industrial plants, power scheduling of renewable energy power stations, and energy management of urban rail transit, and other key fields. In the long-term continuous operation process, the internal mechanical structure (such as the operating mechanism and the contact system) of the electrically controlled switch will be cumulatively worn out due to frequent on-off operation, and the external insulating material will gradually age due to factors such as environmental temperature and humidity fluctuations, electrical load impact, and chemical corrosion. In particular, in scenes with severe load fluctuations, such as arc furnace steel mills or photovoltaic power station grid-connected nodes, the switch action frequency is high, and the aging process is further accelerated. However, the degradation of switch performance is often a nonlinear, gradual, and hidden process, with no significant external signs in the early stage. Traditional maintenance methods that rely on fixed-period manual inspection or simple threshold alarms are difficult to capture the micro signs of performance degradation and cannot effectively warn of sudden failures, posing a serious challenge to continuous production and power supply.

[0003] Currently, the mainstream technology for monitoring and maintaining the operating state of electrically controlled switches still stays at the level of periodic preventive testing and post-maintenance. Specifically, periodic inspection relies on manual measurement of limited parameters such as loop resistance and insulation resistance of the switch, and its cycle is usually several months or even longer, which cannot reflect the state change of the switch between two inspections. Some advanced systems use online monitoring based on fixed thresholds, for example, real-time collection of switch contact temperature and alarm when exceeding the set limit. However, this static threshold method has poor adaptability and cannot distinguish between normal temperature rise caused by transient overload and abnormal overheating caused by poor contact, resulting in high false alarm rate. Moreover, it is essentially a post-alarm and has no predictive ability. To further improve monitoring accuracy, methods such as vibration acoustic analysis and infrared thermal imaging detection are introduced in existing technologies. Vibration analysis installs sensors on the switch body to collect vibration signal waveforms during switching process, and identifies mechanical abnormalities by comparing with standard waveform library. Infrared detection uses a thermal imager to non-contact scan the switch cabinet to locate overheating defects. However, these technologies usually require the installation of high-precision special sensors, which has high deployment and maintenance costs. Moreover, their data analysis relies on expert experience or simple pattern matching algorithms, making it difficult to establish an accurate quantitative relationship model between vibration characteristics, temperature field distribution and switch remaining service life. A more deep-seated technical problem is that the aging state of the switch is the result of the combined action of mechanical, electrical, and insulation factors. Single type of data (such as only temperature or only vibration) is difficult to fully and accurately represent the health status of the switch. Existing solutions generally lack effective fusion and analysis capabilities of multi-source heterogeneous data (such as current and voltage waveforms, operation count records, environmental temperature and humidity, partial discharge signals, and voiceprint features), and cannot construct a precise prediction model that reflects the nonlinear degradation law of the switch. Therefore, the development of preventive maintenance strategies lacks scientific data support, often resulting in insufficient maintenance leading to failure or excessive maintenance causing resource waste. Therefore, there is an urgent need for an automated monitoring technology that can predict the aging trend of electrically controlled switches at low cost and high accuracy and guide preventive maintenance. SUMMARY

[0004] The present application provides an automatic detection and monitoring system for the operating state of electrically controlled switches to solve the problems in the background art.

[0005] The technical solution of the present application to solve the above technical problems is as follows: specifically including: a data fusion module, a life prediction module, a risk quantification module, and a scheduling module, wherein,

[0006] The data fusion module automatically collects multi-source heterogeneous monitoring data, including electrical parameters, mechanical vibration signals, environmental indicators, and historical operation records generated during the operation of the electric control switch after power-on. Based on this multi-source heterogeneous monitoring data, it performs physical feature enhancement calculations to generate arc erosion equivalent and mechanical impact index. Then, it uses an attention weighting mechanism to adaptively weight and fuse the multi-source heterogeneous monitoring data, arc erosion equivalent, and mechanical impact index to generate a health status feature vector of a unified scale. After standardizing this vector, it is marked as a predefined health index vector data object along with the device identifier and timestamp and stored.

[0007] Lifespan prediction module: By calling the health index vector data object, the physical constraint deep learning model is triggered to perform training and inference. The physical constraint deep learning model uses a loss function containing physical law constraint terms to probabilistically predict the remaining lifespan of the electronic control switch and outputs a structured remaining lifespan distribution data object.

[0008] Risk quantification module: By calling the remaining service life distribution data object and combining it with the pre-stored switch criticality level and failure consequence cost data, the dynamic maintenance urgency index is calculated, and this calculation process is performed for each monitored electrical control switch to generate a set of dynamic maintenance urgency indices.

[0009] Scheduling module: By calling the dynamic maintenance urgency index set of all monitored electrical control switches, a multi-agent collaborative learning algorithm is used to adaptively schedule maintenance tasks with the maintenance team as the intelligent agent unit, generating a dynamic maintenance dispatching scheme.

[0010] In a preferred embodiment, the physical feature enhancement calculation in the data fusion module specifically includes arc erosion equivalent calculation and mechanical impact index calculation, wherein the specific process of arc erosion equivalent calculation is as follows:

[0011] Based on electrical parameters, the instantaneous current data sequence and the instantaneous voltage data sequence between contacts are acquired within the time window from the start of contact separation to the complete extinguishing of the arc during a single operation of the electric control switch.

[0012] The instantaneous power absolute value sequence is obtained by calculating the product of the instantaneous current absolute value and the instantaneous voltage absolute value at each sampling moment within the time window;

[0013] Numerical integration is performed on the instantaneous power absolute value sequence, and the integration result is used as the arc erosion equivalent value for this operation;

[0014] The specific process for calculating the mechanical impact index is as follows:

[0015] Based on the mechanical vibration signal, the vibration acceleration data sequence collected in the single operation process of the electric control switch is obtained;

[0016] The fast Fourier transform is performed on the vibration acceleration data sequence to obtain the frequency domain amplitude spectrum;

[0017] The frequency domain amplitude spectrum is multiplied element by element with a pre-generated characteristic frequency mask vector, wherein the characteristic frequency mask vector is set according to the inherent frequency distribution of the mechanical structure of the electric control switch; the value of the maximum element in the result vector obtained after the element-by-element multiplication operation is selected as the mechanical impact index value of this operation;

[0018] The enhanced data set output by the physical feature enhancement calculation includes electrical parameters, mechanical vibration signals, environmental indicators, historical operation records, arc erosion equivalents, and mechanical impact indexes.

[0019] In a preferred embodiment, the specific operation of adaptively weighted fusion using a multi-modal fusion algorithm based on attention weight is as follows:

[0020] First, each type of data in the enhanced data set is mapped to a unified dimensional feature space through a corresponding feature extractor, obtaining an initial feature vector set composed of electrical feature vectors, mechanical vibration feature vectors, environmental feature vectors, historical operation feature vectors, arc erosion equivalent feature vectors, and mechanical impact index feature vectors;

[0021] Subsequently, the initial feature vector set is input into a gated attention network, which first performs linear transformation on the initial feature vector set using a trainable feature transformation weight matrix and a feature transformation bias vector, then performs nonlinear processing on the linear transformation result using a hyperbolic tangent activation function, and then performs linear transformation again on the nonlinearly processed result using a trainable gating weight matrix and a gating bias vector to generate initial attention scores. Finally, the initial attention scores are processed using a normalization exponential function to generate an attention weight vector;

[0022] Finally, the weighted combination operation of the initial feature vector set and the attention weight vector is performed, that is, each initial feature vector is multiplied by its corresponding attention weight and then summed to generate a health state feature vector of a unified scale;

[0023] The health state feature vector is standardized to generate a standardized health state feature vector, wherein the standardization process includes calculating the arithmetic mean and standard deviation of the health state feature vector, and adjusting each element value of the health state feature vector based on the arithmetic mean and the standard deviation, so that the arithmetic mean of the adjusted vector is zero and the standard deviation is one;

[0024] The standardized health state feature vector is associated with the corresponding electrically controlled switch device identifier and the current timestamp, and is stored in the data storage area of the system and marked as a predefined health index vector data object.

[0025] In a preferred embodiment, the specific process of calling the health index vector data object in the life prediction module includes:

[0026] Firstly, the life prediction module retrieves the health index vector data object generated and stored by the data fusion module by accessing the real-time database;

[0027] Then, for an electrically controlled switch device, the life prediction module extracts the health state feature vectors corresponding to the device at multiple historical time points from the pre-stored database according to the chronological order of the timestamps, arranges them in chronological order, and constructs a continuous historical health state feature sequence;

[0028] The physical constraint deep learning model is built using an encoder-decoder architecture, where the encoder part uses a long short-term memory network structure that processes each health state feature vector in the historical health state feature sequence in turn and finally outputs a fixed-dimensional context vector;

[0029] The decoder part receives the context vector output by the encoder and maps the context vector to the predicted value of the remaining useful life through a network composed of multiple fully connected layers;

[0030] In the training process of the physical constraint deep learning model, the loss function used is composed of a data fitting term and a physical law constraint term through weighted summation, where the calculation process of the data fitting term is:

[0031] Using the quantile loss function, for all samples in the current training batch, the quantile loss between the true remaining useful life value of each sample and the predicted value of the physical constraint deep learning model at the predetermined quantile is calculated, and the average of the quantile loss results of all samples is calculated;

[0032] The calculation process of the physical law constraint term is: for the historical health state feature sequence, take the health state value inferred by the physical constraint deep learning model for each pair of adjacent time points, calculate the difference between the health state value at the next time point and the health state value at the previous time point, if the difference is positive, keep the difference, if the difference is negative or zero, take zero, and then sum all the difference calculation results of adjacent time points;

[0033] The weighted summation means that the average value of the data fitting term and the summation result of the physical law constraint term are multiplied by a hyperparameter weight and then added to obtain the total loss function;

[0034] Further, the trained physical constraint deep learning model is obtained.

[0035] In a preferred embodiment, the specific process of outputting a structured remaining useful life distribution data object comprises:

[0036] For the electrically controlled switching device to be predicted, the historical health state feature sequence of the electrically controlled switching device is obtained as a new input sequence by calling the health index vector data object; the trained physical constraint deep learning model is used to perform forward inference on the new input sequence, and a set of remaining useful life prediction values for a plurality of target quantiles are output by minimizing a total loss function comprising a quantile loss function and a physical constraint term, the set of remaining useful life prediction values for the plurality of target quantiles collectively constitute a remaining useful life distribution with uncertainty measurement, and the target quantiles include a 5th percentile, a 50th percentile and a 95th percentile.

[0037] Subsequently, the remaining useful life distribution with uncertainty measurement, the corresponding electrically controlled switching device identifier, the prediction timestamp and the version information of the physical constraint deep learning model are collectively encapsulated into a structured remaining useful life distribution data object.

[0038] In a preferred embodiment, the specific process of the risk quantification module comprises:

[0039] First, the remaining useful life distribution data object generated by the life prediction module is called, and the 50th percentile prediction value of the remaining useful life is extracted from the plurality of quantile prediction values contained in the remaining useful life distribution data object as a representation of the expected value of the remaining useful life, and the difference between the 95th percentile prediction value and the 50th percentile prediction value is calculated as a first representation of the prediction uncertainty.

[0040] Then, the criticality level parameter and the failure consequence cost parameter of the monitored electrically controlled switching device are read from the pre-stored database.

[0041] The calculation process of the dynamic maintenance urgency index is:

[0042] Based on the product of the criticality level parameter and the failure consequence cost parameter, divide the 50th percentile prediction value to obtain a basic urgency value, and multiply the basic urgency value by an uncertainty penalty factor to obtain the final dynamic maintenance urgency index.

[0043] In a preferred embodiment, the specific calculation process of the uncertainty penalty factor is as follows:

[0044] Calculate the ratio of the first representation of the prediction uncertainty and the 50th percentile prediction value, and obtain the relative uncertainty ratio;

[0045] Input the relative uncertainty ratio into the hyperbolic tangent function to obtain a transformed value between zero and one;

[0046] Multiply the transformed value by a predefined uncertainty penalty coefficient greater than zero to obtain a penalty weighted value;

[0047] Add the value one to the penalty weighted value to finally obtain the uncertainty penalty factor.

[0048] In a preferred embodiment, the specific process of performing the calculation process for each monitored electrically controlled switch to generate a dynamic maintenance urgency index set includes:

[0049] Traverse the list of unique device identifiers of all monitored electrically controlled switches; for each device identifier in the list, repeat the data call and index calculation process, that is, call the latest remaining useful life distribution data object corresponding to the device identifier, extract the corresponding percentile prediction value, read the pre-stored criticality level and failure consequence cost data associated with the device identifier, and calculate the dynamic maintenance urgency index value unique to the device according to the above data;

[0050] After completing the calculation of all device identifiers, associate each device identifier with its calculated dynamic maintenance urgency index value and calculation timestamp to form a data set containing multiple records;

[0051] Finally, encapsulate the data set as a dynamic maintenance urgency index set and store it.

[0052] In a preferred embodiment, the specific process of calling the dynamic maintenance urgency index set of all monitored electrically controlled switches in the scheduling module includes:

[0053] First, call the dynamic maintenance urgency index set generated by the risk quantification module, which contains the device identifier of each monitored electrically controlled switch and its corresponding dynamic maintenance urgency index value;

[0054] Then build the global environmental state of the maintenance scheduling problem, which contains the dynamic maintenance urgency index value set of all monitored electrically controlled switches, the current position information and remaining resource state of all maintenance team agents, and external dynamic environmental information including real-time traffic conditions and weather forecast data at each decision-making moment.

[0055] Each maintenance team agent makes decisions based on its local observations, which is a subset of the global environment state, including the dynamic maintenance urgency index values of the electrically controlled switches within a predetermined area near the agent's current location, the agent's own resource state, and the traffic information within the reachable range;

[0056] The adaptive scheduling of maintenance tasks is performed using a multi-agent collaborative learning algorithm, which adopts a centralized training and distributed execution framework. In the training phase, the central trainer uses global environment state information to collaboratively optimize the policy network of all maintenance team agents using a multi-agent proximal policy optimization algorithm. In the decision execution phase, each maintenance team agent independently selects the next electrically controlled switch station to be maintained using the trained policy network based only on its local observation information.

[0057] In a preferred embodiment, the multi-agent collaborative learning algorithm uses a composite reward function containing multiple reward components to guide the collaborative decision optimization of the maintenance team agents. The composite reward function includes a basic risk reduction reward component, a curriculum diversity reward component, a load balancing reward component, and a movement cost penalty component.

[0058] The basic risk reduction reward component is obtained by calculating the sum of the reduction amounts of the dynamic maintenance urgency index values of all electrically controlled switches that have completed maintenance operations within a single decision cycle.

[0059] The curriculum diversity reward component is obtained by calculating the information entropy value of the task type distribution of the electrically controlled switches completed in the current cycle and multiplying it by a weight coefficient.

[0060] The load balancing reward component is obtained by calculating the range of the workloads recently undertaken by all maintenance team agents and multiplying it by a negative weight coefficient.

[0061] The movement cost penalty component is obtained by calculating the total movement cost of all maintenance team agents within the decision cycle and multiplying it by a weight coefficient.

[0062] The process of generating a dynamic maintenance dispatching scheme is as follows: In each scheduling cycle, each maintenance team agent selects actions according to the trained policy network. The set of these action selections constitutes the dynamic maintenance dispatching scheme, and the scheme is packaged together with the generation timestamp and estimated effect as a dispatching scheme data object.

[0063] The beneficial effects of the present application are: through the cooperative work of multiple modules, the full-process automatic monitoring and intelligent operation and maintenance of the operation state of the electric control switch are realized, it can deeply integrate multi-source monitoring data, accurately evaluate the switch health state, and prospectively predict the remaining service life and quantitatively predict the uncertainty, on this basis, the system combines the prediction result with the key of the equipment, dynamically calculates the maintenance priority, and finally generates an adaptive dispatching scheme considering risk reduction and resource efficiency, thereby significantly improving the reliability, safety and operation intelligent level of the power supply system. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 The method flowchart of the present application is provided.

[0065] Figure 2 The system structure block diagram of the present application is provided. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0067] In the description of the present application, the terms "first", "second" are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0068] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or explanation". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given in order to enable any person skilled in the art to implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that those skilled in the art can realize the present application without using these specific details. In other examples, well-known structures and processes will not be described in detail to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the shown embodiments, but is consistent with the broadest scope of principles and features disclosed in the present application.

[0069] Embodiment 1

[0070] The present embodiment provides a method as described above Figures 1-2The illustrated automatic detection and monitoring system for the running state of an electric control switch specifically comprises a data fusion module, a life prediction module, a risk quantification module, and a scheduling module, wherein,

[0071] The data fusion module automatically collects multi-source heterogeneous monitoring data of electrical parameters, mechanical vibration signals, environmental indicators, and historical operation records generated during the running of the electric control switch after power-on; performs physical feature enhancement calculation based on the multi-source heterogeneous monitoring data to generate arc erosion equivalent and mechanical impact index; then uses an attention weight mechanism to adaptively weight and fuse the multi-source heterogeneous monitoring data, the arc erosion equivalent, and the mechanical impact index to generate a health state feature vector of a unified scale; after standardizing the vector, the vector is stored together with the device identifier and the time stamp as a predefined health index vector data object;

[0072] The life prediction module triggers a physical constraint deep learning model to train and infer by calling the health index vector data object, the physical constraint deep learning model probabilistically predicts the remaining useful life of the electric control switch through a loss function containing physical law constraint terms, and outputs a structured remaining useful life distribution data object containing the remaining useful life distribution information with uncertainty measurement;

[0073] The risk quantification module calculates a dynamic maintenance urgency index by calling the remaining useful life distribution data object and combining pre-stored switch criticality level and failure consequence cost data, the dynamic maintenance urgency index is positively correlated with the expected value of the remaining useful life and negatively correlated with the prediction standard deviation, which is used to quantify the maintenance urgency, and the calculation process is performed for each monitored electric control switch to generate a set of dynamic maintenance urgency indexes;

[0074] The scheduling module uses a multi-agent collaborative learning algorithm to adaptively schedule maintenance tasks for a maintenance team as an agent unit by calling the set of dynamic maintenance urgency indexes of all monitored electric control switches, and generates a dynamic maintenance dispatching scheme.

[0075] In this embodiment, it is specifically noted that in the data fusion module, the physical feature enhancement calculation specifically includes arc erosion equivalent calculation and mechanical impact index calculation, wherein the specific process of the arc erosion equivalent calculation is as follows:

[0076] Based on the electrical parameters, the instantaneous current data sequence and the instantaneous voltage data sequence collected within the time window from the contact separation to the complete extinction of the arc in the single operation process of the electric control switch are obtained;

[0077] The instantaneous power absolute value sequence is obtained by calculating the product of the instantaneous current absolute value and the instantaneous voltage absolute value at each sampling time within the time window;

[0078] Numerical integral operation is performed on the absolute value sequence of the instantaneous power, and the integral operation result is taken as the arc erosion equivalent value of this operation, which is used to quantify the cumulative ablation degree of the switch contact caused by arc energy;

[0079] The specific process of mechanical impact index calculation is as follows:

[0080] Based on the mechanical vibration signal, the vibration acceleration data sequence collected in the single operation process of the electric control switch is obtained;

[0081] Fast Fourier transform is performed on the vibration acceleration data sequence to obtain its frequency domain amplitude spectrum;

[0082] The frequency domain amplitude spectrum is multiplied element by element with a pre-generated characteristic frequency mask vector, wherein the characteristic frequency mask vector is set according to the inherent frequency distribution of the mechanical structure of the electric control switch, and has a non-zero weight at the inherent frequency and its adjacent frequency band, and the weight is zero at other frequency bands; The value of the maximum element in the result vector obtained after the element-by-element multiplication operation is selected as the mechanical impact index value of this operation, which is used to represent the mechanical stress intensity generated when the switch operating mechanism acts;

[0083] The enhanced data set output by the physical feature enhancement calculation includes electrical parameters, mechanical vibration signals, environmental indicators, historical operation records, arc erosion equivalent and mechanical impact index;

[0084] The specific operation of adaptive weighted fusion by using the multi-modal fusion algorithm based on attention weight is as follows:

[0085] Firstly, each type of data in the enhanced data set is mapped to a unified dimensional feature space through the corresponding feature extractor, and an initial feature vector set composed of electrical feature vector, mechanical vibration feature vector, environmental feature vector, historical operation feature vector, arc erosion equivalent feature vector and mechanical impact index feature vector is obtained;

[0086] Subsequently, the initial feature vector set is input into the gated attention network, which first performs linear transformation on the initial feature vector set using a trainable feature transformation weight matrix and a feature transformation bias vector, then performs nonlinear processing on the linear transformation result using the hyperbolic tangent activation function, and then performs linear transformation again on the nonlinearly processed result using a trainable gating weight matrix and a gating bias vector to generate initial attention scores, and finally applies the normalization exponential function to the initial attention scores to generate an attention weight vector, each weight value in the vector corresponding to the importance of an initial feature vector;

[0087] Finally, a weighted combination operation of the initial feature vector set and the attention weight vector is performed, i.e., each initial feature vector is multiplied by its corresponding attention weight and then summed, to generate a health state feature vector in a unified scale;

[0088] The health state feature vector is standardized to generate a standardized health state feature vector, wherein the standardization process includes calculating the arithmetic mean and standard deviation of the health state feature vector, and adjusting each element value of the health state feature vector based on the arithmetic mean and standard deviation, so that the adjusted vector has an arithmetic mean of zero and a standard deviation of one;

[0089] The standardized health state feature vector is associated with the corresponding electrically controlled switch device identifier and the current timestamp, and is stored in the data storage area of the system, and is marked as a predefined health index vector data object.

[0090] In this embodiment, it is specifically required to explain that the specific process of calling the health index vector data object in the life prediction module includes:

[0091] First, the life prediction module retrieves the health index vector data objects generated and stored by the data fusion module by accessing the real-time database or message queue, wherein each health index vector data object contains a health state feature vector after standardization, a uniquely associated electrically controlled switch device identifier, and a precisely recorded timestamp;

[0092] Then, for an electrically controlled switch device, the life prediction module extracts the health state feature vectors corresponding to the device at multiple historical time points from the pre-stored database according to the chronological order of the timestamps, arranges them in chronological order, and constructs a continuous historical health state feature sequence, which describes the evolution process of the device health state;

[0093] The physical constraint deep learning model is constructed using an encoder-decoder architecture, wherein the encoder part uses a long short-term memory network structure, which sequentially processes each health state feature vector in the historical health state feature sequence, captures long-term and short-term temporal dependencies in the sequence, and finally outputs a fixed-dimensional context vector that condenses the historical degradation information of the device;

[0094] The decoder part receives the context vector output by the encoder and maps the context vector to the predicted value of the remaining useful life through a network composed of multiple fully connected layers;

[0095] In the training process of the physical constraint deep learning model, the loss function used is composed of a data fitting term and a physical law constraint term through weighted summation, wherein the calculation process of the data fitting term is:

[0096] For all samples in the current training batch, the quantile loss between the true RUL value and the predicted value at the predetermined quantile of the physical constraint deep learning model is calculated for each sample using a quantile loss function, and the average of the quantile loss results of all samples is calculated.

[0097] The calculation process of the physical law constraint term is as follows: for the historical health state feature sequence, the difference between the health state value at the next time point and the health state value at the previous time point is calculated for each pair of adjacent time points inferred by the physical constraint deep learning model, if the difference is positive, the difference is retained, if the difference is negative or zero, zero is taken, and then the sum of all such difference calculation results of adjacent time points is calculated.

[0098] The weighted sum is to add the average value of the data fitting term and the sum of the physical law constraint term after multiplying a hyperparameter weight, to obtain the total loss function; the total loss function is used in the back propagation process of the physical constraint deep learning model training to optimize the model parameters of the physical constraint deep learning model, and the expression of the total loss function is:

[0099]

[0100] wherein, represents the total loss function, the goal of the physical constraint deep learning model training is to minimize this total value, N represents the number of samples in the training batch, i.e. the number of samples used for updating the model parameters of the physical constraint deep learning model at a time, i represents the sample index, taking integer values from 1 to N, used to traverse each sample in the current batch, represents the true RUL value of the i-th sample, which is a target value known based on historical data that needs to be approximated by the physical constraint deep learning model, represents the RUL value predicted by the physical constraint deep learning model for the i-th sample, represents the target quantile, which is a number between 0 and 1 (for example, 0.5 represents the median), used to specify the position of the probability distribution, represents the quantile loss function, which is calculated as wherein represents the indicator function, represents another symbolic representation of the quantile loss function, which is equivalent to represents the prediction error or residual, i.e. the difference between the true value and the predicted value, represents the specific indicator function application, when the condition in the subscript is ​​When the condition "(i.e., the actual value is less than the predicted value, indicating that the physical constraint deep learning model overestimated the actual lifetime)" holds true, OY=1; otherwise (i.e., the actual value is greater than or equal to the predicted value, indicating that the model underestimated or was accurate), OY=0. Its function is to measure the predicted value at the target quantile. Error at that point The weight hyperparameter representing the physical constraint term is a positive real number used to adjust the weight of the physical constraint term in the total loss, balancing the importance of data fitting and adherence to physical laws. This represents the length of the historical health state feature sequence input to the physical constraint deep learning model. This represents the index of a time point in the sequence, with values ​​from 1 to T-1, used to iterate through adjacent time points in the sequence. This represents the scalar value of the health status inferred by the physical constraint deep learning model at the j-th time point in the sequence. This value is obtained by mapping the intermediate states of the model through an auxiliary network and is used to characterize the health level of the device at that moment. This represents the monotonicity constraint term, which calculates the change in the health status estimate between adjacent time steps, and only applies to health status "improving" (i.e., Punishment will be imposed for the following situations. This represents the data fitting term, which is the average value of the quantile loss function. Its purpose is to force the predictions of the physically constrained deep learning model to be as close as possible to the actual observed data. By using quantile loss, the physically constrained deep learning model not only learns to predict the "most likely value" of RUL (such as the median), but also learns to predict its probability distribution (such as the confidence interval), thereby quantifying uncertainty. This represents the physical law constraint term, which is the sum of the monotonicity constraint terms. Its role is not to fit the data, but to embed the physical prior knowledge that "the health status of the equipment should not improve with use" as a constraint into the model. It penalizes any prediction that violates the monotonic degradation law (i.e., the health status estimate increases over time), thereby ensuring that the model output conforms to basic physical common sense and improving the rationality of the prediction.

[0101] This leads to the completed training of a physical constraint deep learning model;

[0102] The specific process of outputting a structured remaining useful life distribution data object includes:

[0103] For the electrically controlled switch device to be predicted, its historical health state feature sequence is obtained as a new input sequence by calling the health index vector data object; the trained physically constrained deep learning model is used to perform forward inference on the new input sequence, and a set of residual useful life prediction values for a plurality of target quantile numbers are output by minimizing the total loss function containing the quantile loss function and the physical constraint term, the set of residual useful life prediction values for the plurality of target quantile numbers together constitute a residual useful life distribution with uncertainty measurement, and the target quantile numbers include the fifth percentile, the fifth percentile and the ninety-fifth percentile, wherein the fifth percentile represents the median estimate of the residual useful life, i.e. the most likely value, and the fifth percentile and the ninety-fifth percentile represent the pessimistic estimate and the optimistic estimate of the residual useful life, respectively, which together constitute a ninety percent confidence interval for quantifying the uncertainty of the prediction.

[0104] Subsequently, the residual useful life distribution with uncertainty measurement, the corresponding electrically controlled switch device identifier, the prediction timestamp and the version information of the physically constrained deep learning model are collectively encapsulated into a structured residual useful life distribution data object; the data object is stored in the system persistent storage with clear labels, so that the risk quantification module can obtain the residual useful life distribution with uncertainty measurement by calling the residual useful life distribution data object, for subsequent risk assessment and maintenance decision analysis.

[0105] In this embodiment, it is specifically required to explain that the specific process of calling the residual useful life distribution data object in the risk quantification module combined with the pre-stored switch criticality level and failure consequence cost data includes:

[0106] First, the residual useful life distribution data object generated by the life prediction module is called, and the fifth percentile prediction value of the residual useful life is extracted as a representation of the expected value of the residual useful life from the plurality of quantile prediction values contained in the residual useful life distribution data object, and the difference between the ninety-fifth percentile prediction value and the fifth percentile prediction value is calculated as the first representation of the prediction uncertainty;

[0107] Then, the criticality level parameter and the failure consequence cost parameter of the monitored electrically controlled switch are read from the pre-stored database, wherein the criticality level parameter is determined according to the importance level of the load carried by the switch in the power supply network, and the failure consequence cost parameter is a dimensionless value quantified after comprehensively estimating the power loss, equipment maintenance cost and safety environmental impact;

[0108] The calculation process of the dynamic maintenance urgency index is:

[0109] A basic urgency value is obtained based on the product of the criticality level parameter and the failure consequence cost parameter, divided by the 50th percentile prediction value, and then the basic urgency value is multiplied by an uncertainty penalty factor to obtain the final dynamic maintenance urgency index; the calculation of the uncertainty penalty factor depends on the ratio of the first representation of prediction uncertainty to the 50th percentile prediction value, and a hyperbolic tangent function is applied to the ratio for nonlinear transformation;

[0110] The specific calculation process of the uncertainty penalty factor is as follows:

[0111] The ratio of the first representation of prediction uncertainty to the 50th percentile prediction value is calculated to obtain a relative uncertainty ratio;

[0112] The relative uncertainty ratio is input into the hyperbolic tangent function to obtain a transformed value between zero and one;

[0113] The transformed value is multiplied by a predefined uncertainty penalty coefficient greater than zero to obtain a penalty weighted value;

[0114] The value one is added to the penalty weighted value to finally obtain the uncertainty penalty factor; the dynamic maintenance urgency index is inversely proportional to the 50th percentile prediction value, and is positively correlated with the first representation of prediction uncertainty through the uncertainty penalty factor;

[0115] The specific process of performing this calculation process for each monitored electrically controlled switch to generate a set of dynamic maintenance urgency indexes includes:

[0116] Traverse the list of unique device identifiers of all monitored electrically controlled switches; for each device identifier in the list, repeat the data call and index calculation process, that is, call the latest remaining useful life distribution data object corresponding to the device identifier, extract the corresponding percentile prediction value, read the pre-stored criticality level and failure consequence cost data associated with the device identifier, and calculate the dynamic maintenance urgency index value unique to the device according to the data;

[0117] After the calculation of all device identifiers is completed, each device identifier is associated with its calculated dynamic maintenance urgency index value and a calculation timestamp to form a data set containing multiple records;

[0118] Finally, the data set is packaged as a dynamic maintenance urgency index set and stored.

[0119] In this embodiment, it is specifically required to explain that the specific process of calling the dynamic maintenance urgency index set of all monitored electrically controlled switches in the scheduling module includes:

[0120] Firstly, a set of dynamically maintained urgency indices generated by the risk quantification module is called, which contains the device identifier and its corresponding dynamically maintained urgency index value of each monitored electrically controlled switch;

[0121] Then, a global environment state of the maintenance scheduling problem is constructed, which contains the set of dynamically maintained urgency index values of all monitored electrically controlled switches, the current position information and remaining resource state of all maintenance team agents, and external dynamic environment information including real-time traffic map and weather forecast data at each decision-making moment;

[0122] Each maintenance team agent makes decisions based on its local observation, which is a subset of the global environment state, including the dynamically maintained urgency index values of electrically controlled switches within a predetermined area near the agent's current position, the agent's own resource state, and the traffic information within its reachable range;

[0123] The multi-agent collaborative learning algorithm is used for adaptive scheduling of maintenance tasks, which adopts a centralized training and distributed execution framework. In the training phase, the central trainer uses global environment state information to collaboratively optimize the policy network of all maintenance team agents through the multi-agent proximal policy optimization algorithm. In the decision execution phase, each maintenance team agent only uses the trained policy network to independently select the next electrically controlled switch station to be maintained based on its local observation information;

[0124] The multi-agent collaborative learning algorithm uses a composite reward function containing multiple reward components to guide the collaborative decision optimization of maintenance team agents. The composite reward function includes a basic risk reduction reward component, a curriculum diversity reward component, a load balancing reward component, and a movement cost penalty component;

[0125] The basic risk reduction reward component is obtained by calculating the sum of the reduction amount of the dynamically maintained urgency index values of all electrically controlled switches that have completed maintenance operations in a single decision-making period, to directly encourage agents to prioritize high-urgency tasks;

[0126] The curriculum diversity reward component is obtained by calculating the information entropy value of the task type distribution of the electrically controlled switches completed in the current period and multiplying it by a weight coefficient, to encourage agents to explore multiple task types and avoid local optimal strategies in the early training stage;

[0127] The load balancing reward component is obtained by calculating the range of workloads recently undertaken by all maintenance team agents and multiplying it by a negative weight coefficient, to punish uneven workloads among agents and promote load balancing;

[0128] The moving cost penalty component is obtained by calculating the total moving cost of all maintenance team agents in a decision cycle and multiplying it by a weight coefficient, so as to optimize resource utilization efficiency;

[0129] The process of generating the dynamic maintenance dispatching scheme is as follows: in each scheduling cycle, each maintenance team agent selects an action according to the trained strategy network, and the set of these action selections constitutes the dynamic maintenance dispatching scheme, and the scheme is encapsulated into a dispatching scheme data object together with a generation timestamp and an estimated effect.

[0130] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0131] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0132] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks. Figure 1 The means for performing the functions specified in one or more flows and / or blocks.

[0133] These computer program instructions can also be stored in a computer-readable memory that can guide the computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks. Figure 1 The means for performing the functions specified in one or more flows and / or blocks.

[0134] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0135] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0136] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An automated detection and monitoring system for the operating status of an electronically controlled switch, characterized in that, Specifically, it includes: The module includes a data fusion module, a lifetime prediction module, a risk quantification module, and a scheduling module. The data fusion module automatically collects multi-source heterogeneous monitoring data, including electrical parameters, mechanical vibration signals, environmental indicators, and historical operation records generated during the operation of the electric control switch after power-on. Based on this multi-source heterogeneous monitoring data, it performs physical feature enhancement calculations to generate arc erosion equivalent and mechanical impact index. Then, it uses an attention weighting mechanism to adaptively weight and fuse the multi-source heterogeneous monitoring data, arc erosion equivalent, and mechanical impact index to generate a health status feature vector of a unified scale. After standardizing this vector, it is marked as a predefined health index vector data object along with the device identifier and timestamp and stored. Lifespan prediction module: By calling the health index vector data object, the physical constraint deep learning model is triggered to perform training and inference. The physical constraint deep learning model uses a loss function containing physical law constraint terms to probabilistically predict the remaining lifespan of the electronic control switch and outputs a structured remaining lifespan distribution data object. Risk quantification module: By calling the remaining service life distribution data object and combining it with the pre-stored switch criticality level and failure consequence cost data, the dynamic maintenance urgency index is calculated, and this calculation process is performed for each monitored electrical control switch to generate a set of dynamic maintenance urgency indices. Scheduling module: By calling the dynamic maintenance urgency index set of all monitored electrical control switches, a multi-agent collaborative learning algorithm is adopted to adaptively schedule maintenance tasks with the maintenance team as the intelligent agent unit, and a dynamic maintenance dispatching scheme is generated.

2. The automated detection and monitoring system for the operating status of an electronically controlled switch according to claim 1, characterized in that: In the data fusion module, the physical feature enhancement calculation specifically includes the calculation of arc erosion equivalent and the calculation of mechanical impact index, wherein the specific process of the arc erosion equivalent calculation is as follows: Based on electrical parameters, the instantaneous current data sequence and the instantaneous voltage data sequence between contacts are collected within the time window from the start of contact separation to the complete extinction of the arc during a single operation of the electric control switch. The instantaneous power absolute value sequence is obtained by calculating the product of the instantaneous current absolute value and the instantaneous voltage absolute value at each sampling moment within the time window; Numerical integration is performed on the instantaneous power absolute value sequence, and the integration result is used as the arc erosion equivalent value for this operation; The specific process for calculating the mechanical impact index is as follows: Based on mechanical vibration signals, the vibration acceleration data sequence collected during a single operation of an electronically controlled switch is obtained; Perform a fast Fourier transform on the vibration acceleration data sequence to obtain its frequency domain amplitude spectrum; The frequency domain amplitude spectrum is multiplied element-wise with a pre-generated characteristic frequency mask vector, wherein the characteristic frequency mask vector is set according to the inherent frequency distribution of the electronically controlled switch mechanical structure; in the result vector obtained after the element-wise multiplication operation, the value of its largest element is selected as the mechanical impact index value of this operation; The enhanced data set output by the physical feature enhancement calculation includes electrical parameters, mechanical vibration signals, environmental indicators, historical operation records, arc erosion equivalent, and mechanical impact index.

3. The automated detection and monitoring system for the operating status of an electronically controlled switch according to claim 2, characterized in that: The specific operation of adaptive weighted fusion using a multimodal fusion algorithm based on attention weights is as follows: First, the various types of data in the enhanced dataset are mapped to a feature space of a unified dimension through the corresponding feature extractors to obtain an initial feature vector set consisting of electrical feature vectors, mechanical vibration feature vectors, environmental feature vectors, historical operation feature vectors, arc erosion equivalent feature vectors, and mechanical impact index feature vectors. Subsequently, the initial feature vector set is input into a gated attention network. This gated attention network first performs a linear transformation on the initial feature vector set using a trainable feature transformation weight matrix and a feature transformation bias vector. Then, it performs nonlinear processing on the linear transformation result using a hyperbolic tangent activation function. Next, it performs a linear transformation on the nonlinearly processed result again using a trainable gate weight matrix and a gate bias vector to generate an initial attention score. Finally, it applies a normalized exponential function to the initial attention score to generate an attention weight vector. Finally, a weighted combination operation is performed on the initial feature vector set and the attention weight vector, that is, each initial feature vector is multiplied by its corresponding attention weight and then summed to generate a health status feature vector of uniform scale. The health status feature vector is standardized to generate a standardized health status feature vector. The standardization process includes calculating the arithmetic mean and standard deviation of the health status feature vector, and adjusting the value of each element of the health status feature vector based on the arithmetic mean and standard deviation, so that the arithmetic mean of the adjusted vector is zero and the standard deviation is one. The standardized health status feature vector is associated with the corresponding electronic control switch device identifier and the current timestamp, and stored in the system's data storage area, and marked as a predefined health index vector data object.

4. The automated detection and monitoring system for the operating status of an electronically controlled switch according to claim 3, characterized in that: The specific process of calling the health index vector data object in the lifespan prediction module includes: First, the lifespan prediction module accesses the real-time database to retrieve health index vector data objects generated and stored by the data fusion module. Then, for an electronically controlled switchgear, the life prediction module extracts the health status feature vectors of the device at multiple historical time points from the pre-stored database according to the order of timestamps, and arranges them in chronological order to construct a continuous historical health status feature sequence. The physical constraint deep learning model is constructed using an encoder-decoder architecture, where the encoder part uses a long short-term memory network structure. This network processes each health state feature vector in the historical health state feature sequence in sequence and finally outputs a fixed-dimensional context vector. The decoder receives the context vector output by the encoder and maps the context vector to a predicted value of the remaining lifetime through a network consisting of multiple fully connected layers. In the training process of a physical constraint deep learning model, the loss function used is composed of a weighted sum of a data fitting term and a physical constraint term. The calculation process of the data fitting term is as follows: The quantile loss function is used to calculate the quantile loss between the actual remaining lifetime value of each sample and the predicted value of the physical constraint deep learning model at a predetermined quantile for all samples in the current training batch, and the average value of the quantile loss results for all samples is calculated. The calculation process of the physical law constraint term is as follows: For the historical health status feature sequence, take the health status value of each pair of adjacent time points inferred by the physical constraint deep learning model, calculate the difference between the health status value of the next time point and the health status value of the previous time point. If the difference is positive, retain the difference; if it is negative or zero, take zero. Then sum the results of such difference calculations for all adjacent time points. The weighted summation refers to multiplying the average value of the data fitting terms and the summation result of the physical law constraint terms by a hyperparameter weight, and then adding them together to obtain the total loss function; This leads to the completed training of a physical constraint deep learning model.

5. The automated detection and monitoring system for the operating status of an electronically controlled switch according to claim 4, characterized in that: The specific process of outputting a structured remaining useful life distribution data object includes: For the electrical control switchgear to be predicted, its historical health status feature sequence is obtained by calling the health index vector data object as a new input sequence; the trained physical constraint deep learning model is used to perform forward inference on the new input sequence, and by minimizing the total loss function including quantile loss function and physical constraint term, a set of remaining useful life prediction values ​​for multiple pre-set target quantiles is output. The set of remaining useful life prediction values ​​for multiple pre-set target quantiles together constitutes the remaining useful life distribution with uncertainty measure. These target quantiles include the 5th percentile, the 50th percentile, and the 95th percentile. Subsequently, the remaining useful life distribution with uncertainty metric, the corresponding electronic control switch equipment identifier, the prediction timestamp, and the version information of the physical constraint deep learning model are encapsulated together into a structured remaining useful life distribution data object.

6. The automated detection and monitoring system for the operating status of an electronically controlled switch according to claim 5, characterized in that: The specific process of the risk quantification module, which combines the remaining useful life distribution data object with pre-stored switch criticality level and failure consequence cost data, includes: First, the remaining useful life distribution data object generated by the useful life prediction module is called. From the multiple quantile prediction values ​​contained in the remaining useful life distribution data object, the 50th percentile prediction value of the remaining useful life is extracted as a representation of the expected value of the remaining useful life. The difference between the 95th percentile prediction value and the 50th percentile prediction value is calculated as the first representation of the prediction uncertainty. Then, the critical level parameters and fault consequence cost parameters of the monitored electronic control switch are read from the pre-stored database; The calculation process for the dynamic maintenance urgency index is as follows: The product of the criticality level parameter and the failure consequence cost parameter is divided by the 50th percentile predicted value to obtain a basic urgency value. The basic urgency value is then multiplied by an uncertainty penalty factor to obtain the final dynamic maintenance urgency index.

7. The automated detection and monitoring system for the operating status of an electronically controlled switch according to claim 6, characterized in that: The specific calculation process for the uncertainty penalty factor is as follows: The ratio of the first characteristic of the prediction uncertainty to the 50th percentile predicted value is calculated to obtain the relative uncertainty ratio; The relative uncertainty ratio is input into the hyperbolic tangent function to obtain a transformation value between zero and one. The transformed value is multiplied by a predefined uncertainty penalty coefficient that is greater than zero to obtain a penalty weighted value; The value is added to the penalty weighting value to obtain the uncertainty penalty factor.

8. The automated detection and monitoring system for the operating status of an electronically controlled switch according to claim 7, characterized in that: The specific process of performing this calculation for each monitored electronic switch to generate a set of dynamic maintenance urgency indices includes: Iterate through the list of unique device identifiers for all monitored electrical control switches; for each device identifier in the list, repeatedly execute the data retrieval and index calculation process, that is, call the latest remaining service life distribution data object corresponding to the device identifier, extract the corresponding percentile prediction value, read the pre-stored criticality level and failure consequence cost data associated with the device identifier, and calculate the unique dynamic maintenance urgency index value of the device. After calculating all device identifiers, each device identifier is associated with and bound to its calculated dynamic maintenance urgency index value and calculation timestamp, forming a data set containing multiple records. Finally, the dataset is encapsulated into a dynamically maintained set of urgency indices and stored.

9. The automated detection and monitoring system for the operating status of an electronically controlled switch according to claim 8, characterized in that: The specific process of calling the dynamic maintenance urgency index set of all monitored electrical control switches in the scheduling module includes: First, the dynamic maintenance urgency index set generated by the risk quantification module is called. This dynamic maintenance urgency index set contains the device identifier of each monitored electrical control switch and its corresponding dynamic maintenance urgency index value. Then, a global environment state for the maintenance scheduling problem is constructed. This global environment state includes, at each decision moment, the set of dynamic maintenance urgency index values ​​of all monitored electronic switches, the current location information and remaining resource status of all maintenance team agents, and external dynamic environment information including real-time traffic maps and weather forecast data. Each maintenance team agent makes decisions based on its local observations, which are a subset of the global environmental state, including the dynamic maintenance urgency index of the electronic switches in a predetermined area near the agent's current location, the agent's own resource status, and traffic information within reach. The adaptive scheduling of maintenance tasks using a multi-agent collaborative learning algorithm adopts a framework of centralized training and distributed execution. During the training phase, the central trainer uses global environmental state information and a multi-agent proximal policy optimization algorithm to collaboratively optimize the policy network of all maintenance team agents. During the decision-making and execution phase, each maintenance team agent independently selects the next electrical control switch site to be maintained based solely on its local observation information and using the trained policy network.

10. The automated detection and monitoring system for the operating status of an electronically controlled switch according to claim 9, characterized in that: The multi-agent collaborative learning algorithm uses a composite reward function containing multiple reward components to guide the collaborative decision optimization of the maintenance team agents. The composite reward function includes a basic risk reduction reward component, a course diversity reward component, a load balancing reward component, and a movement cost penalty component. The basic risk reduction reward is calculated by summing the reduction in the dynamic maintenance urgency index of all electronically controlled switches that have completed maintenance operations within a single decision cycle. The course diversity reward component is obtained by calculating the information entropy value of the distribution of the types of electrical control switch maintenance tasks completed in the current cycle and multiplying it by a weighting coefficient. The load balancing reward component is obtained by calculating the range of the recent workload undertaken by all maintenance team agents and multiplying it by a negative weighting coefficient. The movement cost penalty component is obtained by calculating the total movement cost of all maintenance team agents during the decision-making cycle and multiplying it by a weighting coefficient; The process of generating a dynamic maintenance dispatch plan is as follows: In each scheduling cycle, each maintenance team agent outputs action selections according to the trained policy network. The set of these action selections constitutes a dynamic maintenance dispatch plan, and the plan, together with the generation timestamp and the estimated effect, is encapsulated into a dispatch plan data object.

Images